linux/kernel/bpf/hashtab.c

2643 lines
70 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-only
/* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
* Copyright (c) 2016 Facebook
*/
#include <linux/bpf.h>
#include <linux/btf.h>
#include <linux/jhash.h>
#include <linux/filter.h>
#include <linux/rculist_nulls.h>
#include <linux/rcupdate_wait.h>
#include <linux/random.h>
#include <uapi/linux/btf.h>
bpf: Introduce sleepable BPF programs Introduce sleepable BPF programs that can request such property for themselves via BPF_F_SLEEPABLE flag at program load time. In such case they will be able to use helpers like bpf_copy_from_user() that might sleep. At present only fentry/fexit/fmod_ret and lsm programs can request to be sleepable and only when they are attached to kernel functions that are known to allow sleeping. The non-sleepable programs are relying on implicit rcu_read_lock() and migrate_disable() to protect life time of programs, maps that they use and per-cpu kernel structures used to pass info between bpf programs and the kernel. The sleepable programs cannot be enclosed into rcu_read_lock(). migrate_disable() maps to preempt_disable() in non-RT kernels, so the progs should not be enclosed in migrate_disable() as well. Therefore rcu_read_lock_trace is used to protect the life time of sleepable progs. There are many networking and tracing program types. In many cases the 'struct bpf_prog *' pointer itself is rcu protected within some other kernel data structure and the kernel code is using rcu_dereference() to load that program pointer and call BPF_PROG_RUN() on it. All these cases are not touched. Instead sleepable bpf programs are allowed with bpf trampoline only. The program pointers are hard-coded into generated assembly of bpf trampoline and synchronize_rcu_tasks_trace() is used to protect the life time of the program. The same trampoline can hold both sleepable and non-sleepable progs. When rcu_read_lock_trace is held it means that some sleepable bpf program is running from bpf trampoline. Those programs can use bpf arrays and preallocated hash/lru maps. These map types are waiting on programs to complete via synchronize_rcu_tasks_trace(); Updates to trampoline now has to do synchronize_rcu_tasks_trace() and synchronize_rcu_tasks() to wait for sleepable progs to finish and for trampoline assembly to finish. This is the first step of introducing sleepable progs. Eventually dynamically allocated hash maps can be allowed and networking program types can become sleepable too. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Reviewed-by: Josef Bacik <josef@toxicpanda.com> Acked-by: Andrii Nakryiko <andriin@fb.com> Acked-by: KP Singh <kpsingh@google.com> Link: https://lore.kernel.org/bpf/20200827220114.69225-3-alexei.starovoitov@gmail.com
2020-08-27 22:01:11 +00:00
#include <linux/rcupdate_trace.h>
#include <linux/btf_ids.h>
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
#include "percpu_freelist.h"
#include "bpf_lru_list.h"
#include "map_in_map.h"
#include <linux/bpf_mem_alloc.h>
#define HTAB_CREATE_FLAG_MASK \
(BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE | \
BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
bpf: Allow selecting numa node during map creation The current map creation API does not allow to provide the numa-node preference. The memory usually comes from where the map-creation-process is running. The performance is not ideal if the bpf_prog is known to always run in a numa node different from the map-creation-process. One of the use case is sharding on CPU to different LRU maps (i.e. an array of LRU maps). Here is the test result of map_perf_test on the INNER_LRU_HASH_PREALLOC test if we force the lru map used by CPU0 to be allocated from a remote numa node: [ The machine has 20 cores. CPU0-9 at node 0. CPU10-19 at node 1 ] ># taskset -c 10 ./map_perf_test 512 8 1260000 8000000 5:inner_lru_hash_map_perf pre-alloc 1628380 events per sec 4:inner_lru_hash_map_perf pre-alloc 1626396 events per sec 3:inner_lru_hash_map_perf pre-alloc 1626144 events per sec 6:inner_lru_hash_map_perf pre-alloc 1621657 events per sec 2:inner_lru_hash_map_perf pre-alloc 1621534 events per sec 1:inner_lru_hash_map_perf pre-alloc 1620292 events per sec 7:inner_lru_hash_map_perf pre-alloc 1613305 events per sec 0:inner_lru_hash_map_perf pre-alloc 1239150 events per sec #<<< After specifying numa node: ># taskset -c 10 ./map_perf_test 512 8 1260000 8000000 5:inner_lru_hash_map_perf pre-alloc 1629627 events per sec 3:inner_lru_hash_map_perf pre-alloc 1628057 events per sec 1:inner_lru_hash_map_perf pre-alloc 1623054 events per sec 6:inner_lru_hash_map_perf pre-alloc 1616033 events per sec 2:inner_lru_hash_map_perf pre-alloc 1614630 events per sec 4:inner_lru_hash_map_perf pre-alloc 1612651 events per sec 7:inner_lru_hash_map_perf pre-alloc 1609337 events per sec 0:inner_lru_hash_map_perf pre-alloc 1619340 events per sec #<<< This patch adds one field, numa_node, to the bpf_attr. Since numa node 0 is a valid node, a new flag BPF_F_NUMA_NODE is also added. The numa_node field is honored if and only if the BPF_F_NUMA_NODE flag is set. Numa node selection is not supported for percpu map. This patch does not change all the kmalloc. F.e. 'htab = kzalloc()' is not changed since the object is small enough to stay in the cache. Signed-off-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-18 18:28:00 +00:00
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
#define BATCH_OPS(_name) \
.map_lookup_batch = \
_name##_map_lookup_batch, \
.map_lookup_and_delete_batch = \
_name##_map_lookup_and_delete_batch, \
.map_update_batch = \
generic_map_update_batch, \
.map_delete_batch = \
generic_map_delete_batch
/*
* The bucket lock has two protection scopes:
*
* 1) Serializing concurrent operations from BPF programs on different
* CPUs
*
* 2) Serializing concurrent operations from BPF programs and sys_bpf()
*
* BPF programs can execute in any context including perf, kprobes and
* tracing. As there are almost no limits where perf, kprobes and tracing
* can be invoked from the lock operations need to be protected against
* deadlocks. Deadlocks can be caused by recursion and by an invocation in
* the lock held section when functions which acquire this lock are invoked
* from sys_bpf(). BPF recursion is prevented by incrementing the per CPU
* variable bpf_prog_active, which prevents BPF programs attached to perf
* events, kprobes and tracing to be invoked before the prior invocation
* from one of these contexts completed. sys_bpf() uses the same mechanism
* by pinning the task to the current CPU and incrementing the recursion
* protection across the map operation.
bpf: Prepare hashtab locking for PREEMPT_RT PREEMPT_RT forbids certain operations like memory allocations (even with GFP_ATOMIC) from atomic contexts. This is required because even with GFP_ATOMIC the memory allocator calls into code pathes which acquire locks with long held lock sections. To ensure the deterministic behaviour these locks are regular spinlocks, which are converted to 'sleepable' spinlocks on RT. The only true atomic contexts on an RT kernel are the low level hardware handling, scheduling, low level interrupt handling, NMIs etc. None of these contexts should ever do memory allocations. As regular device interrupt handlers and soft interrupts are forced into thread context, the existing code which does spin_lock*(); alloc(GPF_ATOMIC); spin_unlock*(); just works. In theory the BPF locks could be converted to regular spinlocks as well, but the bucket locks and percpu_freelist locks can be taken from arbitrary contexts (perf, kprobes, tracepoints) which are required to be atomic contexts even on RT. These mechanisms require preallocated maps, so there is no need to invoke memory allocations within the lock held sections. BPF maps which need dynamic allocation are only used from (forced) thread context on RT and can therefore use regular spinlocks which in turn allows to invoke memory allocations from the lock held section. To achieve this make the hash bucket lock a union of a raw and a regular spinlock and initialize and lock/unlock either the raw spinlock for preallocated maps or the regular variant for maps which require memory allocations. On a non RT kernel this distinction is neither possible nor required. spinlock maps to raw_spinlock and the extra code and conditional is optimized out by the compiler. No functional change. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200224145644.509685912@linutronix.de
2020-02-24 14:01:51 +00:00
*
* This has subtle implications on PREEMPT_RT. PREEMPT_RT forbids certain
* operations like memory allocations (even with GFP_ATOMIC) from atomic
* contexts. This is required because even with GFP_ATOMIC the memory
* allocator calls into code paths which acquire locks with long held lock
bpf: Prepare hashtab locking for PREEMPT_RT PREEMPT_RT forbids certain operations like memory allocations (even with GFP_ATOMIC) from atomic contexts. This is required because even with GFP_ATOMIC the memory allocator calls into code pathes which acquire locks with long held lock sections. To ensure the deterministic behaviour these locks are regular spinlocks, which are converted to 'sleepable' spinlocks on RT. The only true atomic contexts on an RT kernel are the low level hardware handling, scheduling, low level interrupt handling, NMIs etc. None of these contexts should ever do memory allocations. As regular device interrupt handlers and soft interrupts are forced into thread context, the existing code which does spin_lock*(); alloc(GPF_ATOMIC); spin_unlock*(); just works. In theory the BPF locks could be converted to regular spinlocks as well, but the bucket locks and percpu_freelist locks can be taken from arbitrary contexts (perf, kprobes, tracepoints) which are required to be atomic contexts even on RT. These mechanisms require preallocated maps, so there is no need to invoke memory allocations within the lock held sections. BPF maps which need dynamic allocation are only used from (forced) thread context on RT and can therefore use regular spinlocks which in turn allows to invoke memory allocations from the lock held section. To achieve this make the hash bucket lock a union of a raw and a regular spinlock and initialize and lock/unlock either the raw spinlock for preallocated maps or the regular variant for maps which require memory allocations. On a non RT kernel this distinction is neither possible nor required. spinlock maps to raw_spinlock and the extra code and conditional is optimized out by the compiler. No functional change. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200224145644.509685912@linutronix.de
2020-02-24 14:01:51 +00:00
* sections. To ensure the deterministic behaviour these locks are regular
* spinlocks, which are converted to 'sleepable' spinlocks on RT. The only
* true atomic contexts on an RT kernel are the low level hardware
* handling, scheduling, low level interrupt handling, NMIs etc. None of
* these contexts should ever do memory allocations.
*
* As regular device interrupt handlers and soft interrupts are forced into
* thread context, the existing code which does
bpf: Make non-preallocated allocation low priority GFP_ATOMIC doesn't cooperate well with memcg pressure so far, especially if we allocate too much GFP_ATOMIC memory. For example, when we set the memcg limit to limit a non-preallocated bpf memory, the GFP_ATOMIC can easily break the memcg limit by force charge. So it is very dangerous to use GFP_ATOMIC in non-preallocated case. One way to make it safe is to remove __GFP_HIGH from GFP_ATOMIC, IOW, use (__GFP_ATOMIC | __GFP_KSWAPD_RECLAIM) instead, then it will be limited if we allocate too much memory. There's a plan to completely remove __GFP_ATOMIC in the mm side[1], so let's use GFP_NOWAIT instead. We introduced BPF_F_NO_PREALLOC is because full map pre-allocation is too memory expensive for some cases. That means removing __GFP_HIGH doesn't break the rule of BPF_F_NO_PREALLOC, but has the same goal with it-avoiding issues caused by too much memory. So let's remove it. This fix can also apply to other run-time allocations, for example, the allocation in lpm trie, local storage and devmap. So let fix it consistently over the bpf code It also fixes a typo in the comment. [1]. https://lore.kernel.org/linux-mm/163712397076.13692.4727608274002939094@noble.neil.brown.name/ Cc: Roman Gushchin <roman.gushchin@linux.dev> Cc: Shakeel Butt <shakeelb@google.com> Cc: NeilBrown <neilb@suse.de> Signed-off-by: Yafang Shao <laoar.shao@gmail.com> Reviewed-by: Shakeel Butt <shakeelb@google.com> Link: https://lore.kernel.org/r/20220709154457.57379-2-laoar.shao@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2022-07-09 15:44:56 +00:00
* spin_lock*(); alloc(GFP_ATOMIC); spin_unlock*();
bpf: Prepare hashtab locking for PREEMPT_RT PREEMPT_RT forbids certain operations like memory allocations (even with GFP_ATOMIC) from atomic contexts. This is required because even with GFP_ATOMIC the memory allocator calls into code pathes which acquire locks with long held lock sections. To ensure the deterministic behaviour these locks are regular spinlocks, which are converted to 'sleepable' spinlocks on RT. The only true atomic contexts on an RT kernel are the low level hardware handling, scheduling, low level interrupt handling, NMIs etc. None of these contexts should ever do memory allocations. As regular device interrupt handlers and soft interrupts are forced into thread context, the existing code which does spin_lock*(); alloc(GPF_ATOMIC); spin_unlock*(); just works. In theory the BPF locks could be converted to regular spinlocks as well, but the bucket locks and percpu_freelist locks can be taken from arbitrary contexts (perf, kprobes, tracepoints) which are required to be atomic contexts even on RT. These mechanisms require preallocated maps, so there is no need to invoke memory allocations within the lock held sections. BPF maps which need dynamic allocation are only used from (forced) thread context on RT and can therefore use regular spinlocks which in turn allows to invoke memory allocations from the lock held section. To achieve this make the hash bucket lock a union of a raw and a regular spinlock and initialize and lock/unlock either the raw spinlock for preallocated maps or the regular variant for maps which require memory allocations. On a non RT kernel this distinction is neither possible nor required. spinlock maps to raw_spinlock and the extra code and conditional is optimized out by the compiler. No functional change. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200224145644.509685912@linutronix.de
2020-02-24 14:01:51 +00:00
* just works.
*
* In theory the BPF locks could be converted to regular spinlocks as well,
* but the bucket locks and percpu_freelist locks can be taken from
* arbitrary contexts (perf, kprobes, tracepoints) which are required to be
* atomic contexts even on RT. Before the introduction of bpf_mem_alloc,
* it is only safe to use raw spinlock for preallocated hash map on a RT kernel,
* because there is no memory allocation within the lock held sections. However
* after hash map was fully converted to use bpf_mem_alloc, there will be
* non-synchronous memory allocation for non-preallocated hash map, so it is
* safe to always use raw spinlock for bucket lock.
*/
struct bucket {
struct hlist_nulls_head head;
raw_spinlock_t raw_lock;
};
#define HASHTAB_MAP_LOCK_COUNT 8
#define HASHTAB_MAP_LOCK_MASK (HASHTAB_MAP_LOCK_COUNT - 1)
struct bpf_htab {
struct bpf_map map;
struct bpf_mem_alloc ma;
struct bpf_mem_alloc pcpu_ma;
struct bucket *buckets;
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
void *elems;
union {
struct pcpu_freelist freelist;
struct bpf_lru lru;
};
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
struct htab_elem *__percpu *extra_elems;
/* number of elements in non-preallocated hashtable are kept
* in either pcount or count
*/
struct percpu_counter pcount;
atomic_t count;
bool use_percpu_counter;
u32 n_buckets; /* number of hash buckets */
u32 elem_size; /* size of each element in bytes */
u32 hashrnd;
bpf: Use separate lockdep class for each hashtab If a hashtab is accessed in both NMI and non-NMI contexts, it may cause deadlock in bucket->lock. LOCKDEP NMI warning highlighted this issue: ./test_progs -t stacktrace [ 74.828970] [ 74.828971] ================================ [ 74.828972] WARNING: inconsistent lock state [ 74.828973] 5.9.0-rc8+ #275 Not tainted [ 74.828974] -------------------------------- [ 74.828975] inconsistent {INITIAL USE} -> {IN-NMI} usage. [ 74.828976] taskset/1174 [HC2[2]:SC0[0]:HE0:SE1] takes: [ 74.828977] ffffc90000ee96b0 (&htab->buckets[i].raw_lock){....}-{2:2}, at: htab_map_update_elem+0x271/0x5a0 [ 74.828981] {INITIAL USE} state was registered at: [ 74.828982] lock_acquire+0x137/0x510 [ 74.828983] _raw_spin_lock_irqsave+0x43/0x90 [ 74.828984] htab_map_update_elem+0x271/0x5a0 [ 74.828984] 0xffffffffa0040b34 [ 74.828985] trace_call_bpf+0x159/0x310 [ 74.828986] perf_trace_run_bpf_submit+0x5f/0xd0 [ 74.828987] perf_trace_urandom_read+0x1be/0x220 [ 74.828988] urandom_read_nowarn.isra.0+0x26f/0x380 [ 74.828989] vfs_read+0xf8/0x280 [ 74.828989] ksys_read+0xc9/0x160 [ 74.828990] do_syscall_64+0x33/0x40 [ 74.828991] entry_SYSCALL_64_after_hwframe+0x44/0xa9 [ 74.828992] irq event stamp: 1766 [ 74.828993] hardirqs last enabled at (1765): [<ffffffff82800ace>] asm_exc_page_fault+0x1e/0x30 [ 74.828994] hardirqs last disabled at (1766): [<ffffffff8267df87>] irqentry_enter+0x37/0x60 [ 74.828995] softirqs last enabled at (856): [<ffffffff81043e7c>] fpu__clear+0xac/0x120 [ 74.828996] softirqs last disabled at (854): [<ffffffff81043df0>] fpu__clear+0x20/0x120 [ 74.828997] [ 74.828998] other info that might help us debug this: [ 74.828999] Possible unsafe locking scenario: [ 74.828999] [ 74.829000] CPU0 [ 74.829001] ---- [ 74.829001] lock(&htab->buckets[i].raw_lock); [ 74.829003] <Interrupt> [ 74.829004] lock(&htab->buckets[i].raw_lock); [ 74.829006] [ 74.829006] *** DEADLOCK *** [ 74.829007] [ 74.829008] 1 lock held by taskset/1174: [ 74.829008] #0: ffff8883ec3fd020 (&cpuctx_lock){-...}-{2:2}, at: perf_event_task_tick+0x101/0x650 [ 74.829012] [ 74.829013] stack backtrace: [ 74.829014] CPU: 0 PID: 1174 Comm: taskset Not tainted 5.9.0-rc8+ #275 [ 74.829015] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.11.0-2.el7 04/01/2014 [ 74.829016] Call Trace: [ 74.829016] <NMI> [ 74.829017] dump_stack+0x9a/0xd0 [ 74.829018] lock_acquire+0x461/0x510 [ 74.829019] ? lock_release+0x6b0/0x6b0 [ 74.829020] ? stack_map_get_build_id_offset+0x45e/0x800 [ 74.829021] ? htab_map_update_elem+0x271/0x5a0 [ 74.829022] ? rcu_read_lock_held_common+0x1a/0x50 [ 74.829022] ? rcu_read_lock_held+0x5f/0xb0 [ 74.829023] _raw_spin_lock_irqsave+0x43/0x90 [ 74.829024] ? htab_map_update_elem+0x271/0x5a0 [ 74.829025] htab_map_update_elem+0x271/0x5a0 [ 74.829026] bpf_prog_1fd9e30e1438d3c5_oncpu+0x9c/0xe88 [ 74.829027] bpf_overflow_handler+0x127/0x320 [ 74.829028] ? perf_event_text_poke_output+0x4d0/0x4d0 [ 74.829029] ? sched_clock_cpu+0x18/0x130 [ 74.829030] __perf_event_overflow+0xae/0x190 [ 74.829030] handle_pmi_common+0x34c/0x470 [ 74.829031] ? intel_pmu_save_and_restart+0x90/0x90 [ 74.829032] ? lock_acquire+0x3f8/0x510 [ 74.829033] ? lock_release+0x6b0/0x6b0 [ 74.829034] intel_pmu_handle_irq+0x11e/0x240 [ 74.829034] perf_event_nmi_handler+0x40/0x60 [ 74.829035] nmi_handle+0x110/0x360 [ 74.829036] ? __intel_pmu_enable_all.constprop.0+0x72/0xf0 [ 74.829037] default_do_nmi+0x6b/0x170 [ 74.829038] exc_nmi+0x106/0x130 [ 74.829038] end_repeat_nmi+0x16/0x55 [ 74.829039] RIP: 0010:__intel_pmu_enable_all.constprop.0+0x72/0xf0 [ 74.829042] Code: 2f 1f 03 48 8d bb b8 0c 00 00 e8 29 09 41 00 48 ... [ 74.829043] RSP: 0000:ffff8880a604fc90 EFLAGS: 00000002 [ 74.829044] RAX: 000000070000000f RBX: ffff8883ec2195a0 RCX: 000000000000038f [ 74.829045] RDX: 0000000000000007 RSI: ffffffff82e72c20 RDI: ffff8883ec21a258 [ 74.829046] RBP: 000000070000000f R08: ffffffff8101b013 R09: fffffbfff0a7982d [ 74.829047] R10: ffffffff853cc167 R11: fffffbfff0a7982c R12: 0000000000000000 [ 74.829049] R13: ffff8883ec3f0af0 R14: ffff8883ec3fd120 R15: ffff8883e9c92098 [ 74.829049] ? intel_pmu_lbr_enable_all+0x43/0x240 [ 74.829050] ? __intel_pmu_enable_all.constprop.0+0x72/0xf0 [ 74.829051] ? __intel_pmu_enable_all.constprop.0+0x72/0xf0 [ 74.829052] </NMI> [ 74.829053] perf_event_task_tick+0x48d/0x650 [ 74.829054] scheduler_tick+0x129/0x210 [ 74.829054] update_process_times+0x37/0x70 [ 74.829055] tick_sched_handle.isra.0+0x35/0x90 [ 74.829056] tick_sched_timer+0x8f/0xb0 [ 74.829057] __hrtimer_run_queues+0x364/0x7d0 [ 74.829058] ? tick_sched_do_timer+0xa0/0xa0 [ 74.829058] ? enqueue_hrtimer+0x1e0/0x1e0 [ 74.829059] ? recalibrate_cpu_khz+0x10/0x10 [ 74.829060] ? ktime_get_update_offsets_now+0x1a3/0x360 [ 74.829061] hrtimer_interrupt+0x1bb/0x360 [ 74.829062] ? rcu_read_lock_sched_held+0xa1/0xd0 [ 74.829063] __sysvec_apic_timer_interrupt+0xed/0x3d0 [ 74.829064] sysvec_apic_timer_interrupt+0x3f/0xd0 [ 74.829064] ? asm_sysvec_apic_timer_interrupt+0xa/0x20 [ 74.829065] asm_sysvec_apic_timer_interrupt+0x12/0x20 [ 74.829066] RIP: 0033:0x7fba18d579b4 [ 74.829068] Code: 74 54 44 0f b6 4a 04 41 83 e1 0f 41 80 f9 ... [ 74.829069] RSP: 002b:00007ffc9ba69570 EFLAGS: 00000206 [ 74.829071] RAX: 00007fba192084c0 RBX: 00007fba18c24d28 RCX: 00000000000007a4 [ 74.829072] RDX: 00007fba18c30488 RSI: 0000000000000000 RDI: 000000000000037b [ 74.829073] RBP: 00007fba18ca5760 R08: 00007fba18c248fc R09: 00007fba18c94c30 [ 74.829074] R10: 000000000000002f R11: 0000000000073c30 R12: 00007ffc9ba695e0 [ 74.829075] R13: 00000000000003f3 R14: 00007fba18c21ac8 R15: 00000000000058d6 However, such warning should not apply across multiple hashtabs. The system will not deadlock if one hashtab is used in NMI, while another hashtab is used in non-NMI. Use separate lockdep class for each hashtab, so that we don't get this false alert. Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20201029071925.3103400-2-songliubraving@fb.com
2020-10-29 07:19:24 +00:00
struct lock_class_key lockdep_key;
int __percpu *map_locked[HASHTAB_MAP_LOCK_COUNT];
};
/* each htab element is struct htab_elem + key + value */
struct htab_elem {
union {
struct hlist_nulls_node hash_node;
struct {
void *padding;
union {
struct pcpu_freelist_node fnode;
bpf: Fix a potential deadlock with bpf_map_do_batch Commit 057996380a42 ("bpf: Add batch ops to all htab bpf map") added lookup_and_delete batch operation for hash table. The current implementation has bpf_lru_push_free() inside the bucket lock, which may cause a deadlock. syzbot reports: -> #2 (&htab->buckets[i].lock#2){....}: __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159 htab_lru_map_delete_node+0xce/0x2f0 kernel/bpf/hashtab.c:593 __bpf_lru_list_shrink_inactive kernel/bpf/bpf_lru_list.c:220 [inline] __bpf_lru_list_shrink+0xf9/0x470 kernel/bpf/bpf_lru_list.c:266 bpf_lru_list_pop_free_to_local kernel/bpf/bpf_lru_list.c:340 [inline] bpf_common_lru_pop_free kernel/bpf/bpf_lru_list.c:447 [inline] bpf_lru_pop_free+0x87c/0x1670 kernel/bpf/bpf_lru_list.c:499 prealloc_lru_pop+0x2c/0xa0 kernel/bpf/hashtab.c:132 __htab_lru_percpu_map_update_elem+0x67e/0xa90 kernel/bpf/hashtab.c:1069 bpf_percpu_hash_update+0x16e/0x210 kernel/bpf/hashtab.c:1585 bpf_map_update_value.isra.0+0x2d7/0x8e0 kernel/bpf/syscall.c:181 generic_map_update_batch+0x41f/0x610 kernel/bpf/syscall.c:1319 bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348 __do_sys_bpf+0x9b7/0x41e0 kernel/bpf/syscall.c:3460 __se_sys_bpf kernel/bpf/syscall.c:3355 [inline] __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355 do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294 entry_SYSCALL_64_after_hwframe+0x49/0xbe -> #0 (&loc_l->lock){....}: check_prev_add kernel/locking/lockdep.c:2475 [inline] check_prevs_add kernel/locking/lockdep.c:2580 [inline] validate_chain kernel/locking/lockdep.c:2970 [inline] __lock_acquire+0x2596/0x4a00 kernel/locking/lockdep.c:3954 lock_acquire+0x190/0x410 kernel/locking/lockdep.c:4484 __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159 bpf_common_lru_push_free kernel/bpf/bpf_lru_list.c:516 [inline] bpf_lru_push_free+0x250/0x5b0 kernel/bpf/bpf_lru_list.c:555 __htab_map_lookup_and_delete_batch+0x8d4/0x1540 kernel/bpf/hashtab.c:1374 htab_lru_map_lookup_and_delete_batch+0x34/0x40 kernel/bpf/hashtab.c:1491 bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348 __do_sys_bpf+0x1f7d/0x41e0 kernel/bpf/syscall.c:3456 __se_sys_bpf kernel/bpf/syscall.c:3355 [inline] __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355 do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294 entry_SYSCALL_64_after_hwframe+0x49/0xbe Possible unsafe locking scenario: CPU0 CPU2 ---- ---- lock(&htab->buckets[i].lock#2); lock(&l->lock); lock(&htab->buckets[i].lock#2); lock(&loc_l->lock); *** DEADLOCK *** To fix the issue, for htab_lru_map_lookup_and_delete_batch() in CPU0, let us do bpf_lru_push_free() out of the htab bucket lock. This can avoid the above deadlock scenario. Fixes: 057996380a42 ("bpf: Add batch ops to all htab bpf map") Reported-by: syzbot+a38ff3d9356388f2fb83@syzkaller.appspotmail.com Reported-by: syzbot+122b5421d14e68f29cd1@syzkaller.appspotmail.com Suggested-by: Hillf Danton <hdanton@sina.com> Suggested-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com> Acked-by: Brian Vazquez <brianvv@google.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Link: https://lore.kernel.org/bpf/20200219234757.3544014-1-yhs@fb.com
2020-02-19 23:47:57 +00:00
struct htab_elem *batch_flink;
};
};
};
union {
/* pointer to per-cpu pointer */
void *ptr_to_pptr;
struct bpf_lru_node lru_node;
};
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
u32 hash;
char key[] __aligned(8);
};
bpf: Prepare hashtab locking for PREEMPT_RT PREEMPT_RT forbids certain operations like memory allocations (even with GFP_ATOMIC) from atomic contexts. This is required because even with GFP_ATOMIC the memory allocator calls into code pathes which acquire locks with long held lock sections. To ensure the deterministic behaviour these locks are regular spinlocks, which are converted to 'sleepable' spinlocks on RT. The only true atomic contexts on an RT kernel are the low level hardware handling, scheduling, low level interrupt handling, NMIs etc. None of these contexts should ever do memory allocations. As regular device interrupt handlers and soft interrupts are forced into thread context, the existing code which does spin_lock*(); alloc(GPF_ATOMIC); spin_unlock*(); just works. In theory the BPF locks could be converted to regular spinlocks as well, but the bucket locks and percpu_freelist locks can be taken from arbitrary contexts (perf, kprobes, tracepoints) which are required to be atomic contexts even on RT. These mechanisms require preallocated maps, so there is no need to invoke memory allocations within the lock held sections. BPF maps which need dynamic allocation are only used from (forced) thread context on RT and can therefore use regular spinlocks which in turn allows to invoke memory allocations from the lock held section. To achieve this make the hash bucket lock a union of a raw and a regular spinlock and initialize and lock/unlock either the raw spinlock for preallocated maps or the regular variant for maps which require memory allocations. On a non RT kernel this distinction is neither possible nor required. spinlock maps to raw_spinlock and the extra code and conditional is optimized out by the compiler. No functional change. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200224145644.509685912@linutronix.de
2020-02-24 14:01:51 +00:00
static inline bool htab_is_prealloc(const struct bpf_htab *htab)
{
return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
}
static void htab_init_buckets(struct bpf_htab *htab)
{
unsigned int i;
for (i = 0; i < htab->n_buckets; i++) {
INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
raw_spin_lock_init(&htab->buckets[i].raw_lock);
lockdep_set_class(&htab->buckets[i].raw_lock,
bpf: Use separate lockdep class for each hashtab If a hashtab is accessed in both NMI and non-NMI contexts, it may cause deadlock in bucket->lock. LOCKDEP NMI warning highlighted this issue: ./test_progs -t stacktrace [ 74.828970] [ 74.828971] ================================ [ 74.828972] WARNING: inconsistent lock state [ 74.828973] 5.9.0-rc8+ #275 Not tainted [ 74.828974] -------------------------------- [ 74.828975] inconsistent {INITIAL USE} -> {IN-NMI} usage. [ 74.828976] taskset/1174 [HC2[2]:SC0[0]:HE0:SE1] takes: [ 74.828977] ffffc90000ee96b0 (&htab->buckets[i].raw_lock){....}-{2:2}, at: htab_map_update_elem+0x271/0x5a0 [ 74.828981] {INITIAL USE} state was registered at: [ 74.828982] lock_acquire+0x137/0x510 [ 74.828983] _raw_spin_lock_irqsave+0x43/0x90 [ 74.828984] htab_map_update_elem+0x271/0x5a0 [ 74.828984] 0xffffffffa0040b34 [ 74.828985] trace_call_bpf+0x159/0x310 [ 74.828986] perf_trace_run_bpf_submit+0x5f/0xd0 [ 74.828987] perf_trace_urandom_read+0x1be/0x220 [ 74.828988] urandom_read_nowarn.isra.0+0x26f/0x380 [ 74.828989] vfs_read+0xf8/0x280 [ 74.828989] ksys_read+0xc9/0x160 [ 74.828990] do_syscall_64+0x33/0x40 [ 74.828991] entry_SYSCALL_64_after_hwframe+0x44/0xa9 [ 74.828992] irq event stamp: 1766 [ 74.828993] hardirqs last enabled at (1765): [<ffffffff82800ace>] asm_exc_page_fault+0x1e/0x30 [ 74.828994] hardirqs last disabled at (1766): [<ffffffff8267df87>] irqentry_enter+0x37/0x60 [ 74.828995] softirqs last enabled at (856): [<ffffffff81043e7c>] fpu__clear+0xac/0x120 [ 74.828996] softirqs last disabled at (854): [<ffffffff81043df0>] fpu__clear+0x20/0x120 [ 74.828997] [ 74.828998] other info that might help us debug this: [ 74.828999] Possible unsafe locking scenario: [ 74.828999] [ 74.829000] CPU0 [ 74.829001] ---- [ 74.829001] lock(&htab->buckets[i].raw_lock); [ 74.829003] <Interrupt> [ 74.829004] lock(&htab->buckets[i].raw_lock); [ 74.829006] [ 74.829006] *** DEADLOCK *** [ 74.829007] [ 74.829008] 1 lock held by taskset/1174: [ 74.829008] #0: ffff8883ec3fd020 (&cpuctx_lock){-...}-{2:2}, at: perf_event_task_tick+0x101/0x650 [ 74.829012] [ 74.829013] stack backtrace: [ 74.829014] CPU: 0 PID: 1174 Comm: taskset Not tainted 5.9.0-rc8+ #275 [ 74.829015] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.11.0-2.el7 04/01/2014 [ 74.829016] Call Trace: [ 74.829016] <NMI> [ 74.829017] dump_stack+0x9a/0xd0 [ 74.829018] lock_acquire+0x461/0x510 [ 74.829019] ? lock_release+0x6b0/0x6b0 [ 74.829020] ? stack_map_get_build_id_offset+0x45e/0x800 [ 74.829021] ? htab_map_update_elem+0x271/0x5a0 [ 74.829022] ? rcu_read_lock_held_common+0x1a/0x50 [ 74.829022] ? rcu_read_lock_held+0x5f/0xb0 [ 74.829023] _raw_spin_lock_irqsave+0x43/0x90 [ 74.829024] ? htab_map_update_elem+0x271/0x5a0 [ 74.829025] htab_map_update_elem+0x271/0x5a0 [ 74.829026] bpf_prog_1fd9e30e1438d3c5_oncpu+0x9c/0xe88 [ 74.829027] bpf_overflow_handler+0x127/0x320 [ 74.829028] ? perf_event_text_poke_output+0x4d0/0x4d0 [ 74.829029] ? sched_clock_cpu+0x18/0x130 [ 74.829030] __perf_event_overflow+0xae/0x190 [ 74.829030] handle_pmi_common+0x34c/0x470 [ 74.829031] ? intel_pmu_save_and_restart+0x90/0x90 [ 74.829032] ? lock_acquire+0x3f8/0x510 [ 74.829033] ? lock_release+0x6b0/0x6b0 [ 74.829034] intel_pmu_handle_irq+0x11e/0x240 [ 74.829034] perf_event_nmi_handler+0x40/0x60 [ 74.829035] nmi_handle+0x110/0x360 [ 74.829036] ? __intel_pmu_enable_all.constprop.0+0x72/0xf0 [ 74.829037] default_do_nmi+0x6b/0x170 [ 74.829038] exc_nmi+0x106/0x130 [ 74.829038] end_repeat_nmi+0x16/0x55 [ 74.829039] RIP: 0010:__intel_pmu_enable_all.constprop.0+0x72/0xf0 [ 74.829042] Code: 2f 1f 03 48 8d bb b8 0c 00 00 e8 29 09 41 00 48 ... [ 74.829043] RSP: 0000:ffff8880a604fc90 EFLAGS: 00000002 [ 74.829044] RAX: 000000070000000f RBX: ffff8883ec2195a0 RCX: 000000000000038f [ 74.829045] RDX: 0000000000000007 RSI: ffffffff82e72c20 RDI: ffff8883ec21a258 [ 74.829046] RBP: 000000070000000f R08: ffffffff8101b013 R09: fffffbfff0a7982d [ 74.829047] R10: ffffffff853cc167 R11: fffffbfff0a7982c R12: 0000000000000000 [ 74.829049] R13: ffff8883ec3f0af0 R14: ffff8883ec3fd120 R15: ffff8883e9c92098 [ 74.829049] ? intel_pmu_lbr_enable_all+0x43/0x240 [ 74.829050] ? __intel_pmu_enable_all.constprop.0+0x72/0xf0 [ 74.829051] ? __intel_pmu_enable_all.constprop.0+0x72/0xf0 [ 74.829052] </NMI> [ 74.829053] perf_event_task_tick+0x48d/0x650 [ 74.829054] scheduler_tick+0x129/0x210 [ 74.829054] update_process_times+0x37/0x70 [ 74.829055] tick_sched_handle.isra.0+0x35/0x90 [ 74.829056] tick_sched_timer+0x8f/0xb0 [ 74.829057] __hrtimer_run_queues+0x364/0x7d0 [ 74.829058] ? tick_sched_do_timer+0xa0/0xa0 [ 74.829058] ? enqueue_hrtimer+0x1e0/0x1e0 [ 74.829059] ? recalibrate_cpu_khz+0x10/0x10 [ 74.829060] ? ktime_get_update_offsets_now+0x1a3/0x360 [ 74.829061] hrtimer_interrupt+0x1bb/0x360 [ 74.829062] ? rcu_read_lock_sched_held+0xa1/0xd0 [ 74.829063] __sysvec_apic_timer_interrupt+0xed/0x3d0 [ 74.829064] sysvec_apic_timer_interrupt+0x3f/0xd0 [ 74.829064] ? asm_sysvec_apic_timer_interrupt+0xa/0x20 [ 74.829065] asm_sysvec_apic_timer_interrupt+0x12/0x20 [ 74.829066] RIP: 0033:0x7fba18d579b4 [ 74.829068] Code: 74 54 44 0f b6 4a 04 41 83 e1 0f 41 80 f9 ... [ 74.829069] RSP: 002b:00007ffc9ba69570 EFLAGS: 00000206 [ 74.829071] RAX: 00007fba192084c0 RBX: 00007fba18c24d28 RCX: 00000000000007a4 [ 74.829072] RDX: 00007fba18c30488 RSI: 0000000000000000 RDI: 000000000000037b [ 74.829073] RBP: 00007fba18ca5760 R08: 00007fba18c248fc R09: 00007fba18c94c30 [ 74.829074] R10: 000000000000002f R11: 0000000000073c30 R12: 00007ffc9ba695e0 [ 74.829075] R13: 00000000000003f3 R14: 00007fba18c21ac8 R15: 00000000000058d6 However, such warning should not apply across multiple hashtabs. The system will not deadlock if one hashtab is used in NMI, while another hashtab is used in non-NMI. Use separate lockdep class for each hashtab, so that we don't get this false alert. Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20201029071925.3103400-2-songliubraving@fb.com
2020-10-29 07:19:24 +00:00
&htab->lockdep_key);
cond_resched();
}
}
static inline int htab_lock_bucket(const struct bpf_htab *htab,
struct bucket *b, u32 hash,
unsigned long *pflags)
{
unsigned long flags;
hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
preempt_disable();
local_irq_save(flags);
if (unlikely(__this_cpu_inc_return(*(htab->map_locked[hash])) != 1)) {
__this_cpu_dec(*(htab->map_locked[hash]));
local_irq_restore(flags);
preempt_enable();
return -EBUSY;
}
raw_spin_lock(&b->raw_lock);
*pflags = flags;
return 0;
}
static inline void htab_unlock_bucket(const struct bpf_htab *htab,
struct bucket *b, u32 hash,
unsigned long flags)
{
hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
raw_spin_unlock(&b->raw_lock);
__this_cpu_dec(*(htab->map_locked[hash]));
local_irq_restore(flags);
preempt_enable();
}
static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
static bool htab_is_lru(const struct bpf_htab *htab)
{
return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
}
static bool htab_is_percpu(const struct bpf_htab *htab)
{
return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
}
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
void __percpu *pptr)
{
*(void __percpu **)(l->key + key_size) = pptr;
}
static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
{
return *(void __percpu **)(l->key + key_size);
}
static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
{
return *(void **)(l->key + roundup(map->key_size, 8));
}
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
{
bpf: Avoid overflows involving hash elem_size Use of bpf_map_charge_init() was making sure hash tables would not use more than 4GB of memory. Since the implicit check disappeared, we have to be more careful about overflows, to support big hash tables. syzbot triggers a panic using : bpf(BPF_MAP_CREATE, {map_type=BPF_MAP_TYPE_LRU_HASH, key_size=16384, value_size=8, max_entries=262200, map_flags=0, inner_map_fd=-1, map_name="", map_ifindex=0, btf_fd=-1, btf_key_type_id=0, btf_value_type_id=0, btf_vmlinux_value_type_id=0}, 64) = ... BUG: KASAN: vmalloc-out-of-bounds in bpf_percpu_lru_populate kernel/bpf/bpf_lru_list.c:594 [inline] BUG: KASAN: vmalloc-out-of-bounds in bpf_lru_populate+0x4ef/0x5e0 kernel/bpf/bpf_lru_list.c:611 Write of size 2 at addr ffffc90017e4a020 by task syz-executor.5/19786 CPU: 0 PID: 19786 Comm: syz-executor.5 Not tainted 5.10.0-rc3-syzkaller #0 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:77 [inline] dump_stack+0x107/0x163 lib/dump_stack.c:118 print_address_description.constprop.0.cold+0x5/0x4c8 mm/kasan/report.c:385 __kasan_report mm/kasan/report.c:545 [inline] kasan_report.cold+0x1f/0x37 mm/kasan/report.c:562 bpf_percpu_lru_populate kernel/bpf/bpf_lru_list.c:594 [inline] bpf_lru_populate+0x4ef/0x5e0 kernel/bpf/bpf_lru_list.c:611 prealloc_init kernel/bpf/hashtab.c:319 [inline] htab_map_alloc+0xf6e/0x1230 kernel/bpf/hashtab.c:507 find_and_alloc_map kernel/bpf/syscall.c:123 [inline] map_create kernel/bpf/syscall.c:829 [inline] __do_sys_bpf+0xa81/0x5170 kernel/bpf/syscall.c:4336 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x45deb9 Code: 0d b4 fb ff c3 66 2e 0f 1f 84 00 00 00 00 00 66 90 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 0f 83 db b3 fb ff c3 66 2e 0f 1f 84 00 00 00 00 RSP: 002b:00007fd93fbc0c78 EFLAGS: 00000246 ORIG_RAX: 0000000000000141 RAX: ffffffffffffffda RBX: 0000000000001a40 RCX: 000000000045deb9 RDX: 0000000000000040 RSI: 0000000020000280 RDI: 0000000000000000 RBP: 000000000119bf60 R08: 0000000000000000 R09: 0000000000000000 R10: 0000000000000000 R11: 0000000000000246 R12: 000000000119bf2c R13: 00007ffc08a7be8f R14: 00007fd93fbc19c0 R15: 000000000119bf2c Fixes: 755e5d55367a ("bpf: Eliminate rlimit-based memory accounting for hashtab maps") Reported-by: syzbot <syzkaller@googlegroups.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Roman Gushchin <guro@fb.com> Link: https://lore.kernel.org/bpf/20201207182821.3940306-1-eric.dumazet@gmail.com
2020-12-07 18:28:21 +00:00
return (struct htab_elem *) (htab->elems + i * (u64)htab->elem_size);
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
}
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
static bool htab_has_extra_elems(struct bpf_htab *htab)
{
return !htab_is_percpu(htab) && !htab_is_lru(htab);
}
static void htab_free_prealloced_timers_and_wq(struct bpf_htab *htab)
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
{
u32 num_entries = htab->map.max_entries;
int i;
if (htab_has_extra_elems(htab))
num_entries += num_possible_cpus();
for (i = 0; i < num_entries; i++) {
struct htab_elem *elem;
elem = get_htab_elem(htab, i);
if (btf_record_has_field(htab->map.record, BPF_TIMER))
bpf_obj_free_timer(htab->map.record,
elem->key + round_up(htab->map.key_size, 8));
if (btf_record_has_field(htab->map.record, BPF_WORKQUEUE))
bpf_obj_free_workqueue(htab->map.record,
elem->key + round_up(htab->map.key_size, 8));
cond_resched();
}
}
static void htab_free_prealloced_fields(struct bpf_htab *htab)
bpf: Wire up freeing of referenced kptr A destructor kfunc can be defined as void func(type *), where type may be void or any other pointer type as per convenience. In this patch, we ensure that the type is sane and capture the function pointer into off_desc of ptr_off_tab for the specific pointer offset, with the invariant that the dtor pointer is always set when 'kptr_ref' tag is applied to the pointer's pointee type, which is indicated by the flag BPF_MAP_VALUE_OFF_F_REF. Note that only BTF IDs whose destructor kfunc is registered, thus become the allowed BTF IDs for embedding as referenced kptr. Hence it serves the purpose of finding dtor kfunc BTF ID, as well acting as a check against the whitelist of allowed BTF IDs for this purpose. Finally, wire up the actual freeing of the referenced pointer if any at all available offsets, so that no references are leaked after the BPF map goes away and the BPF program previously moved the ownership a referenced pointer into it. The behavior is similar to BPF timers, where bpf_map_{update,delete}_elem will free any existing referenced kptr. The same case is with LRU map's bpf_lru_push_free/htab_lru_push_free functions, which are extended to reset unreferenced and free referenced kptr. Note that unlike BPF timers, kptr is not reset or freed when map uref drops to zero. Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20220424214901.2743946-8-memxor@gmail.com
2022-04-24 21:48:55 +00:00
{
u32 num_entries = htab->map.max_entries;
int i;
if (IS_ERR_OR_NULL(htab->map.record))
bpf: Wire up freeing of referenced kptr A destructor kfunc can be defined as void func(type *), where type may be void or any other pointer type as per convenience. In this patch, we ensure that the type is sane and capture the function pointer into off_desc of ptr_off_tab for the specific pointer offset, with the invariant that the dtor pointer is always set when 'kptr_ref' tag is applied to the pointer's pointee type, which is indicated by the flag BPF_MAP_VALUE_OFF_F_REF. Note that only BTF IDs whose destructor kfunc is registered, thus become the allowed BTF IDs for embedding as referenced kptr. Hence it serves the purpose of finding dtor kfunc BTF ID, as well acting as a check against the whitelist of allowed BTF IDs for this purpose. Finally, wire up the actual freeing of the referenced pointer if any at all available offsets, so that no references are leaked after the BPF map goes away and the BPF program previously moved the ownership a referenced pointer into it. The behavior is similar to BPF timers, where bpf_map_{update,delete}_elem will free any existing referenced kptr. The same case is with LRU map's bpf_lru_push_free/htab_lru_push_free functions, which are extended to reset unreferenced and free referenced kptr. Note that unlike BPF timers, kptr is not reset or freed when map uref drops to zero. Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20220424214901.2743946-8-memxor@gmail.com
2022-04-24 21:48:55 +00:00
return;
if (htab_has_extra_elems(htab))
num_entries += num_possible_cpus();
for (i = 0; i < num_entries; i++) {
struct htab_elem *elem;
elem = get_htab_elem(htab, i);
if (htab_is_percpu(htab)) {
void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
int cpu;
for_each_possible_cpu(cpu) {
bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
cond_resched();
}
} else {
bpf_obj_free_fields(htab->map.record, elem->key + round_up(htab->map.key_size, 8));
cond_resched();
}
bpf: Wire up freeing of referenced kptr A destructor kfunc can be defined as void func(type *), where type may be void or any other pointer type as per convenience. In this patch, we ensure that the type is sane and capture the function pointer into off_desc of ptr_off_tab for the specific pointer offset, with the invariant that the dtor pointer is always set when 'kptr_ref' tag is applied to the pointer's pointee type, which is indicated by the flag BPF_MAP_VALUE_OFF_F_REF. Note that only BTF IDs whose destructor kfunc is registered, thus become the allowed BTF IDs for embedding as referenced kptr. Hence it serves the purpose of finding dtor kfunc BTF ID, as well acting as a check against the whitelist of allowed BTF IDs for this purpose. Finally, wire up the actual freeing of the referenced pointer if any at all available offsets, so that no references are leaked after the BPF map goes away and the BPF program previously moved the ownership a referenced pointer into it. The behavior is similar to BPF timers, where bpf_map_{update,delete}_elem will free any existing referenced kptr. The same case is with LRU map's bpf_lru_push_free/htab_lru_push_free functions, which are extended to reset unreferenced and free referenced kptr. Note that unlike BPF timers, kptr is not reset or freed when map uref drops to zero. Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20220424214901.2743946-8-memxor@gmail.com
2022-04-24 21:48:55 +00:00
cond_resched();
}
}
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
static void htab_free_elems(struct bpf_htab *htab)
{
int i;
if (!htab_is_percpu(htab))
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
goto free_elems;
for (i = 0; i < htab->map.max_entries; i++) {
void __percpu *pptr;
pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
htab->map.key_size);
free_percpu(pptr);
cond_resched();
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
}
free_elems:
bpf_map_area_free(htab->elems);
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
}
bpf: Fix a potential deadlock with bpf_map_do_batch Commit 057996380a42 ("bpf: Add batch ops to all htab bpf map") added lookup_and_delete batch operation for hash table. The current implementation has bpf_lru_push_free() inside the bucket lock, which may cause a deadlock. syzbot reports: -> #2 (&htab->buckets[i].lock#2){....}: __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159 htab_lru_map_delete_node+0xce/0x2f0 kernel/bpf/hashtab.c:593 __bpf_lru_list_shrink_inactive kernel/bpf/bpf_lru_list.c:220 [inline] __bpf_lru_list_shrink+0xf9/0x470 kernel/bpf/bpf_lru_list.c:266 bpf_lru_list_pop_free_to_local kernel/bpf/bpf_lru_list.c:340 [inline] bpf_common_lru_pop_free kernel/bpf/bpf_lru_list.c:447 [inline] bpf_lru_pop_free+0x87c/0x1670 kernel/bpf/bpf_lru_list.c:499 prealloc_lru_pop+0x2c/0xa0 kernel/bpf/hashtab.c:132 __htab_lru_percpu_map_update_elem+0x67e/0xa90 kernel/bpf/hashtab.c:1069 bpf_percpu_hash_update+0x16e/0x210 kernel/bpf/hashtab.c:1585 bpf_map_update_value.isra.0+0x2d7/0x8e0 kernel/bpf/syscall.c:181 generic_map_update_batch+0x41f/0x610 kernel/bpf/syscall.c:1319 bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348 __do_sys_bpf+0x9b7/0x41e0 kernel/bpf/syscall.c:3460 __se_sys_bpf kernel/bpf/syscall.c:3355 [inline] __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355 do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294 entry_SYSCALL_64_after_hwframe+0x49/0xbe -> #0 (&loc_l->lock){....}: check_prev_add kernel/locking/lockdep.c:2475 [inline] check_prevs_add kernel/locking/lockdep.c:2580 [inline] validate_chain kernel/locking/lockdep.c:2970 [inline] __lock_acquire+0x2596/0x4a00 kernel/locking/lockdep.c:3954 lock_acquire+0x190/0x410 kernel/locking/lockdep.c:4484 __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159 bpf_common_lru_push_free kernel/bpf/bpf_lru_list.c:516 [inline] bpf_lru_push_free+0x250/0x5b0 kernel/bpf/bpf_lru_list.c:555 __htab_map_lookup_and_delete_batch+0x8d4/0x1540 kernel/bpf/hashtab.c:1374 htab_lru_map_lookup_and_delete_batch+0x34/0x40 kernel/bpf/hashtab.c:1491 bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348 __do_sys_bpf+0x1f7d/0x41e0 kernel/bpf/syscall.c:3456 __se_sys_bpf kernel/bpf/syscall.c:3355 [inline] __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355 do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294 entry_SYSCALL_64_after_hwframe+0x49/0xbe Possible unsafe locking scenario: CPU0 CPU2 ---- ---- lock(&htab->buckets[i].lock#2); lock(&l->lock); lock(&htab->buckets[i].lock#2); lock(&loc_l->lock); *** DEADLOCK *** To fix the issue, for htab_lru_map_lookup_and_delete_batch() in CPU0, let us do bpf_lru_push_free() out of the htab bucket lock. This can avoid the above deadlock scenario. Fixes: 057996380a42 ("bpf: Add batch ops to all htab bpf map") Reported-by: syzbot+a38ff3d9356388f2fb83@syzkaller.appspotmail.com Reported-by: syzbot+122b5421d14e68f29cd1@syzkaller.appspotmail.com Suggested-by: Hillf Danton <hdanton@sina.com> Suggested-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com> Acked-by: Brian Vazquez <brianvv@google.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Link: https://lore.kernel.org/bpf/20200219234757.3544014-1-yhs@fb.com
2020-02-19 23:47:57 +00:00
/* The LRU list has a lock (lru_lock). Each htab bucket has a lock
* (bucket_lock). If both locks need to be acquired together, the lock
* order is always lru_lock -> bucket_lock and this only happens in
* bpf_lru_list.c logic. For example, certain code path of
* bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
* will acquire lru_lock first followed by acquiring bucket_lock.
*
* In hashtab.c, to avoid deadlock, lock acquisition of
* bucket_lock followed by lru_lock is not allowed. In such cases,
* bucket_lock needs to be released first before acquiring lru_lock.
*/
static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
u32 hash)
{
struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
struct htab_elem *l;
if (node) {
bpf_map_inc_elem_count(&htab->map);
l = container_of(node, struct htab_elem, lru_node);
bpf: Don't reinit map value in prealloc_lru_pop The LRU map that is preallocated may have its elements reused while another program holds a pointer to it from bpf_map_lookup_elem. Hence, only check_and_free_fields is appropriate when the element is being deleted, as it ensures proper synchronization against concurrent access of the map value. After that, we cannot call check_and_init_map_value again as it may rewrite bpf_spin_lock, bpf_timer, and kptr fields while they can be concurrently accessed from a BPF program. This is safe to do as when the map entry is deleted, concurrent access is protected against by check_and_free_fields, i.e. an existing timer would be freed, and any existing kptr will be released by it. The program can create further timers and kptrs after check_and_free_fields, but they will eventually be released once the preallocated items are freed on map destruction, even if the item is never reused again. Hence, the deleted item sitting in the free list can still have resources attached to it, and they would never leak. With spin_lock, we never touch the field at all on delete or update, as we may end up modifying the state of the lock. Since the verifier ensures that a bpf_spin_lock call is always paired with bpf_spin_unlock call, the program will eventually release the lock so that on reuse the new user of the value can take the lock. Essentially, for the preallocated case, we must assume that the map value may always be in use by the program, even when it is sitting in the freelist, and handle things accordingly, i.e. use proper synchronization inside check_and_free_fields, and never reinitialize the special fields when it is reused on update. Fixes: 68134668c17f ("bpf: Add map side support for bpf timers.") Acked-by: Yonghong Song <yhs@fb.com> Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Link: https://lore.kernel.org/r/20220809213033.24147-3-memxor@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2022-08-09 21:30:32 +00:00
memcpy(l->key, key, htab->map.key_size);
return l;
}
return NULL;
}
static int prealloc_init(struct bpf_htab *htab)
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
{
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
u32 num_entries = htab->map.max_entries;
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
int err = -ENOMEM, i;
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
if (htab_has_extra_elems(htab))
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
num_entries += num_possible_cpus();
bpf: Avoid overflows involving hash elem_size Use of bpf_map_charge_init() was making sure hash tables would not use more than 4GB of memory. Since the implicit check disappeared, we have to be more careful about overflows, to support big hash tables. syzbot triggers a panic using : bpf(BPF_MAP_CREATE, {map_type=BPF_MAP_TYPE_LRU_HASH, key_size=16384, value_size=8, max_entries=262200, map_flags=0, inner_map_fd=-1, map_name="", map_ifindex=0, btf_fd=-1, btf_key_type_id=0, btf_value_type_id=0, btf_vmlinux_value_type_id=0}, 64) = ... BUG: KASAN: vmalloc-out-of-bounds in bpf_percpu_lru_populate kernel/bpf/bpf_lru_list.c:594 [inline] BUG: KASAN: vmalloc-out-of-bounds in bpf_lru_populate+0x4ef/0x5e0 kernel/bpf/bpf_lru_list.c:611 Write of size 2 at addr ffffc90017e4a020 by task syz-executor.5/19786 CPU: 0 PID: 19786 Comm: syz-executor.5 Not tainted 5.10.0-rc3-syzkaller #0 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:77 [inline] dump_stack+0x107/0x163 lib/dump_stack.c:118 print_address_description.constprop.0.cold+0x5/0x4c8 mm/kasan/report.c:385 __kasan_report mm/kasan/report.c:545 [inline] kasan_report.cold+0x1f/0x37 mm/kasan/report.c:562 bpf_percpu_lru_populate kernel/bpf/bpf_lru_list.c:594 [inline] bpf_lru_populate+0x4ef/0x5e0 kernel/bpf/bpf_lru_list.c:611 prealloc_init kernel/bpf/hashtab.c:319 [inline] htab_map_alloc+0xf6e/0x1230 kernel/bpf/hashtab.c:507 find_and_alloc_map kernel/bpf/syscall.c:123 [inline] map_create kernel/bpf/syscall.c:829 [inline] __do_sys_bpf+0xa81/0x5170 kernel/bpf/syscall.c:4336 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x45deb9 Code: 0d b4 fb ff c3 66 2e 0f 1f 84 00 00 00 00 00 66 90 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 0f 83 db b3 fb ff c3 66 2e 0f 1f 84 00 00 00 00 RSP: 002b:00007fd93fbc0c78 EFLAGS: 00000246 ORIG_RAX: 0000000000000141 RAX: ffffffffffffffda RBX: 0000000000001a40 RCX: 000000000045deb9 RDX: 0000000000000040 RSI: 0000000020000280 RDI: 0000000000000000 RBP: 000000000119bf60 R08: 0000000000000000 R09: 0000000000000000 R10: 0000000000000000 R11: 0000000000000246 R12: 000000000119bf2c R13: 00007ffc08a7be8f R14: 00007fd93fbc19c0 R15: 000000000119bf2c Fixes: 755e5d55367a ("bpf: Eliminate rlimit-based memory accounting for hashtab maps") Reported-by: syzbot <syzkaller@googlegroups.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Roman Gushchin <guro@fb.com> Link: https://lore.kernel.org/bpf/20201207182821.3940306-1-eric.dumazet@gmail.com
2020-12-07 18:28:21 +00:00
htab->elems = bpf_map_area_alloc((u64)htab->elem_size * num_entries,
bpf: Allow selecting numa node during map creation The current map creation API does not allow to provide the numa-node preference. The memory usually comes from where the map-creation-process is running. The performance is not ideal if the bpf_prog is known to always run in a numa node different from the map-creation-process. One of the use case is sharding on CPU to different LRU maps (i.e. an array of LRU maps). Here is the test result of map_perf_test on the INNER_LRU_HASH_PREALLOC test if we force the lru map used by CPU0 to be allocated from a remote numa node: [ The machine has 20 cores. CPU0-9 at node 0. CPU10-19 at node 1 ] ># taskset -c 10 ./map_perf_test 512 8 1260000 8000000 5:inner_lru_hash_map_perf pre-alloc 1628380 events per sec 4:inner_lru_hash_map_perf pre-alloc 1626396 events per sec 3:inner_lru_hash_map_perf pre-alloc 1626144 events per sec 6:inner_lru_hash_map_perf pre-alloc 1621657 events per sec 2:inner_lru_hash_map_perf pre-alloc 1621534 events per sec 1:inner_lru_hash_map_perf pre-alloc 1620292 events per sec 7:inner_lru_hash_map_perf pre-alloc 1613305 events per sec 0:inner_lru_hash_map_perf pre-alloc 1239150 events per sec #<<< After specifying numa node: ># taskset -c 10 ./map_perf_test 512 8 1260000 8000000 5:inner_lru_hash_map_perf pre-alloc 1629627 events per sec 3:inner_lru_hash_map_perf pre-alloc 1628057 events per sec 1:inner_lru_hash_map_perf pre-alloc 1623054 events per sec 6:inner_lru_hash_map_perf pre-alloc 1616033 events per sec 2:inner_lru_hash_map_perf pre-alloc 1614630 events per sec 4:inner_lru_hash_map_perf pre-alloc 1612651 events per sec 7:inner_lru_hash_map_perf pre-alloc 1609337 events per sec 0:inner_lru_hash_map_perf pre-alloc 1619340 events per sec #<<< This patch adds one field, numa_node, to the bpf_attr. Since numa node 0 is a valid node, a new flag BPF_F_NUMA_NODE is also added. The numa_node field is honored if and only if the BPF_F_NUMA_NODE flag is set. Numa node selection is not supported for percpu map. This patch does not change all the kmalloc. F.e. 'htab = kzalloc()' is not changed since the object is small enough to stay in the cache. Signed-off-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-18 18:28:00 +00:00
htab->map.numa_node);
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
if (!htab->elems)
return -ENOMEM;
if (!htab_is_percpu(htab))
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
goto skip_percpu_elems;
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
for (i = 0; i < num_entries; i++) {
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
u32 size = round_up(htab->map.value_size, 8);
void __percpu *pptr;
pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
GFP_USER | __GFP_NOWARN);
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
if (!pptr)
goto free_elems;
htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
pptr);
cond_resched();
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
}
skip_percpu_elems:
if (htab_is_lru(htab))
err = bpf_lru_init(&htab->lru,
htab->map.map_flags & BPF_F_NO_COMMON_LRU,
offsetof(struct htab_elem, hash) -
offsetof(struct htab_elem, lru_node),
htab_lru_map_delete_node,
htab);
else
err = pcpu_freelist_init(&htab->freelist);
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
if (err)
goto free_elems;
if (htab_is_lru(htab))
bpf_lru_populate(&htab->lru, htab->elems,
offsetof(struct htab_elem, lru_node),
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
htab->elem_size, num_entries);
else
pcpu_freelist_populate(&htab->freelist,
htab->elems + offsetof(struct htab_elem, fnode),
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
htab->elem_size, num_entries);
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
return 0;
free_elems:
htab_free_elems(htab);
return err;
}
static void prealloc_destroy(struct bpf_htab *htab)
{
htab_free_elems(htab);
if (htab_is_lru(htab))
bpf_lru_destroy(&htab->lru);
else
pcpu_freelist_destroy(&htab->freelist);
}
static int alloc_extra_elems(struct bpf_htab *htab)
{
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
struct htab_elem *__percpu *pptr, *l_new;
struct pcpu_freelist_node *l;
int cpu;
pptr = bpf_map_alloc_percpu(&htab->map, sizeof(struct htab_elem *), 8,
GFP_USER | __GFP_NOWARN);
if (!pptr)
return -ENOMEM;
for_each_possible_cpu(cpu) {
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
l = pcpu_freelist_pop(&htab->freelist);
/* pop will succeed, since prealloc_init()
* preallocated extra num_possible_cpus elements
*/
l_new = container_of(l, struct htab_elem, fnode);
*per_cpu_ptr(pptr, cpu) = l_new;
}
htab->extra_elems = pptr;
return 0;
}
/* Called from syscall */
static int htab_map_alloc_check(union bpf_attr *attr)
{
bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
/* percpu_lru means each cpu has its own LRU list.
* it is different from BPF_MAP_TYPE_PERCPU_HASH where
* the map's value itself is percpu. percpu_lru has
* nothing to do with the map's value.
*/
bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
bpf: Allow selecting numa node during map creation The current map creation API does not allow to provide the numa-node preference. The memory usually comes from where the map-creation-process is running. The performance is not ideal if the bpf_prog is known to always run in a numa node different from the map-creation-process. One of the use case is sharding on CPU to different LRU maps (i.e. an array of LRU maps). Here is the test result of map_perf_test on the INNER_LRU_HASH_PREALLOC test if we force the lru map used by CPU0 to be allocated from a remote numa node: [ The machine has 20 cores. CPU0-9 at node 0. CPU10-19 at node 1 ] ># taskset -c 10 ./map_perf_test 512 8 1260000 8000000 5:inner_lru_hash_map_perf pre-alloc 1628380 events per sec 4:inner_lru_hash_map_perf pre-alloc 1626396 events per sec 3:inner_lru_hash_map_perf pre-alloc 1626144 events per sec 6:inner_lru_hash_map_perf pre-alloc 1621657 events per sec 2:inner_lru_hash_map_perf pre-alloc 1621534 events per sec 1:inner_lru_hash_map_perf pre-alloc 1620292 events per sec 7:inner_lru_hash_map_perf pre-alloc 1613305 events per sec 0:inner_lru_hash_map_perf pre-alloc 1239150 events per sec #<<< After specifying numa node: ># taskset -c 10 ./map_perf_test 512 8 1260000 8000000 5:inner_lru_hash_map_perf pre-alloc 1629627 events per sec 3:inner_lru_hash_map_perf pre-alloc 1628057 events per sec 1:inner_lru_hash_map_perf pre-alloc 1623054 events per sec 6:inner_lru_hash_map_perf pre-alloc 1616033 events per sec 2:inner_lru_hash_map_perf pre-alloc 1614630 events per sec 4:inner_lru_hash_map_perf pre-alloc 1612651 events per sec 7:inner_lru_hash_map_perf pre-alloc 1609337 events per sec 0:inner_lru_hash_map_perf pre-alloc 1619340 events per sec #<<< This patch adds one field, numa_node, to the bpf_attr. Since numa node 0 is a valid node, a new flag BPF_F_NUMA_NODE is also added. The numa_node field is honored if and only if the BPF_F_NUMA_NODE flag is set. Numa node selection is not supported for percpu map. This patch does not change all the kmalloc. F.e. 'htab = kzalloc()' is not changed since the object is small enough to stay in the cache. Signed-off-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-18 18:28:00 +00:00
int numa_node = bpf_map_attr_numa_node(attr);
BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
offsetof(struct htab_elem, hash_node.pprev));
if (zero_seed && !capable(CAP_SYS_ADMIN))
/* Guard against local DoS, and discourage production use. */
return -EPERM;
if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
!bpf_map_flags_access_ok(attr->map_flags))
return -EINVAL;
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
if (!lru && percpu_lru)
return -EINVAL;
if (lru && !prealloc)
return -ENOTSUPP;
bpf: Allow selecting numa node during map creation The current map creation API does not allow to provide the numa-node preference. The memory usually comes from where the map-creation-process is running. The performance is not ideal if the bpf_prog is known to always run in a numa node different from the map-creation-process. One of the use case is sharding on CPU to different LRU maps (i.e. an array of LRU maps). Here is the test result of map_perf_test on the INNER_LRU_HASH_PREALLOC test if we force the lru map used by CPU0 to be allocated from a remote numa node: [ The machine has 20 cores. CPU0-9 at node 0. CPU10-19 at node 1 ] ># taskset -c 10 ./map_perf_test 512 8 1260000 8000000 5:inner_lru_hash_map_perf pre-alloc 1628380 events per sec 4:inner_lru_hash_map_perf pre-alloc 1626396 events per sec 3:inner_lru_hash_map_perf pre-alloc 1626144 events per sec 6:inner_lru_hash_map_perf pre-alloc 1621657 events per sec 2:inner_lru_hash_map_perf pre-alloc 1621534 events per sec 1:inner_lru_hash_map_perf pre-alloc 1620292 events per sec 7:inner_lru_hash_map_perf pre-alloc 1613305 events per sec 0:inner_lru_hash_map_perf pre-alloc 1239150 events per sec #<<< After specifying numa node: ># taskset -c 10 ./map_perf_test 512 8 1260000 8000000 5:inner_lru_hash_map_perf pre-alloc 1629627 events per sec 3:inner_lru_hash_map_perf pre-alloc 1628057 events per sec 1:inner_lru_hash_map_perf pre-alloc 1623054 events per sec 6:inner_lru_hash_map_perf pre-alloc 1616033 events per sec 2:inner_lru_hash_map_perf pre-alloc 1614630 events per sec 4:inner_lru_hash_map_perf pre-alloc 1612651 events per sec 7:inner_lru_hash_map_perf pre-alloc 1609337 events per sec 0:inner_lru_hash_map_perf pre-alloc 1619340 events per sec #<<< This patch adds one field, numa_node, to the bpf_attr. Since numa node 0 is a valid node, a new flag BPF_F_NUMA_NODE is also added. The numa_node field is honored if and only if the BPF_F_NUMA_NODE flag is set. Numa node selection is not supported for percpu map. This patch does not change all the kmalloc. F.e. 'htab = kzalloc()' is not changed since the object is small enough to stay in the cache. Signed-off-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-18 18:28:00 +00:00
if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
return -EINVAL;
bpf: Allow selecting numa node during map creation The current map creation API does not allow to provide the numa-node preference. The memory usually comes from where the map-creation-process is running. The performance is not ideal if the bpf_prog is known to always run in a numa node different from the map-creation-process. One of the use case is sharding on CPU to different LRU maps (i.e. an array of LRU maps). Here is the test result of map_perf_test on the INNER_LRU_HASH_PREALLOC test if we force the lru map used by CPU0 to be allocated from a remote numa node: [ The machine has 20 cores. CPU0-9 at node 0. CPU10-19 at node 1 ] ># taskset -c 10 ./map_perf_test 512 8 1260000 8000000 5:inner_lru_hash_map_perf pre-alloc 1628380 events per sec 4:inner_lru_hash_map_perf pre-alloc 1626396 events per sec 3:inner_lru_hash_map_perf pre-alloc 1626144 events per sec 6:inner_lru_hash_map_perf pre-alloc 1621657 events per sec 2:inner_lru_hash_map_perf pre-alloc 1621534 events per sec 1:inner_lru_hash_map_perf pre-alloc 1620292 events per sec 7:inner_lru_hash_map_perf pre-alloc 1613305 events per sec 0:inner_lru_hash_map_perf pre-alloc 1239150 events per sec #<<< After specifying numa node: ># taskset -c 10 ./map_perf_test 512 8 1260000 8000000 5:inner_lru_hash_map_perf pre-alloc 1629627 events per sec 3:inner_lru_hash_map_perf pre-alloc 1628057 events per sec 1:inner_lru_hash_map_perf pre-alloc 1623054 events per sec 6:inner_lru_hash_map_perf pre-alloc 1616033 events per sec 2:inner_lru_hash_map_perf pre-alloc 1614630 events per sec 4:inner_lru_hash_map_perf pre-alloc 1612651 events per sec 7:inner_lru_hash_map_perf pre-alloc 1609337 events per sec 0:inner_lru_hash_map_perf pre-alloc 1619340 events per sec #<<< This patch adds one field, numa_node, to the bpf_attr. Since numa node 0 is a valid node, a new flag BPF_F_NUMA_NODE is also added. The numa_node field is honored if and only if the BPF_F_NUMA_NODE flag is set. Numa node selection is not supported for percpu map. This patch does not change all the kmalloc. F.e. 'htab = kzalloc()' is not changed since the object is small enough to stay in the cache. Signed-off-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-18 18:28:00 +00:00
/* check sanity of attributes.
* value_size == 0 may be allowed in the future to use map as a set
*/
if (attr->max_entries == 0 || attr->key_size == 0 ||
attr->value_size == 0)
return -EINVAL;
if ((u64)attr->key_size + attr->value_size >= KMALLOC_MAX_SIZE -
sizeof(struct htab_elem))
/* if key_size + value_size is bigger, the user space won't be
* able to access the elements via bpf syscall. This check
* also makes sure that the elem_size doesn't overflow and it's
* kmalloc-able later in htab_map_update_elem()
*/
return -E2BIG;
return 0;
}
static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
{
bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
/* percpu_lru means each cpu has its own LRU list.
* it is different from BPF_MAP_TYPE_PERCPU_HASH where
* the map's value itself is percpu. percpu_lru has
* nothing to do with the map's value.
*/
bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
struct bpf_htab *htab;
int err, i;
htab = bpf_map_area_alloc(sizeof(*htab), NUMA_NO_NODE);
if (!htab)
return ERR_PTR(-ENOMEM);
bpf: Fix error path in htab_map_alloc() syzbot was able to trigger a use-after-free in htab_map_alloc() [1] htab_map_alloc() lacks a call to lockdep_unregister_key() in its error path. lockdep_register_key() and lockdep_unregister_key() can not fail, it seems better to use them right after htab allocation and before htab freeing, avoiding more goto/labels in htab_map_alloc() [1] BUG: KASAN: use-after-free in lockdep_register_key+0x356/0x3e0 kernel/locking/lockdep.c:1182 Read of size 8 at addr ffff88805fa67ad8 by task syz-executor.3/2356 CPU: 1 PID: 2356 Comm: syz-executor.3 Not tainted 5.9.0-syzkaller #0 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:77 [inline] dump_stack+0x107/0x163 lib/dump_stack.c:118 print_address_description.constprop.0.cold+0xae/0x4c8 mm/kasan/report.c:385 __kasan_report mm/kasan/report.c:545 [inline] kasan_report.cold+0x1f/0x37 mm/kasan/report.c:562 lockdep_register_key+0x356/0x3e0 kernel/locking/lockdep.c:1182 htab_init_buckets kernel/bpf/hashtab.c:144 [inline] htab_map_alloc+0x6c5/0x14a0 kernel/bpf/hashtab.c:521 find_and_alloc_map kernel/bpf/syscall.c:122 [inline] map_create kernel/bpf/syscall.c:825 [inline] __do_sys_bpf+0xa80/0x5180 kernel/bpf/syscall.c:4381 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x45deb9 Code: 0d b4 fb ff c3 66 2e 0f 1f 84 00 00 00 00 00 66 90 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 0f 83 db b3 fb ff c3 66 2e 0f 1f 84 00 00 00 00 RSP: 002b:00007f0eafee1c78 EFLAGS: 00000246 ORIG_RAX: 0000000000000141 RAX: ffffffffffffffda RBX: 0000000000001a00 RCX: 000000000045deb9 RDX: 0000000000000040 RSI: 0000000020000040 RDI: 405a020000000000 RBP: 000000000118bf60 R08: 0000000000000000 R09: 0000000000000000 R10: 0000000000000000 R11: 0000000000000246 R12: 000000000118bf2c R13: 00007ffd3cf9eabf R14: 00007f0eafee29c0 R15: 000000000118bf2c Allocated by task 2053: kasan_save_stack+0x1b/0x40 mm/kasan/common.c:48 kasan_set_track mm/kasan/common.c:56 [inline] __kasan_kmalloc.constprop.0+0xc2/0xd0 mm/kasan/common.c:461 kmalloc include/linux/slab.h:554 [inline] kzalloc include/linux/slab.h:666 [inline] htab_map_alloc+0xdf/0x14a0 kernel/bpf/hashtab.c:454 find_and_alloc_map kernel/bpf/syscall.c:122 [inline] map_create kernel/bpf/syscall.c:825 [inline] __do_sys_bpf+0xa80/0x5180 kernel/bpf/syscall.c:4381 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Freed by task 2053: kasan_save_stack+0x1b/0x40 mm/kasan/common.c:48 kasan_set_track+0x1c/0x30 mm/kasan/common.c:56 kasan_set_free_info+0x1b/0x30 mm/kasan/generic.c:355 __kasan_slab_free+0x102/0x140 mm/kasan/common.c:422 slab_free_hook mm/slub.c:1544 [inline] slab_free_freelist_hook+0x5d/0x150 mm/slub.c:1577 slab_free mm/slub.c:3142 [inline] kfree+0xdb/0x360 mm/slub.c:4124 htab_map_alloc+0x3f9/0x14a0 kernel/bpf/hashtab.c:549 find_and_alloc_map kernel/bpf/syscall.c:122 [inline] map_create kernel/bpf/syscall.c:825 [inline] __do_sys_bpf+0xa80/0x5180 kernel/bpf/syscall.c:4381 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46 entry_SYSCALL_64_after_hwframe+0x44/0xa9 The buggy address belongs to the object at ffff88805fa67800 which belongs to the cache kmalloc-1k of size 1024 The buggy address is located 728 bytes inside of 1024-byte region [ffff88805fa67800, ffff88805fa67c00) The buggy address belongs to the page: page:000000003c5582c4 refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x5fa60 head:000000003c5582c4 order:3 compound_mapcount:0 compound_pincount:0 flags: 0xfff00000010200(slab|head) raw: 00fff00000010200 ffffea0000bc1200 0000000200000002 ffff888010041140 raw: 0000000000000000 0000000000100010 00000001ffffffff 0000000000000000 page dumped because: kasan: bad access detected Memory state around the buggy address: ffff88805fa67980: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ffff88805fa67a00: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ^ ffff88805fa67b00: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ffff88805fa67b80: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb Fixes: c50eb518e262 ("bpf: Use separate lockdep class for each hashtab") Reported-by: syzbot <syzkaller@googlegroups.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20201102114100.3103180-1-eric.dumazet@gmail.com
2020-11-02 11:41:00 +00:00
lockdep_register_key(&htab->lockdep_key);
bpf_map_init_from_attr(&htab->map, attr);
if (percpu_lru) {
/* ensure each CPU's lru list has >=1 elements.
* since we are at it, make each lru list has the same
* number of elements.
*/
htab->map.max_entries = roundup(attr->max_entries,
num_possible_cpus());
if (htab->map.max_entries < attr->max_entries)
htab->map.max_entries = rounddown(attr->max_entries,
num_possible_cpus());
}
/* hash table size must be power of 2; roundup_pow_of_two() can overflow
* into UB on 32-bit arches, so check that first
*/
err = -E2BIG;
if (htab->map.max_entries > 1UL << 31)
goto free_htab;
htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
bpf: fix allocation warnings in bpf maps and integer overflow For large map->value_size the user space can trigger memory allocation warnings like: WARNING: CPU: 2 PID: 11122 at mm/page_alloc.c:2989 __alloc_pages_nodemask+0x695/0x14e0() Call Trace: [< inline >] __dump_stack lib/dump_stack.c:15 [<ffffffff82743b56>] dump_stack+0x68/0x92 lib/dump_stack.c:50 [<ffffffff81244ec9>] warn_slowpath_common+0xd9/0x140 kernel/panic.c:460 [<ffffffff812450f9>] warn_slowpath_null+0x29/0x30 kernel/panic.c:493 [< inline >] __alloc_pages_slowpath mm/page_alloc.c:2989 [<ffffffff81554e95>] __alloc_pages_nodemask+0x695/0x14e0 mm/page_alloc.c:3235 [<ffffffff816188fe>] alloc_pages_current+0xee/0x340 mm/mempolicy.c:2055 [< inline >] alloc_pages include/linux/gfp.h:451 [<ffffffff81550706>] alloc_kmem_pages+0x16/0xf0 mm/page_alloc.c:3414 [<ffffffff815a1c89>] kmalloc_order+0x19/0x60 mm/slab_common.c:1007 [<ffffffff815a1cef>] kmalloc_order_trace+0x1f/0xa0 mm/slab_common.c:1018 [< inline >] kmalloc_large include/linux/slab.h:390 [<ffffffff81627784>] __kmalloc+0x234/0x250 mm/slub.c:3525 [< inline >] kmalloc include/linux/slab.h:463 [< inline >] map_update_elem kernel/bpf/syscall.c:288 [< inline >] SYSC_bpf kernel/bpf/syscall.c:744 To avoid never succeeding kmalloc with order >= MAX_ORDER check that elem->value_size and computed elem_size are within limits for both hash and array type maps. Also add __GFP_NOWARN to kmalloc(value_size | elem_size) to avoid OOM warnings. Note kmalloc(key_size) is highly unlikely to trigger OOM, since key_size <= 512, so keep those kmalloc-s as-is. Large value_size can cause integer overflows in elem_size and map.pages formulas, so check for that as well. Fixes: aaac3ba95e4c ("bpf: charge user for creation of BPF maps and programs") Reported-by: Dmitry Vyukov <dvyukov@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-30 00:59:35 +00:00
htab->elem_size = sizeof(struct htab_elem) +
round_up(htab->map.key_size, 8);
if (percpu)
htab->elem_size += sizeof(void *);
else
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
htab->elem_size += round_up(htab->map.value_size, 8);
bpf: fix allocation warnings in bpf maps and integer overflow For large map->value_size the user space can trigger memory allocation warnings like: WARNING: CPU: 2 PID: 11122 at mm/page_alloc.c:2989 __alloc_pages_nodemask+0x695/0x14e0() Call Trace: [< inline >] __dump_stack lib/dump_stack.c:15 [<ffffffff82743b56>] dump_stack+0x68/0x92 lib/dump_stack.c:50 [<ffffffff81244ec9>] warn_slowpath_common+0xd9/0x140 kernel/panic.c:460 [<ffffffff812450f9>] warn_slowpath_null+0x29/0x30 kernel/panic.c:493 [< inline >] __alloc_pages_slowpath mm/page_alloc.c:2989 [<ffffffff81554e95>] __alloc_pages_nodemask+0x695/0x14e0 mm/page_alloc.c:3235 [<ffffffff816188fe>] alloc_pages_current+0xee/0x340 mm/mempolicy.c:2055 [< inline >] alloc_pages include/linux/gfp.h:451 [<ffffffff81550706>] alloc_kmem_pages+0x16/0xf0 mm/page_alloc.c:3414 [<ffffffff815a1c89>] kmalloc_order+0x19/0x60 mm/slab_common.c:1007 [<ffffffff815a1cef>] kmalloc_order_trace+0x1f/0xa0 mm/slab_common.c:1018 [< inline >] kmalloc_large include/linux/slab.h:390 [<ffffffff81627784>] __kmalloc+0x234/0x250 mm/slub.c:3525 [< inline >] kmalloc include/linux/slab.h:463 [< inline >] map_update_elem kernel/bpf/syscall.c:288 [< inline >] SYSC_bpf kernel/bpf/syscall.c:744 To avoid never succeeding kmalloc with order >= MAX_ORDER check that elem->value_size and computed elem_size are within limits for both hash and array type maps. Also add __GFP_NOWARN to kmalloc(value_size | elem_size) to avoid OOM warnings. Note kmalloc(key_size) is highly unlikely to trigger OOM, since key_size <= 512, so keep those kmalloc-s as-is. Large value_size can cause integer overflows in elem_size and map.pages formulas, so check for that as well. Fixes: aaac3ba95e4c ("bpf: charge user for creation of BPF maps and programs") Reported-by: Dmitry Vyukov <dvyukov@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-30 00:59:35 +00:00
/* check for u32 overflow */
if (htab->n_buckets > U32_MAX / sizeof(struct bucket))
goto free_htab;
err = bpf_map_init_elem_count(&htab->map);
if (err)
goto free_htab;
bpf: fix allocation warnings in bpf maps and integer overflow For large map->value_size the user space can trigger memory allocation warnings like: WARNING: CPU: 2 PID: 11122 at mm/page_alloc.c:2989 __alloc_pages_nodemask+0x695/0x14e0() Call Trace: [< inline >] __dump_stack lib/dump_stack.c:15 [<ffffffff82743b56>] dump_stack+0x68/0x92 lib/dump_stack.c:50 [<ffffffff81244ec9>] warn_slowpath_common+0xd9/0x140 kernel/panic.c:460 [<ffffffff812450f9>] warn_slowpath_null+0x29/0x30 kernel/panic.c:493 [< inline >] __alloc_pages_slowpath mm/page_alloc.c:2989 [<ffffffff81554e95>] __alloc_pages_nodemask+0x695/0x14e0 mm/page_alloc.c:3235 [<ffffffff816188fe>] alloc_pages_current+0xee/0x340 mm/mempolicy.c:2055 [< inline >] alloc_pages include/linux/gfp.h:451 [<ffffffff81550706>] alloc_kmem_pages+0x16/0xf0 mm/page_alloc.c:3414 [<ffffffff815a1c89>] kmalloc_order+0x19/0x60 mm/slab_common.c:1007 [<ffffffff815a1cef>] kmalloc_order_trace+0x1f/0xa0 mm/slab_common.c:1018 [< inline >] kmalloc_large include/linux/slab.h:390 [<ffffffff81627784>] __kmalloc+0x234/0x250 mm/slub.c:3525 [< inline >] kmalloc include/linux/slab.h:463 [< inline >] map_update_elem kernel/bpf/syscall.c:288 [< inline >] SYSC_bpf kernel/bpf/syscall.c:744 To avoid never succeeding kmalloc with order >= MAX_ORDER check that elem->value_size and computed elem_size are within limits for both hash and array type maps. Also add __GFP_NOWARN to kmalloc(value_size | elem_size) to avoid OOM warnings. Note kmalloc(key_size) is highly unlikely to trigger OOM, since key_size <= 512, so keep those kmalloc-s as-is. Large value_size can cause integer overflows in elem_size and map.pages formulas, so check for that as well. Fixes: aaac3ba95e4c ("bpf: charge user for creation of BPF maps and programs") Reported-by: Dmitry Vyukov <dvyukov@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-30 00:59:35 +00:00
err = -ENOMEM;
htab->buckets = bpf_map_area_alloc(htab->n_buckets *
bpf: Allow selecting numa node during map creation The current map creation API does not allow to provide the numa-node preference. The memory usually comes from where the map-creation-process is running. The performance is not ideal if the bpf_prog is known to always run in a numa node different from the map-creation-process. One of the use case is sharding on CPU to different LRU maps (i.e. an array of LRU maps). Here is the test result of map_perf_test on the INNER_LRU_HASH_PREALLOC test if we force the lru map used by CPU0 to be allocated from a remote numa node: [ The machine has 20 cores. CPU0-9 at node 0. CPU10-19 at node 1 ] ># taskset -c 10 ./map_perf_test 512 8 1260000 8000000 5:inner_lru_hash_map_perf pre-alloc 1628380 events per sec 4:inner_lru_hash_map_perf pre-alloc 1626396 events per sec 3:inner_lru_hash_map_perf pre-alloc 1626144 events per sec 6:inner_lru_hash_map_perf pre-alloc 1621657 events per sec 2:inner_lru_hash_map_perf pre-alloc 1621534 events per sec 1:inner_lru_hash_map_perf pre-alloc 1620292 events per sec 7:inner_lru_hash_map_perf pre-alloc 1613305 events per sec 0:inner_lru_hash_map_perf pre-alloc 1239150 events per sec #<<< After specifying numa node: ># taskset -c 10 ./map_perf_test 512 8 1260000 8000000 5:inner_lru_hash_map_perf pre-alloc 1629627 events per sec 3:inner_lru_hash_map_perf pre-alloc 1628057 events per sec 1:inner_lru_hash_map_perf pre-alloc 1623054 events per sec 6:inner_lru_hash_map_perf pre-alloc 1616033 events per sec 2:inner_lru_hash_map_perf pre-alloc 1614630 events per sec 4:inner_lru_hash_map_perf pre-alloc 1612651 events per sec 7:inner_lru_hash_map_perf pre-alloc 1609337 events per sec 0:inner_lru_hash_map_perf pre-alloc 1619340 events per sec #<<< This patch adds one field, numa_node, to the bpf_attr. Since numa node 0 is a valid node, a new flag BPF_F_NUMA_NODE is also added. The numa_node field is honored if and only if the BPF_F_NUMA_NODE flag is set. Numa node selection is not supported for percpu map. This patch does not change all the kmalloc. F.e. 'htab = kzalloc()' is not changed since the object is small enough to stay in the cache. Signed-off-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-18 18:28:00 +00:00
sizeof(struct bucket),
htab->map.numa_node);
if (!htab->buckets)
goto free_elem_count;
for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++) {
htab->map_locked[i] = bpf_map_alloc_percpu(&htab->map,
sizeof(int),
sizeof(int),
GFP_USER);
if (!htab->map_locked[i])
goto free_map_locked;
}
if (htab->map.map_flags & BPF_F_ZERO_SEED)
htab->hashrnd = 0;
else
htab->hashrnd = get_random_u32();
htab_init_buckets(htab);
/* compute_batch_value() computes batch value as num_online_cpus() * 2
* and __percpu_counter_compare() needs
* htab->max_entries - cur_number_of_elems to be more than batch * num_online_cpus()
* for percpu_counter to be faster than atomic_t. In practice the average bpf
* hash map size is 10k, which means that a system with 64 cpus will fill
* hashmap to 20% of 10k before percpu_counter becomes ineffective. Therefore
* define our own batch count as 32 then 10k hash map can be filled up to 80%:
* 10k - 8k > 32 _batch_ * 64 _cpus_
* and __percpu_counter_compare() will still be fast. At that point hash map
* collisions will dominate its performance anyway. Assume that hash map filled
* to 50+% isn't going to be O(1) and use the following formula to choose
* between percpu_counter and atomic_t.
*/
#define PERCPU_COUNTER_BATCH 32
if (attr->max_entries / 2 > num_online_cpus() * PERCPU_COUNTER_BATCH)
htab->use_percpu_counter = true;
if (htab->use_percpu_counter) {
err = percpu_counter_init(&htab->pcount, 0, GFP_KERNEL);
if (err)
goto free_map_locked;
}
if (prealloc) {
err = prealloc_init(htab);
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
if (err)
goto free_map_locked;
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
if (!percpu && !lru) {
/* lru itself can remove the least used element, so
* there is no need for an extra elem during map_update.
*/
err = alloc_extra_elems(htab);
if (err)
goto free_prealloc;
}
} else {
err = bpf_mem_alloc_init(&htab->ma, htab->elem_size, false);
if (err)
goto free_map_locked;
if (percpu) {
err = bpf_mem_alloc_init(&htab->pcpu_ma,
round_up(htab->map.value_size, 8), true);
if (err)
goto free_map_locked;
}
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
}
return &htab->map;
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
free_prealloc:
prealloc_destroy(htab);
free_map_locked:
if (htab->use_percpu_counter)
percpu_counter_destroy(&htab->pcount);
for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
free_percpu(htab->map_locked[i]);
bpf_map_area_free(htab->buckets);
bpf_mem_alloc_destroy(&htab->pcpu_ma);
bpf_mem_alloc_destroy(&htab->ma);
free_elem_count:
bpf_map_free_elem_count(&htab->map);
free_htab:
bpf: Fix error path in htab_map_alloc() syzbot was able to trigger a use-after-free in htab_map_alloc() [1] htab_map_alloc() lacks a call to lockdep_unregister_key() in its error path. lockdep_register_key() and lockdep_unregister_key() can not fail, it seems better to use them right after htab allocation and before htab freeing, avoiding more goto/labels in htab_map_alloc() [1] BUG: KASAN: use-after-free in lockdep_register_key+0x356/0x3e0 kernel/locking/lockdep.c:1182 Read of size 8 at addr ffff88805fa67ad8 by task syz-executor.3/2356 CPU: 1 PID: 2356 Comm: syz-executor.3 Not tainted 5.9.0-syzkaller #0 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:77 [inline] dump_stack+0x107/0x163 lib/dump_stack.c:118 print_address_description.constprop.0.cold+0xae/0x4c8 mm/kasan/report.c:385 __kasan_report mm/kasan/report.c:545 [inline] kasan_report.cold+0x1f/0x37 mm/kasan/report.c:562 lockdep_register_key+0x356/0x3e0 kernel/locking/lockdep.c:1182 htab_init_buckets kernel/bpf/hashtab.c:144 [inline] htab_map_alloc+0x6c5/0x14a0 kernel/bpf/hashtab.c:521 find_and_alloc_map kernel/bpf/syscall.c:122 [inline] map_create kernel/bpf/syscall.c:825 [inline] __do_sys_bpf+0xa80/0x5180 kernel/bpf/syscall.c:4381 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x45deb9 Code: 0d b4 fb ff c3 66 2e 0f 1f 84 00 00 00 00 00 66 90 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 0f 83 db b3 fb ff c3 66 2e 0f 1f 84 00 00 00 00 RSP: 002b:00007f0eafee1c78 EFLAGS: 00000246 ORIG_RAX: 0000000000000141 RAX: ffffffffffffffda RBX: 0000000000001a00 RCX: 000000000045deb9 RDX: 0000000000000040 RSI: 0000000020000040 RDI: 405a020000000000 RBP: 000000000118bf60 R08: 0000000000000000 R09: 0000000000000000 R10: 0000000000000000 R11: 0000000000000246 R12: 000000000118bf2c R13: 00007ffd3cf9eabf R14: 00007f0eafee29c0 R15: 000000000118bf2c Allocated by task 2053: kasan_save_stack+0x1b/0x40 mm/kasan/common.c:48 kasan_set_track mm/kasan/common.c:56 [inline] __kasan_kmalloc.constprop.0+0xc2/0xd0 mm/kasan/common.c:461 kmalloc include/linux/slab.h:554 [inline] kzalloc include/linux/slab.h:666 [inline] htab_map_alloc+0xdf/0x14a0 kernel/bpf/hashtab.c:454 find_and_alloc_map kernel/bpf/syscall.c:122 [inline] map_create kernel/bpf/syscall.c:825 [inline] __do_sys_bpf+0xa80/0x5180 kernel/bpf/syscall.c:4381 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Freed by task 2053: kasan_save_stack+0x1b/0x40 mm/kasan/common.c:48 kasan_set_track+0x1c/0x30 mm/kasan/common.c:56 kasan_set_free_info+0x1b/0x30 mm/kasan/generic.c:355 __kasan_slab_free+0x102/0x140 mm/kasan/common.c:422 slab_free_hook mm/slub.c:1544 [inline] slab_free_freelist_hook+0x5d/0x150 mm/slub.c:1577 slab_free mm/slub.c:3142 [inline] kfree+0xdb/0x360 mm/slub.c:4124 htab_map_alloc+0x3f9/0x14a0 kernel/bpf/hashtab.c:549 find_and_alloc_map kernel/bpf/syscall.c:122 [inline] map_create kernel/bpf/syscall.c:825 [inline] __do_sys_bpf+0xa80/0x5180 kernel/bpf/syscall.c:4381 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46 entry_SYSCALL_64_after_hwframe+0x44/0xa9 The buggy address belongs to the object at ffff88805fa67800 which belongs to the cache kmalloc-1k of size 1024 The buggy address is located 728 bytes inside of 1024-byte region [ffff88805fa67800, ffff88805fa67c00) The buggy address belongs to the page: page:000000003c5582c4 refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x5fa60 head:000000003c5582c4 order:3 compound_mapcount:0 compound_pincount:0 flags: 0xfff00000010200(slab|head) raw: 00fff00000010200 ffffea0000bc1200 0000000200000002 ffff888010041140 raw: 0000000000000000 0000000000100010 00000001ffffffff 0000000000000000 page dumped because: kasan: bad access detected Memory state around the buggy address: ffff88805fa67980: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ffff88805fa67a00: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ^ ffff88805fa67b00: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ffff88805fa67b80: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb Fixes: c50eb518e262 ("bpf: Use separate lockdep class for each hashtab") Reported-by: syzbot <syzkaller@googlegroups.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20201102114100.3103180-1-eric.dumazet@gmail.com
2020-11-02 11:41:00 +00:00
lockdep_unregister_key(&htab->lockdep_key);
bpf_map_area_free(htab);
return ERR_PTR(err);
}
static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
{
bpf: optimize hashmap lookups when key_size is divisible by 4 The BPF hashmap uses the jhash() hash function. There is an optimized version of this hash function which may be used if hash size is a multiple of 4. Apply this optimization to the hashmap in a similar way as it is done in the bloom filter map. On practice the optimization is only noticeable for smaller key sizes, which, however, is sufficient for many applications. An example is listed in the following table of measurements (a hashmap of 65536 elements was used): -------------------------------------------------------------------- | key_size | fullness | lookups /sec | lookups (opt) /sec | gain | -------------------------------------------------------------------- | 4 | 25% | 42.990M | 46.000M | 7.0% | | 4 | 50% | 37.910M | 39.094M | 3.1% | | 4 | 75% | 34.486M | 36.124M | 4.7% | | 4 | 100% | 31.760M | 32.719M | 3.0% | -------------------------------------------------------------------- | 8 | 25% | 43.855M | 49.626M | 13.2% | | 8 | 50% | 38.328M | 42.152M | 10.0% | | 8 | 75% | 34.483M | 38.088M | 10.5% | | 8 | 100% | 31.306M | 34.686M | 10.8% | -------------------------------------------------------------------- | 12 | 25% | 38.398M | 43.770M | 14.0% | | 12 | 50% | 33.336M | 37.712M | 13.1% | | 12 | 75% | 29.917M | 34.440M | 15.1% | | 12 | 100% | 27.322M | 30.480M | 11.6% | -------------------------------------------------------------------- | 16 | 25% | 41.491M | 41.921M | 1.0% | | 16 | 50% | 36.206M | 36.474M | 0.7% | | 16 | 75% | 32.529M | 33.027M | 1.5% | | 16 | 100% | 29.581M | 30.325M | 2.5% | -------------------------------------------------------------------- | 20 | 25% | 34.240M | 36.787M | 7.4% | | 20 | 50% | 30.328M | 32.663M | 7.7% | | 20 | 75% | 27.536M | 29.354M | 6.6% | | 20 | 100% | 24.847M | 26.505M | 6.7% | -------------------------------------------------------------------- | 24 | 25% | 36.329M | 40.608M | 11.8% | | 24 | 50% | 31.444M | 35.059M | 11.5% | | 24 | 75% | 28.426M | 31.452M | 10.6% | | 24 | 100% | 26.278M | 28.741M | 9.4% | -------------------------------------------------------------------- | 28 | 25% | 31.540M | 31.944M | 1.3% | | 28 | 50% | 27.739M | 28.063M | 1.2% | | 28 | 75% | 24.993M | 25.814M | 3.3% | | 28 | 100% | 23.513M | 23.500M | -0.1% | -------------------------------------------------------------------- | 32 | 25% | 32.116M | 33.953M | 5.7% | | 32 | 50% | 28.879M | 29.859M | 3.4% | | 32 | 75% | 26.227M | 26.948M | 2.7% | | 32 | 100% | 23.829M | 24.613M | 3.3% | -------------------------------------------------------------------- | 64 | 25% | 22.535M | 22.554M | 0.1% | | 64 | 50% | 20.471M | 20.675M | 1.0% | | 64 | 75% | 19.077M | 19.146M | 0.4% | | 64 | 100% | 17.710M | 18.131M | 2.4% | -------------------------------------------------------------------- The following script was used to gather the results (SMT & frequency off): cd tools/testing/selftests/bpf for key_size in 4 8 12 16 20 24 28 32 64; do for nr_entries in `seq 16384 16384 65536`; do fullness=$(printf '%3s' $((nr_entries*100/65536))) echo -n "key_size=$key_size: $fullness% full: " sudo ./bench -d2 -a bpf-hashmap-lookup --key_size=$key_size --nr_entries=$nr_entries --max_entries=65536 --nr_loops=2000000 --map_flags=0x40 | grep cpu done echo done Signed-off-by: Anton Protopopov <aspsk@isovalent.com> Link: https://lore.kernel.org/r/20230401200602.3275-1-aspsk@isovalent.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-04-01 20:06:02 +00:00
if (likely(key_len % 4 == 0))
return jhash2(key, key_len / 4, hashrnd);
return jhash(key, key_len, hashrnd);
}
static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
{
return &htab->buckets[hash & (htab->n_buckets - 1)];
}
static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
{
return &__select_bucket(htab, hash)->head;
}
/* this lookup function can only be called with bucket lock taken */
static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
void *key, u32 key_size)
{
struct hlist_nulls_node *n;
struct htab_elem *l;
hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
if (l->hash == hash && !memcmp(&l->key, key, key_size))
return l;
return NULL;
}
/* can be called without bucket lock. it will repeat the loop in
* the unlikely event when elements moved from one bucket into another
* while link list is being walked
*/
static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
u32 hash, void *key,
u32 key_size, u32 n_buckets)
{
struct hlist_nulls_node *n;
struct htab_elem *l;
again:
hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
if (l->hash == hash && !memcmp(&l->key, key, key_size))
return l;
if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
goto again;
return NULL;
}
/* Called from syscall or from eBPF program directly, so
* arguments have to match bpf_map_lookup_elem() exactly.
* The return value is adjusted by BPF instructions
* in htab_map_gen_lookup().
*/
static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
struct hlist_nulls_head *head;
struct htab_elem *l;
u32 hash, key_size;
WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
!rcu_read_lock_bh_held());
key_size = map->key_size;
hash = htab_map_hash(key, key_size, htab->hashrnd);
head = select_bucket(htab, hash);
l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
return l;
}
static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
{
struct htab_elem *l = __htab_map_lookup_elem(map, key);
if (l)
return l->key + round_up(map->key_size, 8);
return NULL;
}
/* inline bpf_map_lookup_elem() call.
* Instead of:
* bpf_prog
* bpf_map_lookup_elem
* map->ops->map_lookup_elem
* htab_map_lookup_elem
* __htab_map_lookup_elem
* do:
* bpf_prog
* __htab_map_lookup_elem
*/
bpf: Allow for map-in-map with dynamic inner array map entries Recent work in f4d05259213f ("bpf: Add map_meta_equal map ops") and 134fede4eecf ("bpf: Relax max_entries check for most of the inner map types") added support for dynamic inner max elements for most map-in-map types. Exceptions were maps like array or prog array where the map_gen_lookup() callback uses the maps' max_entries field as a constant when emitting instructions. We recently implemented Maglev consistent hashing into Cilium's load balancer which uses map-in-map with an outer map being hash and inner being array holding the Maglev backend table for each service. This has been designed this way in order to reduce overall memory consumption given the outer hash map allows to avoid preallocating a large, flat memory area for all services. Also, the number of service mappings is not always known a-priori. The use case for dynamic inner array map entries is to further reduce memory overhead, for example, some services might just have a small number of back ends while others could have a large number. Right now the Maglev backend table for small and large number of backends would need to have the same inner array map entries which adds a lot of unneeded overhead. Dynamic inner array map entries can be realized by avoiding the inlined code generation for their lookup. The lookup will still be efficient since it will be calling into array_map_lookup_elem() directly and thus avoiding retpoline. The patch adds a BPF_F_INNER_MAP flag to map creation which therefore skips inline code generation and relaxes array_map_meta_equal() check to ignore both maps' max_entries. This also still allows to have faster lookups for map-in-map when BPF_F_INNER_MAP is not specified and hence dynamic max_entries not needed. Example code generation where inner map is dynamic sized array: # bpftool p d x i 125 int handle__sys_enter(void * ctx): ; int handle__sys_enter(void *ctx) 0: (b4) w1 = 0 ; int key = 0; 1: (63) *(u32 *)(r10 -4) = r1 2: (bf) r2 = r10 ; 3: (07) r2 += -4 ; inner_map = bpf_map_lookup_elem(&outer_arr_dyn, &key); 4: (18) r1 = map[id:468] 6: (07) r1 += 272 7: (61) r0 = *(u32 *)(r2 +0) 8: (35) if r0 >= 0x3 goto pc+5 9: (67) r0 <<= 3 10: (0f) r0 += r1 11: (79) r0 = *(u64 *)(r0 +0) 12: (15) if r0 == 0x0 goto pc+1 13: (05) goto pc+1 14: (b7) r0 = 0 15: (b4) w6 = -1 ; if (!inner_map) 16: (15) if r0 == 0x0 goto pc+6 17: (bf) r2 = r10 ; 18: (07) r2 += -4 ; val = bpf_map_lookup_elem(inner_map, &key); 19: (bf) r1 = r0 | No inlining but instead 20: (85) call array_map_lookup_elem#149280 | call to array_map_lookup_elem() ; return val ? *val : -1; | for inner array lookup. 21: (15) if r0 == 0x0 goto pc+1 ; return val ? *val : -1; 22: (61) r6 = *(u32 *)(r0 +0) ; } 23: (bc) w0 = w6 24: (95) exit Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Andrii Nakryiko <andrii@kernel.org> Link: https://lore.kernel.org/bpf/20201010234006.7075-4-daniel@iogearbox.net
2020-10-10 23:40:03 +00:00
static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
{
struct bpf_insn *insn = insn_buf;
const int ret = BPF_REG_0;
bpf: avoid retpoline for lookup/update/delete calls on maps While some of the BPF map lookup helpers provide a ->map_gen_lookup() callback for inlining the map lookup altogether it is not available for every map, so the remaining ones have to call bpf_map_lookup_elem() helper which does a dispatch to map->ops->map_lookup_elem(). In times of retpolines, this will control and trap speculative execution rather than letting it do its work for the indirect call and will therefore cause a slowdown. Likewise, bpf_map_update_elem() and bpf_map_delete_elem() do not have an inlined version and need to call into their map->ops->map_update_elem() resp. map->ops->map_delete_elem() handlers. Before: # bpftool prog dump xlated id 1 0: (bf) r2 = r10 1: (07) r2 += -8 2: (7a) *(u64 *)(r2 +0) = 0 3: (18) r1 = map[id:1] 5: (85) call __htab_map_lookup_elem#232656 6: (15) if r0 == 0x0 goto pc+4 7: (71) r1 = *(u8 *)(r0 +35) 8: (55) if r1 != 0x0 goto pc+1 9: (72) *(u8 *)(r0 +35) = 1 10: (07) r0 += 56 11: (15) if r0 == 0x0 goto pc+4 12: (bf) r2 = r0 13: (18) r1 = map[id:1] 15: (85) call bpf_map_delete_elem#215008 <-- indirect call via 16: (95) exit helper After: # bpftool prog dump xlated id 1 0: (bf) r2 = r10 1: (07) r2 += -8 2: (7a) *(u64 *)(r2 +0) = 0 3: (18) r1 = map[id:1] 5: (85) call __htab_map_lookup_elem#233328 6: (15) if r0 == 0x0 goto pc+4 7: (71) r1 = *(u8 *)(r0 +35) 8: (55) if r1 != 0x0 goto pc+1 9: (72) *(u8 *)(r0 +35) = 1 10: (07) r0 += 56 11: (15) if r0 == 0x0 goto pc+4 12: (bf) r2 = r0 13: (18) r1 = map[id:1] 15: (85) call htab_lru_map_delete_elem#238240 <-- direct call 16: (95) exit In all three lookup/update/delete cases however we can use the actual address of the map callback directly if we find that there's only a single path with a map pointer leading to the helper call, meaning when the map pointer has not been poisoned from verifier side. Example code can be seen above for the delete case. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2018-06-02 21:06:35 +00:00
BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
(void *(*)(struct bpf_map *map, void *key))NULL));
*insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
offsetof(struct htab_elem, key) +
round_up(map->key_size, 8));
return insn - insn_buf;
}
bpf, lru: avoid messing with eviction heuristics upon syscall lookup One of the biggest issues we face right now with picking LRU map over regular hash table is that a map walk out of user space, for example, to just dump the existing entries or to remove certain ones, will completely mess up LRU eviction heuristics and wrong entries such as just created ones will get evicted instead. The reason for this is that we mark an entry as "in use" via bpf_lru_node_set_ref() from system call lookup side as well. Thus upon walk, all entries are being marked, so information of actual least recently used ones are "lost". In case of Cilium where it can be used (besides others) as a BPF based connection tracker, this current behavior causes disruption upon control plane changes that need to walk the map from user space to evict certain entries. Discussion result from bpfconf [0] was that we should simply just remove marking from system call side as no good use case could be found where it's actually needed there. Therefore this patch removes marking for regular LRU and per-CPU flavor. If there ever should be a need in future, the behavior could be selected via map creation flag, but due to mentioned reason we avoid this here. [0] http://vger.kernel.org/bpfconf.html Fixes: 29ba732acbee ("bpf: Add BPF_MAP_TYPE_LRU_HASH") Fixes: 8f8449384ec3 ("bpf: Add BPF_MAP_TYPE_LRU_PERCPU_HASH") Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2019-05-13 23:18:56 +00:00
static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
void *key, const bool mark)
{
struct htab_elem *l = __htab_map_lookup_elem(map, key);
if (l) {
bpf, lru: avoid messing with eviction heuristics upon syscall lookup One of the biggest issues we face right now with picking LRU map over regular hash table is that a map walk out of user space, for example, to just dump the existing entries or to remove certain ones, will completely mess up LRU eviction heuristics and wrong entries such as just created ones will get evicted instead. The reason for this is that we mark an entry as "in use" via bpf_lru_node_set_ref() from system call lookup side as well. Thus upon walk, all entries are being marked, so information of actual least recently used ones are "lost". In case of Cilium where it can be used (besides others) as a BPF based connection tracker, this current behavior causes disruption upon control plane changes that need to walk the map from user space to evict certain entries. Discussion result from bpfconf [0] was that we should simply just remove marking from system call side as no good use case could be found where it's actually needed there. Therefore this patch removes marking for regular LRU and per-CPU flavor. If there ever should be a need in future, the behavior could be selected via map creation flag, but due to mentioned reason we avoid this here. [0] http://vger.kernel.org/bpfconf.html Fixes: 29ba732acbee ("bpf: Add BPF_MAP_TYPE_LRU_HASH") Fixes: 8f8449384ec3 ("bpf: Add BPF_MAP_TYPE_LRU_PERCPU_HASH") Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2019-05-13 23:18:56 +00:00
if (mark)
bpf_lru_node_set_ref(&l->lru_node);
return l->key + round_up(map->key_size, 8);
}
return NULL;
}
bpf, lru: avoid messing with eviction heuristics upon syscall lookup One of the biggest issues we face right now with picking LRU map over regular hash table is that a map walk out of user space, for example, to just dump the existing entries or to remove certain ones, will completely mess up LRU eviction heuristics and wrong entries such as just created ones will get evicted instead. The reason for this is that we mark an entry as "in use" via bpf_lru_node_set_ref() from system call lookup side as well. Thus upon walk, all entries are being marked, so information of actual least recently used ones are "lost". In case of Cilium where it can be used (besides others) as a BPF based connection tracker, this current behavior causes disruption upon control plane changes that need to walk the map from user space to evict certain entries. Discussion result from bpfconf [0] was that we should simply just remove marking from system call side as no good use case could be found where it's actually needed there. Therefore this patch removes marking for regular LRU and per-CPU flavor. If there ever should be a need in future, the behavior could be selected via map creation flag, but due to mentioned reason we avoid this here. [0] http://vger.kernel.org/bpfconf.html Fixes: 29ba732acbee ("bpf: Add BPF_MAP_TYPE_LRU_HASH") Fixes: 8f8449384ec3 ("bpf: Add BPF_MAP_TYPE_LRU_PERCPU_HASH") Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2019-05-13 23:18:56 +00:00
static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
{
return __htab_lru_map_lookup_elem(map, key, true);
}
static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
{
return __htab_lru_map_lookup_elem(map, key, false);
}
bpf: Allow for map-in-map with dynamic inner array map entries Recent work in f4d05259213f ("bpf: Add map_meta_equal map ops") and 134fede4eecf ("bpf: Relax max_entries check for most of the inner map types") added support for dynamic inner max elements for most map-in-map types. Exceptions were maps like array or prog array where the map_gen_lookup() callback uses the maps' max_entries field as a constant when emitting instructions. We recently implemented Maglev consistent hashing into Cilium's load balancer which uses map-in-map with an outer map being hash and inner being array holding the Maglev backend table for each service. This has been designed this way in order to reduce overall memory consumption given the outer hash map allows to avoid preallocating a large, flat memory area for all services. Also, the number of service mappings is not always known a-priori. The use case for dynamic inner array map entries is to further reduce memory overhead, for example, some services might just have a small number of back ends while others could have a large number. Right now the Maglev backend table for small and large number of backends would need to have the same inner array map entries which adds a lot of unneeded overhead. Dynamic inner array map entries can be realized by avoiding the inlined code generation for their lookup. The lookup will still be efficient since it will be calling into array_map_lookup_elem() directly and thus avoiding retpoline. The patch adds a BPF_F_INNER_MAP flag to map creation which therefore skips inline code generation and relaxes array_map_meta_equal() check to ignore both maps' max_entries. This also still allows to have faster lookups for map-in-map when BPF_F_INNER_MAP is not specified and hence dynamic max_entries not needed. Example code generation where inner map is dynamic sized array: # bpftool p d x i 125 int handle__sys_enter(void * ctx): ; int handle__sys_enter(void *ctx) 0: (b4) w1 = 0 ; int key = 0; 1: (63) *(u32 *)(r10 -4) = r1 2: (bf) r2 = r10 ; 3: (07) r2 += -4 ; inner_map = bpf_map_lookup_elem(&outer_arr_dyn, &key); 4: (18) r1 = map[id:468] 6: (07) r1 += 272 7: (61) r0 = *(u32 *)(r2 +0) 8: (35) if r0 >= 0x3 goto pc+5 9: (67) r0 <<= 3 10: (0f) r0 += r1 11: (79) r0 = *(u64 *)(r0 +0) 12: (15) if r0 == 0x0 goto pc+1 13: (05) goto pc+1 14: (b7) r0 = 0 15: (b4) w6 = -1 ; if (!inner_map) 16: (15) if r0 == 0x0 goto pc+6 17: (bf) r2 = r10 ; 18: (07) r2 += -4 ; val = bpf_map_lookup_elem(inner_map, &key); 19: (bf) r1 = r0 | No inlining but instead 20: (85) call array_map_lookup_elem#149280 | call to array_map_lookup_elem() ; return val ? *val : -1; | for inner array lookup. 21: (15) if r0 == 0x0 goto pc+1 ; return val ? *val : -1; 22: (61) r6 = *(u32 *)(r0 +0) ; } 23: (bc) w0 = w6 24: (95) exit Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Andrii Nakryiko <andrii@kernel.org> Link: https://lore.kernel.org/bpf/20201010234006.7075-4-daniel@iogearbox.net
2020-10-10 23:40:03 +00:00
static int htab_lru_map_gen_lookup(struct bpf_map *map,
struct bpf_insn *insn_buf)
{
struct bpf_insn *insn = insn_buf;
const int ret = BPF_REG_0;
const int ref_reg = BPF_REG_1;
bpf: avoid retpoline for lookup/update/delete calls on maps While some of the BPF map lookup helpers provide a ->map_gen_lookup() callback for inlining the map lookup altogether it is not available for every map, so the remaining ones have to call bpf_map_lookup_elem() helper which does a dispatch to map->ops->map_lookup_elem(). In times of retpolines, this will control and trap speculative execution rather than letting it do its work for the indirect call and will therefore cause a slowdown. Likewise, bpf_map_update_elem() and bpf_map_delete_elem() do not have an inlined version and need to call into their map->ops->map_update_elem() resp. map->ops->map_delete_elem() handlers. Before: # bpftool prog dump xlated id 1 0: (bf) r2 = r10 1: (07) r2 += -8 2: (7a) *(u64 *)(r2 +0) = 0 3: (18) r1 = map[id:1] 5: (85) call __htab_map_lookup_elem#232656 6: (15) if r0 == 0x0 goto pc+4 7: (71) r1 = *(u8 *)(r0 +35) 8: (55) if r1 != 0x0 goto pc+1 9: (72) *(u8 *)(r0 +35) = 1 10: (07) r0 += 56 11: (15) if r0 == 0x0 goto pc+4 12: (bf) r2 = r0 13: (18) r1 = map[id:1] 15: (85) call bpf_map_delete_elem#215008 <-- indirect call via 16: (95) exit helper After: # bpftool prog dump xlated id 1 0: (bf) r2 = r10 1: (07) r2 += -8 2: (7a) *(u64 *)(r2 +0) = 0 3: (18) r1 = map[id:1] 5: (85) call __htab_map_lookup_elem#233328 6: (15) if r0 == 0x0 goto pc+4 7: (71) r1 = *(u8 *)(r0 +35) 8: (55) if r1 != 0x0 goto pc+1 9: (72) *(u8 *)(r0 +35) = 1 10: (07) r0 += 56 11: (15) if r0 == 0x0 goto pc+4 12: (bf) r2 = r0 13: (18) r1 = map[id:1] 15: (85) call htab_lru_map_delete_elem#238240 <-- direct call 16: (95) exit In all three lookup/update/delete cases however we can use the actual address of the map callback directly if we find that there's only a single path with a map pointer leading to the helper call, meaning when the map pointer has not been poisoned from verifier side. Example code can be seen above for the delete case. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2018-06-02 21:06:35 +00:00
BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
(void *(*)(struct bpf_map *map, void *key))NULL));
*insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
*insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
offsetof(struct htab_elem, lru_node) +
offsetof(struct bpf_lru_node, ref));
*insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
*insn++ = BPF_ST_MEM(BPF_B, ret,
offsetof(struct htab_elem, lru_node) +
offsetof(struct bpf_lru_node, ref),
1);
*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
offsetof(struct htab_elem, key) +
round_up(map->key_size, 8));
return insn - insn_buf;
}
bpf: Wire up freeing of referenced kptr A destructor kfunc can be defined as void func(type *), where type may be void or any other pointer type as per convenience. In this patch, we ensure that the type is sane and capture the function pointer into off_desc of ptr_off_tab for the specific pointer offset, with the invariant that the dtor pointer is always set when 'kptr_ref' tag is applied to the pointer's pointee type, which is indicated by the flag BPF_MAP_VALUE_OFF_F_REF. Note that only BTF IDs whose destructor kfunc is registered, thus become the allowed BTF IDs for embedding as referenced kptr. Hence it serves the purpose of finding dtor kfunc BTF ID, as well acting as a check against the whitelist of allowed BTF IDs for this purpose. Finally, wire up the actual freeing of the referenced pointer if any at all available offsets, so that no references are leaked after the BPF map goes away and the BPF program previously moved the ownership a referenced pointer into it. The behavior is similar to BPF timers, where bpf_map_{update,delete}_elem will free any existing referenced kptr. The same case is with LRU map's bpf_lru_push_free/htab_lru_push_free functions, which are extended to reset unreferenced and free referenced kptr. Note that unlike BPF timers, kptr is not reset or freed when map uref drops to zero. Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20220424214901.2743946-8-memxor@gmail.com
2022-04-24 21:48:55 +00:00
static void check_and_free_fields(struct bpf_htab *htab,
struct htab_elem *elem)
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
{
if (htab_is_percpu(htab)) {
void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
int cpu;
bpf: Wire up freeing of referenced kptr A destructor kfunc can be defined as void func(type *), where type may be void or any other pointer type as per convenience. In this patch, we ensure that the type is sane and capture the function pointer into off_desc of ptr_off_tab for the specific pointer offset, with the invariant that the dtor pointer is always set when 'kptr_ref' tag is applied to the pointer's pointee type, which is indicated by the flag BPF_MAP_VALUE_OFF_F_REF. Note that only BTF IDs whose destructor kfunc is registered, thus become the allowed BTF IDs for embedding as referenced kptr. Hence it serves the purpose of finding dtor kfunc BTF ID, as well acting as a check against the whitelist of allowed BTF IDs for this purpose. Finally, wire up the actual freeing of the referenced pointer if any at all available offsets, so that no references are leaked after the BPF map goes away and the BPF program previously moved the ownership a referenced pointer into it. The behavior is similar to BPF timers, where bpf_map_{update,delete}_elem will free any existing referenced kptr. The same case is with LRU map's bpf_lru_push_free/htab_lru_push_free functions, which are extended to reset unreferenced and free referenced kptr. Note that unlike BPF timers, kptr is not reset or freed when map uref drops to zero. Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20220424214901.2743946-8-memxor@gmail.com
2022-04-24 21:48:55 +00:00
for_each_possible_cpu(cpu)
bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
} else {
void *map_value = elem->key + round_up(htab->map.key_size, 8);
bpf_obj_free_fields(htab->map.record, map_value);
}
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
}
/* It is called from the bpf_lru_list when the LRU needs to delete
* older elements from the htab.
*/
static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
{
struct bpf_htab *htab = arg;
struct htab_elem *l = NULL, *tgt_l;
struct hlist_nulls_head *head;
struct hlist_nulls_node *n;
unsigned long flags;
struct bucket *b;
int ret;
tgt_l = container_of(node, struct htab_elem, lru_node);
b = __select_bucket(htab, tgt_l->hash);
head = &b->head;
ret = htab_lock_bucket(htab, b, tgt_l->hash, &flags);
if (ret)
return false;
hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
if (l == tgt_l) {
hlist_nulls_del_rcu(&l->hash_node);
bpf: Wire up freeing of referenced kptr A destructor kfunc can be defined as void func(type *), where type may be void or any other pointer type as per convenience. In this patch, we ensure that the type is sane and capture the function pointer into off_desc of ptr_off_tab for the specific pointer offset, with the invariant that the dtor pointer is always set when 'kptr_ref' tag is applied to the pointer's pointee type, which is indicated by the flag BPF_MAP_VALUE_OFF_F_REF. Note that only BTF IDs whose destructor kfunc is registered, thus become the allowed BTF IDs for embedding as referenced kptr. Hence it serves the purpose of finding dtor kfunc BTF ID, as well acting as a check against the whitelist of allowed BTF IDs for this purpose. Finally, wire up the actual freeing of the referenced pointer if any at all available offsets, so that no references are leaked after the BPF map goes away and the BPF program previously moved the ownership a referenced pointer into it. The behavior is similar to BPF timers, where bpf_map_{update,delete}_elem will free any existing referenced kptr. The same case is with LRU map's bpf_lru_push_free/htab_lru_push_free functions, which are extended to reset unreferenced and free referenced kptr. Note that unlike BPF timers, kptr is not reset or freed when map uref drops to zero. Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20220424214901.2743946-8-memxor@gmail.com
2022-04-24 21:48:55 +00:00
check_and_free_fields(htab, l);
bpf_map_dec_elem_count(&htab->map);
break;
}
htab_unlock_bucket(htab, b, tgt_l->hash, flags);
return l == tgt_l;
}
/* Called from syscall */
static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
struct hlist_nulls_head *head;
struct htab_elem *l, *next_l;
u32 hash, key_size;
int i = 0;
WARN_ON_ONCE(!rcu_read_lock_held());
key_size = map->key_size;
if (!key)
goto find_first_elem;
hash = htab_map_hash(key, key_size, htab->hashrnd);
head = select_bucket(htab, hash);
/* lookup the key */
l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
if (!l)
goto find_first_elem;
/* key was found, get next key in the same bucket */
next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
struct htab_elem, hash_node);
if (next_l) {
/* if next elem in this hash list is non-zero, just return it */
memcpy(next_key, next_l->key, key_size);
return 0;
}
/* no more elements in this hash list, go to the next bucket */
i = hash & (htab->n_buckets - 1);
i++;
find_first_elem:
/* iterate over buckets */
for (; i < htab->n_buckets; i++) {
head = select_bucket(htab, i);
/* pick first element in the bucket */
next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
struct htab_elem, hash_node);
if (next_l) {
/* if it's not empty, just return it */
memcpy(next_key, next_l->key, key_size);
return 0;
}
}
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
/* iterated over all buckets and all elements */
return -ENOENT;
}
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
{
check_and_free_fields(htab, l);
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
bpf_mem_cache_free(&htab->pcpu_ma, l->ptr_to_pptr);
bpf_mem_cache_free(&htab->ma, l);
}
static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
{
struct bpf_map *map = &htab->map;
void *ptr;
if (map->ops->map_fd_put_ptr) {
ptr = fd_htab_map_get_ptr(map, l);
map->ops->map_fd_put_ptr(map, ptr, true);
}
}
static bool is_map_full(struct bpf_htab *htab)
{
if (htab->use_percpu_counter)
return __percpu_counter_compare(&htab->pcount, htab->map.max_entries,
PERCPU_COUNTER_BATCH) >= 0;
return atomic_read(&htab->count) >= htab->map.max_entries;
}
static void inc_elem_count(struct bpf_htab *htab)
{
bpf_map_inc_elem_count(&htab->map);
if (htab->use_percpu_counter)
percpu_counter_add_batch(&htab->pcount, 1, PERCPU_COUNTER_BATCH);
else
atomic_inc(&htab->count);
}
static void dec_elem_count(struct bpf_htab *htab)
{
bpf_map_dec_elem_count(&htab->map);
if (htab->use_percpu_counter)
percpu_counter_add_batch(&htab->pcount, -1, PERCPU_COUNTER_BATCH);
else
atomic_dec(&htab->count);
}
static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
{
htab_put_fd_value(htab, l);
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
if (htab_is_prealloc(htab)) {
bpf_map_dec_elem_count(&htab->map);
bpf: Wire up freeing of referenced kptr A destructor kfunc can be defined as void func(type *), where type may be void or any other pointer type as per convenience. In this patch, we ensure that the type is sane and capture the function pointer into off_desc of ptr_off_tab for the specific pointer offset, with the invariant that the dtor pointer is always set when 'kptr_ref' tag is applied to the pointer's pointee type, which is indicated by the flag BPF_MAP_VALUE_OFF_F_REF. Note that only BTF IDs whose destructor kfunc is registered, thus become the allowed BTF IDs for embedding as referenced kptr. Hence it serves the purpose of finding dtor kfunc BTF ID, as well acting as a check against the whitelist of allowed BTF IDs for this purpose. Finally, wire up the actual freeing of the referenced pointer if any at all available offsets, so that no references are leaked after the BPF map goes away and the BPF program previously moved the ownership a referenced pointer into it. The behavior is similar to BPF timers, where bpf_map_{update,delete}_elem will free any existing referenced kptr. The same case is with LRU map's bpf_lru_push_free/htab_lru_push_free functions, which are extended to reset unreferenced and free referenced kptr. Note that unlike BPF timers, kptr is not reset or freed when map uref drops to zero. Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20220424214901.2743946-8-memxor@gmail.com
2022-04-24 21:48:55 +00:00
check_and_free_fields(htab, l);
bpf: fix lockdep false positive in percpu_freelist Lockdep warns about false positive: [ 12.492084] 00000000e6b28347 (&head->lock){+...}, at: pcpu_freelist_push+0x2a/0x40 [ 12.492696] but this lock was taken by another, HARDIRQ-safe lock in the past: [ 12.493275] (&rq->lock){-.-.} [ 12.493276] [ 12.493276] [ 12.493276] and interrupts could create inverse lock ordering between them. [ 12.493276] [ 12.494435] [ 12.494435] other info that might help us debug this: [ 12.494979] Possible interrupt unsafe locking scenario: [ 12.494979] [ 12.495518] CPU0 CPU1 [ 12.495879] ---- ---- [ 12.496243] lock(&head->lock); [ 12.496502] local_irq_disable(); [ 12.496969] lock(&rq->lock); [ 12.497431] lock(&head->lock); [ 12.497890] <Interrupt> [ 12.498104] lock(&rq->lock); [ 12.498368] [ 12.498368] *** DEADLOCK *** [ 12.498368] [ 12.498837] 1 lock held by dd/276: [ 12.499110] #0: 00000000c58cb2ee (rcu_read_lock){....}, at: trace_call_bpf+0x5e/0x240 [ 12.499747] [ 12.499747] the shortest dependencies between 2nd lock and 1st lock: [ 12.500389] -> (&rq->lock){-.-.} { [ 12.500669] IN-HARDIRQ-W at: [ 12.500934] _raw_spin_lock+0x2f/0x40 [ 12.501373] scheduler_tick+0x4c/0xf0 [ 12.501812] update_process_times+0x40/0x50 [ 12.502294] tick_periodic+0x27/0xb0 [ 12.502723] tick_handle_periodic+0x1f/0x60 [ 12.503203] timer_interrupt+0x11/0x20 [ 12.503651] __handle_irq_event_percpu+0x43/0x2c0 [ 12.504167] handle_irq_event_percpu+0x20/0x50 [ 12.504674] handle_irq_event+0x37/0x60 [ 12.505139] handle_level_irq+0xa7/0x120 [ 12.505601] handle_irq+0xa1/0x150 [ 12.506018] do_IRQ+0x77/0x140 [ 12.506411] ret_from_intr+0x0/0x1d [ 12.506834] _raw_spin_unlock_irqrestore+0x53/0x60 [ 12.507362] __setup_irq+0x481/0x730 [ 12.507789] setup_irq+0x49/0x80 [ 12.508195] hpet_time_init+0x21/0x32 [ 12.508644] x86_late_time_init+0xb/0x16 [ 12.509106] start_kernel+0x390/0x42a [ 12.509554] secondary_startup_64+0xa4/0xb0 [ 12.510034] IN-SOFTIRQ-W at: [ 12.510305] _raw_spin_lock+0x2f/0x40 [ 12.510772] try_to_wake_up+0x1c7/0x4e0 [ 12.511220] swake_up_locked+0x20/0x40 [ 12.511657] swake_up_one+0x1a/0x30 [ 12.512070] rcu_process_callbacks+0xc5/0x650 [ 12.512553] __do_softirq+0xe6/0x47b [ 12.512978] irq_exit+0xc3/0xd0 [ 12.513372] smp_apic_timer_interrupt+0xa9/0x250 [ 12.513876] apic_timer_interrupt+0xf/0x20 [ 12.514343] default_idle+0x1c/0x170 [ 12.514765] do_idle+0x199/0x240 [ 12.515159] cpu_startup_entry+0x19/0x20 [ 12.515614] start_kernel+0x422/0x42a [ 12.516045] secondary_startup_64+0xa4/0xb0 [ 12.516521] INITIAL USE at: [ 12.516774] _raw_spin_lock_irqsave+0x38/0x50 [ 12.517258] rq_attach_root+0x16/0xd0 [ 12.517685] sched_init+0x2f2/0x3eb [ 12.518096] start_kernel+0x1fb/0x42a [ 12.518525] secondary_startup_64+0xa4/0xb0 [ 12.518986] } [ 12.519132] ... key at: [<ffffffff82b7bc28>] __key.71384+0x0/0x8 [ 12.519649] ... acquired at: [ 12.519892] pcpu_freelist_pop+0x7b/0xd0 [ 12.520221] bpf_get_stackid+0x1d2/0x4d0 [ 12.520563] ___bpf_prog_run+0x8b4/0x11a0 [ 12.520887] [ 12.521008] -> (&head->lock){+...} { [ 12.521292] HARDIRQ-ON-W at: [ 12.521539] _raw_spin_lock+0x2f/0x40 [ 12.521950] pcpu_freelist_push+0x2a/0x40 [ 12.522396] bpf_get_stackid+0x494/0x4d0 [ 12.522828] ___bpf_prog_run+0x8b4/0x11a0 [ 12.523296] INITIAL USE at: [ 12.523537] _raw_spin_lock+0x2f/0x40 [ 12.523944] pcpu_freelist_populate+0xc0/0x120 [ 12.524417] htab_map_alloc+0x405/0x500 [ 12.524835] __do_sys_bpf+0x1a3/0x1a90 [ 12.525253] do_syscall_64+0x4a/0x180 [ 12.525659] entry_SYSCALL_64_after_hwframe+0x49/0xbe [ 12.526167] } [ 12.526311] ... key at: [<ffffffff838f7668>] __key.13130+0x0/0x8 [ 12.526812] ... acquired at: [ 12.527047] __lock_acquire+0x521/0x1350 [ 12.527371] lock_acquire+0x98/0x190 [ 12.527680] _raw_spin_lock+0x2f/0x40 [ 12.527994] pcpu_freelist_push+0x2a/0x40 [ 12.528325] bpf_get_stackid+0x494/0x4d0 [ 12.528645] ___bpf_prog_run+0x8b4/0x11a0 [ 12.528970] [ 12.529092] [ 12.529092] stack backtrace: [ 12.529444] CPU: 0 PID: 276 Comm: dd Not tainted 5.0.0-rc3-00018-g2fa53f892422 #475 [ 12.530043] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.11.0-2.el7 04/01/2014 [ 12.530750] Call Trace: [ 12.530948] dump_stack+0x5f/0x8b [ 12.531248] check_usage_backwards+0x10c/0x120 [ 12.531598] ? ___bpf_prog_run+0x8b4/0x11a0 [ 12.531935] ? mark_lock+0x382/0x560 [ 12.532229] mark_lock+0x382/0x560 [ 12.532496] ? print_shortest_lock_dependencies+0x180/0x180 [ 12.532928] __lock_acquire+0x521/0x1350 [ 12.533271] ? find_get_entry+0x17f/0x2e0 [ 12.533586] ? find_get_entry+0x19c/0x2e0 [ 12.533902] ? lock_acquire+0x98/0x190 [ 12.534196] lock_acquire+0x98/0x190 [ 12.534482] ? pcpu_freelist_push+0x2a/0x40 [ 12.534810] _raw_spin_lock+0x2f/0x40 [ 12.535099] ? pcpu_freelist_push+0x2a/0x40 [ 12.535432] pcpu_freelist_push+0x2a/0x40 [ 12.535750] bpf_get_stackid+0x494/0x4d0 [ 12.536062] ___bpf_prog_run+0x8b4/0x11a0 It has been explained that is a false positive here: https://lkml.org/lkml/2018/7/25/756 Recap: - stackmap uses pcpu_freelist - The lock in pcpu_freelist is a percpu lock - stackmap is only used by tracing bpf_prog - A tracing bpf_prog cannot be run if another bpf_prog has already been running (ensured by the percpu bpf_prog_active counter). Eric pointed out that this lockdep splats stops other legit lockdep splats in selftests/bpf/test_progs.c. Fix this by calling local_irq_save/restore for stackmap. Another false positive had also been worked around by calling local_irq_save in commit 89ad2fa3f043 ("bpf: fix lockdep splat"). That commit added unnecessary irq_save/restore to fast path of bpf hash map. irqs are already disabled at that point, since htab is holding per bucket spin_lock with irqsave. Let's reduce overhead for htab by introducing __pcpu_freelist_push/pop function w/o irqsave and convert pcpu_freelist_push/pop to irqsave to be used elsewhere (right now only in stackmap). It stops lockdep false positive in stackmap with a bit of acceptable overhead. Fixes: 557c0c6e7df8 ("bpf: convert stackmap to pre-allocation") Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org> Reported-by: Eric Dumazet <eric.dumazet@gmail.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2019-01-31 02:12:43 +00:00
__pcpu_freelist_push(&htab->freelist, &l->fnode);
} else {
dec_elem_count(htab);
htab_elem_free(htab, l);
}
}
static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
void *value, bool onallcpus)
{
if (!onallcpus) {
/* copy true value_size bytes */
copy_map_value(&htab->map, this_cpu_ptr(pptr), value);
} else {
u32 size = round_up(htab->map.value_size, 8);
int off = 0, cpu;
for_each_possible_cpu(cpu) {
copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value + off);
off += size;
}
}
}
static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
void *value, bool onallcpus)
{
/* When not setting the initial value on all cpus, zero-fill element
* values for other cpus. Otherwise, bpf program has no way to ensure
* known initial values for cpus other than current one
* (onallcpus=false always when coming from bpf prog).
*/
if (!onallcpus) {
int current_cpu = raw_smp_processor_id();
int cpu;
for_each_possible_cpu(cpu) {
if (cpu == current_cpu)
copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value);
else /* Since elem is preallocated, we cannot touch special fields */
zero_map_value(&htab->map, per_cpu_ptr(pptr, cpu));
}
} else {
pcpu_copy_value(htab, pptr, value, onallcpus);
}
}
bpf: fix map value attribute for hash of maps Currently, iproute2's BPF ELF loader works fine with array of maps when retrieving the fd from a pinned node and doing a selfcheck against the provided map attributes from the object file, but we fail to do the same for hash of maps and thus refuse to get the map from pinned node. Reason is that when allocating hash of maps, fd_htab_map_alloc() will set the value size to sizeof(void *), and any user space map creation requests are forced to set 4 bytes as value size. Thus, selfcheck will complain about exposed 8 bytes on 64 bit archs vs. 4 bytes from object file as value size. Contract is that fdinfo or BPF_MAP_GET_FD_BY_ID returns the value size used to create the map. Fix it by handling it the same way as we do for array of maps, which means that we leave value size at 4 bytes and in the allocation phase round up value size to 8 bytes. alloc_htab_elem() needs an adjustment in order to copy rounded up 8 bytes due to bpf_fd_htab_map_update_elem() calling into htab_map_update_elem() with the pointer of the map pointer as value. Unlike array of maps where we just xchg(), we're using the generic htab_map_update_elem() callback also used from helper calls, which published the key/value already on return, so we need to ensure to memcpy() the right size. Fixes: bcc6b1b7ebf8 ("bpf: Add hash of maps support") Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-22 22:06:09 +00:00
static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
{
return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
BITS_PER_LONG == 64;
}
static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
void *value, u32 key_size, u32 hash,
bool percpu, bool onallcpus,
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
struct htab_elem *old_elem)
{
bpf: introduce bpf_spin_lock Introduce 'struct bpf_spin_lock' and bpf_spin_lock/unlock() helpers to let bpf program serialize access to other variables. Example: struct hash_elem { int cnt; struct bpf_spin_lock lock; }; struct hash_elem * val = bpf_map_lookup_elem(&hash_map, &key); if (val) { bpf_spin_lock(&val->lock); val->cnt++; bpf_spin_unlock(&val->lock); } Restrictions and safety checks: - bpf_spin_lock is only allowed inside HASH and ARRAY maps. - BTF description of the map is mandatory for safety analysis. - bpf program can take one bpf_spin_lock at a time, since two or more can cause dead locks. - only one 'struct bpf_spin_lock' is allowed per map element. It drastically simplifies implementation yet allows bpf program to use any number of bpf_spin_locks. - when bpf_spin_lock is taken the calls (either bpf2bpf or helpers) are not allowed. - bpf program must bpf_spin_unlock() before return. - bpf program can access 'struct bpf_spin_lock' only via bpf_spin_lock()/bpf_spin_unlock() helpers. - load/store into 'struct bpf_spin_lock lock;' field is not allowed. - to use bpf_spin_lock() helper the BTF description of map value must be a struct and have 'struct bpf_spin_lock anyname;' field at the top level. Nested lock inside another struct is not allowed. - syscall map_lookup doesn't copy bpf_spin_lock field to user space. - syscall map_update and program map_update do not update bpf_spin_lock field. - bpf_spin_lock cannot be on the stack or inside networking packet. bpf_spin_lock can only be inside HASH or ARRAY map value. - bpf_spin_lock is available to root only and to all program types. - bpf_spin_lock is not allowed in inner maps of map-in-map. - ld_abs is not allowed inside spin_lock-ed region. - tracing progs and socket filter progs cannot use bpf_spin_lock due to insufficient preemption checks Implementation details: - cgroup-bpf class of programs can nest with xdp/tc programs. Hence bpf_spin_lock is equivalent to spin_lock_irqsave. Other solutions to avoid nested bpf_spin_lock are possible. Like making sure that all networking progs run with softirq disabled. spin_lock_irqsave is the simplest and doesn't add overhead to the programs that don't use it. - arch_spinlock_t is used when its implemented as queued_spin_lock - archs can force their own arch_spinlock_t - on architectures where queued_spin_lock is not available and sizeof(arch_spinlock_t) != sizeof(__u32) trivial lock is used. - presence of bpf_spin_lock inside map value could have been indicated via extra flag during map_create, but specifying it via BTF is cleaner. It provides introspection for map key/value and reduces user mistakes. Next steps: - allow bpf_spin_lock in other map types (like cgroup local storage) - introduce BPF_F_LOCK flag for bpf_map_update() syscall and helper to request kernel to grab bpf_spin_lock before rewriting the value. That will serialize access to map elements. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2019-01-31 23:40:04 +00:00
u32 size = htab->map.value_size;
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
bool prealloc = htab_is_prealloc(htab);
struct htab_elem *l_new, **pl_new;
void __percpu *pptr;
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
if (prealloc) {
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
if (old_elem) {
/* if we're updating the existing element,
* use per-cpu extra elems to avoid freelist_pop/push
*/
pl_new = this_cpu_ptr(htab->extra_elems);
l_new = *pl_new;
htab_put_fd_value(htab, old_elem);
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
*pl_new = old_elem;
} else {
struct pcpu_freelist_node *l;
bpf: fix lockdep false positive in percpu_freelist Lockdep warns about false positive: [ 12.492084] 00000000e6b28347 (&head->lock){+...}, at: pcpu_freelist_push+0x2a/0x40 [ 12.492696] but this lock was taken by another, HARDIRQ-safe lock in the past: [ 12.493275] (&rq->lock){-.-.} [ 12.493276] [ 12.493276] [ 12.493276] and interrupts could create inverse lock ordering between them. [ 12.493276] [ 12.494435] [ 12.494435] other info that might help us debug this: [ 12.494979] Possible interrupt unsafe locking scenario: [ 12.494979] [ 12.495518] CPU0 CPU1 [ 12.495879] ---- ---- [ 12.496243] lock(&head->lock); [ 12.496502] local_irq_disable(); [ 12.496969] lock(&rq->lock); [ 12.497431] lock(&head->lock); [ 12.497890] <Interrupt> [ 12.498104] lock(&rq->lock); [ 12.498368] [ 12.498368] *** DEADLOCK *** [ 12.498368] [ 12.498837] 1 lock held by dd/276: [ 12.499110] #0: 00000000c58cb2ee (rcu_read_lock){....}, at: trace_call_bpf+0x5e/0x240 [ 12.499747] [ 12.499747] the shortest dependencies between 2nd lock and 1st lock: [ 12.500389] -> (&rq->lock){-.-.} { [ 12.500669] IN-HARDIRQ-W at: [ 12.500934] _raw_spin_lock+0x2f/0x40 [ 12.501373] scheduler_tick+0x4c/0xf0 [ 12.501812] update_process_times+0x40/0x50 [ 12.502294] tick_periodic+0x27/0xb0 [ 12.502723] tick_handle_periodic+0x1f/0x60 [ 12.503203] timer_interrupt+0x11/0x20 [ 12.503651] __handle_irq_event_percpu+0x43/0x2c0 [ 12.504167] handle_irq_event_percpu+0x20/0x50 [ 12.504674] handle_irq_event+0x37/0x60 [ 12.505139] handle_level_irq+0xa7/0x120 [ 12.505601] handle_irq+0xa1/0x150 [ 12.506018] do_IRQ+0x77/0x140 [ 12.506411] ret_from_intr+0x0/0x1d [ 12.506834] _raw_spin_unlock_irqrestore+0x53/0x60 [ 12.507362] __setup_irq+0x481/0x730 [ 12.507789] setup_irq+0x49/0x80 [ 12.508195] hpet_time_init+0x21/0x32 [ 12.508644] x86_late_time_init+0xb/0x16 [ 12.509106] start_kernel+0x390/0x42a [ 12.509554] secondary_startup_64+0xa4/0xb0 [ 12.510034] IN-SOFTIRQ-W at: [ 12.510305] _raw_spin_lock+0x2f/0x40 [ 12.510772] try_to_wake_up+0x1c7/0x4e0 [ 12.511220] swake_up_locked+0x20/0x40 [ 12.511657] swake_up_one+0x1a/0x30 [ 12.512070] rcu_process_callbacks+0xc5/0x650 [ 12.512553] __do_softirq+0xe6/0x47b [ 12.512978] irq_exit+0xc3/0xd0 [ 12.513372] smp_apic_timer_interrupt+0xa9/0x250 [ 12.513876] apic_timer_interrupt+0xf/0x20 [ 12.514343] default_idle+0x1c/0x170 [ 12.514765] do_idle+0x199/0x240 [ 12.515159] cpu_startup_entry+0x19/0x20 [ 12.515614] start_kernel+0x422/0x42a [ 12.516045] secondary_startup_64+0xa4/0xb0 [ 12.516521] INITIAL USE at: [ 12.516774] _raw_spin_lock_irqsave+0x38/0x50 [ 12.517258] rq_attach_root+0x16/0xd0 [ 12.517685] sched_init+0x2f2/0x3eb [ 12.518096] start_kernel+0x1fb/0x42a [ 12.518525] secondary_startup_64+0xa4/0xb0 [ 12.518986] } [ 12.519132] ... key at: [<ffffffff82b7bc28>] __key.71384+0x0/0x8 [ 12.519649] ... acquired at: [ 12.519892] pcpu_freelist_pop+0x7b/0xd0 [ 12.520221] bpf_get_stackid+0x1d2/0x4d0 [ 12.520563] ___bpf_prog_run+0x8b4/0x11a0 [ 12.520887] [ 12.521008] -> (&head->lock){+...} { [ 12.521292] HARDIRQ-ON-W at: [ 12.521539] _raw_spin_lock+0x2f/0x40 [ 12.521950] pcpu_freelist_push+0x2a/0x40 [ 12.522396] bpf_get_stackid+0x494/0x4d0 [ 12.522828] ___bpf_prog_run+0x8b4/0x11a0 [ 12.523296] INITIAL USE at: [ 12.523537] _raw_spin_lock+0x2f/0x40 [ 12.523944] pcpu_freelist_populate+0xc0/0x120 [ 12.524417] htab_map_alloc+0x405/0x500 [ 12.524835] __do_sys_bpf+0x1a3/0x1a90 [ 12.525253] do_syscall_64+0x4a/0x180 [ 12.525659] entry_SYSCALL_64_after_hwframe+0x49/0xbe [ 12.526167] } [ 12.526311] ... key at: [<ffffffff838f7668>] __key.13130+0x0/0x8 [ 12.526812] ... acquired at: [ 12.527047] __lock_acquire+0x521/0x1350 [ 12.527371] lock_acquire+0x98/0x190 [ 12.527680] _raw_spin_lock+0x2f/0x40 [ 12.527994] pcpu_freelist_push+0x2a/0x40 [ 12.528325] bpf_get_stackid+0x494/0x4d0 [ 12.528645] ___bpf_prog_run+0x8b4/0x11a0 [ 12.528970] [ 12.529092] [ 12.529092] stack backtrace: [ 12.529444] CPU: 0 PID: 276 Comm: dd Not tainted 5.0.0-rc3-00018-g2fa53f892422 #475 [ 12.530043] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.11.0-2.el7 04/01/2014 [ 12.530750] Call Trace: [ 12.530948] dump_stack+0x5f/0x8b [ 12.531248] check_usage_backwards+0x10c/0x120 [ 12.531598] ? ___bpf_prog_run+0x8b4/0x11a0 [ 12.531935] ? mark_lock+0x382/0x560 [ 12.532229] mark_lock+0x382/0x560 [ 12.532496] ? print_shortest_lock_dependencies+0x180/0x180 [ 12.532928] __lock_acquire+0x521/0x1350 [ 12.533271] ? find_get_entry+0x17f/0x2e0 [ 12.533586] ? find_get_entry+0x19c/0x2e0 [ 12.533902] ? lock_acquire+0x98/0x190 [ 12.534196] lock_acquire+0x98/0x190 [ 12.534482] ? pcpu_freelist_push+0x2a/0x40 [ 12.534810] _raw_spin_lock+0x2f/0x40 [ 12.535099] ? pcpu_freelist_push+0x2a/0x40 [ 12.535432] pcpu_freelist_push+0x2a/0x40 [ 12.535750] bpf_get_stackid+0x494/0x4d0 [ 12.536062] ___bpf_prog_run+0x8b4/0x11a0 It has been explained that is a false positive here: https://lkml.org/lkml/2018/7/25/756 Recap: - stackmap uses pcpu_freelist - The lock in pcpu_freelist is a percpu lock - stackmap is only used by tracing bpf_prog - A tracing bpf_prog cannot be run if another bpf_prog has already been running (ensured by the percpu bpf_prog_active counter). Eric pointed out that this lockdep splats stops other legit lockdep splats in selftests/bpf/test_progs.c. Fix this by calling local_irq_save/restore for stackmap. Another false positive had also been worked around by calling local_irq_save in commit 89ad2fa3f043 ("bpf: fix lockdep splat"). That commit added unnecessary irq_save/restore to fast path of bpf hash map. irqs are already disabled at that point, since htab is holding per bucket spin_lock with irqsave. Let's reduce overhead for htab by introducing __pcpu_freelist_push/pop function w/o irqsave and convert pcpu_freelist_push/pop to irqsave to be used elsewhere (right now only in stackmap). It stops lockdep false positive in stackmap with a bit of acceptable overhead. Fixes: 557c0c6e7df8 ("bpf: convert stackmap to pre-allocation") Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org> Reported-by: Eric Dumazet <eric.dumazet@gmail.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2019-01-31 02:12:43 +00:00
l = __pcpu_freelist_pop(&htab->freelist);
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
if (!l)
return ERR_PTR(-E2BIG);
l_new = container_of(l, struct htab_elem, fnode);
bpf_map_inc_elem_count(&htab->map);
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
}
} else {
if (is_map_full(htab))
if (!old_elem)
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
/* when map is full and update() is replacing
* old element, it's ok to allocate, since
* old element will be freed immediately.
* Otherwise return an error
*/
return ERR_PTR(-E2BIG);
inc_elem_count(htab);
l_new = bpf_mem_cache_alloc(&htab->ma);
if (!l_new) {
l_new = ERR_PTR(-ENOMEM);
goto dec_count;
}
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
}
memcpy(l_new->key, key, key_size);
if (percpu) {
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
if (prealloc) {
pptr = htab_elem_get_ptr(l_new, key_size);
} else {
/* alloc_percpu zero-fills */
pptr = bpf_mem_cache_alloc(&htab->pcpu_ma);
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
if (!pptr) {
bpf_mem_cache_free(&htab->ma, l_new);
l_new = ERR_PTR(-ENOMEM);
goto dec_count;
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
}
l_new->ptr_to_pptr = pptr;
pptr = *(void **)pptr;
}
pcpu_init_value(htab, pptr, value, onallcpus);
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
if (!prealloc)
htab_elem_set_ptr(l_new, key_size, pptr);
bpf: introduce bpf_spin_lock Introduce 'struct bpf_spin_lock' and bpf_spin_lock/unlock() helpers to let bpf program serialize access to other variables. Example: struct hash_elem { int cnt; struct bpf_spin_lock lock; }; struct hash_elem * val = bpf_map_lookup_elem(&hash_map, &key); if (val) { bpf_spin_lock(&val->lock); val->cnt++; bpf_spin_unlock(&val->lock); } Restrictions and safety checks: - bpf_spin_lock is only allowed inside HASH and ARRAY maps. - BTF description of the map is mandatory for safety analysis. - bpf program can take one bpf_spin_lock at a time, since two or more can cause dead locks. - only one 'struct bpf_spin_lock' is allowed per map element. It drastically simplifies implementation yet allows bpf program to use any number of bpf_spin_locks. - when bpf_spin_lock is taken the calls (either bpf2bpf or helpers) are not allowed. - bpf program must bpf_spin_unlock() before return. - bpf program can access 'struct bpf_spin_lock' only via bpf_spin_lock()/bpf_spin_unlock() helpers. - load/store into 'struct bpf_spin_lock lock;' field is not allowed. - to use bpf_spin_lock() helper the BTF description of map value must be a struct and have 'struct bpf_spin_lock anyname;' field at the top level. Nested lock inside another struct is not allowed. - syscall map_lookup doesn't copy bpf_spin_lock field to user space. - syscall map_update and program map_update do not update bpf_spin_lock field. - bpf_spin_lock cannot be on the stack or inside networking packet. bpf_spin_lock can only be inside HASH or ARRAY map value. - bpf_spin_lock is available to root only and to all program types. - bpf_spin_lock is not allowed in inner maps of map-in-map. - ld_abs is not allowed inside spin_lock-ed region. - tracing progs and socket filter progs cannot use bpf_spin_lock due to insufficient preemption checks Implementation details: - cgroup-bpf class of programs can nest with xdp/tc programs. Hence bpf_spin_lock is equivalent to spin_lock_irqsave. Other solutions to avoid nested bpf_spin_lock are possible. Like making sure that all networking progs run with softirq disabled. spin_lock_irqsave is the simplest and doesn't add overhead to the programs that don't use it. - arch_spinlock_t is used when its implemented as queued_spin_lock - archs can force their own arch_spinlock_t - on architectures where queued_spin_lock is not available and sizeof(arch_spinlock_t) != sizeof(__u32) trivial lock is used. - presence of bpf_spin_lock inside map value could have been indicated via extra flag during map_create, but specifying it via BTF is cleaner. It provides introspection for map key/value and reduces user mistakes. Next steps: - allow bpf_spin_lock in other map types (like cgroup local storage) - introduce BPF_F_LOCK flag for bpf_map_update() syscall and helper to request kernel to grab bpf_spin_lock before rewriting the value. That will serialize access to map elements. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2019-01-31 23:40:04 +00:00
} else if (fd_htab_map_needs_adjust(htab)) {
size = round_up(size, 8);
memcpy(l_new->key + round_up(key_size, 8), value, size);
bpf: introduce bpf_spin_lock Introduce 'struct bpf_spin_lock' and bpf_spin_lock/unlock() helpers to let bpf program serialize access to other variables. Example: struct hash_elem { int cnt; struct bpf_spin_lock lock; }; struct hash_elem * val = bpf_map_lookup_elem(&hash_map, &key); if (val) { bpf_spin_lock(&val->lock); val->cnt++; bpf_spin_unlock(&val->lock); } Restrictions and safety checks: - bpf_spin_lock is only allowed inside HASH and ARRAY maps. - BTF description of the map is mandatory for safety analysis. - bpf program can take one bpf_spin_lock at a time, since two or more can cause dead locks. - only one 'struct bpf_spin_lock' is allowed per map element. It drastically simplifies implementation yet allows bpf program to use any number of bpf_spin_locks. - when bpf_spin_lock is taken the calls (either bpf2bpf or helpers) are not allowed. - bpf program must bpf_spin_unlock() before return. - bpf program can access 'struct bpf_spin_lock' only via bpf_spin_lock()/bpf_spin_unlock() helpers. - load/store into 'struct bpf_spin_lock lock;' field is not allowed. - to use bpf_spin_lock() helper the BTF description of map value must be a struct and have 'struct bpf_spin_lock anyname;' field at the top level. Nested lock inside another struct is not allowed. - syscall map_lookup doesn't copy bpf_spin_lock field to user space. - syscall map_update and program map_update do not update bpf_spin_lock field. - bpf_spin_lock cannot be on the stack or inside networking packet. bpf_spin_lock can only be inside HASH or ARRAY map value. - bpf_spin_lock is available to root only and to all program types. - bpf_spin_lock is not allowed in inner maps of map-in-map. - ld_abs is not allowed inside spin_lock-ed region. - tracing progs and socket filter progs cannot use bpf_spin_lock due to insufficient preemption checks Implementation details: - cgroup-bpf class of programs can nest with xdp/tc programs. Hence bpf_spin_lock is equivalent to spin_lock_irqsave. Other solutions to avoid nested bpf_spin_lock are possible. Like making sure that all networking progs run with softirq disabled. spin_lock_irqsave is the simplest and doesn't add overhead to the programs that don't use it. - arch_spinlock_t is used when its implemented as queued_spin_lock - archs can force their own arch_spinlock_t - on architectures where queued_spin_lock is not available and sizeof(arch_spinlock_t) != sizeof(__u32) trivial lock is used. - presence of bpf_spin_lock inside map value could have been indicated via extra flag during map_create, but specifying it via BTF is cleaner. It provides introspection for map key/value and reduces user mistakes. Next steps: - allow bpf_spin_lock in other map types (like cgroup local storage) - introduce BPF_F_LOCK flag for bpf_map_update() syscall and helper to request kernel to grab bpf_spin_lock before rewriting the value. That will serialize access to map elements. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2019-01-31 23:40:04 +00:00
} else {
copy_map_value(&htab->map,
l_new->key + round_up(key_size, 8),
value);
}
l_new->hash = hash;
return l_new;
dec_count:
dec_elem_count(htab);
return l_new;
}
static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
u64 map_flags)
{
if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
/* elem already exists */
return -EEXIST;
if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
/* elem doesn't exist, cannot update it */
return -ENOENT;
return 0;
}
/* Called from syscall or from eBPF program */
bpf: return long from bpf_map_ops funcs This patch changes the return types of bpf_map_ops functions to long, where previously int was returned. Using long allows for bpf programs to maintain the sign bit in the absence of sign extension during situations where inlined bpf helper funcs make calls to the bpf_map_ops funcs and a negative error is returned. The definitions of the helper funcs are generated from comments in the bpf uapi header at `include/uapi/linux/bpf.h`. The return type of these helpers was previously changed from int to long in commit bdb7b79b4ce8. For any case where one of the map helpers call the bpf_map_ops funcs that are still returning 32-bit int, a compiler might not include sign extension instructions to properly convert the 32-bit negative value a 64-bit negative value. For example: bpf assembly excerpt of an inlined helper calling a kernel function and checking for a specific error: ; err = bpf_map_update_elem(&mymap, &key, &val, BPF_NOEXIST); ... 46: call 0xffffffffe103291c ; htab_map_update_elem ; if (err && err != -EEXIST) { 4b: cmp $0xffffffffffffffef,%rax ; cmp -EEXIST,%rax kernel function assembly excerpt of return value from `htab_map_update_elem` returning 32-bit int: movl $0xffffffef, %r9d ... movl %r9d, %eax ...results in the comparison: cmp $0xffffffffffffffef, $0x00000000ffffffef Fixes: bdb7b79b4ce8 ("bpf: Switch most helper return values from 32-bit int to 64-bit long") Tested-by: Eduard Zingerman <eddyz87@gmail.com> Signed-off-by: JP Kobryn <inwardvessel@gmail.com> Link: https://lore.kernel.org/r/20230322194754.185781-3-inwardvessel@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-22 19:47:54 +00:00
static long htab_map_update_elem(struct bpf_map *map, void *key, void *value,
u64 map_flags)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
struct htab_elem *l_new = NULL, *l_old;
struct hlist_nulls_head *head;
unsigned long flags;
struct bucket *b;
u32 key_size, hash;
int ret;
if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
/* unknown flags */
return -EINVAL;
WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
!rcu_read_lock_bh_held());
key_size = map->key_size;
hash = htab_map_hash(key, key_size, htab->hashrnd);
b = __select_bucket(htab, hash);
head = &b->head;
if (unlikely(map_flags & BPF_F_LOCK)) {
if (unlikely(!btf_record_has_field(map->record, BPF_SPIN_LOCK)))
return -EINVAL;
/* find an element without taking the bucket lock */
l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
htab->n_buckets);
ret = check_flags(htab, l_old, map_flags);
if (ret)
return ret;
if (l_old) {
/* grab the element lock and update value in place */
copy_map_value_locked(map,
l_old->key + round_up(key_size, 8),
value, false);
return 0;
}
/* fall through, grab the bucket lock and lookup again.
* 99.9% chance that the element won't be found,
* but second lookup under lock has to be done.
*/
}
ret = htab_lock_bucket(htab, b, hash, &flags);
if (ret)
return ret;
l_old = lookup_elem_raw(head, hash, key, key_size);
ret = check_flags(htab, l_old, map_flags);
if (ret)
goto err;
if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
/* first lookup without the bucket lock didn't find the element,
* but second lookup with the bucket lock found it.
* This case is highly unlikely, but has to be dealt with:
* grab the element lock in addition to the bucket lock
* and update element in place
*/
copy_map_value_locked(map,
l_old->key + round_up(key_size, 8),
value, false);
ret = 0;
goto err;
}
l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
l_old);
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
if (IS_ERR(l_new)) {
/* all pre-allocated elements are in use or memory exhausted */
ret = PTR_ERR(l_new);
goto err;
}
/* add new element to the head of the list, so that
* concurrent search will find it before old elem
*/
hlist_nulls_add_head_rcu(&l_new->hash_node, head);
if (l_old) {
hlist_nulls_del_rcu(&l_old->hash_node);
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
if (!htab_is_prealloc(htab))
free_htab_elem(htab, l_old);
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
else
bpf: Wire up freeing of referenced kptr A destructor kfunc can be defined as void func(type *), where type may be void or any other pointer type as per convenience. In this patch, we ensure that the type is sane and capture the function pointer into off_desc of ptr_off_tab for the specific pointer offset, with the invariant that the dtor pointer is always set when 'kptr_ref' tag is applied to the pointer's pointee type, which is indicated by the flag BPF_MAP_VALUE_OFF_F_REF. Note that only BTF IDs whose destructor kfunc is registered, thus become the allowed BTF IDs for embedding as referenced kptr. Hence it serves the purpose of finding dtor kfunc BTF ID, as well acting as a check against the whitelist of allowed BTF IDs for this purpose. Finally, wire up the actual freeing of the referenced pointer if any at all available offsets, so that no references are leaked after the BPF map goes away and the BPF program previously moved the ownership a referenced pointer into it. The behavior is similar to BPF timers, where bpf_map_{update,delete}_elem will free any existing referenced kptr. The same case is with LRU map's bpf_lru_push_free/htab_lru_push_free functions, which are extended to reset unreferenced and free referenced kptr. Note that unlike BPF timers, kptr is not reset or freed when map uref drops to zero. Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20220424214901.2743946-8-memxor@gmail.com
2022-04-24 21:48:55 +00:00
check_and_free_fields(htab, l_old);
}
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
ret = 0;
err:
htab_unlock_bucket(htab, b, hash, flags);
return ret;
}
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
static void htab_lru_push_free(struct bpf_htab *htab, struct htab_elem *elem)
{
bpf: Wire up freeing of referenced kptr A destructor kfunc can be defined as void func(type *), where type may be void or any other pointer type as per convenience. In this patch, we ensure that the type is sane and capture the function pointer into off_desc of ptr_off_tab for the specific pointer offset, with the invariant that the dtor pointer is always set when 'kptr_ref' tag is applied to the pointer's pointee type, which is indicated by the flag BPF_MAP_VALUE_OFF_F_REF. Note that only BTF IDs whose destructor kfunc is registered, thus become the allowed BTF IDs for embedding as referenced kptr. Hence it serves the purpose of finding dtor kfunc BTF ID, as well acting as a check against the whitelist of allowed BTF IDs for this purpose. Finally, wire up the actual freeing of the referenced pointer if any at all available offsets, so that no references are leaked after the BPF map goes away and the BPF program previously moved the ownership a referenced pointer into it. The behavior is similar to BPF timers, where bpf_map_{update,delete}_elem will free any existing referenced kptr. The same case is with LRU map's bpf_lru_push_free/htab_lru_push_free functions, which are extended to reset unreferenced and free referenced kptr. Note that unlike BPF timers, kptr is not reset or freed when map uref drops to zero. Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20220424214901.2743946-8-memxor@gmail.com
2022-04-24 21:48:55 +00:00
check_and_free_fields(htab, elem);
bpf_map_dec_elem_count(&htab->map);
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
bpf_lru_push_free(&htab->lru, &elem->lru_node);
}
bpf: return long from bpf_map_ops funcs This patch changes the return types of bpf_map_ops functions to long, where previously int was returned. Using long allows for bpf programs to maintain the sign bit in the absence of sign extension during situations where inlined bpf helper funcs make calls to the bpf_map_ops funcs and a negative error is returned. The definitions of the helper funcs are generated from comments in the bpf uapi header at `include/uapi/linux/bpf.h`. The return type of these helpers was previously changed from int to long in commit bdb7b79b4ce8. For any case where one of the map helpers call the bpf_map_ops funcs that are still returning 32-bit int, a compiler might not include sign extension instructions to properly convert the 32-bit negative value a 64-bit negative value. For example: bpf assembly excerpt of an inlined helper calling a kernel function and checking for a specific error: ; err = bpf_map_update_elem(&mymap, &key, &val, BPF_NOEXIST); ... 46: call 0xffffffffe103291c ; htab_map_update_elem ; if (err && err != -EEXIST) { 4b: cmp $0xffffffffffffffef,%rax ; cmp -EEXIST,%rax kernel function assembly excerpt of return value from `htab_map_update_elem` returning 32-bit int: movl $0xffffffef, %r9d ... movl %r9d, %eax ...results in the comparison: cmp $0xffffffffffffffef, $0x00000000ffffffef Fixes: bdb7b79b4ce8 ("bpf: Switch most helper return values from 32-bit int to 64-bit long") Tested-by: Eduard Zingerman <eddyz87@gmail.com> Signed-off-by: JP Kobryn <inwardvessel@gmail.com> Link: https://lore.kernel.org/r/20230322194754.185781-3-inwardvessel@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-22 19:47:54 +00:00
static long htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
u64 map_flags)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
struct htab_elem *l_new, *l_old = NULL;
struct hlist_nulls_head *head;
unsigned long flags;
struct bucket *b;
u32 key_size, hash;
int ret;
if (unlikely(map_flags > BPF_EXIST))
/* unknown flags */
return -EINVAL;
WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
!rcu_read_lock_bh_held());
key_size = map->key_size;
hash = htab_map_hash(key, key_size, htab->hashrnd);
b = __select_bucket(htab, hash);
head = &b->head;
/* For LRU, we need to alloc before taking bucket's
* spinlock because getting free nodes from LRU may need
* to remove older elements from htab and this removal
* operation will need a bucket lock.
*/
l_new = prealloc_lru_pop(htab, key, hash);
if (!l_new)
return -ENOMEM;
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
copy_map_value(&htab->map,
l_new->key + round_up(map->key_size, 8), value);
ret = htab_lock_bucket(htab, b, hash, &flags);
if (ret)
goto err_lock_bucket;
l_old = lookup_elem_raw(head, hash, key, key_size);
ret = check_flags(htab, l_old, map_flags);
if (ret)
goto err;
/* add new element to the head of the list, so that
* concurrent search will find it before old elem
*/
hlist_nulls_add_head_rcu(&l_new->hash_node, head);
if (l_old) {
bpf_lru_node_set_ref(&l_new->lru_node);
hlist_nulls_del_rcu(&l_old->hash_node);
}
ret = 0;
err:
htab_unlock_bucket(htab, b, hash, flags);
err_lock_bucket:
if (ret)
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
htab_lru_push_free(htab, l_new);
else if (l_old)
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
htab_lru_push_free(htab, l_old);
return ret;
}
bpf: return long from bpf_map_ops funcs This patch changes the return types of bpf_map_ops functions to long, where previously int was returned. Using long allows for bpf programs to maintain the sign bit in the absence of sign extension during situations where inlined bpf helper funcs make calls to the bpf_map_ops funcs and a negative error is returned. The definitions of the helper funcs are generated from comments in the bpf uapi header at `include/uapi/linux/bpf.h`. The return type of these helpers was previously changed from int to long in commit bdb7b79b4ce8. For any case where one of the map helpers call the bpf_map_ops funcs that are still returning 32-bit int, a compiler might not include sign extension instructions to properly convert the 32-bit negative value a 64-bit negative value. For example: bpf assembly excerpt of an inlined helper calling a kernel function and checking for a specific error: ; err = bpf_map_update_elem(&mymap, &key, &val, BPF_NOEXIST); ... 46: call 0xffffffffe103291c ; htab_map_update_elem ; if (err && err != -EEXIST) { 4b: cmp $0xffffffffffffffef,%rax ; cmp -EEXIST,%rax kernel function assembly excerpt of return value from `htab_map_update_elem` returning 32-bit int: movl $0xffffffef, %r9d ... movl %r9d, %eax ...results in the comparison: cmp $0xffffffffffffffef, $0x00000000ffffffef Fixes: bdb7b79b4ce8 ("bpf: Switch most helper return values from 32-bit int to 64-bit long") Tested-by: Eduard Zingerman <eddyz87@gmail.com> Signed-off-by: JP Kobryn <inwardvessel@gmail.com> Link: https://lore.kernel.org/r/20230322194754.185781-3-inwardvessel@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-22 19:47:54 +00:00
static long __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
void *value, u64 map_flags,
bool onallcpus)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
struct htab_elem *l_new = NULL, *l_old;
struct hlist_nulls_head *head;
unsigned long flags;
struct bucket *b;
u32 key_size, hash;
int ret;
if (unlikely(map_flags > BPF_EXIST))
/* unknown flags */
return -EINVAL;
WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
!rcu_read_lock_bh_held());
key_size = map->key_size;
hash = htab_map_hash(key, key_size, htab->hashrnd);
b = __select_bucket(htab, hash);
head = &b->head;
ret = htab_lock_bucket(htab, b, hash, &flags);
if (ret)
return ret;
l_old = lookup_elem_raw(head, hash, key, key_size);
ret = check_flags(htab, l_old, map_flags);
if (ret)
goto err;
if (l_old) {
/* per-cpu hash map can update value in-place */
pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
value, onallcpus);
} else {
l_new = alloc_htab_elem(htab, key, value, key_size,
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
hash, true, onallcpus, NULL);
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
if (IS_ERR(l_new)) {
ret = PTR_ERR(l_new);
goto err;
}
hlist_nulls_add_head_rcu(&l_new->hash_node, head);
}
ret = 0;
err:
htab_unlock_bucket(htab, b, hash, flags);
return ret;
}
bpf: return long from bpf_map_ops funcs This patch changes the return types of bpf_map_ops functions to long, where previously int was returned. Using long allows for bpf programs to maintain the sign bit in the absence of sign extension during situations where inlined bpf helper funcs make calls to the bpf_map_ops funcs and a negative error is returned. The definitions of the helper funcs are generated from comments in the bpf uapi header at `include/uapi/linux/bpf.h`. The return type of these helpers was previously changed from int to long in commit bdb7b79b4ce8. For any case where one of the map helpers call the bpf_map_ops funcs that are still returning 32-bit int, a compiler might not include sign extension instructions to properly convert the 32-bit negative value a 64-bit negative value. For example: bpf assembly excerpt of an inlined helper calling a kernel function and checking for a specific error: ; err = bpf_map_update_elem(&mymap, &key, &val, BPF_NOEXIST); ... 46: call 0xffffffffe103291c ; htab_map_update_elem ; if (err && err != -EEXIST) { 4b: cmp $0xffffffffffffffef,%rax ; cmp -EEXIST,%rax kernel function assembly excerpt of return value from `htab_map_update_elem` returning 32-bit int: movl $0xffffffef, %r9d ... movl %r9d, %eax ...results in the comparison: cmp $0xffffffffffffffef, $0x00000000ffffffef Fixes: bdb7b79b4ce8 ("bpf: Switch most helper return values from 32-bit int to 64-bit long") Tested-by: Eduard Zingerman <eddyz87@gmail.com> Signed-off-by: JP Kobryn <inwardvessel@gmail.com> Link: https://lore.kernel.org/r/20230322194754.185781-3-inwardvessel@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-22 19:47:54 +00:00
static long __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
void *value, u64 map_flags,
bool onallcpus)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
struct htab_elem *l_new = NULL, *l_old;
struct hlist_nulls_head *head;
unsigned long flags;
struct bucket *b;
u32 key_size, hash;
int ret;
if (unlikely(map_flags > BPF_EXIST))
/* unknown flags */
return -EINVAL;
WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
!rcu_read_lock_bh_held());
key_size = map->key_size;
hash = htab_map_hash(key, key_size, htab->hashrnd);
b = __select_bucket(htab, hash);
head = &b->head;
/* For LRU, we need to alloc before taking bucket's
* spinlock because LRU's elem alloc may need
* to remove older elem from htab and this removal
* operation will need a bucket lock.
*/
if (map_flags != BPF_EXIST) {
l_new = prealloc_lru_pop(htab, key, hash);
if (!l_new)
return -ENOMEM;
}
ret = htab_lock_bucket(htab, b, hash, &flags);
if (ret)
goto err_lock_bucket;
l_old = lookup_elem_raw(head, hash, key, key_size);
ret = check_flags(htab, l_old, map_flags);
if (ret)
goto err;
if (l_old) {
bpf_lru_node_set_ref(&l_old->lru_node);
/* per-cpu hash map can update value in-place */
pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
value, onallcpus);
} else {
pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
value, onallcpus);
hlist_nulls_add_head_rcu(&l_new->hash_node, head);
l_new = NULL;
}
ret = 0;
err:
htab_unlock_bucket(htab, b, hash, flags);
err_lock_bucket:
if (l_new) {
bpf_map_dec_elem_count(&htab->map);
bpf_lru_push_free(&htab->lru, &l_new->lru_node);
}
return ret;
}
bpf: return long from bpf_map_ops funcs This patch changes the return types of bpf_map_ops functions to long, where previously int was returned. Using long allows for bpf programs to maintain the sign bit in the absence of sign extension during situations where inlined bpf helper funcs make calls to the bpf_map_ops funcs and a negative error is returned. The definitions of the helper funcs are generated from comments in the bpf uapi header at `include/uapi/linux/bpf.h`. The return type of these helpers was previously changed from int to long in commit bdb7b79b4ce8. For any case where one of the map helpers call the bpf_map_ops funcs that are still returning 32-bit int, a compiler might not include sign extension instructions to properly convert the 32-bit negative value a 64-bit negative value. For example: bpf assembly excerpt of an inlined helper calling a kernel function and checking for a specific error: ; err = bpf_map_update_elem(&mymap, &key, &val, BPF_NOEXIST); ... 46: call 0xffffffffe103291c ; htab_map_update_elem ; if (err && err != -EEXIST) { 4b: cmp $0xffffffffffffffef,%rax ; cmp -EEXIST,%rax kernel function assembly excerpt of return value from `htab_map_update_elem` returning 32-bit int: movl $0xffffffef, %r9d ... movl %r9d, %eax ...results in the comparison: cmp $0xffffffffffffffef, $0x00000000ffffffef Fixes: bdb7b79b4ce8 ("bpf: Switch most helper return values from 32-bit int to 64-bit long") Tested-by: Eduard Zingerman <eddyz87@gmail.com> Signed-off-by: JP Kobryn <inwardvessel@gmail.com> Link: https://lore.kernel.org/r/20230322194754.185781-3-inwardvessel@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-22 19:47:54 +00:00
static long htab_percpu_map_update_elem(struct bpf_map *map, void *key,
void *value, u64 map_flags)
{
return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
}
bpf: return long from bpf_map_ops funcs This patch changes the return types of bpf_map_ops functions to long, where previously int was returned. Using long allows for bpf programs to maintain the sign bit in the absence of sign extension during situations where inlined bpf helper funcs make calls to the bpf_map_ops funcs and a negative error is returned. The definitions of the helper funcs are generated from comments in the bpf uapi header at `include/uapi/linux/bpf.h`. The return type of these helpers was previously changed from int to long in commit bdb7b79b4ce8. For any case where one of the map helpers call the bpf_map_ops funcs that are still returning 32-bit int, a compiler might not include sign extension instructions to properly convert the 32-bit negative value a 64-bit negative value. For example: bpf assembly excerpt of an inlined helper calling a kernel function and checking for a specific error: ; err = bpf_map_update_elem(&mymap, &key, &val, BPF_NOEXIST); ... 46: call 0xffffffffe103291c ; htab_map_update_elem ; if (err && err != -EEXIST) { 4b: cmp $0xffffffffffffffef,%rax ; cmp -EEXIST,%rax kernel function assembly excerpt of return value from `htab_map_update_elem` returning 32-bit int: movl $0xffffffef, %r9d ... movl %r9d, %eax ...results in the comparison: cmp $0xffffffffffffffef, $0x00000000ffffffef Fixes: bdb7b79b4ce8 ("bpf: Switch most helper return values from 32-bit int to 64-bit long") Tested-by: Eduard Zingerman <eddyz87@gmail.com> Signed-off-by: JP Kobryn <inwardvessel@gmail.com> Link: https://lore.kernel.org/r/20230322194754.185781-3-inwardvessel@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-22 19:47:54 +00:00
static long htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
void *value, u64 map_flags)
{
return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
false);
}
/* Called from syscall or from eBPF program */
bpf: return long from bpf_map_ops funcs This patch changes the return types of bpf_map_ops functions to long, where previously int was returned. Using long allows for bpf programs to maintain the sign bit in the absence of sign extension during situations where inlined bpf helper funcs make calls to the bpf_map_ops funcs and a negative error is returned. The definitions of the helper funcs are generated from comments in the bpf uapi header at `include/uapi/linux/bpf.h`. The return type of these helpers was previously changed from int to long in commit bdb7b79b4ce8. For any case where one of the map helpers call the bpf_map_ops funcs that are still returning 32-bit int, a compiler might not include sign extension instructions to properly convert the 32-bit negative value a 64-bit negative value. For example: bpf assembly excerpt of an inlined helper calling a kernel function and checking for a specific error: ; err = bpf_map_update_elem(&mymap, &key, &val, BPF_NOEXIST); ... 46: call 0xffffffffe103291c ; htab_map_update_elem ; if (err && err != -EEXIST) { 4b: cmp $0xffffffffffffffef,%rax ; cmp -EEXIST,%rax kernel function assembly excerpt of return value from `htab_map_update_elem` returning 32-bit int: movl $0xffffffef, %r9d ... movl %r9d, %eax ...results in the comparison: cmp $0xffffffffffffffef, $0x00000000ffffffef Fixes: bdb7b79b4ce8 ("bpf: Switch most helper return values from 32-bit int to 64-bit long") Tested-by: Eduard Zingerman <eddyz87@gmail.com> Signed-off-by: JP Kobryn <inwardvessel@gmail.com> Link: https://lore.kernel.org/r/20230322194754.185781-3-inwardvessel@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-22 19:47:54 +00:00
static long htab_map_delete_elem(struct bpf_map *map, void *key)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
struct hlist_nulls_head *head;
struct bucket *b;
struct htab_elem *l;
unsigned long flags;
u32 hash, key_size;
int ret;
WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
!rcu_read_lock_bh_held());
key_size = map->key_size;
hash = htab_map_hash(key, key_size, htab->hashrnd);
b = __select_bucket(htab, hash);
head = &b->head;
ret = htab_lock_bucket(htab, b, hash, &flags);
if (ret)
return ret;
l = lookup_elem_raw(head, hash, key, key_size);
if (l) {
hlist_nulls_del_rcu(&l->hash_node);
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
free_htab_elem(htab, l);
} else {
ret = -ENOENT;
}
htab_unlock_bucket(htab, b, hash, flags);
return ret;
}
bpf: return long from bpf_map_ops funcs This patch changes the return types of bpf_map_ops functions to long, where previously int was returned. Using long allows for bpf programs to maintain the sign bit in the absence of sign extension during situations where inlined bpf helper funcs make calls to the bpf_map_ops funcs and a negative error is returned. The definitions of the helper funcs are generated from comments in the bpf uapi header at `include/uapi/linux/bpf.h`. The return type of these helpers was previously changed from int to long in commit bdb7b79b4ce8. For any case where one of the map helpers call the bpf_map_ops funcs that are still returning 32-bit int, a compiler might not include sign extension instructions to properly convert the 32-bit negative value a 64-bit negative value. For example: bpf assembly excerpt of an inlined helper calling a kernel function and checking for a specific error: ; err = bpf_map_update_elem(&mymap, &key, &val, BPF_NOEXIST); ... 46: call 0xffffffffe103291c ; htab_map_update_elem ; if (err && err != -EEXIST) { 4b: cmp $0xffffffffffffffef,%rax ; cmp -EEXIST,%rax kernel function assembly excerpt of return value from `htab_map_update_elem` returning 32-bit int: movl $0xffffffef, %r9d ... movl %r9d, %eax ...results in the comparison: cmp $0xffffffffffffffef, $0x00000000ffffffef Fixes: bdb7b79b4ce8 ("bpf: Switch most helper return values from 32-bit int to 64-bit long") Tested-by: Eduard Zingerman <eddyz87@gmail.com> Signed-off-by: JP Kobryn <inwardvessel@gmail.com> Link: https://lore.kernel.org/r/20230322194754.185781-3-inwardvessel@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-22 19:47:54 +00:00
static long htab_lru_map_delete_elem(struct bpf_map *map, void *key)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
struct hlist_nulls_head *head;
struct bucket *b;
struct htab_elem *l;
unsigned long flags;
u32 hash, key_size;
int ret;
WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
!rcu_read_lock_bh_held());
key_size = map->key_size;
hash = htab_map_hash(key, key_size, htab->hashrnd);
b = __select_bucket(htab, hash);
head = &b->head;
ret = htab_lock_bucket(htab, b, hash, &flags);
if (ret)
return ret;
l = lookup_elem_raw(head, hash, key, key_size);
if (l)
hlist_nulls_del_rcu(&l->hash_node);
else
ret = -ENOENT;
htab_unlock_bucket(htab, b, hash, flags);
if (l)
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
htab_lru_push_free(htab, l);
return ret;
}
static void delete_all_elements(struct bpf_htab *htab)
{
int i;
/* It's called from a worker thread, so disable migration here,
* since bpf_mem_cache_free() relies on that.
*/
migrate_disable();
for (i = 0; i < htab->n_buckets; i++) {
struct hlist_nulls_head *head = select_bucket(htab, i);
struct hlist_nulls_node *n;
struct htab_elem *l;
hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
hlist_nulls_del_rcu(&l->hash_node);
bpf: fix hashmap extra_elems logic In both kmalloc and prealloc mode the bpf_map_update_elem() is using per-cpu extra_elems to do atomic update when the map is full. There are two issues with it. The logic can be misused, since it allows max_entries+num_cpus elements to be present in the map. And alloc_extra_elems() at map creation time can fail percpu alloc for large map values with a warn: WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60 illegal size (32824) or align (8) for percpu allocation The fixes for both of these issues are different for kmalloc and prealloc modes. For prealloc mode allocate extra num_possible_cpus elements and store their pointers into extra_elems array instead of actual elements. Hence we can use these hidden(spare) elements not only when the map is full but during bpf_map_update_elem() that replaces existing element too. That also improves performance, since pcpu_freelist_pop/push is avoided. Unfortunately this approach cannot be used for kmalloc mode which needs to kfree elements after rcu grace period. Therefore switch it back to normal kmalloc even when full and old element exists like it was prior to commit 6c9059817432 ("bpf: pre-allocate hash map elements"). Add tests to check for over max_entries and large map values. Reported-by: Dave Jones <davej@codemonkey.org.uk> Fixes: 6c9059817432 ("bpf: pre-allocate hash map elements") Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-03-22 02:05:04 +00:00
htab_elem_free(htab, l);
}
cond_resched();
}
migrate_enable();
}
static void htab_free_malloced_timers_and_wq(struct bpf_htab *htab)
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
{
int i;
rcu_read_lock();
for (i = 0; i < htab->n_buckets; i++) {
struct hlist_nulls_head *head = select_bucket(htab, i);
struct hlist_nulls_node *n;
struct htab_elem *l;
bpf: Wire up freeing of referenced kptr A destructor kfunc can be defined as void func(type *), where type may be void or any other pointer type as per convenience. In this patch, we ensure that the type is sane and capture the function pointer into off_desc of ptr_off_tab for the specific pointer offset, with the invariant that the dtor pointer is always set when 'kptr_ref' tag is applied to the pointer's pointee type, which is indicated by the flag BPF_MAP_VALUE_OFF_F_REF. Note that only BTF IDs whose destructor kfunc is registered, thus become the allowed BTF IDs for embedding as referenced kptr. Hence it serves the purpose of finding dtor kfunc BTF ID, as well acting as a check against the whitelist of allowed BTF IDs for this purpose. Finally, wire up the actual freeing of the referenced pointer if any at all available offsets, so that no references are leaked after the BPF map goes away and the BPF program previously moved the ownership a referenced pointer into it. The behavior is similar to BPF timers, where bpf_map_{update,delete}_elem will free any existing referenced kptr. The same case is with LRU map's bpf_lru_push_free/htab_lru_push_free functions, which are extended to reset unreferenced and free referenced kptr. Note that unlike BPF timers, kptr is not reset or freed when map uref drops to zero. Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20220424214901.2743946-8-memxor@gmail.com
2022-04-24 21:48:55 +00:00
hlist_nulls_for_each_entry(l, n, head, hash_node) {
/* We only free timer on uref dropping to zero */
if (btf_record_has_field(htab->map.record, BPF_TIMER))
bpf_obj_free_timer(htab->map.record,
l->key + round_up(htab->map.key_size, 8));
if (btf_record_has_field(htab->map.record, BPF_WORKQUEUE))
bpf_obj_free_workqueue(htab->map.record,
l->key + round_up(htab->map.key_size, 8));
bpf: Wire up freeing of referenced kptr A destructor kfunc can be defined as void func(type *), where type may be void or any other pointer type as per convenience. In this patch, we ensure that the type is sane and capture the function pointer into off_desc of ptr_off_tab for the specific pointer offset, with the invariant that the dtor pointer is always set when 'kptr_ref' tag is applied to the pointer's pointee type, which is indicated by the flag BPF_MAP_VALUE_OFF_F_REF. Note that only BTF IDs whose destructor kfunc is registered, thus become the allowed BTF IDs for embedding as referenced kptr. Hence it serves the purpose of finding dtor kfunc BTF ID, as well acting as a check against the whitelist of allowed BTF IDs for this purpose. Finally, wire up the actual freeing of the referenced pointer if any at all available offsets, so that no references are leaked after the BPF map goes away and the BPF program previously moved the ownership a referenced pointer into it. The behavior is similar to BPF timers, where bpf_map_{update,delete}_elem will free any existing referenced kptr. The same case is with LRU map's bpf_lru_push_free/htab_lru_push_free functions, which are extended to reset unreferenced and free referenced kptr. Note that unlike BPF timers, kptr is not reset or freed when map uref drops to zero. Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20220424214901.2743946-8-memxor@gmail.com
2022-04-24 21:48:55 +00:00
}
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
cond_resched_rcu();
}
rcu_read_unlock();
}
static void htab_map_free_timers_and_wq(struct bpf_map *map)
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
/* We only free timer and workqueue on uref dropping to zero */
if (btf_record_has_field(htab->map.record, BPF_TIMER | BPF_WORKQUEUE)) {
if (!htab_is_prealloc(htab))
htab_free_malloced_timers_and_wq(htab);
else
htab_free_prealloced_timers_and_wq(htab);
}
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
}
/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
static void htab_map_free(struct bpf_map *map)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
int i;
/* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
* bpf_free_used_maps() is called after bpf prog is no longer executing.
* There is no need to synchronize_rcu() here to protect map elements.
*/
/* htab no longer uses call_rcu() directly. bpf_mem_alloc does it
* underneath and is responsible for waiting for callbacks to finish
* during bpf_mem_alloc_destroy().
*/
bpf: Wire up freeing of referenced kptr A destructor kfunc can be defined as void func(type *), where type may be void or any other pointer type as per convenience. In this patch, we ensure that the type is sane and capture the function pointer into off_desc of ptr_off_tab for the specific pointer offset, with the invariant that the dtor pointer is always set when 'kptr_ref' tag is applied to the pointer's pointee type, which is indicated by the flag BPF_MAP_VALUE_OFF_F_REF. Note that only BTF IDs whose destructor kfunc is registered, thus become the allowed BTF IDs for embedding as referenced kptr. Hence it serves the purpose of finding dtor kfunc BTF ID, as well acting as a check against the whitelist of allowed BTF IDs for this purpose. Finally, wire up the actual freeing of the referenced pointer if any at all available offsets, so that no references are leaked after the BPF map goes away and the BPF program previously moved the ownership a referenced pointer into it. The behavior is similar to BPF timers, where bpf_map_{update,delete}_elem will free any existing referenced kptr. The same case is with LRU map's bpf_lru_push_free/htab_lru_push_free functions, which are extended to reset unreferenced and free referenced kptr. Note that unlike BPF timers, kptr is not reset or freed when map uref drops to zero. Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20220424214901.2743946-8-memxor@gmail.com
2022-04-24 21:48:55 +00:00
if (!htab_is_prealloc(htab)) {
bpf: pre-allocate hash map elements If kprobe is placed on spin_unlock then calling kmalloc/kfree from bpf programs is not safe, since the following dead lock is possible: kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe-> bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock) and deadlocks. The following solutions were considered and some implemented, but eventually discarded - kmem_cache_create for every map - add recursion check to slow-path of slub - use reserved memory in bpf_map_update for in_irq or in preempt_disabled - kmalloc via irq_work At the end pre-allocation of all map elements turned out to be the simplest solution and since the user is charged upfront for all the memory, such pre-allocation doesn't affect the user space visible behavior. Since it's impossible to tell whether kprobe is triggered in a safe location from kmalloc point of view, use pre-allocation by default and introduce new BPF_F_NO_PREALLOC flag. While testing of per-cpu hash maps it was discovered that alloc_percpu(GFP_ATOMIC) has odd corner cases and often fails to allocate memory even when 90% of it is free. The pre-allocation of per-cpu hash elements solves this problem as well. Turned out that bpf_map_update() quickly followed by bpf_map_lookup()+bpf_map_delete() is very common pattern used in many of iovisor/bcc/tools, so there is additional benefit of pre-allocation, since such use cases are must faster. Since all hash map elements are now pre-allocated we can remove atomic increment of htab->count and save few more cycles. Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid large malloc/free done by users who don't have sufficient limits. Pre-allocation is done with vmalloc and alloc/free is done via percpu_freelist. Here are performance numbers for different pre-allocation algorithms that were implemented, but discarded in favor of percpu_freelist: 1 cpu: pcpu_ida 2.1M pcpu_ida nolock 2.3M bt 2.4M kmalloc 1.8M hlist+spinlock 2.3M pcpu_freelist 2.6M 4 cpu: pcpu_ida 1.5M pcpu_ida nolock 1.8M bt w/smp_align 1.7M bt no/smp_align 1.1M kmalloc 0.7M hlist+spinlock 0.2M pcpu_freelist 2.0M 8 cpu: pcpu_ida 0.7M bt w/smp_align 0.8M kmalloc 0.4M pcpu_freelist 1.5M 32 cpu: kmalloc 0.13M pcpu_freelist 0.49M pcpu_ida nolock is a modified percpu_ida algorithm without percpu_ida_cpu locks and without cross-cpu tag stealing. It's faster than existing percpu_ida, but not as fast as pcpu_freelist. bt is a variant of block/blk-mq-tag.c simlified and customized for bpf use case. bt w/smp_align is using cache line for every 'long' (similar to blk-mq-tag). bt no/smp_align allocates 'long' bitmasks continuously to save memory. It's comparable to percpu_ida and in some cases faster, but slower than percpu_freelist hlist+spinlock is the simplest free list with single spinlock. As expeceted it has very bad scaling in SMP. kmalloc is existing implementation which is still available via BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist, but saves memory, so in cases where map->max_entries can be large and number of map update/delete per second is low, it may make sense to use it. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 05:57:15 +00:00
delete_all_elements(htab);
bpf: Wire up freeing of referenced kptr A destructor kfunc can be defined as void func(type *), where type may be void or any other pointer type as per convenience. In this patch, we ensure that the type is sane and capture the function pointer into off_desc of ptr_off_tab for the specific pointer offset, with the invariant that the dtor pointer is always set when 'kptr_ref' tag is applied to the pointer's pointee type, which is indicated by the flag BPF_MAP_VALUE_OFF_F_REF. Note that only BTF IDs whose destructor kfunc is registered, thus become the allowed BTF IDs for embedding as referenced kptr. Hence it serves the purpose of finding dtor kfunc BTF ID, as well acting as a check against the whitelist of allowed BTF IDs for this purpose. Finally, wire up the actual freeing of the referenced pointer if any at all available offsets, so that no references are leaked after the BPF map goes away and the BPF program previously moved the ownership a referenced pointer into it. The behavior is similar to BPF timers, where bpf_map_{update,delete}_elem will free any existing referenced kptr. The same case is with LRU map's bpf_lru_push_free/htab_lru_push_free functions, which are extended to reset unreferenced and free referenced kptr. Note that unlike BPF timers, kptr is not reset or freed when map uref drops to zero. Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20220424214901.2743946-8-memxor@gmail.com
2022-04-24 21:48:55 +00:00
} else {
htab_free_prealloced_fields(htab);
prealloc_destroy(htab);
bpf: Wire up freeing of referenced kptr A destructor kfunc can be defined as void func(type *), where type may be void or any other pointer type as per convenience. In this patch, we ensure that the type is sane and capture the function pointer into off_desc of ptr_off_tab for the specific pointer offset, with the invariant that the dtor pointer is always set when 'kptr_ref' tag is applied to the pointer's pointee type, which is indicated by the flag BPF_MAP_VALUE_OFF_F_REF. Note that only BTF IDs whose destructor kfunc is registered, thus become the allowed BTF IDs for embedding as referenced kptr. Hence it serves the purpose of finding dtor kfunc BTF ID, as well acting as a check against the whitelist of allowed BTF IDs for this purpose. Finally, wire up the actual freeing of the referenced pointer if any at all available offsets, so that no references are leaked after the BPF map goes away and the BPF program previously moved the ownership a referenced pointer into it. The behavior is similar to BPF timers, where bpf_map_{update,delete}_elem will free any existing referenced kptr. The same case is with LRU map's bpf_lru_push_free/htab_lru_push_free functions, which are extended to reset unreferenced and free referenced kptr. Note that unlike BPF timers, kptr is not reset or freed when map uref drops to zero. Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20220424214901.2743946-8-memxor@gmail.com
2022-04-24 21:48:55 +00:00
}
bpf_map_free_elem_count(map);
free_percpu(htab->extra_elems);
bpf_map_area_free(htab->buckets);
bpf_mem_alloc_destroy(&htab->pcpu_ma);
bpf_mem_alloc_destroy(&htab->ma);
if (htab->use_percpu_counter)
percpu_counter_destroy(&htab->pcount);
for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
free_percpu(htab->map_locked[i]);
bpf: Fix error path in htab_map_alloc() syzbot was able to trigger a use-after-free in htab_map_alloc() [1] htab_map_alloc() lacks a call to lockdep_unregister_key() in its error path. lockdep_register_key() and lockdep_unregister_key() can not fail, it seems better to use them right after htab allocation and before htab freeing, avoiding more goto/labels in htab_map_alloc() [1] BUG: KASAN: use-after-free in lockdep_register_key+0x356/0x3e0 kernel/locking/lockdep.c:1182 Read of size 8 at addr ffff88805fa67ad8 by task syz-executor.3/2356 CPU: 1 PID: 2356 Comm: syz-executor.3 Not tainted 5.9.0-syzkaller #0 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:77 [inline] dump_stack+0x107/0x163 lib/dump_stack.c:118 print_address_description.constprop.0.cold+0xae/0x4c8 mm/kasan/report.c:385 __kasan_report mm/kasan/report.c:545 [inline] kasan_report.cold+0x1f/0x37 mm/kasan/report.c:562 lockdep_register_key+0x356/0x3e0 kernel/locking/lockdep.c:1182 htab_init_buckets kernel/bpf/hashtab.c:144 [inline] htab_map_alloc+0x6c5/0x14a0 kernel/bpf/hashtab.c:521 find_and_alloc_map kernel/bpf/syscall.c:122 [inline] map_create kernel/bpf/syscall.c:825 [inline] __do_sys_bpf+0xa80/0x5180 kernel/bpf/syscall.c:4381 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x45deb9 Code: 0d b4 fb ff c3 66 2e 0f 1f 84 00 00 00 00 00 66 90 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 0f 83 db b3 fb ff c3 66 2e 0f 1f 84 00 00 00 00 RSP: 002b:00007f0eafee1c78 EFLAGS: 00000246 ORIG_RAX: 0000000000000141 RAX: ffffffffffffffda RBX: 0000000000001a00 RCX: 000000000045deb9 RDX: 0000000000000040 RSI: 0000000020000040 RDI: 405a020000000000 RBP: 000000000118bf60 R08: 0000000000000000 R09: 0000000000000000 R10: 0000000000000000 R11: 0000000000000246 R12: 000000000118bf2c R13: 00007ffd3cf9eabf R14: 00007f0eafee29c0 R15: 000000000118bf2c Allocated by task 2053: kasan_save_stack+0x1b/0x40 mm/kasan/common.c:48 kasan_set_track mm/kasan/common.c:56 [inline] __kasan_kmalloc.constprop.0+0xc2/0xd0 mm/kasan/common.c:461 kmalloc include/linux/slab.h:554 [inline] kzalloc include/linux/slab.h:666 [inline] htab_map_alloc+0xdf/0x14a0 kernel/bpf/hashtab.c:454 find_and_alloc_map kernel/bpf/syscall.c:122 [inline] map_create kernel/bpf/syscall.c:825 [inline] __do_sys_bpf+0xa80/0x5180 kernel/bpf/syscall.c:4381 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Freed by task 2053: kasan_save_stack+0x1b/0x40 mm/kasan/common.c:48 kasan_set_track+0x1c/0x30 mm/kasan/common.c:56 kasan_set_free_info+0x1b/0x30 mm/kasan/generic.c:355 __kasan_slab_free+0x102/0x140 mm/kasan/common.c:422 slab_free_hook mm/slub.c:1544 [inline] slab_free_freelist_hook+0x5d/0x150 mm/slub.c:1577 slab_free mm/slub.c:3142 [inline] kfree+0xdb/0x360 mm/slub.c:4124 htab_map_alloc+0x3f9/0x14a0 kernel/bpf/hashtab.c:549 find_and_alloc_map kernel/bpf/syscall.c:122 [inline] map_create kernel/bpf/syscall.c:825 [inline] __do_sys_bpf+0xa80/0x5180 kernel/bpf/syscall.c:4381 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46 entry_SYSCALL_64_after_hwframe+0x44/0xa9 The buggy address belongs to the object at ffff88805fa67800 which belongs to the cache kmalloc-1k of size 1024 The buggy address is located 728 bytes inside of 1024-byte region [ffff88805fa67800, ffff88805fa67c00) The buggy address belongs to the page: page:000000003c5582c4 refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x5fa60 head:000000003c5582c4 order:3 compound_mapcount:0 compound_pincount:0 flags: 0xfff00000010200(slab|head) raw: 00fff00000010200 ffffea0000bc1200 0000000200000002 ffff888010041140 raw: 0000000000000000 0000000000100010 00000001ffffffff 0000000000000000 page dumped because: kasan: bad access detected Memory state around the buggy address: ffff88805fa67980: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ffff88805fa67a00: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ^ ffff88805fa67b00: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ffff88805fa67b80: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb Fixes: c50eb518e262 ("bpf: Use separate lockdep class for each hashtab") Reported-by: syzbot <syzkaller@googlegroups.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20201102114100.3103180-1-eric.dumazet@gmail.com
2020-11-02 11:41:00 +00:00
lockdep_unregister_key(&htab->lockdep_key);
bpf_map_area_free(htab);
}
static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
struct seq_file *m)
{
void *value;
rcu_read_lock();
value = htab_map_lookup_elem(map, key);
if (!value) {
rcu_read_unlock();
return;
}
btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
seq_puts(m, ": ");
btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
seq_puts(m, "\n");
rcu_read_unlock();
}
static int __htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
void *value, bool is_lru_map,
bool is_percpu, u64 flags)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
struct hlist_nulls_head *head;
unsigned long bflags;
struct htab_elem *l;
u32 hash, key_size;
struct bucket *b;
int ret;
key_size = map->key_size;
hash = htab_map_hash(key, key_size, htab->hashrnd);
b = __select_bucket(htab, hash);
head = &b->head;
ret = htab_lock_bucket(htab, b, hash, &bflags);
if (ret)
return ret;
l = lookup_elem_raw(head, hash, key, key_size);
if (!l) {
ret = -ENOENT;
} else {
if (is_percpu) {
u32 roundup_value_size = round_up(map->value_size, 8);
void __percpu *pptr;
int off = 0, cpu;
pptr = htab_elem_get_ptr(l, key_size);
for_each_possible_cpu(cpu) {
copy_map_value_long(&htab->map, value + off, per_cpu_ptr(pptr, cpu));
check_and_init_map_value(&htab->map, value + off);
off += roundup_value_size;
}
} else {
u32 roundup_key_size = round_up(map->key_size, 8);
if (flags & BPF_F_LOCK)
copy_map_value_locked(map, value, l->key +
roundup_key_size,
true);
else
copy_map_value(map, value, l->key +
roundup_key_size);
/* Zeroing special fields in the temp buffer */
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
check_and_init_map_value(map, value);
}
hlist_nulls_del_rcu(&l->hash_node);
if (!is_lru_map)
free_htab_elem(htab, l);
}
htab_unlock_bucket(htab, b, hash, bflags);
if (is_lru_map && l)
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
htab_lru_push_free(htab, l);
return ret;
}
static int htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
void *value, u64 flags)
{
return __htab_map_lookup_and_delete_elem(map, key, value, false, false,
flags);
}
static int htab_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
void *key, void *value,
u64 flags)
{
return __htab_map_lookup_and_delete_elem(map, key, value, false, true,
flags);
}
static int htab_lru_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
void *value, u64 flags)
{
return __htab_map_lookup_and_delete_elem(map, key, value, true, false,
flags);
}
static int htab_lru_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
void *key, void *value,
u64 flags)
{
return __htab_map_lookup_and_delete_elem(map, key, value, true, true,
flags);
}
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
static int
__htab_map_lookup_and_delete_batch(struct bpf_map *map,
const union bpf_attr *attr,
union bpf_attr __user *uattr,
bool do_delete, bool is_lru_map,
bool is_percpu)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
void __user *uvalues = u64_to_user_ptr(attr->batch.values);
void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
void __user *ubatch = u64_to_user_ptr(attr->batch.in_batch);
u32 batch, max_count, size, bucket_size, map_id;
bpf: Fix a potential deadlock with bpf_map_do_batch Commit 057996380a42 ("bpf: Add batch ops to all htab bpf map") added lookup_and_delete batch operation for hash table. The current implementation has bpf_lru_push_free() inside the bucket lock, which may cause a deadlock. syzbot reports: -> #2 (&htab->buckets[i].lock#2){....}: __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159 htab_lru_map_delete_node+0xce/0x2f0 kernel/bpf/hashtab.c:593 __bpf_lru_list_shrink_inactive kernel/bpf/bpf_lru_list.c:220 [inline] __bpf_lru_list_shrink+0xf9/0x470 kernel/bpf/bpf_lru_list.c:266 bpf_lru_list_pop_free_to_local kernel/bpf/bpf_lru_list.c:340 [inline] bpf_common_lru_pop_free kernel/bpf/bpf_lru_list.c:447 [inline] bpf_lru_pop_free+0x87c/0x1670 kernel/bpf/bpf_lru_list.c:499 prealloc_lru_pop+0x2c/0xa0 kernel/bpf/hashtab.c:132 __htab_lru_percpu_map_update_elem+0x67e/0xa90 kernel/bpf/hashtab.c:1069 bpf_percpu_hash_update+0x16e/0x210 kernel/bpf/hashtab.c:1585 bpf_map_update_value.isra.0+0x2d7/0x8e0 kernel/bpf/syscall.c:181 generic_map_update_batch+0x41f/0x610 kernel/bpf/syscall.c:1319 bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348 __do_sys_bpf+0x9b7/0x41e0 kernel/bpf/syscall.c:3460 __se_sys_bpf kernel/bpf/syscall.c:3355 [inline] __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355 do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294 entry_SYSCALL_64_after_hwframe+0x49/0xbe -> #0 (&loc_l->lock){....}: check_prev_add kernel/locking/lockdep.c:2475 [inline] check_prevs_add kernel/locking/lockdep.c:2580 [inline] validate_chain kernel/locking/lockdep.c:2970 [inline] __lock_acquire+0x2596/0x4a00 kernel/locking/lockdep.c:3954 lock_acquire+0x190/0x410 kernel/locking/lockdep.c:4484 __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159 bpf_common_lru_push_free kernel/bpf/bpf_lru_list.c:516 [inline] bpf_lru_push_free+0x250/0x5b0 kernel/bpf/bpf_lru_list.c:555 __htab_map_lookup_and_delete_batch+0x8d4/0x1540 kernel/bpf/hashtab.c:1374 htab_lru_map_lookup_and_delete_batch+0x34/0x40 kernel/bpf/hashtab.c:1491 bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348 __do_sys_bpf+0x1f7d/0x41e0 kernel/bpf/syscall.c:3456 __se_sys_bpf kernel/bpf/syscall.c:3355 [inline] __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355 do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294 entry_SYSCALL_64_after_hwframe+0x49/0xbe Possible unsafe locking scenario: CPU0 CPU2 ---- ---- lock(&htab->buckets[i].lock#2); lock(&l->lock); lock(&htab->buckets[i].lock#2); lock(&loc_l->lock); *** DEADLOCK *** To fix the issue, for htab_lru_map_lookup_and_delete_batch() in CPU0, let us do bpf_lru_push_free() out of the htab bucket lock. This can avoid the above deadlock scenario. Fixes: 057996380a42 ("bpf: Add batch ops to all htab bpf map") Reported-by: syzbot+a38ff3d9356388f2fb83@syzkaller.appspotmail.com Reported-by: syzbot+122b5421d14e68f29cd1@syzkaller.appspotmail.com Suggested-by: Hillf Danton <hdanton@sina.com> Suggested-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com> Acked-by: Brian Vazquez <brianvv@google.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Link: https://lore.kernel.org/bpf/20200219234757.3544014-1-yhs@fb.com
2020-02-19 23:47:57 +00:00
struct htab_elem *node_to_free = NULL;
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
u64 elem_map_flags, map_flags;
struct hlist_nulls_head *head;
struct hlist_nulls_node *n;
unsigned long flags = 0;
bool locked = false;
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
struct htab_elem *l;
struct bucket *b;
int ret = 0;
elem_map_flags = attr->batch.elem_flags;
if ((elem_map_flags & ~BPF_F_LOCK) ||
((elem_map_flags & BPF_F_LOCK) && !btf_record_has_field(map->record, BPF_SPIN_LOCK)))
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
return -EINVAL;
map_flags = attr->batch.flags;
if (map_flags)
return -EINVAL;
max_count = attr->batch.count;
if (!max_count)
return 0;
if (put_user(0, &uattr->batch.count))
return -EFAULT;
batch = 0;
if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
return -EFAULT;
if (batch >= htab->n_buckets)
return -ENOENT;
key_size = htab->map.key_size;
roundup_key_size = round_up(htab->map.key_size, 8);
value_size = htab->map.value_size;
size = round_up(value_size, 8);
if (is_percpu)
value_size = size * num_possible_cpus();
total = 0;
/* while experimenting with hash tables with sizes ranging from 10 to
* 1000, it was observed that a bucket can have up to 5 entries.
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
*/
bucket_size = 5;
alloc:
/* We cannot do copy_from_user or copy_to_user inside
* the rcu_read_lock. Allocate enough space here.
*/
bpf: Fix integer overflow involving bucket_size In __htab_map_lookup_and_delete_batch(), hash buckets are iterated over to count the number of elements in each bucket (bucket_size). If bucket_size is large enough, the multiplication to calculate kvmalloc() size could overflow, resulting in out-of-bounds write as reported by KASAN: [...] [ 104.986052] BUG: KASAN: vmalloc-out-of-bounds in __htab_map_lookup_and_delete_batch+0x5ce/0xb60 [ 104.986489] Write of size 4194224 at addr ffffc9010503be70 by task crash/112 [ 104.986889] [ 104.987193] CPU: 0 PID: 112 Comm: crash Not tainted 5.14.0-rc4 #13 [ 104.987552] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.13.0-1ubuntu1.1 04/01/2014 [ 104.988104] Call Trace: [ 104.988410] dump_stack_lvl+0x34/0x44 [ 104.988706] print_address_description.constprop.0+0x21/0x140 [ 104.988991] ? __htab_map_lookup_and_delete_batch+0x5ce/0xb60 [ 104.989327] ? __htab_map_lookup_and_delete_batch+0x5ce/0xb60 [ 104.989622] kasan_report.cold+0x7f/0x11b [ 104.989881] ? __htab_map_lookup_and_delete_batch+0x5ce/0xb60 [ 104.990239] kasan_check_range+0x17c/0x1e0 [ 104.990467] memcpy+0x39/0x60 [ 104.990670] __htab_map_lookup_and_delete_batch+0x5ce/0xb60 [ 104.990982] ? __wake_up_common+0x4d/0x230 [ 104.991256] ? htab_of_map_free+0x130/0x130 [ 104.991541] bpf_map_do_batch+0x1fb/0x220 [...] In hashtable, if the elements' keys have the same jhash() value, the elements will be put into the same bucket. By putting a lot of elements into a single bucket, the value of bucket_size can be increased to trigger the integer overflow. Triggering the overflow is possible for both callers with CAP_SYS_ADMIN and callers without CAP_SYS_ADMIN. It will be trivial for a caller with CAP_SYS_ADMIN to intentionally reach this overflow by enabling BPF_F_ZERO_SEED. As this flag will set the random seed passed to jhash() to 0, it will be easy for the caller to prepare keys which will be hashed into the same value, and thus put all the elements into the same bucket. If the caller does not have CAP_SYS_ADMIN, BPF_F_ZERO_SEED cannot be used. However, it will be still technically possible to trigger the overflow, by guessing the random seed value passed to jhash() (32bit) and repeating the attempt to trigger the overflow. In this case, the probability to trigger the overflow will be low and will take a very long time. Fix the integer overflow by calling kvmalloc_array() instead of kvmalloc() to allocate memory. Fixes: 057996380a42 ("bpf: Add batch ops to all htab bpf map") Signed-off-by: Tatsuhiko Yasumatsu <th.yasumatsu@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20210806150419.109658-1-th.yasumatsu@gmail.com
2021-08-06 15:04:18 +00:00
keys = kvmalloc_array(key_size, bucket_size, GFP_USER | __GFP_NOWARN);
values = kvmalloc_array(value_size, bucket_size, GFP_USER | __GFP_NOWARN);
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
if (!keys || !values) {
ret = -ENOMEM;
goto after_loop;
}
again:
bpf_disable_instrumentation();
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
rcu_read_lock();
again_nocopy:
dst_key = keys;
dst_val = values;
b = &htab->buckets[batch];
head = &b->head;
/* do not grab the lock unless need it (bucket_cnt > 0). */
if (locked) {
ret = htab_lock_bucket(htab, b, batch, &flags);
if (ret) {
rcu_read_unlock();
bpf_enable_instrumentation();
goto after_loop;
}
}
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
bucket_cnt = 0;
hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
bucket_cnt++;
if (bucket_cnt && !locked) {
locked = true;
goto again_nocopy;
}
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
if (bucket_cnt > (max_count - total)) {
if (total == 0)
ret = -ENOSPC;
/* Note that since bucket_cnt > 0 here, it is implicit
* that the locked was grabbed, so release it.
*/
htab_unlock_bucket(htab, b, batch, flags);
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
rcu_read_unlock();
bpf_enable_instrumentation();
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
goto after_loop;
}
if (bucket_cnt > bucket_size) {
bucket_size = bucket_cnt;
/* Note that since bucket_cnt > 0 here, it is implicit
* that the locked was grabbed, so release it.
*/
htab_unlock_bucket(htab, b, batch, flags);
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
rcu_read_unlock();
bpf_enable_instrumentation();
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
kvfree(keys);
kvfree(values);
goto alloc;
}
/* Next block is only safe to run if you have grabbed the lock */
if (!locked)
goto next_batch;
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
memcpy(dst_key, l->key, key_size);
if (is_percpu) {
int off = 0, cpu;
void __percpu *pptr;
pptr = htab_elem_get_ptr(l, map->key_size);
for_each_possible_cpu(cpu) {
copy_map_value_long(&htab->map, dst_val + off, per_cpu_ptr(pptr, cpu));
check_and_init_map_value(&htab->map, dst_val + off);
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
off += size;
}
} else {
value = l->key + roundup_key_size;
if (map->map_type == BPF_MAP_TYPE_HASH_OF_MAPS) {
struct bpf_map **inner_map = value;
/* Actual value is the id of the inner map */
map_id = map->ops->map_fd_sys_lookup_elem(*inner_map);
value = &map_id;
}
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
if (elem_map_flags & BPF_F_LOCK)
copy_map_value_locked(map, dst_val, value,
true);
else
copy_map_value(map, dst_val, value);
/* Zeroing special fields in the temp buffer */
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
check_and_init_map_value(map, dst_val);
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
}
if (do_delete) {
hlist_nulls_del_rcu(&l->hash_node);
bpf: Fix a potential deadlock with bpf_map_do_batch Commit 057996380a42 ("bpf: Add batch ops to all htab bpf map") added lookup_and_delete batch operation for hash table. The current implementation has bpf_lru_push_free() inside the bucket lock, which may cause a deadlock. syzbot reports: -> #2 (&htab->buckets[i].lock#2){....}: __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159 htab_lru_map_delete_node+0xce/0x2f0 kernel/bpf/hashtab.c:593 __bpf_lru_list_shrink_inactive kernel/bpf/bpf_lru_list.c:220 [inline] __bpf_lru_list_shrink+0xf9/0x470 kernel/bpf/bpf_lru_list.c:266 bpf_lru_list_pop_free_to_local kernel/bpf/bpf_lru_list.c:340 [inline] bpf_common_lru_pop_free kernel/bpf/bpf_lru_list.c:447 [inline] bpf_lru_pop_free+0x87c/0x1670 kernel/bpf/bpf_lru_list.c:499 prealloc_lru_pop+0x2c/0xa0 kernel/bpf/hashtab.c:132 __htab_lru_percpu_map_update_elem+0x67e/0xa90 kernel/bpf/hashtab.c:1069 bpf_percpu_hash_update+0x16e/0x210 kernel/bpf/hashtab.c:1585 bpf_map_update_value.isra.0+0x2d7/0x8e0 kernel/bpf/syscall.c:181 generic_map_update_batch+0x41f/0x610 kernel/bpf/syscall.c:1319 bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348 __do_sys_bpf+0x9b7/0x41e0 kernel/bpf/syscall.c:3460 __se_sys_bpf kernel/bpf/syscall.c:3355 [inline] __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355 do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294 entry_SYSCALL_64_after_hwframe+0x49/0xbe -> #0 (&loc_l->lock){....}: check_prev_add kernel/locking/lockdep.c:2475 [inline] check_prevs_add kernel/locking/lockdep.c:2580 [inline] validate_chain kernel/locking/lockdep.c:2970 [inline] __lock_acquire+0x2596/0x4a00 kernel/locking/lockdep.c:3954 lock_acquire+0x190/0x410 kernel/locking/lockdep.c:4484 __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159 bpf_common_lru_push_free kernel/bpf/bpf_lru_list.c:516 [inline] bpf_lru_push_free+0x250/0x5b0 kernel/bpf/bpf_lru_list.c:555 __htab_map_lookup_and_delete_batch+0x8d4/0x1540 kernel/bpf/hashtab.c:1374 htab_lru_map_lookup_and_delete_batch+0x34/0x40 kernel/bpf/hashtab.c:1491 bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348 __do_sys_bpf+0x1f7d/0x41e0 kernel/bpf/syscall.c:3456 __se_sys_bpf kernel/bpf/syscall.c:3355 [inline] __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355 do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294 entry_SYSCALL_64_after_hwframe+0x49/0xbe Possible unsafe locking scenario: CPU0 CPU2 ---- ---- lock(&htab->buckets[i].lock#2); lock(&l->lock); lock(&htab->buckets[i].lock#2); lock(&loc_l->lock); *** DEADLOCK *** To fix the issue, for htab_lru_map_lookup_and_delete_batch() in CPU0, let us do bpf_lru_push_free() out of the htab bucket lock. This can avoid the above deadlock scenario. Fixes: 057996380a42 ("bpf: Add batch ops to all htab bpf map") Reported-by: syzbot+a38ff3d9356388f2fb83@syzkaller.appspotmail.com Reported-by: syzbot+122b5421d14e68f29cd1@syzkaller.appspotmail.com Suggested-by: Hillf Danton <hdanton@sina.com> Suggested-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com> Acked-by: Brian Vazquez <brianvv@google.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Link: https://lore.kernel.org/bpf/20200219234757.3544014-1-yhs@fb.com
2020-02-19 23:47:57 +00:00
/* bpf_lru_push_free() will acquire lru_lock, which
* may cause deadlock. See comments in function
* prealloc_lru_pop(). Let us do bpf_lru_push_free()
* after releasing the bucket lock.
*/
if (is_lru_map) {
l->batch_flink = node_to_free;
node_to_free = l;
} else {
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
free_htab_elem(htab, l);
bpf: Fix a potential deadlock with bpf_map_do_batch Commit 057996380a42 ("bpf: Add batch ops to all htab bpf map") added lookup_and_delete batch operation for hash table. The current implementation has bpf_lru_push_free() inside the bucket lock, which may cause a deadlock. syzbot reports: -> #2 (&htab->buckets[i].lock#2){....}: __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159 htab_lru_map_delete_node+0xce/0x2f0 kernel/bpf/hashtab.c:593 __bpf_lru_list_shrink_inactive kernel/bpf/bpf_lru_list.c:220 [inline] __bpf_lru_list_shrink+0xf9/0x470 kernel/bpf/bpf_lru_list.c:266 bpf_lru_list_pop_free_to_local kernel/bpf/bpf_lru_list.c:340 [inline] bpf_common_lru_pop_free kernel/bpf/bpf_lru_list.c:447 [inline] bpf_lru_pop_free+0x87c/0x1670 kernel/bpf/bpf_lru_list.c:499 prealloc_lru_pop+0x2c/0xa0 kernel/bpf/hashtab.c:132 __htab_lru_percpu_map_update_elem+0x67e/0xa90 kernel/bpf/hashtab.c:1069 bpf_percpu_hash_update+0x16e/0x210 kernel/bpf/hashtab.c:1585 bpf_map_update_value.isra.0+0x2d7/0x8e0 kernel/bpf/syscall.c:181 generic_map_update_batch+0x41f/0x610 kernel/bpf/syscall.c:1319 bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348 __do_sys_bpf+0x9b7/0x41e0 kernel/bpf/syscall.c:3460 __se_sys_bpf kernel/bpf/syscall.c:3355 [inline] __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355 do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294 entry_SYSCALL_64_after_hwframe+0x49/0xbe -> #0 (&loc_l->lock){....}: check_prev_add kernel/locking/lockdep.c:2475 [inline] check_prevs_add kernel/locking/lockdep.c:2580 [inline] validate_chain kernel/locking/lockdep.c:2970 [inline] __lock_acquire+0x2596/0x4a00 kernel/locking/lockdep.c:3954 lock_acquire+0x190/0x410 kernel/locking/lockdep.c:4484 __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159 bpf_common_lru_push_free kernel/bpf/bpf_lru_list.c:516 [inline] bpf_lru_push_free+0x250/0x5b0 kernel/bpf/bpf_lru_list.c:555 __htab_map_lookup_and_delete_batch+0x8d4/0x1540 kernel/bpf/hashtab.c:1374 htab_lru_map_lookup_and_delete_batch+0x34/0x40 kernel/bpf/hashtab.c:1491 bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348 __do_sys_bpf+0x1f7d/0x41e0 kernel/bpf/syscall.c:3456 __se_sys_bpf kernel/bpf/syscall.c:3355 [inline] __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355 do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294 entry_SYSCALL_64_after_hwframe+0x49/0xbe Possible unsafe locking scenario: CPU0 CPU2 ---- ---- lock(&htab->buckets[i].lock#2); lock(&l->lock); lock(&htab->buckets[i].lock#2); lock(&loc_l->lock); *** DEADLOCK *** To fix the issue, for htab_lru_map_lookup_and_delete_batch() in CPU0, let us do bpf_lru_push_free() out of the htab bucket lock. This can avoid the above deadlock scenario. Fixes: 057996380a42 ("bpf: Add batch ops to all htab bpf map") Reported-by: syzbot+a38ff3d9356388f2fb83@syzkaller.appspotmail.com Reported-by: syzbot+122b5421d14e68f29cd1@syzkaller.appspotmail.com Suggested-by: Hillf Danton <hdanton@sina.com> Suggested-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com> Acked-by: Brian Vazquez <brianvv@google.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Link: https://lore.kernel.org/bpf/20200219234757.3544014-1-yhs@fb.com
2020-02-19 23:47:57 +00:00
}
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
}
dst_key += key_size;
dst_val += value_size;
}
htab_unlock_bucket(htab, b, batch, flags);
locked = false;
bpf: Fix a potential deadlock with bpf_map_do_batch Commit 057996380a42 ("bpf: Add batch ops to all htab bpf map") added lookup_and_delete batch operation for hash table. The current implementation has bpf_lru_push_free() inside the bucket lock, which may cause a deadlock. syzbot reports: -> #2 (&htab->buckets[i].lock#2){....}: __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159 htab_lru_map_delete_node+0xce/0x2f0 kernel/bpf/hashtab.c:593 __bpf_lru_list_shrink_inactive kernel/bpf/bpf_lru_list.c:220 [inline] __bpf_lru_list_shrink+0xf9/0x470 kernel/bpf/bpf_lru_list.c:266 bpf_lru_list_pop_free_to_local kernel/bpf/bpf_lru_list.c:340 [inline] bpf_common_lru_pop_free kernel/bpf/bpf_lru_list.c:447 [inline] bpf_lru_pop_free+0x87c/0x1670 kernel/bpf/bpf_lru_list.c:499 prealloc_lru_pop+0x2c/0xa0 kernel/bpf/hashtab.c:132 __htab_lru_percpu_map_update_elem+0x67e/0xa90 kernel/bpf/hashtab.c:1069 bpf_percpu_hash_update+0x16e/0x210 kernel/bpf/hashtab.c:1585 bpf_map_update_value.isra.0+0x2d7/0x8e0 kernel/bpf/syscall.c:181 generic_map_update_batch+0x41f/0x610 kernel/bpf/syscall.c:1319 bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348 __do_sys_bpf+0x9b7/0x41e0 kernel/bpf/syscall.c:3460 __se_sys_bpf kernel/bpf/syscall.c:3355 [inline] __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355 do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294 entry_SYSCALL_64_after_hwframe+0x49/0xbe -> #0 (&loc_l->lock){....}: check_prev_add kernel/locking/lockdep.c:2475 [inline] check_prevs_add kernel/locking/lockdep.c:2580 [inline] validate_chain kernel/locking/lockdep.c:2970 [inline] __lock_acquire+0x2596/0x4a00 kernel/locking/lockdep.c:3954 lock_acquire+0x190/0x410 kernel/locking/lockdep.c:4484 __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159 bpf_common_lru_push_free kernel/bpf/bpf_lru_list.c:516 [inline] bpf_lru_push_free+0x250/0x5b0 kernel/bpf/bpf_lru_list.c:555 __htab_map_lookup_and_delete_batch+0x8d4/0x1540 kernel/bpf/hashtab.c:1374 htab_lru_map_lookup_and_delete_batch+0x34/0x40 kernel/bpf/hashtab.c:1491 bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348 __do_sys_bpf+0x1f7d/0x41e0 kernel/bpf/syscall.c:3456 __se_sys_bpf kernel/bpf/syscall.c:3355 [inline] __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355 do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294 entry_SYSCALL_64_after_hwframe+0x49/0xbe Possible unsafe locking scenario: CPU0 CPU2 ---- ---- lock(&htab->buckets[i].lock#2); lock(&l->lock); lock(&htab->buckets[i].lock#2); lock(&loc_l->lock); *** DEADLOCK *** To fix the issue, for htab_lru_map_lookup_and_delete_batch() in CPU0, let us do bpf_lru_push_free() out of the htab bucket lock. This can avoid the above deadlock scenario. Fixes: 057996380a42 ("bpf: Add batch ops to all htab bpf map") Reported-by: syzbot+a38ff3d9356388f2fb83@syzkaller.appspotmail.com Reported-by: syzbot+122b5421d14e68f29cd1@syzkaller.appspotmail.com Suggested-by: Hillf Danton <hdanton@sina.com> Suggested-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com> Acked-by: Brian Vazquez <brianvv@google.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Link: https://lore.kernel.org/bpf/20200219234757.3544014-1-yhs@fb.com
2020-02-19 23:47:57 +00:00
while (node_to_free) {
l = node_to_free;
node_to_free = node_to_free->batch_flink;
bpf: Add map side support for bpf timers. Restrict bpf timers to array, hash (both preallocated and kmalloced), and lru map types. The per-cpu maps with timers don't make sense, since 'struct bpf_timer' is a part of map value. bpf timers in per-cpu maps would mean that the number of timers depends on number of possible cpus and timers would not be accessible from all cpus. lpm map support can be added in the future. The timers in inner maps are supported. The bpf_map_update/delete_elem() helpers and sys_bpf commands cancel and free bpf_timer in a given map element. Similar to 'struct bpf_spin_lock' BTF is required and it is used to validate that map element indeed contains 'struct bpf_timer'. Make check_and_init_map_value() init both bpf_spin_lock and bpf_timer when map element data is reused in preallocated htab and lru maps. Teach copy_map_value() to support both bpf_spin_lock and bpf_timer in a single map element. There could be one of each, but not more than one. Due to 'one bpf_timer in one element' restriction do not support timers in global data, since global data is a map of single element, but from bpf program side it's seen as many global variables and restriction of single global timer would be odd. The sys_bpf map_freeze and sys_mmap syscalls are not allowed on maps with timers, since user space could have corrupted mmap element and crashed the kernel. The maps with timers cannot be readonly. Due to these restrictions search for bpf_timer in datasec BTF in case it was placed in the global data to report clear error. The previous patch allowed 'struct bpf_timer' as a first field in a map element only. Relax this restriction. Refactor lru map to s/bpf_lru_push_free/htab_lru_push_free/ to cancel and free the timer when lru map deletes an element as a part of it eviction algorithm. Make sure that bpf program cannot access 'struct bpf_timer' via direct load/store. The timer operation are done through helpers only. This is similar to 'struct bpf_spin_lock'. Signed-off-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Yonghong Song <yhs@fb.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: Andrii Nakryiko <andrii@kernel.org> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Link: https://lore.kernel.org/bpf/20210715005417.78572-5-alexei.starovoitov@gmail.com
2021-07-15 00:54:10 +00:00
htab_lru_push_free(htab, l);
bpf: Fix a potential deadlock with bpf_map_do_batch Commit 057996380a42 ("bpf: Add batch ops to all htab bpf map") added lookup_and_delete batch operation for hash table. The current implementation has bpf_lru_push_free() inside the bucket lock, which may cause a deadlock. syzbot reports: -> #2 (&htab->buckets[i].lock#2){....}: __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159 htab_lru_map_delete_node+0xce/0x2f0 kernel/bpf/hashtab.c:593 __bpf_lru_list_shrink_inactive kernel/bpf/bpf_lru_list.c:220 [inline] __bpf_lru_list_shrink+0xf9/0x470 kernel/bpf/bpf_lru_list.c:266 bpf_lru_list_pop_free_to_local kernel/bpf/bpf_lru_list.c:340 [inline] bpf_common_lru_pop_free kernel/bpf/bpf_lru_list.c:447 [inline] bpf_lru_pop_free+0x87c/0x1670 kernel/bpf/bpf_lru_list.c:499 prealloc_lru_pop+0x2c/0xa0 kernel/bpf/hashtab.c:132 __htab_lru_percpu_map_update_elem+0x67e/0xa90 kernel/bpf/hashtab.c:1069 bpf_percpu_hash_update+0x16e/0x210 kernel/bpf/hashtab.c:1585 bpf_map_update_value.isra.0+0x2d7/0x8e0 kernel/bpf/syscall.c:181 generic_map_update_batch+0x41f/0x610 kernel/bpf/syscall.c:1319 bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348 __do_sys_bpf+0x9b7/0x41e0 kernel/bpf/syscall.c:3460 __se_sys_bpf kernel/bpf/syscall.c:3355 [inline] __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355 do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294 entry_SYSCALL_64_after_hwframe+0x49/0xbe -> #0 (&loc_l->lock){....}: check_prev_add kernel/locking/lockdep.c:2475 [inline] check_prevs_add kernel/locking/lockdep.c:2580 [inline] validate_chain kernel/locking/lockdep.c:2970 [inline] __lock_acquire+0x2596/0x4a00 kernel/locking/lockdep.c:3954 lock_acquire+0x190/0x410 kernel/locking/lockdep.c:4484 __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:159 bpf_common_lru_push_free kernel/bpf/bpf_lru_list.c:516 [inline] bpf_lru_push_free+0x250/0x5b0 kernel/bpf/bpf_lru_list.c:555 __htab_map_lookup_and_delete_batch+0x8d4/0x1540 kernel/bpf/hashtab.c:1374 htab_lru_map_lookup_and_delete_batch+0x34/0x40 kernel/bpf/hashtab.c:1491 bpf_map_do_batch+0x3f5/0x510 kernel/bpf/syscall.c:3348 __do_sys_bpf+0x1f7d/0x41e0 kernel/bpf/syscall.c:3456 __se_sys_bpf kernel/bpf/syscall.c:3355 [inline] __x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:3355 do_syscall_64+0xfa/0x790 arch/x86/entry/common.c:294 entry_SYSCALL_64_after_hwframe+0x49/0xbe Possible unsafe locking scenario: CPU0 CPU2 ---- ---- lock(&htab->buckets[i].lock#2); lock(&l->lock); lock(&htab->buckets[i].lock#2); lock(&loc_l->lock); *** DEADLOCK *** To fix the issue, for htab_lru_map_lookup_and_delete_batch() in CPU0, let us do bpf_lru_push_free() out of the htab bucket lock. This can avoid the above deadlock scenario. Fixes: 057996380a42 ("bpf: Add batch ops to all htab bpf map") Reported-by: syzbot+a38ff3d9356388f2fb83@syzkaller.appspotmail.com Reported-by: syzbot+122b5421d14e68f29cd1@syzkaller.appspotmail.com Suggested-by: Hillf Danton <hdanton@sina.com> Suggested-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com> Acked-by: Brian Vazquez <brianvv@google.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Link: https://lore.kernel.org/bpf/20200219234757.3544014-1-yhs@fb.com
2020-02-19 23:47:57 +00:00
}
next_batch:
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
/* If we are not copying data, we can go to next bucket and avoid
* unlocking the rcu.
*/
if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
batch++;
goto again_nocopy;
}
rcu_read_unlock();
bpf_enable_instrumentation();
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
key_size * bucket_cnt) ||
copy_to_user(uvalues + total * value_size, values,
value_size * bucket_cnt))) {
ret = -EFAULT;
goto after_loop;
}
total += bucket_cnt;
batch++;
if (batch >= htab->n_buckets) {
ret = -ENOENT;
goto after_loop;
}
goto again;
after_loop:
if (ret == -EFAULT)
goto out;
/* copy # of entries and next batch */
ubatch = u64_to_user_ptr(attr->batch.out_batch);
if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
put_user(total, &uattr->batch.count))
ret = -EFAULT;
out:
kvfree(keys);
kvfree(values);
return ret;
}
static int
htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
union bpf_attr __user *uattr)
{
return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
false, true);
}
static int
htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
const union bpf_attr *attr,
union bpf_attr __user *uattr)
{
return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
false, true);
}
static int
htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
union bpf_attr __user *uattr)
{
return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
false, false);
}
static int
htab_map_lookup_and_delete_batch(struct bpf_map *map,
const union bpf_attr *attr,
union bpf_attr __user *uattr)
{
return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
false, false);
}
static int
htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
const union bpf_attr *attr,
union bpf_attr __user *uattr)
{
return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
true, true);
}
static int
htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
const union bpf_attr *attr,
union bpf_attr __user *uattr)
{
return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
true, true);
}
static int
htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
union bpf_attr __user *uattr)
{
return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
true, false);
}
static int
htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
const union bpf_attr *attr,
union bpf_attr __user *uattr)
{
return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
true, false);
}
struct bpf_iter_seq_hash_map_info {
struct bpf_map *map;
struct bpf_htab *htab;
void *percpu_value_buf; // non-zero means percpu hash
u32 bucket_id;
u32 skip_elems;
};
static struct htab_elem *
bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
struct htab_elem *prev_elem)
{
const struct bpf_htab *htab = info->htab;
u32 skip_elems = info->skip_elems;
u32 bucket_id = info->bucket_id;
struct hlist_nulls_head *head;
struct hlist_nulls_node *n;
struct htab_elem *elem;
struct bucket *b;
u32 i, count;
if (bucket_id >= htab->n_buckets)
return NULL;
/* try to find next elem in the same bucket */
if (prev_elem) {
/* no update/deletion on this bucket, prev_elem should be still valid
* and we won't skip elements.
*/
n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
if (elem)
return elem;
/* not found, unlock and go to the next bucket */
b = &htab->buckets[bucket_id++];
bpf: Do not use bucket_lock for hashmap iterator Currently, for hashmap, the bpf iterator will grab a bucket lock, a spinlock, before traversing the elements in the bucket. This can ensure all bpf visted elements are valid. But this mechanism may cause deadlock if update/deletion happens to the same bucket of the visited map in the program. For example, if we added bpf_map_update_elem() call to the same visited element in selftests bpf_iter_bpf_hash_map.c, we will have the following deadlock: ============================================ WARNING: possible recursive locking detected 5.9.0-rc1+ #841 Not tainted -------------------------------------------- test_progs/1750 is trying to acquire lock: ffff9a5bb73c5e70 (&htab->buckets[i].raw_lock){....}-{2:2}, at: htab_map_update_elem+0x1cf/0x410 but task is already holding lock: ffff9a5bb73c5e20 (&htab->buckets[i].raw_lock){....}-{2:2}, at: bpf_hash_map_seq_find_next+0x94/0x120 other info that might help us debug this: Possible unsafe locking scenario: CPU0 ---- lock(&htab->buckets[i].raw_lock); lock(&htab->buckets[i].raw_lock); *** DEADLOCK *** ... Call Trace: dump_stack+0x78/0xa0 __lock_acquire.cold.74+0x209/0x2e3 lock_acquire+0xba/0x380 ? htab_map_update_elem+0x1cf/0x410 ? __lock_acquire+0x639/0x20c0 _raw_spin_lock_irqsave+0x3b/0x80 ? htab_map_update_elem+0x1cf/0x410 htab_map_update_elem+0x1cf/0x410 ? lock_acquire+0xba/0x380 bpf_prog_ad6dab10433b135d_dump_bpf_hash_map+0x88/0xa9c ? find_held_lock+0x34/0xa0 bpf_iter_run_prog+0x81/0x16e __bpf_hash_map_seq_show+0x145/0x180 bpf_seq_read+0xff/0x3d0 vfs_read+0xad/0x1c0 ksys_read+0x5f/0xe0 do_syscall_64+0x33/0x40 entry_SYSCALL_64_after_hwframe+0x44/0xa9 ... The bucket_lock first grabbed in seq_ops->next() called by bpf_seq_read(), and then grabbed again in htab_map_update_elem() in the bpf program, causing deadlocks. Actually, we do not need bucket_lock here, we can just use rcu_read_lock() similar to netlink iterator where the rcu_read_{lock,unlock} likes below: seq_ops->start(): rcu_read_lock(); seq_ops->next(): rcu_read_unlock(); /* next element */ rcu_read_lock(); seq_ops->stop(); rcu_read_unlock(); Compared to old bucket_lock mechanism, if concurrent updata/delete happens, we may visit stale elements, miss some elements, or repeat some elements. I think this is a reasonable compromise. For users wanting to avoid stale, missing/repeated accesses, bpf_map batch access syscall interface can be used. Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200902235340.2001375-1-yhs@fb.com
2020-09-02 23:53:40 +00:00
rcu_read_unlock();
skip_elems = 0;
}
for (i = bucket_id; i < htab->n_buckets; i++) {
b = &htab->buckets[i];
bpf: Do not use bucket_lock for hashmap iterator Currently, for hashmap, the bpf iterator will grab a bucket lock, a spinlock, before traversing the elements in the bucket. This can ensure all bpf visted elements are valid. But this mechanism may cause deadlock if update/deletion happens to the same bucket of the visited map in the program. For example, if we added bpf_map_update_elem() call to the same visited element in selftests bpf_iter_bpf_hash_map.c, we will have the following deadlock: ============================================ WARNING: possible recursive locking detected 5.9.0-rc1+ #841 Not tainted -------------------------------------------- test_progs/1750 is trying to acquire lock: ffff9a5bb73c5e70 (&htab->buckets[i].raw_lock){....}-{2:2}, at: htab_map_update_elem+0x1cf/0x410 but task is already holding lock: ffff9a5bb73c5e20 (&htab->buckets[i].raw_lock){....}-{2:2}, at: bpf_hash_map_seq_find_next+0x94/0x120 other info that might help us debug this: Possible unsafe locking scenario: CPU0 ---- lock(&htab->buckets[i].raw_lock); lock(&htab->buckets[i].raw_lock); *** DEADLOCK *** ... Call Trace: dump_stack+0x78/0xa0 __lock_acquire.cold.74+0x209/0x2e3 lock_acquire+0xba/0x380 ? htab_map_update_elem+0x1cf/0x410 ? __lock_acquire+0x639/0x20c0 _raw_spin_lock_irqsave+0x3b/0x80 ? htab_map_update_elem+0x1cf/0x410 htab_map_update_elem+0x1cf/0x410 ? lock_acquire+0xba/0x380 bpf_prog_ad6dab10433b135d_dump_bpf_hash_map+0x88/0xa9c ? find_held_lock+0x34/0xa0 bpf_iter_run_prog+0x81/0x16e __bpf_hash_map_seq_show+0x145/0x180 bpf_seq_read+0xff/0x3d0 vfs_read+0xad/0x1c0 ksys_read+0x5f/0xe0 do_syscall_64+0x33/0x40 entry_SYSCALL_64_after_hwframe+0x44/0xa9 ... The bucket_lock first grabbed in seq_ops->next() called by bpf_seq_read(), and then grabbed again in htab_map_update_elem() in the bpf program, causing deadlocks. Actually, we do not need bucket_lock here, we can just use rcu_read_lock() similar to netlink iterator where the rcu_read_{lock,unlock} likes below: seq_ops->start(): rcu_read_lock(); seq_ops->next(): rcu_read_unlock(); /* next element */ rcu_read_lock(); seq_ops->stop(); rcu_read_unlock(); Compared to old bucket_lock mechanism, if concurrent updata/delete happens, we may visit stale elements, miss some elements, or repeat some elements. I think this is a reasonable compromise. For users wanting to avoid stale, missing/repeated accesses, bpf_map batch access syscall interface can be used. Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200902235340.2001375-1-yhs@fb.com
2020-09-02 23:53:40 +00:00
rcu_read_lock();
count = 0;
head = &b->head;
hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
if (count >= skip_elems) {
info->bucket_id = i;
info->skip_elems = count;
return elem;
}
count++;
}
bpf: Do not use bucket_lock for hashmap iterator Currently, for hashmap, the bpf iterator will grab a bucket lock, a spinlock, before traversing the elements in the bucket. This can ensure all bpf visted elements are valid. But this mechanism may cause deadlock if update/deletion happens to the same bucket of the visited map in the program. For example, if we added bpf_map_update_elem() call to the same visited element in selftests bpf_iter_bpf_hash_map.c, we will have the following deadlock: ============================================ WARNING: possible recursive locking detected 5.9.0-rc1+ #841 Not tainted -------------------------------------------- test_progs/1750 is trying to acquire lock: ffff9a5bb73c5e70 (&htab->buckets[i].raw_lock){....}-{2:2}, at: htab_map_update_elem+0x1cf/0x410 but task is already holding lock: ffff9a5bb73c5e20 (&htab->buckets[i].raw_lock){....}-{2:2}, at: bpf_hash_map_seq_find_next+0x94/0x120 other info that might help us debug this: Possible unsafe locking scenario: CPU0 ---- lock(&htab->buckets[i].raw_lock); lock(&htab->buckets[i].raw_lock); *** DEADLOCK *** ... Call Trace: dump_stack+0x78/0xa0 __lock_acquire.cold.74+0x209/0x2e3 lock_acquire+0xba/0x380 ? htab_map_update_elem+0x1cf/0x410 ? __lock_acquire+0x639/0x20c0 _raw_spin_lock_irqsave+0x3b/0x80 ? htab_map_update_elem+0x1cf/0x410 htab_map_update_elem+0x1cf/0x410 ? lock_acquire+0xba/0x380 bpf_prog_ad6dab10433b135d_dump_bpf_hash_map+0x88/0xa9c ? find_held_lock+0x34/0xa0 bpf_iter_run_prog+0x81/0x16e __bpf_hash_map_seq_show+0x145/0x180 bpf_seq_read+0xff/0x3d0 vfs_read+0xad/0x1c0 ksys_read+0x5f/0xe0 do_syscall_64+0x33/0x40 entry_SYSCALL_64_after_hwframe+0x44/0xa9 ... The bucket_lock first grabbed in seq_ops->next() called by bpf_seq_read(), and then grabbed again in htab_map_update_elem() in the bpf program, causing deadlocks. Actually, we do not need bucket_lock here, we can just use rcu_read_lock() similar to netlink iterator where the rcu_read_{lock,unlock} likes below: seq_ops->start(): rcu_read_lock(); seq_ops->next(): rcu_read_unlock(); /* next element */ rcu_read_lock(); seq_ops->stop(); rcu_read_unlock(); Compared to old bucket_lock mechanism, if concurrent updata/delete happens, we may visit stale elements, miss some elements, or repeat some elements. I think this is a reasonable compromise. For users wanting to avoid stale, missing/repeated accesses, bpf_map batch access syscall interface can be used. Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200902235340.2001375-1-yhs@fb.com
2020-09-02 23:53:40 +00:00
rcu_read_unlock();
skip_elems = 0;
}
info->bucket_id = i;
info->skip_elems = 0;
return NULL;
}
static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
{
struct bpf_iter_seq_hash_map_info *info = seq->private;
struct htab_elem *elem;
elem = bpf_hash_map_seq_find_next(info, NULL);
if (!elem)
return NULL;
if (*pos == 0)
++*pos;
return elem;
}
static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
{
struct bpf_iter_seq_hash_map_info *info = seq->private;
++*pos;
++info->skip_elems;
return bpf_hash_map_seq_find_next(info, v);
}
static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
{
struct bpf_iter_seq_hash_map_info *info = seq->private;
u32 roundup_key_size, roundup_value_size;
struct bpf_iter__bpf_map_elem ctx = {};
struct bpf_map *map = info->map;
struct bpf_iter_meta meta;
int ret = 0, off = 0, cpu;
struct bpf_prog *prog;
void __percpu *pptr;
meta.seq = seq;
prog = bpf_iter_get_info(&meta, elem == NULL);
if (prog) {
ctx.meta = &meta;
ctx.map = info->map;
if (elem) {
roundup_key_size = round_up(map->key_size, 8);
ctx.key = elem->key;
if (!info->percpu_value_buf) {
ctx.value = elem->key + roundup_key_size;
} else {
roundup_value_size = round_up(map->value_size, 8);
pptr = htab_elem_get_ptr(elem, map->key_size);
for_each_possible_cpu(cpu) {
copy_map_value_long(map, info->percpu_value_buf + off,
per_cpu_ptr(pptr, cpu));
check_and_init_map_value(map, info->percpu_value_buf + off);
off += roundup_value_size;
}
ctx.value = info->percpu_value_buf;
}
}
ret = bpf_iter_run_prog(prog, &ctx);
}
return ret;
}
static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
{
return __bpf_hash_map_seq_show(seq, v);
}
static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
{
if (!v)
(void)__bpf_hash_map_seq_show(seq, NULL);
else
bpf: Do not use bucket_lock for hashmap iterator Currently, for hashmap, the bpf iterator will grab a bucket lock, a spinlock, before traversing the elements in the bucket. This can ensure all bpf visted elements are valid. But this mechanism may cause deadlock if update/deletion happens to the same bucket of the visited map in the program. For example, if we added bpf_map_update_elem() call to the same visited element in selftests bpf_iter_bpf_hash_map.c, we will have the following deadlock: ============================================ WARNING: possible recursive locking detected 5.9.0-rc1+ #841 Not tainted -------------------------------------------- test_progs/1750 is trying to acquire lock: ffff9a5bb73c5e70 (&htab->buckets[i].raw_lock){....}-{2:2}, at: htab_map_update_elem+0x1cf/0x410 but task is already holding lock: ffff9a5bb73c5e20 (&htab->buckets[i].raw_lock){....}-{2:2}, at: bpf_hash_map_seq_find_next+0x94/0x120 other info that might help us debug this: Possible unsafe locking scenario: CPU0 ---- lock(&htab->buckets[i].raw_lock); lock(&htab->buckets[i].raw_lock); *** DEADLOCK *** ... Call Trace: dump_stack+0x78/0xa0 __lock_acquire.cold.74+0x209/0x2e3 lock_acquire+0xba/0x380 ? htab_map_update_elem+0x1cf/0x410 ? __lock_acquire+0x639/0x20c0 _raw_spin_lock_irqsave+0x3b/0x80 ? htab_map_update_elem+0x1cf/0x410 htab_map_update_elem+0x1cf/0x410 ? lock_acquire+0xba/0x380 bpf_prog_ad6dab10433b135d_dump_bpf_hash_map+0x88/0xa9c ? find_held_lock+0x34/0xa0 bpf_iter_run_prog+0x81/0x16e __bpf_hash_map_seq_show+0x145/0x180 bpf_seq_read+0xff/0x3d0 vfs_read+0xad/0x1c0 ksys_read+0x5f/0xe0 do_syscall_64+0x33/0x40 entry_SYSCALL_64_after_hwframe+0x44/0xa9 ... The bucket_lock first grabbed in seq_ops->next() called by bpf_seq_read(), and then grabbed again in htab_map_update_elem() in the bpf program, causing deadlocks. Actually, we do not need bucket_lock here, we can just use rcu_read_lock() similar to netlink iterator where the rcu_read_{lock,unlock} likes below: seq_ops->start(): rcu_read_lock(); seq_ops->next(): rcu_read_unlock(); /* next element */ rcu_read_lock(); seq_ops->stop(); rcu_read_unlock(); Compared to old bucket_lock mechanism, if concurrent updata/delete happens, we may visit stale elements, miss some elements, or repeat some elements. I think this is a reasonable compromise. For users wanting to avoid stale, missing/repeated accesses, bpf_map batch access syscall interface can be used. Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200902235340.2001375-1-yhs@fb.com
2020-09-02 23:53:40 +00:00
rcu_read_unlock();
}
static int bpf_iter_init_hash_map(void *priv_data,
struct bpf_iter_aux_info *aux)
{
struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
struct bpf_map *map = aux->map;
void *value_buf;
u32 buf_size;
if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
buf_size = round_up(map->value_size, 8) * num_possible_cpus();
value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
if (!value_buf)
return -ENOMEM;
seq_info->percpu_value_buf = value_buf;
}
bpf_map_inc_with_uref(map);
seq_info->map = map;
seq_info->htab = container_of(map, struct bpf_htab, map);
return 0;
}
static void bpf_iter_fini_hash_map(void *priv_data)
{
struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
bpf_map_put_with_uref(seq_info->map);
kfree(seq_info->percpu_value_buf);
}
static const struct seq_operations bpf_hash_map_seq_ops = {
.start = bpf_hash_map_seq_start,
.next = bpf_hash_map_seq_next,
.stop = bpf_hash_map_seq_stop,
.show = bpf_hash_map_seq_show,
};
static const struct bpf_iter_seq_info iter_seq_info = {
.seq_ops = &bpf_hash_map_seq_ops,
.init_seq_private = bpf_iter_init_hash_map,
.fini_seq_private = bpf_iter_fini_hash_map,
.seq_priv_size = sizeof(struct bpf_iter_seq_hash_map_info),
};
bpf: return long from bpf_map_ops funcs This patch changes the return types of bpf_map_ops functions to long, where previously int was returned. Using long allows for bpf programs to maintain the sign bit in the absence of sign extension during situations where inlined bpf helper funcs make calls to the bpf_map_ops funcs and a negative error is returned. The definitions of the helper funcs are generated from comments in the bpf uapi header at `include/uapi/linux/bpf.h`. The return type of these helpers was previously changed from int to long in commit bdb7b79b4ce8. For any case where one of the map helpers call the bpf_map_ops funcs that are still returning 32-bit int, a compiler might not include sign extension instructions to properly convert the 32-bit negative value a 64-bit negative value. For example: bpf assembly excerpt of an inlined helper calling a kernel function and checking for a specific error: ; err = bpf_map_update_elem(&mymap, &key, &val, BPF_NOEXIST); ... 46: call 0xffffffffe103291c ; htab_map_update_elem ; if (err && err != -EEXIST) { 4b: cmp $0xffffffffffffffef,%rax ; cmp -EEXIST,%rax kernel function assembly excerpt of return value from `htab_map_update_elem` returning 32-bit int: movl $0xffffffef, %r9d ... movl %r9d, %eax ...results in the comparison: cmp $0xffffffffffffffef, $0x00000000ffffffef Fixes: bdb7b79b4ce8 ("bpf: Switch most helper return values from 32-bit int to 64-bit long") Tested-by: Eduard Zingerman <eddyz87@gmail.com> Signed-off-by: JP Kobryn <inwardvessel@gmail.com> Link: https://lore.kernel.org/r/20230322194754.185781-3-inwardvessel@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-22 19:47:54 +00:00
static long bpf_for_each_hash_elem(struct bpf_map *map, bpf_callback_t callback_fn,
void *callback_ctx, u64 flags)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
struct hlist_nulls_head *head;
struct hlist_nulls_node *n;
struct htab_elem *elem;
u32 roundup_key_size;
int i, num_elems = 0;
void __percpu *pptr;
struct bucket *b;
void *key, *val;
bool is_percpu;
u64 ret = 0;
if (flags != 0)
return -EINVAL;
is_percpu = htab_is_percpu(htab);
roundup_key_size = round_up(map->key_size, 8);
/* disable migration so percpu value prepared here will be the
* same as the one seen by the bpf program with bpf_map_lookup_elem().
*/
if (is_percpu)
migrate_disable();
for (i = 0; i < htab->n_buckets; i++) {
b = &htab->buckets[i];
rcu_read_lock();
head = &b->head;
hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
key = elem->key;
if (is_percpu) {
/* current cpu value for percpu map */
pptr = htab_elem_get_ptr(elem, map->key_size);
val = this_cpu_ptr(pptr);
} else {
val = elem->key + roundup_key_size;
}
num_elems++;
ret = callback_fn((u64)(long)map, (u64)(long)key,
(u64)(long)val, (u64)(long)callback_ctx, 0);
/* return value: 0 - continue, 1 - stop and return */
if (ret) {
rcu_read_unlock();
goto out;
}
}
rcu_read_unlock();
}
out:
if (is_percpu)
migrate_enable();
return num_elems;
}
bpf: hashtab memory usage htab_map_mem_usage() is introduced to calculate hashmap memory usage. In this helper, some small memory allocations are ignore, as their size is quite small compared with the total size. The inner_map_meta in hash_of_map is also ignored. The result for hashtab as follows, - before this change 1: hash name count_map flags 0x1 <<<< no prealloc, fully set key 16B value 24B max_entries 1048576 memlock 41943040B 2: hash name count_map flags 0x1 <<<< no prealloc, none set key 16B value 24B max_entries 1048576 memlock 41943040B 3: hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 41943040B The memlock is always a fixed size whatever it is preallocated or not, and whatever the count of allocated elements is. - after this change 1: hash name count_map flags 0x1 <<<< non prealloc, fully set key 16B value 24B max_entries 1048576 memlock 117441536B 2: hash name count_map flags 0x1 <<<< non prealloc, non set key 16B value 24B max_entries 1048576 memlock 16778240B 3: hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 109056000B The memlock now is hashtab actually allocated. The result for percpu hash map as follows, - before this change 4: percpu_hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 822083584B 5: percpu_hash name count_map flags 0x1 <<<< no prealloc key 16B value 24B max_entries 1048576 memlock 822083584B - after this change 4: percpu_hash name count_map flags 0x0 key 16B value 24B max_entries 1048576 memlock 897582080B 5: percpu_hash name count_map flags 0x1 key 16B value 24B max_entries 1048576 memlock 922748736B At worst, the difference can be 10x, for example, - before this change 6: hash name count_map flags 0x0 key 4B value 4B max_entries 1048576 memlock 8388608B - after this change 6: hash name count_map flags 0x0 key 4B value 4B max_entries 1048576 memlock 83889408B Signed-off-by: Yafang Shao <laoar.shao@gmail.com> Acked-by: Hou Tao <houtao1@huawei.com> Link: https://lore.kernel.org/r/20230305124615.12358-4-laoar.shao@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-05 12:46:00 +00:00
static u64 htab_map_mem_usage(const struct bpf_map *map)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
u32 value_size = round_up(htab->map.value_size, 8);
bool prealloc = htab_is_prealloc(htab);
bool percpu = htab_is_percpu(htab);
bool lru = htab_is_lru(htab);
u64 num_entries;
u64 usage = sizeof(struct bpf_htab);
usage += sizeof(struct bucket) * htab->n_buckets;
usage += sizeof(int) * num_possible_cpus() * HASHTAB_MAP_LOCK_COUNT;
if (prealloc) {
num_entries = map->max_entries;
if (htab_has_extra_elems(htab))
num_entries += num_possible_cpus();
usage += htab->elem_size * num_entries;
if (percpu)
usage += value_size * num_possible_cpus() * num_entries;
else if (!lru)
usage += sizeof(struct htab_elem *) * num_possible_cpus();
} else {
#define LLIST_NODE_SZ sizeof(struct llist_node)
num_entries = htab->use_percpu_counter ?
percpu_counter_sum(&htab->pcount) :
atomic_read(&htab->count);
usage += (htab->elem_size + LLIST_NODE_SZ) * num_entries;
if (percpu) {
usage += (LLIST_NODE_SZ + sizeof(void *)) * num_entries;
usage += value_size * num_possible_cpus() * num_entries;
}
}
return usage;
}
BTF_ID_LIST_SINGLE(htab_map_btf_ids, struct, bpf_htab)
const struct bpf_map_ops htab_map_ops = {
bpf: Add map_meta_equal map ops Some properties of the inner map is used in the verification time. When an inner map is inserted to an outer map at runtime, bpf_map_meta_equal() is currently used to ensure those properties of the inserting inner map stays the same as the verification time. In particular, the current bpf_map_meta_equal() checks max_entries which turns out to be too restrictive for most of the maps which do not use max_entries during the verification time. It limits the use case that wants to replace a smaller inner map with a larger inner map. There are some maps do use max_entries during verification though. For example, the map_gen_lookup in array_map_ops uses the max_entries to generate the inline lookup code. To accommodate differences between maps, the map_meta_equal is added to bpf_map_ops. Each map-type can decide what to check when its map is used as an inner map during runtime. Also, some map types cannot be used as an inner map and they are currently black listed in bpf_map_meta_alloc() in map_in_map.c. It is not unusual that the new map types may not aware that such blacklist exists. This patch enforces an explicit opt-in and only allows a map to be used as an inner map if it has implemented the map_meta_equal ops. It is based on the discussion in [1]. All maps that support inner map has its map_meta_equal points to bpf_map_meta_equal in this patch. A later patch will relax the max_entries check for most maps. bpf_types.h counts 28 map types. This patch adds 23 ".map_meta_equal" by using coccinelle. -5 for BPF_MAP_TYPE_PROG_ARRAY BPF_MAP_TYPE_(PERCPU)_CGROUP_STORAGE BPF_MAP_TYPE_STRUCT_OPS BPF_MAP_TYPE_ARRAY_OF_MAPS BPF_MAP_TYPE_HASH_OF_MAPS The "if (inner_map->inner_map_meta)" check in bpf_map_meta_alloc() is moved such that the same error is returned. [1]: https://lore.kernel.org/bpf/20200522022342.899756-1-kafai@fb.com/ Signed-off-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20200828011806.1970400-1-kafai@fb.com
2020-08-28 01:18:06 +00:00
.map_meta_equal = bpf_map_meta_equal,
.map_alloc_check = htab_map_alloc_check,
.map_alloc = htab_map_alloc,
.map_free = htab_map_free,
.map_get_next_key = htab_map_get_next_key,
.map_release_uref = htab_map_free_timers_and_wq,
.map_lookup_elem = htab_map_lookup_elem,
.map_lookup_and_delete_elem = htab_map_lookup_and_delete_elem,
.map_update_elem = htab_map_update_elem,
.map_delete_elem = htab_map_delete_elem,
.map_gen_lookup = htab_map_gen_lookup,
.map_seq_show_elem = htab_map_seq_show_elem,
.map_set_for_each_callback_args = map_set_for_each_callback_args,
.map_for_each_callback = bpf_for_each_hash_elem,
bpf: hashtab memory usage htab_map_mem_usage() is introduced to calculate hashmap memory usage. In this helper, some small memory allocations are ignore, as their size is quite small compared with the total size. The inner_map_meta in hash_of_map is also ignored. The result for hashtab as follows, - before this change 1: hash name count_map flags 0x1 <<<< no prealloc, fully set key 16B value 24B max_entries 1048576 memlock 41943040B 2: hash name count_map flags 0x1 <<<< no prealloc, none set key 16B value 24B max_entries 1048576 memlock 41943040B 3: hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 41943040B The memlock is always a fixed size whatever it is preallocated or not, and whatever the count of allocated elements is. - after this change 1: hash name count_map flags 0x1 <<<< non prealloc, fully set key 16B value 24B max_entries 1048576 memlock 117441536B 2: hash name count_map flags 0x1 <<<< non prealloc, non set key 16B value 24B max_entries 1048576 memlock 16778240B 3: hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 109056000B The memlock now is hashtab actually allocated. The result for percpu hash map as follows, - before this change 4: percpu_hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 822083584B 5: percpu_hash name count_map flags 0x1 <<<< no prealloc key 16B value 24B max_entries 1048576 memlock 822083584B - after this change 4: percpu_hash name count_map flags 0x0 key 16B value 24B max_entries 1048576 memlock 897582080B 5: percpu_hash name count_map flags 0x1 key 16B value 24B max_entries 1048576 memlock 922748736B At worst, the difference can be 10x, for example, - before this change 6: hash name count_map flags 0x0 key 4B value 4B max_entries 1048576 memlock 8388608B - after this change 6: hash name count_map flags 0x0 key 4B value 4B max_entries 1048576 memlock 83889408B Signed-off-by: Yafang Shao <laoar.shao@gmail.com> Acked-by: Hou Tao <houtao1@huawei.com> Link: https://lore.kernel.org/r/20230305124615.12358-4-laoar.shao@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-05 12:46:00 +00:00
.map_mem_usage = htab_map_mem_usage,
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
BATCH_OPS(htab),
.map_btf_id = &htab_map_btf_ids[0],
.iter_seq_info = &iter_seq_info,
};
const struct bpf_map_ops htab_lru_map_ops = {
bpf: Add map_meta_equal map ops Some properties of the inner map is used in the verification time. When an inner map is inserted to an outer map at runtime, bpf_map_meta_equal() is currently used to ensure those properties of the inserting inner map stays the same as the verification time. In particular, the current bpf_map_meta_equal() checks max_entries which turns out to be too restrictive for most of the maps which do not use max_entries during the verification time. It limits the use case that wants to replace a smaller inner map with a larger inner map. There are some maps do use max_entries during verification though. For example, the map_gen_lookup in array_map_ops uses the max_entries to generate the inline lookup code. To accommodate differences between maps, the map_meta_equal is added to bpf_map_ops. Each map-type can decide what to check when its map is used as an inner map during runtime. Also, some map types cannot be used as an inner map and they are currently black listed in bpf_map_meta_alloc() in map_in_map.c. It is not unusual that the new map types may not aware that such blacklist exists. This patch enforces an explicit opt-in and only allows a map to be used as an inner map if it has implemented the map_meta_equal ops. It is based on the discussion in [1]. All maps that support inner map has its map_meta_equal points to bpf_map_meta_equal in this patch. A later patch will relax the max_entries check for most maps. bpf_types.h counts 28 map types. This patch adds 23 ".map_meta_equal" by using coccinelle. -5 for BPF_MAP_TYPE_PROG_ARRAY BPF_MAP_TYPE_(PERCPU)_CGROUP_STORAGE BPF_MAP_TYPE_STRUCT_OPS BPF_MAP_TYPE_ARRAY_OF_MAPS BPF_MAP_TYPE_HASH_OF_MAPS The "if (inner_map->inner_map_meta)" check in bpf_map_meta_alloc() is moved such that the same error is returned. [1]: https://lore.kernel.org/bpf/20200522022342.899756-1-kafai@fb.com/ Signed-off-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20200828011806.1970400-1-kafai@fb.com
2020-08-28 01:18:06 +00:00
.map_meta_equal = bpf_map_meta_equal,
.map_alloc_check = htab_map_alloc_check,
.map_alloc = htab_map_alloc,
.map_free = htab_map_free,
.map_get_next_key = htab_map_get_next_key,
.map_release_uref = htab_map_free_timers_and_wq,
.map_lookup_elem = htab_lru_map_lookup_elem,
.map_lookup_and_delete_elem = htab_lru_map_lookup_and_delete_elem,
bpf, lru: avoid messing with eviction heuristics upon syscall lookup One of the biggest issues we face right now with picking LRU map over regular hash table is that a map walk out of user space, for example, to just dump the existing entries or to remove certain ones, will completely mess up LRU eviction heuristics and wrong entries such as just created ones will get evicted instead. The reason for this is that we mark an entry as "in use" via bpf_lru_node_set_ref() from system call lookup side as well. Thus upon walk, all entries are being marked, so information of actual least recently used ones are "lost". In case of Cilium where it can be used (besides others) as a BPF based connection tracker, this current behavior causes disruption upon control plane changes that need to walk the map from user space to evict certain entries. Discussion result from bpfconf [0] was that we should simply just remove marking from system call side as no good use case could be found where it's actually needed there. Therefore this patch removes marking for regular LRU and per-CPU flavor. If there ever should be a need in future, the behavior could be selected via map creation flag, but due to mentioned reason we avoid this here. [0] http://vger.kernel.org/bpfconf.html Fixes: 29ba732acbee ("bpf: Add BPF_MAP_TYPE_LRU_HASH") Fixes: 8f8449384ec3 ("bpf: Add BPF_MAP_TYPE_LRU_PERCPU_HASH") Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2019-05-13 23:18:56 +00:00
.map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
.map_update_elem = htab_lru_map_update_elem,
.map_delete_elem = htab_lru_map_delete_elem,
.map_gen_lookup = htab_lru_map_gen_lookup,
.map_seq_show_elem = htab_map_seq_show_elem,
.map_set_for_each_callback_args = map_set_for_each_callback_args,
.map_for_each_callback = bpf_for_each_hash_elem,
bpf: hashtab memory usage htab_map_mem_usage() is introduced to calculate hashmap memory usage. In this helper, some small memory allocations are ignore, as their size is quite small compared with the total size. The inner_map_meta in hash_of_map is also ignored. The result for hashtab as follows, - before this change 1: hash name count_map flags 0x1 <<<< no prealloc, fully set key 16B value 24B max_entries 1048576 memlock 41943040B 2: hash name count_map flags 0x1 <<<< no prealloc, none set key 16B value 24B max_entries 1048576 memlock 41943040B 3: hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 41943040B The memlock is always a fixed size whatever it is preallocated or not, and whatever the count of allocated elements is. - after this change 1: hash name count_map flags 0x1 <<<< non prealloc, fully set key 16B value 24B max_entries 1048576 memlock 117441536B 2: hash name count_map flags 0x1 <<<< non prealloc, non set key 16B value 24B max_entries 1048576 memlock 16778240B 3: hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 109056000B The memlock now is hashtab actually allocated. The result for percpu hash map as follows, - before this change 4: percpu_hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 822083584B 5: percpu_hash name count_map flags 0x1 <<<< no prealloc key 16B value 24B max_entries 1048576 memlock 822083584B - after this change 4: percpu_hash name count_map flags 0x0 key 16B value 24B max_entries 1048576 memlock 897582080B 5: percpu_hash name count_map flags 0x1 key 16B value 24B max_entries 1048576 memlock 922748736B At worst, the difference can be 10x, for example, - before this change 6: hash name count_map flags 0x0 key 4B value 4B max_entries 1048576 memlock 8388608B - after this change 6: hash name count_map flags 0x0 key 4B value 4B max_entries 1048576 memlock 83889408B Signed-off-by: Yafang Shao <laoar.shao@gmail.com> Acked-by: Hou Tao <houtao1@huawei.com> Link: https://lore.kernel.org/r/20230305124615.12358-4-laoar.shao@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-05 12:46:00 +00:00
.map_mem_usage = htab_map_mem_usage,
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
BATCH_OPS(htab_lru),
.map_btf_id = &htab_map_btf_ids[0],
.iter_seq_info = &iter_seq_info,
};
/* Called from eBPF program */
static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
{
struct htab_elem *l = __htab_map_lookup_elem(map, key);
if (l)
return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
else
return NULL;
}
/* inline bpf_map_lookup_elem() call for per-CPU hashmap */
static int htab_percpu_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
{
struct bpf_insn *insn = insn_buf;
if (!bpf_jit_supports_percpu_insn())
return -EOPNOTSUPP;
BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
(void *(*)(struct bpf_map *map, void *key))NULL));
*insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
*insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 3);
*insn++ = BPF_ALU64_IMM(BPF_ADD, BPF_REG_0,
offsetof(struct htab_elem, key) + map->key_size);
*insn++ = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_0, 0);
*insn++ = BPF_MOV64_PERCPU_REG(BPF_REG_0, BPF_REG_0);
return insn - insn_buf;
}
static void *htab_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
{
struct htab_elem *l;
if (cpu >= nr_cpu_ids)
return NULL;
l = __htab_map_lookup_elem(map, key);
if (l)
return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
else
return NULL;
}
static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
{
struct htab_elem *l = __htab_map_lookup_elem(map, key);
if (l) {
bpf_lru_node_set_ref(&l->lru_node);
return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
}
return NULL;
}
static void *htab_lru_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
{
struct htab_elem *l;
if (cpu >= nr_cpu_ids)
return NULL;
l = __htab_map_lookup_elem(map, key);
if (l) {
bpf_lru_node_set_ref(&l->lru_node);
return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
}
return NULL;
}
int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
{
struct htab_elem *l;
void __percpu *pptr;
int ret = -ENOENT;
int cpu, off = 0;
u32 size;
/* per_cpu areas are zero-filled and bpf programs can only
* access 'value_size' of them, so copying rounded areas
* will not leak any kernel data
*/
size = round_up(map->value_size, 8);
rcu_read_lock();
l = __htab_map_lookup_elem(map, key);
if (!l)
goto out;
bpf, lru: avoid messing with eviction heuristics upon syscall lookup One of the biggest issues we face right now with picking LRU map over regular hash table is that a map walk out of user space, for example, to just dump the existing entries or to remove certain ones, will completely mess up LRU eviction heuristics and wrong entries such as just created ones will get evicted instead. The reason for this is that we mark an entry as "in use" via bpf_lru_node_set_ref() from system call lookup side as well. Thus upon walk, all entries are being marked, so information of actual least recently used ones are "lost". In case of Cilium where it can be used (besides others) as a BPF based connection tracker, this current behavior causes disruption upon control plane changes that need to walk the map from user space to evict certain entries. Discussion result from bpfconf [0] was that we should simply just remove marking from system call side as no good use case could be found where it's actually needed there. Therefore this patch removes marking for regular LRU and per-CPU flavor. If there ever should be a need in future, the behavior could be selected via map creation flag, but due to mentioned reason we avoid this here. [0] http://vger.kernel.org/bpfconf.html Fixes: 29ba732acbee ("bpf: Add BPF_MAP_TYPE_LRU_HASH") Fixes: 8f8449384ec3 ("bpf: Add BPF_MAP_TYPE_LRU_PERCPU_HASH") Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2019-05-13 23:18:56 +00:00
/* We do not mark LRU map element here in order to not mess up
* eviction heuristics when user space does a map walk.
*/
pptr = htab_elem_get_ptr(l, map->key_size);
for_each_possible_cpu(cpu) {
copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));
check_and_init_map_value(map, value + off);
off += size;
}
ret = 0;
out:
rcu_read_unlock();
return ret;
}
int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
u64 map_flags)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
int ret;
rcu_read_lock();
if (htab_is_lru(htab))
ret = __htab_lru_percpu_map_update_elem(map, key, value,
map_flags, true);
else
ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
true);
rcu_read_unlock();
return ret;
}
static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
struct seq_file *m)
{
struct htab_elem *l;
void __percpu *pptr;
int cpu;
rcu_read_lock();
l = __htab_map_lookup_elem(map, key);
if (!l) {
rcu_read_unlock();
return;
}
btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
seq_puts(m, ": {\n");
pptr = htab_elem_get_ptr(l, map->key_size);
for_each_possible_cpu(cpu) {
seq_printf(m, "\tcpu%d: ", cpu);
btf_type_seq_show(map->btf, map->btf_value_type_id,
per_cpu_ptr(pptr, cpu), m);
seq_puts(m, "\n");
}
seq_puts(m, "}\n");
rcu_read_unlock();
}
const struct bpf_map_ops htab_percpu_map_ops = {
bpf: Add map_meta_equal map ops Some properties of the inner map is used in the verification time. When an inner map is inserted to an outer map at runtime, bpf_map_meta_equal() is currently used to ensure those properties of the inserting inner map stays the same as the verification time. In particular, the current bpf_map_meta_equal() checks max_entries which turns out to be too restrictive for most of the maps which do not use max_entries during the verification time. It limits the use case that wants to replace a smaller inner map with a larger inner map. There are some maps do use max_entries during verification though. For example, the map_gen_lookup in array_map_ops uses the max_entries to generate the inline lookup code. To accommodate differences between maps, the map_meta_equal is added to bpf_map_ops. Each map-type can decide what to check when its map is used as an inner map during runtime. Also, some map types cannot be used as an inner map and they are currently black listed in bpf_map_meta_alloc() in map_in_map.c. It is not unusual that the new map types may not aware that such blacklist exists. This patch enforces an explicit opt-in and only allows a map to be used as an inner map if it has implemented the map_meta_equal ops. It is based on the discussion in [1]. All maps that support inner map has its map_meta_equal points to bpf_map_meta_equal in this patch. A later patch will relax the max_entries check for most maps. bpf_types.h counts 28 map types. This patch adds 23 ".map_meta_equal" by using coccinelle. -5 for BPF_MAP_TYPE_PROG_ARRAY BPF_MAP_TYPE_(PERCPU)_CGROUP_STORAGE BPF_MAP_TYPE_STRUCT_OPS BPF_MAP_TYPE_ARRAY_OF_MAPS BPF_MAP_TYPE_HASH_OF_MAPS The "if (inner_map->inner_map_meta)" check in bpf_map_meta_alloc() is moved such that the same error is returned. [1]: https://lore.kernel.org/bpf/20200522022342.899756-1-kafai@fb.com/ Signed-off-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20200828011806.1970400-1-kafai@fb.com
2020-08-28 01:18:06 +00:00
.map_meta_equal = bpf_map_meta_equal,
.map_alloc_check = htab_map_alloc_check,
.map_alloc = htab_map_alloc,
.map_free = htab_map_free,
.map_get_next_key = htab_map_get_next_key,
.map_lookup_elem = htab_percpu_map_lookup_elem,
.map_gen_lookup = htab_percpu_map_gen_lookup,
.map_lookup_and_delete_elem = htab_percpu_map_lookup_and_delete_elem,
.map_update_elem = htab_percpu_map_update_elem,
.map_delete_elem = htab_map_delete_elem,
.map_lookup_percpu_elem = htab_percpu_map_lookup_percpu_elem,
.map_seq_show_elem = htab_percpu_map_seq_show_elem,
.map_set_for_each_callback_args = map_set_for_each_callback_args,
.map_for_each_callback = bpf_for_each_hash_elem,
bpf: hashtab memory usage htab_map_mem_usage() is introduced to calculate hashmap memory usage. In this helper, some small memory allocations are ignore, as their size is quite small compared with the total size. The inner_map_meta in hash_of_map is also ignored. The result for hashtab as follows, - before this change 1: hash name count_map flags 0x1 <<<< no prealloc, fully set key 16B value 24B max_entries 1048576 memlock 41943040B 2: hash name count_map flags 0x1 <<<< no prealloc, none set key 16B value 24B max_entries 1048576 memlock 41943040B 3: hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 41943040B The memlock is always a fixed size whatever it is preallocated or not, and whatever the count of allocated elements is. - after this change 1: hash name count_map flags 0x1 <<<< non prealloc, fully set key 16B value 24B max_entries 1048576 memlock 117441536B 2: hash name count_map flags 0x1 <<<< non prealloc, non set key 16B value 24B max_entries 1048576 memlock 16778240B 3: hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 109056000B The memlock now is hashtab actually allocated. The result for percpu hash map as follows, - before this change 4: percpu_hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 822083584B 5: percpu_hash name count_map flags 0x1 <<<< no prealloc key 16B value 24B max_entries 1048576 memlock 822083584B - after this change 4: percpu_hash name count_map flags 0x0 key 16B value 24B max_entries 1048576 memlock 897582080B 5: percpu_hash name count_map flags 0x1 key 16B value 24B max_entries 1048576 memlock 922748736B At worst, the difference can be 10x, for example, - before this change 6: hash name count_map flags 0x0 key 4B value 4B max_entries 1048576 memlock 8388608B - after this change 6: hash name count_map flags 0x0 key 4B value 4B max_entries 1048576 memlock 83889408B Signed-off-by: Yafang Shao <laoar.shao@gmail.com> Acked-by: Hou Tao <houtao1@huawei.com> Link: https://lore.kernel.org/r/20230305124615.12358-4-laoar.shao@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-05 12:46:00 +00:00
.map_mem_usage = htab_map_mem_usage,
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
BATCH_OPS(htab_percpu),
.map_btf_id = &htab_map_btf_ids[0],
.iter_seq_info = &iter_seq_info,
};
const struct bpf_map_ops htab_lru_percpu_map_ops = {
bpf: Add map_meta_equal map ops Some properties of the inner map is used in the verification time. When an inner map is inserted to an outer map at runtime, bpf_map_meta_equal() is currently used to ensure those properties of the inserting inner map stays the same as the verification time. In particular, the current bpf_map_meta_equal() checks max_entries which turns out to be too restrictive for most of the maps which do not use max_entries during the verification time. It limits the use case that wants to replace a smaller inner map with a larger inner map. There are some maps do use max_entries during verification though. For example, the map_gen_lookup in array_map_ops uses the max_entries to generate the inline lookup code. To accommodate differences between maps, the map_meta_equal is added to bpf_map_ops. Each map-type can decide what to check when its map is used as an inner map during runtime. Also, some map types cannot be used as an inner map and they are currently black listed in bpf_map_meta_alloc() in map_in_map.c. It is not unusual that the new map types may not aware that such blacklist exists. This patch enforces an explicit opt-in and only allows a map to be used as an inner map if it has implemented the map_meta_equal ops. It is based on the discussion in [1]. All maps that support inner map has its map_meta_equal points to bpf_map_meta_equal in this patch. A later patch will relax the max_entries check for most maps. bpf_types.h counts 28 map types. This patch adds 23 ".map_meta_equal" by using coccinelle. -5 for BPF_MAP_TYPE_PROG_ARRAY BPF_MAP_TYPE_(PERCPU)_CGROUP_STORAGE BPF_MAP_TYPE_STRUCT_OPS BPF_MAP_TYPE_ARRAY_OF_MAPS BPF_MAP_TYPE_HASH_OF_MAPS The "if (inner_map->inner_map_meta)" check in bpf_map_meta_alloc() is moved such that the same error is returned. [1]: https://lore.kernel.org/bpf/20200522022342.899756-1-kafai@fb.com/ Signed-off-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20200828011806.1970400-1-kafai@fb.com
2020-08-28 01:18:06 +00:00
.map_meta_equal = bpf_map_meta_equal,
.map_alloc_check = htab_map_alloc_check,
.map_alloc = htab_map_alloc,
.map_free = htab_map_free,
.map_get_next_key = htab_map_get_next_key,
.map_lookup_elem = htab_lru_percpu_map_lookup_elem,
.map_lookup_and_delete_elem = htab_lru_percpu_map_lookup_and_delete_elem,
.map_update_elem = htab_lru_percpu_map_update_elem,
.map_delete_elem = htab_lru_map_delete_elem,
.map_lookup_percpu_elem = htab_lru_percpu_map_lookup_percpu_elem,
.map_seq_show_elem = htab_percpu_map_seq_show_elem,
.map_set_for_each_callback_args = map_set_for_each_callback_args,
.map_for_each_callback = bpf_for_each_hash_elem,
bpf: hashtab memory usage htab_map_mem_usage() is introduced to calculate hashmap memory usage. In this helper, some small memory allocations are ignore, as their size is quite small compared with the total size. The inner_map_meta in hash_of_map is also ignored. The result for hashtab as follows, - before this change 1: hash name count_map flags 0x1 <<<< no prealloc, fully set key 16B value 24B max_entries 1048576 memlock 41943040B 2: hash name count_map flags 0x1 <<<< no prealloc, none set key 16B value 24B max_entries 1048576 memlock 41943040B 3: hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 41943040B The memlock is always a fixed size whatever it is preallocated or not, and whatever the count of allocated elements is. - after this change 1: hash name count_map flags 0x1 <<<< non prealloc, fully set key 16B value 24B max_entries 1048576 memlock 117441536B 2: hash name count_map flags 0x1 <<<< non prealloc, non set key 16B value 24B max_entries 1048576 memlock 16778240B 3: hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 109056000B The memlock now is hashtab actually allocated. The result for percpu hash map as follows, - before this change 4: percpu_hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 822083584B 5: percpu_hash name count_map flags 0x1 <<<< no prealloc key 16B value 24B max_entries 1048576 memlock 822083584B - after this change 4: percpu_hash name count_map flags 0x0 key 16B value 24B max_entries 1048576 memlock 897582080B 5: percpu_hash name count_map flags 0x1 key 16B value 24B max_entries 1048576 memlock 922748736B At worst, the difference can be 10x, for example, - before this change 6: hash name count_map flags 0x0 key 4B value 4B max_entries 1048576 memlock 8388608B - after this change 6: hash name count_map flags 0x0 key 4B value 4B max_entries 1048576 memlock 83889408B Signed-off-by: Yafang Shao <laoar.shao@gmail.com> Acked-by: Hou Tao <houtao1@huawei.com> Link: https://lore.kernel.org/r/20230305124615.12358-4-laoar.shao@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-05 12:46:00 +00:00
.map_mem_usage = htab_map_mem_usage,
bpf: Add batch ops to all htab bpf map htab can't use generic batch support due some problematic behaviours inherent to the data structre, i.e. while iterating the bpf map a concurrent program might delete the next entry that batch was about to use, in that case there's no easy solution to retrieve the next entry, the issue has been discussed multiple times (see [1] and [2]). The only way hmap can be traversed without the problem previously exposed is by making sure that the map is traversing entire buckets. This commit implements those strict requirements for hmap, the implementation follows the same interaction that generic support with some exceptions: - If keys/values buffer are not big enough to traverse a bucket, ENOSPC will be returned. - out_batch contains the value of the next bucket in the iteration, not the next key, but this is transparent for the user since the user should never use out_batch for other than bpf batch syscalls. This commits implements BPF_MAP_LOOKUP_BATCH and adds support for new command BPF_MAP_LOOKUP_AND_DELETE_BATCH. Note that for update/delete batch ops it is possible to use the generic implementations. [1] https://lore.kernel.org/bpf/20190724165803.87470-1-brianvv@google.com/ [2] https://lore.kernel.org/bpf/20190906225434.3635421-1-yhs@fb.com/ Signed-off-by: Yonghong Song <yhs@fb.com> Signed-off-by: Brian Vazquez <brianvv@google.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/20200115184308.162644-6-brianvv@google.com
2020-01-15 18:43:04 +00:00
BATCH_OPS(htab_lru_percpu),
.map_btf_id = &htab_map_btf_ids[0],
.iter_seq_info = &iter_seq_info,
};
static int fd_htab_map_alloc_check(union bpf_attr *attr)
{
if (attr->value_size != sizeof(u32))
return -EINVAL;
return htab_map_alloc_check(attr);
}
static void fd_htab_map_free(struct bpf_map *map)
{
struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
struct hlist_nulls_node *n;
struct hlist_nulls_head *head;
struct htab_elem *l;
int i;
for (i = 0; i < htab->n_buckets; i++) {
head = select_bucket(htab, i);
hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
void *ptr = fd_htab_map_get_ptr(map, l);
map->ops->map_fd_put_ptr(map, ptr, false);
}
}
htab_map_free(map);
}
/* only called from syscall */
int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
{
void **ptr;
int ret = 0;
if (!map->ops->map_fd_sys_lookup_elem)
return -ENOTSUPP;
rcu_read_lock();
ptr = htab_map_lookup_elem(map, key);
if (ptr)
*value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
else
ret = -ENOENT;
rcu_read_unlock();
return ret;
}
/* only called from syscall */
int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
void *key, void *value, u64 map_flags)
{
void *ptr;
int ret;
u32 ufd = *(u32 *)value;
ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
if (IS_ERR(ptr))
return PTR_ERR(ptr);
/* The htab bucket lock is always held during update operations in fd
* htab map, and the following rcu_read_lock() is only used to avoid
* the WARN_ON_ONCE in htab_map_update_elem().
*/
rcu_read_lock();
ret = htab_map_update_elem(map, key, &ptr, map_flags);
rcu_read_unlock();
if (ret)
map->ops->map_fd_put_ptr(map, ptr, false);
return ret;
}
static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
{
struct bpf_map *map, *inner_map_meta;
inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
if (IS_ERR(inner_map_meta))
return inner_map_meta;
map = htab_map_alloc(attr);
if (IS_ERR(map)) {
bpf_map_meta_free(inner_map_meta);
return map;
}
map->inner_map_meta = inner_map_meta;
return map;
}
static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
{
struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
if (!inner_map)
return NULL;
return READ_ONCE(*inner_map);
}
bpf: Allow for map-in-map with dynamic inner array map entries Recent work in f4d05259213f ("bpf: Add map_meta_equal map ops") and 134fede4eecf ("bpf: Relax max_entries check for most of the inner map types") added support for dynamic inner max elements for most map-in-map types. Exceptions were maps like array or prog array where the map_gen_lookup() callback uses the maps' max_entries field as a constant when emitting instructions. We recently implemented Maglev consistent hashing into Cilium's load balancer which uses map-in-map with an outer map being hash and inner being array holding the Maglev backend table for each service. This has been designed this way in order to reduce overall memory consumption given the outer hash map allows to avoid preallocating a large, flat memory area for all services. Also, the number of service mappings is not always known a-priori. The use case for dynamic inner array map entries is to further reduce memory overhead, for example, some services might just have a small number of back ends while others could have a large number. Right now the Maglev backend table for small and large number of backends would need to have the same inner array map entries which adds a lot of unneeded overhead. Dynamic inner array map entries can be realized by avoiding the inlined code generation for their lookup. The lookup will still be efficient since it will be calling into array_map_lookup_elem() directly and thus avoiding retpoline. The patch adds a BPF_F_INNER_MAP flag to map creation which therefore skips inline code generation and relaxes array_map_meta_equal() check to ignore both maps' max_entries. This also still allows to have faster lookups for map-in-map when BPF_F_INNER_MAP is not specified and hence dynamic max_entries not needed. Example code generation where inner map is dynamic sized array: # bpftool p d x i 125 int handle__sys_enter(void * ctx): ; int handle__sys_enter(void *ctx) 0: (b4) w1 = 0 ; int key = 0; 1: (63) *(u32 *)(r10 -4) = r1 2: (bf) r2 = r10 ; 3: (07) r2 += -4 ; inner_map = bpf_map_lookup_elem(&outer_arr_dyn, &key); 4: (18) r1 = map[id:468] 6: (07) r1 += 272 7: (61) r0 = *(u32 *)(r2 +0) 8: (35) if r0 >= 0x3 goto pc+5 9: (67) r0 <<= 3 10: (0f) r0 += r1 11: (79) r0 = *(u64 *)(r0 +0) 12: (15) if r0 == 0x0 goto pc+1 13: (05) goto pc+1 14: (b7) r0 = 0 15: (b4) w6 = -1 ; if (!inner_map) 16: (15) if r0 == 0x0 goto pc+6 17: (bf) r2 = r10 ; 18: (07) r2 += -4 ; val = bpf_map_lookup_elem(inner_map, &key); 19: (bf) r1 = r0 | No inlining but instead 20: (85) call array_map_lookup_elem#149280 | call to array_map_lookup_elem() ; return val ? *val : -1; | for inner array lookup. 21: (15) if r0 == 0x0 goto pc+1 ; return val ? *val : -1; 22: (61) r6 = *(u32 *)(r0 +0) ; } 23: (bc) w0 = w6 24: (95) exit Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Andrii Nakryiko <andrii@kernel.org> Link: https://lore.kernel.org/bpf/20201010234006.7075-4-daniel@iogearbox.net
2020-10-10 23:40:03 +00:00
static int htab_of_map_gen_lookup(struct bpf_map *map,
struct bpf_insn *insn_buf)
{
struct bpf_insn *insn = insn_buf;
const int ret = BPF_REG_0;
bpf: avoid retpoline for lookup/update/delete calls on maps While some of the BPF map lookup helpers provide a ->map_gen_lookup() callback for inlining the map lookup altogether it is not available for every map, so the remaining ones have to call bpf_map_lookup_elem() helper which does a dispatch to map->ops->map_lookup_elem(). In times of retpolines, this will control and trap speculative execution rather than letting it do its work for the indirect call and will therefore cause a slowdown. Likewise, bpf_map_update_elem() and bpf_map_delete_elem() do not have an inlined version and need to call into their map->ops->map_update_elem() resp. map->ops->map_delete_elem() handlers. Before: # bpftool prog dump xlated id 1 0: (bf) r2 = r10 1: (07) r2 += -8 2: (7a) *(u64 *)(r2 +0) = 0 3: (18) r1 = map[id:1] 5: (85) call __htab_map_lookup_elem#232656 6: (15) if r0 == 0x0 goto pc+4 7: (71) r1 = *(u8 *)(r0 +35) 8: (55) if r1 != 0x0 goto pc+1 9: (72) *(u8 *)(r0 +35) = 1 10: (07) r0 += 56 11: (15) if r0 == 0x0 goto pc+4 12: (bf) r2 = r0 13: (18) r1 = map[id:1] 15: (85) call bpf_map_delete_elem#215008 <-- indirect call via 16: (95) exit helper After: # bpftool prog dump xlated id 1 0: (bf) r2 = r10 1: (07) r2 += -8 2: (7a) *(u64 *)(r2 +0) = 0 3: (18) r1 = map[id:1] 5: (85) call __htab_map_lookup_elem#233328 6: (15) if r0 == 0x0 goto pc+4 7: (71) r1 = *(u8 *)(r0 +35) 8: (55) if r1 != 0x0 goto pc+1 9: (72) *(u8 *)(r0 +35) = 1 10: (07) r0 += 56 11: (15) if r0 == 0x0 goto pc+4 12: (bf) r2 = r0 13: (18) r1 = map[id:1] 15: (85) call htab_lru_map_delete_elem#238240 <-- direct call 16: (95) exit In all three lookup/update/delete cases however we can use the actual address of the map callback directly if we find that there's only a single path with a map pointer leading to the helper call, meaning when the map pointer has not been poisoned from verifier side. Example code can be seen above for the delete case. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2018-06-02 21:06:35 +00:00
BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
(void *(*)(struct bpf_map *map, void *key))NULL));
*insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
offsetof(struct htab_elem, key) +
round_up(map->key_size, 8));
*insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
return insn - insn_buf;
}
static void htab_of_map_free(struct bpf_map *map)
{
bpf_map_meta_free(map->inner_map_meta);
fd_htab_map_free(map);
}
const struct bpf_map_ops htab_of_maps_map_ops = {
.map_alloc_check = fd_htab_map_alloc_check,
.map_alloc = htab_of_map_alloc,
.map_free = htab_of_map_free,
.map_get_next_key = htab_map_get_next_key,
.map_lookup_elem = htab_of_map_lookup_elem,
.map_delete_elem = htab_map_delete_elem,
.map_fd_get_ptr = bpf_map_fd_get_ptr,
.map_fd_put_ptr = bpf_map_fd_put_ptr,
.map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
.map_gen_lookup = htab_of_map_gen_lookup,
.map_check_btf = map_check_no_btf,
bpf: hashtab memory usage htab_map_mem_usage() is introduced to calculate hashmap memory usage. In this helper, some small memory allocations are ignore, as their size is quite small compared with the total size. The inner_map_meta in hash_of_map is also ignored. The result for hashtab as follows, - before this change 1: hash name count_map flags 0x1 <<<< no prealloc, fully set key 16B value 24B max_entries 1048576 memlock 41943040B 2: hash name count_map flags 0x1 <<<< no prealloc, none set key 16B value 24B max_entries 1048576 memlock 41943040B 3: hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 41943040B The memlock is always a fixed size whatever it is preallocated or not, and whatever the count of allocated elements is. - after this change 1: hash name count_map flags 0x1 <<<< non prealloc, fully set key 16B value 24B max_entries 1048576 memlock 117441536B 2: hash name count_map flags 0x1 <<<< non prealloc, non set key 16B value 24B max_entries 1048576 memlock 16778240B 3: hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 109056000B The memlock now is hashtab actually allocated. The result for percpu hash map as follows, - before this change 4: percpu_hash name count_map flags 0x0 <<<< prealloc key 16B value 24B max_entries 1048576 memlock 822083584B 5: percpu_hash name count_map flags 0x1 <<<< no prealloc key 16B value 24B max_entries 1048576 memlock 822083584B - after this change 4: percpu_hash name count_map flags 0x0 key 16B value 24B max_entries 1048576 memlock 897582080B 5: percpu_hash name count_map flags 0x1 key 16B value 24B max_entries 1048576 memlock 922748736B At worst, the difference can be 10x, for example, - before this change 6: hash name count_map flags 0x0 key 4B value 4B max_entries 1048576 memlock 8388608B - after this change 6: hash name count_map flags 0x0 key 4B value 4B max_entries 1048576 memlock 83889408B Signed-off-by: Yafang Shao <laoar.shao@gmail.com> Acked-by: Hou Tao <houtao1@huawei.com> Link: https://lore.kernel.org/r/20230305124615.12358-4-laoar.shao@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-03-05 12:46:00 +00:00
.map_mem_usage = htab_map_mem_usage,
BATCH_OPS(htab),
.map_btf_id = &htab_map_btf_ids[0],
};