Whamcloud - gitweb
LU-2924 Only wake up ldlm_poold as frequently as the check interval
[fs/lustre-release.git] / lustre / ldlm / ldlm_pool.c
index 7aca642..5289a4e 100644 (file)
@@ -1,29 +1,45 @@
-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
+/*
+ * GPL HEADER START
  *
- *  Copyright (c) 2007 Cluster File Systems, Inc.
- *   Author: Yury Umanets <umka@clusterfs.com>
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
- *   This file is part of the Lustre file system, http://www.lustre.org
- *   Lustre is a trademark of Cluster File Systems, Inc.
+ * 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.
  *
- *   You may have signed or agreed to another license before downloading
- *   this software.  If so, you are bound by the terms and conditions
- *   of that agreement, and the following does not apply to you.  See the
- *   LICENSE file included with this distribution for more information.
+ * 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).
  *
- *   If you did not agree to a different license, then this copy of Lustre
- *   is open source software; you can redistribute it and/or modify it
- *   under the terms of version 2 of the GNU General Public License as
- *   published by the Free Software Foundation.
+ * You should have received a copy of the GNU General Public License
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  *
- *   In either case, Lustre 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
- *   license text for more details.
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
+ *
+ * GPL HEADER END
+ */
+/*
+ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Use is subject to license terms.
+ *
+ * Copyright (c) 2010, 2012, Intel Corporation.
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ * Lustre is a trademark of Sun Microsystems, Inc.
+ *
+ * lustre/ldlm/ldlm_pool.c
+ *
+ * Author: Yury Umanets <umka@clusterfs.com>
  */
 
-/* Idea of this code is rather simple. Each second, for each server namespace
+/*
+ * Idea of this code is rather simple. Each second, for each server namespace
  * we have SLV - server lock volume which is calculated on current number of
  * granted locks, grant speed for past period, etc - that is, locking load.
  * This SLV number may be thought as a flow definition for simplicity. It is
  * pl_cancel_rate - Number of canceled locks for last T (calculated);
  * pl_grant_speed - Grant speed (GR - CR) for last T (calculated);
  * pl_grant_plan - Planned number of granted locks for next T (calculated);
- *
- * pl_grant_step - Grant plan step, that is how ->pl_grant_plan
- * will change in next T (tunable);
- *
  * pl_server_lock_volume - Current server lock volume (calculated);
  *
  * As it may be seen from list above, we have few possible tunables which may
 # include <lustre_dlm.h>
 #else
 # include <liblustre.h>
-# include <libcfs/kp30.h>
 #endif
 
+#include <cl_object.h>
+
 #include <obd_class.h>
 #include <obd_support.h>
 #include "ldlm_internal.h"
 
 #ifdef HAVE_LRU_RESIZE_SUPPORT
 
-/* 50 ldlm locks for 1MB of RAM. */
-#define LDLM_POOL_HOST_L ((num_physpages >> (20 - PAGE_SHIFT)) * 50)
+/*
+ * 50 ldlm locks for 1MB of RAM.
+ */
+#define LDLM_POOL_HOST_L ((CFS_NUM_CACHEPAGES >> (20 - CFS_PAGE_SHIFT)) * 50)
+
+/*
+ * Maximal possible grant step plan in %.
+ */
+#define LDLM_POOL_MAX_GSP (30)
+
+/*
+ * Minimal possible grant step plan in %.
+ */
+#define LDLM_POOL_MIN_GSP (1)
 
-/* Default step in % for grant plan. */
-#define LDLM_POOL_GSP (5)
+/*
+ * This controls the speed of reaching LDLM_POOL_MAX_GSP
+ * with increasing thread period.
+ */
+#define LDLM_POOL_GSP_STEP_SHIFT (2)
 
-/* LDLM_POOL_GSP% of all locks is default GP. */
-#define LDLM_POOL_GP(L)   ((L) * LDLM_POOL_GSP / 100)
+/*
+ * LDLM_POOL_GSP% of all locks is default GP.
+ */
+#define LDLM_POOL_GP(L)   (((L) * LDLM_POOL_MAX_GSP) / 100)
 
-/* Max age for locks on clients. */
+/*
+ * Max age for locks on clients.
+ */
 #define LDLM_POOL_MAX_AGE (36000)
 
+/*
+ * The granularity of SLV calculation.
+ */
+#define LDLM_POOL_SLV_SHIFT (10)
+
 #ifdef __KERNEL__
 extern cfs_proc_dir_entry_t *ldlm_ns_proc_dir;
 #endif
 
-#define avg(src, add) \
-        ((src) = ((src) + (add)) / 2)
-
-static inline __u64 dru(__u64 val, __u32 div)
+static inline __u64 dru(__u64 val, __u32 shift, int round_up)
 {
-        __u64 ret = val + (div - 1);
-        do_div(ret, div);
-        return ret;
+        return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift;
 }
 
 static inline __u64 ldlm_pool_slv_max(__u32 L)
 {
-        /* Allow to have all locks for 1 client for 10 hrs.
-         * Formula is the following: limit * 10h / 1 client. */
-        __u64 lim = L *  LDLM_POOL_MAX_AGE / 1;
+        /*
+         * Allow to have all locks for 1 client for 10 hrs.
+         * Formula is the following: limit * 10h / 1 client.
+         */
+        __u64 lim = (__u64)L *  LDLM_POOL_MAX_AGE / 1;
         return lim;
 }
 
@@ -149,6 +183,7 @@ enum {
         LDLM_POOL_SHRINK_REQTD_STAT,
         LDLM_POOL_SHRINK_FREED_STAT,
         LDLM_POOL_RECALC_STAT,
+        LDLM_POOL_TIMING_STAT,
         LDLM_POOL_LAST_STAT
 };
 
@@ -157,37 +192,92 @@ static inline struct ldlm_namespace *ldlm_pl2ns(struct ldlm_pool *pl)
         return container_of(pl, struct ldlm_namespace, ns_pool);
 }
 
-/* Should be called under ->pl_lock taken */
-static inline void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl)
+/**
+ * Calculates suggested grant_step in % of available locks for passed
+ * \a period. This is later used in grant_plan calculations.
+ */
+static inline int ldlm_pool_t2gsp(unsigned int t)
 {
-        int grant_plan, granted;
-        __u32 limit;
-        
-        limit = ldlm_pool_get_limit(pl);
-        granted = atomic_read(&pl->pl_granted);
-
-        grant_plan = granted + ((limit - granted) *
-                atomic_read(&pl->pl_grant_step)) / 100;
-        atomic_set(&pl->pl_grant_plan, grant_plan);
+        /*
+         * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP
+         * and up to 30% for anything higher than LDLM_POOL_GSP_STEP.
+         *
+         * How this will affect execution is the following:
+         *
+         * - for thread period 1s we will have grant_step 1% which good from
+         * pov of taking some load off from server and push it out to clients.
+         * This is like that because 1% for grant_step means that server will
+         * not allow clients to get lots of locks in short period of time and
+         * keep all old locks in their caches. Clients will always have to
+         * get some locks back if they want to take some new;
+         *
+         * - for thread period 10s (which is default) we will have 23% which
+         * means that clients will have enough of room to take some new locks
+         * without getting some back. All locks from this 23% which were not
+         * taken by clients in current period will contribute in SLV growing.
+         * SLV growing means more locks cached on clients until limit or grant
+         * plan is reached.
+         */
+        return LDLM_POOL_MAX_GSP -
+                ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >>
+                 (t >> LDLM_POOL_GSP_STEP_SHIFT));
 }
 
-/* Should be called under ->pl_lock taken */
-static inline void ldlm_pool_recalc_slv(struct ldlm_pool *pl)
+/**
+ * Recalculates next grant limit on passed \a pl.
+ *
+ * \pre ->pl_lock is locked.
+ */
+static void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl)
 {
-        int slv_factor, granted, grant_plan;
-        __u32 limit;
-        __u64 slv;
+        int granted, grant_step, limit;
 
-        slv = ldlm_pool_get_slv(pl);
         limit = ldlm_pool_get_limit(pl);
-        granted = atomic_read(&pl->pl_granted);
-        grant_plan = atomic_read(&pl->pl_grant_plan);
+        granted = cfs_atomic_read(&pl->pl_granted);
+
+        grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
+        grant_step = ((limit - granted) * grant_step) / 100;
+        pl->pl_grant_plan = granted + grant_step;
+        limit = (limit * 5) >> 2;
+        if (pl->pl_grant_plan > limit)
+                pl->pl_grant_plan = limit;
+}
 
-        if ((slv_factor = limit - (granted - grant_plan)) <= 0)
-                slv_factor = 1;
+/**
+ * Recalculates next SLV on passed \a pl.
+ *
+ * \pre ->pl_lock is locked.
+ */
+static void ldlm_pool_recalc_slv(struct ldlm_pool *pl)
+{
+        int granted;
+        int grant_plan;
+        int round_up;
+        __u64 slv;
+        __u64 slv_factor;
+        __u64 grant_usage;
+        __u32 limit;
 
-        slv = (slv * ((slv_factor * 100) / limit));
-        slv = dru(slv, 100);
+        slv = pl->pl_server_lock_volume;
+        grant_plan = pl->pl_grant_plan;
+        limit = ldlm_pool_get_limit(pl);
+        granted = cfs_atomic_read(&pl->pl_granted);
+        round_up = granted < limit;
+
+        grant_usage = max_t(int, limit - (granted - grant_plan), 1);
+
+        /*
+         * Find out SLV change factor which is the ratio of grant usage
+         * from limit. SLV changes as fast as the ratio of grant plan
+         * consumption. The more locks from grant plan are not consumed
+         * by clients in last interval (idle time), the faster grows
+         * SLV. And the opposite, the more grant plan is over-consumed
+         * (load time) the faster drops SLV.
+         */
+        slv_factor = (grant_usage << LDLM_POOL_SLV_SHIFT);
+        do_div(slv_factor, limit);
+        slv = slv * slv_factor;
+        slv = dru(slv, LDLM_POOL_SLV_SHIFT, round_up);
 
         if (slv > ldlm_pool_slv_max(limit)) {
                 slv = ldlm_pool_slv_max(limit);
@@ -195,18 +285,23 @@ static inline void ldlm_pool_recalc_slv(struct ldlm_pool *pl)
                 slv = ldlm_pool_slv_min(limit);
         }
 
-        ldlm_pool_set_slv(pl, slv);
+        pl->pl_server_lock_volume = slv;
 }
 
-static inline void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
+/**
+ * Recalculates next stats on passed \a pl.
+ *
+ * \pre ->pl_lock is locked.
+ */
+static void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
 {
-        __u64 slv = ldlm_pool_get_slv(pl);
-        __u32 granted = atomic_read(&pl->pl_granted);
-        __u32 grant_rate = atomic_read(&pl->pl_grant_rate);
-        __u32 grant_plan = atomic_read(&pl->pl_grant_plan);
-        __u32 cancel_rate = atomic_read(&pl->pl_cancel_rate);
+        int grant_plan = pl->pl_grant_plan;
+        __u64 slv = pl->pl_server_lock_volume;
+        int granted = cfs_atomic_read(&pl->pl_granted);
+        int grant_rate = cfs_atomic_read(&pl->pl_grant_rate);
+        int cancel_rate = cfs_atomic_read(&pl->pl_cancel_rate);
 
-        lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT, 
+        lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT,
                             slv);
         lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
                             granted);
@@ -218,133 +313,255 @@ static inline void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
                             cancel_rate);
 }
 
+/**
+ * Sets current SLV into obd accessible via ldlm_pl2ns(pl)->ns_obd.
+ */
+static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl)
+{
+        struct obd_device *obd;
+
+        /*
+         * Set new SLV in obd field for using it later without accessing the
+         * pool. This is required to avoid race between sending reply to client
+         * with new SLV and cleanup server stack in which we can't guarantee
+         * that namespace is still alive. We know only that obd is alive as
+         * long as valid export is alive.
+         */
+        obd = ldlm_pl2ns(pl)->ns_obd;
+        LASSERT(obd != NULL);
+       write_lock(&obd->obd_pool_lock);
+        obd->obd_pool_slv = pl->pl_server_lock_volume;
+       write_unlock(&obd->obd_pool_lock);
+}
+
+/**
+ * Recalculates all pool fields on passed \a pl.
+ *
+ * \pre ->pl_lock is not locked.
+ */
 static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
 {
         time_t recalc_interval_sec;
         ENTRY;
 
-        spin_lock(&pl->pl_lock);
         recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
-        if (recalc_interval_sec > 0) {
-                /* Update statistics */
-                ldlm_pool_recalc_stats(pl);
-
-                /* Recalc SLV after last period. This should be done
-                 * _before_ recalculating new grant plan. */
-                ldlm_pool_recalc_slv(pl);
+        if (recalc_interval_sec < pl->pl_recalc_period)
+                RETURN(0);
 
-                /* Update grant_plan for new period. */
-                ldlm_pool_recalc_grant_plan(pl);
+       spin_lock(&pl->pl_lock);
+       recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
+       if (recalc_interval_sec < pl->pl_recalc_period) {
+               spin_unlock(&pl->pl_lock);
+               RETURN(0);
+       }
+        /*
+         * Recalc SLV after last period. This should be done
+         * _before_ recalculating new grant plan.
+         */
+        ldlm_pool_recalc_slv(pl);
+
+        /*
+         * Make sure that pool informed obd of last SLV changes.
+         */
+        ldlm_srv_pool_push_slv(pl);
+
+        /*
+         * Update grant_plan for new period.
+         */
+        ldlm_pool_recalc_grant_plan(pl);
 
-                /* Zero out all rates and speed for the last period. */
-                atomic_set(&pl->pl_grant_rate, 0);
-                atomic_set(&pl->pl_cancel_rate, 0);
-                atomic_set(&pl->pl_grant_speed, 0);
-                pl->pl_recalc_time = cfs_time_current_sec();
-        }
-        spin_unlock(&pl->pl_lock);
-        RETURN(0);
+        pl->pl_recalc_time = cfs_time_current_sec();
+        lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
+                            recalc_interval_sec);
+       spin_unlock(&pl->pl_lock);
+       RETURN(0);
 }
 
-/* Our goal here is to decrease SLV the way to make a client hold
- * @nr locks smaller in next 10h. */
+/**
+ * This function is used on server side as main entry point for memory
+ * pressure handling. It decreases SLV on \a pl according to passed
+ * \a nr and \a gfp_mask.
+ *
+ * Our goal here is to decrease SLV such a way that clients hold \a nr
+ * locks smaller in next 10h.
+ */
 static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
                                 int nr, unsigned int gfp_mask)
 {
         __u32 limit;
-        ENTRY;
 
-        /* VM is asking how many entries may be potentially freed. */
+        /*
+         * VM is asking how many entries may be potentially freed.
+         */
         if (nr == 0)
-                RETURN(atomic_read(&pl->pl_granted));
+                return cfs_atomic_read(&pl->pl_granted);
 
-        /* Client already canceled locks but server is already in shrinker
-         * and can't cancel anything. Let's catch this race. */
-        if (atomic_read(&pl->pl_granted) == 0)
+        /*
+         * Client already canceled locks but server is already in shrinker
+         * and can't cancel anything. Let's catch this race.
+         */
+        if (cfs_atomic_read(&pl->pl_granted) == 0)
                 RETURN(0);
 
-        spin_lock(&pl->pl_lock);
+       spin_lock(&pl->pl_lock);
 
-        /* We want shrinker to possibly cause cancelation of @nr locks from
+        /*
+         * We want shrinker to possibly cause cancellation of @nr locks from
          * clients or grant approximately @nr locks smaller next intervals.
          *
-         * This is why we decresed SLV by @nr. This effect will only be as
+         * This is why we decreased SLV by @nr. This effect will only be as
          * long as one re-calc interval (1s these days) and this should be
          * enough to pass this decreased SLV to all clients. On next recalc
          * interval pool will either increase SLV if locks load is not high
          * or will keep on same level or even decrease again, thus, shrinker
          * decreased SLV will affect next recalc intervals and this way will
-         * make locking load lower. */
-        if (nr < ldlm_pool_get_slv(pl)) {
-                ldlm_pool_set_slv(pl, ldlm_pool_get_slv(pl) - nr);
+         * make locking load lower.
+         */
+        if (nr < pl->pl_server_lock_volume) {
+                pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr;
         } else {
                 limit = ldlm_pool_get_limit(pl);
-                ldlm_pool_set_slv(pl, ldlm_pool_slv_min(limit));
+                pl->pl_server_lock_volume = ldlm_pool_slv_min(limit);
         }
-        spin_unlock(&pl->pl_lock);
 
-        /* We did not really free any memory here so far, it only will be
-         * freed later may be, so that we return 0 to not confuse VM. */
-        RETURN(0);
+        /*
+         * Make sure that pool informed obd of last SLV changes.
+         */
+        ldlm_srv_pool_push_slv(pl);
+       spin_unlock(&pl->pl_lock);
+
+       /*
+        * We did not really free any memory here so far, it only will be
+        * freed later may be, so that we return 0 to not confuse VM.
+        */
+       return 0;
 }
 
+/**
+ * Setup server side pool \a pl with passed \a limit.
+ */
 static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit)
 {
-        ENTRY;
+        struct obd_device *obd;
+
+        obd = ldlm_pl2ns(pl)->ns_obd;
+        LASSERT(obd != NULL && obd != LP_POISON);
+        LASSERT(obd->obd_type != LP_POISON);
+       write_lock(&obd->obd_pool_lock);
+        obd->obd_pool_limit = limit;
+       write_unlock(&obd->obd_pool_lock);
+
         ldlm_pool_set_limit(pl, limit);
-        RETURN(0);
+        return 0;
+}
+
+/**
+ * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl.
+ */
+static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
+{
+        struct obd_device *obd;
+
+        /*
+         * Get new SLV and Limit from obd which is updated with coming
+         * RPCs.
+         */
+        obd = ldlm_pl2ns(pl)->ns_obd;
+        LASSERT(obd != NULL);
+       read_lock(&obd->obd_pool_lock);
+        pl->pl_server_lock_volume = obd->obd_pool_slv;
+        ldlm_pool_set_limit(pl, obd->obd_pool_limit);
+       read_unlock(&obd->obd_pool_lock);
 }
 
+/**
+ * Recalculates client size pool \a pl according to current SLV and Limit.
+ */
 static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
 {
         time_t recalc_interval_sec;
         ENTRY;
 
-        spin_lock(&pl->pl_lock);
-
         recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
-        if (recalc_interval_sec > 0) {
-                /* Update statistics only every T */
-                ldlm_pool_recalc_stats(pl);
+        if (recalc_interval_sec < pl->pl_recalc_period)
+                RETURN(0);
 
-                /* Zero out grant/cancel rates and speed for last period. */
-                atomic_set(&pl->pl_grant_rate, 0);
-                atomic_set(&pl->pl_cancel_rate, 0);
-                atomic_set(&pl->pl_grant_speed, 0);
-                pl->pl_recalc_time = cfs_time_current_sec();
+       spin_lock(&pl->pl_lock);
+       /*
+        * Check if we need to recalc lists now.
+        */
+       recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
+       if (recalc_interval_sec < pl->pl_recalc_period) {
+               spin_unlock(&pl->pl_lock);
+                RETURN(0);
         }
-        spin_unlock(&pl->pl_lock);
 
-        /* Do not cancel locks in case lru resize is disabled for this ns */
+        /*
+         * Make sure that pool knows last SLV and Limit from obd.
+         */
+        ldlm_cli_pool_pop_slv(pl);
+
+        pl->pl_recalc_time = cfs_time_current_sec();
+        lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
+                            recalc_interval_sec);
+       spin_unlock(&pl->pl_lock);
+
+        /*
+         * Do not cancel locks in case lru resize is disabled for this ns.
+         */
         if (!ns_connect_lru_resize(ldlm_pl2ns(pl)))
                 RETURN(0);
 
-        /* In the time of canceling locks on client we do not need to maintain
+        /*
+         * In the time of canceling locks on client we do not need to maintain
          * sharp timing, we only want to cancel locks asap according to new SLV.
-         * This may be called when SLV has changed much, this is why we do not
-         * take into account pl->pl_recalc_time here. */
-        RETURN(ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LDLM_ASYNC, 
-                               LDLM_CANCEL_LRUR));
+         * It may be called when SLV has changed much, this is why we do not
+         * take into account pl->pl_recalc_time here.
+         */
+       RETURN(ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC,
+                              LDLM_CANCEL_LRUR));
 }
 
+/**
+ * This function is main entry point for memory pressure handling on client
+ * side.  Main goal of this function is to cancel some number of locks on
+ * passed \a pl according to \a nr and \a gfp_mask.
+ */
 static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
                                 int nr, unsigned int gfp_mask)
 {
-        ENTRY;
-        
-        /* Do not cancel locks in case lru resize is disabled for this ns */
-        if (!ns_connect_lru_resize(ldlm_pl2ns(pl)))
+        struct ldlm_namespace *ns;
+        int canceled = 0, unused;
+
+        ns = ldlm_pl2ns(pl);
+
+        /*
+         * Do not cancel locks in case lru resize is disabled for this ns.
+         */
+        if (!ns_connect_lru_resize(ns))
                 RETURN(0);
 
-        /* Find out how many locks may be released according to shrink 
-         * policy. */
-        if (nr == 0)
-                RETURN(ldlm_cancel_lru_local(ldlm_pl2ns(pl), NULL, 0, 
-                                             0, 0, LDLM_CANCEL_SHRINK));
+        /*
+         * Make sure that pool knows last SLV and Limit from obd.
+         */
+        ldlm_cli_pool_pop_slv(pl);
+
+       spin_lock(&ns->ns_lock);
+       unused = ns->ns_nr_unused;
+       spin_unlock(&ns->ns_lock);
 
-        /* Cancel @nr locks accoding to shrink policy */
-        RETURN(ldlm_cancel_lru(ldlm_pl2ns(pl), nr, LDLM_SYNC, 
-                               LDLM_CANCEL_SHRINK));
+        if (nr) {
+               canceled = ldlm_cancel_lru(ns, nr, LCF_ASYNC,
+                                          LDLM_CANCEL_SHRINK);
+        }
+#ifdef __KERNEL__
+        /*
+         * Return the number of potentially reclaimable locks.
+         */
+        return ((unused - canceled) / 100) * sysctl_vfs_cache_pressure;
+#else
+        return unused - canceled;
+#endif
 }
 
 struct ldlm_pool_ops ldlm_srv_pool_ops = {
@@ -358,32 +575,62 @@ struct ldlm_pool_ops ldlm_cli_pool_ops = {
         .po_shrink = ldlm_cli_pool_shrink
 };
 
+/**
+ * Pool recalc wrapper. Will call either client or server pool recalc callback
+ * depending what pool \a pl is used.
+ */
 int ldlm_pool_recalc(struct ldlm_pool *pl)
 {
+        time_t recalc_interval_sec;
         int count;
 
-        if (pl->pl_ops->po_recalc != NULL && pool_recalc_enabled(pl)) {
+        recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
+        if (recalc_interval_sec <= 0)
+                goto recalc;
+
+       spin_lock(&pl->pl_lock);
+        if (recalc_interval_sec > 0) {
+                /*
+                 * Update pool statistics every 1s.
+                 */
+                ldlm_pool_recalc_stats(pl);
+
+                /*
+                 * Zero out all rates and speed for the last period.
+                 */
+                cfs_atomic_set(&pl->pl_grant_rate, 0);
+                cfs_atomic_set(&pl->pl_cancel_rate, 0);
+        }
+       spin_unlock(&pl->pl_lock);
+
+ recalc:
+        if (pl->pl_ops->po_recalc != NULL) {
                 count = pl->pl_ops->po_recalc(pl);
-                lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT, 
+                lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
                                     count);
-                return count;
         }
-        return 0;
+       recalc_interval_sec = pl->pl_recalc_time - cfs_time_current_sec() +
+                             pl->pl_recalc_period;
+
+        return recalc_interval_sec;
 }
-EXPORT_SYMBOL(ldlm_pool_recalc);
 
+/**
+ * Pool shrink wrapper. Will call either client or server pool recalc callback
+ * depending what pool \a pl is used.
+ */
 int ldlm_pool_shrink(struct ldlm_pool *pl, int nr,
                      unsigned int gfp_mask)
 {
         int cancel = 0;
-        
-        if (pl->pl_ops->po_shrink != NULL && pool_shrink_enabled(pl)) {
+
+        if (pl->pl_ops->po_shrink != NULL) {
                 cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask);
                 if (nr > 0) {
-                        lprocfs_counter_add(pl->pl_stats, 
+                        lprocfs_counter_add(pl->pl_stats,
                                             LDLM_POOL_SHRINK_REQTD_STAT,
                                             nr);
-                        lprocfs_counter_add(pl->pl_stats, 
+                        lprocfs_counter_add(pl->pl_stats,
                                             LDLM_POOL_SHRINK_FREED_STAT,
                                             cancel);
                         CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, "
@@ -394,14 +641,17 @@ int ldlm_pool_shrink(struct ldlm_pool *pl, int nr,
 }
 EXPORT_SYMBOL(ldlm_pool_shrink);
 
-/* The purpose of this function is to re-setup limit and maximal allowed
- * slv according to the passed limit. */
+/**
+ * Pool setup wrapper. Will call either client or server pool recalc callback
+ * depending what pool \a pl is used.
+ *
+ * Sets passed \a limit into pool \a pl.
+ */
 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
 {
-        ENTRY;
         if (pl->pl_ops->po_setup != NULL)
-                RETURN(pl->pl_ops->po_setup(pl, limit));
-        RETURN(0);
+                return(pl->pl_ops->po_setup(pl, limit));
+        return 0;
 }
 EXPORT_SYMBOL(ldlm_pool_setup);
 
@@ -409,34 +659,37 @@ EXPORT_SYMBOL(ldlm_pool_setup);
 static int lprocfs_rd_pool_state(char *page, char **start, off_t off,
                                  int count, int *eof, void *data)
 {
-        __u32 granted, grant_rate, cancel_rate, grant_step;
-        int nr = 0, grant_speed, grant_plan;
+        int granted, grant_rate, cancel_rate, grant_step;
+        int nr = 0, grant_speed, grant_plan, lvf;
         struct ldlm_pool *pl = data;
+        __u64 slv, clv;
         __u32 limit;
-        __u64 slv;
 
-        spin_lock(&pl->pl_lock);
-        slv = ldlm_pool_get_slv(pl);
+       spin_lock(&pl->pl_lock);
+        slv = pl->pl_server_lock_volume;
+        clv = pl->pl_client_lock_volume;
         limit = ldlm_pool_get_limit(pl);
-        granted = atomic_read(&pl->pl_granted);
-        grant_rate = atomic_read(&pl->pl_grant_rate);
-        grant_plan = atomic_read(&pl->pl_grant_plan);
-        grant_step = atomic_read(&pl->pl_grant_step);
-        grant_speed = atomic_read(&pl->pl_grant_speed);
-        cancel_rate = atomic_read(&pl->pl_cancel_rate);
-        spin_unlock(&pl->pl_lock);
+        grant_plan = pl->pl_grant_plan;
+        granted = cfs_atomic_read(&pl->pl_granted);
+        grant_rate = cfs_atomic_read(&pl->pl_grant_rate);
+        cancel_rate = cfs_atomic_read(&pl->pl_cancel_rate);
+        grant_speed = grant_rate - cancel_rate;
+        lvf = cfs_atomic_read(&pl->pl_lock_volume_factor);
+        grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
+       spin_unlock(&pl->pl_lock);
 
         nr += snprintf(page + nr, count - nr, "LDLM pool state (%s):\n",
                        pl->pl_name);
         nr += snprintf(page + nr, count - nr, "  SLV: "LPU64"\n", slv);
-
-        nr += snprintf(page + nr, count - nr, "  LVF: %d\n",
-                       atomic_read(&pl->pl_lock_volume_factor));
-
-        nr += snprintf(page + nr, count - nr, "  GSP: %d%%\n",
-                       grant_step);
-        nr += snprintf(page + nr, count - nr, "  GP:  %d\n",
-                       grant_plan);
+        nr += snprintf(page + nr, count - nr, "  CLV: "LPU64"\n", clv);
+        nr += snprintf(page + nr, count - nr, "  LVF: %d\n", lvf);
+
+        if (ns_is_server(ldlm_pl2ns(pl))) {
+                nr += snprintf(page + nr, count - nr, "  GSP: %d%%\n",
+                               grant_step);
+                nr += snprintf(page + nr, count - nr, "  GP:  %d\n",
+                               grant_plan);
+        }
         nr += snprintf(page + nr, count - nr, "  GR:  %d\n",
                        grant_rate);
         nr += snprintf(page + nr, count - nr, "  CR:  %d\n",
@@ -450,6 +703,24 @@ static int lprocfs_rd_pool_state(char *page, char **start, off_t off,
         return nr;
 }
 
+static int lprocfs_rd_grant_speed(char *page, char **start, off_t off,
+                                 int count, int *eof, void *data)
+{
+       struct ldlm_pool *pl = data;
+       int               grant_speed;
+
+       spin_lock(&pl->pl_lock);
+       /* serialize with ldlm_pool_recalc */
+       grant_speed = cfs_atomic_read(&pl->pl_grant_rate) -
+                       cfs_atomic_read(&pl->pl_cancel_rate);
+       spin_unlock(&pl->pl_lock);
+       return lprocfs_rd_uint(page, start, off, count, eof, &grant_speed);
+}
+
+LDLM_POOL_PROC_READER(grant_plan, int);
+LDLM_POOL_PROC_READER(recalc_period, int);
+LDLM_POOL_PROC_WRITER(recalc_period, int);
+
 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
 {
         struct ldlm_namespace *ns = ldlm_pl2ns(pl);
@@ -463,10 +734,11 @@ static int ldlm_pool_proc_init(struct ldlm_pool *pl)
         if (!var_name)
                 RETURN(-ENOMEM);
 
-        parent_ns_proc = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
+        parent_ns_proc = lprocfs_srch(ldlm_ns_proc_dir,
+                                      ldlm_ns_name(ns));
         if (parent_ns_proc == NULL) {
                 CERROR("%s: proc entry is not initialized\n",
-                       ns->ns_name);
+                       ldlm_ns_name(ns));
                 GOTO(out_free_name, rc = -EINVAL);
         }
         pl->pl_proc_dir = lprocfs_register("pool", parent_ns_proc,
@@ -497,15 +769,9 @@ static int ldlm_pool_proc_init(struct ldlm_pool *pl)
         pool_vars[0].read_fptr = lprocfs_rd_atomic;
         lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
 
-        snprintf(var_name, MAX_STRING_SIZE, "control");
-        pool_vars[0].data = &pl->pl_control;
-        pool_vars[0].read_fptr = lprocfs_rd_uint;
-        pool_vars[0].write_fptr = lprocfs_wr_uint;
-        lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
-
         snprintf(var_name, MAX_STRING_SIZE, "grant_speed");
-        pool_vars[0].data = &pl->pl_grant_speed;
-        pool_vars[0].read_fptr = lprocfs_rd_atomic;
+        pool_vars[0].data = pl;
+        pool_vars[0].read_fptr = lprocfs_rd_grant_speed;
         lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
 
         snprintf(var_name, MAX_STRING_SIZE, "cancel_rate");
@@ -519,21 +785,20 @@ static int ldlm_pool_proc_init(struct ldlm_pool *pl)
         lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
 
         snprintf(var_name, MAX_STRING_SIZE, "grant_plan");
-        pool_vars[0].data = &pl->pl_grant_plan;
-        pool_vars[0].read_fptr = lprocfs_rd_atomic;
+        pool_vars[0].data = pl;
+        pool_vars[0].read_fptr = lprocfs_rd_grant_plan;
         lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
 
-        snprintf(var_name, MAX_STRING_SIZE, "grant_step");
-        pool_vars[0].data = &pl->pl_grant_step;
-        pool_vars[0].read_fptr = lprocfs_rd_atomic;
-        if (ns_is_server(ns))
-                pool_vars[0].write_fptr = lprocfs_wr_atomic;
+        snprintf(var_name, MAX_STRING_SIZE, "recalc_period");
+        pool_vars[0].data = pl;
+        pool_vars[0].read_fptr = lprocfs_rd_recalc_period;
+        pool_vars[0].write_fptr = lprocfs_wr_recalc_period;
         lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
 
         snprintf(var_name, MAX_STRING_SIZE, "lock_volume_factor");
         pool_vars[0].data = &pl->pl_lock_volume_factor;
-        pool_vars[0].read_fptr = lprocfs_rd_uint;
-        pool_vars[0].write_fptr = lprocfs_wr_uint;
+        pool_vars[0].read_fptr = lprocfs_rd_atomic;
+        pool_vars[0].write_fptr = lprocfs_wr_atomic;
         lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
 
         snprintf(var_name, MAX_STRING_SIZE, "state");
@@ -549,9 +814,11 @@ static int ldlm_pool_proc_init(struct ldlm_pool *pl)
         lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
                              LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
                              "granted", "locks");
-        lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT, 0,
+        lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
+                             LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
                              "grant", "locks");
-        lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT, 0,
+        lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
+                             LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
                              "cancel", "locks");
         lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
                              LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
@@ -574,7 +841,10 @@ static int ldlm_pool_proc_init(struct ldlm_pool *pl)
         lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
                              LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
                              "recalc_freed", "locks");
-        lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
+        lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
+                             LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
+                             "recalc_timing", "sec");
+       rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
 
         EXIT;
 out_free_name:
@@ -601,34 +871,33 @@ static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
                    int idx, ldlm_side_t client)
 {
-        int rc;
-        ENTRY;
+       int rc;
+       ENTRY;
 
-        spin_lock_init(&pl->pl_lock);
-        atomic_set(&pl->pl_granted, 0);
+       spin_lock_init(&pl->pl_lock);
+        cfs_atomic_set(&pl->pl_granted, 0);
         pl->pl_recalc_time = cfs_time_current_sec();
-        atomic_set(&pl->pl_lock_volume_factor, 1);
+        cfs_atomic_set(&pl->pl_lock_volume_factor, 1);
 
-        atomic_set(&pl->pl_grant_rate, 0);
-        atomic_set(&pl->pl_cancel_rate, 0);
-        atomic_set(&pl->pl_grant_speed, 0);
-        pl->pl_control = LDLM_POOL_CTL_FULL;
-        atomic_set(&pl->pl_grant_step, LDLM_POOL_GSP);
-        atomic_set(&pl->pl_grant_plan, LDLM_POOL_GP(LDLM_POOL_HOST_L));
+        cfs_atomic_set(&pl->pl_grant_rate, 0);
+        cfs_atomic_set(&pl->pl_cancel_rate, 0);
+        pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
 
         snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
-                 ns->ns_name, idx);
+                 ldlm_ns_name(ns), idx);
 
         if (client == LDLM_NAMESPACE_SERVER) {
                 pl->pl_ops = &ldlm_srv_pool_ops;
                 ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L);
-                ldlm_pool_set_slv(pl, ldlm_pool_slv_max(LDLM_POOL_HOST_L));
+                pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD;
+                pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L);
         } else {
-                ldlm_pool_set_slv(pl, 1);
                 ldlm_pool_set_limit(pl, 1);
+                pl->pl_server_lock_volume = 0;
                 pl->pl_ops = &ldlm_cli_pool_ops;
+                pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
         }
-
+        pl->pl_client_lock_volume = 0;
         rc = ldlm_pool_proc_init(pl);
         if (rc)
                 RETURN(rc);
@@ -643,182 +912,294 @@ void ldlm_pool_fini(struct ldlm_pool *pl)
 {
         ENTRY;
         ldlm_pool_proc_fini(pl);
-        pl->pl_ops = NULL;
+
+        /*
+         * Pool should not be used after this point. We can't free it here as
+         * it lives in struct ldlm_namespace, but still interested in catching
+         * any abnormal using cases.
+         */
+        POISON(pl, 0x5a, sizeof(*pl));
         EXIT;
 }
 EXPORT_SYMBOL(ldlm_pool_fini);
 
+/**
+ * Add new taken ldlm lock \a lock into pool \a pl accounting.
+ */
 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
 {
-        ENTRY;
-        atomic_inc(&pl->pl_granted);
-        atomic_inc(&pl->pl_grant_rate);
-        atomic_inc(&pl->pl_grant_speed);
+        /*
+         * FLOCK locks are special in a sense that they are almost never
+         * cancelled, instead special kind of lock is used to drop them.
+         * also there is no LRU for flock locks, so no point in tracking
+         * them anyway.
+         */
+        if (lock->l_resource->lr_type == LDLM_FLOCK)
+                return;
 
+        cfs_atomic_inc(&pl->pl_granted);
+        cfs_atomic_inc(&pl->pl_grant_rate);
         lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
-        /* Do not do pool recalc for client side as all locks which
-         * potentially may be canceled has already been packed into 
+        /*
+         * Do not do pool recalc for client side as all locks which
+         * potentially may be canceled has already been packed into
          * enqueue/cancel rpc. Also we do not want to run out of stack
-         * with too long call paths. */
+         * with too long call paths.
+         */
         if (ns_is_server(ldlm_pl2ns(pl)))
                 ldlm_pool_recalc(pl);
-        EXIT;
 }
 EXPORT_SYMBOL(ldlm_pool_add);
 
+/**
+ * Remove ldlm lock \a lock from pool \a pl accounting.
+ */
 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
 {
-        ENTRY;
-        LASSERT(atomic_read(&pl->pl_granted) > 0);
-        atomic_dec(&pl->pl_granted);
-        atomic_inc(&pl->pl_cancel_rate);
-        atomic_dec(&pl->pl_grant_speed);
-        
+        /*
+         * Filter out FLOCK locks. Read above comment in ldlm_pool_add().
+         */
+        if (lock->l_resource->lr_type == LDLM_FLOCK)
+                return;
+
+        LASSERT(cfs_atomic_read(&pl->pl_granted) > 0);
+        cfs_atomic_dec(&pl->pl_granted);
+        cfs_atomic_inc(&pl->pl_cancel_rate);
+
         lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
 
         if (ns_is_server(ldlm_pl2ns(pl)))
                 ldlm_pool_recalc(pl);
-        EXIT;
 }
 EXPORT_SYMBOL(ldlm_pool_del);
 
-/* ->pl_lock should be taken. */
+/**
+ * Returns current \a pl SLV.
+ *
+ * \pre ->pl_lock is not locked.
+ */
 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
 {
-        return pl->pl_server_lock_volume;
+       __u64 slv;
+       spin_lock(&pl->pl_lock);
+       slv = pl->pl_server_lock_volume;
+       spin_unlock(&pl->pl_lock);
+       return slv;
 }
 EXPORT_SYMBOL(ldlm_pool_get_slv);
 
-/* ->pl_lock should be taken. */
+/**
+ * Sets passed \a slv to \a pl.
+ *
+ * \pre ->pl_lock is not locked.
+ */
 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
 {
-        pl->pl_server_lock_volume = slv;
+       spin_lock(&pl->pl_lock);
+       pl->pl_server_lock_volume = slv;
+       spin_unlock(&pl->pl_lock);
 }
 EXPORT_SYMBOL(ldlm_pool_set_slv);
 
+/**
+ * Returns current \a pl CLV.
+ *
+ * \pre ->pl_lock is not locked.
+ */
+__u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
+{
+       __u64 slv;
+       spin_lock(&pl->pl_lock);
+       slv = pl->pl_client_lock_volume;
+       spin_unlock(&pl->pl_lock);
+       return slv;
+}
+EXPORT_SYMBOL(ldlm_pool_get_clv);
+
+/**
+ * Sets passed \a clv to \a pl.
+ *
+ * \pre ->pl_lock is not locked.
+ */
+void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
+{
+       spin_lock(&pl->pl_lock);
+       pl->pl_client_lock_volume = clv;
+       spin_unlock(&pl->pl_lock);
+}
+EXPORT_SYMBOL(ldlm_pool_set_clv);
+
+/**
+ * Returns current \a pl limit.
+ */
 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
 {
-        return atomic_read(&pl->pl_limit);
+        return cfs_atomic_read(&pl->pl_limit);
 }
 EXPORT_SYMBOL(ldlm_pool_get_limit);
 
+/**
+ * Sets passed \a limit to \a pl.
+ */
 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
 {
-        atomic_set(&pl->pl_limit, limit);
+        cfs_atomic_set(&pl->pl_limit, limit);
 }
 EXPORT_SYMBOL(ldlm_pool_set_limit);
 
-/* Server side is only enabled for kernel space for now. */
+/**
+ * Returns current LVF from \a pl.
+ */
+__u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
+{
+        return cfs_atomic_read(&pl->pl_lock_volume_factor);
+}
+EXPORT_SYMBOL(ldlm_pool_get_lvf);
+
 #ifdef __KERNEL__
 static int ldlm_pool_granted(struct ldlm_pool *pl)
 {
-        return atomic_read(&pl->pl_granted);
+        return cfs_atomic_read(&pl->pl_granted);
 }
 
 static struct ptlrpc_thread *ldlm_pools_thread;
-static struct shrinker *ldlm_pools_srv_shrinker;
-static struct shrinker *ldlm_pools_cli_shrinker;
+static struct cfs_shrinker *ldlm_pools_srv_shrinker;
+static struct cfs_shrinker *ldlm_pools_cli_shrinker;
 static struct completion ldlm_pools_comp;
 
-void ldlm_pools_wakeup(void)
-{
-        ENTRY;
-        if (ldlm_pools_thread == NULL)
-                return;
-        ldlm_pools_thread->t_flags |= SVC_EVENT;
-        cfs_waitq_signal(&ldlm_pools_thread->t_ctl_waitq);
-        EXIT;
-}
-EXPORT_SYMBOL(ldlm_pools_wakeup);
-
-/* Cancel @nr locks from all namespaces (if possible). Returns number of
+/*
+ * Cancel \a nr locks from all namespaces (if possible). Returns number of
  * cached locks after shrink is finished. All namespaces are asked to
- * cancel approximately equal amount of locks. */
-static int ldlm_pools_shrink(ldlm_side_t client, int nr, 
+ * cancel approximately equal amount of locks to keep balancing.
+ */
+static int ldlm_pools_shrink(ldlm_side_t client, int nr,
                              unsigned int gfp_mask)
 {
         int total = 0, cached = 0, nr_ns;
         struct ldlm_namespace *ns;
+       struct ldlm_namespace *ns_old = NULL; /* loop detection */
+        void *cookie;
 
-        if (nr != 0 && !(gfp_mask & __GFP_FS))
+        if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
+            !(gfp_mask & __GFP_FS))
                 return -1;
 
-        CDEBUG(D_DLMTRACE, "request to shrink %d %s locks from all pools\n",
+        CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n",
                nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
 
-        /* Find out how many resources we may release. */
-        for (nr_ns = atomic_read(ldlm_namespace_nr(client)); 
-             nr_ns > 0; nr_ns--) 
+        cookie = cl_env_reenter();
+
+        /*
+         * Find out how many resources we may release.
+         */
+       for (nr_ns = ldlm_namespace_nr_read(client);
+            nr_ns > 0; nr_ns--)
         {
-                mutex_down(ldlm_namespace_lock(client));
-                if (list_empty(ldlm_namespace_list(client))) {
-                        mutex_up(ldlm_namespace_lock(client));
+               mutex_lock(ldlm_namespace_lock(client));
+                if (cfs_list_empty(ldlm_namespace_list(client))) {
+                       mutex_unlock(ldlm_namespace_lock(client));
+                        cl_env_reexit(cookie);
                         return 0;
                 }
-                ns = ldlm_namespace_first(client);
+                ns = ldlm_namespace_first_locked(client);
+
+               if (ns == ns_old) {
+                       mutex_unlock(ldlm_namespace_lock(client));
+                       break;
+               }
+
+               if (ldlm_ns_empty(ns)) {
+                       ldlm_namespace_move_to_inactive_locked(ns, client);
+                       mutex_unlock(ldlm_namespace_lock(client));
+                       continue;
+               }
+
+               if (ns_old == NULL)
+                       ns_old = ns;
+
                 ldlm_namespace_get(ns);
-                ldlm_namespace_move(ns, client);
-                mutex_up(ldlm_namespace_lock(client));
+                ldlm_namespace_move_to_active_locked(ns, client);
+               mutex_unlock(ldlm_namespace_lock(client));
                 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
-                ldlm_namespace_put(ns, 1);
+                ldlm_namespace_put(ns);
         }
-        if (nr == 0 || total == 0)
+
+        if (nr == 0 || total == 0) {
+                cl_env_reexit(cookie);
                 return total;
+        }
 
-        /* Shrink at least ldlm_namespace_nr(client) namespaces. */
-        for (nr_ns = atomic_read(ldlm_namespace_nr(client)); 
-             nr_ns > 0; nr_ns--) 
+        /*
+         * Shrink at least ldlm_namespace_nr(client) namespaces.
+         */
+       for (nr_ns = ldlm_namespace_nr_read(client) - nr_ns;
+            nr_ns > 0; nr_ns--)
         {
                 int cancel, nr_locks;
 
-                /* Do not call shrink under ldlm_namespace_lock(client) */
-                mutex_down(ldlm_namespace_lock(client));
-                if (list_empty(ldlm_namespace_list(client))) {
-                        mutex_up(ldlm_namespace_lock(client));
-                        /* If list is empty, we can't return any @cached > 0,
+                /*
+                 * Do not call shrink under ldlm_namespace_lock(client)
+                 */
+               mutex_lock(ldlm_namespace_lock(client));
+                if (cfs_list_empty(ldlm_namespace_list(client))) {
+                       mutex_unlock(ldlm_namespace_lock(client));
+                        /*
+                         * If list is empty, we can't return any @cached > 0,
                          * that probably would cause needless shrinker
-                         * call. */
+                         * call.
+                         */
                         cached = 0;
                         break;
                 }
-                ns = ldlm_namespace_first(client);
+                ns = ldlm_namespace_first_locked(client);
                 ldlm_namespace_get(ns);
-                ldlm_namespace_move(ns, client);
-                mutex_up(ldlm_namespace_lock(client));
-                
+                ldlm_namespace_move_to_active_locked(ns, client);
+               mutex_unlock(ldlm_namespace_lock(client));
+
                 nr_locks = ldlm_pool_granted(&ns->ns_pool);
                 cancel = 1 + nr_locks * nr / total;
                 ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
                 cached += ldlm_pool_granted(&ns->ns_pool);
-                ldlm_namespace_put(ns, 1);
+                ldlm_namespace_put(ns);
         }
-        return cached;
+        cl_env_reexit(cookie);
+        /* we only decrease the SLV in server pools shrinker, return -1 to
+         * kernel to avoid needless loop. LU-1128 */
+        return (client == LDLM_NAMESPACE_SERVER) ? -1 : cached;
 }
 
-static int ldlm_pools_srv_shrink(int nr, unsigned int gfp_mask)
+static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
 {
-        return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER, nr, gfp_mask);
+        return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER,
+                                 shrink_param(sc, nr_to_scan),
+                                 shrink_param(sc, gfp_mask));
 }
 
-static int ldlm_pools_cli_shrink(int nr, unsigned int gfp_mask)
+static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
 {
-        return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT, nr, gfp_mask);
+        return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT,
+                                 shrink_param(sc, nr_to_scan),
+                                 shrink_param(sc, gfp_mask));
 }
 
-void ldlm_pools_recalc(ldlm_side_t client)
+int ldlm_pools_recalc(ldlm_side_t client)
 {
         __u32 nr_l = 0, nr_p = 0, l;
         struct ldlm_namespace *ns;
+        struct ldlm_namespace *ns_old = NULL;
         int nr, equal = 0;
+       int time = 50; /* seconds of sleep if no active namespaces */
 
-        /* No need to setup pool limit for client pools. */
+        /*
+         * No need to setup pool limit for client pools.
+         */
         if (client == LDLM_NAMESPACE_SERVER) {
-                /* Check all modest namespaces first. */
-                mutex_down(ldlm_namespace_lock(client));
-                list_for_each_entry(ns, ldlm_namespace_list(client), 
-                                    ns_list_chain) 
+                /*
+                 * Check all modest namespaces first.
+                 */
+               mutex_lock(ldlm_namespace_lock(client));
+                cfs_list_for_each_entry(ns, ldlm_namespace_list(client),
+                                        ns_list_chain)
                 {
                         if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
                                 continue;
@@ -827,16 +1208,20 @@ void ldlm_pools_recalc(ldlm_side_t client)
                         if (l == 0)
                                 l = 1;
 
-                        /* Set the modest pools limit equal to their avg granted
-                         * locks + 5%. */
-                        l += dru(l * LDLM_POOLS_MODEST_MARGIN, 100);
+                        /*
+                         * Set the modest pools limit equal to their avg granted
+                         * locks + ~6%.
+                         */
+                        l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0);
                         ldlm_pool_setup(&ns->ns_pool, l);
                         nr_l += l;
                         nr_p++;
                 }
 
-                /* Make sure that modest namespaces did not eat more that 2/3 
-                 * of limit */
+                /*
+                 * Make sure that modest namespaces did not eat more that 2/3
+                 * of limit.
+                 */
                 if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) {
                         CWARN("\"Modest\" pools eat out 2/3 of server locks "
                               "limit (%d of %lu). This means that you have too "
@@ -845,52 +1230,109 @@ void ldlm_pools_recalc(ldlm_side_t client)
                         equal = 1;
                 }
 
-                /* The rest is given to greedy namespaces. */
-                list_for_each_entry(ns, ldlm_namespace_list(client), 
-                                    ns_list_chain) 
+                /*
+                 * The rest is given to greedy namespaces.
+                 */
+                cfs_list_for_each_entry(ns, ldlm_namespace_list(client),
+                                        ns_list_chain)
                 {
                         if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
                                 continue;
 
                         if (equal) {
-                                /* In the case 2/3 locks are eaten out by
+                                /*
+                                 * In the case 2/3 locks are eaten out by
                                  * modest pools, we re-setup equal limit
-                                 * for _all_ pools. */
+                                 * for _all_ pools.
+                                 */
                                 l = LDLM_POOL_HOST_L /
-                                        atomic_read(ldlm_namespace_nr(client));
+                                       ldlm_namespace_nr_read(client);
                         } else {
-                                /* All the rest of greedy pools will have
-                                 * all locks in equal parts.*/
+                                /*
+                                 * All the rest of greedy pools will have
+                                 * all locks in equal parts.
+                                 */
                                 l = (LDLM_POOL_HOST_L - nr_l) /
-                                        (atomic_read(ldlm_namespace_nr(client)) -
+                                       (ldlm_namespace_nr_read(client) -
                                          nr_p);
                         }
                         ldlm_pool_setup(&ns->ns_pool, l);
                 }
-                mutex_up(ldlm_namespace_lock(client));
+               mutex_unlock(ldlm_namespace_lock(client));
         }
 
-        /* Recalc at least ldlm_namespace_nr(client) namespaces. */
-        for (nr = atomic_read(ldlm_namespace_nr(client)); nr > 0; nr--) {
-                /* Lock the list, get first @ns in the list, getref, move it
+        /*
+         * Recalc at least ldlm_namespace_nr(client) namespaces.
+         */
+       for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) {
+                int     skip;
+                /*
+                 * Lock the list, get first @ns in the list, getref, move it
                  * to the tail, unlock and call pool recalc. This way we avoid
                  * calling recalc under @ns lock what is really good as we get
                  * rid of potential deadlock on client nodes when canceling
-                 * locks synchronously. */
-                mutex_down(ldlm_namespace_lock(client));
-                if (list_empty(ldlm_namespace_list(client))) {
-                        mutex_up(ldlm_namespace_lock(client));
-                        break;
-                }
-                ns = ldlm_namespace_first(client);
-                ldlm_namespace_get(ns);
-                ldlm_namespace_move(ns, client);
-                mutex_up(ldlm_namespace_lock(client));
-
-                /* After setup is done - recalc the pool. */
-                ldlm_pool_recalc(&ns->ns_pool);
-                ldlm_namespace_put(ns, 1);
+                 * locks synchronously.
+                 */
+               mutex_lock(ldlm_namespace_lock(client));
+               if (cfs_list_empty(ldlm_namespace_list(client))) {
+                       mutex_unlock(ldlm_namespace_lock(client));
+                       break;
+               }
+               ns = ldlm_namespace_first_locked(client);
+
+               if (ns_old == ns) { /* Full pass complete */
+                       mutex_unlock(ldlm_namespace_lock(client));
+                       break;
+               }
+
+               /* We got an empty namespace, need to move it back to inactive
+                * list.
+                * The race with parallel resource creation is fine:
+                * - If they do namespace_get before our check, we fail the
+                *   check and they move this item to the end of the list anyway
+                * - If we do the check and then they do namespace_get, then
+                *   we move the namespace to inactive and they will move
+                *   it back to active (synchronised by the lock, so no clash
+                *   there).
+                */
+               if (ldlm_ns_empty(ns)) {
+                       ldlm_namespace_move_to_inactive_locked(ns, client);
+                       mutex_unlock(ldlm_namespace_lock(client));
+                       continue;
+               }
+
+               if (ns_old == NULL)
+                       ns_old = ns;
+
+               spin_lock(&ns->ns_lock);
+               /*
+                * skip ns which is being freed, and we don't want to increase
+                * its refcount again, not even temporarily. bz21519 & LU-499.
+                */
+               if (ns->ns_stopping) {
+                       skip = 1;
+               } else {
+                       skip = 0;
+                       ldlm_namespace_get(ns);
+               }
+               spin_unlock(&ns->ns_lock);
+
+               ldlm_namespace_move_to_active_locked(ns, client);
+               mutex_unlock(ldlm_namespace_lock(client));
+
+               /*
+                * After setup is done - recalc the pool.
+                */
+               if (!skip) {
+                       int ttime = ldlm_pool_recalc(&ns->ns_pool);
+
+                       if (ttime < time)
+                               time = ttime;
+
+                       ldlm_namespace_put(ns);
+               }
         }
+       return time;
 }
 EXPORT_SYMBOL(ldlm_pools_recalc);
 
@@ -898,10 +1340,11 @@ static int ldlm_pools_thread_main(void *arg)
 {
         struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
         char *t_name = "ldlm_poold";
+       int s_time, c_time;
         ENTRY;
 
         cfs_daemonize(t_name);
-        thread->t_flags = SVC_RUNNING;
+        thread_set_flags(thread, SVC_RUNNING);
         cfs_waitq_signal(&thread->t_ctl_waitq);
 
         CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n",
@@ -910,33 +1353,36 @@ static int ldlm_pools_thread_main(void *arg)
         while (1) {
                 struct l_wait_info lwi;
 
-                /* Recal all pools on this tick. */
-                ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
-                ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
-                
-                /* Wait until the next check time, or until we're
-                 * stopped. */
-                lwi = LWI_TIMEOUT(cfs_time_seconds(LDLM_POOLS_THREAD_PERIOD),
-                                  NULL, NULL);
-                l_wait_event(thread->t_ctl_waitq, (thread->t_flags &
-                                                   (SVC_STOPPING|SVC_EVENT)),
+               /*
+                * Recal all pools on this tick.
+                */
+               s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
+               c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
+
+               /*
+                * Wait until the next check time, or until we're
+                * stopped.
+                */
+               lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)),
+                                 NULL, NULL);
+                l_wait_event(thread->t_ctl_waitq,
+                             thread_is_stopping(thread) ||
+                             thread_is_event(thread),
                              &lwi);
 
-                if (thread->t_flags & SVC_STOPPING) {
-                        thread->t_flags &= ~SVC_STOPPING;
+                if (thread_test_and_clear_flags(thread, SVC_STOPPING))
                         break;
-                } else if (thread->t_flags & SVC_EVENT) {
-                        thread->t_flags &= ~SVC_EVENT;
-                }
+                else
+                        thread_test_and_clear_flags(thread, SVC_EVENT);
         }
 
-        thread->t_flags = SVC_STOPPED;
+        thread_set_flags(thread, SVC_STOPPED);
         cfs_waitq_signal(&thread->t_ctl_waitq);
 
         CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n",
                t_name, cfs_curproc_pid());
 
-        complete_and_exit(&ldlm_pools_comp, 0);
+       complete_and_exit(&ldlm_pools_comp, 0);
 }
 
 static int ldlm_pools_thread_start(void)
@@ -952,13 +1398,15 @@ static int ldlm_pools_thread_start(void)
         if (ldlm_pools_thread == NULL)
                 RETURN(-ENOMEM);
 
-        init_completion(&ldlm_pools_comp);
+       init_completion(&ldlm_pools_comp);
         cfs_waitq_init(&ldlm_pools_thread->t_ctl_waitq);
 
-        /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
-         * just drop the VM and FILES in ptlrpc_daemonize() right away. */
-        rc = cfs_kernel_thread(ldlm_pools_thread_main, ldlm_pools_thread,
-                               CLONE_VM | CLONE_FILES);
+        /*
+         * CLONE_VM and CLONE_FILES just avoid a needless copy, because we
+         * just drop the VM and FILES in cfs_daemonize() right away.
+         */
+        rc = cfs_create_thread(ldlm_pools_thread_main, ldlm_pools_thread,
+                               CFS_DAEMON_FLAGS);
         if (rc < 0) {
                 CERROR("Can't start pool thread, error %d\n",
                        rc);
@@ -967,7 +1415,7 @@ static int ldlm_pools_thread_start(void)
                 RETURN(rc);
         }
         l_wait_event(ldlm_pools_thread->t_ctl_waitq,
-                     (ldlm_pools_thread->t_flags & SVC_RUNNING), &lwi);
+                     thread_is_running(ldlm_pools_thread), &lwi);
         RETURN(0);
 }
 
@@ -980,13 +1428,15 @@ static void ldlm_pools_thread_stop(void)
                 return;
         }
 
-        ldlm_pools_thread->t_flags = SVC_STOPPING;
+        thread_set_flags(ldlm_pools_thread, SVC_STOPPING);
         cfs_waitq_signal(&ldlm_pools_thread->t_ctl_waitq);
 
-        /* Make sure that pools thread is finished before freeing @thread.
+        /*
+         * Make sure that pools thread is finished before freeing @thread.
          * This fixes possible race and oops due to accessing freed memory
-         * in pools thread. */
-        wait_for_completion(&ldlm_pools_comp);
+         * in pools thread.
+         */
+       wait_for_completion(&ldlm_pools_comp);
         OBD_FREE_PTR(ldlm_pools_thread);
         ldlm_pools_thread = NULL;
         EXIT;
@@ -999,10 +1449,12 @@ int ldlm_pools_init(void)
 
         rc = ldlm_pools_thread_start();
         if (rc == 0) {
-                ldlm_pools_srv_shrinker = set_shrinker(DEFAULT_SEEKS,
-                                                       ldlm_pools_srv_shrink);
-                ldlm_pools_cli_shrinker = set_shrinker(DEFAULT_SEEKS,
-                                                       ldlm_pools_cli_shrink);
+                ldlm_pools_srv_shrinker =
+                        cfs_set_shrinker(CFS_DEFAULT_SEEKS,
+                                         ldlm_pools_srv_shrink);
+                ldlm_pools_cli_shrinker =
+                        cfs_set_shrinker(CFS_DEFAULT_SEEKS,
+                                         ldlm_pools_cli_shrink);
         }
         RETURN(rc);
 }
@@ -1011,11 +1463,11 @@ EXPORT_SYMBOL(ldlm_pools_init);
 void ldlm_pools_fini(void)
 {
         if (ldlm_pools_srv_shrinker != NULL) {
-                remove_shrinker(ldlm_pools_srv_shrinker);
+                cfs_remove_shrinker(ldlm_pools_srv_shrinker);
                 ldlm_pools_srv_shrinker = NULL;
         }
         if (ldlm_pools_cli_shrinker != NULL) {
-                remove_shrinker(ldlm_pools_cli_shrinker);
+                cfs_remove_shrinker(ldlm_pools_cli_shrinker);
                 ldlm_pools_cli_shrinker = NULL;
         }
         ldlm_pools_thread_stop();
@@ -1024,7 +1476,7 @@ EXPORT_SYMBOL(ldlm_pools_fini);
 #endif /* __KERNEL__ */
 
 #else /* !HAVE_LRU_RESIZE_SUPPORT */
-int ldlm_pool_setup(struct ldlm_pool *pl, __u32 limit)
+int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
 {
         return 0;
 }
@@ -1080,6 +1532,18 @@ void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
 }
 EXPORT_SYMBOL(ldlm_pool_set_slv);
 
+__u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
+{
+        return 1;
+}
+EXPORT_SYMBOL(ldlm_pool_get_clv);
+
+void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
+{
+        return;
+}
+EXPORT_SYMBOL(ldlm_pool_set_clv);
+
 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
 {
         return 0;
@@ -1092,6 +1556,12 @@ void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
 }
 EXPORT_SYMBOL(ldlm_pool_set_limit);
 
+__u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
+{
+        return 0;
+}
+EXPORT_SYMBOL(ldlm_pool_get_lvf);
+
 int ldlm_pools_init(void)
 {
         return 0;
@@ -1104,15 +1574,9 @@ void ldlm_pools_fini(void)
 }
 EXPORT_SYMBOL(ldlm_pools_fini);
 
-void ldlm_pools_wakeup(void)
+int ldlm_pools_recalc(ldlm_side_t client)
 {
-        return;
-}
-EXPORT_SYMBOL(ldlm_pools_wakeup);
-
-void ldlm_pools_recalc(ldlm_side_t client)
-{
-        return;
+        return 0;
 }
 EXPORT_SYMBOL(ldlm_pools_recalc);
 #endif /* HAVE_LRU_RESIZE_SUPPORT */