LU-15642 obdclass: use consistent stats units

[fs/lustre-release.git] / lustre / obdclass / lu_object.c

/*
 * GPL HEADER START
 *
 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License version 2 only,
 * as published by the Free Software Foundation.
 *
 * This program is distributed in the hope that it will be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * General Public License version 2 for more details (a copy is included
 * in the LICENSE file that accompanied this code).
 *
 * You should have received a copy of the GNU General Public License
 * version 2 along with this program; If not, see
 * http://www.gnu.org/licenses/gpl-2.0.html
 *
 * GPL HEADER END
 */
/*
 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
 * Use is subject to license terms.
 *
 * Copyright (c) 2011, 2017, Intel Corporation.
 */
/*
 * This file is part of Lustre, http://www.lustre.org/
 *
 * lustre/obdclass/lu_object.c
 *
 * Lustre Object.
 * These are the only exported functions, they provide some generic
 * infrastructure for managing object devices
 *
 *   Author: Nikita Danilov <nikita.danilov@sun.com>
 */

#define DEBUG_SUBSYSTEM S_CLASS

#include <linux/delay.h>
#include <linux/module.h>
#include <linux/list.h>
#include <linux/processor.h>
#include <linux/random.h>

#include <libcfs/libcfs.h>
#include <libcfs/linux/linux-mem.h>
#include <libcfs/linux/linux-hash.h>
#include <obd_class.h>
#include <obd_support.h>
#include <lustre_disk.h>
#include <lustre_fid.h>
#include <lu_object.h>
#include <lu_ref.h>

struct lu_site_bkt_data {
        /**
         * LRU list, updated on each access to object. Protected by
         * lsb_waitq.lock.
         *
         * "Cold" end of LRU is lu_site::ls_lru.next. Accessed object are
         * moved to the lu_site::ls_lru.prev
         */
        struct list_head                lsb_lru;
        /**
         * Wait-queue signaled when an object in this site is ultimately
         * destroyed (lu_object_free()) or initialized (lu_object_start()).
         * It is used by lu_object_find() to wait before re-trying when
         * object in the process of destruction is found in the hash table;
         * or wait object to be initialized by the allocator.
         *
         * \see htable_lookup().
         */
        wait_queue_head_t               lsb_waitq;
};

enum {
        LU_CACHE_PERCENT_MAX     = 50,
        LU_CACHE_PERCENT_DEFAULT = 20
};

#define LU_CACHE_NR_MAX_ADJUST          512
#define LU_CACHE_NR_UNLIMITED           -1
#define LU_CACHE_NR_DEFAULT             LU_CACHE_NR_UNLIMITED
/** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
#define LU_CACHE_NR_ZFS_LIMIT           10240

#define LU_CACHE_NR_MIN                 4096
#define LU_CACHE_NR_MAX                 0x80000000UL

/**
 * Max 256 buckets, we don't want too many buckets because:
 * - consume too much memory (currently max 16K)
 * - avoid unbalanced LRU list
 * With few cpus there is little gain from extra buckets, so
 * we treat this as a maximum in lu_site_init().
 */
#define LU_SITE_BKT_BITS    8

static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
module_param(lu_cache_percent, int, 0644);
MODULE_PARM_DESC(lu_cache_percent, "Percentage of memory to be used as lu_object cache");

static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
module_param(lu_cache_nr, long, 0644);
MODULE_PARM_DESC(lu_cache_nr, "Maximum number of objects in lu_object cache");

static void lu_object_free(const struct lu_env *env, struct lu_object *o);
static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx);

static u32 lu_fid_hash(const void *data, u32 len, u32 seed)
{
        const struct lu_fid *fid = data;

        seed = cfs_hash_32(seed ^ fid->f_oid, 32);
        seed ^= cfs_hash_64(fid->f_seq, 32);
        return seed;
}

static const struct rhashtable_params obj_hash_params = {
        .key_len        = sizeof(struct lu_fid),
        .key_offset     = offsetof(struct lu_object_header, loh_fid),
        .head_offset    = offsetof(struct lu_object_header, loh_hash),
        .hashfn         = lu_fid_hash,
        .automatic_shrinking = true,
};

static inline int lu_bkt_hash(struct lu_site *s, const struct lu_fid *fid)
{
        return lu_fid_hash(fid, sizeof(*fid), s->ls_bkt_seed) &
               (s->ls_bkt_cnt - 1);
}

wait_queue_head_t *
lu_site_wq_from_fid(struct lu_site *site, struct lu_fid *fid)
{
        struct lu_site_bkt_data *bkt;

        bkt = &site->ls_bkts[lu_bkt_hash(site, fid)];
        return &bkt->lsb_waitq;
}
EXPORT_SYMBOL(lu_site_wq_from_fid);

/**
 * Decrease reference counter on object. If last reference is freed, return
 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
 * case, free object immediately.
 */
void lu_object_put(const struct lu_env *env, struct lu_object *o)
{
        struct lu_site_bkt_data *bkt;
        struct lu_object_header *top = o->lo_header;
        struct lu_site *site = o->lo_dev->ld_site;
        struct lu_object *orig = o;
        const struct lu_fid *fid = lu_object_fid(o);

        /*
         * till we have full fids-on-OST implemented anonymous objects
         * are possible in OSP. such an object isn't listed in the site
         * so we should not remove it from the site.
         */
        if (fid_is_zero(fid)) {
                LASSERT(list_empty(&top->loh_lru));
                if (!atomic_dec_and_test(&top->loh_ref))
                        return;
                list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
                        if (o->lo_ops->loo_object_release != NULL)
                                o->lo_ops->loo_object_release(env, o);
                }
                lu_object_free(env, orig);
                return;
        }

        bkt = &site->ls_bkts[lu_bkt_hash(site, &top->loh_fid)];
        if (atomic_add_unless(&top->loh_ref, -1, 1)) {
still_active:
                /*
                 * At this point the object reference is dropped and lock is
                 * not taken, so lu_object should not be touched because it
                 * can be freed by concurrent thread.
                 *
                 * Somebody may be waiting for this, currently only used for
                 * cl_object, see cl_object_put_last().
                 */
                wake_up(&bkt->lsb_waitq);

                return;
        }

        spin_lock(&bkt->lsb_waitq.lock);
        if (!atomic_dec_and_test(&top->loh_ref)) {
                spin_unlock(&bkt->lsb_waitq.lock);
                goto still_active;
        }

        /*
         * Refcount is zero, and cannot be incremented without taking the bkt
         * lock, so object is stable.
         */

        /*
         * When last reference is released, iterate over object layers, and
         * notify them that object is no longer busy.
         */
        list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
                if (o->lo_ops->loo_object_release != NULL)
                        o->lo_ops->loo_object_release(env, o);
        }

        /*
         * Don't use local 'is_dying' here because if was taken without lock but
         * here we need the latest actual value of it so check lu_object
         * directly here.
         */
        if (!lu_object_is_dying(top) &&
            (lu_object_exists(orig) || lu_object_is_cl(orig))) {
                LASSERT(list_empty(&top->loh_lru));
                list_add_tail(&top->loh_lru, &bkt->lsb_lru);
                spin_unlock(&bkt->lsb_waitq.lock);
                percpu_counter_inc(&site->ls_lru_len_counter);
                CDEBUG(D_INODE, "Add %p/%p to site lru. bkt: %p\n",
                       orig, top, bkt);
                return;
        }

        /*
         * If object is dying (will not be cached) then remove it from hash
         * table (it is already not on the LRU).
         *
         * This is done with bucket lock held.  As the only way to acquire first
         * reference to previously unreferenced object is through hash-table
         * lookup (lu_object_find()) which takes the lock for first reference,
         * no race with concurrent object lookup is possible and we can safely
         * destroy object below.
         */
        if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
                rhashtable_remove_fast(&site->ls_obj_hash, &top->loh_hash,
                                       obj_hash_params);

        spin_unlock(&bkt->lsb_waitq.lock);
        /* Object was already removed from hash above, can kill it. */
        lu_object_free(env, orig);
}
EXPORT_SYMBOL(lu_object_put);

/**
 * Put object and don't keep in cache. This is temporary solution for
 * multi-site objects when its layering is not constant.
 */
void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
{
        set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
        return lu_object_put(env, o);
}
EXPORT_SYMBOL(lu_object_put_nocache);

/**
 * Kill the object and take it out of LRU cache.
 * Currently used by client code for layout change.
 */
void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
{
        struct lu_object_header *top;

        top = o->lo_header;
        set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
        if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
                struct lu_site *site = o->lo_dev->ld_site;
                struct rhashtable *obj_hash = &site->ls_obj_hash;
                struct lu_site_bkt_data *bkt;

                bkt = &site->ls_bkts[lu_bkt_hash(site, &top->loh_fid)];
                spin_lock(&bkt->lsb_waitq.lock);
                if (!list_empty(&top->loh_lru)) {
                        list_del_init(&top->loh_lru);
                        percpu_counter_dec(&site->ls_lru_len_counter);
                }
                spin_unlock(&bkt->lsb_waitq.lock);

                rhashtable_remove_fast(obj_hash, &top->loh_hash,
                                       obj_hash_params);
        }
}
EXPORT_SYMBOL(lu_object_unhash);

/**
 * Allocate new object.
 *
 * This follows object creation protocol, described in the comment within
 * struct lu_device_operations definition.
 */
static struct lu_object *lu_object_alloc(const struct lu_env *env,
                                         struct lu_device *dev,
                                         const struct lu_fid *f)
{
        struct lu_object *top;

        /*
         * Create top-level object slice. This will also create
         * lu_object_header.
         */
        top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
        if (top == NULL)
                return ERR_PTR(-ENOMEM);
        if (IS_ERR(top))
                return top;
        /*
         * This is the only place where object fid is assigned. It's constant
         * after this point.
         */
        top->lo_header->loh_fid = *f;

        return top;
}

/**
 * Initialize object.
 *
 * This is called after object hash insertion to avoid returning an object with
 * stale attributes.
 */
static int lu_object_start(const struct lu_env *env, struct lu_device *dev,
                           struct lu_object *top,
                           const struct lu_object_conf *conf)
{
        struct lu_object *scan;
        struct list_head *layers;
        unsigned int init_mask = 0;
        unsigned int init_flag;
        int clean;
        int result;

        layers = &top->lo_header->loh_layers;

        do {
                /*
                 * Call ->loo_object_init() repeatedly, until no more new
                 * object slices are created.
                 */
                clean = 1;
                init_flag = 1;
                list_for_each_entry(scan, layers, lo_linkage) {
                        if (init_mask & init_flag)
                                goto next;
                        clean = 0;
                        scan->lo_header = top->lo_header;
                        result = scan->lo_ops->loo_object_init(env, scan, conf);
                        if (result)
                                return result;

                        init_mask |= init_flag;
next:
                        init_flag <<= 1;
                }
        } while (!clean);

        list_for_each_entry_reverse(scan, layers, lo_linkage) {
                if (scan->lo_ops->loo_object_start != NULL) {
                        result = scan->lo_ops->loo_object_start(env, scan);
                        if (result)
                                return result;
                }
        }

        lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);

        set_bit(LU_OBJECT_INITED, &top->lo_header->loh_flags);

        return 0;
}

/**
 * Free an object.
 */
static void lu_object_free(const struct lu_env *env, struct lu_object *o)
{
        wait_queue_head_t *wq;
        struct lu_site *site;
        struct lu_object *scan;
        struct list_head *layers;
        LIST_HEAD(splice);

        site = o->lo_dev->ld_site;
        layers = &o->lo_header->loh_layers;
        wq = lu_site_wq_from_fid(site, &o->lo_header->loh_fid);
        /*
         * First call ->loo_object_delete() method to release all resources.
         */
        list_for_each_entry_reverse(scan, layers, lo_linkage) {
                if (scan->lo_ops->loo_object_delete != NULL)
                        scan->lo_ops->loo_object_delete(env, scan);
        }

        /*
         * Then, splice object layers into stand-alone list, and call
         * ->loo_object_free() on all layers to free memory. Splice is
         * necessary, because lu_object_header is freed together with the
         * top-level slice.
         */
        list_splice_init(layers, &splice);
        while (!list_empty(&splice)) {
                /*
                 * Free layers in bottom-to-top order, so that object header
                 * lives as long as possible and ->loo_object_free() methods
                 * can look at its contents.
                 */
                o = container_of(splice.prev, struct lu_object, lo_linkage);
                list_del_init(&o->lo_linkage);
                LASSERT(o->lo_ops->loo_object_free != NULL);
                o->lo_ops->loo_object_free(env, o);
        }

        if (waitqueue_active(wq))
                wake_up(wq);
}

/**
 * Free \a nr objects from the cold end of the site LRU list.
 * if canblock is 0, then don't block awaiting for another
 * instance of lu_site_purge() to complete
 */
int lu_site_purge_objects(const struct lu_env *env, struct lu_site *s,
                          int nr, int canblock)
{
        struct lu_object_header *h;
        struct lu_object_header *temp;
        struct lu_site_bkt_data *bkt;
        LIST_HEAD(dispose);
        int                      did_sth;
        unsigned int             start = 0;
        int                      count;
        int                      bnr;
        unsigned int             i;

        if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
                RETURN(0);

        /*
         * Under LRU list lock, scan LRU list and move unreferenced objects to
         * the dispose list, removing them from LRU and hash table.
         */
        if (nr != ~0)
                start = s->ls_purge_start;
        bnr = (nr == ~0) ? -1 : nr / s->ls_bkt_cnt + 1;
again:
        /*
         * It doesn't make any sense to make purge threads parallel, that can
         * only bring troubles to us.  See LU-5331.
         */
        if (canblock != 0)
                mutex_lock(&s->ls_purge_mutex);
        else if (mutex_trylock(&s->ls_purge_mutex) == 0)
                goto out;

        did_sth = 0;
        for (i = start; i < s->ls_bkt_cnt ; i++) {
                count = bnr;
                bkt = &s->ls_bkts[i];
                spin_lock(&bkt->lsb_waitq.lock);

                list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
                        LASSERT(atomic_read(&h->loh_ref) == 0);

                        LINVRNT(lu_bkt_hash(s, &h->loh_fid) == i);

                        set_bit(LU_OBJECT_UNHASHED, &h->loh_flags);
                        rhashtable_remove_fast(&s->ls_obj_hash, &h->loh_hash,
                                               obj_hash_params);
                        list_move(&h->loh_lru, &dispose);
                        percpu_counter_dec(&s->ls_lru_len_counter);
                        if (did_sth == 0)
                                did_sth = 1;

                        if (nr != ~0 && --nr == 0)
                                break;

                        if (count > 0 && --count == 0)
                                break;

                }
                spin_unlock(&bkt->lsb_waitq.lock);
                cond_resched();
                /*
                 * Free everything on the dispose list. This is safe against
                 * races due to the reasons described in lu_object_put().
                 */
                while ((h = list_first_entry_or_null(&dispose,
                                                     struct lu_object_header,
                                                     loh_lru)) != NULL) {
                        list_del_init(&h->loh_lru);
                        lu_object_free(env, lu_object_top(h));
                        lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
                }

                if (nr == 0)
                        break;
        }
        mutex_unlock(&s->ls_purge_mutex);

        if (nr != 0 && did_sth && start != 0) {
                start = 0; /* restart from the first bucket */
                goto again;
        }
        /* race on s->ls_purge_start, but nobody cares */
        s->ls_purge_start = i & (s->ls_bkt_cnt - 1);
out:
        return nr;
}
EXPORT_SYMBOL(lu_site_purge_objects);

/*
 * Object printing.
 *
 * Code below has to jump through certain loops to output object description
 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
 * composes object description from strings that are parts of _lines_ of
 * output (i.e., strings that are not terminated by newline). This doesn't fit
 * very well into libcfs_debug_msg() interface that assumes that each message
 * supplied to it is a self-contained output line.
 *
 * To work around this, strings are collected in a temporary buffer
 * (implemented as a value of lu_cdebug_key key), until terminating newline
 * character is detected.
 *
 */

enum {
        /**
         * Maximal line size.
         *
         * XXX overflow is not handled correctly.
         */
        LU_CDEBUG_LINE = 512
};

struct lu_cdebug_data {
        /**
         * Temporary buffer.
         */
        char lck_area[LU_CDEBUG_LINE];
};

/* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);

/**
 * Key, holding temporary buffer. This key is registered very early by
 * lu_global_init().
 */
static struct lu_context_key lu_global_key = {
        .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
                    LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
        .lct_init = lu_global_key_init,
        .lct_fini = lu_global_key_fini
};

/**
 * Printer function emitting messages through libcfs_debug_msg().
 */
int lu_cdebug_printer(const struct lu_env *env,
                      void *cookie, const char *format, ...)
{
        struct libcfs_debug_msg_data *msgdata = cookie;
        struct lu_cdebug_data        *key;
        int used;
        int complete;
        va_list args;

        va_start(args, format);

        key = lu_context_key_get(&env->le_ctx, &lu_global_key);
        LASSERT(key != NULL);

        used = strlen(key->lck_area);
        complete = format[strlen(format) - 1] == '\n';
        /*
         * Append new chunk to the buffer.
         */
        vsnprintf(key->lck_area + used,
                  ARRAY_SIZE(key->lck_area) - used, format, args);
        if (complete) {
                if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
                        libcfs_debug_msg(msgdata, "%s\n", key->lck_area);
                key->lck_area[0] = 0;
        }
        va_end(args);
        return 0;
}
EXPORT_SYMBOL(lu_cdebug_printer);

/**
 * Print object header.
 */
void lu_object_header_print(const struct lu_env *env, void *cookie,
                            lu_printer_t printer,
                            const struct lu_object_header *hdr)
{
        (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
                   hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
                   PFID(&hdr->loh_fid),
                   test_bit(LU_OBJECT_UNHASHED,
                            &hdr->loh_flags) ? "" : " hash",
                   list_empty(&hdr->loh_lru) ? "" : " lru",
                   hdr->loh_attr & LOHA_EXISTS ? " exist" : "");
}
EXPORT_SYMBOL(lu_object_header_print);

/**
 * Print human readable representation of the \a o to the \a printer.
 */
void lu_object_print(const struct lu_env *env, void *cookie,
                     lu_printer_t printer, const struct lu_object *o)
{
        static const char ruler[] = "........................................";
        struct lu_object_header *top;
        int depth = 4;

        top = o->lo_header;
        lu_object_header_print(env, cookie, printer, top);
        (*printer)(env, cookie, "{\n");

        list_for_each_entry(o, &top->loh_layers, lo_linkage) {
                /*
                 * print `.' \a depth times followed by type name and address
                 */
                (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
                           o->lo_dev->ld_type->ldt_name, o);

                if (o->lo_ops->loo_object_print != NULL)
                        (*o->lo_ops->loo_object_print)(env, cookie, printer, o);

                (*printer)(env, cookie, "\n");
        }

        (*printer)(env, cookie, "} header@%p\n", top);
}
EXPORT_SYMBOL(lu_object_print);

/**
 * Check object consistency.
 */
int lu_object_invariant(const struct lu_object *o)
{
        struct lu_object_header *top;

        top = o->lo_header;
        list_for_each_entry(o, &top->loh_layers, lo_linkage) {
                if (o->lo_ops->loo_object_invariant != NULL &&
                    !o->lo_ops->loo_object_invariant(o))
                        return 0;
        }
        return 1;
}

/*
 * Limit the lu_object cache to a maximum of lu_cache_nr objects.  Because the
 * calculation for the number of objects to reclaim is not covered by a lock the
 * maximum number of objects is capped by LU_CACHE_MAX_ADJUST.  This ensures
 * that many concurrent threads will not accidentally purge the entire cache.
 */
static void lu_object_limit(const struct lu_env *env,
                            struct lu_device *dev)
{
        u64 size, nr;

        if (lu_cache_nr == LU_CACHE_NR_UNLIMITED)
                return;

        size = atomic_read(&dev->ld_site->ls_obj_hash.nelems);
        nr = (u64)lu_cache_nr;
        if (size <= nr)
                return;

        lu_site_purge_objects(env, dev->ld_site,
                              min_t(u64, size - nr, LU_CACHE_NR_MAX_ADJUST),
                              0);
}

static struct lu_object *htable_lookup(const struct lu_env *env,
                                       struct lu_device *dev,
                                       struct lu_site_bkt_data *bkt,
                                       const struct lu_fid *f,
                                       struct lu_object_header *new)
{
        struct lu_site *s = dev->ld_site;
        struct lu_object_header *h;

try_again:
        rcu_read_lock();
        if (new)
                h = rhashtable_lookup_get_insert_fast(&s->ls_obj_hash,
                                                      &new->loh_hash,
                                                      obj_hash_params);
        else
                h = rhashtable_lookup(&s->ls_obj_hash, f, obj_hash_params);

        if (IS_ERR_OR_NULL(h)) {
                /* Not found */
                if (!new)
                        lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
                rcu_read_unlock();
                if (PTR_ERR(h) == -ENOMEM) {
                        msleep(20);
                        goto try_again;
                }
                lu_object_limit(env, dev);
                if (PTR_ERR(h) == -E2BIG)
                        goto try_again;

                return ERR_PTR(-ENOENT);
        }

        if (atomic_inc_not_zero(&h->loh_ref)) {
                rcu_read_unlock();
                return lu_object_top(h);
        }

        spin_lock(&bkt->lsb_waitq.lock);
        if (lu_object_is_dying(h) ||
            test_bit(LU_OBJECT_UNHASHED, &h->loh_flags)) {
                spin_unlock(&bkt->lsb_waitq.lock);
                rcu_read_unlock();
                if (new) {
                        /*
                         * Old object might have already been removed, or will
                         * be soon.  We need to insert our new object, so
                         * remove the old one just in case it is still there.
                         */
                        rhashtable_remove_fast(&s->ls_obj_hash, &h->loh_hash,
                                               obj_hash_params);
                        goto try_again;
                }
                lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
                return ERR_PTR(-ENOENT);
        }
        /* Now protected by spinlock */
        rcu_read_unlock();

        if (!list_empty(&h->loh_lru)) {
                list_del_init(&h->loh_lru);
                percpu_counter_dec(&s->ls_lru_len_counter);
        }
        atomic_inc(&h->loh_ref);
        spin_unlock(&bkt->lsb_waitq.lock);
        lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
        return lu_object_top(h);
}

/**
 * Search cache for an object with the fid \a f. If such object is found,
 * return it. Otherwise, create new object, insert it into cache and return
 * it. In any case, additional reference is acquired on the returned object.
 */
struct lu_object *lu_object_find(const struct lu_env *env,
                                 struct lu_device *dev, const struct lu_fid *f,
                                 const struct lu_object_conf *conf)
{
        return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
}
EXPORT_SYMBOL(lu_object_find);

/*
 * Get a 'first' reference to an object that was found while looking through the
 * hash table.
 */
struct lu_object *lu_object_get_first(struct lu_object_header *h,
                                      struct lu_device *dev)
{
        struct lu_site *s = dev->ld_site;
        struct lu_object *ret;

        if (IS_ERR_OR_NULL(h) || lu_object_is_dying(h))
                return NULL;

        ret = lu_object_locate(h, dev->ld_type);
        if (!ret)
                return ret;

        if (!atomic_inc_not_zero(&h->loh_ref)) {
                struct lu_site_bkt_data *bkt;

                bkt = &s->ls_bkts[lu_bkt_hash(s, &h->loh_fid)];
                spin_lock(&bkt->lsb_waitq.lock);
                if (!lu_object_is_dying(h) &&
                    !test_bit(LU_OBJECT_UNHASHED, &h->loh_flags))
                        atomic_inc(&h->loh_ref);
                else
                        ret = NULL;
                spin_unlock(&bkt->lsb_waitq.lock);
        }
        return ret;
}
EXPORT_SYMBOL(lu_object_get_first);

/**
 * Core logic of lu_object_find*() functions.
 *
 * Much like lu_object_find(), but top level device of object is specifically
 * \a dev rather than top level device of the site. This interface allows
 * objects of different "stacking" to be created within the same site.
 */
struct lu_object *lu_object_find_at(const struct lu_env *env,
                                    struct lu_device *dev,
                                    const struct lu_fid *f,
                                    const struct lu_object_conf *conf)
{
        struct lu_object *o;
        struct lu_object *shadow;
        struct lu_site *s;
        struct lu_site_bkt_data *bkt;
        struct rhashtable *hs;
        int rc;

        ENTRY;

        /* FID is from disk or network, zero FID is meaningless, return error
         * early to avoid assertion in lu_object_put. If a zero FID is wanted,
         * it should be allocated via lu_object_anon().
         */
        if (fid_is_zero(f))
                RETURN(ERR_PTR(-EINVAL));

        /*
         * This uses standard index maintenance protocol:
         *
         *     - search index under lock, and return object if found;
         *     - otherwise, unlock index, allocate new object;
         *     - lock index and search again;
         *     - if nothing is found (usual case), insert newly created
         *       object into index;
         *     - otherwise (race: other thread inserted object), free
         *       object just allocated.
         *     - unlock index;
         *     - return object.
         *
         * For "LOC_F_NEW" case, we are sure the object is new established.
         * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
         * just alloc and insert directly.
         *
         */
        s  = dev->ld_site;
        hs = &s->ls_obj_hash;

        if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_OBD_ZERO_NLINK_RACE)))
                lu_site_purge(env, s, -1);

        bkt = &s->ls_bkts[lu_bkt_hash(s, f)];
        if (!(conf && conf->loc_flags & LOC_F_NEW)) {
                o = htable_lookup(env, dev, bkt, f, NULL);

                if (!IS_ERR(o)) {
                        if (likely(lu_object_is_inited(o->lo_header)))
                                RETURN(o);

                        wait_event_idle(bkt->lsb_waitq,
                                        lu_object_is_inited(o->lo_header) ||
                                        lu_object_is_dying(o->lo_header));

                        if (lu_object_is_dying(o->lo_header)) {
                                lu_object_put(env, o);

                                RETURN(ERR_PTR(-ENOENT));
                        }

                        RETURN(o);
                }

                if (PTR_ERR(o) != -ENOENT)
                        RETURN(o);
        }

        /*
         * Allocate new object, NB, object is unitialized in case object
         * is changed between allocation and hash insertion, thus the object
         * with stale attributes is returned.
         */
        o = lu_object_alloc(env, dev, f);
        if (IS_ERR(o))
                RETURN(o);

        LASSERT(lu_fid_eq(lu_object_fid(o), f));

        CFS_RACE_WAIT(OBD_FAIL_OBD_ZERO_NLINK_RACE);

        if (conf && conf->loc_flags & LOC_F_NEW) {
                int status = rhashtable_insert_fast(hs, &o->lo_header->loh_hash,
                                                    obj_hash_params);
                if (status)
                        /* Strange error - go the slow way */
                        shadow = htable_lookup(env, dev, bkt, f, o->lo_header);
                else
                        shadow = ERR_PTR(-ENOENT);
        } else {
                shadow = htable_lookup(env, dev, bkt, f, o->lo_header);
        }
        if (likely(PTR_ERR(shadow) == -ENOENT)) {
                /*
                 * The new object has been successfully inserted.
                 *
                 * This may result in rather complicated operations, including
                 * fld queries, inode loading, etc.
                 */
                rc = lu_object_start(env, dev, o, conf);
                if (rc) {
                        lu_object_put_nocache(env, o);
                        RETURN(ERR_PTR(rc));
                }

                wake_up(&bkt->lsb_waitq);

                lu_object_limit(env, dev);

                RETURN(o);
        }

        lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
        lu_object_free(env, o);

        if (!(conf && conf->loc_flags & LOC_F_NEW) &&
            !IS_ERR(shadow) &&
            !lu_object_is_inited(shadow->lo_header)) {
                wait_event_idle(bkt->lsb_waitq,
                                lu_object_is_inited(shadow->lo_header) ||
                                lu_object_is_dying(shadow->lo_header));

                if (lu_object_is_dying(shadow->lo_header)) {
                        lu_object_put(env, shadow);

                        RETURN(ERR_PTR(-ENOENT));
                }
        }

        RETURN(shadow);
}
EXPORT_SYMBOL(lu_object_find_at);

/**
 * Find object with given fid, and return its slice belonging to given device.
 */
struct lu_object *lu_object_find_slice(const struct lu_env *env,
                                       struct lu_device *dev,
                                       const struct lu_fid *f,
                                       const struct lu_object_conf *conf)
{
        struct lu_object *top;
        struct lu_object *obj;

        top = lu_object_find(env, dev, f, conf);
        if (IS_ERR(top))
                return top;

        obj = lu_object_locate(top->lo_header, dev->ld_type);
        if (unlikely(obj == NULL)) {
                lu_object_put(env, top);
                obj = ERR_PTR(-ENOENT);
        }

        return obj;
}
EXPORT_SYMBOL(lu_object_find_slice);

int lu_device_type_init(struct lu_device_type *ldt)
{
        int result = 0;

        atomic_set(&ldt->ldt_device_nr, 0);
        if (ldt->ldt_ops->ldto_init)
                result = ldt->ldt_ops->ldto_init(ldt);

        return result;
}
EXPORT_SYMBOL(lu_device_type_init);

void lu_device_type_fini(struct lu_device_type *ldt)
{
        if (ldt->ldt_ops->ldto_fini)
                ldt->ldt_ops->ldto_fini(ldt);
}
EXPORT_SYMBOL(lu_device_type_fini);

/**
 * Global list of all sites on this node
 */
static LIST_HEAD(lu_sites);
static DECLARE_RWSEM(lu_sites_guard);

/**
 * Global environment used by site shrinker.
 */
static struct lu_env lu_shrink_env;

struct lu_site_print_arg {
        struct lu_env   *lsp_env;
        void            *lsp_cookie;
        lu_printer_t     lsp_printer;
};

static void
lu_site_obj_print(struct lu_object_header *h, struct lu_site_print_arg *arg)
{
        if (!list_empty(&h->loh_layers)) {
                const struct lu_object *o;

                o = lu_object_top(h);
                lu_object_print(arg->lsp_env, arg->lsp_cookie,
                                arg->lsp_printer, o);
        } else {
                lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
                                       arg->lsp_printer, h);
        }
}

/**
 * Print all objects in \a s.
 */
void lu_site_print(const struct lu_env *env, struct lu_site *s, atomic_t *ref,
                   int msg_flag, lu_printer_t printer)
{
        struct lu_site_print_arg arg = {
                .lsp_env     = (struct lu_env *)env,
                .lsp_printer = printer,
        };
        struct rhashtable_iter iter;
        struct lu_object_header *h;
        LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, msg_flag, NULL);

        if (!s || !atomic_read(ref))
                return;

        arg.lsp_cookie = (void *)&msgdata;

        rhashtable_walk_enter(&s->ls_obj_hash, &iter);
        rhashtable_walk_start(&iter);
        while ((h = rhashtable_walk_next(&iter)) != NULL) {
                if (IS_ERR(h))
                        continue;
                lu_site_obj_print(h, &arg);
        }
        rhashtable_walk_stop(&iter);
        rhashtable_walk_exit(&iter);
}
EXPORT_SYMBOL(lu_site_print);

/**
 * Return desired hash table order.
 */
static void lu_htable_limits(struct lu_device *top)
{
        unsigned long cache_size;

        /*
         * For ZFS based OSDs the cache should be disabled by default.  This
         * allows the ZFS ARC maximum flexibility in determining what buffers
         * to cache.  If Lustre has objects or buffer which it wants to ensure
         * always stay cached it must maintain a hold on them.
         */
        if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
                lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
                return;
        }

        /*
         * Calculate hash table size, assuming that we want reasonable
         * performance when 20% of total memory is occupied by cache of
         * lu_objects.
         *
         * Size of lu_object is (arbitrary) taken as 1K (together with inode).
         */
        cache_size = cfs_totalram_pages();

#if BITS_PER_LONG == 32
        /* limit hashtable size for lowmem systems to low RAM */
        if (cache_size > 1 << (30 - PAGE_SHIFT))
                cache_size = 1 << (30 - PAGE_SHIFT) * 3 / 4;
#endif

        /* clear off unreasonable cache setting. */
        if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
                CWARN("obdclass: invalid lu_cache_percent: %u, it must be in the range of (0, %u]. Will use default value: %u.\n",
                      lu_cache_percent, LU_CACHE_PERCENT_MAX,
                      LU_CACHE_PERCENT_DEFAULT);

                lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
        }
        cache_size = cache_size / 100 * lu_cache_percent *
                (PAGE_SIZE / 1024);

        lu_cache_nr = clamp_t(typeof(cache_size), cache_size,
                              LU_CACHE_NR_MIN, LU_CACHE_NR_MAX);
}

void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
{
        spin_lock(&s->ls_ld_lock);
        if (list_empty(&d->ld_linkage))
                list_add(&d->ld_linkage, &s->ls_ld_linkage);
        spin_unlock(&s->ls_ld_lock);
}
EXPORT_SYMBOL(lu_dev_add_linkage);

void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
{
        spin_lock(&s->ls_ld_lock);
        list_del_init(&d->ld_linkage);
        spin_unlock(&s->ls_ld_lock);
}
EXPORT_SYMBOL(lu_dev_del_linkage);

/**
  * Initialize site \a s, with \a d as the top level device.
  */
int lu_site_init(struct lu_site *s, struct lu_device *top)
{
        struct lu_site_bkt_data *bkt;
        unsigned int i;
        int rc;
        ENTRY;

        memset(s, 0, sizeof *s);
        mutex_init(&s->ls_purge_mutex);
        lu_htable_limits(top);

#ifdef HAVE_PERCPU_COUNTER_INIT_GFP_FLAG
        rc = percpu_counter_init(&s->ls_lru_len_counter, 0, GFP_NOFS);
#else
        rc = percpu_counter_init(&s->ls_lru_len_counter, 0);
#endif
        if (rc)
                return -ENOMEM;

        if (rhashtable_init(&s->ls_obj_hash, &obj_hash_params) != 0) {
                CERROR("failed to create lu_site hash\n");
                return -ENOMEM;
        }

        s->ls_bkt_seed = prandom_u32();
        s->ls_bkt_cnt = max_t(long, 1 << LU_SITE_BKT_BITS,
                              2 * num_possible_cpus());
        s->ls_bkt_cnt = roundup_pow_of_two(s->ls_bkt_cnt);
        OBD_ALLOC_PTR_ARRAY_LARGE(s->ls_bkts, s->ls_bkt_cnt);
        if (!s->ls_bkts) {
                rhashtable_destroy(&s->ls_obj_hash);
                s->ls_bkts = NULL;
                return -ENOMEM;
        }

        for (i = 0; i < s->ls_bkt_cnt; i++) {
                bkt = &s->ls_bkts[i];
                INIT_LIST_HEAD(&bkt->lsb_lru);
                init_waitqueue_head(&bkt->lsb_waitq);
        }

        s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
        if (s->ls_stats == NULL) {
                OBD_FREE_PTR_ARRAY_LARGE(s->ls_bkts, s->ls_bkt_cnt);
                s->ls_bkts = NULL;
                rhashtable_destroy(&s->ls_obj_hash);
                return -ENOMEM;
        }

        lprocfs_counter_init(s->ls_stats, LU_SS_CREATED, 0, "created");
        lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT, 0, "cache_hit");
        lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS, 0, "cache_miss");
        lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE, 0, "cache_race");
        lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
                             0, "cache_death_race");
        lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED, 0, "lru_purged");

        INIT_LIST_HEAD(&s->ls_linkage);
        s->ls_top_dev = top;
        top->ld_site = s;
        lu_device_get(top);
        lu_ref_add(&top->ld_reference, "site-top", s);

        INIT_LIST_HEAD(&s->ls_ld_linkage);
        spin_lock_init(&s->ls_ld_lock);

        lu_dev_add_linkage(s, top);

        RETURN(0);
}
EXPORT_SYMBOL(lu_site_init);

/**
 * Finalize \a s and release its resources.
 */
void lu_site_fini(struct lu_site *s)
{
        down_write(&lu_sites_guard);
        list_del_init(&s->ls_linkage);
        up_write(&lu_sites_guard);

        percpu_counter_destroy(&s->ls_lru_len_counter);

        if (s->ls_bkts) {
                rhashtable_destroy(&s->ls_obj_hash);
                OBD_FREE_PTR_ARRAY_LARGE(s->ls_bkts, s->ls_bkt_cnt);
                s->ls_bkts = NULL;
        }

        if (s->ls_top_dev != NULL) {
                s->ls_top_dev->ld_site = NULL;
                lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
                lu_device_put(s->ls_top_dev);
                s->ls_top_dev = NULL;
        }

        if (s->ls_stats != NULL)
                lprocfs_free_stats(&s->ls_stats);
}
EXPORT_SYMBOL(lu_site_fini);

/**
 * Called when initialization of stack for this site is completed.
 */
int lu_site_init_finish(struct lu_site *s)
{
        int result;
        down_write(&lu_sites_guard);
        result = lu_context_refill(&lu_shrink_env.le_ctx);
        if (result == 0)
                list_add(&s->ls_linkage, &lu_sites);
        up_write(&lu_sites_guard);
        return result;
}
EXPORT_SYMBOL(lu_site_init_finish);

/**
 * Acquire additional reference on device \a d
 */
void lu_device_get(struct lu_device *d)
{
        atomic_inc(&d->ld_ref);
}
EXPORT_SYMBOL(lu_device_get);

/**
 * Release reference on device \a d.
 */
void lu_device_put(struct lu_device *d)
{
        LASSERT(atomic_read(&d->ld_ref) > 0);
        atomic_dec(&d->ld_ref);
}
EXPORT_SYMBOL(lu_device_put);

enum { /* Maximal number of tld slots. */
        LU_CONTEXT_KEY_NR = 40
};
static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
static DECLARE_RWSEM(lu_key_initing);

/**
 * Initialize device \a d of type \a t.
 */
int lu_device_init(struct lu_device *d, struct lu_device_type *t)
{
        if (atomic_add_unless(&t->ldt_device_nr, 1, 0) == 0) {
                down_write(&lu_key_initing);
                if (t->ldt_ops->ldto_start &&
                    atomic_read(&t->ldt_device_nr) == 0)
                        t->ldt_ops->ldto_start(t);
                atomic_inc(&t->ldt_device_nr);
                up_write(&lu_key_initing);
        }

        memset(d, 0, sizeof *d);
        d->ld_type = t;
        lu_ref_init(&d->ld_reference);
        INIT_LIST_HEAD(&d->ld_linkage);

        return 0;
}
EXPORT_SYMBOL(lu_device_init);

/**
 * Finalize device \a d.
 */
void lu_device_fini(struct lu_device *d)
{
        struct lu_device_type *t = d->ld_type;

        if (d->ld_obd != NULL) {
                d->ld_obd->obd_lu_dev = NULL;
                d->ld_obd = NULL;
        }

        lu_ref_fini(&d->ld_reference);
        LASSERTF(atomic_read(&d->ld_ref) == 0,
                 "Refcount is %u\n", atomic_read(&d->ld_ref));
        LASSERT(atomic_read(&t->ldt_device_nr) > 0);

        if (atomic_dec_and_test(&t->ldt_device_nr) &&
            t->ldt_ops->ldto_stop != NULL)
                t->ldt_ops->ldto_stop(t);
}
EXPORT_SYMBOL(lu_device_fini);

/**
 * Initialize object \a o that is part of compound object \a h and was created
 * by device \a d.
 */
int lu_object_init(struct lu_object *o, struct lu_object_header *h,
                   struct lu_device *d)
{
        memset(o, 0, sizeof(*o));
        o->lo_header = h;
        o->lo_dev = d;
        lu_device_get(d);
        lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
        INIT_LIST_HEAD(&o->lo_linkage);

        return 0;
}
EXPORT_SYMBOL(lu_object_init);

/**
 * Finalize object and release its resources.
 */
void lu_object_fini(struct lu_object *o)
{
        struct lu_device *dev = o->lo_dev;

        LASSERT(list_empty(&o->lo_linkage));

        if (dev != NULL) {
                lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
                              "lu_object", o);
                lu_device_put(dev);
                o->lo_dev = NULL;
        }
}
EXPORT_SYMBOL(lu_object_fini);

/**
 * Add object \a o as first layer of compound object \a h
 *
 * This is typically called by the ->ldo_object_alloc() method of top-level
 * device.
 */
void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
{
        list_move(&o->lo_linkage, &h->loh_layers);
}
EXPORT_SYMBOL(lu_object_add_top);

/**
 * Add object \a o as a layer of compound object, going after \a before.
 *
 * This is typically called by the ->ldo_object_alloc() method of \a
 * before->lo_dev.
 */
void lu_object_add(struct lu_object *before, struct lu_object *o)
{
        list_move(&o->lo_linkage, &before->lo_linkage);
}
EXPORT_SYMBOL(lu_object_add);

/**
 * Initialize compound object.
 */
int lu_object_header_init(struct lu_object_header *h)
{
        memset(h, 0, sizeof *h);
        atomic_set(&h->loh_ref, 1);
        INIT_LIST_HEAD(&h->loh_lru);
        INIT_LIST_HEAD(&h->loh_layers);
        lu_ref_init(&h->loh_reference);
        return 0;
}
EXPORT_SYMBOL(lu_object_header_init);

/**
 * Finalize compound object.
 */
void lu_object_header_fini(struct lu_object_header *h)
{
        LASSERT(list_empty(&h->loh_layers));
        LASSERT(list_empty(&h->loh_lru));
        lu_ref_fini(&h->loh_reference);
}
EXPORT_SYMBOL(lu_object_header_fini);

/**
 * Free lu_object_header with proper RCU handling
 */
void lu_object_header_free(struct lu_object_header *h)
{
        lu_object_header_fini(h);
        OBD_FREE_PRE(h, sizeof(*h), "kfreed");
        kfree_rcu(h, loh_rcu);
}
EXPORT_SYMBOL(lu_object_header_free);

/**
 * Given a compound object, find its slice, corresponding to the device type
 * \a dtype.
 */
struct lu_object *lu_object_locate(struct lu_object_header *h,
                                   const struct lu_device_type *dtype)
{
        struct lu_object *o;

        list_for_each_entry(o, &h->loh_layers, lo_linkage) {
                if (o->lo_dev->ld_type == dtype)
                        return o;
        }
        return NULL;
}
EXPORT_SYMBOL(lu_object_locate);

/**
 * Finalize and free devices in the device stack.
 *
 * Finalize device stack by purging object cache, and calling
 * lu_device_type_operations::ldto_device_fini() and
 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
 */
void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
{
        struct lu_site   *site = top->ld_site;
        struct lu_device *scan;
        struct lu_device *next;

        lu_site_purge(env, site, ~0);
        for (scan = top; scan != NULL; scan = next) {
                next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
                lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
                lu_device_put(scan);
        }

        /* purge again. */
        lu_site_purge(env, site, ~0);

        for (scan = top; scan != NULL; scan = next) {
                const struct lu_device_type *ldt = scan->ld_type;

                next = ldt->ldt_ops->ldto_device_free(env, scan);
        }
}

/**
 * Global counter incremented whenever key is registered, unregistered,
 * revived or quiesced. This is used to void unnecessary calls to
 * lu_context_refill(). No locking is provided, as initialization and shutdown
 * are supposed to be externally serialized.
 */
static atomic_t key_set_version = ATOMIC_INIT(0);

/**
 * Register new key.
 */
int lu_context_key_register(struct lu_context_key *key)
{
        int result;
        unsigned int i;

        LASSERT(key->lct_init != NULL);
        LASSERT(key->lct_fini != NULL);
        LASSERT(key->lct_tags != 0);
        LASSERT(key->lct_owner != NULL);

        result = -ENFILE;
        atomic_set(&key->lct_used, 1);
        lu_ref_init(&key->lct_reference);
        for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
                if (lu_keys[i])
                        continue;
                key->lct_index = i;

                if (strncmp("osd_", module_name(key->lct_owner), 4) == 0)
                        CFS_RACE_WAIT(OBD_FAIL_OBD_SETUP);

                if (cmpxchg(&lu_keys[i], NULL, key) != NULL)
                        continue;

                result = 0;
                atomic_inc(&key_set_version);
                break;
        }
        if (result) {
                lu_ref_fini(&key->lct_reference);
                atomic_set(&key->lct_used, 0);
        }
        return result;
}
EXPORT_SYMBOL(lu_context_key_register);

static void key_fini(struct lu_context *ctx, int index)
{
        if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
                struct lu_context_key *key;

                key = lu_keys[index];
                LASSERT(key != NULL);
                LASSERT(key->lct_fini != NULL);
                LASSERT(atomic_read(&key->lct_used) > 0);

                key->lct_fini(ctx, key, ctx->lc_value[index]);
                lu_ref_del(&key->lct_reference, "ctx", ctx);
                if (atomic_dec_and_test(&key->lct_used))
                        wake_up_var(&key->lct_used);

                LASSERT(key->lct_owner != NULL);
                if ((ctx->lc_tags & LCT_NOREF) == 0) {
                        LINVRNT(module_refcount(key->lct_owner) > 0);
                        module_put(key->lct_owner);
                }
                ctx->lc_value[index] = NULL;
        }
}

/**
 * Deregister key.
 */
void lu_context_key_degister(struct lu_context_key *key)
{
        LASSERT(atomic_read(&key->lct_used) >= 1);
        LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));

        lu_context_key_quiesce(NULL, key);

        key_fini(&lu_shrink_env.le_ctx, key->lct_index);

        /**
         * Wait until all transient contexts referencing this key have
         * run lu_context_key::lct_fini() method.
         */
        atomic_dec(&key->lct_used);
        wait_var_event(&key->lct_used, atomic_read(&key->lct_used) == 0);

        if (!WARN_ON(lu_keys[key->lct_index] == NULL))
                lu_ref_fini(&key->lct_reference);

        smp_store_release(&lu_keys[key->lct_index], NULL);
}
EXPORT_SYMBOL(lu_context_key_degister);

/**
 * Register a number of keys. This has to be called after all keys have been
 * initialized by a call to LU_CONTEXT_KEY_INIT().
 */
int lu_context_key_register_many(struct lu_context_key *k, ...)
{
        struct lu_context_key *key = k;
        va_list args;
        int result;

        va_start(args, k);
        do {
                result = lu_context_key_register(key);
                if (result)
                        break;
                key = va_arg(args, struct lu_context_key *);
        } while (key != NULL);
        va_end(args);

        if (result != 0) {
                va_start(args, k);
                while (k != key) {
                        lu_context_key_degister(k);
                        k = va_arg(args, struct lu_context_key *);
                }
                va_end(args);
        }

        return result;
}
EXPORT_SYMBOL(lu_context_key_register_many);

/**
 * De-register a number of keys. This is a dual to
 * lu_context_key_register_many().
 */
void lu_context_key_degister_many(struct lu_context_key *k, ...)
{
        va_list args;

        va_start(args, k);
        do {
                lu_context_key_degister(k);
                k = va_arg(args, struct lu_context_key*);
        } while (k != NULL);
        va_end(args);
}
EXPORT_SYMBOL(lu_context_key_degister_many);

/**
 * Revive a number of keys.
 */
void lu_context_key_revive_many(struct lu_context_key *k, ...)
{
        va_list args;

        va_start(args, k);
        do {
                lu_context_key_revive(k);
                k = va_arg(args, struct lu_context_key*);
        } while (k != NULL);
        va_end(args);
}
EXPORT_SYMBOL(lu_context_key_revive_many);

/**
 * Quiescent a number of keys.
 */
void lu_context_key_quiesce_many(struct lu_device_type *t,
                                 struct lu_context_key *k, ...)
{
        va_list args;

        va_start(args, k);
        do {
                lu_context_key_quiesce(t, k);
                k = va_arg(args, struct lu_context_key*);
        } while (k != NULL);
        va_end(args);
}
EXPORT_SYMBOL(lu_context_key_quiesce_many);

/**
 * Return value associated with key \a key in context \a ctx.
 */
void *lu_context_key_get(const struct lu_context *ctx,
                         const struct lu_context_key *key)
{
        LINVRNT(ctx->lc_state == LCS_ENTERED);
        LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
        LASSERT(lu_keys[key->lct_index] == key);
        return ctx->lc_value[key->lct_index];
}
EXPORT_SYMBOL(lu_context_key_get);

/**
 * List of remembered contexts. XXX document me.
 */
static LIST_HEAD(lu_context_remembered);
static DEFINE_SPINLOCK(lu_context_remembered_guard);

/**
 * Destroy \a key in all remembered contexts. This is used to destroy key
 * values in "shared" contexts (like service threads), when a module owning
 * the key is about to be unloaded.
 */
void lu_context_key_quiesce(struct lu_device_type *t,
                            struct lu_context_key *key)
{
        struct lu_context *ctx;

        if (key->lct_tags & LCT_QUIESCENT)
                return;
        /*
         * The write-lock on lu_key_initing will ensure that any
         * keys_fill() which didn't see LCT_QUIESCENT will have
         * finished before we call key_fini().
         */
        down_write(&lu_key_initing);
        if (!(key->lct_tags & LCT_QUIESCENT)) {
                if (t == NULL || atomic_read(&t->ldt_device_nr) == 0)
                        key->lct_tags |= LCT_QUIESCENT;
                up_write(&lu_key_initing);

                spin_lock(&lu_context_remembered_guard);
                list_for_each_entry(ctx, &lu_context_remembered, lc_remember) {
                        spin_until_cond(READ_ONCE(ctx->lc_state) != LCS_LEAVING);
                        key_fini(ctx, key->lct_index);
                }
                spin_unlock(&lu_context_remembered_guard);

                return;
        }
        up_write(&lu_key_initing);
}

void lu_context_key_revive(struct lu_context_key *key)
{
        key->lct_tags &= ~LCT_QUIESCENT;
        atomic_inc(&key_set_version);
}

static void keys_fini(struct lu_context *ctx)
{
        unsigned int i;

        if (ctx->lc_value == NULL)
                return;

        for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
                key_fini(ctx, i);

        OBD_FREE_PTR_ARRAY(ctx->lc_value, ARRAY_SIZE(lu_keys));
        ctx->lc_value = NULL;
}

static int keys_fill(struct lu_context *ctx)
{
        unsigned int i;
        int rc = 0;

        /*
         * A serialisation with lu_context_key_quiesce() is needed, to
         * ensure we see LCT_QUIESCENT and don't allocate a new value
         * after it freed one.  The rwsem provides this.  As down_read()
         * does optimistic spinning while the writer is active, this is
         * unlikely to ever sleep.
         */
        down_read(&lu_key_initing);
        ctx->lc_version = atomic_read(&key_set_version);

        LINVRNT(ctx->lc_value);
        for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
                struct lu_context_key *key;

                key = lu_keys[i];
                if (!ctx->lc_value[i] && key &&
                    (key->lct_tags & ctx->lc_tags) &&
                    /*
                     * Don't create values for a LCT_QUIESCENT key, as this
                     * will pin module owning a key.
                     */
                    !(key->lct_tags & LCT_QUIESCENT)) {
                        void *value;

                        LINVRNT(key->lct_init != NULL);
                        LINVRNT(key->lct_index == i);

                        LASSERT(key->lct_owner != NULL);
                        if (!(ctx->lc_tags & LCT_NOREF) &&
                            try_module_get(key->lct_owner) == 0) {
                                /* module is unloading, skip this key */
                                continue;
                        }

                        value = key->lct_init(ctx, key);
                        if (unlikely(IS_ERR(value))) {
                                rc = PTR_ERR(value);
                                break;
                        }

                        lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
                        atomic_inc(&key->lct_used);
                        /*
                         * This is the only place in the code, where an
                         * element of ctx->lc_value[] array is set to non-NULL
                         * value.
                         */
                        ctx->lc_value[i] = value;
                        if (key->lct_exit != NULL)
                                ctx->lc_tags |= LCT_HAS_EXIT;
                }
        }

        up_read(&lu_key_initing);
        return rc;
}

static int keys_init(struct lu_context *ctx)
{
        OBD_ALLOC_PTR_ARRAY(ctx->lc_value, ARRAY_SIZE(lu_keys));
        if (likely(ctx->lc_value != NULL))
                return keys_fill(ctx);

        return -ENOMEM;
}

/**
 * Initialize context data-structure. Create values for all keys.
 */
int lu_context_init(struct lu_context *ctx, __u32 tags)
{
        int     rc;

        memset(ctx, 0, sizeof *ctx);
        ctx->lc_state = LCS_INITIALIZED;
        ctx->lc_tags = tags;
        if (tags & LCT_REMEMBER) {
                spin_lock(&lu_context_remembered_guard);
                list_add(&ctx->lc_remember, &lu_context_remembered);
                spin_unlock(&lu_context_remembered_guard);
        } else {
                INIT_LIST_HEAD(&ctx->lc_remember);
        }

        rc = keys_init(ctx);
        if (rc != 0)
                lu_context_fini(ctx);

        return rc;
}
EXPORT_SYMBOL(lu_context_init);

/**
 * Finalize context data-structure. Destroy key values.
 */
void lu_context_fini(struct lu_context *ctx)
{
        LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
        ctx->lc_state = LCS_FINALIZED;

        if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
                LASSERT(list_empty(&ctx->lc_remember));
        } else {
                /* could race with key degister */
                spin_lock(&lu_context_remembered_guard);
                list_del_init(&ctx->lc_remember);
                spin_unlock(&lu_context_remembered_guard);
        }
        keys_fini(ctx);
}
EXPORT_SYMBOL(lu_context_fini);

/**
 * Called before entering context.
 */
void lu_context_enter(struct lu_context *ctx)
{
        LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
        ctx->lc_state = LCS_ENTERED;
}
EXPORT_SYMBOL(lu_context_enter);

/**
 * Called after exiting from \a ctx
 */
void lu_context_exit(struct lu_context *ctx)
{
        unsigned int i;

        LINVRNT(ctx->lc_state == LCS_ENTERED);
        /*
         * Disable preempt to ensure we get a warning if
         * any lct_exit ever tries to sleep.  That would hurt
         * lu_context_key_quiesce() which spins waiting for us.
         * This also ensure we aren't preempted while the state
         * is LCS_LEAVING, as that too would cause problems for
         * lu_context_key_quiesce().
         */
        preempt_disable();
        /*
         * Ensure lu_context_key_quiesce() sees LCS_LEAVING
         * or we see LCT_QUIESCENT
         */
        smp_store_mb(ctx->lc_state, LCS_LEAVING);
        if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value) {
                for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
                        struct lu_context_key *key;

                        key = lu_keys[i];
                        if (ctx->lc_value[i] &&
                            !(key->lct_tags & LCT_QUIESCENT) &&
                            key->lct_exit)
                                key->lct_exit(ctx, key, ctx->lc_value[i]);
                }
        }

        smp_store_release(&ctx->lc_state, LCS_LEFT);
        preempt_enable();
}
EXPORT_SYMBOL(lu_context_exit);

/**
 * Allocate for context all missing keys that were registered after context
 * creation. key_set_version is only changed in rare cases when modules
 * are loaded and removed.
 */
int lu_context_refill(struct lu_context *ctx)
{
        if (likely(ctx->lc_version == atomic_read(&key_set_version)))
                return 0;

        return keys_fill(ctx);
}

/**
 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
 * obd being added. Currently, this is only used on client side, specifically
 * for echo device client, for other stack (like ptlrpc threads), context are
 * predefined when the lu_device type are registered, during the module probe
 * phase.
 */
u32 lu_context_tags_default = LCT_CL_THREAD;
u32 lu_session_tags_default = LCT_SESSION;

void lu_context_tags_update(__u32 tags)
{
        spin_lock(&lu_context_remembered_guard);
        lu_context_tags_default |= tags;
        atomic_inc(&key_set_version);
        spin_unlock(&lu_context_remembered_guard);
}
EXPORT_SYMBOL(lu_context_tags_update);

void lu_context_tags_clear(__u32 tags)
{
        spin_lock(&lu_context_remembered_guard);
        lu_context_tags_default &= ~tags;
        atomic_inc(&key_set_version);
        spin_unlock(&lu_context_remembered_guard);
}
EXPORT_SYMBOL(lu_context_tags_clear);

void lu_session_tags_update(__u32 tags)
{
        spin_lock(&lu_context_remembered_guard);
        lu_session_tags_default |= tags;
        atomic_inc(&key_set_version);
        spin_unlock(&lu_context_remembered_guard);
}
EXPORT_SYMBOL(lu_session_tags_update);

void lu_session_tags_clear(__u32 tags)
{
        spin_lock(&lu_context_remembered_guard);
        lu_session_tags_default &= ~tags;
        atomic_inc(&key_set_version);
        spin_unlock(&lu_context_remembered_guard);
}
EXPORT_SYMBOL(lu_session_tags_clear);

int lu_env_init(struct lu_env *env, __u32 tags)
{
        int result;

        env->le_ses = NULL;
        result = lu_context_init(&env->le_ctx, tags);
        if (likely(result == 0))
                lu_context_enter(&env->le_ctx);
        return result;
}
EXPORT_SYMBOL(lu_env_init);

void lu_env_fini(struct lu_env *env)
{
        lu_context_exit(&env->le_ctx);
        lu_context_fini(&env->le_ctx);
        env->le_ses = NULL;
}
EXPORT_SYMBOL(lu_env_fini);

int lu_env_refill(struct lu_env *env)
{
        int result;

        result = lu_context_refill(&env->le_ctx);
        if (result == 0 && env->le_ses != NULL)
                result = lu_context_refill(env->le_ses);
        return result;
}
EXPORT_SYMBOL(lu_env_refill);

/**
 * Currently, this API will only be used by echo client.
 * Because echo client and normal lustre client will share
 * same cl_env cache. So echo client needs to refresh
 * the env context after it get one from the cache, especially
 * when normal client and echo client co-exist in the same client.
 */
int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
                          __u32 stags)
{
        int    result;

        if ((env->le_ctx.lc_tags & ctags) != ctags) {
                env->le_ctx.lc_version = 0;
                env->le_ctx.lc_tags |= ctags;
        }

        if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
                env->le_ses->lc_version = 0;
                env->le_ses->lc_tags |= stags;
        }

        result = lu_env_refill(env);

        return result;
}
EXPORT_SYMBOL(lu_env_refill_by_tags);


struct lu_env_item {
        struct task_struct *lei_task;   /* rhashtable key */
        struct rhash_head lei_linkage;
        struct lu_env *lei_env;
        struct rcu_head lei_rcu_head;
};

static const struct rhashtable_params lu_env_rhash_params = {
        .key_len     = sizeof(struct task_struct *),
        .key_offset  = offsetof(struct lu_env_item, lei_task),
        .head_offset = offsetof(struct lu_env_item, lei_linkage),
    };

struct rhashtable lu_env_rhash;

struct lu_env_percpu {
        struct task_struct *lep_task;
        struct lu_env *lep_env ____cacheline_aligned_in_smp;
};

static struct lu_env_percpu lu_env_percpu[NR_CPUS];

int lu_env_add_task(struct lu_env *env, struct task_struct *task)
{
        struct lu_env_item *lei, *old;

        LASSERT(env);

        OBD_ALLOC_PTR(lei);
        if (!lei)
                return -ENOMEM;

        lei->lei_task = task;
        lei->lei_env = env;

        old = rhashtable_lookup_get_insert_fast(&lu_env_rhash,
                                                &lei->lei_linkage,
                                                lu_env_rhash_params);
        LASSERT(!old);

        return 0;
}
EXPORT_SYMBOL(lu_env_add_task);

int lu_env_add(struct lu_env *env)
{
        return lu_env_add_task(env, current);
}
EXPORT_SYMBOL(lu_env_add);

static void lu_env_item_free(struct rcu_head *head)
{
        struct lu_env_item *lei;

        lei = container_of(head, struct lu_env_item, lei_rcu_head);
        OBD_FREE_PTR(lei);
}

void lu_env_remove(struct lu_env *env)
{
        struct lu_env_item *lei;
        const void *task = current;
        int i;

        for_each_possible_cpu(i) {
                if (lu_env_percpu[i].lep_env == env) {
                        LASSERT(lu_env_percpu[i].lep_task == task);
                        lu_env_percpu[i].lep_task = NULL;
                        lu_env_percpu[i].lep_env = NULL;
                }
        }

        /* The rcu_lock is not taking in this case since the key
         * used is the actual task_struct. This implies that each
         * object is only removed by the owning thread, so there
         * can never be a race on a particular object.
         */
        lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
                                     lu_env_rhash_params);
        if (lei && rhashtable_remove_fast(&lu_env_rhash, &lei->lei_linkage,
                                          lu_env_rhash_params) == 0)
                call_rcu(&lei->lei_rcu_head, lu_env_item_free);
}
EXPORT_SYMBOL(lu_env_remove);

struct lu_env *lu_env_find(void)
{
        struct lu_env *env = NULL;
        struct lu_env_item *lei;
        const void *task = current;
        int i = get_cpu();

        if (lu_env_percpu[i].lep_task == current) {
                env = lu_env_percpu[i].lep_env;
                put_cpu();
                LASSERT(env);
                return env;
        }

        lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
                                     lu_env_rhash_params);
        if (lei) {
                env = lei->lei_env;
                lu_env_percpu[i].lep_task = current;
                lu_env_percpu[i].lep_env = env;
        }
        put_cpu();

        return env;
}
EXPORT_SYMBOL(lu_env_find);

typedef struct lu_site_stats{
        unsigned        lss_populated;
        unsigned        lss_max_search;
        unsigned        lss_total;
        unsigned        lss_busy;
} lu_site_stats_t;

static void lu_site_stats_get(const struct lu_site *s,
                              lu_site_stats_t *stats)
{
        int cnt = atomic_read(&s->ls_obj_hash.nelems);
        /*
         * percpu_counter_sum_positive() won't accept a const pointer
         * as it does modify the struct by taking a spinlock
         */
        struct lu_site *s2 = (struct lu_site *)s;

        stats->lss_busy += cnt -
                percpu_counter_sum_positive(&s2->ls_lru_len_counter);

        stats->lss_total += cnt;
        stats->lss_max_search = 0;
        stats->lss_populated = 0;
}


/*
 * lu_cache_shrink_count() returns an approximate number of cached objects
 * that can be freed by shrink_slab(). A counter, which tracks the
 * number of items in the site's lru, is maintained in a percpu_counter
 * for each site. The percpu values are incremented and decremented as
 * objects are added or removed from the lru. The percpu values are summed
 * and saved whenever a percpu value exceeds a threshold. Thus the saved,
 * summed value at any given time may not accurately reflect the current
 * lru length. But this value is sufficiently accurate for the needs of
 * a shrinker.
 *
 * Using a per cpu counter is a compromise solution to concurrent access:
 * lu_object_put() can update the counter without locking the site and
 * lu_cache_shrink_count can sum the counters without locking each
 * ls_obj_hash bucket.
 */
static unsigned long lu_cache_shrink_count(struct shrinker *sk,
                                           struct shrink_control *sc)
{
        struct lu_site *s;
        struct lu_site *tmp;
        unsigned long cached = 0;

        if (!(sc->gfp_mask & __GFP_FS))
                return 0;

        down_read(&lu_sites_guard);
        list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage)
                cached += percpu_counter_read_positive(&s->ls_lru_len_counter);
        up_read(&lu_sites_guard);

        cached = (cached / 100) * sysctl_vfs_cache_pressure;
        CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n",
               cached, sysctl_vfs_cache_pressure);

        return cached;
}

static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
                                          struct shrink_control *sc)
{
        struct lu_site *s;
        struct lu_site *tmp;
        unsigned long remain = sc->nr_to_scan;
        LIST_HEAD(splice);

        if (!(sc->gfp_mask & __GFP_FS))
                /* We must not take the lu_sites_guard lock when
                 * __GFP_FS is *not* set because of the deadlock
                 * possibility detailed above. Additionally,
                 * since we cannot determine the number of
                 * objects in the cache without taking this
                 * lock, we're in a particularly tough spot. As
                 * a result, we'll just lie and say our cache is
                 * empty. This _should_ be ok, as we can't
                 * reclaim objects when __GFP_FS is *not* set
                 * anyways.
                 */
                return SHRINK_STOP;

        down_write(&lu_sites_guard);
        list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
                remain = lu_site_purge(&lu_shrink_env, s, remain);
                /*
                 * Move just shrunk site to the tail of site list to
                 * assure shrinking fairness.
                 */
                list_move_tail(&s->ls_linkage, &splice);
        }
        list_splice(&splice, lu_sites.prev);
        up_write(&lu_sites_guard);

        return sc->nr_to_scan - remain;
}

#ifdef HAVE_SHRINKER_COUNT
static struct shrinker lu_site_shrinker = {
        .count_objects  = lu_cache_shrink_count,
        .scan_objects   = lu_cache_shrink_scan,
        .seeks          = DEFAULT_SEEKS,
};

#else
/*
 * There exists a potential lock inversion deadlock scenario when using
 * Lustre on top of ZFS. This occurs between one of ZFS's
 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
 * while thread B will take the ht_lock and sleep on the lu_sites_guard
 * lock. Obviously neither thread will wake and drop their respective hold
 * on their lock.
 *
 * To prevent this from happening we must ensure the lu_sites_guard lock is
 * not taken while down this code path. ZFS reliably does not set the
 * __GFP_FS bit in its code paths, so this can be used to determine if it
 * is safe to take the lu_sites_guard lock.
 *
 * Ideally we should accurately return the remaining number of cached
 * objects without taking the lu_sites_guard lock, but this is not
 * possible in the current implementation.
 */
static int lu_cache_shrink(struct shrinker *shrinker,
                           struct shrink_control *sc)
{
        int cached = 0;

        CDEBUG(D_INODE, "Shrink %lu objects\n", sc->nr_to_scan);

        if (sc->nr_to_scan != 0)
                lu_cache_shrink_scan(shrinker, sc);

        cached = lu_cache_shrink_count(shrinker, sc);
        return cached;
}

static struct shrinker lu_site_shrinker = {
        .shrink  = lu_cache_shrink,
        .seeks   = DEFAULT_SEEKS,
};

#endif /* HAVE_SHRINKER_COUNT */


/*
 * Debugging stuff.
 */

/**
 * Environment to be used in debugger, contains all tags.
 */
static struct lu_env lu_debugging_env;

/**
 * Debugging printer function using printk().
 */
int lu_printk_printer(const struct lu_env *env,
                      void *unused, const char *format, ...)
{
        va_list args;

        va_start(args, format);
        vprintk(format, args);
        va_end(args);
        return 0;
}

int lu_debugging_setup(void)
{
        return lu_env_init(&lu_debugging_env, ~0);
}

void lu_context_keys_dump(void)
{
        unsigned int i;

        for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
                struct lu_context_key *key;

                key = lu_keys[i];
                if (key != NULL) {
                        CERROR("LU context keys [%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
                                i, key, key->lct_tags,
                                key->lct_init, key->lct_fini, key->lct_exit,
                                key->lct_index, atomic_read(&key->lct_used),
                                key->lct_owner ? key->lct_owner->name : "",
                                key->lct_owner);
                        lu_ref_print(&key->lct_reference);
                }
        }
}

/**
 * Initialization of global lu_* data.
 */
int lu_global_init(void)
{
        int result;

        CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);

        result = lu_ref_global_init();
        if (result != 0)
                return result;

        LU_CONTEXT_KEY_INIT(&lu_global_key);
        result = lu_context_key_register(&lu_global_key);
        if (result)
                goto out_lu_ref;

        /*
         * At this level, we don't know what tags are needed, so allocate them
         * conservatively. This should not be too bad, because this
         * environment is global.
         */
        down_write(&lu_sites_guard);
        result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
        up_write(&lu_sites_guard);
        if (result) {
                lu_context_key_degister(&lu_global_key);
                goto out_lu_ref;
        }

        /*
         * seeks estimation: 3 seeks to read a record from oi, one to read
         * inode, one for ea. Unfortunately setting this high value results in
         * lu_object/inode cache consuming all the memory.
         */
        result = register_shrinker(&lu_site_shrinker);
        if (result)
                goto out_env;

        result = rhashtable_init(&lu_env_rhash, &lu_env_rhash_params);

        if (result)
                goto out_shrinker;

        return result;

out_shrinker:
        unregister_shrinker(&lu_site_shrinker);
out_env:
        /* ordering here is explained in lu_global_fini() */
        lu_context_key_degister(&lu_global_key);
        down_write(&lu_sites_guard);
        lu_env_fini(&lu_shrink_env);
        up_write(&lu_sites_guard);
out_lu_ref:
        lu_ref_global_fini();
        return result;
}

/**
 * Dual to lu_global_init().
 */
void lu_global_fini(void)
{
        unregister_shrinker(&lu_site_shrinker);

        lu_context_key_degister(&lu_global_key);

        /*
         * Tear shrinker environment down _after_ de-registering
         * lu_global_key, because the latter has a value in the former.
         */
        down_write(&lu_sites_guard);
        lu_env_fini(&lu_shrink_env);
        up_write(&lu_sites_guard);

        rhashtable_destroy(&lu_env_rhash);

        lu_ref_global_fini();
}

static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
{
#ifdef CONFIG_PROC_FS
        struct lprocfs_counter ret;

        lprocfs_stats_collect(stats, idx, &ret);
        return (__u32)ret.lc_count;
#else
        return 0;
#endif
}

/**
 * Output site statistical counters into a buffer. Suitable for
 * lprocfs_rd_*()-style functions.
 */
int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
{
        const struct bucket_table *tbl;
        lu_site_stats_t stats;
        unsigned int chains;

        memset(&stats, 0, sizeof(stats));
        lu_site_stats_get(s, &stats);

        rcu_read_lock();
        tbl = rht_dereference_rcu(s->ls_obj_hash.tbl,
                                  &((struct lu_site *)s)->ls_obj_hash);
        chains = tbl->size;
        rcu_read_unlock();
        seq_printf(m, "%d/%d %d/%u %d %d %d %d %d %d %d\n",
                   stats.lss_busy,
                   stats.lss_total,
                   stats.lss_populated,
                   chains,
                   stats.lss_max_search,
                   ls_stats_read(s->ls_stats, LU_SS_CREATED),
                   ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
                   ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
                   ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
                   ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
                   ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
        return 0;
}
EXPORT_SYMBOL(lu_site_stats_seq_print);

/**
 * Helper function to initialize a number of kmem slab caches at once.
 */
int lu_kmem_init(struct lu_kmem_descr *caches)
{
        int result;
        struct lu_kmem_descr *iter = caches;

        for (result = 0; iter->ckd_cache != NULL; ++iter) {
                *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
                                                     iter->ckd_size,
                                                     0, 0, NULL);
                if (*iter->ckd_cache == NULL) {
                        result = -ENOMEM;
                        /* free all previously allocated caches */
                        lu_kmem_fini(caches);
                        break;
                }
        }
        return result;
}
EXPORT_SYMBOL(lu_kmem_init);

/**
 * Helper function to finalize a number of kmem slab cached at once. Dual to
 * lu_kmem_init().
 */
void lu_kmem_fini(struct lu_kmem_descr *caches)
{
        for (; caches->ckd_cache != NULL; ++caches) {
                if (*caches->ckd_cache != NULL) {
                        kmem_cache_destroy(*caches->ckd_cache);
                        *caches->ckd_cache = NULL;
                }
        }
}
EXPORT_SYMBOL(lu_kmem_fini);

/**
 * Temporary solution to be able to assign fid in ->do_create()
 * till we have fully-functional OST fids
 */
void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
                          const struct lu_fid *fid)
{
        struct lu_site          *s = o->lo_dev->ld_site;
        struct lu_fid           *old = &o->lo_header->loh_fid;
        int rc;

        LASSERT(fid_is_zero(old));
        *old = *fid;
try_again:
        rc = rhashtable_lookup_insert_fast(&s->ls_obj_hash,
                                           &o->lo_header->loh_hash,
                                           obj_hash_params);
        /* supposed to be unique */
        LASSERT(rc != -EEXIST);
        /* handle hash table resizing */
        if (rc == -ENOMEM || rc == -EBUSY) {
                msleep(20);
                goto try_again;
        }
        /* trim the hash if its growing to big */
        lu_object_limit(env, o->lo_dev);
        if (rc == -E2BIG)
                goto try_again;

        LASSERTF(rc == 0, "failed hashtable insertion: rc = %d\n", rc);
}
EXPORT_SYMBOL(lu_object_assign_fid);

/**
 * allocates object with 0 (non-assiged) fid
 * XXX: temporary solution to be able to assign fid in ->do_create()
 *      till we have fully-functional OST fids
 */
struct lu_object *lu_object_anon(const struct lu_env *env,
                                 struct lu_device *dev,
                                 const struct lu_object_conf *conf)
{
        struct lu_fid fid;
        struct lu_object *o;
        int rc;

        fid_zero(&fid);
        o = lu_object_alloc(env, dev, &fid);
        if (!IS_ERR(o)) {
                rc = lu_object_start(env, dev, o, conf);
                if (rc) {
                        lu_object_free(env, o);
                        return ERR_PTR(rc);
                }
        }

        return o;
}
EXPORT_SYMBOL(lu_object_anon);

struct lu_buf LU_BUF_NULL = {
        .lb_buf = NULL,
        .lb_len = 0
};
EXPORT_SYMBOL(LU_BUF_NULL);

void lu_buf_free(struct lu_buf *buf)
{
        LASSERT(buf);
        if (buf->lb_buf) {
                LASSERT(buf->lb_len > 0);
                OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
                buf->lb_buf = NULL;
                buf->lb_len = 0;
        }
}
EXPORT_SYMBOL(lu_buf_free);

void lu_buf_alloc(struct lu_buf *buf, size_t size)
{
        LASSERT(buf);
        LASSERT(buf->lb_buf == NULL);
        LASSERT(buf->lb_len == 0);
        OBD_ALLOC_LARGE(buf->lb_buf, size);
        if (likely(buf->lb_buf))
                buf->lb_len = size;
}
EXPORT_SYMBOL(lu_buf_alloc);

void lu_buf_realloc(struct lu_buf *buf, size_t size)
{
        lu_buf_free(buf);
        lu_buf_alloc(buf, size);
}
EXPORT_SYMBOL(lu_buf_realloc);

struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
{
        if (buf->lb_buf == NULL && buf->lb_len == 0)
                lu_buf_alloc(buf, len);

        if ((len > buf->lb_len) && (buf->lb_buf != NULL))
                lu_buf_realloc(buf, len);

        return buf;
}
EXPORT_SYMBOL(lu_buf_check_and_alloc);

/**
 * Increase the size of the \a buf.
 * preserves old data in buffer
 * old buffer remains unchanged on error
 * \retval 0 or -ENOMEM
 */
int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
{
        char *ptr;

        if (len <= buf->lb_len)
                return 0;

        OBD_ALLOC_LARGE(ptr, len);
        if (ptr == NULL)
                return -ENOMEM;

        /* Free the old buf */
        if (buf->lb_buf != NULL) {
                memcpy(ptr, buf->lb_buf, buf->lb_len);
                OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
        }

        buf->lb_buf = ptr;
        buf->lb_len = len;
        return 0;
}
EXPORT_SYMBOL(lu_buf_check_and_grow);