linux/kernel/bpf/devmap.c

1144 lines
29 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-only
/* Copyright (c) 2017 Covalent IO, Inc. http://covalent.io
*/
/* Devmaps primary use is as a backend map for XDP BPF helper call
* bpf_redirect_map(). Because XDP is mostly concerned with performance we
* spent some effort to ensure the datapath with redirect maps does not use
* any locking. This is a quick note on the details.
*
* We have three possible paths to get into the devmap control plane bpf
* syscalls, bpf programs, and driver side xmit/flush operations. A bpf syscall
* will invoke an update, delete, or lookup operation. To ensure updates and
* deletes appear atomic from the datapath side xchg() is used to modify the
* netdev_map array. Then because the datapath does a lookup into the netdev_map
* array (read-only) from an RCU critical section we use call_rcu() to wait for
* an rcu grace period before free'ing the old data structures. This ensures the
* datapath always has a valid copy. However, the datapath does a "flush"
* operation that pushes any pending packets in the driver outside the RCU
* critical section. Each bpf_dtab_netdev tracks these pending operations using
* a per-cpu flush list. The bpf_dtab_netdev object will not be destroyed until
* this list is empty, indicating outstanding flush operations have completed.
*
* BPF syscalls may race with BPF program calls on any of the update, delete
* or lookup operations. As noted above the xchg() operation also keep the
* netdev_map consistent in this case. From the devmap side BPF programs
* calling into these operations are the same as multiple user space threads
* making system calls.
*
* Finally, any of the above may race with a netdev_unregister notifier. The
* unregister notifier must search for net devices in the map structure that
* contain a reference to the net device and remove them. This is a two step
* process (a) dereference the bpf_dtab_netdev object in netdev_map and (b)
* check to see if the ifindex is the same as the net_device being removed.
bpf: devmap fix mutex in rcu critical section Originally we used a mutex to protect concurrent devmap update and delete operations from racing with netdev unregister notifier callbacks. The notifier hook is needed because we increment the netdev ref count when a dev is added to the devmap. This ensures the netdev reference is valid in the datapath. However, we don't want to block unregister events, hence the initial mutex and notifier handler. The concern was in the notifier hook we search the map for dev entries that hold a refcnt on the net device being torn down. But, in order to do this we require two steps, (i) dereference the netdev: dev = rcu_dereference(map[i]) (ii) test ifindex: dev->ifindex == removing_ifindex and then finally we can swap in the NULL dev in the map via an xchg operation, xchg(map[i], NULL) The danger here is a concurrent update could run a different xchg op concurrently leading us to replace the new dev with a NULL dev incorrectly. CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) dev = rcu_dereference(map[i]) xchg(map[i]), new_dev); rcu_call(dev,...) xchg(map[i], NULL) The above flow would create the incorrect state with the dev reference in the update path being lost. To resolve this the original code used a mutex around the above block. However, updates, deletes, and lookups occur inside rcu critical sections so we can't use a mutex in this context safely. Fortunately, by writing slightly better code we can avoid the mutex altogether. If CPU 1 in the above example uses a cmpxchg and _only_ replaces the dev reference in the map when it is in fact the expected dev the race is removed completely. The two cases being illustrated here, first the race condition, CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) dev = rcu_dereference(map[i]) xchg(map[i]), new_dev); rcu_call(dev,...) odev = cmpxchg(map[i], dev, NULL) Now we can test the cmpxchg return value, detect odev != dev and abort. Or in the good case, CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) odev = cmpxchg(map[i], dev, NULL) [...] Now 'odev == dev' and we can do proper cleanup. And viola the original race we tried to solve with a mutex is corrected and the trace noted by Sasha below is resolved due to removal of the mutex. Note: When walking the devmap and removing dev references as needed we depend on the core to fail any calls to dev_get_by_index() using the ifindex of the device being removed. This way we do not race with the user while searching the devmap. Additionally, the mutex was also protecting list add/del/read on the list of maps in-use. This patch converts this to an RCU list and spinlock implementation. This protects the list from concurrent alloc/free operations. The notifier hook walks this list so it uses RCU read semantics. BUG: sleeping function called from invalid context at kernel/locking/mutex.c:747 in_atomic(): 1, irqs_disabled(): 0, pid: 16315, name: syz-executor1 1 lock held by syz-executor1/16315: #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] map_delete_elem kernel/bpf/syscall.c:577 [inline] #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] SYSC_bpf kernel/bpf/syscall.c:1427 [inline] #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] SyS_bpf+0x1d32/0x4ba0 kernel/bpf/syscall.c:1388 Fixes: 2ddf71e23cc2 ("net: add notifier hooks for devmap bpf map") Reported-by: Sasha Levin <alexander.levin@verizon.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: John Fastabend <john.fastabend@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-05 05:02:19 +00:00
* When removing the dev a cmpxchg() is used to ensure the correct dev is
* removed, in the case of a concurrent update or delete operation it is
* possible that the initially referenced dev is no longer in the map. As the
* notifier hook walks the map we know that new dev references can not be
* added by the user because core infrastructure ensures dev_get_by_index()
* calls will fail at this point.
*
* The devmap_hash type is a map type which interprets keys as ifindexes and
* indexes these using a hashmap. This allows maps that use ifindex as key to be
* densely packed instead of having holes in the lookup array for unused
* ifindexes. The setup and packet enqueue/send code is shared between the two
* types of devmap; only the lookup and insertion is different.
*/
#include <linux/bpf.h>
#include <net/xdp.h>
#include <linux/filter.h>
#include <trace/events/xdp.h>
#include <linux/btf_ids.h>
#define DEV_CREATE_FLAG_MASK \
(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
xdp: Move devmap bulk queue into struct net_device Commit 96360004b862 ("xdp: Make devmap flush_list common for all map instances"), changed devmap flushing to be a global operation instead of a per-map operation. However, the queue structure used for bulking was still allocated as part of the containing map. This patch moves the devmap bulk queue into struct net_device. The motivation for this is reusing it for the non-map variant of XDP_REDIRECT, which will be changed in a subsequent commit. To avoid other fields of struct net_device moving to different cache lines, we also move a couple of other members around. We defer the actual allocation of the bulk queue structure until the NETDEV_REGISTER notification devmap.c. This makes it possible to check for ndo_xdp_xmit support before allocating the structure, which is not possible at the time struct net_device is allocated. However, we keep the freeing in free_netdev() to avoid adding another RCU callback on NETDEV_UNREGISTER. Because of this change, we lose the reference back to the map that originated the redirect, so change the tracepoint to always return 0 as the map ID and index. Otherwise no functional change is intended with this patch. After this patch, the relevant part of struct net_device looks like this, according to pahole: /* --- cacheline 14 boundary (896 bytes) --- */ struct netdev_queue * _tx __attribute__((__aligned__(64))); /* 896 8 */ unsigned int num_tx_queues; /* 904 4 */ unsigned int real_num_tx_queues; /* 908 4 */ struct Qdisc * qdisc; /* 912 8 */ unsigned int tx_queue_len; /* 920 4 */ spinlock_t tx_global_lock; /* 924 4 */ struct xdp_dev_bulk_queue * xdp_bulkq; /* 928 8 */ struct xps_dev_maps * xps_cpus_map; /* 936 8 */ struct xps_dev_maps * xps_rxqs_map; /* 944 8 */ struct mini_Qdisc * miniq_egress; /* 952 8 */ /* --- cacheline 15 boundary (960 bytes) --- */ struct hlist_head qdisc_hash[16]; /* 960 128 */ /* --- cacheline 17 boundary (1088 bytes) --- */ struct timer_list watchdog_timer; /* 1088 40 */ /* XXX last struct has 4 bytes of padding */ int watchdog_timeo; /* 1128 4 */ /* XXX 4 bytes hole, try to pack */ struct list_head todo_list; /* 1136 16 */ /* --- cacheline 18 boundary (1152 bytes) --- */ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Björn Töpel <bjorn.topel@intel.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768397.1458396.12673224324627072349.stgit@toke.dk
2020-01-16 15:14:44 +00:00
struct xdp_dev_bulk_queue {
struct xdp_frame *q[DEV_MAP_BULK_SIZE];
struct list_head flush_node;
xdp: Move devmap bulk queue into struct net_device Commit 96360004b862 ("xdp: Make devmap flush_list common for all map instances"), changed devmap flushing to be a global operation instead of a per-map operation. However, the queue structure used for bulking was still allocated as part of the containing map. This patch moves the devmap bulk queue into struct net_device. The motivation for this is reusing it for the non-map variant of XDP_REDIRECT, which will be changed in a subsequent commit. To avoid other fields of struct net_device moving to different cache lines, we also move a couple of other members around. We defer the actual allocation of the bulk queue structure until the NETDEV_REGISTER notification devmap.c. This makes it possible to check for ndo_xdp_xmit support before allocating the structure, which is not possible at the time struct net_device is allocated. However, we keep the freeing in free_netdev() to avoid adding another RCU callback on NETDEV_UNREGISTER. Because of this change, we lose the reference back to the map that originated the redirect, so change the tracepoint to always return 0 as the map ID and index. Otherwise no functional change is intended with this patch. After this patch, the relevant part of struct net_device looks like this, according to pahole: /* --- cacheline 14 boundary (896 bytes) --- */ struct netdev_queue * _tx __attribute__((__aligned__(64))); /* 896 8 */ unsigned int num_tx_queues; /* 904 4 */ unsigned int real_num_tx_queues; /* 908 4 */ struct Qdisc * qdisc; /* 912 8 */ unsigned int tx_queue_len; /* 920 4 */ spinlock_t tx_global_lock; /* 924 4 */ struct xdp_dev_bulk_queue * xdp_bulkq; /* 928 8 */ struct xps_dev_maps * xps_cpus_map; /* 936 8 */ struct xps_dev_maps * xps_rxqs_map; /* 944 8 */ struct mini_Qdisc * miniq_egress; /* 952 8 */ /* --- cacheline 15 boundary (960 bytes) --- */ struct hlist_head qdisc_hash[16]; /* 960 128 */ /* --- cacheline 17 boundary (1088 bytes) --- */ struct timer_list watchdog_timer; /* 1088 40 */ /* XXX last struct has 4 bytes of padding */ int watchdog_timeo; /* 1128 4 */ /* XXX 4 bytes hole, try to pack */ struct list_head todo_list; /* 1136 16 */ /* --- cacheline 18 boundary (1152 bytes) --- */ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Björn Töpel <bjorn.topel@intel.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768397.1458396.12673224324627072349.stgit@toke.dk
2020-01-16 15:14:44 +00:00
struct net_device *dev;
struct net_device *dev_rx;
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
struct bpf_prog *xdp_prog;
unsigned int count;
};
struct bpf_dtab_netdev {
struct net_device *dev; /* must be first member, due to tracepoint */
struct hlist_node index_hlist;
struct bpf_dtab *dtab;
struct bpf_prog *xdp_prog;
struct rcu_head rcu;
xdp: Move devmap bulk queue into struct net_device Commit 96360004b862 ("xdp: Make devmap flush_list common for all map instances"), changed devmap flushing to be a global operation instead of a per-map operation. However, the queue structure used for bulking was still allocated as part of the containing map. This patch moves the devmap bulk queue into struct net_device. The motivation for this is reusing it for the non-map variant of XDP_REDIRECT, which will be changed in a subsequent commit. To avoid other fields of struct net_device moving to different cache lines, we also move a couple of other members around. We defer the actual allocation of the bulk queue structure until the NETDEV_REGISTER notification devmap.c. This makes it possible to check for ndo_xdp_xmit support before allocating the structure, which is not possible at the time struct net_device is allocated. However, we keep the freeing in free_netdev() to avoid adding another RCU callback on NETDEV_UNREGISTER. Because of this change, we lose the reference back to the map that originated the redirect, so change the tracepoint to always return 0 as the map ID and index. Otherwise no functional change is intended with this patch. After this patch, the relevant part of struct net_device looks like this, according to pahole: /* --- cacheline 14 boundary (896 bytes) --- */ struct netdev_queue * _tx __attribute__((__aligned__(64))); /* 896 8 */ unsigned int num_tx_queues; /* 904 4 */ unsigned int real_num_tx_queues; /* 908 4 */ struct Qdisc * qdisc; /* 912 8 */ unsigned int tx_queue_len; /* 920 4 */ spinlock_t tx_global_lock; /* 924 4 */ struct xdp_dev_bulk_queue * xdp_bulkq; /* 928 8 */ struct xps_dev_maps * xps_cpus_map; /* 936 8 */ struct xps_dev_maps * xps_rxqs_map; /* 944 8 */ struct mini_Qdisc * miniq_egress; /* 952 8 */ /* --- cacheline 15 boundary (960 bytes) --- */ struct hlist_head qdisc_hash[16]; /* 960 128 */ /* --- cacheline 17 boundary (1088 bytes) --- */ struct timer_list watchdog_timer; /* 1088 40 */ /* XXX last struct has 4 bytes of padding */ int watchdog_timeo; /* 1128 4 */ /* XXX 4 bytes hole, try to pack */ struct list_head todo_list; /* 1136 16 */ /* --- cacheline 18 boundary (1152 bytes) --- */ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Björn Töpel <bjorn.topel@intel.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768397.1458396.12673224324627072349.stgit@toke.dk
2020-01-16 15:14:44 +00:00
unsigned int idx;
struct bpf_devmap_val val;
};
struct bpf_dtab {
struct bpf_map map;
xdp: Add proper __rcu annotations to redirect map entries XDP_REDIRECT works by a three-step process: the bpf_redirect() and bpf_redirect_map() helpers will lookup the target of the redirect and store it (along with some other metadata) in a per-CPU struct bpf_redirect_info. Next, when the program returns the XDP_REDIRECT return code, the driver will call xdp_do_redirect() which will use the information thus stored to actually enqueue the frame into a bulk queue structure (that differs slightly by map type, but shares the same principle). Finally, before exiting its NAPI poll loop, the driver will call xdp_do_flush(), which will flush all the different bulk queues, thus completing the redirect. Pointers to the map entries will be kept around for this whole sequence of steps, protected by RCU. However, there is no top-level rcu_read_lock() in the core code; instead drivers add their own rcu_read_lock() around the XDP portions of the code, but somewhat inconsistently as Martin discovered[0]. However, things still work because everything happens inside a single NAPI poll sequence, which means it's between a pair of calls to local_bh_disable()/local_bh_enable(). So Paul suggested[1] that we could document this intention by using rcu_dereference_check() with rcu_read_lock_bh_held() as a second parameter, thus allowing sparse and lockdep to verify that everything is done correctly. This patch does just that: we add an __rcu annotation to the map entry pointers and remove the various comments explaining the NAPI poll assurance strewn through devmap.c in favour of a longer explanation in filter.c. The goal is to have one coherent documentation of the entire flow, and rely on the RCU annotations as a "standard" way of communicating the flow in the map code (which can additionally be understood by sparse and lockdep). The RCU annotation replacements result in a fairly straight-forward replacement where READ_ONCE() becomes rcu_dereference_check(), WRITE_ONCE() becomes rcu_assign_pointer() and xchg() and cmpxchg() gets wrapped in the proper constructs to cast the pointer back and forth between __rcu and __kernel address space (for the benefit of sparse). The one complication is that xskmap has a few constructions where double-pointers are passed back and forth; these simply all gain __rcu annotations, and only the final reference/dereference to the inner-most pointer gets changed. With this, everything can be run through sparse without eliciting complaints, and lockdep can verify correctness even without the use of rcu_read_lock() in the drivers. Subsequent patches will clean these up from the drivers. [0] https://lore.kernel.org/bpf/20210415173551.7ma4slcbqeyiba2r@kafai-mbp.dhcp.thefacebook.com/ [1] https://lore.kernel.org/bpf/20210419165837.GA975577@paulmck-ThinkPad-P17-Gen-1/ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20210624160609.292325-6-toke@redhat.com
2021-06-24 16:05:55 +00:00
struct bpf_dtab_netdev __rcu **netdev_map; /* DEVMAP type only */
struct list_head list;
/* these are only used for DEVMAP_HASH type maps */
struct hlist_head *dev_index_head;
spinlock_t index_lock;
unsigned int items;
u32 n_buckets;
};
xdp: Use bulking for non-map XDP_REDIRECT and consolidate code paths Since the bulk queue used by XDP_REDIRECT now lives in struct net_device, we can re-use the bulking for the non-map version of the bpf_redirect() helper. This is a simple matter of having xdp_do_redirect_slow() queue the frame on the bulk queue instead of sending it out with __bpf_tx_xdp(). Unfortunately we can't make the bpf_redirect() helper return an error if the ifindex doesn't exit (as bpf_redirect_map() does), because we don't have a reference to the network namespace of the ingress device at the time the helper is called. So we have to leave it as-is and keep the device lookup in xdp_do_redirect_slow(). Since this leaves less reason to have the non-map redirect code in a separate function, so we get rid of the xdp_do_redirect_slow() function entirely. This does lose us the tracepoint disambiguation, but fortunately the xdp_redirect and xdp_redirect_map tracepoints use the same tracepoint entry structures. This means both can contain a map index, so we can just amend the tracepoint definitions so we always emit the xdp_redirect(_err) tracepoints, but with the map ID only populated if a map is present. This means we retire the xdp_redirect_map(_err) tracepoints entirely, but keep the definitions around in case someone is still listening for them. With this change, the performance of the xdp_redirect sample program goes from 5Mpps to 8.4Mpps (a 68% increase). Since the flush functions are no longer map-specific, rename the flush() functions to drop _map from their names. One of the renamed functions is the xdp_do_flush_map() callback used in all the xdp-enabled drivers. To keep from having to update all drivers, use a #define to keep the old name working, and only update the virtual drivers in this patch. Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768505.1458396.17518057312953572912.stgit@toke.dk
2020-01-16 15:14:45 +00:00
static DEFINE_PER_CPU(struct list_head, dev_flush_list);
bpf: devmap fix mutex in rcu critical section Originally we used a mutex to protect concurrent devmap update and delete operations from racing with netdev unregister notifier callbacks. The notifier hook is needed because we increment the netdev ref count when a dev is added to the devmap. This ensures the netdev reference is valid in the datapath. However, we don't want to block unregister events, hence the initial mutex and notifier handler. The concern was in the notifier hook we search the map for dev entries that hold a refcnt on the net device being torn down. But, in order to do this we require two steps, (i) dereference the netdev: dev = rcu_dereference(map[i]) (ii) test ifindex: dev->ifindex == removing_ifindex and then finally we can swap in the NULL dev in the map via an xchg operation, xchg(map[i], NULL) The danger here is a concurrent update could run a different xchg op concurrently leading us to replace the new dev with a NULL dev incorrectly. CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) dev = rcu_dereference(map[i]) xchg(map[i]), new_dev); rcu_call(dev,...) xchg(map[i], NULL) The above flow would create the incorrect state with the dev reference in the update path being lost. To resolve this the original code used a mutex around the above block. However, updates, deletes, and lookups occur inside rcu critical sections so we can't use a mutex in this context safely. Fortunately, by writing slightly better code we can avoid the mutex altogether. If CPU 1 in the above example uses a cmpxchg and _only_ replaces the dev reference in the map when it is in fact the expected dev the race is removed completely. The two cases being illustrated here, first the race condition, CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) dev = rcu_dereference(map[i]) xchg(map[i]), new_dev); rcu_call(dev,...) odev = cmpxchg(map[i], dev, NULL) Now we can test the cmpxchg return value, detect odev != dev and abort. Or in the good case, CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) odev = cmpxchg(map[i], dev, NULL) [...] Now 'odev == dev' and we can do proper cleanup. And viola the original race we tried to solve with a mutex is corrected and the trace noted by Sasha below is resolved due to removal of the mutex. Note: When walking the devmap and removing dev references as needed we depend on the core to fail any calls to dev_get_by_index() using the ifindex of the device being removed. This way we do not race with the user while searching the devmap. Additionally, the mutex was also protecting list add/del/read on the list of maps in-use. This patch converts this to an RCU list and spinlock implementation. This protects the list from concurrent alloc/free operations. The notifier hook walks this list so it uses RCU read semantics. BUG: sleeping function called from invalid context at kernel/locking/mutex.c:747 in_atomic(): 1, irqs_disabled(): 0, pid: 16315, name: syz-executor1 1 lock held by syz-executor1/16315: #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] map_delete_elem kernel/bpf/syscall.c:577 [inline] #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] SYSC_bpf kernel/bpf/syscall.c:1427 [inline] #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] SyS_bpf+0x1d32/0x4ba0 kernel/bpf/syscall.c:1388 Fixes: 2ddf71e23cc2 ("net: add notifier hooks for devmap bpf map") Reported-by: Sasha Levin <alexander.levin@verizon.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: John Fastabend <john.fastabend@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-05 05:02:19 +00:00
static DEFINE_SPINLOCK(dev_map_lock);
static LIST_HEAD(dev_map_list);
static struct hlist_head *dev_map_create_hash(unsigned int entries,
int numa_node)
{
int i;
struct hlist_head *hash;
hash = bpf_map_area_alloc((u64) entries * sizeof(*hash), numa_node);
if (hash != NULL)
for (i = 0; i < entries; i++)
INIT_HLIST_HEAD(&hash[i]);
return hash;
}
static inline struct hlist_head *dev_map_index_hash(struct bpf_dtab *dtab,
int idx)
{
return &dtab->dev_index_head[idx & (dtab->n_buckets - 1)];
}
static int dev_map_init_map(struct bpf_dtab *dtab, union bpf_attr *attr)
{
u32 valsize = attr->value_size;
/* check sanity of attributes. 2 value sizes supported:
* 4 bytes: ifindex
* 8 bytes: ifindex + prog fd
*/
if (attr->max_entries == 0 || attr->key_size != 4 ||
(valsize != offsetofend(struct bpf_devmap_val, ifindex) &&
valsize != offsetofend(struct bpf_devmap_val, bpf_prog.fd)) ||
attr->map_flags & ~DEV_CREATE_FLAG_MASK)
return -EINVAL;
/* Lookup returns a pointer straight to dev->ifindex, so make sure the
* verifier prevents writes from the BPF side
*/
attr->map_flags |= BPF_F_RDONLY_PROG;
bpf_map_init_from_attr(&dtab->map, attr);
if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
dtab->n_buckets = roundup_pow_of_two(dtab->map.max_entries);
if (!dtab->n_buckets) /* Overflow check */
return -EINVAL;
}
if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
dtab->dev_index_head = dev_map_create_hash(dtab->n_buckets,
dtab->map.numa_node);
if (!dtab->dev_index_head)
return -ENOMEM;
spin_lock_init(&dtab->index_lock);
} else {
dtab->netdev_map = bpf_map_area_alloc((u64) dtab->map.max_entries *
sizeof(struct bpf_dtab_netdev *),
dtab->map.numa_node);
if (!dtab->netdev_map)
return -ENOMEM;
}
return 0;
}
static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
{
struct bpf_dtab *dtab;
int err;
if (!capable(CAP_NET_ADMIN))
return ERR_PTR(-EPERM);
dtab = bpf_map_area_alloc(sizeof(*dtab), NUMA_NO_NODE);
if (!dtab)
return ERR_PTR(-ENOMEM);
err = dev_map_init_map(dtab, attr);
if (err) {
bpf_map_area_free(dtab);
return ERR_PTR(err);
}
spin_lock(&dev_map_lock);
list_add_tail_rcu(&dtab->list, &dev_map_list);
spin_unlock(&dev_map_lock);
return &dtab->map;
}
static void dev_map_free(struct bpf_map *map)
{
struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
int i;
/* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
* so the programs (can be more than one that used this map) were
* disconnected from events. The following synchronize_rcu() guarantees
* both rcu read critical sections complete and waits for
* preempt-disable regions (NAPI being the relevant context here) so we
* are certain there will be no further reads against the netdev_map and
* all flush operations are complete. Flush operations can only be done
* from NAPI context for this reason.
*/
spin_lock(&dev_map_lock);
list_del_rcu(&dtab->list);
spin_unlock(&dev_map_lock);
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
bpf_clear_redirect_map(map);
synchronize_rcu();
bpf: devmap: fix use-after-free Read in __dev_map_entry_free synchronize_rcu() is fine when the rcu callbacks only need to free memory (kfree_rcu() or direct kfree() call rcu call backs) __dev_map_entry_free() is a bit more complex, so we need to make sure that call queued __dev_map_entry_free() callbacks have completed. sysbot report: BUG: KASAN: use-after-free in dev_map_flush_old kernel/bpf/devmap.c:365 [inline] BUG: KASAN: use-after-free in __dev_map_entry_free+0x2a8/0x300 kernel/bpf/devmap.c:379 Read of size 8 at addr ffff8801b8da38c8 by task ksoftirqd/1/18 CPU: 1 PID: 18 Comm: ksoftirqd/1 Not tainted 4.17.0+ #39 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+0x1b9/0x294 lib/dump_stack.c:113 print_address_description+0x6c/0x20b mm/kasan/report.c:256 kasan_report_error mm/kasan/report.c:354 [inline] kasan_report.cold.7+0x242/0x2fe mm/kasan/report.c:412 __asan_report_load8_noabort+0x14/0x20 mm/kasan/report.c:433 dev_map_flush_old kernel/bpf/devmap.c:365 [inline] __dev_map_entry_free+0x2a8/0x300 kernel/bpf/devmap.c:379 __rcu_reclaim kernel/rcu/rcu.h:178 [inline] rcu_do_batch kernel/rcu/tree.c:2558 [inline] invoke_rcu_callbacks kernel/rcu/tree.c:2818 [inline] __rcu_process_callbacks kernel/rcu/tree.c:2785 [inline] rcu_process_callbacks+0xe9d/0x1760 kernel/rcu/tree.c:2802 __do_softirq+0x2e0/0xaf5 kernel/softirq.c:284 run_ksoftirqd+0x86/0x100 kernel/softirq.c:645 smpboot_thread_fn+0x417/0x870 kernel/smpboot.c:164 kthread+0x345/0x410 kernel/kthread.c:240 ret_from_fork+0x3a/0x50 arch/x86/entry/entry_64.S:412 Allocated by task 6675: save_stack+0x43/0xd0 mm/kasan/kasan.c:448 set_track mm/kasan/kasan.c:460 [inline] kasan_kmalloc+0xc4/0xe0 mm/kasan/kasan.c:553 kmem_cache_alloc_trace+0x152/0x780 mm/slab.c:3620 kmalloc include/linux/slab.h:513 [inline] kzalloc include/linux/slab.h:706 [inline] dev_map_alloc+0x208/0x7f0 kernel/bpf/devmap.c:102 find_and_alloc_map kernel/bpf/syscall.c:129 [inline] map_create+0x393/0x1010 kernel/bpf/syscall.c:453 __do_sys_bpf kernel/bpf/syscall.c:2351 [inline] __se_sys_bpf kernel/bpf/syscall.c:2328 [inline] __x64_sys_bpf+0x303/0x510 kernel/bpf/syscall.c:2328 do_syscall_64+0x1b1/0x800 arch/x86/entry/common.c:290 entry_SYSCALL_64_after_hwframe+0x49/0xbe Freed by task 26: save_stack+0x43/0xd0 mm/kasan/kasan.c:448 set_track mm/kasan/kasan.c:460 [inline] __kasan_slab_free+0x11a/0x170 mm/kasan/kasan.c:521 kasan_slab_free+0xe/0x10 mm/kasan/kasan.c:528 __cache_free mm/slab.c:3498 [inline] kfree+0xd9/0x260 mm/slab.c:3813 dev_map_free+0x4fa/0x670 kernel/bpf/devmap.c:191 bpf_map_free_deferred+0xba/0xf0 kernel/bpf/syscall.c:262 process_one_work+0xc64/0x1b70 kernel/workqueue.c:2153 worker_thread+0x181/0x13a0 kernel/workqueue.c:2296 kthread+0x345/0x410 kernel/kthread.c:240 ret_from_fork+0x3a/0x50 arch/x86/entry/entry_64.S:412 The buggy address belongs to the object at ffff8801b8da37c0 which belongs to the cache kmalloc-512 of size 512 The buggy address is located 264 bytes inside of 512-byte region [ffff8801b8da37c0, ffff8801b8da39c0) The buggy address belongs to the page: page:ffffea0006e368c0 count:1 mapcount:0 mapping:ffff8801da800940 index:0xffff8801b8da3540 flags: 0x2fffc0000000100(slab) raw: 02fffc0000000100 ffffea0007217b88 ffffea0006e30cc8 ffff8801da800940 raw: ffff8801b8da3540 ffff8801b8da3040 0000000100000004 0000000000000000 page dumped because: kasan: bad access detected Memory state around the buggy address: ffff8801b8da3780: fc fc fc fc fc fc fc fc fb fb fb fb fb fb fb fb ffff8801b8da3800: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb > ffff8801b8da3880: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ^ ffff8801b8da3900: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ffff8801b8da3980: fb fb fb fb fb fb fb fb fc fc fc fc fc fc fc fc Fixes: 546ac1ffb70d ("bpf: add devmap, a map for storing net device references") Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: syzbot+457d3e2ffbcf31aee5c0@syzkaller.appspotmail.com Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2019-05-13 16:59:16 +00:00
/* Make sure prior __dev_map_entry_free() have completed. */
rcu_barrier();
if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
for (i = 0; i < dtab->n_buckets; i++) {
struct bpf_dtab_netdev *dev;
struct hlist_head *head;
struct hlist_node *next;
head = dev_map_index_hash(dtab, i);
hlist_for_each_entry_safe(dev, next, head, index_hlist) {
hlist_del_rcu(&dev->index_hlist);
if (dev->xdp_prog)
bpf_prog_put(dev->xdp_prog);
dev_put(dev->dev);
kfree(dev);
}
}
bpf_map_area_free(dtab->dev_index_head);
} else {
for (i = 0; i < dtab->map.max_entries; i++) {
struct bpf_dtab_netdev *dev;
xdp: Add proper __rcu annotations to redirect map entries XDP_REDIRECT works by a three-step process: the bpf_redirect() and bpf_redirect_map() helpers will lookup the target of the redirect and store it (along with some other metadata) in a per-CPU struct bpf_redirect_info. Next, when the program returns the XDP_REDIRECT return code, the driver will call xdp_do_redirect() which will use the information thus stored to actually enqueue the frame into a bulk queue structure (that differs slightly by map type, but shares the same principle). Finally, before exiting its NAPI poll loop, the driver will call xdp_do_flush(), which will flush all the different bulk queues, thus completing the redirect. Pointers to the map entries will be kept around for this whole sequence of steps, protected by RCU. However, there is no top-level rcu_read_lock() in the core code; instead drivers add their own rcu_read_lock() around the XDP portions of the code, but somewhat inconsistently as Martin discovered[0]. However, things still work because everything happens inside a single NAPI poll sequence, which means it's between a pair of calls to local_bh_disable()/local_bh_enable(). So Paul suggested[1] that we could document this intention by using rcu_dereference_check() with rcu_read_lock_bh_held() as a second parameter, thus allowing sparse and lockdep to verify that everything is done correctly. This patch does just that: we add an __rcu annotation to the map entry pointers and remove the various comments explaining the NAPI poll assurance strewn through devmap.c in favour of a longer explanation in filter.c. The goal is to have one coherent documentation of the entire flow, and rely on the RCU annotations as a "standard" way of communicating the flow in the map code (which can additionally be understood by sparse and lockdep). The RCU annotation replacements result in a fairly straight-forward replacement where READ_ONCE() becomes rcu_dereference_check(), WRITE_ONCE() becomes rcu_assign_pointer() and xchg() and cmpxchg() gets wrapped in the proper constructs to cast the pointer back and forth between __rcu and __kernel address space (for the benefit of sparse). The one complication is that xskmap has a few constructions where double-pointers are passed back and forth; these simply all gain __rcu annotations, and only the final reference/dereference to the inner-most pointer gets changed. With this, everything can be run through sparse without eliciting complaints, and lockdep can verify correctness even without the use of rcu_read_lock() in the drivers. Subsequent patches will clean these up from the drivers. [0] https://lore.kernel.org/bpf/20210415173551.7ma4slcbqeyiba2r@kafai-mbp.dhcp.thefacebook.com/ [1] https://lore.kernel.org/bpf/20210419165837.GA975577@paulmck-ThinkPad-P17-Gen-1/ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20210624160609.292325-6-toke@redhat.com
2021-06-24 16:05:55 +00:00
dev = rcu_dereference_raw(dtab->netdev_map[i]);
if (!dev)
continue;
if (dev->xdp_prog)
bpf_prog_put(dev->xdp_prog);
dev_put(dev->dev);
kfree(dev);
}
bpf_map_area_free(dtab->netdev_map);
}
bpf_map_area_free(dtab);
}
static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
{
struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
u32 index = key ? *(u32 *)key : U32_MAX;
u32 *next = next_key;
if (index >= dtab->map.max_entries) {
*next = 0;
return 0;
}
if (index == dtab->map.max_entries - 1)
return -ENOENT;
*next = index + 1;
return 0;
}
xdp: Add proper __rcu annotations to redirect map entries XDP_REDIRECT works by a three-step process: the bpf_redirect() and bpf_redirect_map() helpers will lookup the target of the redirect and store it (along with some other metadata) in a per-CPU struct bpf_redirect_info. Next, when the program returns the XDP_REDIRECT return code, the driver will call xdp_do_redirect() which will use the information thus stored to actually enqueue the frame into a bulk queue structure (that differs slightly by map type, but shares the same principle). Finally, before exiting its NAPI poll loop, the driver will call xdp_do_flush(), which will flush all the different bulk queues, thus completing the redirect. Pointers to the map entries will be kept around for this whole sequence of steps, protected by RCU. However, there is no top-level rcu_read_lock() in the core code; instead drivers add their own rcu_read_lock() around the XDP portions of the code, but somewhat inconsistently as Martin discovered[0]. However, things still work because everything happens inside a single NAPI poll sequence, which means it's between a pair of calls to local_bh_disable()/local_bh_enable(). So Paul suggested[1] that we could document this intention by using rcu_dereference_check() with rcu_read_lock_bh_held() as a second parameter, thus allowing sparse and lockdep to verify that everything is done correctly. This patch does just that: we add an __rcu annotation to the map entry pointers and remove the various comments explaining the NAPI poll assurance strewn through devmap.c in favour of a longer explanation in filter.c. The goal is to have one coherent documentation of the entire flow, and rely on the RCU annotations as a "standard" way of communicating the flow in the map code (which can additionally be understood by sparse and lockdep). The RCU annotation replacements result in a fairly straight-forward replacement where READ_ONCE() becomes rcu_dereference_check(), WRITE_ONCE() becomes rcu_assign_pointer() and xchg() and cmpxchg() gets wrapped in the proper constructs to cast the pointer back and forth between __rcu and __kernel address space (for the benefit of sparse). The one complication is that xskmap has a few constructions where double-pointers are passed back and forth; these simply all gain __rcu annotations, and only the final reference/dereference to the inner-most pointer gets changed. With this, everything can be run through sparse without eliciting complaints, and lockdep can verify correctness even without the use of rcu_read_lock() in the drivers. Subsequent patches will clean these up from the drivers. [0] https://lore.kernel.org/bpf/20210415173551.7ma4slcbqeyiba2r@kafai-mbp.dhcp.thefacebook.com/ [1] https://lore.kernel.org/bpf/20210419165837.GA975577@paulmck-ThinkPad-P17-Gen-1/ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20210624160609.292325-6-toke@redhat.com
2021-06-24 16:05:55 +00:00
/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
* by local_bh_disable() (from XDP calls inside NAPI). The
* rcu_read_lock_bh_held() below makes lockdep accept both.
*/
static void *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
{
struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
struct hlist_head *head = dev_map_index_hash(dtab, key);
struct bpf_dtab_netdev *dev;
hlist_for_each_entry_rcu(dev, head, index_hlist,
lockdep_is_held(&dtab->index_lock))
if (dev->idx == key)
return dev;
return NULL;
}
static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
void *next_key)
{
struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
u32 idx, *next = next_key;
struct bpf_dtab_netdev *dev, *next_dev;
struct hlist_head *head;
int i = 0;
if (!key)
goto find_first;
idx = *(u32 *)key;
dev = __dev_map_hash_lookup_elem(map, idx);
if (!dev)
goto find_first;
next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
struct bpf_dtab_netdev, index_hlist);
if (next_dev) {
*next = next_dev->idx;
return 0;
}
i = idx & (dtab->n_buckets - 1);
i++;
find_first:
for (; i < dtab->n_buckets; i++) {
head = dev_map_index_hash(dtab, i);
next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
struct bpf_dtab_netdev,
index_hlist);
if (next_dev) {
*next = next_dev->idx;
return 0;
}
}
return -ENOENT;
}
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
static int dev_map_bpf_prog_run(struct bpf_prog *xdp_prog,
struct xdp_frame **frames, int n,
struct net_device *dev)
{
struct xdp_txq_info txq = { .dev = dev };
struct xdp_buff xdp;
int i, nframes = 0;
for (i = 0; i < n; i++) {
struct xdp_frame *xdpf = frames[i];
u32 act;
int err;
xdp_convert_frame_to_buff(xdpf, &xdp);
xdp.txq = &txq;
act = bpf_prog_run_xdp(xdp_prog, &xdp);
switch (act) {
case XDP_PASS:
err = xdp_update_frame_from_buff(&xdp, xdpf);
if (unlikely(err < 0))
xdp_return_frame_rx_napi(xdpf);
else
frames[nframes++] = xdpf;
break;
default:
bpf_warn_invalid_xdp_action(NULL, xdp_prog, act);
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
fallthrough;
case XDP_ABORTED:
trace_xdp_exception(dev, xdp_prog, act);
fallthrough;
case XDP_DROP:
xdp_return_frame_rx_napi(xdpf);
break;
}
}
return nframes; /* sent frames count */
}
static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
{
xdp: Move devmap bulk queue into struct net_device Commit 96360004b862 ("xdp: Make devmap flush_list common for all map instances"), changed devmap flushing to be a global operation instead of a per-map operation. However, the queue structure used for bulking was still allocated as part of the containing map. This patch moves the devmap bulk queue into struct net_device. The motivation for this is reusing it for the non-map variant of XDP_REDIRECT, which will be changed in a subsequent commit. To avoid other fields of struct net_device moving to different cache lines, we also move a couple of other members around. We defer the actual allocation of the bulk queue structure until the NETDEV_REGISTER notification devmap.c. This makes it possible to check for ndo_xdp_xmit support before allocating the structure, which is not possible at the time struct net_device is allocated. However, we keep the freeing in free_netdev() to avoid adding another RCU callback on NETDEV_UNREGISTER. Because of this change, we lose the reference back to the map that originated the redirect, so change the tracepoint to always return 0 as the map ID and index. Otherwise no functional change is intended with this patch. After this patch, the relevant part of struct net_device looks like this, according to pahole: /* --- cacheline 14 boundary (896 bytes) --- */ struct netdev_queue * _tx __attribute__((__aligned__(64))); /* 896 8 */ unsigned int num_tx_queues; /* 904 4 */ unsigned int real_num_tx_queues; /* 908 4 */ struct Qdisc * qdisc; /* 912 8 */ unsigned int tx_queue_len; /* 920 4 */ spinlock_t tx_global_lock; /* 924 4 */ struct xdp_dev_bulk_queue * xdp_bulkq; /* 928 8 */ struct xps_dev_maps * xps_cpus_map; /* 936 8 */ struct xps_dev_maps * xps_rxqs_map; /* 944 8 */ struct mini_Qdisc * miniq_egress; /* 952 8 */ /* --- cacheline 15 boundary (960 bytes) --- */ struct hlist_head qdisc_hash[16]; /* 960 128 */ /* --- cacheline 17 boundary (1088 bytes) --- */ struct timer_list watchdog_timer; /* 1088 40 */ /* XXX last struct has 4 bytes of padding */ int watchdog_timeo; /* 1128 4 */ /* XXX 4 bytes hole, try to pack */ struct list_head todo_list; /* 1136 16 */ /* --- cacheline 18 boundary (1152 bytes) --- */ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Björn Töpel <bjorn.topel@intel.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768397.1458396.12673224324627072349.stgit@toke.dk
2020-01-16 15:14:44 +00:00
struct net_device *dev = bq->dev;
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
unsigned int cnt = bq->count;
int sent = 0, err = 0;
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
int to_send = cnt;
int i;
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
if (unlikely(!cnt))
return;
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
for (i = 0; i < cnt; i++) {
struct xdp_frame *xdpf = bq->q[i];
prefetch(xdpf);
}
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
if (bq->xdp_prog) {
to_send = dev_map_bpf_prog_run(bq->xdp_prog, bq->q, cnt, dev);
if (!to_send)
goto out;
}
sent = dev->netdev_ops->ndo_xdp_xmit(dev, to_send, bq->q, flags);
if (sent < 0) {
/* If ndo_xdp_xmit fails with an errno, no frames have
* been xmit'ed.
*/
err = sent;
sent = 0;
}
/* If not all frames have been transmitted, it is our
* responsibility to free them
*/
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
for (i = sent; unlikely(i < to_send); i++)
xdp_return_frame_rx_napi(bq->q[i]);
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
out:
bq->count = 0;
trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, cnt - sent, err);
}
xdp: Add proper __rcu annotations to redirect map entries XDP_REDIRECT works by a three-step process: the bpf_redirect() and bpf_redirect_map() helpers will lookup the target of the redirect and store it (along with some other metadata) in a per-CPU struct bpf_redirect_info. Next, when the program returns the XDP_REDIRECT return code, the driver will call xdp_do_redirect() which will use the information thus stored to actually enqueue the frame into a bulk queue structure (that differs slightly by map type, but shares the same principle). Finally, before exiting its NAPI poll loop, the driver will call xdp_do_flush(), which will flush all the different bulk queues, thus completing the redirect. Pointers to the map entries will be kept around for this whole sequence of steps, protected by RCU. However, there is no top-level rcu_read_lock() in the core code; instead drivers add their own rcu_read_lock() around the XDP portions of the code, but somewhat inconsistently as Martin discovered[0]. However, things still work because everything happens inside a single NAPI poll sequence, which means it's between a pair of calls to local_bh_disable()/local_bh_enable(). So Paul suggested[1] that we could document this intention by using rcu_dereference_check() with rcu_read_lock_bh_held() as a second parameter, thus allowing sparse and lockdep to verify that everything is done correctly. This patch does just that: we add an __rcu annotation to the map entry pointers and remove the various comments explaining the NAPI poll assurance strewn through devmap.c in favour of a longer explanation in filter.c. The goal is to have one coherent documentation of the entire flow, and rely on the RCU annotations as a "standard" way of communicating the flow in the map code (which can additionally be understood by sparse and lockdep). The RCU annotation replacements result in a fairly straight-forward replacement where READ_ONCE() becomes rcu_dereference_check(), WRITE_ONCE() becomes rcu_assign_pointer() and xchg() and cmpxchg() gets wrapped in the proper constructs to cast the pointer back and forth between __rcu and __kernel address space (for the benefit of sparse). The one complication is that xskmap has a few constructions where double-pointers are passed back and forth; these simply all gain __rcu annotations, and only the final reference/dereference to the inner-most pointer gets changed. With this, everything can be run through sparse without eliciting complaints, and lockdep can verify correctness even without the use of rcu_read_lock() in the drivers. Subsequent patches will clean these up from the drivers. [0] https://lore.kernel.org/bpf/20210415173551.7ma4slcbqeyiba2r@kafai-mbp.dhcp.thefacebook.com/ [1] https://lore.kernel.org/bpf/20210419165837.GA975577@paulmck-ThinkPad-P17-Gen-1/ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20210624160609.292325-6-toke@redhat.com
2021-06-24 16:05:55 +00:00
/* __dev_flush is called from xdp_do_flush() which _must_ be signalled from the
* driver before returning from its napi->poll() routine. See the comment above
* xdp_do_flush() in filter.c.
*/
xdp: Use bulking for non-map XDP_REDIRECT and consolidate code paths Since the bulk queue used by XDP_REDIRECT now lives in struct net_device, we can re-use the bulking for the non-map version of the bpf_redirect() helper. This is a simple matter of having xdp_do_redirect_slow() queue the frame on the bulk queue instead of sending it out with __bpf_tx_xdp(). Unfortunately we can't make the bpf_redirect() helper return an error if the ifindex doesn't exit (as bpf_redirect_map() does), because we don't have a reference to the network namespace of the ingress device at the time the helper is called. So we have to leave it as-is and keep the device lookup in xdp_do_redirect_slow(). Since this leaves less reason to have the non-map redirect code in a separate function, so we get rid of the xdp_do_redirect_slow() function entirely. This does lose us the tracepoint disambiguation, but fortunately the xdp_redirect and xdp_redirect_map tracepoints use the same tracepoint entry structures. This means both can contain a map index, so we can just amend the tracepoint definitions so we always emit the xdp_redirect(_err) tracepoints, but with the map ID only populated if a map is present. This means we retire the xdp_redirect_map(_err) tracepoints entirely, but keep the definitions around in case someone is still listening for them. With this change, the performance of the xdp_redirect sample program goes from 5Mpps to 8.4Mpps (a 68% increase). Since the flush functions are no longer map-specific, rename the flush() functions to drop _map from their names. One of the renamed functions is the xdp_do_flush_map() callback used in all the xdp-enabled drivers. To keep from having to update all drivers, use a #define to keep the old name working, and only update the virtual drivers in this patch. Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768505.1458396.17518057312953572912.stgit@toke.dk
2020-01-16 15:14:45 +00:00
void __dev_flush(void)
{
xdp: Use bulking for non-map XDP_REDIRECT and consolidate code paths Since the bulk queue used by XDP_REDIRECT now lives in struct net_device, we can re-use the bulking for the non-map version of the bpf_redirect() helper. This is a simple matter of having xdp_do_redirect_slow() queue the frame on the bulk queue instead of sending it out with __bpf_tx_xdp(). Unfortunately we can't make the bpf_redirect() helper return an error if the ifindex doesn't exit (as bpf_redirect_map() does), because we don't have a reference to the network namespace of the ingress device at the time the helper is called. So we have to leave it as-is and keep the device lookup in xdp_do_redirect_slow(). Since this leaves less reason to have the non-map redirect code in a separate function, so we get rid of the xdp_do_redirect_slow() function entirely. This does lose us the tracepoint disambiguation, but fortunately the xdp_redirect and xdp_redirect_map tracepoints use the same tracepoint entry structures. This means both can contain a map index, so we can just amend the tracepoint definitions so we always emit the xdp_redirect(_err) tracepoints, but with the map ID only populated if a map is present. This means we retire the xdp_redirect_map(_err) tracepoints entirely, but keep the definitions around in case someone is still listening for them. With this change, the performance of the xdp_redirect sample program goes from 5Mpps to 8.4Mpps (a 68% increase). Since the flush functions are no longer map-specific, rename the flush() functions to drop _map from their names. One of the renamed functions is the xdp_do_flush_map() callback used in all the xdp-enabled drivers. To keep from having to update all drivers, use a #define to keep the old name working, and only update the virtual drivers in this patch. Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768505.1458396.17518057312953572912.stgit@toke.dk
2020-01-16 15:14:45 +00:00
struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
xdp: Move devmap bulk queue into struct net_device Commit 96360004b862 ("xdp: Make devmap flush_list common for all map instances"), changed devmap flushing to be a global operation instead of a per-map operation. However, the queue structure used for bulking was still allocated as part of the containing map. This patch moves the devmap bulk queue into struct net_device. The motivation for this is reusing it for the non-map variant of XDP_REDIRECT, which will be changed in a subsequent commit. To avoid other fields of struct net_device moving to different cache lines, we also move a couple of other members around. We defer the actual allocation of the bulk queue structure until the NETDEV_REGISTER notification devmap.c. This makes it possible to check for ndo_xdp_xmit support before allocating the structure, which is not possible at the time struct net_device is allocated. However, we keep the freeing in free_netdev() to avoid adding another RCU callback on NETDEV_UNREGISTER. Because of this change, we lose the reference back to the map that originated the redirect, so change the tracepoint to always return 0 as the map ID and index. Otherwise no functional change is intended with this patch. After this patch, the relevant part of struct net_device looks like this, according to pahole: /* --- cacheline 14 boundary (896 bytes) --- */ struct netdev_queue * _tx __attribute__((__aligned__(64))); /* 896 8 */ unsigned int num_tx_queues; /* 904 4 */ unsigned int real_num_tx_queues; /* 908 4 */ struct Qdisc * qdisc; /* 912 8 */ unsigned int tx_queue_len; /* 920 4 */ spinlock_t tx_global_lock; /* 924 4 */ struct xdp_dev_bulk_queue * xdp_bulkq; /* 928 8 */ struct xps_dev_maps * xps_cpus_map; /* 936 8 */ struct xps_dev_maps * xps_rxqs_map; /* 944 8 */ struct mini_Qdisc * miniq_egress; /* 952 8 */ /* --- cacheline 15 boundary (960 bytes) --- */ struct hlist_head qdisc_hash[16]; /* 960 128 */ /* --- cacheline 17 boundary (1088 bytes) --- */ struct timer_list watchdog_timer; /* 1088 40 */ /* XXX last struct has 4 bytes of padding */ int watchdog_timeo; /* 1128 4 */ /* XXX 4 bytes hole, try to pack */ struct list_head todo_list; /* 1136 16 */ /* --- cacheline 18 boundary (1152 bytes) --- */ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Björn Töpel <bjorn.topel@intel.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768397.1458396.12673224324627072349.stgit@toke.dk
2020-01-16 15:14:44 +00:00
struct xdp_dev_bulk_queue *bq, *tmp;
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
bq_xmit_all(bq, XDP_XMIT_FLUSH);
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
bq->dev_rx = NULL;
bq->xdp_prog = NULL;
__list_del_clearprev(&bq->flush_node);
}
}
xdp: Add proper __rcu annotations to redirect map entries XDP_REDIRECT works by a three-step process: the bpf_redirect() and bpf_redirect_map() helpers will lookup the target of the redirect and store it (along with some other metadata) in a per-CPU struct bpf_redirect_info. Next, when the program returns the XDP_REDIRECT return code, the driver will call xdp_do_redirect() which will use the information thus stored to actually enqueue the frame into a bulk queue structure (that differs slightly by map type, but shares the same principle). Finally, before exiting its NAPI poll loop, the driver will call xdp_do_flush(), which will flush all the different bulk queues, thus completing the redirect. Pointers to the map entries will be kept around for this whole sequence of steps, protected by RCU. However, there is no top-level rcu_read_lock() in the core code; instead drivers add their own rcu_read_lock() around the XDP portions of the code, but somewhat inconsistently as Martin discovered[0]. However, things still work because everything happens inside a single NAPI poll sequence, which means it's between a pair of calls to local_bh_disable()/local_bh_enable(). So Paul suggested[1] that we could document this intention by using rcu_dereference_check() with rcu_read_lock_bh_held() as a second parameter, thus allowing sparse and lockdep to verify that everything is done correctly. This patch does just that: we add an __rcu annotation to the map entry pointers and remove the various comments explaining the NAPI poll assurance strewn through devmap.c in favour of a longer explanation in filter.c. The goal is to have one coherent documentation of the entire flow, and rely on the RCU annotations as a "standard" way of communicating the flow in the map code (which can additionally be understood by sparse and lockdep). The RCU annotation replacements result in a fairly straight-forward replacement where READ_ONCE() becomes rcu_dereference_check(), WRITE_ONCE() becomes rcu_assign_pointer() and xchg() and cmpxchg() gets wrapped in the proper constructs to cast the pointer back and forth between __rcu and __kernel address space (for the benefit of sparse). The one complication is that xskmap has a few constructions where double-pointers are passed back and forth; these simply all gain __rcu annotations, and only the final reference/dereference to the inner-most pointer gets changed. With this, everything can be run through sparse without eliciting complaints, and lockdep can verify correctness even without the use of rcu_read_lock() in the drivers. Subsequent patches will clean these up from the drivers. [0] https://lore.kernel.org/bpf/20210415173551.7ma4slcbqeyiba2r@kafai-mbp.dhcp.thefacebook.com/ [1] https://lore.kernel.org/bpf/20210419165837.GA975577@paulmck-ThinkPad-P17-Gen-1/ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20210624160609.292325-6-toke@redhat.com
2021-06-24 16:05:55 +00:00
/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
* by local_bh_disable() (from XDP calls inside NAPI). The
* rcu_read_lock_bh_held() below makes lockdep accept both.
*/
static void *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
{
struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
struct bpf_dtab_netdev *obj;
if (key >= map->max_entries)
return NULL;
xdp: Add proper __rcu annotations to redirect map entries XDP_REDIRECT works by a three-step process: the bpf_redirect() and bpf_redirect_map() helpers will lookup the target of the redirect and store it (along with some other metadata) in a per-CPU struct bpf_redirect_info. Next, when the program returns the XDP_REDIRECT return code, the driver will call xdp_do_redirect() which will use the information thus stored to actually enqueue the frame into a bulk queue structure (that differs slightly by map type, but shares the same principle). Finally, before exiting its NAPI poll loop, the driver will call xdp_do_flush(), which will flush all the different bulk queues, thus completing the redirect. Pointers to the map entries will be kept around for this whole sequence of steps, protected by RCU. However, there is no top-level rcu_read_lock() in the core code; instead drivers add their own rcu_read_lock() around the XDP portions of the code, but somewhat inconsistently as Martin discovered[0]. However, things still work because everything happens inside a single NAPI poll sequence, which means it's between a pair of calls to local_bh_disable()/local_bh_enable(). So Paul suggested[1] that we could document this intention by using rcu_dereference_check() with rcu_read_lock_bh_held() as a second parameter, thus allowing sparse and lockdep to verify that everything is done correctly. This patch does just that: we add an __rcu annotation to the map entry pointers and remove the various comments explaining the NAPI poll assurance strewn through devmap.c in favour of a longer explanation in filter.c. The goal is to have one coherent documentation of the entire flow, and rely on the RCU annotations as a "standard" way of communicating the flow in the map code (which can additionally be understood by sparse and lockdep). The RCU annotation replacements result in a fairly straight-forward replacement where READ_ONCE() becomes rcu_dereference_check(), WRITE_ONCE() becomes rcu_assign_pointer() and xchg() and cmpxchg() gets wrapped in the proper constructs to cast the pointer back and forth between __rcu and __kernel address space (for the benefit of sparse). The one complication is that xskmap has a few constructions where double-pointers are passed back and forth; these simply all gain __rcu annotations, and only the final reference/dereference to the inner-most pointer gets changed. With this, everything can be run through sparse without eliciting complaints, and lockdep can verify correctness even without the use of rcu_read_lock() in the drivers. Subsequent patches will clean these up from the drivers. [0] https://lore.kernel.org/bpf/20210415173551.7ma4slcbqeyiba2r@kafai-mbp.dhcp.thefacebook.com/ [1] https://lore.kernel.org/bpf/20210419165837.GA975577@paulmck-ThinkPad-P17-Gen-1/ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20210624160609.292325-6-toke@redhat.com
2021-06-24 16:05:55 +00:00
obj = rcu_dereference_check(dtab->netdev_map[key],
rcu_read_lock_bh_held());
return obj;
}
xdp: Add proper __rcu annotations to redirect map entries XDP_REDIRECT works by a three-step process: the bpf_redirect() and bpf_redirect_map() helpers will lookup the target of the redirect and store it (along with some other metadata) in a per-CPU struct bpf_redirect_info. Next, when the program returns the XDP_REDIRECT return code, the driver will call xdp_do_redirect() which will use the information thus stored to actually enqueue the frame into a bulk queue structure (that differs slightly by map type, but shares the same principle). Finally, before exiting its NAPI poll loop, the driver will call xdp_do_flush(), which will flush all the different bulk queues, thus completing the redirect. Pointers to the map entries will be kept around for this whole sequence of steps, protected by RCU. However, there is no top-level rcu_read_lock() in the core code; instead drivers add their own rcu_read_lock() around the XDP portions of the code, but somewhat inconsistently as Martin discovered[0]. However, things still work because everything happens inside a single NAPI poll sequence, which means it's between a pair of calls to local_bh_disable()/local_bh_enable(). So Paul suggested[1] that we could document this intention by using rcu_dereference_check() with rcu_read_lock_bh_held() as a second parameter, thus allowing sparse and lockdep to verify that everything is done correctly. This patch does just that: we add an __rcu annotation to the map entry pointers and remove the various comments explaining the NAPI poll assurance strewn through devmap.c in favour of a longer explanation in filter.c. The goal is to have one coherent documentation of the entire flow, and rely on the RCU annotations as a "standard" way of communicating the flow in the map code (which can additionally be understood by sparse and lockdep). The RCU annotation replacements result in a fairly straight-forward replacement where READ_ONCE() becomes rcu_dereference_check(), WRITE_ONCE() becomes rcu_assign_pointer() and xchg() and cmpxchg() gets wrapped in the proper constructs to cast the pointer back and forth between __rcu and __kernel address space (for the benefit of sparse). The one complication is that xskmap has a few constructions where double-pointers are passed back and forth; these simply all gain __rcu annotations, and only the final reference/dereference to the inner-most pointer gets changed. With this, everything can be run through sparse without eliciting complaints, and lockdep can verify correctness even without the use of rcu_read_lock() in the drivers. Subsequent patches will clean these up from the drivers. [0] https://lore.kernel.org/bpf/20210415173551.7ma4slcbqeyiba2r@kafai-mbp.dhcp.thefacebook.com/ [1] https://lore.kernel.org/bpf/20210419165837.GA975577@paulmck-ThinkPad-P17-Gen-1/ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20210624160609.292325-6-toke@redhat.com
2021-06-24 16:05:55 +00:00
/* Runs in NAPI, i.e., softirq under local_bh_disable(). Thus, safe percpu
* variable access, and map elements stick around. See comment above
* xdp_do_flush() in filter.c.
*/
static void bq_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
struct net_device *dev_rx, struct bpf_prog *xdp_prog)
{
xdp: Use bulking for non-map XDP_REDIRECT and consolidate code paths Since the bulk queue used by XDP_REDIRECT now lives in struct net_device, we can re-use the bulking for the non-map version of the bpf_redirect() helper. This is a simple matter of having xdp_do_redirect_slow() queue the frame on the bulk queue instead of sending it out with __bpf_tx_xdp(). Unfortunately we can't make the bpf_redirect() helper return an error if the ifindex doesn't exit (as bpf_redirect_map() does), because we don't have a reference to the network namespace of the ingress device at the time the helper is called. So we have to leave it as-is and keep the device lookup in xdp_do_redirect_slow(). Since this leaves less reason to have the non-map redirect code in a separate function, so we get rid of the xdp_do_redirect_slow() function entirely. This does lose us the tracepoint disambiguation, but fortunately the xdp_redirect and xdp_redirect_map tracepoints use the same tracepoint entry structures. This means both can contain a map index, so we can just amend the tracepoint definitions so we always emit the xdp_redirect(_err) tracepoints, but with the map ID only populated if a map is present. This means we retire the xdp_redirect_map(_err) tracepoints entirely, but keep the definitions around in case someone is still listening for them. With this change, the performance of the xdp_redirect sample program goes from 5Mpps to 8.4Mpps (a 68% increase). Since the flush functions are no longer map-specific, rename the flush() functions to drop _map from their names. One of the renamed functions is the xdp_do_flush_map() callback used in all the xdp-enabled drivers. To keep from having to update all drivers, use a #define to keep the old name working, and only update the virtual drivers in this patch. Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768505.1458396.17518057312953572912.stgit@toke.dk
2020-01-16 15:14:45 +00:00
struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
xdp: Move devmap bulk queue into struct net_device Commit 96360004b862 ("xdp: Make devmap flush_list common for all map instances"), changed devmap flushing to be a global operation instead of a per-map operation. However, the queue structure used for bulking was still allocated as part of the containing map. This patch moves the devmap bulk queue into struct net_device. The motivation for this is reusing it for the non-map variant of XDP_REDIRECT, which will be changed in a subsequent commit. To avoid other fields of struct net_device moving to different cache lines, we also move a couple of other members around. We defer the actual allocation of the bulk queue structure until the NETDEV_REGISTER notification devmap.c. This makes it possible to check for ndo_xdp_xmit support before allocating the structure, which is not possible at the time struct net_device is allocated. However, we keep the freeing in free_netdev() to avoid adding another RCU callback on NETDEV_UNREGISTER. Because of this change, we lose the reference back to the map that originated the redirect, so change the tracepoint to always return 0 as the map ID and index. Otherwise no functional change is intended with this patch. After this patch, the relevant part of struct net_device looks like this, according to pahole: /* --- cacheline 14 boundary (896 bytes) --- */ struct netdev_queue * _tx __attribute__((__aligned__(64))); /* 896 8 */ unsigned int num_tx_queues; /* 904 4 */ unsigned int real_num_tx_queues; /* 908 4 */ struct Qdisc * qdisc; /* 912 8 */ unsigned int tx_queue_len; /* 920 4 */ spinlock_t tx_global_lock; /* 924 4 */ struct xdp_dev_bulk_queue * xdp_bulkq; /* 928 8 */ struct xps_dev_maps * xps_cpus_map; /* 936 8 */ struct xps_dev_maps * xps_rxqs_map; /* 944 8 */ struct mini_Qdisc * miniq_egress; /* 952 8 */ /* --- cacheline 15 boundary (960 bytes) --- */ struct hlist_head qdisc_hash[16]; /* 960 128 */ /* --- cacheline 17 boundary (1088 bytes) --- */ struct timer_list watchdog_timer; /* 1088 40 */ /* XXX last struct has 4 bytes of padding */ int watchdog_timeo; /* 1128 4 */ /* XXX 4 bytes hole, try to pack */ struct list_head todo_list; /* 1136 16 */ /* --- cacheline 18 boundary (1152 bytes) --- */ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Björn Töpel <bjorn.topel@intel.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768397.1458396.12673224324627072349.stgit@toke.dk
2020-01-16 15:14:44 +00:00
struct xdp_dev_bulk_queue *bq = this_cpu_ptr(dev->xdp_bulkq);
if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
bq_xmit_all(bq, 0);
/* Ingress dev_rx will be the same for all xdp_frame's in
* bulk_queue, because bq stored per-CPU and must be flushed
* from net_device drivers NAPI func end.
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
*
* Do the same with xdp_prog and flush_list since these fields
* are only ever modified together.
*/
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
if (!bq->dev_rx) {
bq->dev_rx = dev_rx;
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
bq->xdp_prog = xdp_prog;
list_add(&bq->flush_node, flush_list);
}
bq->q[bq->count++] = xdpf;
}
static inline int __xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
struct net_device *dev_rx,
struct bpf_prog *xdp_prog)
{
int err;
if (!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
return -EOPNOTSUPP;
if (unlikely(!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
xdp_frame_has_frags(xdpf)))
return -EOPNOTSUPP;
err = xdp_ok_fwd_dev(dev, xdp_get_frame_len(xdpf));
if (unlikely(err))
return err;
bpf: Run devmap xdp_prog on flush instead of bulk enqueue This changes the devmap XDP program support to run the program when the bulk queue is flushed instead of before the frame is enqueued. This has a couple of benefits: - It "sorts" the packets by destination devmap entry, and then runs the same BPF program on all the packets in sequence. This ensures that we keep the XDP program and destination device properties hot in I-cache. - It makes the multicast implementation simpler because it can just enqueue packets using bq_enqueue() without having to deal with the devmap program at all. The drawback is that if the devmap program drops the packet, the enqueue step is redundant. However, arguably this is mostly visible in a micro-benchmark, and with more mixed traffic the I-cache benefit should win out. The performance impact of just this patch is as follows: Using 2 10Gb i40e NIC, redirecting one to another, or into a veth interface, which do XDP_DROP on veth peer. With xdp_redirect_map in sample/bpf, send pkts via pktgen cmd: ./pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -t 10 -s 64 There are about +/- 0.1M deviation for native testing, the performance improved for the base-case, but some drop back with xdp devmap prog attached. Version | Test | Generic | Native | Native + 2nd xdp_prog 5.12 rc4 | xdp_redirect_map i40e->i40e | 1.9M | 9.6M | 8.4M 5.12 rc4 | xdp_redirect_map i40e->veth | 1.7M | 11.7M | 9.8M 5.12 rc4 + patch | xdp_redirect_map i40e->i40e | 1.9M | 9.8M | 8.0M 5.12 rc4 + patch | xdp_redirect_map i40e->veth | 1.7M | 12.0M | 9.4M When bq_xmit_all() is called from bq_enqueue(), another packet will always be enqueued immediately after, so clearing dev_rx, xdp_prog and flush_node in bq_xmit_all() is redundant. Move the clear to __dev_flush(), and only check them once in bq_enqueue() since they are all modified together. This change also has the side effect of extending the lifetime of the RCU-protected xdp_prog that lives inside the devmap entries: Instead of just living for the duration of the XDP program invocation, the reference now lives all the way until the bq is flushed. This is safe because the bq flush happens at the end of the NAPI poll loop, so everything happens between a local_bh_disable()/local_bh_enable() pair. However, this is by no means obvious from looking at the call sites; in particular, some drivers have an additional rcu_read_lock() around only the XDP program invocation, which only confuses matters further. Cleaning this up will be done in a separate patch series. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-2-liuhangbin@gmail.com
2021-05-19 09:07:44 +00:00
bq_enqueue(dev, xdpf, dev_rx, xdp_prog);
return 0;
}
static u32 dev_map_bpf_prog_run_skb(struct sk_buff *skb, struct bpf_dtab_netdev *dst)
{
struct xdp_txq_info txq = { .dev = dst->dev };
struct xdp_buff xdp;
u32 act;
if (!dst->xdp_prog)
return XDP_PASS;
__skb_pull(skb, skb->mac_len);
xdp.txq = &txq;
act = bpf_prog_run_generic_xdp(skb, &xdp, dst->xdp_prog);
switch (act) {
case XDP_PASS:
__skb_push(skb, skb->mac_len);
break;
default:
bpf_warn_invalid_xdp_action(NULL, dst->xdp_prog, act);
fallthrough;
case XDP_ABORTED:
trace_xdp_exception(dst->dev, dst->xdp_prog, act);
fallthrough;
case XDP_DROP:
kfree_skb(skb);
break;
}
return act;
}
int dev_xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
xdp: Use bulking for non-map XDP_REDIRECT and consolidate code paths Since the bulk queue used by XDP_REDIRECT now lives in struct net_device, we can re-use the bulking for the non-map version of the bpf_redirect() helper. This is a simple matter of having xdp_do_redirect_slow() queue the frame on the bulk queue instead of sending it out with __bpf_tx_xdp(). Unfortunately we can't make the bpf_redirect() helper return an error if the ifindex doesn't exit (as bpf_redirect_map() does), because we don't have a reference to the network namespace of the ingress device at the time the helper is called. So we have to leave it as-is and keep the device lookup in xdp_do_redirect_slow(). Since this leaves less reason to have the non-map redirect code in a separate function, so we get rid of the xdp_do_redirect_slow() function entirely. This does lose us the tracepoint disambiguation, but fortunately the xdp_redirect and xdp_redirect_map tracepoints use the same tracepoint entry structures. This means both can contain a map index, so we can just amend the tracepoint definitions so we always emit the xdp_redirect(_err) tracepoints, but with the map ID only populated if a map is present. This means we retire the xdp_redirect_map(_err) tracepoints entirely, but keep the definitions around in case someone is still listening for them. With this change, the performance of the xdp_redirect sample program goes from 5Mpps to 8.4Mpps (a 68% increase). Since the flush functions are no longer map-specific, rename the flush() functions to drop _map from their names. One of the renamed functions is the xdp_do_flush_map() callback used in all the xdp-enabled drivers. To keep from having to update all drivers, use a #define to keep the old name working, and only update the virtual drivers in this patch. Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768505.1458396.17518057312953572912.stgit@toke.dk
2020-01-16 15:14:45 +00:00
struct net_device *dev_rx)
{
return __xdp_enqueue(dev, xdpf, dev_rx, NULL);
xdp: Use bulking for non-map XDP_REDIRECT and consolidate code paths Since the bulk queue used by XDP_REDIRECT now lives in struct net_device, we can re-use the bulking for the non-map version of the bpf_redirect() helper. This is a simple matter of having xdp_do_redirect_slow() queue the frame on the bulk queue instead of sending it out with __bpf_tx_xdp(). Unfortunately we can't make the bpf_redirect() helper return an error if the ifindex doesn't exit (as bpf_redirect_map() does), because we don't have a reference to the network namespace of the ingress device at the time the helper is called. So we have to leave it as-is and keep the device lookup in xdp_do_redirect_slow(). Since this leaves less reason to have the non-map redirect code in a separate function, so we get rid of the xdp_do_redirect_slow() function entirely. This does lose us the tracepoint disambiguation, but fortunately the xdp_redirect and xdp_redirect_map tracepoints use the same tracepoint entry structures. This means both can contain a map index, so we can just amend the tracepoint definitions so we always emit the xdp_redirect(_err) tracepoints, but with the map ID only populated if a map is present. This means we retire the xdp_redirect_map(_err) tracepoints entirely, but keep the definitions around in case someone is still listening for them. With this change, the performance of the xdp_redirect sample program goes from 5Mpps to 8.4Mpps (a 68% increase). Since the flush functions are no longer map-specific, rename the flush() functions to drop _map from their names. One of the renamed functions is the xdp_do_flush_map() callback used in all the xdp-enabled drivers. To keep from having to update all drivers, use a #define to keep the old name working, and only update the virtual drivers in this patch. Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768505.1458396.17518057312953572912.stgit@toke.dk
2020-01-16 15:14:45 +00:00
}
int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_frame *xdpf,
xdp: Use bulking for non-map XDP_REDIRECT and consolidate code paths Since the bulk queue used by XDP_REDIRECT now lives in struct net_device, we can re-use the bulking for the non-map version of the bpf_redirect() helper. This is a simple matter of having xdp_do_redirect_slow() queue the frame on the bulk queue instead of sending it out with __bpf_tx_xdp(). Unfortunately we can't make the bpf_redirect() helper return an error if the ifindex doesn't exit (as bpf_redirect_map() does), because we don't have a reference to the network namespace of the ingress device at the time the helper is called. So we have to leave it as-is and keep the device lookup in xdp_do_redirect_slow(). Since this leaves less reason to have the non-map redirect code in a separate function, so we get rid of the xdp_do_redirect_slow() function entirely. This does lose us the tracepoint disambiguation, but fortunately the xdp_redirect and xdp_redirect_map tracepoints use the same tracepoint entry structures. This means both can contain a map index, so we can just amend the tracepoint definitions so we always emit the xdp_redirect(_err) tracepoints, but with the map ID only populated if a map is present. This means we retire the xdp_redirect_map(_err) tracepoints entirely, but keep the definitions around in case someone is still listening for them. With this change, the performance of the xdp_redirect sample program goes from 5Mpps to 8.4Mpps (a 68% increase). Since the flush functions are no longer map-specific, rename the flush() functions to drop _map from their names. One of the renamed functions is the xdp_do_flush_map() callback used in all the xdp-enabled drivers. To keep from having to update all drivers, use a #define to keep the old name working, and only update the virtual drivers in this patch. Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768505.1458396.17518057312953572912.stgit@toke.dk
2020-01-16 15:14:45 +00:00
struct net_device *dev_rx)
{
struct net_device *dev = dst->dev;
return __xdp_enqueue(dev, xdpf, dev_rx, dst->xdp_prog);
xdp: Use bulking for non-map XDP_REDIRECT and consolidate code paths Since the bulk queue used by XDP_REDIRECT now lives in struct net_device, we can re-use the bulking for the non-map version of the bpf_redirect() helper. This is a simple matter of having xdp_do_redirect_slow() queue the frame on the bulk queue instead of sending it out with __bpf_tx_xdp(). Unfortunately we can't make the bpf_redirect() helper return an error if the ifindex doesn't exit (as bpf_redirect_map() does), because we don't have a reference to the network namespace of the ingress device at the time the helper is called. So we have to leave it as-is and keep the device lookup in xdp_do_redirect_slow(). Since this leaves less reason to have the non-map redirect code in a separate function, so we get rid of the xdp_do_redirect_slow() function entirely. This does lose us the tracepoint disambiguation, but fortunately the xdp_redirect and xdp_redirect_map tracepoints use the same tracepoint entry structures. This means both can contain a map index, so we can just amend the tracepoint definitions so we always emit the xdp_redirect(_err) tracepoints, but with the map ID only populated if a map is present. This means we retire the xdp_redirect_map(_err) tracepoints entirely, but keep the definitions around in case someone is still listening for them. With this change, the performance of the xdp_redirect sample program goes from 5Mpps to 8.4Mpps (a 68% increase). Since the flush functions are no longer map-specific, rename the flush() functions to drop _map from their names. One of the renamed functions is the xdp_do_flush_map() callback used in all the xdp-enabled drivers. To keep from having to update all drivers, use a #define to keep the old name working, and only update the virtual drivers in this patch. Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768505.1458396.17518057312953572912.stgit@toke.dk
2020-01-16 15:14:45 +00:00
}
static bool is_valid_dst(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf)
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
{
if (!obj)
return false;
if (!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
return false;
if (unlikely(!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
xdp_frame_has_frags(xdpf)))
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
return false;
if (xdp_ok_fwd_dev(obj->dev, xdp_get_frame_len(xdpf)))
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
return false;
return true;
}
static int dev_map_enqueue_clone(struct bpf_dtab_netdev *obj,
struct net_device *dev_rx,
struct xdp_frame *xdpf)
{
struct xdp_frame *nxdpf;
nxdpf = xdpf_clone(xdpf);
if (!nxdpf)
return -ENOMEM;
bq_enqueue(obj->dev, nxdpf, dev_rx, obj->xdp_prog);
return 0;
}
static inline bool is_ifindex_excluded(int *excluded, int num_excluded, int ifindex)
{
while (num_excluded--) {
if (ifindex == excluded[num_excluded])
return true;
}
return false;
}
/* Get ifindex of each upper device. 'indexes' must be able to hold at
* least MAX_NEST_DEV elements.
* Returns the number of ifindexes added.
*/
static int get_upper_ifindexes(struct net_device *dev, int *indexes)
{
struct net_device *upper;
struct list_head *iter;
int n = 0;
netdev_for_each_upper_dev_rcu(dev, upper, iter) {
indexes[n++] = upper->ifindex;
}
return n;
}
int dev_map_enqueue_multi(struct xdp_frame *xdpf, struct net_device *dev_rx,
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
struct bpf_map *map, bool exclude_ingress)
{
struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
struct bpf_dtab_netdev *dst, *last_dst = NULL;
int excluded_devices[1+MAX_NEST_DEV];
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
struct hlist_head *head;
int num_excluded = 0;
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
unsigned int i;
int err;
if (exclude_ingress) {
num_excluded = get_upper_ifindexes(dev_rx, excluded_devices);
excluded_devices[num_excluded++] = dev_rx->ifindex;
}
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
for (i = 0; i < map->max_entries; i++) {
dst = rcu_dereference_check(dtab->netdev_map[i],
rcu_read_lock_bh_held());
if (!is_valid_dst(dst, xdpf))
continue;
if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
continue;
/* we only need n-1 clones; last_dst enqueued below */
if (!last_dst) {
last_dst = dst;
continue;
}
err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
if (err)
return err;
last_dst = dst;
}
} else { /* BPF_MAP_TYPE_DEVMAP_HASH */
for (i = 0; i < dtab->n_buckets; i++) {
head = dev_map_index_hash(dtab, i);
hlist_for_each_entry_rcu(dst, head, index_hlist,
lockdep_is_held(&dtab->index_lock)) {
if (!is_valid_dst(dst, xdpf))
continue;
if (is_ifindex_excluded(excluded_devices, num_excluded,
dst->dev->ifindex))
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
continue;
/* we only need n-1 clones; last_dst enqueued below */
if (!last_dst) {
last_dst = dst;
continue;
}
err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
if (err)
return err;
last_dst = dst;
}
}
}
/* consume the last copy of the frame */
if (last_dst)
bq_enqueue(last_dst->dev, xdpf, dev_rx, last_dst->xdp_prog);
else
xdp_return_frame_rx_napi(xdpf); /* dtab is empty */
return 0;
}
int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
struct bpf_prog *xdp_prog)
{
int err;
err = xdp_ok_fwd_dev(dst->dev, skb->len);
if (unlikely(err))
return err;
/* Redirect has already succeeded semantically at this point, so we just
* return 0 even if packet is dropped. Helper below takes care of
* freeing skb.
*/
if (dev_map_bpf_prog_run_skb(skb, dst) != XDP_PASS)
return 0;
skb->dev = dst->dev;
generic_xdp_tx(skb, xdp_prog);
return 0;
}
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
static int dev_map_redirect_clone(struct bpf_dtab_netdev *dst,
struct sk_buff *skb,
struct bpf_prog *xdp_prog)
{
struct sk_buff *nskb;
int err;
nskb = skb_clone(skb, GFP_ATOMIC);
if (!nskb)
return -ENOMEM;
err = dev_map_generic_redirect(dst, nskb, xdp_prog);
if (unlikely(err)) {
consume_skb(nskb);
return err;
}
return 0;
}
int dev_map_redirect_multi(struct net_device *dev, struct sk_buff *skb,
struct bpf_prog *xdp_prog, struct bpf_map *map,
bool exclude_ingress)
{
struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
struct bpf_dtab_netdev *dst, *last_dst = NULL;
int excluded_devices[1+MAX_NEST_DEV];
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
struct hlist_head *head;
struct hlist_node *next;
int num_excluded = 0;
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
unsigned int i;
int err;
if (exclude_ingress) {
num_excluded = get_upper_ifindexes(dev, excluded_devices);
excluded_devices[num_excluded++] = dev->ifindex;
}
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
for (i = 0; i < map->max_entries; i++) {
dst = rcu_dereference_check(dtab->netdev_map[i],
rcu_read_lock_bh_held());
if (!dst)
continue;
if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
continue;
/* we only need n-1 clones; last_dst enqueued below */
if (!last_dst) {
last_dst = dst;
continue;
}
err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
if (err)
return err;
last_dst = dst;
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
}
} else { /* BPF_MAP_TYPE_DEVMAP_HASH */
for (i = 0; i < dtab->n_buckets; i++) {
head = dev_map_index_hash(dtab, i);
hlist_for_each_entry_safe(dst, next, head, index_hlist) {
if (!dst)
continue;
if (is_ifindex_excluded(excluded_devices, num_excluded,
dst->dev->ifindex))
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
continue;
/* we only need n-1 clones; last_dst enqueued below */
if (!last_dst) {
last_dst = dst;
continue;
}
err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
if (err)
return err;
last_dst = dst;
}
}
}
/* consume the first skb and return */
if (last_dst)
return dev_map_generic_redirect(last_dst, skb, xdp_prog);
/* dtab is empty */
consume_skb(skb);
return 0;
}
static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
{
struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
return obj ? &obj->val : NULL;
}
static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
{
struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
*(u32 *)key);
return obj ? &obj->val : NULL;
}
static void __dev_map_entry_free(struct rcu_head *rcu)
{
struct bpf_dtab_netdev *dev;
dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
if (dev->xdp_prog)
bpf_prog_put(dev->xdp_prog);
dev_put(dev->dev);
kfree(dev);
}
static int dev_map_delete_elem(struct bpf_map *map, void *key)
{
struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
struct bpf_dtab_netdev *old_dev;
int k = *(u32 *)key;
if (k >= map->max_entries)
return -EINVAL;
xdp: Add proper __rcu annotations to redirect map entries XDP_REDIRECT works by a three-step process: the bpf_redirect() and bpf_redirect_map() helpers will lookup the target of the redirect and store it (along with some other metadata) in a per-CPU struct bpf_redirect_info. Next, when the program returns the XDP_REDIRECT return code, the driver will call xdp_do_redirect() which will use the information thus stored to actually enqueue the frame into a bulk queue structure (that differs slightly by map type, but shares the same principle). Finally, before exiting its NAPI poll loop, the driver will call xdp_do_flush(), which will flush all the different bulk queues, thus completing the redirect. Pointers to the map entries will be kept around for this whole sequence of steps, protected by RCU. However, there is no top-level rcu_read_lock() in the core code; instead drivers add their own rcu_read_lock() around the XDP portions of the code, but somewhat inconsistently as Martin discovered[0]. However, things still work because everything happens inside a single NAPI poll sequence, which means it's between a pair of calls to local_bh_disable()/local_bh_enable(). So Paul suggested[1] that we could document this intention by using rcu_dereference_check() with rcu_read_lock_bh_held() as a second parameter, thus allowing sparse and lockdep to verify that everything is done correctly. This patch does just that: we add an __rcu annotation to the map entry pointers and remove the various comments explaining the NAPI poll assurance strewn through devmap.c in favour of a longer explanation in filter.c. The goal is to have one coherent documentation of the entire flow, and rely on the RCU annotations as a "standard" way of communicating the flow in the map code (which can additionally be understood by sparse and lockdep). The RCU annotation replacements result in a fairly straight-forward replacement where READ_ONCE() becomes rcu_dereference_check(), WRITE_ONCE() becomes rcu_assign_pointer() and xchg() and cmpxchg() gets wrapped in the proper constructs to cast the pointer back and forth between __rcu and __kernel address space (for the benefit of sparse). The one complication is that xskmap has a few constructions where double-pointers are passed back and forth; these simply all gain __rcu annotations, and only the final reference/dereference to the inner-most pointer gets changed. With this, everything can be run through sparse without eliciting complaints, and lockdep can verify correctness even without the use of rcu_read_lock() in the drivers. Subsequent patches will clean these up from the drivers. [0] https://lore.kernel.org/bpf/20210415173551.7ma4slcbqeyiba2r@kafai-mbp.dhcp.thefacebook.com/ [1] https://lore.kernel.org/bpf/20210419165837.GA975577@paulmck-ThinkPad-P17-Gen-1/ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20210624160609.292325-6-toke@redhat.com
2021-06-24 16:05:55 +00:00
old_dev = unrcu_pointer(xchg(&dtab->netdev_map[k], NULL));
if (old_dev)
call_rcu(&old_dev->rcu, __dev_map_entry_free);
return 0;
}
static int dev_map_hash_delete_elem(struct bpf_map *map, void *key)
{
struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
struct bpf_dtab_netdev *old_dev;
int k = *(u32 *)key;
unsigned long flags;
int ret = -ENOENT;
spin_lock_irqsave(&dtab->index_lock, flags);
old_dev = __dev_map_hash_lookup_elem(map, k);
if (old_dev) {
dtab->items--;
hlist_del_init_rcu(&old_dev->index_hlist);
call_rcu(&old_dev->rcu, __dev_map_entry_free);
ret = 0;
}
spin_unlock_irqrestore(&dtab->index_lock, flags);
return ret;
}
static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
struct bpf_dtab *dtab,
struct bpf_devmap_val *val,
unsigned int idx)
{
struct bpf_prog *prog = NULL;
struct bpf_dtab_netdev *dev;
dev = bpf_map_kmalloc_node(&dtab->map, sizeof(*dev),
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
GFP_NOWAIT | __GFP_NOWARN,
dtab->map.numa_node);
if (!dev)
return ERR_PTR(-ENOMEM);
dev->dev = dev_get_by_index(net, val->ifindex);
if (!dev->dev)
goto err_out;
if (val->bpf_prog.fd > 0) {
prog = bpf_prog_get_type_dev(val->bpf_prog.fd,
BPF_PROG_TYPE_XDP, false);
if (IS_ERR(prog))
goto err_put_dev;
if (prog->expected_attach_type != BPF_XDP_DEVMAP ||
!bpf_prog_map_compatible(&dtab->map, prog))
goto err_put_prog;
}
dev->idx = idx;
dev->dtab = dtab;
if (prog) {
dev->xdp_prog = prog;
dev->val.bpf_prog.id = prog->aux->id;
} else {
dev->xdp_prog = NULL;
dev->val.bpf_prog.id = 0;
}
dev->val.ifindex = val->ifindex;
return dev;
err_put_prog:
bpf_prog_put(prog);
err_put_dev:
dev_put(dev->dev);
err_out:
kfree(dev);
return ERR_PTR(-EINVAL);
}
static int __dev_map_update_elem(struct net *net, struct bpf_map *map,
void *key, void *value, u64 map_flags)
{
struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
struct bpf_dtab_netdev *dev, *old_dev;
struct bpf_devmap_val val = {};
u32 i = *(u32 *)key;
if (unlikely(map_flags > BPF_EXIST))
return -EINVAL;
if (unlikely(i >= dtab->map.max_entries))
return -E2BIG;
if (unlikely(map_flags == BPF_NOEXIST))
return -EEXIST;
/* already verified value_size <= sizeof val */
memcpy(&val, value, map->value_size);
if (!val.ifindex) {
dev = NULL;
/* can not specify fd if ifindex is 0 */
if (val.bpf_prog.fd > 0)
return -EINVAL;
} else {
dev = __dev_map_alloc_node(net, dtab, &val, i);
if (IS_ERR(dev))
return PTR_ERR(dev);
}
/* Use call_rcu() here to ensure rcu critical sections have completed
* Remembering the driver side flush operation will happen before the
* net device is removed.
*/
xdp: Add proper __rcu annotations to redirect map entries XDP_REDIRECT works by a three-step process: the bpf_redirect() and bpf_redirect_map() helpers will lookup the target of the redirect and store it (along with some other metadata) in a per-CPU struct bpf_redirect_info. Next, when the program returns the XDP_REDIRECT return code, the driver will call xdp_do_redirect() which will use the information thus stored to actually enqueue the frame into a bulk queue structure (that differs slightly by map type, but shares the same principle). Finally, before exiting its NAPI poll loop, the driver will call xdp_do_flush(), which will flush all the different bulk queues, thus completing the redirect. Pointers to the map entries will be kept around for this whole sequence of steps, protected by RCU. However, there is no top-level rcu_read_lock() in the core code; instead drivers add their own rcu_read_lock() around the XDP portions of the code, but somewhat inconsistently as Martin discovered[0]. However, things still work because everything happens inside a single NAPI poll sequence, which means it's between a pair of calls to local_bh_disable()/local_bh_enable(). So Paul suggested[1] that we could document this intention by using rcu_dereference_check() with rcu_read_lock_bh_held() as a second parameter, thus allowing sparse and lockdep to verify that everything is done correctly. This patch does just that: we add an __rcu annotation to the map entry pointers and remove the various comments explaining the NAPI poll assurance strewn through devmap.c in favour of a longer explanation in filter.c. The goal is to have one coherent documentation of the entire flow, and rely on the RCU annotations as a "standard" way of communicating the flow in the map code (which can additionally be understood by sparse and lockdep). The RCU annotation replacements result in a fairly straight-forward replacement where READ_ONCE() becomes rcu_dereference_check(), WRITE_ONCE() becomes rcu_assign_pointer() and xchg() and cmpxchg() gets wrapped in the proper constructs to cast the pointer back and forth between __rcu and __kernel address space (for the benefit of sparse). The one complication is that xskmap has a few constructions where double-pointers are passed back and forth; these simply all gain __rcu annotations, and only the final reference/dereference to the inner-most pointer gets changed. With this, everything can be run through sparse without eliciting complaints, and lockdep can verify correctness even without the use of rcu_read_lock() in the drivers. Subsequent patches will clean these up from the drivers. [0] https://lore.kernel.org/bpf/20210415173551.7ma4slcbqeyiba2r@kafai-mbp.dhcp.thefacebook.com/ [1] https://lore.kernel.org/bpf/20210419165837.GA975577@paulmck-ThinkPad-P17-Gen-1/ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20210624160609.292325-6-toke@redhat.com
2021-06-24 16:05:55 +00:00
old_dev = unrcu_pointer(xchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev)));
if (old_dev)
call_rcu(&old_dev->rcu, __dev_map_entry_free);
return 0;
}
static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
u64 map_flags)
{
return __dev_map_update_elem(current->nsproxy->net_ns,
map, key, value, map_flags);
}
static int __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
void *key, void *value, u64 map_flags)
{
struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
struct bpf_dtab_netdev *dev, *old_dev;
struct bpf_devmap_val val = {};
u32 idx = *(u32 *)key;
unsigned long flags;
int err = -EEXIST;
/* already verified value_size <= sizeof val */
memcpy(&val, value, map->value_size);
if (unlikely(map_flags > BPF_EXIST || !val.ifindex))
return -EINVAL;
spin_lock_irqsave(&dtab->index_lock, flags);
old_dev = __dev_map_hash_lookup_elem(map, idx);
if (old_dev && (map_flags & BPF_NOEXIST))
goto out_err;
dev = __dev_map_alloc_node(net, dtab, &val, idx);
if (IS_ERR(dev)) {
err = PTR_ERR(dev);
goto out_err;
}
if (old_dev) {
hlist_del_rcu(&old_dev->index_hlist);
} else {
if (dtab->items >= dtab->map.max_entries) {
spin_unlock_irqrestore(&dtab->index_lock, flags);
call_rcu(&dev->rcu, __dev_map_entry_free);
return -E2BIG;
}
dtab->items++;
}
hlist_add_head_rcu(&dev->index_hlist,
dev_map_index_hash(dtab, idx));
spin_unlock_irqrestore(&dtab->index_lock, flags);
if (old_dev)
call_rcu(&old_dev->rcu, __dev_map_entry_free);
return 0;
out_err:
spin_unlock_irqrestore(&dtab->index_lock, flags);
return err;
}
static int dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
u64 map_flags)
{
return __dev_map_hash_update_elem(current->nsproxy->net_ns,
map, key, value, map_flags);
}
static int dev_map_redirect(struct bpf_map *map, u64 ifindex, u64 flags)
{
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
return __bpf_xdp_redirect_map(map, ifindex, flags,
BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
__dev_map_lookup_elem);
}
static int dev_hash_map_redirect(struct bpf_map *map, u64 ifindex, u64 flags)
{
xdp: Extend xdp_redirect_map with broadcast support This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to extend xdp_redirect_map for broadcast support. With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be excluded when do broadcasting. When getting the devices in dev hash map via dev_map_hash_get_next_key(), there is a possibility that we fall back to the first key when a device was removed. This will duplicate packets on some interfaces. So just walk the whole buckets to avoid this issue. For dev array map, we also walk the whole map to find valid interfaces. Function bpf_clear_redirect_map() was removed in commit ee75aef23afe ("bpf, xdp: Restructure redirect actions"). Add it back as we need to use ri->map again. With test topology: +-------------------+ +-------------------+ | Host A (i40e 10G) | ---------- | eno1(i40e 10G) | +-------------------+ | | | Host B | +-------------------+ | | | Host C (i40e 10G) | ---------- | eno2(i40e 10G) | +-------------------+ | | | +------+ | | veth0 -- | Peer | | | veth1 -- | | | | veth2 -- | NS | | | +------+ | +-------------------+ On Host A: # pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64 On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory): Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing. All the veth peers in the NS have a XDP_DROP program loaded. The forward_map max_entries in xdp_redirect_map_multi is modify to 4. Testing the performance impact on the regular xdp_redirect path with and without patch (to check impact of additional check for broadcast mode): 5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M 5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M 5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M 5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M Testing the performance when cloning packets with the redirect_map_multi test, using a redirect map size of 4, filled with 1-3 devices: 5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M 5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M 5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M Signed-off-by: Hangbin Liu <liuhangbin@gmail.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
2021-05-19 09:07:45 +00:00
return __bpf_xdp_redirect_map(map, ifindex, flags,
BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
__dev_map_hash_lookup_elem);
}
BTF_ID_LIST_SINGLE(dev_map_btf_ids, struct, bpf_dtab)
const struct bpf_map_ops dev_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 = dev_map_alloc,
.map_free = dev_map_free,
.map_get_next_key = dev_map_get_next_key,
.map_lookup_elem = dev_map_lookup_elem,
.map_update_elem = dev_map_update_elem,
.map_delete_elem = dev_map_delete_elem,
.map_check_btf = map_check_no_btf,
.map_btf_id = &dev_map_btf_ids[0],
.map_redirect = dev_map_redirect,
};
const struct bpf_map_ops dev_map_hash_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 = dev_map_alloc,
.map_free = dev_map_free,
.map_get_next_key = dev_map_hash_get_next_key,
.map_lookup_elem = dev_map_hash_lookup_elem,
.map_update_elem = dev_map_hash_update_elem,
.map_delete_elem = dev_map_hash_delete_elem,
.map_check_btf = map_check_no_btf,
.map_btf_id = &dev_map_btf_ids[0],
.map_redirect = dev_hash_map_redirect,
};
static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
struct net_device *netdev)
{
unsigned long flags;
u32 i;
spin_lock_irqsave(&dtab->index_lock, flags);
for (i = 0; i < dtab->n_buckets; i++) {
struct bpf_dtab_netdev *dev;
struct hlist_head *head;
struct hlist_node *next;
head = dev_map_index_hash(dtab, i);
hlist_for_each_entry_safe(dev, next, head, index_hlist) {
if (netdev != dev->dev)
continue;
dtab->items--;
hlist_del_rcu(&dev->index_hlist);
call_rcu(&dev->rcu, __dev_map_entry_free);
}
}
spin_unlock_irqrestore(&dtab->index_lock, flags);
}
static int dev_map_notification(struct notifier_block *notifier,
ulong event, void *ptr)
{
struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
struct bpf_dtab *dtab;
xdp: Move devmap bulk queue into struct net_device Commit 96360004b862 ("xdp: Make devmap flush_list common for all map instances"), changed devmap flushing to be a global operation instead of a per-map operation. However, the queue structure used for bulking was still allocated as part of the containing map. This patch moves the devmap bulk queue into struct net_device. The motivation for this is reusing it for the non-map variant of XDP_REDIRECT, which will be changed in a subsequent commit. To avoid other fields of struct net_device moving to different cache lines, we also move a couple of other members around. We defer the actual allocation of the bulk queue structure until the NETDEV_REGISTER notification devmap.c. This makes it possible to check for ndo_xdp_xmit support before allocating the structure, which is not possible at the time struct net_device is allocated. However, we keep the freeing in free_netdev() to avoid adding another RCU callback on NETDEV_UNREGISTER. Because of this change, we lose the reference back to the map that originated the redirect, so change the tracepoint to always return 0 as the map ID and index. Otherwise no functional change is intended with this patch. After this patch, the relevant part of struct net_device looks like this, according to pahole: /* --- cacheline 14 boundary (896 bytes) --- */ struct netdev_queue * _tx __attribute__((__aligned__(64))); /* 896 8 */ unsigned int num_tx_queues; /* 904 4 */ unsigned int real_num_tx_queues; /* 908 4 */ struct Qdisc * qdisc; /* 912 8 */ unsigned int tx_queue_len; /* 920 4 */ spinlock_t tx_global_lock; /* 924 4 */ struct xdp_dev_bulk_queue * xdp_bulkq; /* 928 8 */ struct xps_dev_maps * xps_cpus_map; /* 936 8 */ struct xps_dev_maps * xps_rxqs_map; /* 944 8 */ struct mini_Qdisc * miniq_egress; /* 952 8 */ /* --- cacheline 15 boundary (960 bytes) --- */ struct hlist_head qdisc_hash[16]; /* 960 128 */ /* --- cacheline 17 boundary (1088 bytes) --- */ struct timer_list watchdog_timer; /* 1088 40 */ /* XXX last struct has 4 bytes of padding */ int watchdog_timeo; /* 1128 4 */ /* XXX 4 bytes hole, try to pack */ struct list_head todo_list; /* 1136 16 */ /* --- cacheline 18 boundary (1152 bytes) --- */ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Björn Töpel <bjorn.topel@intel.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768397.1458396.12673224324627072349.stgit@toke.dk
2020-01-16 15:14:44 +00:00
int i, cpu;
switch (event) {
xdp: Move devmap bulk queue into struct net_device Commit 96360004b862 ("xdp: Make devmap flush_list common for all map instances"), changed devmap flushing to be a global operation instead of a per-map operation. However, the queue structure used for bulking was still allocated as part of the containing map. This patch moves the devmap bulk queue into struct net_device. The motivation for this is reusing it for the non-map variant of XDP_REDIRECT, which will be changed in a subsequent commit. To avoid other fields of struct net_device moving to different cache lines, we also move a couple of other members around. We defer the actual allocation of the bulk queue structure until the NETDEV_REGISTER notification devmap.c. This makes it possible to check for ndo_xdp_xmit support before allocating the structure, which is not possible at the time struct net_device is allocated. However, we keep the freeing in free_netdev() to avoid adding another RCU callback on NETDEV_UNREGISTER. Because of this change, we lose the reference back to the map that originated the redirect, so change the tracepoint to always return 0 as the map ID and index. Otherwise no functional change is intended with this patch. After this patch, the relevant part of struct net_device looks like this, according to pahole: /* --- cacheline 14 boundary (896 bytes) --- */ struct netdev_queue * _tx __attribute__((__aligned__(64))); /* 896 8 */ unsigned int num_tx_queues; /* 904 4 */ unsigned int real_num_tx_queues; /* 908 4 */ struct Qdisc * qdisc; /* 912 8 */ unsigned int tx_queue_len; /* 920 4 */ spinlock_t tx_global_lock; /* 924 4 */ struct xdp_dev_bulk_queue * xdp_bulkq; /* 928 8 */ struct xps_dev_maps * xps_cpus_map; /* 936 8 */ struct xps_dev_maps * xps_rxqs_map; /* 944 8 */ struct mini_Qdisc * miniq_egress; /* 952 8 */ /* --- cacheline 15 boundary (960 bytes) --- */ struct hlist_head qdisc_hash[16]; /* 960 128 */ /* --- cacheline 17 boundary (1088 bytes) --- */ struct timer_list watchdog_timer; /* 1088 40 */ /* XXX last struct has 4 bytes of padding */ int watchdog_timeo; /* 1128 4 */ /* XXX 4 bytes hole, try to pack */ struct list_head todo_list; /* 1136 16 */ /* --- cacheline 18 boundary (1152 bytes) --- */ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Björn Töpel <bjorn.topel@intel.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768397.1458396.12673224324627072349.stgit@toke.dk
2020-01-16 15:14:44 +00:00
case NETDEV_REGISTER:
if (!netdev->netdev_ops->ndo_xdp_xmit || netdev->xdp_bulkq)
break;
/* will be freed in free_netdev() */
netdev->xdp_bulkq = alloc_percpu(struct xdp_dev_bulk_queue);
xdp: Move devmap bulk queue into struct net_device Commit 96360004b862 ("xdp: Make devmap flush_list common for all map instances"), changed devmap flushing to be a global operation instead of a per-map operation. However, the queue structure used for bulking was still allocated as part of the containing map. This patch moves the devmap bulk queue into struct net_device. The motivation for this is reusing it for the non-map variant of XDP_REDIRECT, which will be changed in a subsequent commit. To avoid other fields of struct net_device moving to different cache lines, we also move a couple of other members around. We defer the actual allocation of the bulk queue structure until the NETDEV_REGISTER notification devmap.c. This makes it possible to check for ndo_xdp_xmit support before allocating the structure, which is not possible at the time struct net_device is allocated. However, we keep the freeing in free_netdev() to avoid adding another RCU callback on NETDEV_UNREGISTER. Because of this change, we lose the reference back to the map that originated the redirect, so change the tracepoint to always return 0 as the map ID and index. Otherwise no functional change is intended with this patch. After this patch, the relevant part of struct net_device looks like this, according to pahole: /* --- cacheline 14 boundary (896 bytes) --- */ struct netdev_queue * _tx __attribute__((__aligned__(64))); /* 896 8 */ unsigned int num_tx_queues; /* 904 4 */ unsigned int real_num_tx_queues; /* 908 4 */ struct Qdisc * qdisc; /* 912 8 */ unsigned int tx_queue_len; /* 920 4 */ spinlock_t tx_global_lock; /* 924 4 */ struct xdp_dev_bulk_queue * xdp_bulkq; /* 928 8 */ struct xps_dev_maps * xps_cpus_map; /* 936 8 */ struct xps_dev_maps * xps_rxqs_map; /* 944 8 */ struct mini_Qdisc * miniq_egress; /* 952 8 */ /* --- cacheline 15 boundary (960 bytes) --- */ struct hlist_head qdisc_hash[16]; /* 960 128 */ /* --- cacheline 17 boundary (1088 bytes) --- */ struct timer_list watchdog_timer; /* 1088 40 */ /* XXX last struct has 4 bytes of padding */ int watchdog_timeo; /* 1128 4 */ /* XXX 4 bytes hole, try to pack */ struct list_head todo_list; /* 1136 16 */ /* --- cacheline 18 boundary (1152 bytes) --- */ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: Björn Töpel <bjorn.topel@intel.com> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768397.1458396.12673224324627072349.stgit@toke.dk
2020-01-16 15:14:44 +00:00
if (!netdev->xdp_bulkq)
return NOTIFY_BAD;
for_each_possible_cpu(cpu)
per_cpu_ptr(netdev->xdp_bulkq, cpu)->dev = netdev;
break;
case NETDEV_UNREGISTER:
bpf: devmap fix mutex in rcu critical section Originally we used a mutex to protect concurrent devmap update and delete operations from racing with netdev unregister notifier callbacks. The notifier hook is needed because we increment the netdev ref count when a dev is added to the devmap. This ensures the netdev reference is valid in the datapath. However, we don't want to block unregister events, hence the initial mutex and notifier handler. The concern was in the notifier hook we search the map for dev entries that hold a refcnt on the net device being torn down. But, in order to do this we require two steps, (i) dereference the netdev: dev = rcu_dereference(map[i]) (ii) test ifindex: dev->ifindex == removing_ifindex and then finally we can swap in the NULL dev in the map via an xchg operation, xchg(map[i], NULL) The danger here is a concurrent update could run a different xchg op concurrently leading us to replace the new dev with a NULL dev incorrectly. CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) dev = rcu_dereference(map[i]) xchg(map[i]), new_dev); rcu_call(dev,...) xchg(map[i], NULL) The above flow would create the incorrect state with the dev reference in the update path being lost. To resolve this the original code used a mutex around the above block. However, updates, deletes, and lookups occur inside rcu critical sections so we can't use a mutex in this context safely. Fortunately, by writing slightly better code we can avoid the mutex altogether. If CPU 1 in the above example uses a cmpxchg and _only_ replaces the dev reference in the map when it is in fact the expected dev the race is removed completely. The two cases being illustrated here, first the race condition, CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) dev = rcu_dereference(map[i]) xchg(map[i]), new_dev); rcu_call(dev,...) odev = cmpxchg(map[i], dev, NULL) Now we can test the cmpxchg return value, detect odev != dev and abort. Or in the good case, CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) odev = cmpxchg(map[i], dev, NULL) [...] Now 'odev == dev' and we can do proper cleanup. And viola the original race we tried to solve with a mutex is corrected and the trace noted by Sasha below is resolved due to removal of the mutex. Note: When walking the devmap and removing dev references as needed we depend on the core to fail any calls to dev_get_by_index() using the ifindex of the device being removed. This way we do not race with the user while searching the devmap. Additionally, the mutex was also protecting list add/del/read on the list of maps in-use. This patch converts this to an RCU list and spinlock implementation. This protects the list from concurrent alloc/free operations. The notifier hook walks this list so it uses RCU read semantics. BUG: sleeping function called from invalid context at kernel/locking/mutex.c:747 in_atomic(): 1, irqs_disabled(): 0, pid: 16315, name: syz-executor1 1 lock held by syz-executor1/16315: #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] map_delete_elem kernel/bpf/syscall.c:577 [inline] #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] SYSC_bpf kernel/bpf/syscall.c:1427 [inline] #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] SyS_bpf+0x1d32/0x4ba0 kernel/bpf/syscall.c:1388 Fixes: 2ddf71e23cc2 ("net: add notifier hooks for devmap bpf map") Reported-by: Sasha Levin <alexander.levin@verizon.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: John Fastabend <john.fastabend@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-05 05:02:19 +00:00
/* This rcu_read_lock/unlock pair is needed because
* dev_map_list is an RCU list AND to ensure a delete
* operation does not free a netdev_map entry while we
* are comparing it against the netdev being unregistered.
*/
rcu_read_lock();
list_for_each_entry_rcu(dtab, &dev_map_list, list) {
if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
dev_map_hash_remove_netdev(dtab, netdev);
continue;
}
for (i = 0; i < dtab->map.max_entries; i++) {
bpf: devmap fix mutex in rcu critical section Originally we used a mutex to protect concurrent devmap update and delete operations from racing with netdev unregister notifier callbacks. The notifier hook is needed because we increment the netdev ref count when a dev is added to the devmap. This ensures the netdev reference is valid in the datapath. However, we don't want to block unregister events, hence the initial mutex and notifier handler. The concern was in the notifier hook we search the map for dev entries that hold a refcnt on the net device being torn down. But, in order to do this we require two steps, (i) dereference the netdev: dev = rcu_dereference(map[i]) (ii) test ifindex: dev->ifindex == removing_ifindex and then finally we can swap in the NULL dev in the map via an xchg operation, xchg(map[i], NULL) The danger here is a concurrent update could run a different xchg op concurrently leading us to replace the new dev with a NULL dev incorrectly. CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) dev = rcu_dereference(map[i]) xchg(map[i]), new_dev); rcu_call(dev,...) xchg(map[i], NULL) The above flow would create the incorrect state with the dev reference in the update path being lost. To resolve this the original code used a mutex around the above block. However, updates, deletes, and lookups occur inside rcu critical sections so we can't use a mutex in this context safely. Fortunately, by writing slightly better code we can avoid the mutex altogether. If CPU 1 in the above example uses a cmpxchg and _only_ replaces the dev reference in the map when it is in fact the expected dev the race is removed completely. The two cases being illustrated here, first the race condition, CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) dev = rcu_dereference(map[i]) xchg(map[i]), new_dev); rcu_call(dev,...) odev = cmpxchg(map[i], dev, NULL) Now we can test the cmpxchg return value, detect odev != dev and abort. Or in the good case, CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) odev = cmpxchg(map[i], dev, NULL) [...] Now 'odev == dev' and we can do proper cleanup. And viola the original race we tried to solve with a mutex is corrected and the trace noted by Sasha below is resolved due to removal of the mutex. Note: When walking the devmap and removing dev references as needed we depend on the core to fail any calls to dev_get_by_index() using the ifindex of the device being removed. This way we do not race with the user while searching the devmap. Additionally, the mutex was also protecting list add/del/read on the list of maps in-use. This patch converts this to an RCU list and spinlock implementation. This protects the list from concurrent alloc/free operations. The notifier hook walks this list so it uses RCU read semantics. BUG: sleeping function called from invalid context at kernel/locking/mutex.c:747 in_atomic(): 1, irqs_disabled(): 0, pid: 16315, name: syz-executor1 1 lock held by syz-executor1/16315: #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] map_delete_elem kernel/bpf/syscall.c:577 [inline] #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] SYSC_bpf kernel/bpf/syscall.c:1427 [inline] #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] SyS_bpf+0x1d32/0x4ba0 kernel/bpf/syscall.c:1388 Fixes: 2ddf71e23cc2 ("net: add notifier hooks for devmap bpf map") Reported-by: Sasha Levin <alexander.levin@verizon.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: John Fastabend <john.fastabend@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-05 05:02:19 +00:00
struct bpf_dtab_netdev *dev, *odev;
xdp: Add proper __rcu annotations to redirect map entries XDP_REDIRECT works by a three-step process: the bpf_redirect() and bpf_redirect_map() helpers will lookup the target of the redirect and store it (along with some other metadata) in a per-CPU struct bpf_redirect_info. Next, when the program returns the XDP_REDIRECT return code, the driver will call xdp_do_redirect() which will use the information thus stored to actually enqueue the frame into a bulk queue structure (that differs slightly by map type, but shares the same principle). Finally, before exiting its NAPI poll loop, the driver will call xdp_do_flush(), which will flush all the different bulk queues, thus completing the redirect. Pointers to the map entries will be kept around for this whole sequence of steps, protected by RCU. However, there is no top-level rcu_read_lock() in the core code; instead drivers add their own rcu_read_lock() around the XDP portions of the code, but somewhat inconsistently as Martin discovered[0]. However, things still work because everything happens inside a single NAPI poll sequence, which means it's between a pair of calls to local_bh_disable()/local_bh_enable(). So Paul suggested[1] that we could document this intention by using rcu_dereference_check() with rcu_read_lock_bh_held() as a second parameter, thus allowing sparse and lockdep to verify that everything is done correctly. This patch does just that: we add an __rcu annotation to the map entry pointers and remove the various comments explaining the NAPI poll assurance strewn through devmap.c in favour of a longer explanation in filter.c. The goal is to have one coherent documentation of the entire flow, and rely on the RCU annotations as a "standard" way of communicating the flow in the map code (which can additionally be understood by sparse and lockdep). The RCU annotation replacements result in a fairly straight-forward replacement where READ_ONCE() becomes rcu_dereference_check(), WRITE_ONCE() becomes rcu_assign_pointer() and xchg() and cmpxchg() gets wrapped in the proper constructs to cast the pointer back and forth between __rcu and __kernel address space (for the benefit of sparse). The one complication is that xskmap has a few constructions where double-pointers are passed back and forth; these simply all gain __rcu annotations, and only the final reference/dereference to the inner-most pointer gets changed. With this, everything can be run through sparse without eliciting complaints, and lockdep can verify correctness even without the use of rcu_read_lock() in the drivers. Subsequent patches will clean these up from the drivers. [0] https://lore.kernel.org/bpf/20210415173551.7ma4slcbqeyiba2r@kafai-mbp.dhcp.thefacebook.com/ [1] https://lore.kernel.org/bpf/20210419165837.GA975577@paulmck-ThinkPad-P17-Gen-1/ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20210624160609.292325-6-toke@redhat.com
2021-06-24 16:05:55 +00:00
dev = rcu_dereference(dtab->netdev_map[i]);
if (!dev || netdev != dev->dev)
continue;
xdp: Add proper __rcu annotations to redirect map entries XDP_REDIRECT works by a three-step process: the bpf_redirect() and bpf_redirect_map() helpers will lookup the target of the redirect and store it (along with some other metadata) in a per-CPU struct bpf_redirect_info. Next, when the program returns the XDP_REDIRECT return code, the driver will call xdp_do_redirect() which will use the information thus stored to actually enqueue the frame into a bulk queue structure (that differs slightly by map type, but shares the same principle). Finally, before exiting its NAPI poll loop, the driver will call xdp_do_flush(), which will flush all the different bulk queues, thus completing the redirect. Pointers to the map entries will be kept around for this whole sequence of steps, protected by RCU. However, there is no top-level rcu_read_lock() in the core code; instead drivers add their own rcu_read_lock() around the XDP portions of the code, but somewhat inconsistently as Martin discovered[0]. However, things still work because everything happens inside a single NAPI poll sequence, which means it's between a pair of calls to local_bh_disable()/local_bh_enable(). So Paul suggested[1] that we could document this intention by using rcu_dereference_check() with rcu_read_lock_bh_held() as a second parameter, thus allowing sparse and lockdep to verify that everything is done correctly. This patch does just that: we add an __rcu annotation to the map entry pointers and remove the various comments explaining the NAPI poll assurance strewn through devmap.c in favour of a longer explanation in filter.c. The goal is to have one coherent documentation of the entire flow, and rely on the RCU annotations as a "standard" way of communicating the flow in the map code (which can additionally be understood by sparse and lockdep). The RCU annotation replacements result in a fairly straight-forward replacement where READ_ONCE() becomes rcu_dereference_check(), WRITE_ONCE() becomes rcu_assign_pointer() and xchg() and cmpxchg() gets wrapped in the proper constructs to cast the pointer back and forth between __rcu and __kernel address space (for the benefit of sparse). The one complication is that xskmap has a few constructions where double-pointers are passed back and forth; these simply all gain __rcu annotations, and only the final reference/dereference to the inner-most pointer gets changed. With this, everything can be run through sparse without eliciting complaints, and lockdep can verify correctness even without the use of rcu_read_lock() in the drivers. Subsequent patches will clean these up from the drivers. [0] https://lore.kernel.org/bpf/20210415173551.7ma4slcbqeyiba2r@kafai-mbp.dhcp.thefacebook.com/ [1] https://lore.kernel.org/bpf/20210419165837.GA975577@paulmck-ThinkPad-P17-Gen-1/ Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Link: https://lore.kernel.org/bpf/20210624160609.292325-6-toke@redhat.com
2021-06-24 16:05:55 +00:00
odev = unrcu_pointer(cmpxchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev), NULL));
bpf: devmap fix mutex in rcu critical section Originally we used a mutex to protect concurrent devmap update and delete operations from racing with netdev unregister notifier callbacks. The notifier hook is needed because we increment the netdev ref count when a dev is added to the devmap. This ensures the netdev reference is valid in the datapath. However, we don't want to block unregister events, hence the initial mutex and notifier handler. The concern was in the notifier hook we search the map for dev entries that hold a refcnt on the net device being torn down. But, in order to do this we require two steps, (i) dereference the netdev: dev = rcu_dereference(map[i]) (ii) test ifindex: dev->ifindex == removing_ifindex and then finally we can swap in the NULL dev in the map via an xchg operation, xchg(map[i], NULL) The danger here is a concurrent update could run a different xchg op concurrently leading us to replace the new dev with a NULL dev incorrectly. CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) dev = rcu_dereference(map[i]) xchg(map[i]), new_dev); rcu_call(dev,...) xchg(map[i], NULL) The above flow would create the incorrect state with the dev reference in the update path being lost. To resolve this the original code used a mutex around the above block. However, updates, deletes, and lookups occur inside rcu critical sections so we can't use a mutex in this context safely. Fortunately, by writing slightly better code we can avoid the mutex altogether. If CPU 1 in the above example uses a cmpxchg and _only_ replaces the dev reference in the map when it is in fact the expected dev the race is removed completely. The two cases being illustrated here, first the race condition, CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) dev = rcu_dereference(map[i]) xchg(map[i]), new_dev); rcu_call(dev,...) odev = cmpxchg(map[i], dev, NULL) Now we can test the cmpxchg return value, detect odev != dev and abort. Or in the good case, CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) odev = cmpxchg(map[i], dev, NULL) [...] Now 'odev == dev' and we can do proper cleanup. And viola the original race we tried to solve with a mutex is corrected and the trace noted by Sasha below is resolved due to removal of the mutex. Note: When walking the devmap and removing dev references as needed we depend on the core to fail any calls to dev_get_by_index() using the ifindex of the device being removed. This way we do not race with the user while searching the devmap. Additionally, the mutex was also protecting list add/del/read on the list of maps in-use. This patch converts this to an RCU list and spinlock implementation. This protects the list from concurrent alloc/free operations. The notifier hook walks this list so it uses RCU read semantics. BUG: sleeping function called from invalid context at kernel/locking/mutex.c:747 in_atomic(): 1, irqs_disabled(): 0, pid: 16315, name: syz-executor1 1 lock held by syz-executor1/16315: #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] map_delete_elem kernel/bpf/syscall.c:577 [inline] #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] SYSC_bpf kernel/bpf/syscall.c:1427 [inline] #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] SyS_bpf+0x1d32/0x4ba0 kernel/bpf/syscall.c:1388 Fixes: 2ddf71e23cc2 ("net: add notifier hooks for devmap bpf map") Reported-by: Sasha Levin <alexander.levin@verizon.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: John Fastabend <john.fastabend@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-05 05:02:19 +00:00
if (dev == odev)
call_rcu(&dev->rcu,
__dev_map_entry_free);
}
}
bpf: devmap fix mutex in rcu critical section Originally we used a mutex to protect concurrent devmap update and delete operations from racing with netdev unregister notifier callbacks. The notifier hook is needed because we increment the netdev ref count when a dev is added to the devmap. This ensures the netdev reference is valid in the datapath. However, we don't want to block unregister events, hence the initial mutex and notifier handler. The concern was in the notifier hook we search the map for dev entries that hold a refcnt on the net device being torn down. But, in order to do this we require two steps, (i) dereference the netdev: dev = rcu_dereference(map[i]) (ii) test ifindex: dev->ifindex == removing_ifindex and then finally we can swap in the NULL dev in the map via an xchg operation, xchg(map[i], NULL) The danger here is a concurrent update could run a different xchg op concurrently leading us to replace the new dev with a NULL dev incorrectly. CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) dev = rcu_dereference(map[i]) xchg(map[i]), new_dev); rcu_call(dev,...) xchg(map[i], NULL) The above flow would create the incorrect state with the dev reference in the update path being lost. To resolve this the original code used a mutex around the above block. However, updates, deletes, and lookups occur inside rcu critical sections so we can't use a mutex in this context safely. Fortunately, by writing slightly better code we can avoid the mutex altogether. If CPU 1 in the above example uses a cmpxchg and _only_ replaces the dev reference in the map when it is in fact the expected dev the race is removed completely. The two cases being illustrated here, first the race condition, CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) dev = rcu_dereference(map[i]) xchg(map[i]), new_dev); rcu_call(dev,...) odev = cmpxchg(map[i], dev, NULL) Now we can test the cmpxchg return value, detect odev != dev and abort. Or in the good case, CPU 1 CPU 2 notifier hook bpf devmap update dev = rcu_dereference(map[i]) odev = cmpxchg(map[i], dev, NULL) [...] Now 'odev == dev' and we can do proper cleanup. And viola the original race we tried to solve with a mutex is corrected and the trace noted by Sasha below is resolved due to removal of the mutex. Note: When walking the devmap and removing dev references as needed we depend on the core to fail any calls to dev_get_by_index() using the ifindex of the device being removed. This way we do not race with the user while searching the devmap. Additionally, the mutex was also protecting list add/del/read on the list of maps in-use. This patch converts this to an RCU list and spinlock implementation. This protects the list from concurrent alloc/free operations. The notifier hook walks this list so it uses RCU read semantics. BUG: sleeping function called from invalid context at kernel/locking/mutex.c:747 in_atomic(): 1, irqs_disabled(): 0, pid: 16315, name: syz-executor1 1 lock held by syz-executor1/16315: #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] map_delete_elem kernel/bpf/syscall.c:577 [inline] #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] SYSC_bpf kernel/bpf/syscall.c:1427 [inline] #0: (rcu_read_lock){......}, at: [<ffffffff8c363bc2>] SyS_bpf+0x1d32/0x4ba0 kernel/bpf/syscall.c:1388 Fixes: 2ddf71e23cc2 ("net: add notifier hooks for devmap bpf map") Reported-by: Sasha Levin <alexander.levin@verizon.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: John Fastabend <john.fastabend@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-05 05:02:19 +00:00
rcu_read_unlock();
break;
default:
break;
}
return NOTIFY_OK;
}
static struct notifier_block dev_map_notifier = {
.notifier_call = dev_map_notification,
};
static int __init dev_map_init(void)
{
int cpu;
/* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
offsetof(struct _bpf_dtab_netdev, dev));
register_netdevice_notifier(&dev_map_notifier);
for_each_possible_cpu(cpu)
xdp: Use bulking for non-map XDP_REDIRECT and consolidate code paths Since the bulk queue used by XDP_REDIRECT now lives in struct net_device, we can re-use the bulking for the non-map version of the bpf_redirect() helper. This is a simple matter of having xdp_do_redirect_slow() queue the frame on the bulk queue instead of sending it out with __bpf_tx_xdp(). Unfortunately we can't make the bpf_redirect() helper return an error if the ifindex doesn't exit (as bpf_redirect_map() does), because we don't have a reference to the network namespace of the ingress device at the time the helper is called. So we have to leave it as-is and keep the device lookup in xdp_do_redirect_slow(). Since this leaves less reason to have the non-map redirect code in a separate function, so we get rid of the xdp_do_redirect_slow() function entirely. This does lose us the tracepoint disambiguation, but fortunately the xdp_redirect and xdp_redirect_map tracepoints use the same tracepoint entry structures. This means both can contain a map index, so we can just amend the tracepoint definitions so we always emit the xdp_redirect(_err) tracepoints, but with the map ID only populated if a map is present. This means we retire the xdp_redirect_map(_err) tracepoints entirely, but keep the definitions around in case someone is still listening for them. With this change, the performance of the xdp_redirect sample program goes from 5Mpps to 8.4Mpps (a 68% increase). Since the flush functions are no longer map-specific, rename the flush() functions to drop _map from their names. One of the renamed functions is the xdp_do_flush_map() callback used in all the xdp-enabled drivers. To keep from having to update all drivers, use a #define to keep the old name working, and only update the virtual drivers in this patch. Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Link: https://lore.kernel.org/bpf/157918768505.1458396.17518057312953572912.stgit@toke.dk
2020-01-16 15:14:45 +00:00
INIT_LIST_HEAD(&per_cpu(dev_flush_list, cpu));
return 0;
}
subsys_initcall(dev_map_init);