Whamcloud - gitweb
LU-2622 obdclass: Remove the global cl_env list
[fs/lustre-release.git] / lustre / obdclass / cl_object.c
index 0310b75..1eb8e76 100644 (file)
@@ -1,6 +1,4 @@
-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
  * GPL HEADER START
  *
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  * GPL HEADER END
  */
 /*
- * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
+ * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
+ *
+ * Copyright (c) 2011, 2012, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -50,9 +50,6 @@
  */
 
 #define DEBUG_SUBSYSTEM S_CLASS
-#ifndef EXPORT_SYMTAB
-# define EXPORT_SYMTAB
-#endif
 
 #include <libcfs/libcfs.h>
 /* class_put_type() */
@@ -60,9 +57,7 @@
 #include <obd_support.h>
 #include <lustre_fid.h>
 #include <libcfs/list.h>
-/* lu_time_global_{init,fini}() */
-#include <lu_time.h>
-
+#include <libcfs/libcfs_hash.h> /* for cfs_hash stuff */
 #include <cl_object.h>
 #include "cl_internal.h"
 
@@ -75,6 +70,8 @@ static struct lock_class_key cl_lock_guard_class;
 /** Lock class of cl_object_header::coh_attr_guard */
 static struct lock_class_key cl_attr_guard_class;
 
+extern __u32 lu_context_tags_default;
+extern __u32 lu_session_tags_default;
 /**
  * Initialize cl_object_header.
  */
@@ -85,16 +82,17 @@ int cl_object_header_init(struct cl_object_header *h)
         ENTRY;
         result = lu_object_header_init(&h->coh_lu);
         if (result == 0) {
-                spin_lock_init(&h->coh_page_guard);
-                spin_lock_init(&h->coh_lock_guard);
-                spin_lock_init(&h->coh_attr_guard);
-                lockdep_set_class(&h->coh_attr_guard, &cl_page_guard_class);
-                lockdep_set_class(&h->coh_attr_guard, &cl_lock_guard_class);
-                lockdep_set_class(&h->coh_attr_guard, &cl_attr_guard_class);
+               spin_lock_init(&h->coh_page_guard);
+               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);
 }
@@ -105,7 +103,7 @@ EXPORT_SYMBOL(cl_object_header_init);
  */
 void cl_object_header_fini(struct cl_object_header *h)
 {
-        LASSERT(list_empty(&h->coh_locks));
+        LASSERT(cfs_list_empty(&h->coh_locks));
         lu_object_header_fini(&h->coh_lu);
 }
 EXPORT_SYMBOL(cl_object_header_fini);
@@ -122,7 +120,7 @@ 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)
 {
-        might_sleep();
+        cfs_might_sleep();
         return lu2cl(lu_object_find_slice(env, cl2lu_dev(cd), fid, &c->coc_lu));
 }
 EXPORT_SYMBOL(cl_object_find);
@@ -185,7 +183,7 @@ EXPORT_SYMBOL(cl_object_top);
  */
 static spinlock_t *cl_object_attr_guard(struct cl_object *o)
 {
-        return &cl_object_header(cl_object_top(o))->coh_attr_guard;
+       return &cl_object_header(cl_object_top(o))->coh_attr_guard;
 }
 
 /**
@@ -197,7 +195,7 @@ static spinlock_t *cl_object_attr_guard(struct cl_object *o)
  */
 void cl_object_attr_lock(struct cl_object *o)
 {
-        spin_lock(cl_object_attr_guard(o));
+       spin_lock(cl_object_attr_guard(o));
 }
 EXPORT_SYMBOL(cl_object_attr_lock);
 
@@ -206,7 +204,7 @@ EXPORT_SYMBOL(cl_object_attr_lock);
  */
 void cl_object_attr_unlock(struct cl_object *o)
 {
-        spin_unlock(cl_object_attr_guard(o));
+       spin_unlock(cl_object_attr_guard(o));
 }
 EXPORT_SYMBOL(cl_object_attr_unlock);
 
@@ -228,7 +226,7 @@ int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
 
         top = obj->co_lu.lo_header;
         result = 0;
-        list_for_each_entry(obj, &top->loh_layers, co_lu.lo_linkage) {
+        cfs_list_for_each_entry(obj, &top->loh_layers, co_lu.lo_linkage) {
                 if (obj->co_ops->coo_attr_get != NULL) {
                         result = obj->co_ops->coo_attr_get(env, obj, attr);
                         if (result != 0) {
@@ -260,7 +258,8 @@ int cl_object_attr_set(const struct lu_env *env, struct cl_object *obj,
 
         top = obj->co_lu.lo_header;
         result = 0;
-        list_for_each_entry_reverse(obj, &top->loh_layers, co_lu.lo_linkage) {
+        cfs_list_for_each_entry_reverse(obj, &top->loh_layers,
+                                        co_lu.lo_linkage) {
                 if (obj->co_ops->coo_attr_set != NULL) {
                         result = obj->co_ops->coo_attr_set(env, obj, attr, v);
                         if (result != 0) {
@@ -291,7 +290,8 @@ int cl_object_glimpse(const struct lu_env *env, struct cl_object *obj,
         ENTRY;
         top = obj->co_lu.lo_header;
         result = 0;
-        list_for_each_entry_reverse(obj, &top->loh_layers, co_lu.lo_linkage) {
+        cfs_list_for_each_entry_reverse(obj, &top->loh_layers,
+                                        co_lu.lo_linkage) {
                 if (obj->co_ops->coo_glimpse != NULL) {
                         result = obj->co_ops->coo_glimpse(env, obj, lvb);
                         if (result != 0)
@@ -319,7 +319,7 @@ int cl_conf_set(const struct lu_env *env, struct cl_object *obj,
         ENTRY;
         top = obj->co_lu.lo_header;
         result = 0;
-        list_for_each_entry(obj, &top->loh_layers, co_lu.lo_linkage) {
+        cfs_list_for_each_entry(obj, &top->loh_layers, co_lu.lo_linkage) {
                 if (obj->co_ops->coo_conf_set != NULL) {
                         result = obj->co_ops->coo_conf_set(env, obj, conf);
                         if (result != 0)
@@ -345,7 +345,7 @@ void cl_object_kill(const struct lu_env *env, struct cl_object *obj)
         LASSERT(hdr->coh_tree.rnode == NULL);
         LASSERT(hdr->coh_pages == 0);
 
-        set_bit(LU_OBJECT_HEARD_BANSHEE, &hdr->coh_lu.loh_flags);
+       set_bit(LU_OBJECT_HEARD_BANSHEE, &hdr->coh_lu.loh_flags);
         /*
          * Destroy all locks. Object destruction (including cl_inode_fini())
          * cannot cancel the locks, because in the case of a local client,
@@ -369,36 +369,55 @@ void cl_object_prune(const struct lu_env *env, struct cl_object *obj)
 }
 EXPORT_SYMBOL(cl_object_prune);
 
+/**
+ * Check if the object has locks.
+ */
+int cl_object_has_locks(struct cl_object *obj)
+{
+       struct cl_object_header *head = cl_object_header(obj);
+       int has;
+
+       spin_lock(&head->coh_lock_guard);
+       has = cfs_list_empty(&head->coh_locks);
+       spin_unlock(&head->coh_lock_guard);
+
+       return (has == 0);
+}
+EXPORT_SYMBOL(cl_object_has_locks);
+
 void cache_stats_init(struct cache_stats *cs, const char *name)
 {
+       int i;
+
         cs->cs_name = name;
-        atomic_set(&cs->cs_lookup, 0);
-        atomic_set(&cs->cs_hit,    0);
-        atomic_set(&cs->cs_total,  0);
-        atomic_set(&cs->cs_busy,   0);
+       for (i = 0; i < CS_NR; i++)
+               cfs_atomic_set(&cs->cs_stats[i], 0);
 }
 
 int cache_stats_print(const struct cache_stats *cs,
                       char *page, int count, int h)
 {
-        int nob = 0;
-/*
-       lookup    hit  total cached create
-  env: ...... ...... ...... ...... ......
-*/
-        if (h)
-                nob += snprintf(page, count,
-                                "       lookup    hit  total   busy create\n");
-
-        nob += snprintf(page + nob, count - nob,
-                        "%5.5s: %6u %6u %6u %6u %6u",
-                        cs->cs_name,
-                        atomic_read(&cs->cs_lookup),
-                        atomic_read(&cs->cs_hit),
-                        atomic_read(&cs->cs_total),
-                        atomic_read(&cs->cs_busy),
-                        atomic_read(&cs->cs_created));
-        return nob;
+       int nob = 0;
+       int i;
+       /*
+        *   lookup    hit    total  cached create
+        * env: ...... ...... ...... ...... ......
+        */
+       if (h) {
+               const char *names[CS_NR] = CS_NAMES;
+
+               nob += snprintf(page + nob, count - nob, "%6s", " ");
+               for (i = 0; i < CS_NR; i++)
+                       nob += snprintf(page + nob, count - nob,
+                                       "%8s", names[i]);
+               nob += snprintf(page + nob, count - nob, "\n");
+       }
+
+       nob += snprintf(page + nob, count - nob, "%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;
 }
 
 /**
@@ -417,9 +436,9 @@ 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)
-                        atomic_set(&s->cs_pages_state[0], 0);
+                        cfs_atomic_set(&s->cs_pages_state[0], 0);
                 for (i = 0; i < ARRAY_SIZE(s->cs_locks_state); ++i)
-                        atomic_set(&s->cs_locks_state[i], 0);
+                        cfs_atomic_set(&s->cs_locks_state[i], 0);
         }
         return result;
 }
@@ -436,11 +455,7 @@ EXPORT_SYMBOL(cl_site_fini);
 
 static struct cache_stats cl_env_stats = {
         .cs_name    = "envs",
-        .cs_created = ATOMIC_INIT(0),
-        .cs_lookup  = ATOMIC_INIT(0),
-        .cs_hit     = ATOMIC_INIT(0),
-        .cs_total   = ATOMIC_INIT(0),
-        .cs_busy    = ATOMIC_INIT(0)
+       .cs_stats = { CFS_ATOMIC_INIT(0), }
 };
 
 /**
@@ -463,7 +478,7 @@ int cl_site_stats_print(const struct cl_site *site, char *page, int count)
                 [CLS_QUEUING]   = "q",
                 [CLS_ENQUEUED]  = "e",
                 [CLS_HELD]      = "h",
-                [CLS_UNLOCKING] = "u",
+                [CLS_INTRANSIT] = "t",
                 [CLS_CACHED]    = "c",
                 [CLS_FREEING]   = "f"
         };
@@ -479,14 +494,14 @@ locks: ...... ...... ...... ...... ...... [...... ...... ...... ...... ......]
         for (i = 0; i < ARRAY_SIZE(site->cs_pages_state); ++i)
                 nob += snprintf(page + nob, count - nob, "%s: %u ",
                                 pstate[i],
-                                atomic_read(&site->cs_pages_state[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],
-                                atomic_read(&site->cs_locks_state[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");
@@ -500,48 +515,78 @@ EXPORT_SYMBOL(cl_site_stats_print);
  *
  */
 
-/*
- * TBD: Description.
+/**
+ * The most efficient way is to store cl_env pointer in task specific
+ * structures. On Linux, it wont' be easy to use task_struct->journal_info
+ * because Lustre code may call into other fs which has certain assumptions
+ * about journal_info. Currently following fields in task_struct are identified
+ * can be used for this purpose:
+ *  - cl_env: for liblustre.
+ *  - tux_info: ony on RedHat kernel.
+ *  - ...
+ * \note As long as we use task_struct to store cl_env, we assume that once
+ * called into Lustre, we'll never call into the other part of the kernel
+ * which will use those fields in task_struct without explicitly exiting
+ * Lustre.
  *
- * XXX: this assumes that re-entrant file system calls (e.g., ->writepage())
- * do not modify already existing current->journal_info.
+ * If there's no space in task_struct is available, hash will be used.
+ * 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 spinlock_t cl_envs_guard = SPIN_LOCK_UNLOCKED;
-
 struct cl_env {
         void             *ce_magic;
         struct lu_env     ce_lu;
         struct lu_context ce_ses;
+
+#ifdef LL_TASK_CL_ENV
+        void             *ce_prev;
+#else
+        /**
+         * This allows cl_env to be entered into cl_env_hash which implements
+         * the current thread -> client environment lookup.
+         */
+        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;
+
         /*
          * Linkage into global list of all client environments. Used for
          * garbage collection.
          */
-        struct list_head  ce_linkage;
+        cfs_list_t        ce_linkage;
         /*
          *
          */
         int               ce_ref;
-        void             *ce_prev;
         /*
          * Debugging field: address of the caller who made original
          * allocation.
          */
         void             *ce_debug;
-        void             *ce_owner;
 };
 
-#define CL_ENV_INC(counter) atomic_inc(&cl_env_stats.counter)
+#ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
+#define CL_ENV_INC(counter) cfs_atomic_inc(&cl_env_stats.cs_stats[CS_##counter])
 
-#define CL_ENV_DEC(counter)                                             \
-        do {                                                            \
-                LASSERT(atomic_read(&cl_env_stats.counter) > 0);        \
-                atomic_dec(&cl_env_stats.counter);                      \
-        } while (0)
+#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]);                 \
+} while (0)
+#else
+#define CL_ENV_INC(counter)
+#define CL_ENV_DEC(counter)
+#endif
 
 static void cl_env_init0(struct cl_env *cle, void *debug)
 {
@@ -550,28 +595,171 @@ static void cl_env_init0(struct cl_env *cle, void *debug)
         LASSERT(cle->ce_debug == NULL && cle->ce_owner == NULL);
 
         cle->ce_ref = 1;
-        cle->ce_prev = current->journal_info;
         cle->ce_debug = debug;
-        cle->ce_owner = current;
-        current->journal_info = cle;
-        CL_ENV_INC(cs_busy);
+        CL_ENV_INC(busy);
+}
+
+
+#ifndef LL_TASK_CL_ENV
+/*
+ * The implementation of using hash table to connect cl_env and thread
+ */
+
+static cfs_hash_t *cl_env_hash;
+
+static unsigned cl_env_hops_hash(cfs_hash_t *lh,
+                                 const void *key, unsigned mask)
+{
+#if BITS_PER_LONG == 64
+        return cfs_hash_u64_hash((__u64)key, mask);
+#else
+        return cfs_hash_u32_hash((__u32)key, mask);
+#endif
+}
+
+static void *cl_env_hops_obj(cfs_hlist_node_t *hn)
+{
+        struct cl_env *cle = cfs_hlist_entry(hn, struct cl_env, ce_node);
+        LASSERT(cle->ce_magic == &cl_env_init0);
+        return (void *)cle;
+}
+
+static int cl_env_hops_keycmp(const void *key, cfs_hlist_node_t *hn)
+{
+        struct cl_env *cle = cl_env_hops_obj(hn);
+
+        LASSERT(cle->ce_owner != NULL);
+        return (key == cle->ce_owner);
+}
+
+static void cl_env_hops_noop(cfs_hash_t *hs, cfs_hlist_node_t *hn)
+{
+        struct cl_env *cle = cfs_hlist_entry(hn, struct cl_env, ce_node);
+        LASSERT(cle->ce_magic == &cl_env_init0);
+}
+
+static cfs_hash_ops_t cl_env_hops = {
+        .hs_hash        = cl_env_hops_hash,
+        .hs_key         = cl_env_hops_obj,
+        .hs_keycmp      = cl_env_hops_keycmp,
+        .hs_object      = cl_env_hops_obj,
+        .hs_get         = cl_env_hops_noop,
+        .hs_put_locked  = cl_env_hops_noop,
+};
+
+static inline struct cl_env *cl_env_fetch(void)
+{
+        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;
+}
+
+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);
+        }
+}
+
+static inline void cl_env_do_detach(struct cl_env *cle)
+{
+        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;
+}
+
+static int cl_env_store_init(void) {
+        cl_env_hash = cfs_hash_create("cl_env",
+                                      HASH_CL_ENV_BITS, HASH_CL_ENV_BITS,
+                                      HASH_CL_ENV_BKT_BITS, 0,
+                                      CFS_HASH_MIN_THETA,
+                                      CFS_HASH_MAX_THETA,
+                                      &cl_env_hops,
+                                      CFS_HASH_RW_BKTLOCK);
+        return cl_env_hash != NULL ? 0 :-ENOMEM;
+}
+
+static void cl_env_store_fini(void) {
+        cfs_hash_putref(cl_env_hash);
+}
+
+#else /* LL_TASK_CL_ENV */
+/*
+ * The implementation of store cl_env directly in thread structure.
+ */
+
+static inline struct cl_env *cl_env_fetch(void)
+{
+        struct cl_env *cle;
+
+        cle = cfs_current()->LL_TASK_CL_ENV;
+        if (cle && cle->ce_magic != &cl_env_init0)
+                cle = NULL;
+        return cle;
 }
 
-static struct lu_env *cl_env_new(__u32 tags, void *debug)
+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;
+        }
+}
+
+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;
+}
+
+static int cl_env_store_init(void) { return 0; }
+static void cl_env_store_fini(void) { }
+
+#endif /* LL_TASK_CL_ENV */
+
+static inline struct cl_env *cl_env_detach(struct cl_env *cle)
+{
+        if (cle == NULL)
+                cle = cl_env_fetch();
+
+        if (cle && cle->ce_owner)
+                cl_env_do_detach(cle);
+
+        return 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(cle, cl_env_kmem);
+        OBD_SLAB_ALLOC_PTR_GFP(cle, cl_env_kmem, CFS_ALLOC_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|tags);
+                rc = lu_env_init(env, LCT_CL_THREAD|ctx_tags);
                 if (rc == 0) {
-                        rc = lu_context_init(&cle->ce_ses, LCT_SESSION|tags);
+                        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;
@@ -583,8 +771,8 @@ static struct lu_env *cl_env_new(__u32 tags, void *debug)
                         OBD_SLAB_FREE_PTR(cle, cl_env_kmem);
                         env = ERR_PTR(rc);
                 } else {
-                        CL_ENV_INC(cs_created);
-                        CL_ENV_INC(cs_total);
+                        CL_ENV_INC(create);
+                        CL_ENV_INC(total);
                 }
         } else
                 env = ERR_PTR(-ENOMEM);
@@ -593,45 +781,12 @@ static struct lu_env *cl_env_new(__u32 tags, void *debug)
 
 static void cl_env_fini(struct cl_env *cle)
 {
-        CL_ENV_DEC(cs_total);
+        CL_ENV_DEC(total);
         lu_context_fini(&cle->ce_lu.le_ctx);
         lu_context_fini(&cle->ce_ses);
         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, list_empty(&cl_envs)));
-        if (cl_envs_cached_nr > 0) {
-                int rc;
-
-                cle = container_of(cl_envs.next, struct cl_env, ce_linkage);
-                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(0, debug);
-        }
-        RETURN(env);
-}
-
 static inline struct cl_env *cl_env_container(struct lu_env *env)
 {
         return container_of(env, struct cl_env, ce_lu);
@@ -642,19 +797,19 @@ struct lu_env *cl_env_peek(int *refcheck)
         struct lu_env *env;
         struct cl_env *cle;
 
-        CL_ENV_INC(cs_lookup);
+        CL_ENV_INC(lookup);
 
         /* check that we don't go far from untrusted pointer */
         CLASSERT(offsetof(struct cl_env, ce_magic) == 0);
 
         env = NULL;
-        cle = current->journal_info;
-        if (cle != NULL && cle->ce_magic == &cl_env_init0) {
-                CL_ENV_INC(cs_hit);
+        cle = cl_env_fetch();
+        if (cle != NULL) {
+                CL_ENV_INC(hit);
                 env = &cle->ce_lu;
                 *refcheck = ++cle->ce_ref;
         }
-        CDEBUG(D_OTHER, "%i@%p\n", cle ? cle->ce_ref : 0, cle);
+        CDEBUG(D_OTHER, "%d@%p\n", cle ? cle->ce_ref : 0, cle);
         return env;
 }
 EXPORT_SYMBOL(cl_env_peek);
@@ -663,8 +818,6 @@ 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
@@ -678,13 +831,17 @@ struct lu_env *cl_env_get(int *refcheck)
 
         env = cl_env_peek(refcheck);
         if (env == NULL) {
-                env = cl_env_obtain(__builtin_return_address(0));
+               env = cl_env_new(lu_context_tags_default,
+                                lu_session_tags_default,
+                                __builtin_return_address(0));
+
                 if (!IS_ERR(env)) {
                         struct cl_env *cle;
 
                         cle = cl_env_container(env);
+                        cl_env_attach(cle);
                         *refcheck = cle->ce_ref;
-                        CDEBUG(D_OTHER, "%i@%p\n", cle->ce_ref, cle);
+                        CDEBUG(D_OTHER, "%d@%p\n", cle->ce_ref, cle);
                 }
         }
         return env;
@@ -701,13 +858,13 @@ struct lu_env *cl_env_alloc(int *refcheck, __u32 tags)
         struct lu_env *env;
 
         LASSERT(cl_env_peek(refcheck) == NULL);
-        env = cl_env_new(tags, __builtin_return_address(0));
+        env = cl_env_new(tags, tags, __builtin_return_address(0));
         if (!IS_ERR(env)) {
                 struct cl_env *cle;
 
                 cle = cl_env_container(env);
                 *refcheck = cle->ce_ref;
-                CDEBUG(D_OTHER, "%i@%p\n", cle->ce_ref, cle);
+                CDEBUG(D_OTHER, "%d@%p\n", cle->ce_ref, cle);
         }
         return env;
 }
@@ -715,38 +872,12 @@ EXPORT_SYMBOL(cl_env_alloc);
 
 static void cl_env_exit(struct cl_env *cle)
 {
+        LASSERT(cle->ce_owner == NULL);
         lu_context_exit(&cle->ce_lu.le_ctx);
         lu_context_exit(&cle->ce_ses);
 }
 
 /**
- * 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 (; !list_empty(&cl_envs) && nr > 0; --nr) {
-                cle = container_of(cl_envs.next, struct cl_env, ce_linkage);
-                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, 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
@@ -762,68 +893,35 @@ void cl_env_put(struct lu_env *env, int *refcheck)
         LASSERT(cle->ce_ref > 0);
         LASSERT(ergo(refcheck != NULL, cle->ce_ref == *refcheck));
 
-        CDEBUG(D_OTHER, "%i@%p\n", cle->ce_ref, cle);
+        CDEBUG(D_OTHER, "%d@%p\n", cle->ce_ref, cle);
         if (--cle->ce_ref == 0) {
-                CL_ENV_DEC(cs_busy);
-                current->journal_info = cle->ce_prev;
-                LASSERT(cle->ce_prev == NULL ||
-                        cl_env_container(cle->ce_prev)->ce_magic !=
-                        &cl_env_init0);
+                CL_ENV_DEC(busy);
+                cl_env_detach(cle);
                 cle->ce_debug = NULL;
-                cle->ce_owner = NULL;
                 cl_env_exit(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);
-                        list_add(&cle->ce_linkage, &cl_envs);
-                        cl_envs_cached_nr++;
-                        spin_unlock(&cl_envs_guard);
-                } else
-                        cl_env_fini(cle);
-        }
+               cl_env_fini(cle);
+       }
 }
 EXPORT_SYMBOL(cl_env_put);
 
 /**
  * Declares a point of re-entrancy.
  *
- * In Linux kernel environments are attached to the thread through
- * current->journal_info pointer that is used by other sub-systems also. When
- * lustre code is invoked in the situation where current->journal_info is
- * potentially already set, cl_env_reenter() is called to save
- * current->journal_info value, so that current->journal_info field can be
- * used to store pointer to the environment.
- *
  * \see cl_env_reexit()
  */
 void *cl_env_reenter(void)
 {
-        void *cookie;
-
-        cookie = current->journal_info;
-        current->journal_info = NULL;
-        CDEBUG(D_OTHER, "cookie: %p\n", cookie);
-        return cookie;
+        return cl_env_detach(NULL);
 }
 EXPORT_SYMBOL(cl_env_reenter);
 
 /**
  * Exits re-entrancy.
- *
- * This restores old value of current->journal_info that was saved by
- * cl_env_reenter().
  */
 void cl_env_reexit(void *cookie)
 {
-        current->journal_info = cookie;
-        CDEBUG(D_OTHER, "cookie: %p\n", cookie);
+        cl_env_detach(NULL);
+        cl_env_attach(cookie);
 }
 EXPORT_SYMBOL(cl_env_reexit);
 
@@ -838,12 +936,11 @@ void cl_env_implant(struct lu_env *env, int *refcheck)
 {
         struct cl_env *cle = cl_env_container(env);
 
-        LASSERT(current->journal_info == NULL);
         LASSERT(cle->ce_ref > 0);
 
-        current->journal_info = cle;
+        cl_env_attach(cle);
         cl_env_get(refcheck);
-        CDEBUG(D_OTHER, "%i@%p\n", cle->ce_ref, cle);
+        CDEBUG(D_OTHER, "%d@%p\n", cle->ce_ref, cle);
 }
 EXPORT_SYMBOL(cl_env_implant);
 
@@ -854,13 +951,12 @@ void cl_env_unplant(struct lu_env *env, int *refcheck)
 {
         struct cl_env *cle = cl_env_container(env);
 
-        LASSERT(cle == current->journal_info);
         LASSERT(cle->ce_ref > 1);
 
-        CDEBUG(D_OTHER, "%i@%p\n", cle->ce_ref, cle);
+        CDEBUG(D_OTHER, "%d@%p\n", cle->ce_ref, cle);
 
+        cl_env_detach(cle);
         cl_env_put(env, refcheck);
-        current->journal_info = NULL;
 }
 EXPORT_SYMBOL(cl_env_unplant);
 
@@ -879,7 +975,12 @@ struct lu_env *cl_env_nested_get(struct cl_env_nest *nest)
                 }
         }
         env = cl_env_get(&nest->cen_refcheck);
-        LASSERT(ergo(!IS_ERR(env), !cl_io_is_going(env)));
+        if (IS_ERR(env)) {
+                cl_env_reexit(nest->cen_cookie);
+                return env;
+        }
+
+        LASSERT(!cl_io_is_going(env));
         return env;
 }
 EXPORT_SYMBOL(cl_env_nested_get);
@@ -925,7 +1026,6 @@ void cl_lvb2attr(struct cl_attr *attr, const struct ost_lvb *lvb)
 }
 EXPORT_SYMBOL(cl_lvb2attr);
 
-
 /*****************************************************************************
  *
  * Temporary prototype thing: mirror obd-devices into cl devices.
@@ -995,8 +1095,12 @@ static void *cl_key_init(const struct lu_context *ctx,
         struct cl_thread_info *info;
 
         info = cl0_key_init(ctx, key);
-        if (!IS_ERR(info))
-                lu_ref_init(&info->clt_locks_locked);
+        if (!IS_ERR(info)) {
+                int i;
+
+                for (i = 0; i < ARRAY_SIZE(info->clt_counters); ++i)
+                        lu_ref_init(&info->clt_counters[i].ctc_locks_locked);
+        }
         return info;
 }
 
@@ -1004,9 +1108,11 @@ static void cl_key_fini(const struct lu_context *ctx,
                         struct lu_context_key *key, void *data)
 {
         struct cl_thread_info *info;
+        int i;
 
         info = data;
-        lu_ref_fini(&info->clt_locks_locked);
+        for (i = 0; i < ARRAY_SIZE(info->clt_counters); ++i)
+                lu_ref_fini(&info->clt_counters[i].ctc_locks_locked);
         cl0_key_fini(ctx, key, data);
 }
 
@@ -1014,14 +1120,16 @@ static void cl_key_exit(const struct lu_context *ctx,
                         struct lu_context_key *key, void *data)
 {
         struct cl_thread_info *info = data;
+        int i;
 
-        LASSERT(info->clt_nr_locks_locked == 0);
-        LASSERT(info->clt_nr_held == 0);
-        LASSERT(info->clt_nr_used == 0);
-        LASSERT(info->clt_nr_locks_acquired == 0);
-
-        lu_ref_fini(&info->clt_locks_locked);
-        lu_ref_init(&info->clt_locks_locked);
+        for (i = 0; i < ARRAY_SIZE(info->clt_counters); ++i) {
+                LASSERT(info->clt_counters[i].ctc_nr_held == 0);
+                LASSERT(info->clt_counters[i].ctc_nr_used == 0);
+                LASSERT(info->clt_counters[i].ctc_nr_locks_acquired == 0);
+                LASSERT(info->clt_counters[i].ctc_nr_locks_locked == 0);
+                lu_ref_fini(&info->clt_counters[i].ctc_locks_locked);
+                lu_ref_init(&info->clt_counters[i].ctc_locks_locked);
+        }
 }
 
 static struct lu_context_key cl_key = {
@@ -1052,16 +1160,36 @@ int cl_global_init(void)
 {
         int result;
 
+        result = cl_env_store_init();
+        if (result)
+                return result;
+
         result = lu_kmem_init(cl_object_caches);
-        if (result == 0) {
-                LU_CONTEXT_KEY_INIT(&cl_key);
-                result = lu_context_key_register(&cl_key);
-                if (result == 0) {
-                        result = cl_lock_init();
-                        if (result == 0)
-                                result = cl_page_init();
-                }
-        }
+        if (result)
+                goto out_store;
+
+        LU_CONTEXT_KEY_INIT(&cl_key);
+        result = lu_context_key_register(&cl_key);
+        if (result)
+                goto out_kmem;
+
+        result = cl_lock_init();
+        if (result)
+                goto out_context;
+
+        result = cl_page_init();
+        if (result)
+                goto out_lock;
+
+        return 0;
+out_lock:
+        cl_lock_fini();
+out_context:
+        lu_context_key_degister(&cl_key);
+out_kmem:
+        lu_kmem_fini(cl_object_caches);
+out_store:
+        cl_env_store_fini();
         return result;
 }
 
@@ -1074,4 +1202,5 @@ void cl_global_fini(void)
         cl_page_fini();
         lu_context_key_degister(&cl_key);
         lu_kmem_fini(cl_object_caches);
+        cl_env_store_fini();
 }