2005-04-17 06:20:36 +08:00
|
|
|
/*
|
|
|
|
* 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/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>
|
2005-08-16 13:18:02 +08:00
|
|
|
#include <net/ip.h>
|
2005-04-17 06:20:36 +08:00
|
|
|
#include <net/inetpeer.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.
|
|
|
|
* At this moment this information consists only of ID field for the next
|
|
|
|
* outgoing IP packet. This field is incremented with each packet as encoded
|
|
|
|
* in inet_getid() function (include/net/inetpeer.h).
|
|
|
|
* At the moment of writing this notes identifier of IP packets is generated
|
|
|
|
* to be unpredictable using this code only for packets subjected
|
|
|
|
* (actually or potentially) to defragmentation. I.e. DF packets less than
|
|
|
|
* PMTU in size uses a constant ID and do not use this code (see
|
|
|
|
* ip_select_ident() in include/net/ip.h).
|
|
|
|
*
|
|
|
|
* Route cache entries hold references to our nodes.
|
|
|
|
* New cache entries get references via lookup by destination IP address in
|
|
|
|
* the avl tree. The reference is grabbed only when it's needed i.e. only
|
|
|
|
* when we try to output IP packet which needs an unpredictable ID (see
|
|
|
|
* __ip_select_ident() in net/ipv4/route.c).
|
|
|
|
* 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 AVL 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 16:23:14 +08:00
|
|
|
* 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
|
2005-04-17 06:20:36 +08:00
|
|
|
* AND reference count being 0.
|
|
|
|
* 3. Nodes appears and disappears from unused node list only under
|
|
|
|
* "inet_peer_unused_lock".
|
|
|
|
* 4. Global variable peer_total is modified under the pool lock.
|
|
|
|
* 5. struct inet_peer fields modification:
|
|
|
|
* avl_left, avl_right, avl_parent, avl_height: pool lock
|
2007-11-13 13:27:28 +08:00
|
|
|
* unused: unused node list lock
|
2005-04-17 06:20:36 +08:00
|
|
|
* refcnt: atomically against modifications on other CPU;
|
|
|
|
* usually under some other lock to prevent node disappearing
|
|
|
|
* dtime: unused node list lock
|
2010-12-01 03:53:55 +08:00
|
|
|
* daddr: unchangeable
|
2010-06-16 12:52:13 +08:00
|
|
|
* ip_id_count: atomic value (no lock needed)
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
|
|
|
|
2006-12-07 12:33:20 +08:00
|
|
|
static struct kmem_cache *peer_cachep __read_mostly;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
#define node_height(x) x->avl_height
|
2010-06-15 03:35:21 +08:00
|
|
|
|
|
|
|
#define peer_avl_empty ((struct inet_peer *)&peer_fake_node)
|
2010-10-26 07:55:38 +08:00
|
|
|
#define peer_avl_empty_rcu ((struct inet_peer __rcu __force *)&peer_fake_node)
|
2010-06-15 03:35:21 +08:00
|
|
|
static const struct inet_peer peer_fake_node = {
|
2010-10-26 07:55:38 +08:00
|
|
|
.avl_left = peer_avl_empty_rcu,
|
|
|
|
.avl_right = peer_avl_empty_rcu,
|
2005-04-17 06:20:36 +08:00
|
|
|
.avl_height = 0
|
|
|
|
};
|
2010-06-15 03:35:21 +08:00
|
|
|
|
2010-12-01 04:12:23 +08:00
|
|
|
struct inet_peer_base {
|
2010-10-26 07:55:38 +08:00
|
|
|
struct inet_peer __rcu *root;
|
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 16:23:14 +08:00
|
|
|
spinlock_t lock;
|
2010-06-15 03:35:21 +08:00
|
|
|
int total;
|
2010-12-01 04:12:23 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
static struct inet_peer_base v4_peers = {
|
2010-10-26 07:55:38 +08:00
|
|
|
.root = peer_avl_empty_rcu,
|
2010-12-01 03:41:59 +08:00
|
|
|
.lock = __SPIN_LOCK_UNLOCKED(v4_peers.lock),
|
2010-06-15 03:35:21 +08:00
|
|
|
.total = 0,
|
|
|
|
};
|
2010-12-01 04:12:23 +08:00
|
|
|
|
|
|
|
static struct inet_peer_base v6_peers = {
|
|
|
|
.root = peer_avl_empty_rcu,
|
|
|
|
.lock = __SPIN_LOCK_UNLOCKED(v6_peers.lock),
|
|
|
|
.total = 0,
|
|
|
|
};
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
#define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
|
|
|
|
|
|
|
|
/* Exported for sysctl_net_ipv4. */
|
2007-03-07 12:23:10 +08:00
|
|
|
int inet_peer_threshold __read_mostly = 65536 + 128; /* start to throw entries more
|
2005-04-17 06:20:36 +08:00
|
|
|
* aggressively at this stage */
|
2007-03-07 12:23:10 +08:00
|
|
|
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 */
|
|
|
|
int inet_peer_gc_mintime __read_mostly = 10 * HZ;
|
|
|
|
int inet_peer_gc_maxtime __read_mostly = 120 * HZ;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2010-06-15 03:35:21 +08:00
|
|
|
static struct {
|
|
|
|
struct list_head list;
|
|
|
|
spinlock_t lock;
|
|
|
|
} unused_peers = {
|
|
|
|
.list = LIST_HEAD_INIT(unused_peers.list),
|
|
|
|
.lock = __SPIN_LOCK_UNLOCKED(unused_peers.lock),
|
|
|
|
};
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
static void peer_check_expire(unsigned long dummy);
|
2005-09-10 04:10:40 +08:00
|
|
|
static DEFINE_TIMER(peer_periodic_timer, peer_check_expire, 0, 0);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
|
|
|
/* 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),
|
2010-06-16 12:52:13 +08:00
|
|
|
0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
|
2007-07-20 09:11:58 +08:00
|
|
|
NULL);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
/* All the timers, started at system startup tend
|
|
|
|
to synchronize. Perturb it a bit.
|
|
|
|
*/
|
|
|
|
peer_periodic_timer.expires = jiffies
|
|
|
|
+ net_random() % inet_peer_gc_maxtime
|
|
|
|
+ inet_peer_gc_maxtime;
|
|
|
|
add_timer(&peer_periodic_timer);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Called with or without local BH being disabled. */
|
|
|
|
static void unlink_from_unused(struct inet_peer *p)
|
|
|
|
{
|
2010-06-15 03:35:21 +08:00
|
|
|
if (!list_empty(&p->unused)) {
|
|
|
|
spin_lock_bh(&unused_peers.lock);
|
|
|
|
list_del_init(&p->unused);
|
|
|
|
spin_unlock_bh(&unused_peers.lock);
|
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
2010-12-02 09:28:18 +08:00
|
|
|
static int addr_compare(const struct inetpeer_addr *a,
|
|
|
|
const struct inetpeer_addr *b)
|
2010-12-01 04:08:53 +08:00
|
|
|
{
|
|
|
|
int i, n = (a->family == AF_INET ? 1 : 4);
|
|
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
|
|
if (a->a6[i] == b->a6[i])
|
|
|
|
continue;
|
|
|
|
if (a->a6[i] < b->a6[i])
|
|
|
|
return -1;
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2007-03-07 12:23:10 +08:00
|
|
|
/*
|
|
|
|
* Called with local BH disabled and the pool lock held.
|
|
|
|
*/
|
2010-12-01 03:41:59 +08:00
|
|
|
#define lookup(_daddr, _stack, _base) \
|
2005-04-17 06:20:36 +08:00
|
|
|
({ \
|
2010-10-26 07:55:38 +08:00
|
|
|
struct inet_peer *u; \
|
|
|
|
struct inet_peer __rcu **v; \
|
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 16:23:14 +08:00
|
|
|
\
|
|
|
|
stackptr = _stack; \
|
2010-12-01 03:41:59 +08:00
|
|
|
*stackptr++ = &_base->root; \
|
|
|
|
for (u = rcu_dereference_protected(_base->root, \
|
|
|
|
lockdep_is_held(&_base->lock)); \
|
2010-10-26 07:55:38 +08:00
|
|
|
u != peer_avl_empty; ) { \
|
2010-12-01 04:08:53 +08:00
|
|
|
int cmp = addr_compare(_daddr, &u->daddr); \
|
|
|
|
if (cmp == 0) \
|
2005-04-17 06:20:36 +08:00
|
|
|
break; \
|
2010-12-01 04:08:53 +08:00
|
|
|
if (cmp == -1) \
|
2005-04-17 06:20:36 +08:00
|
|
|
v = &u->avl_left; \
|
|
|
|
else \
|
|
|
|
v = &u->avl_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 16:23:14 +08:00
|
|
|
*stackptr++ = v; \
|
2010-10-26 07:55:38 +08:00
|
|
|
u = rcu_dereference_protected(*v, \
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&_base->lock)); \
|
2005-04-17 06:20:36 +08:00
|
|
|
} \
|
|
|
|
u; \
|
|
|
|
})
|
|
|
|
|
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 16:23:14 +08:00
|
|
|
/*
|
|
|
|
* Called with rcu_read_lock_bh()
|
|
|
|
* Because we hold no lock against a writer, its quite possible we fall
|
|
|
|
* in an endless loop.
|
|
|
|
* But every pointer we follow is guaranteed to be valid thanks to RCU.
|
|
|
|
* We exit from this function if number of links exceeds PEER_MAXDEPTH
|
|
|
|
*/
|
2010-12-02 09:28:18 +08:00
|
|
|
static struct inet_peer *lookup_rcu_bh(const struct inetpeer_addr *daddr,
|
2010-12-01 04:08:53 +08:00
|
|
|
struct inet_peer_base *base)
|
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 16:23:14 +08:00
|
|
|
{
|
2010-12-01 03:41:59 +08:00
|
|
|
struct inet_peer *u = rcu_dereference_bh(base->root);
|
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 16:23:14 +08:00
|
|
|
int count = 0;
|
|
|
|
|
|
|
|
while (u != peer_avl_empty) {
|
2010-12-01 04:08:53 +08:00
|
|
|
int cmp = addr_compare(daddr, &u->daddr);
|
|
|
|
if (cmp == 0) {
|
2010-06-16 12:47:39 +08:00
|
|
|
/* Before taking a reference, check if this entry was
|
|
|
|
* deleted, unlink_from_pool() sets refcnt=-1 to make
|
|
|
|
* distinction between an unused entry (refcnt=0) and
|
|
|
|
* a freed one.
|
|
|
|
*/
|
|
|
|
if (unlikely(!atomic_add_unless(&u->refcnt, 1, -1)))
|
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 16:23:14 +08:00
|
|
|
u = NULL;
|
|
|
|
return u;
|
|
|
|
}
|
2010-12-01 04:08:53 +08:00
|
|
|
if (cmp == -1)
|
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 16:23:14 +08:00
|
|
|
u = rcu_dereference_bh(u->avl_left);
|
|
|
|
else
|
|
|
|
u = rcu_dereference_bh(u->avl_right);
|
|
|
|
if (unlikely(++count == PEER_MAXDEPTH))
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Called with local BH disabled and the pool lock held. */
|
2010-12-01 03:41:59 +08:00
|
|
|
#define lookup_rightempty(start, base) \
|
2005-04-17 06:20:36 +08:00
|
|
|
({ \
|
2010-10-26 07:55:38 +08:00
|
|
|
struct inet_peer *u; \
|
|
|
|
struct inet_peer __rcu **v; \
|
2005-04-17 06:20:36 +08:00
|
|
|
*stackptr++ = &start->avl_left; \
|
|
|
|
v = &start->avl_left; \
|
2010-10-26 07:55:38 +08:00
|
|
|
for (u = rcu_dereference_protected(*v, \
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&base->lock)); \
|
2010-10-26 07:55:38 +08:00
|
|
|
u->avl_right != peer_avl_empty_rcu; ) { \
|
2005-04-17 06:20:36 +08:00
|
|
|
v = &u->avl_right; \
|
|
|
|
*stackptr++ = v; \
|
2010-10-26 07:55:38 +08:00
|
|
|
u = rcu_dereference_protected(*v, \
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&base->lock)); \
|
2005-04-17 06:20:36 +08:00
|
|
|
} \
|
|
|
|
u; \
|
|
|
|
})
|
|
|
|
|
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 16:23:14 +08:00
|
|
|
/* Called with local BH disabled and the pool lock held.
|
2005-04-17 06:20:36 +08:00
|
|
|
* Variable names are the proof of operation correctness.
|
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 16:23:14 +08:00
|
|
|
* Look into mm/map_avl.c for more detail description of the ideas.
|
|
|
|
*/
|
2010-10-26 07:55:38 +08:00
|
|
|
static void peer_avl_rebalance(struct inet_peer __rcu **stack[],
|
2010-12-01 03:41:59 +08:00
|
|
|
struct inet_peer __rcu ***stackend,
|
|
|
|
struct inet_peer_base *base)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2010-10-26 07:55:38 +08:00
|
|
|
struct inet_peer __rcu **nodep;
|
|
|
|
struct inet_peer *node, *l, *r;
|
2005-04-17 06:20:36 +08:00
|
|
|
int lh, rh;
|
|
|
|
|
|
|
|
while (stackend > stack) {
|
|
|
|
nodep = *--stackend;
|
2010-10-26 07:55:38 +08:00
|
|
|
node = rcu_dereference_protected(*nodep,
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&base->lock));
|
2010-10-26 07:55:38 +08:00
|
|
|
l = rcu_dereference_protected(node->avl_left,
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&base->lock));
|
2010-10-26 07:55:38 +08:00
|
|
|
r = rcu_dereference_protected(node->avl_right,
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&base->lock));
|
2005-04-17 06:20:36 +08:00
|
|
|
lh = node_height(l);
|
|
|
|
rh = node_height(r);
|
|
|
|
if (lh > rh + 1) { /* l: RH+2 */
|
|
|
|
struct inet_peer *ll, *lr, *lrl, *lrr;
|
|
|
|
int lrh;
|
2010-10-26 07:55:38 +08:00
|
|
|
ll = rcu_dereference_protected(l->avl_left,
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&base->lock));
|
2010-10-26 07:55:38 +08:00
|
|
|
lr = rcu_dereference_protected(l->avl_right,
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&base->lock));
|
2005-04-17 06:20:36 +08:00
|
|
|
lrh = node_height(lr);
|
|
|
|
if (lrh <= node_height(ll)) { /* ll: RH+1 */
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(node->avl_left, lr); /* lr: RH or RH+1 */
|
|
|
|
RCU_INIT_POINTER(node->avl_right, r); /* r: RH */
|
2005-04-17 06:20:36 +08:00
|
|
|
node->avl_height = lrh + 1; /* RH+1 or RH+2 */
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH+1 */
|
|
|
|
RCU_INIT_POINTER(l->avl_right, node); /* node: RH+1 or RH+2 */
|
2005-04-17 06:20:36 +08:00
|
|
|
l->avl_height = node->avl_height + 1;
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(*nodep, l);
|
2005-04-17 06:20:36 +08:00
|
|
|
} else { /* ll: RH, lr: RH+1 */
|
2010-10-26 07:55:38 +08:00
|
|
|
lrl = rcu_dereference_protected(lr->avl_left,
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&base->lock)); /* lrl: RH or RH-1 */
|
2010-10-26 07:55:38 +08:00
|
|
|
lrr = rcu_dereference_protected(lr->avl_right,
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&base->lock)); /* lrr: RH or RH-1 */
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(node->avl_left, lrr); /* lrr: RH or RH-1 */
|
|
|
|
RCU_INIT_POINTER(node->avl_right, r); /* r: RH */
|
2005-04-17 06:20:36 +08:00
|
|
|
node->avl_height = rh + 1; /* node: RH+1 */
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH */
|
|
|
|
RCU_INIT_POINTER(l->avl_right, lrl); /* lrl: RH or RH-1 */
|
2005-04-17 06:20:36 +08:00
|
|
|
l->avl_height = rh + 1; /* l: RH+1 */
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(lr->avl_left, l); /* l: RH+1 */
|
|
|
|
RCU_INIT_POINTER(lr->avl_right, node); /* node: RH+1 */
|
2005-04-17 06:20:36 +08:00
|
|
|
lr->avl_height = rh + 2;
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(*nodep, lr);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
} else if (rh > lh + 1) { /* r: LH+2 */
|
|
|
|
struct inet_peer *rr, *rl, *rlr, *rll;
|
|
|
|
int rlh;
|
2010-10-26 07:55:38 +08:00
|
|
|
rr = rcu_dereference_protected(r->avl_right,
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&base->lock));
|
2010-10-26 07:55:38 +08:00
|
|
|
rl = rcu_dereference_protected(r->avl_left,
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&base->lock));
|
2005-04-17 06:20:36 +08:00
|
|
|
rlh = node_height(rl);
|
|
|
|
if (rlh <= node_height(rr)) { /* rr: LH+1 */
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(node->avl_right, rl); /* rl: LH or LH+1 */
|
|
|
|
RCU_INIT_POINTER(node->avl_left, l); /* l: LH */
|
2005-04-17 06:20:36 +08:00
|
|
|
node->avl_height = rlh + 1; /* LH+1 or LH+2 */
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH+1 */
|
|
|
|
RCU_INIT_POINTER(r->avl_left, node); /* node: LH+1 or LH+2 */
|
2005-04-17 06:20:36 +08:00
|
|
|
r->avl_height = node->avl_height + 1;
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(*nodep, r);
|
2005-04-17 06:20:36 +08:00
|
|
|
} else { /* rr: RH, rl: RH+1 */
|
2010-10-26 07:55:38 +08:00
|
|
|
rlr = rcu_dereference_protected(rl->avl_right,
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&base->lock)); /* rlr: LH or LH-1 */
|
2010-10-26 07:55:38 +08:00
|
|
|
rll = rcu_dereference_protected(rl->avl_left,
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&base->lock)); /* rll: LH or LH-1 */
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(node->avl_right, rll); /* rll: LH or LH-1 */
|
|
|
|
RCU_INIT_POINTER(node->avl_left, l); /* l: LH */
|
2005-04-17 06:20:36 +08:00
|
|
|
node->avl_height = lh + 1; /* node: LH+1 */
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH */
|
|
|
|
RCU_INIT_POINTER(r->avl_left, rlr); /* rlr: LH or LH-1 */
|
2005-04-17 06:20:36 +08:00
|
|
|
r->avl_height = lh + 1; /* r: LH+1 */
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(rl->avl_right, r); /* r: LH+1 */
|
|
|
|
RCU_INIT_POINTER(rl->avl_left, node); /* node: LH+1 */
|
2005-04-17 06:20:36 +08:00
|
|
|
rl->avl_height = lh + 2;
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(*nodep, rl);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
} else {
|
|
|
|
node->avl_height = (lh > rh ? lh : rh) + 1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
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 16:23:14 +08:00
|
|
|
/* Called with local BH disabled and the pool lock held. */
|
2010-12-01 03:41:59 +08:00
|
|
|
#define link_to_pool(n, base) \
|
2005-04-17 06:20:36 +08:00
|
|
|
do { \
|
|
|
|
n->avl_height = 1; \
|
2010-10-26 07:55:38 +08:00
|
|
|
n->avl_left = peer_avl_empty_rcu; \
|
|
|
|
n->avl_right = peer_avl_empty_rcu; \
|
|
|
|
/* lockless readers can catch us now */ \
|
|
|
|
rcu_assign_pointer(**--stackptr, n); \
|
2010-12-01 03:41:59 +08:00
|
|
|
peer_avl_rebalance(stack, stackptr, base); \
|
2010-06-15 03:35:21 +08:00
|
|
|
} while (0)
|
2005-04-17 06:20:36 +08:00
|
|
|
|
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 16:23:14 +08:00
|
|
|
static void inetpeer_free_rcu(struct rcu_head *head)
|
|
|
|
{
|
|
|
|
kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
|
|
|
|
}
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/* May be called with local BH enabled. */
|
2010-12-01 03:41:59 +08:00
|
|
|
static void unlink_from_pool(struct inet_peer *p, struct inet_peer_base *base)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
int do_free;
|
|
|
|
|
|
|
|
do_free = 0;
|
|
|
|
|
2010-12-01 03:41:59 +08:00
|
|
|
spin_lock_bh(&base->lock);
|
2005-04-17 06:20:36 +08:00
|
|
|
/* Check the reference counter. It was artificially incremented by 1
|
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 16:23:14 +08:00
|
|
|
* in cleanup() function to prevent sudden disappearing. If we can
|
|
|
|
* atomically (because of lockless readers) take this last reference,
|
|
|
|
* it's safe to remove the node and free it later.
|
2010-06-16 12:47:39 +08:00
|
|
|
* We use refcnt=-1 to alert lockless readers this entry is deleted.
|
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 16:23:14 +08:00
|
|
|
*/
|
2010-06-16 12:47:39 +08:00
|
|
|
if (atomic_cmpxchg(&p->refcnt, 1, -1) == 1) {
|
2010-10-26 07:55:38 +08:00
|
|
|
struct inet_peer __rcu **stack[PEER_MAXDEPTH];
|
|
|
|
struct inet_peer __rcu ***stackptr, ***delp;
|
2010-12-01 04:08:53 +08:00
|
|
|
if (lookup(&p->daddr, stack, base) != p)
|
2005-04-17 06:20:36 +08:00
|
|
|
BUG();
|
|
|
|
delp = stackptr - 1; /* *delp[0] == p */
|
2010-10-26 07:55:38 +08:00
|
|
|
if (p->avl_left == peer_avl_empty_rcu) {
|
2005-04-17 06:20:36 +08:00
|
|
|
*delp[0] = p->avl_right;
|
|
|
|
--stackptr;
|
|
|
|
} else {
|
|
|
|
/* look for a node to insert instead of p */
|
|
|
|
struct inet_peer *t;
|
2010-12-01 03:41:59 +08:00
|
|
|
t = lookup_rightempty(p, base);
|
2010-10-26 07:55:38 +08:00
|
|
|
BUG_ON(rcu_dereference_protected(*stackptr[-1],
|
2010-12-01 03:41:59 +08:00
|
|
|
lockdep_is_held(&base->lock)) != t);
|
2005-04-17 06:20:36 +08:00
|
|
|
**--stackptr = t->avl_left;
|
2010-12-01 03:53:55 +08:00
|
|
|
/* t is removed, t->daddr > x->daddr for any
|
2005-04-17 06:20:36 +08:00
|
|
|
* x in p->avl_left subtree.
|
|
|
|
* Put t in the old place of p. */
|
2010-10-26 07:55:38 +08:00
|
|
|
RCU_INIT_POINTER(*delp[0], t);
|
2005-04-17 06:20:36 +08:00
|
|
|
t->avl_left = p->avl_left;
|
|
|
|
t->avl_right = p->avl_right;
|
|
|
|
t->avl_height = p->avl_height;
|
2006-01-09 14:24:28 +08:00
|
|
|
BUG_ON(delp[1] != &p->avl_left);
|
2005-04-17 06:20:36 +08:00
|
|
|
delp[1] = &t->avl_left; /* was &p->avl_left */
|
|
|
|
}
|
2010-12-01 03:41:59 +08:00
|
|
|
peer_avl_rebalance(stack, stackptr, base);
|
|
|
|
base->total--;
|
2005-04-17 06:20:36 +08:00
|
|
|
do_free = 1;
|
|
|
|
}
|
2010-12-01 03:41:59 +08:00
|
|
|
spin_unlock_bh(&base->lock);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
if (do_free)
|
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 16:23:14 +08:00
|
|
|
call_rcu_bh(&p->rcu, inetpeer_free_rcu);
|
2005-04-17 06:20:36 +08:00
|
|
|
else
|
|
|
|
/* The node is used again. Decrease the reference counter
|
|
|
|
* back. The loop "cleanup -> unlink_from_unused
|
|
|
|
* -> unlink_from_pool -> putpeer -> link_to_unused
|
|
|
|
* -> cleanup (for the same node)"
|
|
|
|
* doesn't really exist because the entry will have a
|
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 16:23:14 +08:00
|
|
|
* recent deletion time and will not be cleaned again soon.
|
|
|
|
*/
|
2005-04-17 06:20:36 +08:00
|
|
|
inet_putpeer(p);
|
|
|
|
}
|
|
|
|
|
2010-12-01 04:12:23 +08:00
|
|
|
static struct inet_peer_base *family_to_base(int family)
|
|
|
|
{
|
|
|
|
return (family == AF_INET ? &v4_peers : &v6_peers);
|
|
|
|
}
|
|
|
|
|
2010-12-01 03:41:59 +08:00
|
|
|
static struct inet_peer_base *peer_to_base(struct inet_peer *p)
|
|
|
|
{
|
2010-12-01 04:12:23 +08:00
|
|
|
return family_to_base(p->daddr.family);
|
2010-12-01 03:41:59 +08:00
|
|
|
}
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/* May be called with local BH enabled. */
|
|
|
|
static int cleanup_once(unsigned long ttl)
|
|
|
|
{
|
2007-11-13 13:27:28 +08:00
|
|
|
struct inet_peer *p = NULL;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
/* Remove the first entry from the list of unused nodes. */
|
2010-06-15 03:35:21 +08:00
|
|
|
spin_lock_bh(&unused_peers.lock);
|
|
|
|
if (!list_empty(&unused_peers.list)) {
|
2007-11-13 13:27:28 +08:00
|
|
|
__u32 delta;
|
|
|
|
|
2010-06-15 03:35:21 +08:00
|
|
|
p = list_first_entry(&unused_peers.list, struct inet_peer, unused);
|
2007-11-13 13:27:28 +08:00
|
|
|
delta = (__u32)jiffies - p->dtime;
|
|
|
|
|
2006-10-13 12:21:06 +08:00
|
|
|
if (delta < ttl) {
|
2005-04-17 06:20:36 +08:00
|
|
|
/* Do not prune fresh entries. */
|
2010-06-15 03:35:21 +08:00
|
|
|
spin_unlock_bh(&unused_peers.lock);
|
2005-04-17 06:20:36 +08:00
|
|
|
return -1;
|
|
|
|
}
|
2007-11-13 13:27:28 +08:00
|
|
|
|
|
|
|
list_del_init(&p->unused);
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/* Grab an extra reference to prevent node disappearing
|
|
|
|
* before unlink_from_pool() call. */
|
|
|
|
atomic_inc(&p->refcnt);
|
|
|
|
}
|
2010-06-15 03:35:21 +08:00
|
|
|
spin_unlock_bh(&unused_peers.lock);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
if (p == NULL)
|
|
|
|
/* It means that the total number of USED entries has
|
|
|
|
* grown over inet_peer_threshold. It shouldn't really
|
|
|
|
* happen because of entry limits in route cache. */
|
|
|
|
return -1;
|
|
|
|
|
2010-12-01 03:41:59 +08:00
|
|
|
unlink_from_pool(p, peer_to_base(p));
|
2005-04-17 06:20:36 +08:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Called with or without local BH being disabled. */
|
2010-12-02 09:28:18 +08:00
|
|
|
struct inet_peer *inet_getpeer(struct inetpeer_addr *daddr, int create)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2010-10-26 07:55:38 +08:00
|
|
|
struct inet_peer __rcu **stack[PEER_MAXDEPTH], ***stackptr;
|
2010-12-01 03:41:59 +08:00
|
|
|
struct inet_peer_base *base = family_to_base(AF_INET);
|
|
|
|
struct inet_peer *p;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
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 16:23:14 +08:00
|
|
|
/* Look up for the address quickly, lockless.
|
|
|
|
* Because of a concurrent writer, we might not find an existing entry.
|
|
|
|
*/
|
|
|
|
rcu_read_lock_bh();
|
2010-12-01 04:08:53 +08:00
|
|
|
p = lookup_rcu_bh(daddr, base);
|
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 16:23:14 +08:00
|
|
|
rcu_read_unlock_bh();
|
|
|
|
|
|
|
|
if (p) {
|
|
|
|
/* The existing node has been found.
|
|
|
|
* Remove the entry from unused list if it was there.
|
|
|
|
*/
|
|
|
|
unlink_from_unused(p);
|
|
|
|
return p;
|
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
|
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 16:23:14 +08:00
|
|
|
/* retry an exact lookup, taking the lock before.
|
|
|
|
* At least, nodes should be hot in our cache.
|
|
|
|
*/
|
2010-12-01 03:41:59 +08:00
|
|
|
spin_lock_bh(&base->lock);
|
2010-12-01 04:08:53 +08:00
|
|
|
p = lookup(daddr, stack, base);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (p != peer_avl_empty) {
|
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 16:23:14 +08:00
|
|
|
atomic_inc(&p->refcnt);
|
2010-12-01 03:41:59 +08:00
|
|
|
spin_unlock_bh(&base->lock);
|
2005-04-17 06:20:36 +08:00
|
|
|
/* Remove the entry from unused list if it was there. */
|
|
|
|
unlink_from_unused(p);
|
|
|
|
return 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 16:23:14 +08:00
|
|
|
p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL;
|
|
|
|
if (p) {
|
2010-12-01 03:54:19 +08:00
|
|
|
p->daddr = *daddr;
|
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 16:23:14 +08:00
|
|
|
atomic_set(&p->refcnt, 1);
|
|
|
|
atomic_set(&p->rid, 0);
|
2010-12-01 03:54:19 +08:00
|
|
|
atomic_set(&p->ip_id_count, secure_ip_id(daddr->a4));
|
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 16:23:14 +08:00
|
|
|
p->tcp_ts_stamp = 0;
|
|
|
|
INIT_LIST_HEAD(&p->unused);
|
|
|
|
|
|
|
|
|
|
|
|
/* Link the node. */
|
2010-12-01 03:41:59 +08:00
|
|
|
link_to_pool(p, base);
|
|
|
|
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 16:23:14 +08:00
|
|
|
}
|
2010-12-01 03:41:59 +08:00
|
|
|
spin_unlock_bh(&base->lock);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2010-12-01 03:41:59 +08:00
|
|
|
if (base->total >= inet_peer_threshold)
|
2005-04-17 06:20:36 +08:00
|
|
|
/* Remove one less-recently-used entry. */
|
|
|
|
cleanup_once(0);
|
|
|
|
|
|
|
|
return p;
|
|
|
|
}
|
|
|
|
|
2010-12-01 03:41:59 +08:00
|
|
|
static int compute_total(void)
|
|
|
|
{
|
2010-12-01 04:12:23 +08:00
|
|
|
return v4_peers.total + v6_peers.total;
|
2010-12-01 03:41:59 +08:00
|
|
|
}
|
2010-12-01 04:27:11 +08:00
|
|
|
EXPORT_SYMBOL_GPL(inet_getpeer);
|
2010-12-01 03:41:59 +08:00
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/* Called with local BH disabled. */
|
|
|
|
static void peer_check_expire(unsigned long dummy)
|
|
|
|
{
|
2006-10-13 12:21:06 +08:00
|
|
|
unsigned long now = jiffies;
|
2010-12-01 03:41:59 +08:00
|
|
|
int ttl, total;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2010-12-01 03:41:59 +08:00
|
|
|
total = compute_total();
|
|
|
|
if (total >= inet_peer_threshold)
|
2005-04-17 06:20:36 +08:00
|
|
|
ttl = inet_peer_minttl;
|
|
|
|
else
|
|
|
|
ttl = inet_peer_maxttl
|
|
|
|
- (inet_peer_maxttl - inet_peer_minttl) / HZ *
|
2010-12-01 03:41:59 +08:00
|
|
|
total / inet_peer_threshold * HZ;
|
2006-10-13 12:21:06 +08:00
|
|
|
while (!cleanup_once(ttl)) {
|
|
|
|
if (jiffies != now)
|
|
|
|
break;
|
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
/* Trigger the timer after inet_peer_gc_mintime .. inet_peer_gc_maxtime
|
|
|
|
* interval depending on the total number of entries (more entries,
|
|
|
|
* less interval). */
|
2010-12-01 03:41:59 +08:00
|
|
|
total = compute_total();
|
|
|
|
if (total >= inet_peer_threshold)
|
2005-08-24 01:10:15 +08:00
|
|
|
peer_periodic_timer.expires = jiffies + inet_peer_gc_mintime;
|
|
|
|
else
|
|
|
|
peer_periodic_timer.expires = jiffies
|
|
|
|
+ inet_peer_gc_maxtime
|
|
|
|
- (inet_peer_gc_maxtime - inet_peer_gc_mintime) / HZ *
|
2010-12-01 03:41:59 +08:00
|
|
|
total / inet_peer_threshold * HZ;
|
2005-04-17 06:20:36 +08:00
|
|
|
add_timer(&peer_periodic_timer);
|
|
|
|
}
|
2006-10-13 12:21:06 +08:00
|
|
|
|
|
|
|
void inet_putpeer(struct inet_peer *p)
|
|
|
|
{
|
2010-06-15 03:35:21 +08:00
|
|
|
local_bh_disable();
|
|
|
|
|
|
|
|
if (atomic_dec_and_lock(&p->refcnt, &unused_peers.lock)) {
|
|
|
|
list_add_tail(&p->unused, &unused_peers.list);
|
2006-10-13 12:21:06 +08:00
|
|
|
p->dtime = (__u32)jiffies;
|
2010-06-15 03:35:21 +08:00
|
|
|
spin_unlock(&unused_peers.lock);
|
2006-10-13 12:21:06 +08:00
|
|
|
}
|
2010-06-15 03:35:21 +08:00
|
|
|
|
|
|
|
local_bh_enable();
|
2006-10-13 12:21:06 +08:00
|
|
|
}
|
2010-12-01 04:27:11 +08:00
|
|
|
EXPORT_SYMBOL_GPL(inet_putpeer);
|