linux/net/ipv4/inetpeer.c

304 lines
8.3 KiB
C
Raw Normal View History

/*
* INETPEER - A storage for permanent information about peers
*
* This source is covered by the GNU GPL, the same as all kernel sources.
*
* Authors: Andrey V. Savochkin <saw@msu.ru>
*/
#include <linux/cache.h>
#include <linux/module.h>
#include <linux/types.h>
#include <linux/slab.h>
#include <linux/interrupt.h>
#include <linux/spinlock.h>
#include <linux/random.h>
#include <linux/timer.h>
#include <linux/time.h>
#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/net.h>
#include <linux/workqueue.h>
#include <net/ip.h>
#include <net/inetpeer.h>
#include <net/secure_seq.h>
/*
* Theory of operations.
* We keep one entry for each peer IP address. The nodes contains long-living
* information about the peer which doesn't depend on routes.
*
* Nodes are removed only when reference counter goes to 0.
* When it's happened the node may be removed when a sufficient amount of
* time has been passed since its last use. The less-recently-used entry can
* also be removed if the pool is overloaded i.e. if the total amount of
* entries is greater-or-equal than the threshold.
*
* Node pool is organised as an RB tree.
* Such an implementation has been chosen not just for fun. It's a way to
* prevent easy and efficient DoS attacks by creating hash collisions. A huge
* amount of long living nodes in a single hash slot would significantly delay
* lookups performed with disabled BHs.
*
* Serialisation issues.
inetpeer: RCU conversion inetpeer currently uses an AVL tree protected by an rwlock. It's possible to make most lookups use RCU 1) Add a struct rcu_head to struct inet_peer 2) add a lookup_rcu_bh() helper to perform lockless and opportunistic lookup. This is a normal function, not a macro like lookup(). 3) Add a limit to number of links followed by lookup_rcu_bh(). This is needed in case we fall in a loop. 4) add an smp_wmb() in link_to_pool() right before node insert. 5) make unlink_from_pool() use atomic_cmpxchg() to make sure it can take last reference to an inet_peer, since lockless readers could increase refcount, even while we hold peers.lock. 6) Delay struct inet_peer freeing after rcu grace period so that lookup_rcu_bh() cannot crash. 7) inet_getpeer() first attempts lockless lookup. Note this lookup can fail even if target is in AVL tree, but a concurrent writer can let tree in a non correct form. If this attemps fails, lock is taken a regular lookup is performed again. 8) convert peers.lock from rwlock to a spinlock 9) Remove SLAB_HWCACHE_ALIGN when peer_cachep is created, because rcu_head adds 16 bytes on 64bit arches, doubling effective size (64 -> 128 bytes) In a future patch, this is probably possible to revert this part, if rcu field is put in an union to share space with rid, ip_id_count, tcp_ts & tcp_ts_stamp. These fields being manipulated only with refcnt > 0. Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-06-15 08:23:14 +00:00
* 1. Nodes may appear in the tree only with the pool lock held.
* 2. Nodes may disappear from the tree only with the pool lock held
* AND reference count being 0.
* 3. Global variable peer_total is modified under the pool lock.
* 4. struct inet_peer fields modification:
* rb_node: pool lock
* refcnt: atomically against modifications on other CPU;
* usually under some other lock to prevent node disappearing
* daddr: unchangeable
*/
static struct kmem_cache *peer_cachep __ro_after_init;
void inet_peer_base_init(struct inet_peer_base *bp)
{
bp->rb_root = RB_ROOT;
seqlock_init(&bp->lock);
bp->total = 0;
}
EXPORT_SYMBOL_GPL(inet_peer_base_init);
#define PEER_MAX_GC 32
/* Exported for sysctl_net_ipv4. */
int inet_peer_threshold __read_mostly = 65536 + 128; /* start to throw entries more
* aggressively at this stage */
int inet_peer_minttl __read_mostly = 120 * HZ; /* TTL under high load: 120 sec */
int inet_peer_maxttl __read_mostly = 10 * 60 * HZ; /* usual time to live: 10 min */
/* Called from ip_output.c:ip_init */
void __init inet_initpeers(void)
{
struct sysinfo si;
/* Use the straight interface to information about memory. */
si_meminfo(&si);
/* The values below were suggested by Alexey Kuznetsov
* <kuznet@ms2.inr.ac.ru>. I don't have any opinion about the values
* myself. --SAW
*/
if (si.totalram <= (32768*1024)/PAGE_SIZE)
inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
if (si.totalram <= (16384*1024)/PAGE_SIZE)
inet_peer_threshold >>= 1; /* about 512KB */
if (si.totalram <= (8192*1024)/PAGE_SIZE)
inet_peer_threshold >>= 2; /* about 128KB */
peer_cachep = kmem_cache_create("inet_peer_cache",
sizeof(struct inet_peer),
0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
NULL);
}
/* Called with rcu_read_lock() or base->lock held */
static struct inet_peer *lookup(const struct inetpeer_addr *daddr,
struct inet_peer_base *base,
unsigned int seq,
struct inet_peer *gc_stack[],
unsigned int *gc_cnt,
struct rb_node **parent_p,
struct rb_node ***pp_p)
inetpeer: RCU conversion inetpeer currently uses an AVL tree protected by an rwlock. It's possible to make most lookups use RCU 1) Add a struct rcu_head to struct inet_peer 2) add a lookup_rcu_bh() helper to perform lockless and opportunistic lookup. This is a normal function, not a macro like lookup(). 3) Add a limit to number of links followed by lookup_rcu_bh(). This is needed in case we fall in a loop. 4) add an smp_wmb() in link_to_pool() right before node insert. 5) make unlink_from_pool() use atomic_cmpxchg() to make sure it can take last reference to an inet_peer, since lockless readers could increase refcount, even while we hold peers.lock. 6) Delay struct inet_peer freeing after rcu grace period so that lookup_rcu_bh() cannot crash. 7) inet_getpeer() first attempts lockless lookup. Note this lookup can fail even if target is in AVL tree, but a concurrent writer can let tree in a non correct form. If this attemps fails, lock is taken a regular lookup is performed again. 8) convert peers.lock from rwlock to a spinlock 9) Remove SLAB_HWCACHE_ALIGN when peer_cachep is created, because rcu_head adds 16 bytes on 64bit arches, doubling effective size (64 -> 128 bytes) In a future patch, this is probably possible to revert this part, if rcu field is put in an union to share space with rid, ip_id_count, tcp_ts & tcp_ts_stamp. These fields being manipulated only with refcnt > 0. Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-06-15 08:23:14 +00:00
{
struct rb_node **pp, *parent, *next;
struct inet_peer *p;
pp = &base->rb_root.rb_node;
parent = NULL;
while (1) {
int cmp;
inetpeer: RCU conversion inetpeer currently uses an AVL tree protected by an rwlock. It's possible to make most lookups use RCU 1) Add a struct rcu_head to struct inet_peer 2) add a lookup_rcu_bh() helper to perform lockless and opportunistic lookup. This is a normal function, not a macro like lookup(). 3) Add a limit to number of links followed by lookup_rcu_bh(). This is needed in case we fall in a loop. 4) add an smp_wmb() in link_to_pool() right before node insert. 5) make unlink_from_pool() use atomic_cmpxchg() to make sure it can take last reference to an inet_peer, since lockless readers could increase refcount, even while we hold peers.lock. 6) Delay struct inet_peer freeing after rcu grace period so that lookup_rcu_bh() cannot crash. 7) inet_getpeer() first attempts lockless lookup. Note this lookup can fail even if target is in AVL tree, but a concurrent writer can let tree in a non correct form. If this attemps fails, lock is taken a regular lookup is performed again. 8) convert peers.lock from rwlock to a spinlock 9) Remove SLAB_HWCACHE_ALIGN when peer_cachep is created, because rcu_head adds 16 bytes on 64bit arches, doubling effective size (64 -> 128 bytes) In a future patch, this is probably possible to revert this part, if rcu field is put in an union to share space with rid, ip_id_count, tcp_ts & tcp_ts_stamp. These fields being manipulated only with refcnt > 0. Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-06-15 08:23:14 +00:00
next = rcu_dereference_raw(*pp);
if (!next)
break;
parent = next;
p = rb_entry(parent, struct inet_peer, rb_node);
cmp = inetpeer_addr_cmp(daddr, &p->daddr);
if (cmp == 0) {
if (!refcount_inc_not_zero(&p->refcnt))
break;
return p;
}
if (gc_stack) {
if (*gc_cnt < PEER_MAX_GC)
gc_stack[(*gc_cnt)++] = p;
} else if (unlikely(read_seqretry(&base->lock, seq))) {
break;
inetpeer: RCU conversion inetpeer currently uses an AVL tree protected by an rwlock. It's possible to make most lookups use RCU 1) Add a struct rcu_head to struct inet_peer 2) add a lookup_rcu_bh() helper to perform lockless and opportunistic lookup. This is a normal function, not a macro like lookup(). 3) Add a limit to number of links followed by lookup_rcu_bh(). This is needed in case we fall in a loop. 4) add an smp_wmb() in link_to_pool() right before node insert. 5) make unlink_from_pool() use atomic_cmpxchg() to make sure it can take last reference to an inet_peer, since lockless readers could increase refcount, even while we hold peers.lock. 6) Delay struct inet_peer freeing after rcu grace period so that lookup_rcu_bh() cannot crash. 7) inet_getpeer() first attempts lockless lookup. Note this lookup can fail even if target is in AVL tree, but a concurrent writer can let tree in a non correct form. If this attemps fails, lock is taken a regular lookup is performed again. 8) convert peers.lock from rwlock to a spinlock 9) Remove SLAB_HWCACHE_ALIGN when peer_cachep is created, because rcu_head adds 16 bytes on 64bit arches, doubling effective size (64 -> 128 bytes) In a future patch, this is probably possible to revert this part, if rcu field is put in an union to share space with rid, ip_id_count, tcp_ts & tcp_ts_stamp. These fields being manipulated only with refcnt > 0. Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-06-15 08:23:14 +00:00
}
if (cmp == -1)
pp = &next->rb_left;
inetpeer: RCU conversion inetpeer currently uses an AVL tree protected by an rwlock. It's possible to make most lookups use RCU 1) Add a struct rcu_head to struct inet_peer 2) add a lookup_rcu_bh() helper to perform lockless and opportunistic lookup. This is a normal function, not a macro like lookup(). 3) Add a limit to number of links followed by lookup_rcu_bh(). This is needed in case we fall in a loop. 4) add an smp_wmb() in link_to_pool() right before node insert. 5) make unlink_from_pool() use atomic_cmpxchg() to make sure it can take last reference to an inet_peer, since lockless readers could increase refcount, even while we hold peers.lock. 6) Delay struct inet_peer freeing after rcu grace period so that lookup_rcu_bh() cannot crash. 7) inet_getpeer() first attempts lockless lookup. Note this lookup can fail even if target is in AVL tree, but a concurrent writer can let tree in a non correct form. If this attemps fails, lock is taken a regular lookup is performed again. 8) convert peers.lock from rwlock to a spinlock 9) Remove SLAB_HWCACHE_ALIGN when peer_cachep is created, because rcu_head adds 16 bytes on 64bit arches, doubling effective size (64 -> 128 bytes) In a future patch, this is probably possible to revert this part, if rcu field is put in an union to share space with rid, ip_id_count, tcp_ts & tcp_ts_stamp. These fields being manipulated only with refcnt > 0. Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-06-15 08:23:14 +00:00
else
pp = &next->rb_right;
inetpeer: RCU conversion inetpeer currently uses an AVL tree protected by an rwlock. It's possible to make most lookups use RCU 1) Add a struct rcu_head to struct inet_peer 2) add a lookup_rcu_bh() helper to perform lockless and opportunistic lookup. This is a normal function, not a macro like lookup(). 3) Add a limit to number of links followed by lookup_rcu_bh(). This is needed in case we fall in a loop. 4) add an smp_wmb() in link_to_pool() right before node insert. 5) make unlink_from_pool() use atomic_cmpxchg() to make sure it can take last reference to an inet_peer, since lockless readers could increase refcount, even while we hold peers.lock. 6) Delay struct inet_peer freeing after rcu grace period so that lookup_rcu_bh() cannot crash. 7) inet_getpeer() first attempts lockless lookup. Note this lookup can fail even if target is in AVL tree, but a concurrent writer can let tree in a non correct form. If this attemps fails, lock is taken a regular lookup is performed again. 8) convert peers.lock from rwlock to a spinlock 9) Remove SLAB_HWCACHE_ALIGN when peer_cachep is created, because rcu_head adds 16 bytes on 64bit arches, doubling effective size (64 -> 128 bytes) In a future patch, this is probably possible to revert this part, if rcu field is put in an union to share space with rid, ip_id_count, tcp_ts & tcp_ts_stamp. These fields being manipulated only with refcnt > 0. Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-06-15 08:23:14 +00:00
}
*parent_p = parent;
*pp_p = pp;
inetpeer: RCU conversion inetpeer currently uses an AVL tree protected by an rwlock. It's possible to make most lookups use RCU 1) Add a struct rcu_head to struct inet_peer 2) add a lookup_rcu_bh() helper to perform lockless and opportunistic lookup. This is a normal function, not a macro like lookup(). 3) Add a limit to number of links followed by lookup_rcu_bh(). This is needed in case we fall in a loop. 4) add an smp_wmb() in link_to_pool() right before node insert. 5) make unlink_from_pool() use atomic_cmpxchg() to make sure it can take last reference to an inet_peer, since lockless readers could increase refcount, even while we hold peers.lock. 6) Delay struct inet_peer freeing after rcu grace period so that lookup_rcu_bh() cannot crash. 7) inet_getpeer() first attempts lockless lookup. Note this lookup can fail even if target is in AVL tree, but a concurrent writer can let tree in a non correct form. If this attemps fails, lock is taken a regular lookup is performed again. 8) convert peers.lock from rwlock to a spinlock 9) Remove SLAB_HWCACHE_ALIGN when peer_cachep is created, because rcu_head adds 16 bytes on 64bit arches, doubling effective size (64 -> 128 bytes) In a future patch, this is probably possible to revert this part, if rcu field is put in an union to share space with rid, ip_id_count, tcp_ts & tcp_ts_stamp. These fields being manipulated only with refcnt > 0. Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-06-15 08:23:14 +00:00
return NULL;
}
static void inetpeer_free_rcu(struct rcu_head *head)
{
kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
}
/* perform garbage collect on all items stacked during a lookup */
static void inet_peer_gc(struct inet_peer_base *base,
struct inet_peer *gc_stack[],
unsigned int gc_cnt)
{
struct inet_peer *p;
__u32 delta, ttl;
int i;
if (base->total >= inet_peer_threshold)
ttl = 0; /* be aggressive */
else
ttl = inet_peer_maxttl
- (inet_peer_maxttl - inet_peer_minttl) / HZ *
base->total / inet_peer_threshold * HZ;
for (i = 0; i < gc_cnt; i++) {
p = gc_stack[i];
delta = (__u32)jiffies - p->dtime;
if (delta < ttl || !refcount_dec_if_one(&p->refcnt))
gc_stack[i] = NULL;
}
for (i = 0; i < gc_cnt; i++) {
p = gc_stack[i];
if (p) {
rb_erase(&p->rb_node, &base->rb_root);
base->total--;
call_rcu(&p->rcu, inetpeer_free_rcu);
}
}
}
struct inet_peer *inet_getpeer(struct inet_peer_base *base,
const struct inetpeer_addr *daddr,
int create)
{
struct inet_peer *p, *gc_stack[PEER_MAX_GC];
struct rb_node **pp, *parent;
unsigned int gc_cnt, seq;
int invalidated;
/* Attempt a lockless lookup first.
inetpeer: RCU conversion inetpeer currently uses an AVL tree protected by an rwlock. It's possible to make most lookups use RCU 1) Add a struct rcu_head to struct inet_peer 2) add a lookup_rcu_bh() helper to perform lockless and opportunistic lookup. This is a normal function, not a macro like lookup(). 3) Add a limit to number of links followed by lookup_rcu_bh(). This is needed in case we fall in a loop. 4) add an smp_wmb() in link_to_pool() right before node insert. 5) make unlink_from_pool() use atomic_cmpxchg() to make sure it can take last reference to an inet_peer, since lockless readers could increase refcount, even while we hold peers.lock. 6) Delay struct inet_peer freeing after rcu grace period so that lookup_rcu_bh() cannot crash. 7) inet_getpeer() first attempts lockless lookup. Note this lookup can fail even if target is in AVL tree, but a concurrent writer can let tree in a non correct form. If this attemps fails, lock is taken a regular lookup is performed again. 8) convert peers.lock from rwlock to a spinlock 9) Remove SLAB_HWCACHE_ALIGN when peer_cachep is created, because rcu_head adds 16 bytes on 64bit arches, doubling effective size (64 -> 128 bytes) In a future patch, this is probably possible to revert this part, if rcu field is put in an union to share space with rid, ip_id_count, tcp_ts & tcp_ts_stamp. These fields being manipulated only with refcnt > 0. Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-06-15 08:23:14 +00:00
* Because of a concurrent writer, we might not find an existing entry.
*/
rcu_read_lock();
seq = read_seqbegin(&base->lock);
p = lookup(daddr, base, seq, NULL, &gc_cnt, &parent, &pp);
invalidated = read_seqretry(&base->lock, seq);
rcu_read_unlock();
inetpeer: RCU conversion inetpeer currently uses an AVL tree protected by an rwlock. It's possible to make most lookups use RCU 1) Add a struct rcu_head to struct inet_peer 2) add a lookup_rcu_bh() helper to perform lockless and opportunistic lookup. This is a normal function, not a macro like lookup(). 3) Add a limit to number of links followed by lookup_rcu_bh(). This is needed in case we fall in a loop. 4) add an smp_wmb() in link_to_pool() right before node insert. 5) make unlink_from_pool() use atomic_cmpxchg() to make sure it can take last reference to an inet_peer, since lockless readers could increase refcount, even while we hold peers.lock. 6) Delay struct inet_peer freeing after rcu grace period so that lookup_rcu_bh() cannot crash. 7) inet_getpeer() first attempts lockless lookup. Note this lookup can fail even if target is in AVL tree, but a concurrent writer can let tree in a non correct form. If this attemps fails, lock is taken a regular lookup is performed again. 8) convert peers.lock from rwlock to a spinlock 9) Remove SLAB_HWCACHE_ALIGN when peer_cachep is created, because rcu_head adds 16 bytes on 64bit arches, doubling effective size (64 -> 128 bytes) In a future patch, this is probably possible to revert this part, if rcu field is put in an union to share space with rid, ip_id_count, tcp_ts & tcp_ts_stamp. These fields being manipulated only with refcnt > 0. Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-06-15 08:23:14 +00:00
if (p)
inetpeer: RCU conversion inetpeer currently uses an AVL tree protected by an rwlock. It's possible to make most lookups use RCU 1) Add a struct rcu_head to struct inet_peer 2) add a lookup_rcu_bh() helper to perform lockless and opportunistic lookup. This is a normal function, not a macro like lookup(). 3) Add a limit to number of links followed by lookup_rcu_bh(). This is needed in case we fall in a loop. 4) add an smp_wmb() in link_to_pool() right before node insert. 5) make unlink_from_pool() use atomic_cmpxchg() to make sure it can take last reference to an inet_peer, since lockless readers could increase refcount, even while we hold peers.lock. 6) Delay struct inet_peer freeing after rcu grace period so that lookup_rcu_bh() cannot crash. 7) inet_getpeer() first attempts lockless lookup. Note this lookup can fail even if target is in AVL tree, but a concurrent writer can let tree in a non correct form. If this attemps fails, lock is taken a regular lookup is performed again. 8) convert peers.lock from rwlock to a spinlock 9) Remove SLAB_HWCACHE_ALIGN when peer_cachep is created, because rcu_head adds 16 bytes on 64bit arches, doubling effective size (64 -> 128 bytes) In a future patch, this is probably possible to revert this part, if rcu field is put in an union to share space with rid, ip_id_count, tcp_ts & tcp_ts_stamp. These fields being manipulated only with refcnt > 0. Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-06-15 08:23:14 +00:00
return p;
/* If no writer did a change during our lookup, we can return early. */
if (!create && !invalidated)
return NULL;
inetpeer: RCU conversion inetpeer currently uses an AVL tree protected by an rwlock. It's possible to make most lookups use RCU 1) Add a struct rcu_head to struct inet_peer 2) add a lookup_rcu_bh() helper to perform lockless and opportunistic lookup. This is a normal function, not a macro like lookup(). 3) Add a limit to number of links followed by lookup_rcu_bh(). This is needed in case we fall in a loop. 4) add an smp_wmb() in link_to_pool() right before node insert. 5) make unlink_from_pool() use atomic_cmpxchg() to make sure it can take last reference to an inet_peer, since lockless readers could increase refcount, even while we hold peers.lock. 6) Delay struct inet_peer freeing after rcu grace period so that lookup_rcu_bh() cannot crash. 7) inet_getpeer() first attempts lockless lookup. Note this lookup can fail even if target is in AVL tree, but a concurrent writer can let tree in a non correct form. If this attemps fails, lock is taken a regular lookup is performed again. 8) convert peers.lock from rwlock to a spinlock 9) Remove SLAB_HWCACHE_ALIGN when peer_cachep is created, because rcu_head adds 16 bytes on 64bit arches, doubling effective size (64 -> 128 bytes) In a future patch, this is probably possible to revert this part, if rcu field is put in an union to share space with rid, ip_id_count, tcp_ts & tcp_ts_stamp. These fields being manipulated only with refcnt > 0. Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-06-15 08:23:14 +00:00
/* retry an exact lookup, taking the lock before.
* At least, nodes should be hot in our cache.
*/
parent = NULL;
write_seqlock_bh(&base->lock);
inetpeer: RCU conversion inetpeer currently uses an AVL tree protected by an rwlock. It's possible to make most lookups use RCU 1) Add a struct rcu_head to struct inet_peer 2) add a lookup_rcu_bh() helper to perform lockless and opportunistic lookup. This is a normal function, not a macro like lookup(). 3) Add a limit to number of links followed by lookup_rcu_bh(). This is needed in case we fall in a loop. 4) add an smp_wmb() in link_to_pool() right before node insert. 5) make unlink_from_pool() use atomic_cmpxchg() to make sure it can take last reference to an inet_peer, since lockless readers could increase refcount, even while we hold peers.lock. 6) Delay struct inet_peer freeing after rcu grace period so that lookup_rcu_bh() cannot crash. 7) inet_getpeer() first attempts lockless lookup. Note this lookup can fail even if target is in AVL tree, but a concurrent writer can let tree in a non correct form. If this attemps fails, lock is taken a regular lookup is performed again. 8) convert peers.lock from rwlock to a spinlock 9) Remove SLAB_HWCACHE_ALIGN when peer_cachep is created, because rcu_head adds 16 bytes on 64bit arches, doubling effective size (64 -> 128 bytes) In a future patch, this is probably possible to revert this part, if rcu field is put in an union to share space with rid, ip_id_count, tcp_ts & tcp_ts_stamp. These fields being manipulated only with refcnt > 0. Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-06-15 08:23:14 +00:00
gc_cnt = 0;
p = lookup(daddr, base, seq, gc_stack, &gc_cnt, &parent, &pp);
if (!p && create) {
p = kmem_cache_alloc(peer_cachep, GFP_ATOMIC);
if (p) {
p->daddr = *daddr;
inetpeer: fix uninit-value in inet_getpeer syzbot/KMSAN reported that p->dtime was read while it was not yet initialized in : delta = (__u32)jiffies - p->dtime; if (delta < ttl || !refcount_dec_if_one(&p->refcnt)) gc_stack[i] = NULL; This is a false positive, because the inetpeer wont be erased from rb-tree if the refcount_dec_if_one(&p->refcnt) does not succeed. And this wont happen before first inet_putpeer() call for this inetpeer has been done, and ->dtime field is written exactly before the refcount_dec_and_test(&p->refcnt). The KMSAN report was : BUG: KMSAN: uninit-value in inet_peer_gc net/ipv4/inetpeer.c:163 [inline] BUG: KMSAN: uninit-value in inet_getpeer+0x1567/0x1e70 net/ipv4/inetpeer.c:228 CPU: 0 PID: 9494 Comm: syz-executor5 Not tainted 4.16.0+ #82 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:17 [inline] dump_stack+0x185/0x1d0 lib/dump_stack.c:53 kmsan_report+0x142/0x240 mm/kmsan/kmsan.c:1067 __msan_warning_32+0x6c/0xb0 mm/kmsan/kmsan_instr.c:676 inet_peer_gc net/ipv4/inetpeer.c:163 [inline] inet_getpeer+0x1567/0x1e70 net/ipv4/inetpeer.c:228 inet_getpeer_v4 include/net/inetpeer.h:110 [inline] icmpv4_xrlim_allow net/ipv4/icmp.c:330 [inline] icmp_send+0x2b44/0x3050 net/ipv4/icmp.c:725 ip_options_compile+0x237c/0x29f0 net/ipv4/ip_options.c:472 ip_rcv_options net/ipv4/ip_input.c:284 [inline] ip_rcv_finish+0xda8/0x16d0 net/ipv4/ip_input.c:365 NF_HOOK include/linux/netfilter.h:288 [inline] ip_rcv+0x119d/0x16f0 net/ipv4/ip_input.c:493 __netif_receive_skb_core+0x47cf/0x4a80 net/core/dev.c:4562 __netif_receive_skb net/core/dev.c:4627 [inline] netif_receive_skb_internal+0x49d/0x630 net/core/dev.c:4701 netif_receive_skb+0x230/0x240 net/core/dev.c:4725 tun_rx_batched drivers/net/tun.c:1555 [inline] tun_get_user+0x6d88/0x7580 drivers/net/tun.c:1962 tun_chr_write_iter+0x1d4/0x330 drivers/net/tun.c:1990 do_iter_readv_writev+0x7bb/0x970 include/linux/fs.h:1776 do_iter_write+0x30d/0xd40 fs/read_write.c:932 vfs_writev fs/read_write.c:977 [inline] do_writev+0x3c9/0x830 fs/read_write.c:1012 SYSC_writev+0x9b/0xb0 fs/read_write.c:1085 SyS_writev+0x56/0x80 fs/read_write.c:1082 do_syscall_64+0x309/0x430 arch/x86/entry/common.c:287 entry_SYSCALL_64_after_hwframe+0x3d/0xa2 RIP: 0033:0x455111 RSP: 002b:00007fae0365cba0 EFLAGS: 00000293 ORIG_RAX: 0000000000000014 RAX: ffffffffffffffda RBX: 000000000000002e RCX: 0000000000455111 RDX: 0000000000000001 RSI: 00007fae0365cbf0 RDI: 00000000000000fc RBP: 0000000020000040 R08: 00000000000000fc R09: 0000000000000000 R10: 000000000000002e R11: 0000000000000293 R12: 00000000ffffffff R13: 0000000000000658 R14: 00000000006fc8e0 R15: 0000000000000000 Uninit was created at: kmsan_save_stack_with_flags mm/kmsan/kmsan.c:278 [inline] kmsan_internal_poison_shadow+0xb8/0x1b0 mm/kmsan/kmsan.c:188 kmsan_kmalloc+0x94/0x100 mm/kmsan/kmsan.c:314 kmem_cache_alloc+0xaab/0xb90 mm/slub.c:2756 inet_getpeer+0xed8/0x1e70 net/ipv4/inetpeer.c:210 inet_getpeer_v4 include/net/inetpeer.h:110 [inline] ip4_frag_init+0x4d1/0x740 net/ipv4/ip_fragment.c:153 inet_frag_alloc net/ipv4/inet_fragment.c:369 [inline] inet_frag_create net/ipv4/inet_fragment.c:385 [inline] inet_frag_find+0x7da/0x1610 net/ipv4/inet_fragment.c:418 ip_find net/ipv4/ip_fragment.c:275 [inline] ip_defrag+0x448/0x67a0 net/ipv4/ip_fragment.c:676 ip_check_defrag+0x775/0xda0 net/ipv4/ip_fragment.c:724 packet_rcv_fanout+0x2a8/0x8d0 net/packet/af_packet.c:1447 deliver_skb net/core/dev.c:1897 [inline] deliver_ptype_list_skb net/core/dev.c:1912 [inline] __netif_receive_skb_core+0x314a/0x4a80 net/core/dev.c:4545 __netif_receive_skb net/core/dev.c:4627 [inline] netif_receive_skb_internal+0x49d/0x630 net/core/dev.c:4701 netif_receive_skb+0x230/0x240 net/core/dev.c:4725 tun_rx_batched drivers/net/tun.c:1555 [inline] tun_get_user+0x6d88/0x7580 drivers/net/tun.c:1962 tun_chr_write_iter+0x1d4/0x330 drivers/net/tun.c:1990 do_iter_readv_writev+0x7bb/0x970 include/linux/fs.h:1776 do_iter_write+0x30d/0xd40 fs/read_write.c:932 vfs_writev fs/read_write.c:977 [inline] do_writev+0x3c9/0x830 fs/read_write.c:1012 SYSC_writev+0x9b/0xb0 fs/read_write.c:1085 SyS_writev+0x56/0x80 fs/read_write.c:1082 do_syscall_64+0x309/0x430 arch/x86/entry/common.c:287 entry_SYSCALL_64_after_hwframe+0x3d/0xa2 Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: syzbot <syzkaller@googlegroups.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2018-04-09 13:43:27 +00:00
p->dtime = (__u32)jiffies;
refcount_set(&p->refcnt, 2);
atomic_set(&p->rid, 0);
p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW;
p->rate_tokens = 0;
p->n_redirects = 0;
/* 60*HZ is arbitrary, but chosen enough high so that the first
* calculation of tokens is at its maximum.
*/
p->rate_last = jiffies - 60*HZ;
rb_link_node(&p->rb_node, parent, pp);
rb_insert_color(&p->rb_node, &base->rb_root);
base->total++;
}
inetpeer: RCU conversion inetpeer currently uses an AVL tree protected by an rwlock. It's possible to make most lookups use RCU 1) Add a struct rcu_head to struct inet_peer 2) add a lookup_rcu_bh() helper to perform lockless and opportunistic lookup. This is a normal function, not a macro like lookup(). 3) Add a limit to number of links followed by lookup_rcu_bh(). This is needed in case we fall in a loop. 4) add an smp_wmb() in link_to_pool() right before node insert. 5) make unlink_from_pool() use atomic_cmpxchg() to make sure it can take last reference to an inet_peer, since lockless readers could increase refcount, even while we hold peers.lock. 6) Delay struct inet_peer freeing after rcu grace period so that lookup_rcu_bh() cannot crash. 7) inet_getpeer() first attempts lockless lookup. Note this lookup can fail even if target is in AVL tree, but a concurrent writer can let tree in a non correct form. If this attemps fails, lock is taken a regular lookup is performed again. 8) convert peers.lock from rwlock to a spinlock 9) Remove SLAB_HWCACHE_ALIGN when peer_cachep is created, because rcu_head adds 16 bytes on 64bit arches, doubling effective size (64 -> 128 bytes) In a future patch, this is probably possible to revert this part, if rcu field is put in an union to share space with rid, ip_id_count, tcp_ts & tcp_ts_stamp. These fields being manipulated only with refcnt > 0. Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-06-15 08:23:14 +00:00
}
if (gc_cnt)
inet_peer_gc(base, gc_stack, gc_cnt);
write_sequnlock_bh(&base->lock);
return p;
}
EXPORT_SYMBOL_GPL(inet_getpeer);
void inet_putpeer(struct inet_peer *p)
{
p->dtime = (__u32)jiffies;
if (refcount_dec_and_test(&p->refcnt))
call_rcu(&p->rcu, inetpeer_free_rcu);
}
EXPORT_SYMBOL_GPL(inet_putpeer);
/*
* Check transmit rate limitation for given message.
* The rate information is held in the inet_peer entries now.
* This function is generic and could be used for other purposes
* too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
*
* Note that the same inet_peer fields are modified by functions in
* route.c too, but these work for packet destinations while xrlim_allow
* works for icmp destinations. This means the rate limiting information
* for one "ip object" is shared - and these ICMPs are twice limited:
* by source and by destination.
*
* RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
* SHOULD allow setting of rate limits
*
* Shared between ICMPv4 and ICMPv6.
*/
#define XRLIM_BURST_FACTOR 6
bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
{
unsigned long now, token;
bool rc = false;
if (!peer)
return true;
token = peer->rate_tokens;
now = jiffies;
token += now - peer->rate_last;
peer->rate_last = now;
if (token > XRLIM_BURST_FACTOR * timeout)
token = XRLIM_BURST_FACTOR * timeout;
if (token >= timeout) {
token -= timeout;
rc = true;
}
peer->rate_tokens = token;
return rc;
}
EXPORT_SYMBOL(inet_peer_xrlim_allow);
void inetpeer_invalidate_tree(struct inet_peer_base *base)
{
struct rb_node *p = rb_first(&base->rb_root);
while (p) {
struct inet_peer *peer = rb_entry(p, struct inet_peer, rb_node);
p = rb_next(p);
rb_erase(&peer->rb_node, &base->rb_root);
inet_putpeer(peer);
cond_resched();
}
base->total = 0;
}
EXPORT_SYMBOL(inetpeer_invalidate_tree);