X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fobdclass%2Fcl_object.c;h=ac46d074b24e3ce0c1d1f7ebb453a17b7972022f;hb=cc2ff1bfd66a5c004eb6ed61fc2dac3f1ab49d3a;hp=6073df50063f8c72da886f1a24f68f2e885c93a4;hpb=9c6ebfd11be1cc11629a2960c8c818492e1f0980;p=fs%2Flustre-release.git

diff --git a/lustre/obdclass/cl_object.c b/lustre/obdclass/cl_object.c
index 6073df5..ac46d07 100644
--- a/lustre/obdclass/cl_object.c
+++ b/lustre/obdclass/cl_object.c
@@ -36,6 +36,7 @@
  * Client Lustre Object.
  *
  *   Author: Nikita Danilov <nikita.danilov@sun.com>
+ *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  */
 
 /*
@@ -43,7 +44,6 @@
  *
  *  i_mutex
  *      PG_locked
- *          ->coh_page_guard
  *          ->coh_lock_guard
  *          ->coh_attr_guard
  *          ->ls_guard
@@ -63,8 +63,6 @@
 
 static struct kmem_cache *cl_env_kmem;
 
-/** Lock class of cl_object_header::coh_page_guard */
-static struct lock_class_key cl_page_guard_class;
 /** Lock class of cl_object_header::coh_lock_guard */
 static struct lock_class_key cl_lock_guard_class;
 /** Lock class of cl_object_header::coh_attr_guard */
@@ -77,24 +75,19 @@ extern __u32 lu_session_tags_default;
  */
 int cl_object_header_init(struct cl_object_header *h)
 {
-        int result;
+	int result;
 
-        ENTRY;
-        result = lu_object_header_init(&h->coh_lu);
-        if (result == 0) {
-		spin_lock_init(&h->coh_page_guard);
+	ENTRY;
+	result = lu_object_header_init(&h->coh_lu);
+	if (result == 0) {
 		spin_lock_init(&h->coh_lock_guard);
 		spin_lock_init(&h->coh_attr_guard);
-		lockdep_set_class(&h->coh_page_guard, &cl_page_guard_class);
 		lockdep_set_class(&h->coh_lock_guard, &cl_lock_guard_class);
 		lockdep_set_class(&h->coh_attr_guard, &cl_attr_guard_class);
-                h->coh_pages = 0;
-                /* XXX hard coded GFP_* mask. */
-                INIT_RADIX_TREE(&h->coh_tree, GFP_ATOMIC);
-                CFS_INIT_LIST_HEAD(&h->coh_locks);
-		h->coh_page_bufsize = ALIGN(sizeof(struct cl_page), 8);
-        }
-        RETURN(result);
+		CFS_INIT_LIST_HEAD(&h->coh_locks);
+		h->coh_page_bufsize = 0;
+	}
+	RETURN(result);
 }
 EXPORT_SYMBOL(cl_object_header_init);
 
@@ -120,7 +113,7 @@ struct cl_object *cl_object_find(const struct lu_env *env,
                                  struct cl_device *cd, const struct lu_fid *fid,
                                  const struct cl_object_conf *c)
 {
-        cfs_might_sleep();
+	might_sleep();
         return lu2cl(lu_object_find_slice(env, cl2lu_dev(cd), fid, &c->coc_lu));
 }
 EXPORT_SYMBOL(cl_object_find);
@@ -156,7 +149,7 @@ EXPORT_SYMBOL(cl_object_get);
 /**
  * Returns the top-object for a given \a o.
  *
- * \see cl_page_top(), cl_io_top()
+ * \see cl_io_top()
  */
 struct cl_object *cl_object_top(struct cl_object *o)
 {
@@ -221,7 +214,7 @@ int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
 	struct lu_object_header *top;
 	int result;
 
-	LASSERT(spin_is_locked(cl_object_attr_guard(obj)));
+	assert_spin_locked(cl_object_attr_guard(obj));
 	ENTRY;
 
         top = obj->co_lu.lo_header;
@@ -253,7 +246,7 @@ int cl_object_attr_set(const struct lu_env *env, struct cl_object *obj,
 	struct lu_object_header *top;
 	int result;
 
-	LASSERT(spin_is_locked(cl_object_attr_guard(obj)));
+	assert_spin_locked(cl_object_attr_guard(obj));
 	ENTRY;
 
         top = obj->co_lu.lo_header;
@@ -331,6 +324,33 @@ int cl_conf_set(const struct lu_env *env, struct cl_object *obj,
 EXPORT_SYMBOL(cl_conf_set);
 
 /**
+ * Prunes caches of pages and locks for this object.
+ */
+void cl_object_prune(const struct lu_env *env, struct cl_object *obj)
+{
+	struct lu_object_header *top;
+	struct cl_object *o;
+	int result;
+	ENTRY;
+
+	top = obj->co_lu.lo_header;
+	result = 0;
+	cfs_list_for_each_entry(o, &top->loh_layers, co_lu.lo_linkage) {
+		if (o->co_ops->coo_prune != NULL) {
+			result = o->co_ops->coo_prune(env, o);
+			if (result != 0)
+				break;
+		}
+	}
+
+	/* TODO: pruning locks will be moved into layers after cl_lock
+	 * simplification is done */
+	cl_locks_prune(env, obj, 1);
+	EXIT;
+}
+EXPORT_SYMBOL(cl_object_prune);
+
+/**
  * Helper function removing all object locks, and marking object for
  * deletion. All object pages must have been deleted at this point.
  *
@@ -342,8 +362,6 @@ void cl_object_kill(const struct lu_env *env, struct cl_object *obj)
         struct cl_object_header *hdr;
 
         hdr = cl_object_header(obj);
-        LASSERT(hdr->coh_tree.rnode == NULL);
-        LASSERT(hdr->coh_pages == 0);
 
 	set_bit(LU_OBJECT_HEARD_BANSHEE, &hdr->coh_lu.loh_flags);
         /*
@@ -358,18 +376,6 @@ void cl_object_kill(const struct lu_env *env, struct cl_object *obj)
 EXPORT_SYMBOL(cl_object_kill);
 
 /**
- * Prunes caches of pages and locks for this object.
- */
-void cl_object_prune(const struct lu_env *env, struct cl_object *obj)
-{
-        ENTRY;
-        cl_pages_prune(env, obj);
-        cl_locks_prune(env, obj, 1);
-        EXIT;
-}
-EXPORT_SYMBOL(cl_object_prune);
-
-/**
  * Check if the object has locks.
  */
 int cl_object_has_locks(struct cl_object *obj)
@@ -391,14 +397,13 @@ void cache_stats_init(struct cache_stats *cs, const char *name)
 
         cs->cs_name = name;
 	for (i = 0; i < CS_NR; i++)
-		cfs_atomic_set(&cs->cs_stats[i], 0);
+		atomic_set(&cs->cs_stats[i], 0);
 }
 
-int cache_stats_print(const struct cache_stats *cs,
-                      char *page, int count, int h)
+int cache_stats_print(const struct cache_stats *cs, struct seq_file *m, int h)
 {
-	int nob = 0;
 	int i;
+
 	/*
 	 *   lookup    hit    total  cached create
 	 * env: ...... ...... ...... ...... ......
@@ -406,20 +411,20 @@ int cache_stats_print(const struct cache_stats *cs,
 	if (h) {
 		const char *names[CS_NR] = CS_NAMES;
 
-		nob += snprintf(page + nob, count - nob, "%6s", " ");
+		seq_printf(m, "%6s", " ");
 		for (i = 0; i < CS_NR; i++)
-			nob += snprintf(page + nob, count - nob,
-					"%8s", names[i]);
-		nob += snprintf(page + nob, count - nob, "\n");
+			seq_printf(m, "%8s", names[i]);
+		seq_printf(m, "\n");
 	}
 
-	nob += snprintf(page + nob, count - nob, "%5.5s:", cs->cs_name);
+	seq_printf(m, "%5.5s:", cs->cs_name);
 	for (i = 0; i < CS_NR; i++)
-		nob += snprintf(page + nob, count - nob, "%8u",
-				cfs_atomic_read(&cs->cs_stats[i]));
-	return nob;
+		seq_printf(m, "%8u", atomic_read(&cs->cs_stats[i]));
+	return 0;
 }
 
+static void cl_env_percpu_refill(void);
+
 /**
  * Initialize client site.
  *
@@ -436,11 +441,12 @@ int cl_site_init(struct cl_site *s, struct cl_device *d)
                 cache_stats_init(&s->cs_pages, "pages");
                 cache_stats_init(&s->cs_locks, "locks");
                 for (i = 0; i < ARRAY_SIZE(s->cs_pages_state); ++i)
-                        cfs_atomic_set(&s->cs_pages_state[0], 0);
+			atomic_set(&s->cs_pages_state[0], 0);
                 for (i = 0; i < ARRAY_SIZE(s->cs_locks_state); ++i)
-                        cfs_atomic_set(&s->cs_locks_state[i], 0);
-        }
-        return result;
+			atomic_set(&s->cs_locks_state[i], 0);
+		cl_env_percpu_refill();
+	}
+	return result;
 }
 EXPORT_SYMBOL(cl_site_init);
 
@@ -455,57 +461,55 @@ EXPORT_SYMBOL(cl_site_fini);
 
 static struct cache_stats cl_env_stats = {
         .cs_name    = "envs",
-	.cs_stats = { CFS_ATOMIC_INIT(0), }
+	.cs_stats = { ATOMIC_INIT(0), }
 };
 
 /**
  * Outputs client site statistical counters into a buffer. Suitable for
  * ll_rd_*()-style functions.
  */
-int cl_site_stats_print(const struct cl_site *site, char *page, int count)
+int cl_site_stats_print(const struct cl_site *site, struct seq_file *m)
 {
-        int nob;
-        int i;
-        static const char *pstate[] = {
-                [CPS_CACHED]  = "c",
-                [CPS_OWNED]   = "o",
-                [CPS_PAGEOUT] = "w",
-                [CPS_PAGEIN]  = "r",
-                [CPS_FREEING] = "f"
-        };
-        static const char *lstate[] = {
-                [CLS_NEW]       = "n",
-                [CLS_QUEUING]   = "q",
-                [CLS_ENQUEUED]  = "e",
-                [CLS_HELD]      = "h",
-                [CLS_INTRANSIT] = "t",
-                [CLS_CACHED]    = "c",
-                [CLS_FREEING]   = "f"
-        };
+	static const char *pstate[] = {
+		[CPS_CACHED]	= "c",
+		[CPS_OWNED]	= "o",
+		[CPS_PAGEOUT]	= "w",
+		[CPS_PAGEIN]	= "r",
+		[CPS_FREEING]	= "f"
+	};
+	static const char *lstate[] = {
+		[CLS_NEW]	= "n",
+		[CLS_QUEUING]	= "q",
+		[CLS_ENQUEUED]	= "e",
+		[CLS_HELD]	= "h",
+		[CLS_INTRANSIT]	= "t",
+		[CLS_CACHED]	= "c",
+		[CLS_FREEING]	= "f"
+	};
+	int i;
+
 /*
        lookup    hit  total   busy create
 pages: ...... ...... ...... ...... ...... [...... ...... ...... ......]
 locks: ...... ...... ...... ...... ...... [...... ...... ...... ...... ......]
   env: ...... ...... ...... ...... ......
  */
-        nob = lu_site_stats_print(&site->cs_lu, page, count);
-        nob += cache_stats_print(&site->cs_pages, page + nob, count - nob, 1);
-        nob += snprintf(page + nob, count - nob, " [");
-        for (i = 0; i < ARRAY_SIZE(site->cs_pages_state); ++i)
-                nob += snprintf(page + nob, count - nob, "%s: %u ",
-                                pstate[i],
-                                cfs_atomic_read(&site->cs_pages_state[i]));
-        nob += snprintf(page + nob, count - nob, "]\n");
-        nob += cache_stats_print(&site->cs_locks, page + nob, count - nob, 0);
-        nob += snprintf(page + nob, count - nob, " [");
-        for (i = 0; i < ARRAY_SIZE(site->cs_locks_state); ++i)
-                nob += snprintf(page + nob, count - nob, "%s: %u ",
-                                lstate[i],
-                                cfs_atomic_read(&site->cs_locks_state[i]));
-        nob += snprintf(page + nob, count - nob, "]\n");
-        nob += cache_stats_print(&cl_env_stats, page + nob, count - nob, 0);
-        nob += snprintf(page + nob, count - nob, "\n");
-        return nob;
+	lu_site_stats_seq_print(&site->cs_lu, m);
+	cache_stats_print(&site->cs_pages, m, 1);
+	seq_printf(m, " [");
+	for (i = 0; i < ARRAY_SIZE(site->cs_pages_state); ++i)
+		seq_printf(m, "%s: %u ", pstate[i],
+			   atomic_read(&site->cs_pages_state[i]));
+	seq_printf(m, "]\n");
+	cache_stats_print(&site->cs_locks, m, 0);
+	seq_printf(m, " [");
+	for (i = 0; i < ARRAY_SIZE(site->cs_locks_state); ++i)
+		seq_printf(m, "%s: %u ", lstate[i],
+			   atomic_read(&site->cs_locks_state[i]));
+	seq_printf(m, "]\n");
+	cache_stats_print(&cl_env_stats, m, 0);
+	seq_printf(m, "\n");
+	return 0;
 }
 EXPORT_SYMBOL(cl_site_stats_print);
 
@@ -533,6 +537,12 @@ EXPORT_SYMBOL(cl_site_stats_print);
  * bz20044, bz22683.
  */
 
+static CFS_LIST_HEAD(cl_envs);
+static unsigned cl_envs_cached_nr  = 0;
+static unsigned cl_envs_cached_max = 128; /* XXX: prototype: arbitrary limit
+                                           * for now. */
+static DEFINE_SPINLOCK(cl_envs_guard);
+
 struct cl_env {
         void             *ce_magic;
         struct lu_env     ce_lu;
@@ -547,18 +557,18 @@ struct cl_env {
          */
         cfs_hlist_node_t  ce_node;
 #endif
-        /**
-         * Owner for the current cl_env.
-         *
-         * If LL_TASK_CL_ENV is defined, this point to the owning cfs_current(),
-         * only for debugging purpose ;
-         * Otherwise hash is used, and this is the key for cfs_hash.
-         * Now current thread pid is stored. Note using thread pointer would
-         * lead to unbalanced hash because of its specific allocation locality
-         * and could be varied for different platforms and OSes, even different
-         * OS versions.
-         */
-        void             *ce_owner;
+	/**
+	 * Owner for the current cl_env.
+	 *
+	 * If LL_TASK_CL_ENV is defined, this point to the owning current,
+	 * only for debugging purpose ;
+	 * Otherwise hash is used, and this is the key for cfs_hash.
+	 * Now current thread pid is stored. Note using thread pointer would
+	 * lead to unbalanced hash because of its specific allocation locality
+	 * and could be varied for different platforms and OSes, even different
+	 * OS versions.
+	 */
+	void             *ce_owner;
 
         /*
          * Linkage into global list of all client environments. Used for
@@ -577,11 +587,11 @@ struct cl_env {
 };
 
 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
-#define CL_ENV_INC(counter) cfs_atomic_inc(&cl_env_stats.cs_stats[CS_##counter])
+#define CL_ENV_INC(counter) atomic_inc(&cl_env_stats.cs_stats[CS_##counter])
 
 #define CL_ENV_DEC(counter) do {                                              \
-	LASSERT(cfs_atomic_read(&cl_env_stats.cs_stats[CS_##counter]) > 0);   \
-	cfs_atomic_dec(&cl_env_stats.cs_stats[CS_##counter]);                 \
+	LASSERT(atomic_read(&cl_env_stats.cs_stats[CS_##counter]) > 0);   \
+	atomic_dec(&cl_env_stats.cs_stats[CS_##counter]);                 \
 } while (0)
 #else
 #define CL_ENV_INC(counter)
@@ -649,35 +659,35 @@ static cfs_hash_ops_t cl_env_hops = {
 
 static inline struct cl_env *cl_env_fetch(void)
 {
-        struct cl_env *cle;
+	struct cl_env *cle;
 
-        cle = cfs_hash_lookup(cl_env_hash, (void *) (long) cfs_current()->pid);
-        LASSERT(ergo(cle, cle->ce_magic == &cl_env_init0));
-        return cle;
+	cle = cfs_hash_lookup(cl_env_hash, (void *) (long) current->pid);
+	LASSERT(ergo(cle, cle->ce_magic == &cl_env_init0));
+	return cle;
 }
 
 static inline void cl_env_attach(struct cl_env *cle)
 {
-        if (cle) {
-                int rc;
-
-                LASSERT(cle->ce_owner == NULL);
-                cle->ce_owner = (void *) (long) cfs_current()->pid;
-                rc = cfs_hash_add_unique(cl_env_hash, cle->ce_owner,
-                                         &cle->ce_node);
-                LASSERT(rc == 0);
-        }
+	if (cle) {
+		int rc;
+
+		LASSERT(cle->ce_owner == NULL);
+		cle->ce_owner = (void *) (long) current->pid;
+		rc = cfs_hash_add_unique(cl_env_hash, cle->ce_owner,
+					 &cle->ce_node);
+		LASSERT(rc == 0);
+	}
 }
 
 static inline void cl_env_do_detach(struct cl_env *cle)
 {
-        void *cookie;
+	void *cookie;
 
-        LASSERT(cle->ce_owner == (void *) (long) cfs_current()->pid);
-        cookie = cfs_hash_del(cl_env_hash, cle->ce_owner,
-                              &cle->ce_node);
-        LASSERT(cookie == cle);
-        cle->ce_owner = NULL;
+	LASSERT(cle->ce_owner == (void *) (long) current->pid);
+	cookie = cfs_hash_del(cl_env_hash, cle->ce_owner,
+			      &cle->ce_node);
+	LASSERT(cookie == cle);
+	cle->ce_owner = NULL;
 }
 
 static int cl_env_store_init(void) {
@@ -702,30 +712,30 @@ static void cl_env_store_fini(void) {
 
 static inline struct cl_env *cl_env_fetch(void)
 {
-        struct cl_env *cle;
+	struct cl_env *cle;
 
-        cle = cfs_current()->LL_TASK_CL_ENV;
-        if (cle && cle->ce_magic != &cl_env_init0)
-                cle = NULL;
-        return cle;
+	cle = current->LL_TASK_CL_ENV;
+	if (cle && cle->ce_magic != &cl_env_init0)
+		cle = NULL;
+	return cle;
 }
 
 static inline void cl_env_attach(struct cl_env *cle)
 {
-        if (cle) {
-                LASSERT(cle->ce_owner == NULL);
-                cle->ce_owner = cfs_current();
-                cle->ce_prev = cfs_current()->LL_TASK_CL_ENV;
-                cfs_current()->LL_TASK_CL_ENV = cle;
-        }
+	if (cle) {
+		LASSERT(cle->ce_owner == NULL);
+		cle->ce_owner = current;
+		cle->ce_prev = current->LL_TASK_CL_ENV;
+		current->LL_TASK_CL_ENV = cle;
+	}
 }
 
 static inline void cl_env_do_detach(struct cl_env *cle)
 {
-        LASSERT(cle->ce_owner == cfs_current());
-        LASSERT(cfs_current()->LL_TASK_CL_ENV == cle);
-        cfs_current()->LL_TASK_CL_ENV = cle->ce_prev;
-        cle->ce_owner = NULL;
+	LASSERT(cle->ce_owner == current);
+	LASSERT(current->LL_TASK_CL_ENV == cle);
+	current->LL_TASK_CL_ENV = cle->ce_prev;
+	cle->ce_owner = NULL;
 }
 
 static int cl_env_store_init(void) { return 0; }
@@ -746,37 +756,37 @@ static inline struct cl_env *cl_env_detach(struct cl_env *cle)
 
 static struct lu_env *cl_env_new(__u32 ctx_tags, __u32 ses_tags, void *debug)
 {
-        struct lu_env *env;
-        struct cl_env *cle;
-
-	OBD_SLAB_ALLOC_PTR_GFP(cle, cl_env_kmem, __GFP_IO);
-        if (cle != NULL) {
-                int rc;
-
-                CFS_INIT_LIST_HEAD(&cle->ce_linkage);
-                cle->ce_magic = &cl_env_init0;
-                env = &cle->ce_lu;
-                rc = lu_env_init(env, LCT_CL_THREAD|ctx_tags);
-                if (rc == 0) {
-                        rc = lu_context_init(&cle->ce_ses,
-                                             LCT_SESSION | ses_tags);
-                        if (rc == 0) {
-                                lu_context_enter(&cle->ce_ses);
-                                env->le_ses = &cle->ce_ses;
-                                cl_env_init0(cle, debug);
-                        } else
-                                lu_env_fini(env);
-                }
-                if (rc != 0) {
-                        OBD_SLAB_FREE_PTR(cle, cl_env_kmem);
-                        env = ERR_PTR(rc);
-                } else {
-                        CL_ENV_INC(create);
-                        CL_ENV_INC(total);
-                }
-        } else
-                env = ERR_PTR(-ENOMEM);
-        return env;
+	struct lu_env *env;
+	struct cl_env *cle;
+
+	OBD_SLAB_ALLOC_PTR_GFP(cle, cl_env_kmem, GFP_NOFS);
+	if (cle != NULL) {
+		int rc;
+
+		CFS_INIT_LIST_HEAD(&cle->ce_linkage);
+		cle->ce_magic = &cl_env_init0;
+		env = &cle->ce_lu;
+		rc = lu_env_init(env, LCT_CL_THREAD|ctx_tags);
+		if (rc == 0) {
+			rc = lu_context_init(&cle->ce_ses,
+					     LCT_SESSION | ses_tags);
+			if (rc == 0) {
+				lu_context_enter(&cle->ce_ses);
+				env->le_ses = &cle->ce_ses;
+				cl_env_init0(cle, debug);
+			} else
+				lu_env_fini(env);
+		}
+		if (rc != 0) {
+			OBD_SLAB_FREE_PTR(cle, cl_env_kmem);
+			env = ERR_PTR(rc);
+		} else {
+			CL_ENV_INC(create);
+			CL_ENV_INC(total);
+		}
+	} else
+		env = ERR_PTR(-ENOMEM);
+	return env;
 }
 
 static void cl_env_fini(struct cl_env *cle)
@@ -787,6 +797,40 @@ static void cl_env_fini(struct cl_env *cle)
         OBD_SLAB_FREE_PTR(cle, cl_env_kmem);
 }
 
+static struct lu_env *cl_env_obtain(void *debug)
+{
+	struct cl_env *cle;
+	struct lu_env *env;
+
+	ENTRY;
+	spin_lock(&cl_envs_guard);
+	LASSERT(equi(cl_envs_cached_nr == 0, cfs_list_empty(&cl_envs)));
+	if (cl_envs_cached_nr > 0) {
+		int rc;
+
+		cle = container_of(cl_envs.next, struct cl_env, ce_linkage);
+		cfs_list_del_init(&cle->ce_linkage);
+		cl_envs_cached_nr--;
+		spin_unlock(&cl_envs_guard);
+
+                env = &cle->ce_lu;
+                rc = lu_env_refill(env);
+                if (rc == 0) {
+                        cl_env_init0(cle, debug);
+                        lu_context_enter(&env->le_ctx);
+                        lu_context_enter(&cle->ce_ses);
+                } else {
+                        cl_env_fini(cle);
+                        env = ERR_PTR(rc);
+                }
+        } else {
+		spin_unlock(&cl_envs_guard);
+		env = cl_env_new(lu_context_tags_default,
+				 lu_session_tags_default, debug);
+	}
+	RETURN(env);
+}
+
 static inline struct cl_env *cl_env_container(struct lu_env *env)
 {
         return container_of(env, struct cl_env, ce_lu);
@@ -818,6 +862,8 @@ EXPORT_SYMBOL(cl_env_peek);
  * Returns lu_env: if there already is an environment associated with the
  * current thread, it is returned, otherwise, new environment is allocated.
  *
+ * Allocations are amortized through the global cache of environments.
+ *
  * \param refcheck pointer to a counter used to detect environment leaks. In
  * the usual case cl_env_get() and cl_env_put() are called in the same lexical
  * scope and pointer to the same integer is passed as \a refcheck. This is
@@ -831,10 +877,7 @@ struct lu_env *cl_env_get(int *refcheck)
 
         env = cl_env_peek(refcheck);
         if (env == NULL) {
-		env = cl_env_new(lu_context_tags_default,
-				 lu_session_tags_default,
-				 __builtin_return_address(0));
-
+                env = cl_env_obtain(__builtin_return_address(0));
                 if (!IS_ERR(env)) {
                         struct cl_env *cle;
 
@@ -878,6 +921,33 @@ static void cl_env_exit(struct cl_env *cle)
 }
 
 /**
+ * Finalizes and frees a given number of cached environments. This is done to
+ * (1) free some memory (not currently hooked into VM), or (2) release
+ * references to modules.
+ */
+unsigned cl_env_cache_purge(unsigned nr)
+{
+	struct cl_env *cle;
+
+	ENTRY;
+	spin_lock(&cl_envs_guard);
+	for (; !cfs_list_empty(&cl_envs) && nr > 0; --nr) {
+		cle = container_of(cl_envs.next, struct cl_env, ce_linkage);
+		cfs_list_del_init(&cle->ce_linkage);
+		LASSERT(cl_envs_cached_nr > 0);
+		cl_envs_cached_nr--;
+		spin_unlock(&cl_envs_guard);
+
+		cl_env_fini(cle);
+		spin_lock(&cl_envs_guard);
+	}
+	LASSERT(equi(cl_envs_cached_nr == 0, cfs_list_empty(&cl_envs)));
+	spin_unlock(&cl_envs_guard);
+	RETURN(nr);
+}
+EXPORT_SYMBOL(cl_env_cache_purge);
+
+/**
  * Release an environment.
  *
  * Decrement \a env reference counter. When counter drops to 0, nothing in
@@ -899,7 +969,21 @@ void cl_env_put(struct lu_env *env, int *refcheck)
                 cl_env_detach(cle);
                 cle->ce_debug = NULL;
                 cl_env_exit(cle);
-		cl_env_fini(cle);
+                /*
+                 * Don't bother to take a lock here.
+                 *
+                 * Return environment to the cache only when it was allocated
+                 * with the standard tags.
+                 */
+                if (cl_envs_cached_nr < cl_envs_cached_max &&
+                    (env->le_ctx.lc_tags & ~LCT_HAS_EXIT) == LCT_CL_THREAD &&
+                    (env->le_ses->lc_tags & ~LCT_HAS_EXIT) == LCT_SESSION) {
+			spin_lock(&cl_envs_guard);
+			cfs_list_add(&cle->ce_linkage, &cl_envs);
+			cl_envs_cached_nr++;
+			spin_unlock(&cl_envs_guard);
+		} else
+			cl_env_fini(cle);
 	}
 }
 EXPORT_SYMBOL(cl_env_put);
@@ -1026,6 +1110,103 @@ void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb)
 }
 EXPORT_SYMBOL(cl_lvb2attr);
 
+static struct cl_env cl_env_percpu[NR_CPUS];
+
+static int cl_env_percpu_init(void)
+{
+	struct cl_env *cle;
+	int tags = LCT_REMEMBER | LCT_NOREF;
+	int i, j;
+	int rc = 0;
+
+	for_each_possible_cpu(i) {
+		struct lu_env *env;
+
+		cle = &cl_env_percpu[i];
+		env = &cle->ce_lu;
+
+		CFS_INIT_LIST_HEAD(&cle->ce_linkage);
+		cle->ce_magic = &cl_env_init0;
+		rc = lu_env_init(env, LCT_CL_THREAD | tags);
+		if (rc == 0) {
+			rc = lu_context_init(&cle->ce_ses, LCT_SESSION | tags);
+                        if (rc == 0) {
+                                lu_context_enter(&cle->ce_ses);
+                                env->le_ses = &cle->ce_ses;
+			} else {
+				lu_env_fini(env);
+			}
+		}
+		if (rc != 0)
+			break;
+	}
+	if (rc != 0) {
+		/* Indices 0 to i (excluding i) were correctly initialized,
+		 * thus we must uninitialize up to i, the rest are undefined. */
+		for (j = 0; j < i; j++) {
+			cle = &cl_env_percpu[i];
+			lu_context_exit(&cle->ce_ses);
+			lu_context_fini(&cle->ce_ses);
+			lu_env_fini(&cle->ce_lu);
+		}
+	}
+
+	return rc;
+}
+
+static void cl_env_percpu_fini(void)
+{
+	int i;
+
+	for_each_possible_cpu(i) {
+		struct cl_env *cle = &cl_env_percpu[i];
+
+		lu_context_exit(&cle->ce_ses);
+		lu_context_fini(&cle->ce_ses);
+		lu_env_fini(&cle->ce_lu);
+	}
+}
+
+static void cl_env_percpu_refill(void)
+{
+	int i;
+
+	for_each_possible_cpu(i)
+		lu_env_refill(&cl_env_percpu[i].ce_lu);
+}
+
+void cl_env_percpu_put(struct lu_env *env)
+{
+	struct cl_env *cle;
+	int cpu;
+
+	cpu = smp_processor_id();
+	cle = cl_env_container(env);
+	LASSERT(cle == &cl_env_percpu[cpu]);
+
+	cle->ce_ref--;
+	LASSERT(cle->ce_ref == 0);
+
+	CL_ENV_DEC(busy);
+	cl_env_detach(cle);
+	cle->ce_debug = NULL;
+
+	put_cpu();
+}
+EXPORT_SYMBOL(cl_env_percpu_put);
+
+struct lu_env *cl_env_percpu_get()
+{
+	struct cl_env *cle;
+
+	cle = &cl_env_percpu[get_cpu()];
+	cl_env_init0(cle, __builtin_return_address(0));
+
+	cl_env_attach(cle);
+	return &cle->ce_lu;
+}
+EXPORT_SYMBOL(cl_env_percpu_get);
+
 /*****************************************************************************
  *
  * Temporary prototype thing: mirror obd-devices into cl devices.
@@ -1181,6 +1362,11 @@ int cl_global_init(void)
         if (result)
                 goto out_lock;
 
+	result = cl_env_percpu_init();
+	if (result)
+		/* no cl_env_percpu_fini on error */
+		goto out_lock;
+
         return 0;
 out_lock:
         cl_lock_fini();
@@ -1198,6 +1384,7 @@ out_store:
  */
 void cl_global_fini(void)
 {
+	cl_env_percpu_fini();
         cl_lock_fini();
         cl_page_fini();
         lu_context_key_degister(&cl_key);