Whamcloud - gitweb
LU-5823 clio: add coo_obd_info_get and coo_data_version
[fs/lustre-release.git] / lustre / obdclass / cl_object.c
index a7236fc..b003952 100644 (file)
@@ -27,7 +27,7 @@
  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2011, 2013, Intel Corporation.
+ * Copyright (c) 2011, 2014, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -36,6 +36,7 @@
  * Client Lustre Object.
  *
  *   Author: Nikita Danilov <nikita.danilov@sun.com>
+ *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  */
 
 /*
@@ -43,8 +44,6 @@
  *
  *  i_mutex
  *      PG_locked
- *          ->coh_page_guard
- *          ->coh_lock_guard
  *          ->coh_attr_guard
  *          ->ls_guard
  */
 #include <libcfs/list.h>
 #include <libcfs/libcfs_hash.h> /* for cfs_hash stuff */
 #include <cl_object.h>
+#include <lu_object.h>
 #include "cl_internal.h"
 
 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 */
 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.
  */
 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);
-               spin_lock_init(&h->coh_lock_guard);
+       ENTRY;
+       result = lu_object_header_init(&h->coh_lu);
+       if (result == 0) {
                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);
+               h->coh_page_bufsize = 0;
+       }
+       RETURN(result);
 }
 EXPORT_SYMBOL(cl_object_header_init);
 
@@ -103,10 +89,8 @@ EXPORT_SYMBOL(cl_object_header_init);
  */
 void cl_object_header_fini(struct cl_object_header *h)
 {
-        LASSERT(cfs_list_empty(&h->coh_locks));
         lu_object_header_fini(&h->coh_lu);
 }
-EXPORT_SYMBOL(cl_object_header_fini);
 
 /**
  * Returns a cl_object with a given \a fid.
@@ -156,7 +140,7 @@ 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)
 {
@@ -191,9 +175,10 @@ static spinlock_t *cl_object_attr_guard(struct cl_object *o)
  *
  * Prevents data-attributes from changing, until lock is released by
  * cl_object_attr_unlock(). This has to be called before calls to
- * cl_object_attr_get(), cl_object_attr_set().
+ * cl_object_attr_get(), cl_object_attr_update().
  */
 void cl_object_attr_lock(struct cl_object *o)
+__acquires(cl_object_attr_guard(o))
 {
        spin_lock(cl_object_attr_guard(o));
 }
@@ -203,6 +188,7 @@ EXPORT_SYMBOL(cl_object_attr_lock);
  * Releases data-attributes lock, acquired by cl_object_attr_lock().
  */
 void cl_object_attr_unlock(struct cl_object *o)
+__releases(cl_object_attr_guard(o))
 {
        spin_unlock(cl_object_attr_guard(o));
 }
@@ -216,27 +202,27 @@ EXPORT_SYMBOL(cl_object_attr_unlock);
  * for.
  */
 int cl_object_attr_get(const struct lu_env *env, struct cl_object *obj,
-                       struct cl_attr *attr)
+                       struct cl_attr *attr)
 {
        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;
-        result = 0;
-        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) {
-                                if (result > 0)
-                                        result = 0;
-                                break;
-                        }
-                }
-        }
-        RETURN(result);
+       top = obj->co_lu.lo_header;
+       result = 0;
+       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) {
+                               if (result > 0)
+                                       result = 0;
+                               break;
+                       }
+               }
+       }
+       RETURN(result);
 }
 EXPORT_SYMBOL(cl_object_attr_get);
 
@@ -244,34 +230,34 @@ EXPORT_SYMBOL(cl_object_attr_get);
  * Updates data-attributes of an object \a obj.
  *
  * Only attributes, mentioned in a validness bit-mask \a v are
- * updated. Calls cl_object_operations::coo_attr_set() on every layer, bottom
+ * updated. Calls cl_object_operations::coo_upd_attr() on every layer, bottom
  * to top.
  */
-int cl_object_attr_set(const struct lu_env *env, struct cl_object *obj,
-                       const struct cl_attr *attr, unsigned v)
+int cl_object_attr_update(const struct lu_env *env, struct cl_object *obj,
+                         const struct cl_attr *attr, unsigned v)
 {
        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;
-        result = 0;
-        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) {
-                                if (result > 0)
-                                        result = 0;
-                                break;
-                        }
-                }
-        }
-        RETURN(result);
+       top = obj->co_lu.lo_header;
+       result = 0;
+       list_for_each_entry_reverse(obj, &top->loh_layers, co_lu.lo_linkage) {
+               if (obj->co_ops->coo_attr_update != NULL) {
+                       result = obj->co_ops->coo_attr_update(env, obj, attr,
+                                                             v);
+                       if (result != 0) {
+                               if (result > 0)
+                                       result = 0;
+                               break;
+                       }
+               }
+       }
+       RETURN(result);
 }
-EXPORT_SYMBOL(cl_object_attr_set);
+EXPORT_SYMBOL(cl_object_attr_update);
 
 /**
  * Notifies layers (bottom-to-top) that glimpse AST was received.
@@ -290,8 +276,7 @@ int cl_object_glimpse(const struct lu_env *env, struct cl_object *obj,
         ENTRY;
         top = obj->co_lu.lo_header;
         result = 0;
-        cfs_list_for_each_entry_reverse(obj, &top->loh_layers,
-                                        co_lu.lo_linkage) {
+       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 +304,7 @@ int cl_conf_set(const struct lu_env *env, struct cl_object *obj,
         ENTRY;
         top = obj->co_lu.lo_header;
         result = 0;
-        cfs_list_for_each_entry(obj, &top->loh_layers, co_lu.lo_linkage) {
+       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)
@@ -331,59 +316,163 @@ int cl_conf_set(const struct lu_env *env, struct cl_object *obj,
 EXPORT_SYMBOL(cl_conf_set);
 
 /**
- * Helper function removing all object locks, and marking object for
- * deletion. All object pages must have been deleted at this point.
- *
- * This is called by cl_inode_fini() and lov_object_delete() to destroy top-
- * and sub- objects respectively.
+ * Prunes caches of pages and locks for this object.
  */
-void cl_object_kill(const struct lu_env *env, struct cl_object *obj)
+int cl_object_prune(const struct lu_env *env, struct cl_object *obj)
 {
-        struct cl_object_header *hdr;
+       struct lu_object_header *top;
+       struct cl_object *o;
+       int result;
+       ENTRY;
 
-        hdr = cl_object_header(obj);
-        LASSERT(hdr->coh_tree.rnode == NULL);
-        LASSERT(hdr->coh_pages == 0);
+       top = obj->co_lu.lo_header;
+       result = 0;
+       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;
+               }
+       }
 
-       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,
-         * where client and server share the same thread running
-         * prune_icache(), this can dead-lock with ldlm_cancel_handler()
-         * waiting on __wait_on_freeing_inode().
-         */
-        cl_locks_prune(env, obj, 0);
+       RETURN(result);
 }
-EXPORT_SYMBOL(cl_object_kill);
+EXPORT_SYMBOL(cl_object_prune);
 
 /**
- * Prunes caches of pages and locks for this object.
+ * Get stripe information of this object.
  */
-void cl_object_prune(const struct lu_env *env, struct cl_object *obj)
+int cl_object_getstripe(const struct lu_env *env, struct cl_object *obj,
+                       struct lov_user_md __user *uarg)
 {
-        ENTRY;
-        cl_pages_prune(env, obj);
-        cl_locks_prune(env, obj, 1);
-        EXIT;
+       struct lu_object_header *top;
+       int                     result = 0;
+       ENTRY;
+
+       top = obj->co_lu.lo_header;
+       list_for_each_entry(obj, &top->loh_layers, co_lu.lo_linkage) {
+               if (obj->co_ops->coo_getstripe != NULL) {
+                       result = obj->co_ops->coo_getstripe(env, obj, uarg);
+                       if (result != 0)
+                               break;
+               }
+       }
+       RETURN(result);
 }
-EXPORT_SYMBOL(cl_object_prune);
+EXPORT_SYMBOL(cl_object_getstripe);
+
+/**
+ * Find whether there is any callback data (ldlm lock) attached upon this
+ * object.
+ */
+int cl_object_find_cbdata(const struct lu_env *env, struct cl_object *obj,
+                          ldlm_iterator_t iter, void *data)
+{
+       struct lu_object_header *top;
+       int                     result = 0;
+       ENTRY;
+
+       top = obj->co_lu.lo_header;
+       list_for_each_entry(obj, &top->loh_layers, co_lu.lo_linkage) {
+               if (obj->co_ops->coo_find_cbdata != NULL) {
+                       result = obj->co_ops->coo_find_cbdata(env, obj, iter,
+                                                             data);
+                       if (result != 0)
+                               break;
+               }
+       }
+       RETURN(result);
+}
+EXPORT_SYMBOL(cl_object_find_cbdata);
 
 /**
- * Check if the object has locks.
+ * Get fiemap extents from file object.
+ *
+ * \param env [in]     lustre environment
+ * \param obj [in]     file object
+ * \param key [in]     fiemap request argument
+ * \param fiemap [out] fiemap extents mapping retrived
+ * \param buflen [in]  max buffer length of @fiemap
+ *
+ * \retval 0   success
+ * \retval < 0 error
  */
-int cl_object_has_locks(struct cl_object *obj)
+int cl_object_fiemap(const struct lu_env *env, struct cl_object *obj,
+                    struct ll_fiemap_info_key *key,
+                    struct fiemap *fiemap, size_t *buflen)
+{
+       struct lu_object_header *top;
+       int                     result = 0;
+       ENTRY;
+
+       top = obj->co_lu.lo_header;
+       list_for_each_entry(obj, &top->loh_layers, co_lu.lo_linkage) {
+               if (obj->co_ops->coo_fiemap != NULL) {
+                       result = obj->co_ops->coo_fiemap(env, obj, key, fiemap,
+                                                        buflen);
+                       if (result != 0)
+                               break;
+               }
+       }
+       RETURN(result);
+}
+EXPORT_SYMBOL(cl_object_fiemap);
+
+int cl_object_obd_info_get(const struct lu_env *env, struct cl_object *obj,
+                          struct obd_info *oinfo,
+                          struct ptlrpc_request_set *set)
+{
+       struct lu_object_header *top;
+       int                     result = 0;
+       ENTRY;
+
+       top = obj->co_lu.lo_header;
+       list_for_each_entry(obj, &top->loh_layers, co_lu.lo_linkage) {
+               if (obj->co_ops->coo_obd_info_get != NULL) {
+                       result = obj->co_ops->coo_obd_info_get(env, obj, oinfo,
+                                                              set);
+                       if (result != 0)
+                               break;
+               }
+       }
+       RETURN(result);
+}
+EXPORT_SYMBOL(cl_object_obd_info_get);
+
+int cl_object_data_version(const struct lu_env *env, struct cl_object *obj,
+                          __u64 *data_version, int flags)
 {
-       struct cl_object_header *head = cl_object_header(obj);
-       int has;
+       struct lu_object_header *top;
+       int                     result = 0;
+       ENTRY;
+
+       top = obj->co_lu.lo_header;
+       list_for_each_entry(obj, &top->loh_layers, co_lu.lo_linkage) {
+               if (obj->co_ops->coo_data_version != NULL) {
+                       result = obj->co_ops->coo_data_version(env, obj,
+                                                       data_version, flags);
+                       if (result != 0)
+                               break;
+               }
+       }
+       RETURN(result);
+}
+EXPORT_SYMBOL(cl_object_data_version);
 
-       spin_lock(&head->coh_lock_guard);
-       has = cfs_list_empty(&head->coh_locks);
-       spin_unlock(&head->coh_lock_guard);
+/**
+ * Helper function removing all object locks, and marking object for
+ * deletion. All object pages must have been deleted at this point.
+ *
+ * This is called by cl_inode_fini() and lov_object_delete() to destroy top-
+ * and sub- objects respectively.
+ */
+void cl_object_kill(const struct lu_env *env, struct cl_object *obj)
+{
+       struct cl_object_header *hdr = cl_object_header(obj);
 
-       return (has == 0);
+       set_bit(LU_OBJECT_HEARD_BANSHEE, &hdr->coh_lu.loh_flags);
 }
-EXPORT_SYMBOL(cl_object_has_locks);
+EXPORT_SYMBOL(cl_object_kill);
 
 void cache_stats_init(struct cache_stats *cs, const char *name)
 {
@@ -391,14 +480,14 @@ 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)
+static 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 +495,20 @@ 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.
  *
@@ -428,19 +517,17 @@ int cache_stats_print(const struct cache_stats *cs,
  */
 int cl_site_init(struct cl_site *s, struct cl_device *d)
 {
-        int i;
+       size_t i;
         int result;
 
         result = lu_site_init(&s->cs_lu, &d->cd_lu_dev);
         if (result == 0) {
                 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);
-                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_pages_state[0], 0);
+               cl_env_percpu_refill();
+       }
+       return result;
 }
 EXPORT_SYMBOL(cl_site_init);
 
@@ -455,57 +542,40 @@ 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 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"
-        };
+int cl_site_stats_print(const struct cl_site *site, struct seq_file *m)
+{
+       static const char *pstate[] = {
+               [CPS_CACHED]    = "c",
+               [CPS_OWNED]     = "o",
+               [CPS_PAGEOUT]   = "w",
+               [CPS_PAGEIN]    = "r",
+               [CPS_FREEING]   = "f"
+       };
+       size_t 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(&cl_env_stats, m, 0);
+       seq_printf(m, "\n");
+       return 0;
 }
 EXPORT_SYMBOL(cl_site_stats_print);
 
@@ -533,6 +603,12 @@ EXPORT_SYMBOL(cl_site_stats_print);
  * bz20044, bz22683.
  */
 
+static struct 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;
@@ -545,7 +621,7 @@ struct cl_env {
          * 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;
+       struct hlist_node  ce_node;
 #endif
        /**
         * Owner for the current cl_env.
@@ -564,7 +640,7 @@ struct cl_env {
          * Linkage into global list of all client environments. Used for
          * garbage collection.
          */
-        cfs_list_t        ce_linkage;
+       struct list_head  ce_linkage;
         /*
          *
          */
@@ -577,11 +653,11 @@ 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)
@@ -617,14 +693,15 @@ static unsigned cl_env_hops_hash(cfs_hash_t *lh,
 #endif
 }
 
-static void *cl_env_hops_obj(cfs_hlist_node_t *hn)
+static void *cl_env_hops_obj(struct hlist_node *hn)
 {
-        struct cl_env *cle = cfs_hlist_entry(hn, struct cl_env, ce_node);
-        LASSERT(cle->ce_magic == &cl_env_init0);
-        return (void *)cle;
+       struct cl_env *cle = 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)
+static int cl_env_hops_keycmp(const void *key, struct hlist_node *hn)
 {
         struct cl_env *cle = cl_env_hops_obj(hn);
 
@@ -632,9 +709,9 @@ static int cl_env_hops_keycmp(const void *key, cfs_hlist_node_t *hn)
         return (key == cle->ce_owner);
 }
 
-static void cl_env_hops_noop(cfs_hash_t *hs, cfs_hlist_node_t *hn)
+static void cl_env_hops_noop(cfs_hash_t *hs, struct hlist_node *hn)
 {
-        struct cl_env *cle = cfs_hlist_entry(hn, struct cl_env, ce_node);
+       struct cl_env *cle = hlist_entry(hn, struct cl_env, ce_node);
         LASSERT(cle->ce_magic == &cl_env_init0);
 }
 
@@ -746,37 +823,37 @@ 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;
+       struct lu_env *env;
+       struct cl_env *cle;
 
-       OBD_SLAB_ALLOC_PTR_GFP(cle, cl_env_kmem, __GFP_IO);
-        if (cle != NULL) {
-                int rc;
+       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;
+               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 +864,40 @@ 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, 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(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);
@@ -812,12 +923,13 @@ struct lu_env *cl_env_peek(int *refcheck)
         CDEBUG(D_OTHER, "%d@%p\n", cle ? cle->ce_ref : 0, cle);
         return env;
 }
-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 +943,7 @@ 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 +987,33 @@ 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 (; !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
@@ -899,7 +1035,21 @@ 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);
+                       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);
@@ -1007,7 +1157,6 @@ void cl_attr2lvb(struct ost_lvb *lvb, const struct cl_attr *attr)
         lvb->lvb_blocks = attr->cat_blocks;
         EXIT;
 }
-EXPORT_SYMBOL(cl_attr2lvb);
 
 /**
  * Converts struct ost_lvb to struct cl_attr.
@@ -1026,6 +1175,103 @@ 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;
+
+               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.
@@ -1073,12 +1319,6 @@ void cl_stack_fini(const struct lu_env *env, struct cl_device *cl)
 }
 EXPORT_SYMBOL(cl_stack_fini);
 
-int  cl_lock_init(void);
-void cl_lock_fini(void);
-
-int  cl_page_init(void);
-void cl_page_fini(void);
-
 static struct lu_context_key cl_key;
 
 struct cl_thread_info *cl_env_info(const struct lu_env *env)
@@ -1096,7 +1336,7 @@ static void *cl_key_init(const struct lu_context *ctx,
 
         info = cl0_key_init(ctx, key);
         if (!IS_ERR(info)) {
-                int i;
+               size_t i;
 
                 for (i = 0; i < ARRAY_SIZE(info->clt_counters); ++i)
                         lu_ref_init(&info->clt_counters[i].ctc_locks_locked);
@@ -1107,8 +1347,8 @@ static void *cl_key_init(const struct lu_context *ctx,
 static void cl_key_fini(const struct lu_context *ctx,
                         struct lu_context_key *key, void *data)
 {
-        struct cl_thread_info *info;
-        int i;
+       struct cl_thread_info *info;
+       size_t i;
 
         info = data;
         for (i = 0; i < ARRAY_SIZE(info->clt_counters); ++i)
@@ -1119,8 +1359,8 @@ static void cl_key_fini(const struct lu_context *ctx,
 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;
+       struct cl_thread_info *info = data;
+       size_t i;
 
         for (i = 0; i < ARRAY_SIZE(info->clt_counters); ++i) {
                 LASSERT(info->clt_counters[i].ctc_nr_held == 0);
@@ -1158,11 +1398,13 @@ static struct lu_kmem_descr cl_object_caches[] = {
  */
 int cl_global_init(void)
 {
-        int result;
+       int result;
 
-        result = cl_env_store_init();
-        if (result)
-                return result;
+       INIT_LIST_HEAD(&cl_envs);
+
+       result = cl_env_store_init();
+       if (result)
+               return result;
 
         result = lu_kmem_init(cl_object_caches);
         if (result)
@@ -1173,17 +1415,13 @@ int cl_global_init(void)
         if (result)
                 goto out_kmem;
 
-        result = cl_lock_init();
-        if (result)
-                goto out_context;
-
-        result = cl_page_init();
-        if (result)
-                goto out_lock;
+       result = cl_env_percpu_init();
+       if (result)
+               /* no cl_env_percpu_fini on error */
+               goto out_context;
 
         return 0;
-out_lock:
-        cl_lock_fini();
+
 out_context:
         lu_context_key_degister(&cl_key);
 out_kmem:
@@ -1198,8 +1436,7 @@ out_store:
  */
 void cl_global_fini(void)
 {
-        cl_lock_fini();
-        cl_page_fini();
+       cl_env_percpu_fini();
         lu_context_key_degister(&cl_key);
         lu_kmem_fini(cl_object_caches);
         cl_env_store_fini();