| blob | bc4aa14506394f97e5e3e25552d46fa61e205a72 |
1 fs/mbcache.c: doucple the locking of local from global data
3 From: T Makphaibulchoke <tmac@hp.com>
5 The patch increases the parallelism of mbcache by using the built-in
6 lock in the hlist_bl_node to protect the mb_cache's local block and
7 index hash chains. The global data mb_cache_lru_list and
8 mb_cache_list continue to be protected by the global
9 mb_cache_spinlock.
11 New block group spinlock, mb_cache_bg_lock is also added to serialize
12 accesses to mb_cache_entry's local data.
14 A new member e_refcnt is added to the mb_cache_entry structure to help
15 preventing an mb_cache_entry from being deallocated by a free while it
16 is being referenced by either mb_cache_entry_get() or
17 mb_cache_entry_find().
19 Signed-off-by: T. Makphaibulchoke <tmac@hp.com>
20 Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
21 ---
22 fs/mbcache.c | 417 ++++++++++++++++++++++++++++++++++++++++++-----------------
23 1 file changed, 301 insertions(+), 116 deletions(-)
25 diff --git a/fs/mbcache.c b/fs/mbcache.c
26 index 55db0da..786ecab 100644
27 --- a/fs/mbcache.c
28 +++ b/fs/mbcache.c
29 @@ -26,6 +26,41 @@
30 * back on the lru list.
31 */
33 +/*
34 + * Lock descriptions and usage:
35 + *
36 + * Each hash chain of both the block and index hash tables now contains
37 + * a built-in lock used to serialize accesses to the hash chain.
38 + *
39 + * Accesses to global data structures mb_cache_list and mb_cache_lru_list
40 + * are serialized via the global spinlock mb_cache_spinlock.
41 + *
42 + * Each mb_cache_entry contains a spinlock, e_entry_lock, to serialize
43 + * accesses to its local data, such as e_used and e_queued.
44 + *
45 + * Lock ordering:
46 + *
47 + * Each block hash chain's lock has the highest lock order, followed by an
48 + * index hash chain's lock, mb_cache_bg_lock (used to implement mb_cache_entry's
49 + * lock), and mb_cach_spinlock, with the lowest order. While holding
50 + * either a block or index hash chain lock, a thread can acquire an
51 + * mc_cache_bg_lock, which in turn can also acquire mb_cache_spinlock.
52 + *
53 + * Synchronization:
54 + *
55 + * Since both mb_cache_entry_get and mb_cache_entry_find scan the block and
56 + * index hash chian, it needs to lock the corresponding hash chain. For each
57 + * mb_cache_entry within the chain, it needs to lock the mb_cache_entry to
58 + * prevent either any simultaneous release or free on the entry and also
59 + * to serialize accesses to either the e_used or e_queued member of the entry.
60 + *
61 + * To avoid having a dangling reference to an already freed
62 + * mb_cache_entry, an mb_cache_entry is only freed when it is not on a
63 + * block hash chain and also no longer being referenced, both e_used,
64 + * and e_queued are 0's. When an mb_cache_entry is explicitly freed it is
65 + * first removed from a block hash chain.
66 + */
67 +
68 #include <linux/kernel.h>
69 #include <linux/module.h>
71 @@ -37,6 +72,7 @@
72 #include <linux/list_bl.h>
73 #include <linux/mbcache.h>
74 #include <linux/init.h>
75 +#include <linux/blockgroup_lock.h>
77 #ifdef MB_CACHE_DEBUG
78 # define mb_debug(f...) do { \
79 @@ -57,8 +93,13 @@
81 #define MB_CACHE_WRITER ((unsigned short)~0U >> 1)
83 +#define MB_CACHE_ENTRY_LOCK_BITS __builtin_log2(NR_BG_LOCKS)
84 +#define MB_CACHE_ENTRY_LOCK_INDEX(ce) \
85 + (hash_long((unsigned long)ce, MB_CACHE_ENTRY_LOCK_BITS))
86 +
87 static DECLARE_WAIT_QUEUE_HEAD(mb_cache_queue);
88 -
89 +static struct blockgroup_lock *mb_cache_bg_lock;
90 +
91 MODULE_AUTHOR("Andreas Gruenbacher <a.gruenbacher@computer.org>");
92 MODULE_DESCRIPTION("Meta block cache (for extended attributes)");
93 MODULE_LICENSE("GPL");
94 @@ -86,6 +127,20 @@ static LIST_HEAD(mb_cache_list);
95 static LIST_HEAD(mb_cache_lru_list);
96 static DEFINE_SPINLOCK(mb_cache_spinlock);
98 +static inline void
99 +__spin_lock_mb_cache_entry(struct mb_cache_entry *ce)
100 +{
101 + spin_lock(bgl_lock_ptr(mb_cache_bg_lock,
102 + MB_CACHE_ENTRY_LOCK_INDEX(ce)));
103 +}
104 +
105 +static inline void
106 +__spin_unlock_mb_cache_entry(struct mb_cache_entry *ce)
107 +{
108 + spin_unlock(bgl_lock_ptr(mb_cache_bg_lock,
109 + MB_CACHE_ENTRY_LOCK_INDEX(ce)));
110 +}
111 +
112 static inline int
113 __mb_cache_entry_is_block_hashed(struct mb_cache_entry *ce)
114 {
115 @@ -113,11 +168,21 @@ __mb_cache_entry_unhash_index(struct mb_cache_entry *ce)
116 hlist_bl_del_init(&ce->e_index.o_list);
117 }
119 +/*
120 + * __mb_cache_entry_unhash_unlock()
121 + *
122 + * This function is called to unhash both the block and index hash
123 + * chain.
124 + * It assumes both the block and index hash chain is locked upon entry.
125 + * It also unlock both hash chains both exit
126 + */
127 static inline void
128 -__mb_cache_entry_unhash(struct mb_cache_entry *ce)
129 +__mb_cache_entry_unhash_unlock(struct mb_cache_entry *ce)
130 {
131 __mb_cache_entry_unhash_index(ce);
132 + hlist_bl_unlock(ce->e_index_hash_p);
133 __mb_cache_entry_unhash_block(ce);
134 + hlist_bl_unlock(ce->e_block_hash_p);
135 }
137 static void
138 @@ -125,36 +190,47 @@ __mb_cache_entry_forget(struct mb_cache_entry *ce, gfp_t gfp_mask)
139 {
140 struct mb_cache *cache = ce->e_cache;
142 - mb_assert(!(ce->e_used || ce->e_queued));
143 + mb_assert(!(ce->e_used || ce->e_queued || atomic_read(&ce->e_refcnt)));
144 kmem_cache_free(cache->c_entry_cache, ce);
145 atomic_dec(&cache->c_entry_count);
146 }
148 -
149 static void
150 -__mb_cache_entry_release_unlock(struct mb_cache_entry *ce)
151 - __releases(mb_cache_spinlock)
152 +__mb_cache_entry_release(struct mb_cache_entry *ce)
153 {
154 + /* First lock the entry to serialize access to its local data. */
155 + __spin_lock_mb_cache_entry(ce);
156 /* Wake up all processes queuing for this cache entry. */
157 if (ce->e_queued)
158 wake_up_all(&mb_cache_queue);
159 if (ce->e_used >= MB_CACHE_WRITER)
160 ce->e_used -= MB_CACHE_WRITER;
161 + /*
162 + * Make sure that all cache entries on lru_list have
163 + * both e_used and e_qued of 0s.
164 + */
165 ce->e_used--;
166 - if (!(ce->e_used || ce->e_queued)) {
167 - if (!__mb_cache_entry_is_block_hashed(ce))
168 + if (!(ce->e_used || ce->e_queued || atomic_read(&ce->e_refcnt))) {
169 + if (!__mb_cache_entry_is_block_hashed(ce)) {
170 + __spin_unlock_mb_cache_entry(ce);
171 goto forget;
172 - mb_assert(list_empty(&ce->e_lru_list));
173 - list_add_tail(&ce->e_lru_list, &mb_cache_lru_list);
174 + }
175 + /*
176 + * Need access to lru list, first drop entry lock,
177 + * then reacquire the lock in the proper order.
178 + */
179 + spin_lock(&mb_cache_spinlock);
180 + if (list_empty(&ce->e_lru_list))
181 + list_add_tail(&ce->e_lru_list, &mb_cache_lru_list);
182 + spin_unlock(&mb_cache_spinlock);
183 }
184 - spin_unlock(&mb_cache_spinlock);
185 + __spin_unlock_mb_cache_entry(ce);
186 return;
187 forget:
188 - spin_unlock(&mb_cache_spinlock);
189 + mb_assert(list_empty(&ce->e_lru_list));
190 __mb_cache_entry_forget(ce, GFP_KERNEL);
191 }
193 -
194 /*
195 * mb_cache_shrink_scan() memory pressure callback
196 *
197 @@ -177,17 +253,34 @@ mb_cache_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
199 mb_debug("trying to free %d entries", nr_to_scan);
200 spin_lock(&mb_cache_spinlock);
201 - while (nr_to_scan-- && !list_empty(&mb_cache_lru_list)) {
202 + while ((nr_to_scan-- > 0) && !list_empty(&mb_cache_lru_list)) {
203 struct mb_cache_entry *ce =
204 list_entry(mb_cache_lru_list.next,
205 - struct mb_cache_entry, e_lru_list);
206 - list_move_tail(&ce->e_lru_list, &free_list);
207 - __mb_cache_entry_unhash(ce);
208 - freed++;
209 + struct mb_cache_entry, e_lru_list);
210 + list_del_init(&ce->e_lru_list);
211 + if (ce->e_used || ce->e_queued || atomic_read(&ce->e_refcnt))
212 + continue;
213 + spin_unlock(&mb_cache_spinlock);
214 + /* Prevent any find or get operation on the entry */
215 + hlist_bl_lock(ce->e_block_hash_p);
216 + hlist_bl_lock(ce->e_index_hash_p);
217 + /* Ignore if it is touched by a find/get */
218 + if (ce->e_used || ce->e_queued || atomic_read(&ce->e_refcnt) ||
219 + !list_empty(&ce->e_lru_list)) {
220 + hlist_bl_unlock(ce->e_index_hash_p);
221 + hlist_bl_unlock(ce->e_block_hash_p);
222 + spin_lock(&mb_cache_spinlock);
223 + continue;
224 + }
225 + __mb_cache_entry_unhash_unlock(ce);
226 + list_add_tail(&ce->e_lru_list, &free_list);
227 + spin_lock(&mb_cache_spinlock);
228 }
229 spin_unlock(&mb_cache_spinlock);
230 +
231 list_for_each_entry_safe(entry, tmp, &free_list, e_lru_list) {
232 __mb_cache_entry_forget(entry, gfp_mask);
233 + freed++;
234 }
235 return freed;
236 }
237 @@ -232,6 +325,14 @@ mb_cache_create(const char *name, int bucket_bits)
238 int n, bucket_count = 1 << bucket_bits;
239 struct mb_cache *cache = NULL;
241 + if (!mb_cache_bg_lock) {
242 + mb_cache_bg_lock = kmalloc(sizeof(struct blockgroup_lock),
243 + GFP_KERNEL);
244 + if (!mb_cache_bg_lock)
245 + return NULL;
246 + bgl_lock_init(mb_cache_bg_lock);
247 + }
248 +
249 cache = kmalloc(sizeof(struct mb_cache), GFP_KERNEL);
250 if (!cache)
251 return NULL;
252 @@ -290,21 +391,47 @@ void
253 mb_cache_shrink(struct block_device *bdev)
254 {
255 LIST_HEAD(free_list);
256 - struct list_head *l, *ltmp;
257 + struct list_head *l;
258 + struct mb_cache_entry *ce, *tmp;
260 + l = &mb_cache_lru_list;
261 spin_lock(&mb_cache_spinlock);
262 - list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
263 - struct mb_cache_entry *ce =
264 - list_entry(l, struct mb_cache_entry, e_lru_list);
265 + while (!list_is_last(l, &mb_cache_lru_list)) {
266 + l = l->next;
267 + ce = list_entry(l, struct mb_cache_entry, e_lru_list);
268 if (ce->e_bdev == bdev) {
269 - list_move_tail(&ce->e_lru_list, &free_list);
270 - __mb_cache_entry_unhash(ce);
271 + list_del_init(&ce->e_lru_list);
272 + if (ce->e_used || ce->e_queued ||
273 + atomic_read(&ce->e_refcnt))
274 + continue;
275 + spin_unlock(&mb_cache_spinlock);
276 + /*
277 + * Prevent any find or get operation on the entry.
278 + */
279 + hlist_bl_lock(ce->e_block_hash_p);
280 + hlist_bl_lock(ce->e_index_hash_p);
281 + /* Ignore if it is touched by a find/get */
282 + if (ce->e_used || ce->e_queued ||
283 + atomic_read(&ce->e_refcnt) ||
284 + !list_empty(&ce->e_lru_list)) {
285 + hlist_bl_unlock(ce->e_index_hash_p);
286 + hlist_bl_unlock(ce->e_block_hash_p);
287 + l = &mb_cache_lru_list;
288 + spin_lock(&mb_cache_spinlock);
289 + continue;
290 + }
291 + __mb_cache_entry_unhash_unlock(ce);
292 + mb_assert(!(ce->e_used || ce->e_queued ||
293 + atomic_read(&ce->e_refcnt)));
294 + list_add_tail(&ce->e_lru_list, &free_list);
295 + l = &mb_cache_lru_list;
296 + spin_lock(&mb_cache_spinlock);
297 }
298 }
299 spin_unlock(&mb_cache_spinlock);
300 - list_for_each_safe(l, ltmp, &free_list) {
301 - __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
302 - e_lru_list), GFP_KERNEL);
303 +
304 + list_for_each_entry_safe(ce, tmp, &free_list, e_lru_list) {
305 + __mb_cache_entry_forget(ce, GFP_KERNEL);
306 }
307 }
309 @@ -320,23 +447,27 @@ void
310 mb_cache_destroy(struct mb_cache *cache)
311 {
312 LIST_HEAD(free_list);
313 - struct list_head *l, *ltmp;
314 + struct mb_cache_entry *ce, *tmp;
316 spin_lock(&mb_cache_spinlock);
317 - list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
318 - struct mb_cache_entry *ce =
319 - list_entry(l, struct mb_cache_entry, e_lru_list);
320 - if (ce->e_cache == cache) {
321 + list_for_each_entry_safe(ce, tmp, &mb_cache_lru_list, e_lru_list) {
322 + if (ce->e_cache == cache)
323 list_move_tail(&ce->e_lru_list, &free_list);
324 - __mb_cache_entry_unhash(ce);
325 - }
326 }
327 list_del(&cache->c_cache_list);
328 spin_unlock(&mb_cache_spinlock);
330 - list_for_each_safe(l, ltmp, &free_list) {
331 - __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
332 - e_lru_list), GFP_KERNEL);
333 + list_for_each_entry_safe(ce, tmp, &free_list, e_lru_list) {
334 + list_del_init(&ce->e_lru_list);
335 + /*
336 + * Prevent any find or get operation on the entry.
337 + */
338 + hlist_bl_lock(ce->e_block_hash_p);
339 + hlist_bl_lock(ce->e_index_hash_p);
340 + mb_assert(!(ce->e_used || ce->e_queued ||
341 + atomic_read(&ce->e_refcnt)));
342 + __mb_cache_entry_unhash_unlock(ce);
343 + __mb_cache_entry_forget(ce, GFP_KERNEL);
344 }
346 if (atomic_read(&cache->c_entry_count) > 0) {
347 @@ -345,8 +476,6 @@ mb_cache_destroy(struct mb_cache *cache)
348 atomic_read(&cache->c_entry_count));
349 }
351 - kmem_cache_destroy(cache->c_entry_cache);
352 -
353 kfree(cache->c_index_hash);
354 kfree(cache->c_block_hash);
355 kfree(cache);
356 @@ -363,29 +492,59 @@ mb_cache_destroy(struct mb_cache *cache)
357 struct mb_cache_entry *
358 mb_cache_entry_alloc(struct mb_cache *cache, gfp_t gfp_flags)
359 {
360 - struct mb_cache_entry *ce = NULL;
361 + struct mb_cache_entry *ce;
363 if (atomic_read(&cache->c_entry_count) >= cache->c_max_entries) {
364 + struct list_head *l;
365 +
366 + l = &mb_cache_lru_list;
367 spin_lock(&mb_cache_spinlock);
368 - if (!list_empty(&mb_cache_lru_list)) {
369 - ce = list_entry(mb_cache_lru_list.next,
370 - struct mb_cache_entry, e_lru_list);
371 - list_del_init(&ce->e_lru_list);
372 - __mb_cache_entry_unhash(ce);
373 + while (!list_is_last(l, &mb_cache_lru_list)) {
374 + l = l->next;
375 + ce = list_entry(l, struct mb_cache_entry, e_lru_list);
376 + if (ce->e_cache == cache) {
377 + list_del_init(&ce->e_lru_list);
378 + if (ce->e_used || ce->e_queued ||
379 + atomic_read(&ce->e_refcnt))
380 + continue;
381 + spin_unlock(&mb_cache_spinlock);
382 + /*
383 + * Prevent any find or get operation on the
384 + * entry.
385 + */
386 + hlist_bl_lock(ce->e_block_hash_p);
387 + hlist_bl_lock(ce->e_index_hash_p);
388 + /* Ignore if it is touched by a find/get */
389 + if (ce->e_used || ce->e_queued ||
390 + atomic_read(&ce->e_refcnt) ||
391 + !list_empty(&ce->e_lru_list)) {
392 + hlist_bl_unlock(ce->e_index_hash_p);
393 + hlist_bl_unlock(ce->e_block_hash_p);
394 + l = &mb_cache_lru_list;
395 + spin_lock(&mb_cache_spinlock);
396 + continue;
397 + }
398 + mb_assert(list_empty(&ce->e_lru_list));
399 + mb_assert(!(ce->e_used || ce->e_queued ||
400 + atomic_read(&ce->e_refcnt)));
401 + __mb_cache_entry_unhash_unlock(ce);
402 + goto found;
403 + }
404 }
405 spin_unlock(&mb_cache_spinlock);
406 }
407 - if (!ce) {
408 - ce = kmem_cache_alloc(cache->c_entry_cache, gfp_flags);
409 - if (!ce)
410 - return NULL;
411 - atomic_inc(&cache->c_entry_count);
412 - INIT_LIST_HEAD(&ce->e_lru_list);
413 - INIT_HLIST_BL_NODE(&ce->e_block_list);
414 - INIT_HLIST_BL_NODE(&ce->e_index.o_list);
415 - ce->e_cache = cache;
416 - ce->e_queued = 0;
417 - }
418 +
419 + ce = kmem_cache_alloc(cache->c_entry_cache, gfp_flags);
420 + if (!ce)
421 + return NULL;
422 + atomic_inc(&cache->c_entry_count);
423 + INIT_LIST_HEAD(&ce->e_lru_list);
424 + INIT_HLIST_BL_NODE(&ce->e_block_list);
425 + INIT_HLIST_BL_NODE(&ce->e_index.o_list);
426 + ce->e_cache = cache;
427 + ce->e_queued = 0;
428 + atomic_set(&ce->e_refcnt, 0);
429 +found:
430 ce->e_block_hash_p = &cache->c_block_hash[0];
431 ce->e_index_hash_p = &cache->c_index_hash[0];
432 ce->e_used = 1 + MB_CACHE_WRITER;
433 @@ -414,7 +573,6 @@ mb_cache_entry_insert(struct mb_cache_entry *ce, struct block_device *bdev,
434 struct mb_cache *cache = ce->e_cache;
435 unsigned int bucket;
436 struct hlist_bl_node *l;
437 - int error = -EBUSY;
438 struct hlist_bl_head *block_hash_p;
439 struct hlist_bl_head *index_hash_p;
440 struct mb_cache_entry *lce;
441 @@ -423,26 +581,29 @@ mb_cache_entry_insert(struct mb_cache_entry *ce, struct block_device *bdev,
442 bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
443 cache->c_bucket_bits);
444 block_hash_p = &cache->c_block_hash[bucket];
445 - spin_lock(&mb_cache_spinlock);
446 + hlist_bl_lock(block_hash_p);
447 hlist_bl_for_each_entry(lce, l, block_hash_p, e_block_list) {
448 - if (lce->e_bdev == bdev && lce->e_block == block)
449 - goto out;
450 + if (lce->e_bdev == bdev && lce->e_block == block) {
451 + hlist_bl_unlock(block_hash_p);
452 + return -EBUSY;
453 + }
454 }
455 mb_assert(!__mb_cache_entry_is_block_hashed(ce));
456 - __mb_cache_entry_unhash(ce);
457 + __mb_cache_entry_unhash_block(ce);
458 + __mb_cache_entry_unhash_index(ce);
459 ce->e_bdev = bdev;
460 ce->e_block = block;
461 ce->e_block_hash_p = block_hash_p;
462 ce->e_index.o_key = key;
463 + hlist_bl_add_head(&ce->e_block_list, block_hash_p);
464 + hlist_bl_unlock(block_hash_p);
465 bucket = hash_long(key, cache->c_bucket_bits);
466 index_hash_p = &cache->c_index_hash[bucket];
467 + hlist_bl_lock(index_hash_p);
468 ce->e_index_hash_p = index_hash_p;
469 hlist_bl_add_head(&ce->e_index.o_list, index_hash_p);
470 - hlist_bl_add_head(&ce->e_block_list, block_hash_p);
471 - error = 0;
472 -out:
473 - spin_unlock(&mb_cache_spinlock);
474 - return error;
475 + hlist_bl_unlock(index_hash_p);
476 + return 0;
477 }
480 @@ -456,24 +617,26 @@ out:
481 void
482 mb_cache_entry_release(struct mb_cache_entry *ce)
483 {
484 - spin_lock(&mb_cache_spinlock);
485 - __mb_cache_entry_release_unlock(ce);
486 + __mb_cache_entry_release(ce);
487 }
490 /*
491 * mb_cache_entry_free()
492 *
493 - * This is equivalent to the sequence mb_cache_entry_takeout() --
494 - * mb_cache_entry_release().
495 */
496 void
497 mb_cache_entry_free(struct mb_cache_entry *ce)
498 {
499 - spin_lock(&mb_cache_spinlock);
500 + mb_assert(ce);
501 mb_assert(list_empty(&ce->e_lru_list));
502 - __mb_cache_entry_unhash(ce);
503 - __mb_cache_entry_release_unlock(ce);
504 + hlist_bl_lock(ce->e_index_hash_p);
505 + __mb_cache_entry_unhash_index(ce);
506 + hlist_bl_unlock(ce->e_index_hash_p);
507 + hlist_bl_lock(ce->e_block_hash_p);
508 + __mb_cache_entry_unhash_block(ce);
509 + hlist_bl_unlock(ce->e_block_hash_p);
510 + __mb_cache_entry_release(ce);
511 }
514 @@ -497,39 +660,48 @@ mb_cache_entry_get(struct mb_cache *cache, struct block_device *bdev,
515 bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
516 cache->c_bucket_bits);
517 block_hash_p = &cache->c_block_hash[bucket];
518 - spin_lock(&mb_cache_spinlock);
519 + /* First serialize access to the block corresponding hash chain. */
520 + hlist_bl_lock(block_hash_p);
521 hlist_bl_for_each_entry(ce, l, block_hash_p, e_block_list) {
522 mb_assert(ce->e_block_hash_p == block_hash_p);
523 if (ce->e_bdev == bdev && ce->e_block == block) {
524 - DEFINE_WAIT(wait);
525 + /*
526 + * Prevent a free from removing the entry.
527 + */
528 + atomic_inc(&ce->e_refcnt);
529 + hlist_bl_unlock(block_hash_p);
530 + __spin_lock_mb_cache_entry(ce);
531 + atomic_dec(&ce->e_refcnt);
532 + if (ce->e_used > 0) {
533 + DEFINE_WAIT(wait);
534 + while (ce->e_used > 0) {
535 + ce->e_queued++;
536 + prepare_to_wait(&mb_cache_queue, &wait,
537 + TASK_UNINTERRUPTIBLE);
538 + __spin_unlock_mb_cache_entry(ce);
539 + schedule();
540 + __spin_lock_mb_cache_entry(ce);
541 + ce->e_queued--;
542 + }
543 + finish_wait(&mb_cache_queue, &wait);
544 + }
545 + ce->e_used += 1 + MB_CACHE_WRITER;
546 + __spin_unlock_mb_cache_entry(ce);
548 - if (!list_empty(&ce->e_lru_list))
549 + if (!list_empty(&ce->e_lru_list)) {
550 + spin_lock(&mb_cache_spinlock);
551 list_del_init(&ce->e_lru_list);
552 -
553 - while (ce->e_used > 0) {
554 - ce->e_queued++;
555 - prepare_to_wait(&mb_cache_queue, &wait,
556 - TASK_UNINTERRUPTIBLE);
557 spin_unlock(&mb_cache_spinlock);
558 - schedule();
559 - spin_lock(&mb_cache_spinlock);
560 - ce->e_queued--;
561 }
562 - finish_wait(&mb_cache_queue, &wait);
563 - ce->e_used += 1 + MB_CACHE_WRITER;
564 -
565 if (!__mb_cache_entry_is_block_hashed(ce)) {
566 - __mb_cache_entry_release_unlock(ce);
567 + __mb_cache_entry_release(ce);
568 return NULL;
569 }
570 - goto cleanup;
571 + return ce;
572 }
573 }
574 - ce = NULL;
575 -
576 -cleanup:
577 - spin_unlock(&mb_cache_spinlock);
578 - return ce;
579 + hlist_bl_unlock(block_hash_p);
580 + return NULL;
581 }
583 #if !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0)
584 @@ -538,40 +710,53 @@ static struct mb_cache_entry *
585 __mb_cache_entry_find(struct hlist_bl_node *l, struct hlist_bl_head *head,
586 struct block_device *bdev, unsigned int key)
587 {
588 +
589 + /* The index hash chain is alredy acquire by caller. */
590 while (l != NULL) {
591 struct mb_cache_entry *ce =
592 hlist_bl_entry(l, struct mb_cache_entry,
593 e_index.o_list);
594 mb_assert(ce->e_index_hash_p == head);
595 if (ce->e_bdev == bdev && ce->e_index.o_key == key) {
596 - DEFINE_WAIT(wait);
597 -
598 - if (!list_empty(&ce->e_lru_list))
599 - list_del_init(&ce->e_lru_list);
600 -
601 + /*
602 + * Prevent a free from removing the entry.
603 + */
604 + atomic_inc(&ce->e_refcnt);
605 + hlist_bl_unlock(head);
606 + __spin_lock_mb_cache_entry(ce);
607 + atomic_dec(&ce->e_refcnt);
608 + ce->e_used++;
609 /* Incrementing before holding the lock gives readers
610 priority over writers. */
611 - ce->e_used++;
612 - while (ce->e_used >= MB_CACHE_WRITER) {
613 - ce->e_queued++;
614 - prepare_to_wait(&mb_cache_queue, &wait,
615 - TASK_UNINTERRUPTIBLE);
616 - spin_unlock(&mb_cache_spinlock);
617 - schedule();
618 + if (ce->e_used >= MB_CACHE_WRITER) {
619 + DEFINE_WAIT(wait);
620 +
621 + while (ce->e_used >= MB_CACHE_WRITER) {
622 + ce->e_queued++;
623 + prepare_to_wait(&mb_cache_queue, &wait,
624 + TASK_UNINTERRUPTIBLE);
625 + __spin_unlock_mb_cache_entry(ce);
626 + schedule();
627 + __spin_lock_mb_cache_entry(ce);
628 + ce->e_queued--;
629 + }
630 + finish_wait(&mb_cache_queue, &wait);
631 + }
632 + __spin_unlock_mb_cache_entry(ce);
633 + if (!list_empty(&ce->e_lru_list)) {
634 spin_lock(&mb_cache_spinlock);
635 - ce->e_queued--;
636 + list_del_init(&ce->e_lru_list);
637 + spin_unlock(&mb_cache_spinlock);
638 }
639 - finish_wait(&mb_cache_queue, &wait);
640 -
641 if (!__mb_cache_entry_is_block_hashed(ce)) {
642 - __mb_cache_entry_release_unlock(ce);
643 - spin_lock(&mb_cache_spinlock);
644 + __mb_cache_entry_release(ce);
645 return ERR_PTR(-EAGAIN);
646 }
647 return ce;
648 }
649 l = l->next;
650 }
651 + hlist_bl_unlock(head);
652 return NULL;
653 }
655 @@ -598,12 +783,12 @@ mb_cache_entry_find_first(struct mb_cache *cache, struct block_device *bdev,
656 struct hlist_bl_head *index_hash_p;
658 index_hash_p = &cache->c_index_hash[bucket];
659 - spin_lock(&mb_cache_spinlock);
660 + hlist_bl_lock(index_hash_p);
661 if (!hlist_bl_empty(index_hash_p)) {
662 l = hlist_bl_first(index_hash_p);
663 ce = __mb_cache_entry_find(l, index_hash_p, bdev, key);
664 - }
665 - spin_unlock(&mb_cache_spinlock);
666 + } else
667 + hlist_bl_unlock(index_hash_p);
668 return ce;
669 }
671 @@ -638,11 +823,11 @@ mb_cache_entry_find_next(struct mb_cache_entry *prev,
673 index_hash_p = &cache->c_index_hash[bucket];
674 mb_assert(prev->e_index_hash_p == index_hash_p);
675 - spin_lock(&mb_cache_spinlock);
676 + hlist_bl_lock(index_hash_p);
677 mb_assert(!hlist_bl_empty(index_hash_p));
678 l = prev->e_index.o_list.next;
679 ce = __mb_cache_entry_find(l, index_hash_p, bdev, key);
680 - __mb_cache_entry_release_unlock(prev);
681 + __mb_cache_entry_release(prev);
682 return ce;
683 }
685 --
686 1.7.11.3