fs/mbcache.c: doucple the locking of local from global data
authorT Makphaibulchoke <tmac@hp.com>
Tue, 18 Mar 2014 23:23:20 +0000 (19:23 -0400)
committerTheodore Ts'o <tytso@mit.edu>
Tue, 18 Mar 2014 23:23:20 +0000 (19:23 -0400)
The patch increases the parallelism of mbcache by using the built-in
lock in the hlist_bl_node to protect the mb_cache's local block and
index hash chains.  The global data mb_cache_lru_list and
mb_cache_list continue to be protected by the global
mb_cache_spinlock.

New block group spinlock, mb_cache_bg_lock is also added to serialize
accesses to mb_cache_entry's local data.

A new member e_refcnt is added to the mb_cache_entry structure to help
preventing an mb_cache_entry from being deallocated by a free while it
is being referenced by either mb_cache_entry_get() or
mb_cache_entry_find().

Signed-off-by: T. Makphaibulchoke <tmac@hp.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
fs/mbcache.c

index 55db0daaca74c2f8ee66ca04f7e5900b0594ddc1..786ecab81c99c5c4765f8d976c32a1d4029b9179 100644 (file)
  * back on the lru list.
  */
 
+/*
+ * Lock descriptions and usage:
+ *
+ * Each hash chain of both the block and index hash tables now contains
+ * a built-in lock used to serialize accesses to the hash chain.
+ *
+ * Accesses to global data structures mb_cache_list and mb_cache_lru_list
+ * are serialized via the global spinlock mb_cache_spinlock.
+ *
+ * Each mb_cache_entry contains a spinlock, e_entry_lock, to serialize
+ * accesses to its local data, such as e_used and e_queued.
+ *
+ * Lock ordering:
+ *
+ * Each block hash chain's lock has the highest lock order, followed by an
+ * index hash chain's lock, mb_cache_bg_lock (used to implement mb_cache_entry's
+ * lock), and mb_cach_spinlock, with the lowest order.  While holding
+ * either a block or index hash chain lock, a thread can acquire an
+ * mc_cache_bg_lock, which in turn can also acquire mb_cache_spinlock.
+ *
+ * Synchronization:
+ *
+ * Since both mb_cache_entry_get and mb_cache_entry_find scan the block and
+ * index hash chian, it needs to lock the corresponding hash chain.  For each
+ * mb_cache_entry within the chain, it needs to lock the mb_cache_entry to
+ * prevent either any simultaneous release or free on the entry and also
+ * to serialize accesses to either the e_used or e_queued member of the entry.
+ *
+ * To avoid having a dangling reference to an already freed
+ * mb_cache_entry, an mb_cache_entry is only freed when it is not on a
+ * block hash chain and also no longer being referenced, both e_used,
+ * and e_queued are 0's.  When an mb_cache_entry is explicitly freed it is
+ * first removed from a block hash chain.
+ */
+
 #include <linux/kernel.h>
 #include <linux/module.h>
 
@@ -37,6 +72,7 @@
 #include <linux/list_bl.h>
 #include <linux/mbcache.h>
 #include <linux/init.h>
+#include <linux/blockgroup_lock.h>
 
 #ifdef MB_CACHE_DEBUG
 # define mb_debug(f...) do { \
 
 #define MB_CACHE_WRITER ((unsigned short)~0U >> 1)
 
+#define MB_CACHE_ENTRY_LOCK_BITS       __builtin_log2(NR_BG_LOCKS)
+#define        MB_CACHE_ENTRY_LOCK_INDEX(ce)                   \
+       (hash_long((unsigned long)ce, MB_CACHE_ENTRY_LOCK_BITS))
+
 static DECLARE_WAIT_QUEUE_HEAD(mb_cache_queue);
-               
+static struct blockgroup_lock *mb_cache_bg_lock;
+
 MODULE_AUTHOR("Andreas Gruenbacher <a.gruenbacher@computer.org>");
 MODULE_DESCRIPTION("Meta block cache (for extended attributes)");
 MODULE_LICENSE("GPL");
@@ -86,6 +127,20 @@ static LIST_HEAD(mb_cache_list);
 static LIST_HEAD(mb_cache_lru_list);
 static DEFINE_SPINLOCK(mb_cache_spinlock);
 
+static inline void
+__spin_lock_mb_cache_entry(struct mb_cache_entry *ce)
+{
+       spin_lock(bgl_lock_ptr(mb_cache_bg_lock,
+               MB_CACHE_ENTRY_LOCK_INDEX(ce)));
+}
+
+static inline void
+__spin_unlock_mb_cache_entry(struct mb_cache_entry *ce)
+{
+       spin_unlock(bgl_lock_ptr(mb_cache_bg_lock,
+               MB_CACHE_ENTRY_LOCK_INDEX(ce)));
+}
+
 static inline int
 __mb_cache_entry_is_block_hashed(struct mb_cache_entry *ce)
 {
@@ -113,11 +168,21 @@ __mb_cache_entry_unhash_index(struct mb_cache_entry *ce)
                hlist_bl_del_init(&ce->e_index.o_list);
 }
 
+/*
+ * __mb_cache_entry_unhash_unlock()
+ *
+ * This function is called to unhash both the block and index hash
+ * chain.
+ * It assumes both the block and index hash chain is locked upon entry.
+ * It also unlock both hash chains both exit
+ */
 static inline void
-__mb_cache_entry_unhash(struct mb_cache_entry *ce)
+__mb_cache_entry_unhash_unlock(struct mb_cache_entry *ce)
 {
        __mb_cache_entry_unhash_index(ce);
+       hlist_bl_unlock(ce->e_index_hash_p);
        __mb_cache_entry_unhash_block(ce);
+       hlist_bl_unlock(ce->e_block_hash_p);
 }
 
 static void
@@ -125,36 +190,47 @@ __mb_cache_entry_forget(struct mb_cache_entry *ce, gfp_t gfp_mask)
 {
        struct mb_cache *cache = ce->e_cache;
 
-       mb_assert(!(ce->e_used || ce->e_queued));
+       mb_assert(!(ce->e_used || ce->e_queued || atomic_read(&ce->e_refcnt)));
        kmem_cache_free(cache->c_entry_cache, ce);
        atomic_dec(&cache->c_entry_count);
 }
 
-
 static void
-__mb_cache_entry_release_unlock(struct mb_cache_entry *ce)
-       __releases(mb_cache_spinlock)
+__mb_cache_entry_release(struct mb_cache_entry *ce)
 {
+       /* First lock the entry to serialize access to its local data. */
+       __spin_lock_mb_cache_entry(ce);
        /* Wake up all processes queuing for this cache entry. */
        if (ce->e_queued)
                wake_up_all(&mb_cache_queue);
        if (ce->e_used >= MB_CACHE_WRITER)
                ce->e_used -= MB_CACHE_WRITER;
+       /*
+        * Make sure that all cache entries on lru_list have
+        * both e_used and e_qued of 0s.
+        */
        ce->e_used--;
-       if (!(ce->e_used || ce->e_queued)) {
-               if (!__mb_cache_entry_is_block_hashed(ce))
+       if (!(ce->e_used || ce->e_queued || atomic_read(&ce->e_refcnt))) {
+               if (!__mb_cache_entry_is_block_hashed(ce)) {
+                       __spin_unlock_mb_cache_entry(ce);
                        goto forget;
-               mb_assert(list_empty(&ce->e_lru_list));
-               list_add_tail(&ce->e_lru_list, &mb_cache_lru_list);
+               }
+               /*
+                * Need access to lru list, first drop entry lock,
+                * then reacquire the lock in the proper order.
+                */
+               spin_lock(&mb_cache_spinlock);
+               if (list_empty(&ce->e_lru_list))
+                       list_add_tail(&ce->e_lru_list, &mb_cache_lru_list);
+               spin_unlock(&mb_cache_spinlock);
        }
-       spin_unlock(&mb_cache_spinlock);
+       __spin_unlock_mb_cache_entry(ce);
        return;
 forget:
-       spin_unlock(&mb_cache_spinlock);
+       mb_assert(list_empty(&ce->e_lru_list));
        __mb_cache_entry_forget(ce, GFP_KERNEL);
 }
 
-
 /*
  * mb_cache_shrink_scan()  memory pressure callback
  *
@@ -177,17 +253,34 @@ mb_cache_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
 
        mb_debug("trying to free %d entries", nr_to_scan);
        spin_lock(&mb_cache_spinlock);
-       while (nr_to_scan-- && !list_empty(&mb_cache_lru_list)) {
+       while ((nr_to_scan-- > 0) && !list_empty(&mb_cache_lru_list)) {
                struct mb_cache_entry *ce =
                        list_entry(mb_cache_lru_list.next,
-                                  struct mb_cache_entry, e_lru_list);
-               list_move_tail(&ce->e_lru_list, &free_list);
-               __mb_cache_entry_unhash(ce);
-               freed++;
+                               struct mb_cache_entry, e_lru_list);
+               list_del_init(&ce->e_lru_list);
+               if (ce->e_used || ce->e_queued || atomic_read(&ce->e_refcnt))
+                       continue;
+               spin_unlock(&mb_cache_spinlock);
+               /* Prevent any find or get operation on the entry */
+               hlist_bl_lock(ce->e_block_hash_p);
+               hlist_bl_lock(ce->e_index_hash_p);
+               /* Ignore if it is touched by a find/get */
+               if (ce->e_used || ce->e_queued || atomic_read(&ce->e_refcnt) ||
+                       !list_empty(&ce->e_lru_list)) {
+                       hlist_bl_unlock(ce->e_index_hash_p);
+                       hlist_bl_unlock(ce->e_block_hash_p);
+                       spin_lock(&mb_cache_spinlock);
+                       continue;
+               }
+               __mb_cache_entry_unhash_unlock(ce);
+               list_add_tail(&ce->e_lru_list, &free_list);
+               spin_lock(&mb_cache_spinlock);
        }
        spin_unlock(&mb_cache_spinlock);
+
        list_for_each_entry_safe(entry, tmp, &free_list, e_lru_list) {
                __mb_cache_entry_forget(entry, gfp_mask);
+               freed++;
        }
        return freed;
 }
@@ -232,6 +325,14 @@ mb_cache_create(const char *name, int bucket_bits)
        int n, bucket_count = 1 << bucket_bits;
        struct mb_cache *cache = NULL;
 
+       if (!mb_cache_bg_lock) {
+               mb_cache_bg_lock = kmalloc(sizeof(struct blockgroup_lock),
+                       GFP_KERNEL);
+               if (!mb_cache_bg_lock)
+                       return NULL;
+               bgl_lock_init(mb_cache_bg_lock);
+       }
+
        cache = kmalloc(sizeof(struct mb_cache), GFP_KERNEL);
        if (!cache)
                return NULL;
@@ -290,21 +391,47 @@ void
 mb_cache_shrink(struct block_device *bdev)
 {
        LIST_HEAD(free_list);
-       struct list_head *l, *ltmp;
+       struct list_head *l;
+       struct mb_cache_entry *ce, *tmp;
 
+       l = &mb_cache_lru_list;
        spin_lock(&mb_cache_spinlock);
-       list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
-               struct mb_cache_entry *ce =
-                       list_entry(l, struct mb_cache_entry, e_lru_list);
+       while (!list_is_last(l, &mb_cache_lru_list)) {
+               l = l->next;
+               ce = list_entry(l, struct mb_cache_entry, e_lru_list);
                if (ce->e_bdev == bdev) {
-                       list_move_tail(&ce->e_lru_list, &free_list);
-                       __mb_cache_entry_unhash(ce);
+                       list_del_init(&ce->e_lru_list);
+                       if (ce->e_used || ce->e_queued ||
+                               atomic_read(&ce->e_refcnt))
+                               continue;
+                       spin_unlock(&mb_cache_spinlock);
+                       /*
+                        * Prevent any find or get operation on the entry.
+                        */
+                       hlist_bl_lock(ce->e_block_hash_p);
+                       hlist_bl_lock(ce->e_index_hash_p);
+                       /* Ignore if it is touched by a find/get */
+                       if (ce->e_used || ce->e_queued ||
+                               atomic_read(&ce->e_refcnt) ||
+                               !list_empty(&ce->e_lru_list)) {
+                               hlist_bl_unlock(ce->e_index_hash_p);
+                               hlist_bl_unlock(ce->e_block_hash_p);
+                               l = &mb_cache_lru_list;
+                               spin_lock(&mb_cache_spinlock);
+                               continue;
+                       }
+                       __mb_cache_entry_unhash_unlock(ce);
+                       mb_assert(!(ce->e_used || ce->e_queued ||
+                               atomic_read(&ce->e_refcnt)));
+                       list_add_tail(&ce->e_lru_list, &free_list);
+                       l = &mb_cache_lru_list;
+                       spin_lock(&mb_cache_spinlock);
                }
        }
        spin_unlock(&mb_cache_spinlock);
-       list_for_each_safe(l, ltmp, &free_list) {
-               __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
-                                                  e_lru_list), GFP_KERNEL);
+
+       list_for_each_entry_safe(ce, tmp, &free_list, e_lru_list) {
+               __mb_cache_entry_forget(ce, GFP_KERNEL);
        }
 }
 
@@ -320,23 +447,27 @@ void
 mb_cache_destroy(struct mb_cache *cache)
 {
        LIST_HEAD(free_list);
-       struct list_head *l, *ltmp;
+       struct mb_cache_entry *ce, *tmp;
 
        spin_lock(&mb_cache_spinlock);
-       list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
-               struct mb_cache_entry *ce =
-                       list_entry(l, struct mb_cache_entry, e_lru_list);
-               if (ce->e_cache == cache) {
+       list_for_each_entry_safe(ce, tmp, &mb_cache_lru_list, e_lru_list) {
+               if (ce->e_cache == cache)
                        list_move_tail(&ce->e_lru_list, &free_list);
-                       __mb_cache_entry_unhash(ce);
-               }
        }
        list_del(&cache->c_cache_list);
        spin_unlock(&mb_cache_spinlock);
 
-       list_for_each_safe(l, ltmp, &free_list) {
-               __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
-                                                  e_lru_list), GFP_KERNEL);
+       list_for_each_entry_safe(ce, tmp, &free_list, e_lru_list) {
+               list_del_init(&ce->e_lru_list);
+               /*
+                * Prevent any find or get operation on the entry.
+                */
+               hlist_bl_lock(ce->e_block_hash_p);
+               hlist_bl_lock(ce->e_index_hash_p);
+               mb_assert(!(ce->e_used || ce->e_queued ||
+                       atomic_read(&ce->e_refcnt)));
+               __mb_cache_entry_unhash_unlock(ce);
+               __mb_cache_entry_forget(ce, GFP_KERNEL);
        }
 
        if (atomic_read(&cache->c_entry_count) > 0) {
@@ -345,8 +476,6 @@ mb_cache_destroy(struct mb_cache *cache)
                          atomic_read(&cache->c_entry_count));
        }
 
-       kmem_cache_destroy(cache->c_entry_cache);
-
        kfree(cache->c_index_hash);
        kfree(cache->c_block_hash);
        kfree(cache);
@@ -363,29 +492,59 @@ mb_cache_destroy(struct mb_cache *cache)
 struct mb_cache_entry *
 mb_cache_entry_alloc(struct mb_cache *cache, gfp_t gfp_flags)
 {
-       struct mb_cache_entry *ce = NULL;
+       struct mb_cache_entry *ce;
 
        if (atomic_read(&cache->c_entry_count) >= cache->c_max_entries) {
+               struct list_head *l;
+
+               l = &mb_cache_lru_list;
                spin_lock(&mb_cache_spinlock);
-               if (!list_empty(&mb_cache_lru_list)) {
-                       ce = list_entry(mb_cache_lru_list.next,
-                                       struct mb_cache_entry, e_lru_list);
-                       list_del_init(&ce->e_lru_list);
-                       __mb_cache_entry_unhash(ce);
+               while (!list_is_last(l, &mb_cache_lru_list)) {
+                       l = l->next;
+                       ce = list_entry(l, struct mb_cache_entry, e_lru_list);
+                       if (ce->e_cache == cache) {
+                               list_del_init(&ce->e_lru_list);
+                               if (ce->e_used || ce->e_queued ||
+                                       atomic_read(&ce->e_refcnt))
+                                       continue;
+                               spin_unlock(&mb_cache_spinlock);
+                               /*
+                                * Prevent any find or get operation on the
+                                * entry.
+                                */
+                               hlist_bl_lock(ce->e_block_hash_p);
+                               hlist_bl_lock(ce->e_index_hash_p);
+                               /* Ignore if it is touched by a find/get */
+                               if (ce->e_used || ce->e_queued ||
+                                       atomic_read(&ce->e_refcnt) ||
+                                       !list_empty(&ce->e_lru_list)) {
+                                       hlist_bl_unlock(ce->e_index_hash_p);
+                                       hlist_bl_unlock(ce->e_block_hash_p);
+                                       l = &mb_cache_lru_list;
+                                       spin_lock(&mb_cache_spinlock);
+                                       continue;
+                               }
+                               mb_assert(list_empty(&ce->e_lru_list));
+                               mb_assert(!(ce->e_used || ce->e_queued ||
+                                       atomic_read(&ce->e_refcnt)));
+                               __mb_cache_entry_unhash_unlock(ce);
+                               goto found;
+                       }
                }
                spin_unlock(&mb_cache_spinlock);
        }
-       if (!ce) {
-               ce = kmem_cache_alloc(cache->c_entry_cache, gfp_flags);
-               if (!ce)
-                       return NULL;
-               atomic_inc(&cache->c_entry_count);
-               INIT_LIST_HEAD(&ce->e_lru_list);
-               INIT_HLIST_BL_NODE(&ce->e_block_list);
-               INIT_HLIST_BL_NODE(&ce->e_index.o_list);
-               ce->e_cache = cache;
-               ce->e_queued = 0;
-       }
+
+       ce = kmem_cache_alloc(cache->c_entry_cache, gfp_flags);
+       if (!ce)
+               return NULL;
+       atomic_inc(&cache->c_entry_count);
+       INIT_LIST_HEAD(&ce->e_lru_list);
+       INIT_HLIST_BL_NODE(&ce->e_block_list);
+       INIT_HLIST_BL_NODE(&ce->e_index.o_list);
+       ce->e_cache = cache;
+       ce->e_queued = 0;
+       atomic_set(&ce->e_refcnt, 0);
+found:
        ce->e_block_hash_p = &cache->c_block_hash[0];
        ce->e_index_hash_p = &cache->c_index_hash[0];
        ce->e_used = 1 + MB_CACHE_WRITER;
@@ -414,7 +573,6 @@ mb_cache_entry_insert(struct mb_cache_entry *ce, struct block_device *bdev,
        struct mb_cache *cache = ce->e_cache;
        unsigned int bucket;
        struct hlist_bl_node *l;
-       int error = -EBUSY;
        struct hlist_bl_head *block_hash_p;
        struct hlist_bl_head *index_hash_p;
        struct mb_cache_entry *lce;
@@ -423,26 +581,29 @@ mb_cache_entry_insert(struct mb_cache_entry *ce, struct block_device *bdev,
        bucket = hash_long((unsigned long)bdev + (block & 0xffffffff), 
                           cache->c_bucket_bits);
        block_hash_p = &cache->c_block_hash[bucket];
-       spin_lock(&mb_cache_spinlock);
+       hlist_bl_lock(block_hash_p);
        hlist_bl_for_each_entry(lce, l, block_hash_p, e_block_list) {
-               if (lce->e_bdev == bdev && lce->e_block == block)
-                       goto out;
+               if (lce->e_bdev == bdev && lce->e_block == block) {
+                       hlist_bl_unlock(block_hash_p);
+                       return -EBUSY;
+               }
        }
        mb_assert(!__mb_cache_entry_is_block_hashed(ce));
-       __mb_cache_entry_unhash(ce);
+       __mb_cache_entry_unhash_block(ce);
+       __mb_cache_entry_unhash_index(ce);
        ce->e_bdev = bdev;
        ce->e_block = block;
        ce->e_block_hash_p = block_hash_p;
        ce->e_index.o_key = key;
+       hlist_bl_add_head(&ce->e_block_list, block_hash_p);
+       hlist_bl_unlock(block_hash_p);
        bucket = hash_long(key, cache->c_bucket_bits);
        index_hash_p = &cache->c_index_hash[bucket];
+       hlist_bl_lock(index_hash_p);
        ce->e_index_hash_p = index_hash_p;
        hlist_bl_add_head(&ce->e_index.o_list, index_hash_p);
-       hlist_bl_add_head(&ce->e_block_list, block_hash_p);
-       error = 0;
-out:
-       spin_unlock(&mb_cache_spinlock);
-       return error;
+       hlist_bl_unlock(index_hash_p);
+       return 0;
 }
 
 
@@ -456,24 +617,26 @@ out:
 void
 mb_cache_entry_release(struct mb_cache_entry *ce)
 {
-       spin_lock(&mb_cache_spinlock);
-       __mb_cache_entry_release_unlock(ce);
+       __mb_cache_entry_release(ce);
 }
 
 
 /*
  * mb_cache_entry_free()
  *
- * This is equivalent to the sequence mb_cache_entry_takeout() --
- * mb_cache_entry_release().
  */
 void
 mb_cache_entry_free(struct mb_cache_entry *ce)
 {
-       spin_lock(&mb_cache_spinlock);
+       mb_assert(ce);
        mb_assert(list_empty(&ce->e_lru_list));
-       __mb_cache_entry_unhash(ce);
-       __mb_cache_entry_release_unlock(ce);
+       hlist_bl_lock(ce->e_index_hash_p);
+       __mb_cache_entry_unhash_index(ce);
+       hlist_bl_unlock(ce->e_index_hash_p);
+       hlist_bl_lock(ce->e_block_hash_p);
+       __mb_cache_entry_unhash_block(ce);
+       hlist_bl_unlock(ce->e_block_hash_p);
+       __mb_cache_entry_release(ce);
 }
 
 
@@ -497,39 +660,48 @@ mb_cache_entry_get(struct mb_cache *cache, struct block_device *bdev,
        bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
                           cache->c_bucket_bits);
        block_hash_p = &cache->c_block_hash[bucket];
-       spin_lock(&mb_cache_spinlock);
+       /* First serialize access to the block corresponding hash chain. */
+       hlist_bl_lock(block_hash_p);
        hlist_bl_for_each_entry(ce, l, block_hash_p, e_block_list) {
                mb_assert(ce->e_block_hash_p == block_hash_p);
                if (ce->e_bdev == bdev && ce->e_block == block) {
-                       DEFINE_WAIT(wait);
+                       /*
+                        * Prevent a free from removing the entry.
+                        */
+                       atomic_inc(&ce->e_refcnt);
+                       hlist_bl_unlock(block_hash_p);
+                       __spin_lock_mb_cache_entry(ce);
+                       atomic_dec(&ce->e_refcnt);
+                       if (ce->e_used > 0) {
+                               DEFINE_WAIT(wait);
+                               while (ce->e_used > 0) {
+                                       ce->e_queued++;
+                                       prepare_to_wait(&mb_cache_queue, &wait,
+                                                       TASK_UNINTERRUPTIBLE);
+                                       __spin_unlock_mb_cache_entry(ce);
+                                       schedule();
+                                       __spin_lock_mb_cache_entry(ce);
+                                       ce->e_queued--;
+                               }
+                               finish_wait(&mb_cache_queue, &wait);
+                       }
+                       ce->e_used += 1 + MB_CACHE_WRITER;
+                       __spin_unlock_mb_cache_entry(ce);
 
-                       if (!list_empty(&ce->e_lru_list))
+                       if (!list_empty(&ce->e_lru_list)) {
+                               spin_lock(&mb_cache_spinlock);
                                list_del_init(&ce->e_lru_list);
-
-                       while (ce->e_used > 0) {
-                               ce->e_queued++;
-                               prepare_to_wait(&mb_cache_queue, &wait,
-                                               TASK_UNINTERRUPTIBLE);
                                spin_unlock(&mb_cache_spinlock);
-                               schedule();
-                               spin_lock(&mb_cache_spinlock);
-                               ce->e_queued--;
                        }
-                       finish_wait(&mb_cache_queue, &wait);
-                       ce->e_used += 1 + MB_CACHE_WRITER;
-
                        if (!__mb_cache_entry_is_block_hashed(ce)) {
-                               __mb_cache_entry_release_unlock(ce);
+                               __mb_cache_entry_release(ce);
                                return NULL;
                        }
-                       goto cleanup;
+                       return ce;
                }
        }
-       ce = NULL;
-
-cleanup:
-       spin_unlock(&mb_cache_spinlock);
-       return ce;
+       hlist_bl_unlock(block_hash_p);
+       return NULL;
 }
 
 #if !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0)
@@ -538,40 +710,53 @@ static struct mb_cache_entry *
 __mb_cache_entry_find(struct hlist_bl_node *l, struct hlist_bl_head *head,
                      struct block_device *bdev, unsigned int key)
 {
+
+       /* The index hash chain is alredy acquire by caller. */
        while (l != NULL) {
                struct mb_cache_entry *ce =
                        hlist_bl_entry(l, struct mb_cache_entry,
                                e_index.o_list);
                mb_assert(ce->e_index_hash_p == head);
                if (ce->e_bdev == bdev && ce->e_index.o_key == key) {
-                       DEFINE_WAIT(wait);
-
-                       if (!list_empty(&ce->e_lru_list))
-                               list_del_init(&ce->e_lru_list);
-
+                       /*
+                        * Prevent a free from removing the entry.
+                        */
+                       atomic_inc(&ce->e_refcnt);
+                       hlist_bl_unlock(head);
+                       __spin_lock_mb_cache_entry(ce);
+                       atomic_dec(&ce->e_refcnt);
+                       ce->e_used++;
                        /* Incrementing before holding the lock gives readers
                           priority over writers. */
-                       ce->e_used++;
-                       while (ce->e_used >= MB_CACHE_WRITER) {
-                               ce->e_queued++;
-                               prepare_to_wait(&mb_cache_queue, &wait,
-                                               TASK_UNINTERRUPTIBLE);
-                               spin_unlock(&mb_cache_spinlock);
-                               schedule();
+                       if (ce->e_used >= MB_CACHE_WRITER) {
+                               DEFINE_WAIT(wait);
+
+                               while (ce->e_used >= MB_CACHE_WRITER) {
+                                       ce->e_queued++;
+                                       prepare_to_wait(&mb_cache_queue, &wait,
+                                                       TASK_UNINTERRUPTIBLE);
+                                       __spin_unlock_mb_cache_entry(ce);
+                                       schedule();
+                                       __spin_lock_mb_cache_entry(ce);
+                                       ce->e_queued--;
+                               }
+                               finish_wait(&mb_cache_queue, &wait);
+                       }
+                       __spin_unlock_mb_cache_entry(ce);
+                       if (!list_empty(&ce->e_lru_list)) {
                                spin_lock(&mb_cache_spinlock);
-                               ce->e_queued--;
+                               list_del_init(&ce->e_lru_list);
+                               spin_unlock(&mb_cache_spinlock);
                        }
-                       finish_wait(&mb_cache_queue, &wait);
-
                        if (!__mb_cache_entry_is_block_hashed(ce)) {
-                               __mb_cache_entry_release_unlock(ce);
-                               spin_lock(&mb_cache_spinlock);
+                               __mb_cache_entry_release(ce);
                                return ERR_PTR(-EAGAIN);
                        }
                        return ce;
                }
                l = l->next;
        }
+       hlist_bl_unlock(head);
        return NULL;
 }
 
@@ -598,12 +783,12 @@ mb_cache_entry_find_first(struct mb_cache *cache, struct block_device *bdev,
        struct hlist_bl_head *index_hash_p;
 
        index_hash_p = &cache->c_index_hash[bucket];
-       spin_lock(&mb_cache_spinlock);
+       hlist_bl_lock(index_hash_p);
        if (!hlist_bl_empty(index_hash_p)) {
                l = hlist_bl_first(index_hash_p);
                ce = __mb_cache_entry_find(l, index_hash_p, bdev, key);
-       }
-       spin_unlock(&mb_cache_spinlock);
+       } else
+               hlist_bl_unlock(index_hash_p);
        return ce;
 }
 
@@ -638,11 +823,11 @@ mb_cache_entry_find_next(struct mb_cache_entry *prev,
 
        index_hash_p = &cache->c_index_hash[bucket];
        mb_assert(prev->e_index_hash_p == index_hash_p);
-       spin_lock(&mb_cache_spinlock);
+       hlist_bl_lock(index_hash_p);
        mb_assert(!hlist_bl_empty(index_hash_p));
        l = prev->e_index.o_list.next;
        ce = __mb_cache_entry_find(l, index_hash_p, bdev, key);
-       __mb_cache_entry_release_unlock(prev);
+       __mb_cache_entry_release(prev);
        return ce;
 }