Whamcloud - gitweb
LU-7689 obdclass: limit lu_site hash table size on clients
[fs/lustre-release.git] / lustre / obdclass / lu_object.c
index 2053bd3..e04654c 100644 (file)
@@ -27,7 +27,7 @@
  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2011, 2014, Intel Corporation.
+ * Copyright (c) 2011, 2015, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -69,6 +69,7 @@ enum {
 
 #define LU_SITE_BITS_MIN    12
 #define LU_SITE_BITS_MAX    24
+#define LU_SITE_BITS_MAX_CL 19
 /**
  * total 256 buckets, we don't want too many buckets because:
  * - consume too much memory
@@ -86,6 +87,7 @@ CFS_MODULE_PARM(lu_cache_nr, "l", long, 0644,
                "Maximum number of objects in lu_object cache");
 
 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
+static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx);
 
 /**
  * Decrease reference counter on object. If last reference is freed, return
@@ -98,7 +100,7 @@ void lu_object_put(const struct lu_env *env, struct lu_object *o)
         struct lu_object_header *top;
         struct lu_site          *site;
         struct lu_object        *orig;
-        cfs_hash_bd_t            bd;
+       struct cfs_hash_bd            bd;
        const struct lu_fid     *fid;
 
         top  = o->lo_header;
@@ -140,8 +142,6 @@ void lu_object_put(const struct lu_env *env, struct lu_object *o)
                return;
        }
 
-        LASSERT(bkt->lsb_busy > 0);
-        bkt->lsb_busy--;
         /*
          * When last reference is released, iterate over object
          * layers, and notify them that object is no longer busy.
@@ -151,15 +151,21 @@ void lu_object_put(const struct lu_env *env, struct lu_object *o)
                         o->lo_ops->loo_object_release(env, o);
         }
 
-        if (!lu_object_is_dying(top)) {
+       if (!lu_object_is_dying(top) &&
+           (lu_object_exists(orig) || lu_object_is_cl(orig))) {
                LASSERT(list_empty(&top->loh_lru));
                list_add_tail(&top->loh_lru, &bkt->lsb_lru);
+               bkt->lsb_lru_len++;
+               lprocfs_counter_incr(site->ls_stats, LU_SS_LRU_LEN);
+               CDEBUG(D_INODE, "Add %p to site lru. hash: %p, bkt: %p, "
+                      "lru_len: %ld\n",
+                      o, site->ls_obj_hash, bkt, bkt->lsb_lru_len);
                 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
                 return;
         }
 
         /*
-         * If object is dying (will not be cached), removed it
+        * If object is dying (will not be cached) then remove it
          * from hash table and LRU.
          *
          * This is done with hash table and LRU lists locked. As the only
@@ -202,11 +208,19 @@ void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
        top = o->lo_header;
        set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
        if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
-               cfs_hash_t *obj_hash = o->lo_dev->ld_site->ls_obj_hash;
-               cfs_hash_bd_t bd;
+               struct lu_site *site = o->lo_dev->ld_site;
+               struct cfs_hash *obj_hash = site->ls_obj_hash;
+               struct cfs_hash_bd bd;
 
                cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
-               list_del_init(&top->loh_lru);
+               if (!list_empty(&top->loh_lru)) {
+                       struct lu_site_bkt_data *bkt;
+
+                       list_del_init(&top->loh_lru);
+                       bkt = cfs_hash_bd_extra_get(obj_hash, &bd);
+                       bkt->lsb_lru_len--;
+                       lprocfs_counter_decr(site->ls_stats, LU_SS_LRU_LEN);
+               }
                cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
                cfs_hash_bd_unlock(obj_hash, &bd, 1);
        }
@@ -340,11 +354,11 @@ int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
         struct lu_object_header *h;
         struct lu_object_header *temp;
         struct lu_site_bkt_data *bkt;
-        cfs_hash_bd_t            bd;
-        cfs_hash_bd_t            bd2;
+       struct cfs_hash_bd            bd;
+       struct cfs_hash_bd            bd2;
        struct list_head         dispose;
        int                      did_sth;
-       unsigned int             start;
+       unsigned int             start = 0;
         int                      count;
         int                      bnr;
        unsigned int             i;
@@ -357,7 +371,8 @@ int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
          * Under LRU list lock, scan LRU list and move unreferenced objects to
          * the dispose list, removing them from LRU and hash table.
          */
-        start = s->ls_purge_start;
+       if (nr != ~0)
+               start = s->ls_purge_start;
        bnr = (nr == ~0) ? -1 : nr / (int)CFS_HASH_NBKT(s->ls_obj_hash) + 1;
  again:
        /*
@@ -382,6 +397,8 @@ int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
                         cfs_hash_bd_del_locked(s->ls_obj_hash,
                                                &bd2, &h->loh_hash);
                        list_move(&h->loh_lru, &dispose);
+                       bkt->lsb_lru_len--;
+                       lprocfs_counter_decr(s->ls_stats, LU_SS_LRU_LEN);
                         if (did_sth == 0)
                                 did_sth = 1;
 
@@ -565,10 +582,9 @@ int lu_object_invariant(const struct lu_object *o)
         }
         return 1;
 }
-EXPORT_SYMBOL(lu_object_invariant);
 
 static struct lu_object *htable_lookup(struct lu_site *s,
-                                      cfs_hash_bd_t *bd,
+                                      struct cfs_hash_bd *bd,
                                       const struct lu_fid *f,
                                       wait_queue_t *waiter,
                                       __u64 *version)
@@ -595,7 +611,11 @@ static struct lu_object *htable_lookup(struct lu_site *s,
         if (likely(!lu_object_is_dying(h))) {
                cfs_hash_get(s->ls_obj_hash, hnode);
                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
-               list_del_init(&h->loh_lru);
+               if (!list_empty(&h->loh_lru)) {
+                       list_del_init(&h->loh_lru);
+                       bkt->lsb_lru_len--;
+                       lprocfs_counter_decr(s->ls_stats, LU_SS_LRU_LEN);
+               }
                 return lu_object_top(h);
         }
 
@@ -606,7 +626,7 @@ static struct lu_object *htable_lookup(struct lu_site *s,
          */
 
        if (likely(waiter != NULL)) {
-               init_waitqueue_entry_current(waiter);
+               init_waitqueue_entry(waiter, current);
                add_wait_queue(&bkt->lsb_marche_funebre, waiter);
                set_current_state(TASK_UNINTERRUPTIBLE);
                lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
@@ -615,31 +635,6 @@ static struct lu_object *htable_lookup(struct lu_site *s,
        return ERR_PTR(-EAGAIN);
 }
 
-static struct lu_object *htable_lookup_nowait(struct lu_site *s,
-                                             cfs_hash_bd_t *bd,
-                                             const struct lu_fid *f)
-{
-       struct hlist_node       *hnode;
-       struct lu_object_header *h;
-
-       /* cfs_hash_bd_peek_locked is a somehow "internal" function
-        * of cfs_hash, it doesn't add refcount on object. */
-       hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
-       if (hnode == NULL) {
-               lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
-               return ERR_PTR(-ENOENT);
-       }
-
-       h = container_of0(hnode, struct lu_object_header, loh_hash);
-       if (unlikely(lu_object_is_dying(h)))
-               return ERR_PTR(-ENOENT);
-
-       cfs_hash_get(s->ls_obj_hash, hnode);
-       lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
-       list_del_init(&h->loh_lru);
-       return lu_object_top(h);
-}
-
 /**
  * Search cache for an object with the fid \a f. If such object is found,
  * return it. Otherwise, create new object, insert it into cache and return
@@ -683,9 +678,8 @@ static struct lu_object *lu_object_new(const struct lu_env *env,
                                        const struct lu_object_conf *conf)
 {
         struct lu_object        *o;
-        cfs_hash_t              *hs;
-        cfs_hash_bd_t            bd;
-        struct lu_site_bkt_data *bkt;
+       struct cfs_hash              *hs;
+       struct cfs_hash_bd            bd;
 
         o = lu_object_alloc(env, dev, f, conf);
         if (unlikely(IS_ERR(o)))
@@ -693,9 +687,7 @@ static struct lu_object *lu_object_new(const struct lu_env *env,
 
         hs = dev->ld_site->ls_obj_hash;
         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
-        bkt = cfs_hash_bd_extra_get(hs, &bd);
         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
-        bkt->lsb_busy++;
         cfs_hash_bd_unlock(hs, &bd, 1);
 
        lu_object_limit(env, dev);
@@ -715,8 +707,8 @@ static struct lu_object *lu_object_find_try(const struct lu_env *env,
        struct lu_object      *o;
        struct lu_object      *shadow;
        struct lu_site        *s;
-       cfs_hash_t            *hs;
-       cfs_hash_bd_t          bd;
+       struct cfs_hash            *hs;
+       struct cfs_hash_bd          bd;
        __u64                  version = 0;
 
         /*
@@ -764,11 +756,7 @@ static struct lu_object *lu_object_find_try(const struct lu_env *env,
 
         shadow = htable_lookup(s, &bd, f, waiter, &version);
        if (likely(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT)) {
-                struct lu_site_bkt_data *bkt;
-
-                bkt = cfs_hash_bd_extra_get(hs, &bd);
                 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
-                bkt->lsb_busy++;
                 cfs_hash_bd_unlock(hs, &bd, 1);
 
                lu_object_limit(env, dev);
@@ -807,7 +795,7 @@ struct lu_object *lu_object_find_at(const struct lu_env *env,
                 * lu_object_find_try() already added waiter into the
                 * wait queue.
                 */
-               waitq_wait(&wait, TASK_UNINTERRUPTIBLE);
+               schedule();
                bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
                remove_wait_queue(&bkt->lsb_marche_funebre, &wait);
        }
@@ -815,30 +803,6 @@ struct lu_object *lu_object_find_at(const struct lu_env *env,
 EXPORT_SYMBOL(lu_object_find_at);
 
 /**
- * Try to find the object in cache without waiting for the dead object
- * to be released nor allocating object if no cached one was found.
- *
- * The found object will be set as LU_OBJECT_HEARD_BANSHEE for purging.
- */
-void lu_object_purge(const struct lu_env *env, struct lu_device *dev,
-                    const struct lu_fid *f)
-{
-       struct lu_site          *s  = dev->ld_site;
-       cfs_hash_t              *hs = s->ls_obj_hash;
-       cfs_hash_bd_t            bd;
-       struct lu_object        *o;
-
-       cfs_hash_bd_get_and_lock(hs, f, &bd, 1);
-       o = htable_lookup_nowait(s, &bd, f);
-       cfs_hash_bd_unlock(hs, &bd, 1);
-       if (!IS_ERR(o)) {
-               set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
-               lu_object_put(env, o);
-       }
-}
-EXPORT_SYMBOL(lu_object_purge);
-
-/**
  * Find object with given fid, and return its slice belonging to given device.
  */
 struct lu_object *lu_object_find_slice(const struct lu_env *env,
@@ -915,7 +879,7 @@ struct lu_site_print_arg {
 };
 
 static int
-lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+lu_site_obj_print(struct cfs_hash *hs, struct cfs_hash_bd *bd,
                  struct hlist_node *hnode, void *data)
 {
        struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
@@ -958,6 +922,7 @@ static unsigned long lu_htable_order(struct lu_device *top)
 {
        unsigned long cache_size;
        unsigned long bits;
+       unsigned long bits_max = LU_SITE_BITS_MAX;
 
        /*
         * For ZFS based OSDs the cache should be disabled by default.  This
@@ -971,6 +936,9 @@ static unsigned long lu_htable_order(struct lu_device *top)
                return LU_SITE_BITS_MIN;
        }
 
+       if (strcmp(top->ld_type->ldt_name, LUSTRE_VVP_NAME) == 0)
+               bits_max = LU_SITE_BITS_MAX_CL;
+
         /*
          * Calculate hash table size, assuming that we want reasonable
          * performance when 20% of total memory is occupied by cache of
@@ -1001,10 +969,11 @@ static unsigned long lu_htable_order(struct lu_device *top)
         for (bits = 1; (1 << bits) < cache_size; ++bits) {
                 ;
         }
-        return bits;
+
+       return clamp_t(typeof(bits), bits, LU_SITE_BITS_MIN, bits_max);
 }
 
-static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
+static unsigned lu_obj_hop_hash(struct cfs_hash *hs,
                                const void *key, unsigned mask)
 {
        struct lu_fid  *fid = (struct lu_fid *)key;
@@ -1044,27 +1013,20 @@ static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
        return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
 }
 
-static void lu_obj_hop_get(cfs_hash_t *hs, struct hlist_node *hnode)
+static void lu_obj_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
 {
        struct lu_object_header *h;
 
        h = hlist_entry(hnode, struct lu_object_header, loh_hash);
-       if (atomic_add_return(1, &h->loh_ref) == 1) {
-               struct lu_site_bkt_data *bkt;
-               cfs_hash_bd_t            bd;
-
-               cfs_hash_bd_get(hs, &h->loh_fid, &bd);
-               bkt = cfs_hash_bd_extra_get(hs, &bd);
-               bkt->lsb_busy++;
-       }
+       atomic_inc(&h->loh_ref);
 }
 
-static void lu_obj_hop_put_locked(cfs_hash_t *hs, struct hlist_node *hnode)
+static void lu_obj_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 {
         LBUG(); /* we should never called it */
 }
 
-static cfs_hash_ops_t lu_site_hash_ops = {
+static struct cfs_hash_ops lu_site_hash_ops = {
         .hs_hash        = lu_obj_hop_hash,
         .hs_key         = lu_obj_hop_key,
         .hs_keycmp      = lu_obj_hop_keycmp,
@@ -1096,7 +1058,7 @@ EXPORT_SYMBOL(lu_dev_del_linkage);
 int lu_site_init(struct lu_site *s, struct lu_device *top)
 {
        struct lu_site_bkt_data *bkt;
-       cfs_hash_bd_t bd;
+       struct cfs_hash_bd bd;
        char name[16];
        unsigned long bits;
        unsigned int i;
@@ -1104,10 +1066,8 @@ int lu_site_init(struct lu_site *s, struct lu_device *top)
 
        memset(s, 0, sizeof *s);
        mutex_init(&s->ls_purge_mutex);
-       bits = lu_htable_order(top);
        snprintf(name, sizeof(name), "lu_site_%s", top->ld_type->ldt_name);
-       for (bits = clamp_t(typeof(bits), bits,
-                           LU_SITE_BITS_MIN, LU_SITE_BITS_MAX);
+       for (bits = lu_htable_order(top);
             bits >= LU_SITE_BITS_MIN; bits--) {
                s->ls_obj_hash = cfs_hash_create(name, bits, bits,
                                                 bits - LU_SITE_BKT_BITS,
@@ -1152,6 +1112,12 @@ int lu_site_init(struct lu_site *s, struct lu_device *top)
                              0, "cache_death_race", "cache_death_race");
         lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
                              0, "lru_purged", "lru_purged");
+       /*
+        * Unlike other counters, lru_len can be decremented so
+        * need lc_sum instead of just lc_count
+        */
+       lprocfs_counter_init(s->ls_stats, LU_SS_LRU_LEN,
+                            LPROCFS_CNTR_AVGMINMAX, "lru_len", "lru_len");
 
        INIT_LIST_HEAD(&s->ls_linkage);
         s->ls_top_dev = top;
@@ -1408,7 +1374,6 @@ void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
                 }
         }
 }
-EXPORT_SYMBOL(lu_stack_fini);
 
 enum {
         /**
@@ -1419,7 +1384,8 @@ enum {
 
 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
 
-static DEFINE_SPINLOCK(lu_keys_guard);
+DEFINE_RWLOCK(lu_keys_guard);
+static atomic_t lu_key_initing_cnt = ATOMIC_INIT(0);
 
 /**
  * Global counter incremented whenever key is registered, unregistered,
@@ -1443,7 +1409,7 @@ int lu_context_key_register(struct lu_context_key *key)
         LASSERT(key->lct_owner != NULL);
 
         result = -ENFILE;
-       spin_lock(&lu_keys_guard);
+       write_lock(&lu_keys_guard);
         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
                 if (lu_keys[i] == NULL) {
                         key->lct_index = i;
@@ -1455,7 +1421,7 @@ int lu_context_key_register(struct lu_context_key *key)
                         break;
                 }
         }
-       spin_unlock(&lu_keys_guard);
+       write_unlock(&lu_keys_guard);
        return result;
 }
 EXPORT_SYMBOL(lu_context_key_register);
@@ -1494,13 +1460,26 @@ void lu_context_key_degister(struct lu_context_key *key)
        lu_context_key_quiesce(key);
 
        ++key_set_version;
-       spin_lock(&lu_keys_guard);
+       write_lock(&lu_keys_guard);
        key_fini(&lu_shrink_env.le_ctx, key->lct_index);
+
+       /**
+        * Wait until all transient contexts referencing this key have
+        * run lu_context_key::lct_fini() method.
+        */
+       while (atomic_read(&key->lct_used) > 1) {
+               write_unlock(&lu_keys_guard);
+               CDEBUG(D_INFO, "lu_context_key_degister: \"%s\" %p, %d\n",
+                      key->lct_owner ? key->lct_owner->name : "", key,
+                      atomic_read(&key->lct_used));
+               schedule();
+               write_lock(&lu_keys_guard);
+       }
        if (lu_keys[key->lct_index]) {
                lu_keys[key->lct_index] = NULL;
                lu_ref_fini(&key->lct_reference);
        }
-       spin_unlock(&lu_keys_guard);
+       write_unlock(&lu_keys_guard);
 
        LASSERTF(atomic_read(&key->lct_used) == 1,
                 "key has instances: %d\n",
@@ -1622,26 +1601,40 @@ void lu_context_key_quiesce(struct lu_context_key *key)
                  * XXX layering violation.
                  */
                 cl_env_cache_purge(~0);
-                key->lct_tags |= LCT_QUIESCENT;
                 /*
                  * XXX memory barrier has to go here.
                  */
-               spin_lock(&lu_keys_guard);
+               write_lock(&lu_keys_guard);
+               key->lct_tags |= LCT_QUIESCENT;
+
+               /**
+                * Wait until all lu_context_key::lct_init() methods
+                * have completed.
+                */
+               while (atomic_read(&lu_key_initing_cnt) > 0) {
+                       write_unlock(&lu_keys_guard);
+                       CDEBUG(D_INFO, "lu_context_key_quiesce: \"%s\""
+                              " %p, %d (%d)\n",
+                              key->lct_owner ? key->lct_owner->name : "",
+                              key, atomic_read(&key->lct_used),
+                              atomic_read(&lu_key_initing_cnt));
+                       schedule();
+                       write_lock(&lu_keys_guard);
+               }
+
                list_for_each_entry(ctx, &lu_context_remembered,
                                    lc_remember)
                        key_fini(ctx, key->lct_index);
-               spin_unlock(&lu_keys_guard);
+               write_unlock(&lu_keys_guard);
                ++key_set_version;
        }
 }
-EXPORT_SYMBOL(lu_context_key_quiesce);
 
 void lu_context_key_revive(struct lu_context_key *key)
 {
         key->lct_tags &= ~LCT_QUIESCENT;
         ++key_set_version;
 }
-EXPORT_SYMBOL(lu_context_key_revive);
 
 static void keys_fini(struct lu_context *ctx)
 {
@@ -1661,6 +1654,19 @@ static int keys_fill(struct lu_context *ctx)
 {
        unsigned int i;
 
+       /*
+        * A serialisation with lu_context_key_quiesce() is needed, but some
+        * "key->lct_init()" are calling kernel memory allocation routine and
+        * can't be called while holding a spin_lock.
+        * "lu_keys_guard" is held while incrementing "lu_key_initing_cnt"
+        * to ensure the start of the serialisation.
+        * An atomic_t variable is still used, in order not to reacquire the
+        * lock when decrementing the counter.
+        */
+       read_lock(&lu_keys_guard);
+       atomic_inc(&lu_key_initing_cnt);
+       read_unlock(&lu_keys_guard);
+
         LINVRNT(ctx->lc_value != NULL);
         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
                 struct lu_context_key *key;
@@ -1678,13 +1684,19 @@ static int keys_fill(struct lu_context *ctx)
                         LINVRNT(key->lct_init != NULL);
                         LINVRNT(key->lct_index == i);
 
-                        value = key->lct_init(ctx, key);
-                        if (unlikely(IS_ERR(value)))
-                                return PTR_ERR(value);
-
                        LASSERT(key->lct_owner != NULL);
-                       if (!(ctx->lc_tags & LCT_NOREF))
-                               try_module_get(key->lct_owner);
+                       if (!(ctx->lc_tags & LCT_NOREF) &&
+                           try_module_get(key->lct_owner) == 0) {
+                               /* module is unloading, skip this key */
+                               continue;
+                       }
+
+                       value = key->lct_init(ctx, key);
+                       if (unlikely(IS_ERR(value))) {
+                               atomic_dec(&lu_key_initing_cnt);
+                               return PTR_ERR(value);
+                       }
+
                        lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
                        atomic_inc(&key->lct_used);
                         /*
@@ -1698,6 +1710,7 @@ static int keys_fill(struct lu_context *ctx)
                 }
                 ctx->lc_version = key_set_version;
         }
+       atomic_dec(&lu_key_initing_cnt);
         return 0;
 }
 
@@ -1721,9 +1734,9 @@ int lu_context_init(struct lu_context *ctx, __u32 tags)
        ctx->lc_state = LCS_INITIALIZED;
        ctx->lc_tags = tags;
        if (tags & LCT_REMEMBER) {
-               spin_lock(&lu_keys_guard);
+               write_lock(&lu_keys_guard);
                list_add(&ctx->lc_remember, &lu_context_remembered);
-               spin_unlock(&lu_keys_guard);
+               write_unlock(&lu_keys_guard);
        } else {
                INIT_LIST_HEAD(&ctx->lc_remember);
        }
@@ -1749,10 +1762,10 @@ void lu_context_fini(struct lu_context *ctx)
                keys_fini(ctx);
 
        } else { /* could race with key degister */
-               spin_lock(&lu_keys_guard);
+               write_lock(&lu_keys_guard);
                keys_fini(ctx);
                list_del_init(&ctx->lc_remember);
-               spin_unlock(&lu_keys_guard);
+               write_unlock(&lu_keys_guard);
        }
 }
 EXPORT_SYMBOL(lu_context_fini);
@@ -1778,15 +1791,20 @@ void lu_context_exit(struct lu_context *ctx)
         ctx->lc_state = LCS_LEFT;
         if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
-                        if (ctx->lc_value[i] != NULL) {
-                                struct lu_context_key *key;
-
-                                key = lu_keys[i];
-                                LASSERT(key != NULL);
-                                if (key->lct_exit != NULL)
-                                        key->lct_exit(ctx,
-                                                      key, ctx->lc_value[i]);
-                        }
+                       /* could race with key quiescency */
+                       if (ctx->lc_tags & LCT_REMEMBER)
+                               read_lock(&lu_keys_guard);
+                       if (ctx->lc_value[i] != NULL) {
+                               struct lu_context_key *key;
+
+                               key = lu_keys[i];
+                               LASSERT(key != NULL);
+                               if (key->lct_exit != NULL)
+                                       key->lct_exit(ctx,
+                                                     key, ctx->lc_value[i]);
+                       }
+                       if (ctx->lc_tags & LCT_REMEMBER)
+                               read_unlock(&lu_keys_guard);
                 }
         }
 }
@@ -1801,7 +1819,6 @@ int lu_context_refill(struct lu_context *ctx)
 {
         return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
 }
-EXPORT_SYMBOL(lu_context_refill);
 
 /**
  * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
@@ -1815,37 +1832,37 @@ __u32 lu_session_tags_default = 0;
 
 void lu_context_tags_update(__u32 tags)
 {
-       spin_lock(&lu_keys_guard);
+       write_lock(&lu_keys_guard);
        lu_context_tags_default |= tags;
        key_set_version++;
-       spin_unlock(&lu_keys_guard);
+       write_unlock(&lu_keys_guard);
 }
 EXPORT_SYMBOL(lu_context_tags_update);
 
 void lu_context_tags_clear(__u32 tags)
 {
-       spin_lock(&lu_keys_guard);
+       write_lock(&lu_keys_guard);
        lu_context_tags_default &= ~tags;
        key_set_version++;
-       spin_unlock(&lu_keys_guard);
+       write_unlock(&lu_keys_guard);
 }
 EXPORT_SYMBOL(lu_context_tags_clear);
 
 void lu_session_tags_update(__u32 tags)
 {
-       spin_lock(&lu_keys_guard);
+       write_lock(&lu_keys_guard);
        lu_session_tags_default |= tags;
        key_set_version++;
-       spin_unlock(&lu_keys_guard);
+       write_unlock(&lu_keys_guard);
 }
 EXPORT_SYMBOL(lu_session_tags_update);
 
 void lu_session_tags_clear(__u32 tags)
 {
-       spin_lock(&lu_keys_guard);
+       write_lock(&lu_keys_guard);
        lu_session_tags_default &= ~tags;
        key_set_version++;
-       spin_unlock(&lu_keys_guard);
+       write_unlock(&lu_keys_guard);
 }
 EXPORT_SYMBOL(lu_session_tags_clear);
 
@@ -1917,10 +1934,10 @@ typedef struct lu_site_stats{
         unsigned        lss_busy;
 } lu_site_stats_t;
 
-static void lu_site_stats_get(cfs_hash_t *hs,
+static void lu_site_stats_get(struct cfs_hash *hs,
                               lu_site_stats_t *stats, int populated)
 {
-       cfs_hash_bd_t bd;
+       struct cfs_hash_bd bd;
        unsigned int  i;
 
         cfs_hash_for_each_bucket(hs, &bd, i) {
@@ -1928,7 +1945,8 @@ static void lu_site_stats_get(cfs_hash_t *hs,
                struct hlist_head       *hhead;
 
                 cfs_hash_bd_lock(hs, &bd, 1);
-                stats->lss_busy  += bkt->lsb_busy;
+               stats->lss_busy  +=
+                       cfs_hash_bd_count_get(&bd) - bkt->lsb_lru_len;
                 stats->lss_total += cfs_hash_bd_count_get(&bd);
                 stats->lss_max_search = max((int)stats->lss_max_search,
                                             cfs_hash_bd_depmax_get(&bd));
@@ -1946,10 +1964,22 @@ static void lu_site_stats_get(cfs_hash_t *hs,
 }
 
 
+/*
+ * lu_cache_shrink_count returns the number of cached objects that are
+ * candidates to be freed by shrink_slab(). A counter, which tracks
+ * the number of items in the site's lru, is maintained in the per cpu
+ * stats of each site. The counter is incremented when an object is added
+ * to a site's lru and decremented when one is removed. The number of
+ * free-able objects is the sum of all per cpu counters for all sites.
+ *
+ * Using a per cpu counter is a compromise solution to concurrent access:
+ * lu_object_put() can update the counter without locking the site and
+ * lu_cache_shrink_count can sum the counters without locking each
+ * ls_obj_hash bucket.
+ */
 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
                                           struct shrink_control *sc)
 {
-       lu_site_stats_t stats;
        struct lu_site *s;
        struct lu_site *tmp;
        unsigned long cached = 0;
@@ -1959,14 +1989,14 @@ static unsigned long lu_cache_shrink_count(struct shrinker *sk,
 
        mutex_lock(&lu_sites_guard);
        list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
-               memset(&stats, 0, sizeof(stats));
-               lu_site_stats_get(s->ls_obj_hash, &stats, 0);
-               cached += stats.lss_total - stats.lss_busy;
+               cached += ls_stats_read(s->ls_stats, LU_SS_LRU_LEN);
        }
        mutex_unlock(&lu_sites_guard);
 
        cached = (cached / 100) * sysctl_vfs_cache_pressure;
-       CDEBUG(D_INODE, "%ld objects cached\n", cached);
+       CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n",
+              cached, sysctl_vfs_cache_pressure);
+
        return cached;
 }
 
@@ -2023,7 +2053,7 @@ static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
  * is safe to take the lu_sites_guard lock.
  *
  * Ideally we should accurately return the remaining number of cached
- * objects without taking the  lu_sites_guard lock, but this is not
+ * objects without taking the lu_sites_guard lock, but this is not
  * possible in the current implementation.
  */
 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
@@ -2040,11 +2070,10 @@ static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
 
        CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
 
-       lu_cache_shrink_scan(shrinker, &scv);
+       if (scv.nr_to_scan != 0)
+               lu_cache_shrink_scan(shrinker, &scv);
 
        cached = lu_cache_shrink_count(shrinker, &scv);
-       if (scv.nr_to_scan == 0)
-               CDEBUG(D_INODE, "%d objects cached\n", cached);
        return cached;
 }
 
@@ -2098,7 +2127,6 @@ void lu_context_keys_dump(void)
                 }
         }
 }
-EXPORT_SYMBOL(lu_context_keys_dump);
 
 /**
  * Initialization of global lu_* data.
@@ -2173,12 +2201,19 @@ void lu_global_fini(void)
 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
 {
 #ifdef CONFIG_PROC_FS
-        struct lprocfs_counter ret;
+       struct lprocfs_counter ret;
 
-        lprocfs_stats_collect(stats, idx, &ret);
-        return (__u32)ret.lc_count;
+       lprocfs_stats_collect(stats, idx, &ret);
+       if (idx == LU_SS_LRU_LEN)
+               /*
+                * protect against counter on cpu A being decremented
+                * before counter is incremented on cpu B; unlikely
+                */
+               return (__u32)((ret.lc_sum > 0) ? ret.lc_sum : 0);
+       else
+               return (__u32)ret.lc_count;
 #else
-        return 0;
+       return 0;
 #endif
 }
 
@@ -2193,7 +2228,7 @@ int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
        memset(&stats, 0, sizeof(stats));
        lu_site_stats_get(s->ls_obj_hash, &stats, 1);
 
-       return seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
+       return seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d %d\n",
                          stats.lss_busy,
                          stats.lss_total,
                          stats.lss_populated,
@@ -2204,32 +2239,11 @@ int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
                          ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
                          ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
                          ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
-                         ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
+                         ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED),
+                         ls_stats_read(s->ls_stats, LU_SS_LRU_LEN));
 }
 EXPORT_SYMBOL(lu_site_stats_seq_print);
 
-int lu_site_stats_print(const struct lu_site *s, char *page, int count)
-{
-        lu_site_stats_t stats;
-
-        memset(&stats, 0, sizeof(stats));
-        lu_site_stats_get(s->ls_obj_hash, &stats, 1);
-
-        return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
-                        stats.lss_busy,
-                        stats.lss_total,
-                        stats.lss_populated,
-                        CFS_HASH_NHLIST(s->ls_obj_hash),
-                        stats.lss_max_search,
-                        ls_stats_read(s->ls_stats, LU_SS_CREATED),
-                        ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
-                        ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
-                        ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
-                        ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
-                        ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
-}
-EXPORT_SYMBOL(lu_site_stats_print);
-
 /**
  * Helper function to initialize a number of kmem slab caches at once.
  */
@@ -2277,11 +2291,10 @@ void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
 {
        struct lu_site          *s = o->lo_dev->ld_site;
        struct lu_fid           *old = &o->lo_header->loh_fid;
-       struct lu_site_bkt_data *bkt;
        struct lu_object        *shadow;
        wait_queue_t             waiter;
-       cfs_hash_t              *hs;
-       cfs_hash_bd_t            bd;
+       struct cfs_hash         *hs;
+       struct cfs_hash_bd       bd;
        __u64                    version = 0;
 
        LASSERT(fid_is_zero(old));
@@ -2292,9 +2305,7 @@ void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
        /* supposed to be unique */
        LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
        *old = *fid;
-       bkt = cfs_hash_bd_extra_get(hs, &bd);
        cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
-       bkt->lsb_busy++;
        cfs_hash_bd_unlock(hs, &bd, 1);
 }
 EXPORT_SYMBOL(lu_object_assign_fid);
@@ -2393,4 +2404,3 @@ int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
        buf->lb_len = len;
        return 0;
 }
-EXPORT_SYMBOL(lu_buf_check_and_grow);