aboutsummaryrefslogtreecommitdiffstats
path: root/net/ipv4
diff options
context:
space:
mode:
authorEric Dumazet <eric.dumazet@gmail.com>2010-06-15 04:23:14 -0400
committerDavid S. Miller <davem@davemloft.net>2010-06-15 17:23:38 -0400
commitaa1039e73cc2cf834e99c09d2033d5d2675357b9 (patch)
tree0db06e4adddaf0f77b4e8de170710b74a17375e4 /net/ipv4
parent7b34a4644b4342896e0c1967b8f953213ea4a990 (diff)
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>
Diffstat (limited to 'net/ipv4')
-rw-r--r--net/ipv4/inetpeer.c164
1 files changed, 95 insertions, 69 deletions
diff --git a/net/ipv4/inetpeer.c b/net/ipv4/inetpeer.c
index 035673fd42d4..58fbc7e2475e 100644
--- a/net/ipv4/inetpeer.c
+++ b/net/ipv4/inetpeer.c
@@ -51,8 +51,8 @@
51 * lookups performed with disabled BHs. 51 * lookups performed with disabled BHs.
52 * 52 *
53 * Serialisation issues. 53 * Serialisation issues.
54 * 1. Nodes may appear in the tree only with the pool write lock held. 54 * 1. Nodes may appear in the tree only with the pool lock held.
55 * 2. Nodes may disappear from the tree only with the pool write lock held 55 * 2. Nodes may disappear from the tree only with the pool lock held
56 * AND reference count being 0. 56 * AND reference count being 0.
57 * 3. Nodes appears and disappears from unused node list only under 57 * 3. Nodes appears and disappears from unused node list only under
58 * "inet_peer_unused_lock". 58 * "inet_peer_unused_lock".
@@ -80,11 +80,11 @@ static const struct inet_peer peer_fake_node = {
80 80
81static struct { 81static struct {
82 struct inet_peer *root; 82 struct inet_peer *root;
83 rwlock_t lock; 83 spinlock_t lock;
84 int total; 84 int total;
85} peers = { 85} peers = {
86 .root = peer_avl_empty, 86 .root = peer_avl_empty,
87 .lock = __RW_LOCK_UNLOCKED(peers.lock), 87 .lock = __SPIN_LOCK_UNLOCKED(peers.lock),
88 .total = 0, 88 .total = 0,
89}; 89};
90#define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */ 90#define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
@@ -129,7 +129,7 @@ void __init inet_initpeers(void)
129 129
130 peer_cachep = kmem_cache_create("inet_peer_cache", 130 peer_cachep = kmem_cache_create("inet_peer_cache",
131 sizeof(struct inet_peer), 131 sizeof(struct inet_peer),
132 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, 132 0, SLAB_PANIC,
133 NULL); 133 NULL);
134 134
135 /* All the timers, started at system startup tend 135 /* All the timers, started at system startup tend
@@ -153,16 +153,13 @@ static void unlink_from_unused(struct inet_peer *p)
153 153
154/* 154/*
155 * Called with local BH disabled and the pool lock held. 155 * Called with local BH disabled and the pool lock held.
156 * _stack is known to be NULL or not at compile time,
157 * so compiler will optimize the if (_stack) tests.
158 */ 156 */
159#define lookup(_daddr, _stack) \ 157#define lookup(_daddr, _stack) \
160({ \ 158({ \
161 struct inet_peer *u, **v; \ 159 struct inet_peer *u, **v; \
162 if (_stack != NULL) { \ 160 \
163 stackptr = _stack; \ 161 stackptr = _stack; \
164 *stackptr++ = &peers.root; \ 162 *stackptr++ = &peers.root; \
165 } \
166 for (u = peers.root; u != peer_avl_empty; ) { \ 163 for (u = peers.root; u != peer_avl_empty; ) { \
167 if (_daddr == u->v4daddr) \ 164 if (_daddr == u->v4daddr) \
168 break; \ 165 break; \
@@ -170,14 +167,41 @@ static void unlink_from_unused(struct inet_peer *p)
170 v = &u->avl_left; \ 167 v = &u->avl_left; \
171 else \ 168 else \
172 v = &u->avl_right; \ 169 v = &u->avl_right; \
173 if (_stack != NULL) \ 170 *stackptr++ = v; \
174 *stackptr++ = v; \
175 u = *v; \ 171 u = *v; \
176 } \ 172 } \
177 u; \ 173 u; \
178}) 174})
179 175
180/* Called with local BH disabled and the pool write lock held. */ 176/*
177 * Called with rcu_read_lock_bh()
178 * Because we hold no lock against a writer, its quite possible we fall
179 * in an endless loop.
180 * But every pointer we follow is guaranteed to be valid thanks to RCU.
181 * We exit from this function if number of links exceeds PEER_MAXDEPTH
182 */
183static struct inet_peer *lookup_rcu_bh(__be32 daddr)
184{
185 struct inet_peer *u = rcu_dereference_bh(peers.root);
186 int count = 0;
187
188 while (u != peer_avl_empty) {
189 if (daddr == u->v4daddr) {
190 if (unlikely(!atomic_inc_not_zero(&u->refcnt)))
191 u = NULL;
192 return u;
193 }
194 if ((__force __u32)daddr < (__force __u32)u->v4daddr)
195 u = rcu_dereference_bh(u->avl_left);
196 else
197 u = rcu_dereference_bh(u->avl_right);
198 if (unlikely(++count == PEER_MAXDEPTH))
199 break;
200 }
201 return NULL;
202}
203
204/* Called with local BH disabled and the pool lock held. */
181#define lookup_rightempty(start) \ 205#define lookup_rightempty(start) \
182({ \ 206({ \
183 struct inet_peer *u, **v; \ 207 struct inet_peer *u, **v; \
@@ -191,9 +215,10 @@ static void unlink_from_unused(struct inet_peer *p)
191 u; \ 215 u; \
192}) 216})
193 217
194/* Called with local BH disabled and the pool write lock held. 218/* Called with local BH disabled and the pool lock held.
195 * Variable names are the proof of operation correctness. 219 * Variable names are the proof of operation correctness.
196 * Look into mm/map_avl.c for more detail description of the ideas. */ 220 * Look into mm/map_avl.c for more detail description of the ideas.
221 */
197static void peer_avl_rebalance(struct inet_peer **stack[], 222static void peer_avl_rebalance(struct inet_peer **stack[],
198 struct inet_peer ***stackend) 223 struct inet_peer ***stackend)
199{ 224{
@@ -269,16 +294,22 @@ static void peer_avl_rebalance(struct inet_peer **stack[],
269 } 294 }
270} 295}
271 296
272/* Called with local BH disabled and the pool write lock held. */ 297/* Called with local BH disabled and the pool lock held. */
273#define link_to_pool(n) \ 298#define link_to_pool(n) \
274do { \ 299do { \
275 n->avl_height = 1; \ 300 n->avl_height = 1; \
276 n->avl_left = peer_avl_empty; \ 301 n->avl_left = peer_avl_empty; \
277 n->avl_right = peer_avl_empty; \ 302 n->avl_right = peer_avl_empty; \
303 smp_wmb(); /* lockless readers can catch us now */ \
278 **--stackptr = n; \ 304 **--stackptr = n; \
279 peer_avl_rebalance(stack, stackptr); \ 305 peer_avl_rebalance(stack, stackptr); \
280} while (0) 306} while (0)
281 307
308static void inetpeer_free_rcu(struct rcu_head *head)
309{
310 kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
311}
312
282/* May be called with local BH enabled. */ 313/* May be called with local BH enabled. */
283static void unlink_from_pool(struct inet_peer *p) 314static void unlink_from_pool(struct inet_peer *p)
284{ 315{
@@ -286,13 +317,13 @@ static void unlink_from_pool(struct inet_peer *p)
286 317
287 do_free = 0; 318 do_free = 0;
288 319
289 write_lock_bh(&peers.lock); 320 spin_lock_bh(&peers.lock);
290 /* Check the reference counter. It was artificially incremented by 1 321 /* Check the reference counter. It was artificially incremented by 1
291 * in cleanup() function to prevent sudden disappearing. If the 322 * in cleanup() function to prevent sudden disappearing. If we can
292 * reference count is still 1 then the node is referenced only as `p' 323 * atomically (because of lockless readers) take this last reference,
293 * here and from the pool. So under the exclusive pool lock it's safe 324 * it's safe to remove the node and free it later.
294 * to remove the node and free it later. */ 325 */
295 if (atomic_read(&p->refcnt) == 1) { 326 if (atomic_cmpxchg(&p->refcnt, 1, 0) == 1) {
296 struct inet_peer **stack[PEER_MAXDEPTH]; 327 struct inet_peer **stack[PEER_MAXDEPTH];
297 struct inet_peer ***stackptr, ***delp; 328 struct inet_peer ***stackptr, ***delp;
298 if (lookup(p->v4daddr, stack) != p) 329 if (lookup(p->v4daddr, stack) != p)
@@ -321,17 +352,18 @@ static void unlink_from_pool(struct inet_peer *p)
321 peers.total--; 352 peers.total--;
322 do_free = 1; 353 do_free = 1;
323 } 354 }
324 write_unlock_bh(&peers.lock); 355 spin_unlock_bh(&peers.lock);
325 356
326 if (do_free) 357 if (do_free)
327 kmem_cache_free(peer_cachep, p); 358 call_rcu_bh(&p->rcu, inetpeer_free_rcu);
328 else 359 else
329 /* The node is used again. Decrease the reference counter 360 /* The node is used again. Decrease the reference counter
330 * back. The loop "cleanup -> unlink_from_unused 361 * back. The loop "cleanup -> unlink_from_unused
331 * -> unlink_from_pool -> putpeer -> link_to_unused 362 * -> unlink_from_pool -> putpeer -> link_to_unused
332 * -> cleanup (for the same node)" 363 * -> cleanup (for the same node)"
333 * doesn't really exist because the entry will have a 364 * doesn't really exist because the entry will have a
334 * recent deletion time and will not be cleaned again soon. */ 365 * recent deletion time and will not be cleaned again soon.
366 */
335 inet_putpeer(p); 367 inet_putpeer(p);
336} 368}
337 369
@@ -375,62 +407,56 @@ static int cleanup_once(unsigned long ttl)
375/* Called with or without local BH being disabled. */ 407/* Called with or without local BH being disabled. */
376struct inet_peer *inet_getpeer(__be32 daddr, int create) 408struct inet_peer *inet_getpeer(__be32 daddr, int create)
377{ 409{
378 struct inet_peer *p, *n; 410 struct inet_peer *p;
379 struct inet_peer **stack[PEER_MAXDEPTH], ***stackptr; 411 struct inet_peer **stack[PEER_MAXDEPTH], ***stackptr;
380 412
381 /* Look up for the address quickly. */ 413 /* Look up for the address quickly, lockless.
382 read_lock_bh(&peers.lock); 414 * Because of a concurrent writer, we might not find an existing entry.
383 p = lookup(daddr, NULL); 415 */
384 if (p != peer_avl_empty) 416 rcu_read_lock_bh();
385 atomic_inc(&p->refcnt); 417 p = lookup_rcu_bh(daddr);
386 read_unlock_bh(&peers.lock); 418 rcu_read_unlock_bh();
419
420 if (p) {
421 /* The existing node has been found.
422 * Remove the entry from unused list if it was there.
423 */
424 unlink_from_unused(p);
425 return p;
426 }
387 427
428 /* retry an exact lookup, taking the lock before.
429 * At least, nodes should be hot in our cache.
430 */
431 spin_lock_bh(&peers.lock);
432 p = lookup(daddr, stack);
388 if (p != peer_avl_empty) { 433 if (p != peer_avl_empty) {
389 /* The existing node has been found. */ 434 atomic_inc(&p->refcnt);
435 spin_unlock_bh(&peers.lock);
390 /* Remove the entry from unused list if it was there. */ 436 /* Remove the entry from unused list if it was there. */
391 unlink_from_unused(p); 437 unlink_from_unused(p);
392 return p; 438 return p;
393 } 439 }
394 440 p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL;
395 if (!create) 441 if (p) {
396 return NULL; 442 p->v4daddr = daddr;
397 443 atomic_set(&p->refcnt, 1);
398 /* Allocate the space outside the locked region. */ 444 atomic_set(&p->rid, 0);
399 n = kmem_cache_alloc(peer_cachep, GFP_ATOMIC); 445 atomic_set(&p->ip_id_count, secure_ip_id(daddr));
400 if (n == NULL) 446 p->tcp_ts_stamp = 0;
401 return NULL; 447 INIT_LIST_HEAD(&p->unused);
402 n->v4daddr = daddr; 448
403 atomic_set(&n->refcnt, 1); 449
404 atomic_set(&n->rid, 0); 450 /* Link the node. */
405 atomic_set(&n->ip_id_count, secure_ip_id(daddr)); 451 link_to_pool(p);
406 n->tcp_ts_stamp = 0; 452 peers.total++;
407 453 }
408 write_lock_bh(&peers.lock); 454 spin_unlock_bh(&peers.lock);
409 /* Check if an entry has suddenly appeared. */
410 p = lookup(daddr, stack);
411 if (p != peer_avl_empty)
412 goto out_free;
413
414 /* Link the node. */
415 link_to_pool(n);
416 INIT_LIST_HEAD(&n->unused);
417 peers.total++;
418 write_unlock_bh(&peers.lock);
419 455
420 if (peers.total >= inet_peer_threshold) 456 if (peers.total >= inet_peer_threshold)
421 /* Remove one less-recently-used entry. */ 457 /* Remove one less-recently-used entry. */
422 cleanup_once(0); 458 cleanup_once(0);
423 459
424 return n;
425
426out_free:
427 /* The appropriate node is already in the pool. */
428 atomic_inc(&p->refcnt);
429 write_unlock_bh(&peers.lock);
430 /* Remove the entry from unused list if it was there. */
431 unlink_from_unused(p);
432 /* Free preallocated the preallocated node. */
433 kmem_cache_free(peer_cachep, n);
434 return p; 460 return p;
435} 461}
436 462