4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_pool.c
34 * Author: Yury Umanets <umka@clusterfs.com>
38 * Idea of this code is rather simple. Each second, for each server namespace
39 * we have SLV - server lock volume which is calculated on current number of
40 * granted locks, grant speed for past period, etc - that is, locking load.
41 * This SLV number may be thought as a flow definition for simplicity. It is
42 * sent to clients with each occasion to let them know what is current load
43 * situation on the server. By default, at the beginning, SLV on server is
44 * set max value which is calculated as the following: allow to one client
45 * have all locks of limit ->pl_limit for 10h.
47 * Next, on clients, number of cached locks is not limited artificially in any
48 * way as it was before. Instead, client calculates CLV, that is, client lock
49 * volume for each lock and compares it with last SLV from the server. CLV is
50 * calculated as the number of locks in LRU * lock live time in seconds. If
51 * CLV > SLV - lock is canceled.
53 * Client has LVF, that is, lock volume factor which regulates how much sensitive
54 * client should be about last SLV from server. The higher LVF is the more locks
55 * will be canceled on client. Default value for it is 1. Setting LVF to 2 means
56 * that client will cancel locks 2 times faster.
58 * Locks on a client will be canceled more intensively in these cases:
59 * (1) if SLV is smaller, that is, load is higher on the server;
60 * (2) client has a lot of locks (the more locks are held by client, the bigger
61 * chances that some of them should be canceled);
62 * (3) client has old locks (taken some time ago);
64 * Thus, according to flow paradigm that we use for better understanding SLV,
65 * CLV is the volume of particle in flow described by SLV. According to this,
66 * if flow is getting thinner, more and more particles become outside of it and
67 * as particles are locks, they should be canceled.
69 * General idea of this belongs to Vitaly Fertman (vitaly@clusterfs.com). Andreas
70 * Dilger (adilger@clusterfs.com) proposed few nice ideas like using LVF and many
71 * cleanups. Flow definition to allow more easy understanding of the logic belongs
72 * to Nikita Danilov (nikita@clusterfs.com) as well as many cleanups and fixes.
73 * And design and implementation are done by Yury Umanets (umka@clusterfs.com).
75 * Glossary for terms used:
77 * pl_limit - Number of allowed locks in pool. Applies to server and client
80 * pl_granted - Number of granted locks (calculated);
81 * pl_grant_rate - Number of granted locks for last T (calculated);
82 * pl_cancel_rate - Number of canceled locks for last T (calculated);
83 * pl_grant_speed - Grant speed (GR - CR) for last T (calculated);
84 * pl_grant_plan - Planned number of granted locks for next T (calculated);
85 * pl_server_lock_volume - Current server lock volume (calculated);
87 * As it may be seen from list above, we have few possible tunables which may
88 * affect behavior much. They all may be modified via sysfs. However, they also
89 * give a possibility for constructing few pre-defined behavior policies. If
90 * none of predefines is suitable for a working pattern being used, new one may
91 * be "constructed" via sysfs tunables.
94 #define DEBUG_SUBSYSTEM S_LDLM
96 #include <linux/kthread.h>
97 #include <lustre_dlm.h>
98 #include <cl_object.h>
99 #include <obd_class.h>
100 #include <obd_support.h>
101 #include "ldlm_internal.h"
103 #ifdef HAVE_LRU_RESIZE_SUPPORT
106 * 50 ldlm locks for 1MB of RAM.
108 #define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_SHIFT)) * 50)
111 * Maximal possible grant step plan in %.
113 #define LDLM_POOL_MAX_GSP (30)
116 * Minimal possible grant step plan in %.
118 #define LDLM_POOL_MIN_GSP (1)
121 * This controls the speed of reaching LDLM_POOL_MAX_GSP
122 * with increasing thread period.
124 #define LDLM_POOL_GSP_STEP_SHIFT (2)
127 * LDLM_POOL_GSP% of all locks is default GP.
129 #define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100)
132 * Max age for locks on clients.
134 #define LDLM_POOL_MAX_AGE (36000)
137 * The granularity of SLV calculation.
139 #define LDLM_POOL_SLV_SHIFT (10)
141 extern struct proc_dir_entry *ldlm_ns_proc_dir;
143 static inline __u64 dru(__u64 val, __u32 shift, int round_up)
145 return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift;
148 static inline __u64 ldlm_pool_slv_max(__u32 L)
151 * Allow to have all locks for 1 client for 10 hrs.
152 * Formula is the following: limit * 10h / 1 client.
154 __u64 lim = (__u64)L * LDLM_POOL_MAX_AGE / 1;
158 static inline __u64 ldlm_pool_slv_min(__u32 L)
164 LDLM_POOL_FIRST_STAT = 0,
165 LDLM_POOL_GRANTED_STAT = LDLM_POOL_FIRST_STAT,
166 LDLM_POOL_GRANT_STAT,
167 LDLM_POOL_CANCEL_STAT,
168 LDLM_POOL_GRANT_RATE_STAT,
169 LDLM_POOL_CANCEL_RATE_STAT,
170 LDLM_POOL_GRANT_PLAN_STAT,
172 LDLM_POOL_SHRINK_REQTD_STAT,
173 LDLM_POOL_SHRINK_FREED_STAT,
174 LDLM_POOL_RECALC_STAT,
175 LDLM_POOL_TIMING_STAT,
179 static inline struct ldlm_namespace *ldlm_pl2ns(struct ldlm_pool *pl)
181 return container_of(pl, struct ldlm_namespace, ns_pool);
185 * Calculates suggested grant_step in % of available locks for passed
186 * \a period. This is later used in grant_plan calculations.
188 static inline int ldlm_pool_t2gsp(unsigned int t)
191 * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP
192 * and up to 30% for anything higher than LDLM_POOL_GSP_STEP.
194 * How this will affect execution is the following:
196 * - for thread period 1s we will have grant_step 1% which good from
197 * pov of taking some load off from server and push it out to clients.
198 * This is like that because 1% for grant_step means that server will
199 * not allow clients to get lots of locks in short period of time and
200 * keep all old locks in their caches. Clients will always have to
201 * get some locks back if they want to take some new;
203 * - for thread period 10s (which is default) we will have 23% which
204 * means that clients will have enough of room to take some new locks
205 * without getting some back. All locks from this 23% which were not
206 * taken by clients in current period will contribute in SLV growing.
207 * SLV growing means more locks cached on clients until limit or grant
210 return LDLM_POOL_MAX_GSP -
211 ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >>
212 (t >> LDLM_POOL_GSP_STEP_SHIFT));
215 static inline int ldlm_pool_granted(struct ldlm_pool *pl)
217 return atomic_read(&pl->pl_granted);
221 * Recalculates next grant limit on passed \a pl.
223 * \pre ->pl_lock is locked.
225 static void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl)
227 int granted, grant_step, limit;
229 limit = ldlm_pool_get_limit(pl);
230 granted = ldlm_pool_granted(pl);
232 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
233 grant_step = ((limit - granted) * grant_step) / 100;
234 pl->pl_grant_plan = granted + grant_step;
235 limit = (limit * 5) >> 2;
236 if (pl->pl_grant_plan > limit)
237 pl->pl_grant_plan = limit;
241 * Recalculates next SLV on passed \a pl.
243 * \pre ->pl_lock is locked.
245 static void ldlm_pool_recalc_slv(struct ldlm_pool *pl)
255 slv = pl->pl_server_lock_volume;
256 grant_plan = pl->pl_grant_plan;
257 limit = ldlm_pool_get_limit(pl);
258 granted = ldlm_pool_granted(pl);
259 round_up = granted < limit;
261 grant_usage = max_t(int, limit - (granted - grant_plan), 1);
264 * Find out SLV change factor which is the ratio of grant usage
265 * from limit. SLV changes as fast as the ratio of grant plan
266 * consumption. The more locks from grant plan are not consumed
267 * by clients in last interval (idle time), the faster grows
268 * SLV. And the opposite, the more grant plan is over-consumed
269 * (load time) the faster drops SLV.
271 slv_factor = (grant_usage << LDLM_POOL_SLV_SHIFT);
272 do_div(slv_factor, limit);
273 slv = slv * slv_factor;
274 slv = dru(slv, LDLM_POOL_SLV_SHIFT, round_up);
276 if (slv > ldlm_pool_slv_max(limit)) {
277 slv = ldlm_pool_slv_max(limit);
278 } else if (slv < ldlm_pool_slv_min(limit)) {
279 slv = ldlm_pool_slv_min(limit);
282 pl->pl_server_lock_volume = slv;
286 * Recalculates next stats on passed \a pl.
288 * \pre ->pl_lock is locked.
290 static void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
292 int grant_plan = pl->pl_grant_plan;
293 __u64 slv = pl->pl_server_lock_volume;
294 int granted = ldlm_pool_granted(pl);
295 int grant_rate = atomic_read(&pl->pl_grant_rate);
296 int cancel_rate = atomic_read(&pl->pl_cancel_rate);
298 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT,
300 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
302 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
304 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
306 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
311 * Sets current SLV into obd accessible via ldlm_pl2ns(pl)->ns_obd.
313 static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl)
315 struct obd_device *obd;
318 * Set new SLV in obd field for using it later without accessing the
319 * pool. This is required to avoid race between sending reply to client
320 * with new SLV and cleanup server stack in which we can't guarantee
321 * that namespace is still alive. We know only that obd is alive as
322 * long as valid export is alive.
324 obd = ldlm_pl2ns(pl)->ns_obd;
325 LASSERT(obd != NULL);
326 write_lock(&obd->obd_pool_lock);
327 obd->obd_pool_slv = pl->pl_server_lock_volume;
328 write_unlock(&obd->obd_pool_lock);
332 * Recalculates all pool fields on passed \a pl.
334 * \pre ->pl_lock is not locked.
336 static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
338 time64_t recalc_interval_sec;
341 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time;
342 if (recalc_interval_sec < pl->pl_recalc_period)
345 spin_lock(&pl->pl_lock);
346 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time;
347 if (recalc_interval_sec < pl->pl_recalc_period) {
348 spin_unlock(&pl->pl_lock);
352 * Recalc SLV after last period. This should be done
353 * _before_ recalculating new grant plan.
355 ldlm_pool_recalc_slv(pl);
358 * Make sure that pool informed obd of last SLV changes.
360 ldlm_srv_pool_push_slv(pl);
363 * Update grant_plan for new period.
365 ldlm_pool_recalc_grant_plan(pl);
367 pl->pl_recalc_time = ktime_get_real_seconds();
368 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
369 recalc_interval_sec);
370 spin_unlock(&pl->pl_lock);
375 * This function is used on server side as main entry point for memory
376 * pressure handling. It decreases SLV on \a pl according to passed
377 * \a nr and \a gfp_mask.
379 * Our goal here is to decrease SLV such a way that clients hold \a nr
380 * locks smaller in next 10h.
382 static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
383 int nr, gfp_t gfp_mask)
388 * VM is asking how many entries may be potentially freed.
391 return ldlm_pool_granted(pl);
394 * Client already canceled locks but server is already in shrinker
395 * and can't cancel anything. Let's catch this race.
397 if (ldlm_pool_granted(pl) == 0)
400 spin_lock(&pl->pl_lock);
403 * We want shrinker to possibly cause cancellation of @nr locks from
404 * clients or grant approximately @nr locks smaller next intervals.
406 * This is why we decreased SLV by @nr. This effect will only be as
407 * long as one re-calc interval (1s these days) and this should be
408 * enough to pass this decreased SLV to all clients. On next recalc
409 * interval pool will either increase SLV if locks load is not high
410 * or will keep on same level or even decrease again, thus, shrinker
411 * decreased SLV will affect next recalc intervals and this way will
412 * make locking load lower.
414 if (nr < pl->pl_server_lock_volume) {
415 pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr;
417 limit = ldlm_pool_get_limit(pl);
418 pl->pl_server_lock_volume = ldlm_pool_slv_min(limit);
422 * Make sure that pool informed obd of last SLV changes.
424 ldlm_srv_pool_push_slv(pl);
425 spin_unlock(&pl->pl_lock);
428 * We did not really free any memory here so far, it only will be
429 * freed later may be, so that we return 0 to not confuse VM.
435 * Setup server side pool \a pl with passed \a limit.
437 static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit)
439 struct obd_device *obd;
441 obd = ldlm_pl2ns(pl)->ns_obd;
442 LASSERT(obd != NULL && obd != LP_POISON);
443 LASSERT(obd->obd_type != LP_POISON);
444 write_lock(&obd->obd_pool_lock);
445 obd->obd_pool_limit = limit;
446 write_unlock(&obd->obd_pool_lock);
448 ldlm_pool_set_limit(pl, limit);
453 * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl.
455 static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
457 struct obd_device *obd;
460 * Get new SLV and Limit from obd which is updated with coming
463 obd = ldlm_pl2ns(pl)->ns_obd;
464 LASSERT(obd != NULL);
465 read_lock(&obd->obd_pool_lock);
466 pl->pl_server_lock_volume = obd->obd_pool_slv;
467 ldlm_pool_set_limit(pl, obd->obd_pool_limit);
468 read_unlock(&obd->obd_pool_lock);
472 * Recalculates client size pool \a pl according to current SLV and Limit.
474 static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
476 time64_t recalc_interval_sec;
480 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time;
481 if (recalc_interval_sec < pl->pl_recalc_period)
484 spin_lock(&pl->pl_lock);
486 * Check if we need to recalc lists now.
488 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time;
489 if (recalc_interval_sec < pl->pl_recalc_period) {
490 spin_unlock(&pl->pl_lock);
495 * Make sure that pool knows last SLV and Limit from obd.
497 ldlm_cli_pool_pop_slv(pl);
498 spin_unlock(&pl->pl_lock);
501 * Do not cancel locks in case lru resize is disabled for this ns.
503 if (!ns_connect_lru_resize(ldlm_pl2ns(pl)))
507 * In the time of canceling locks on client we do not need to maintain
508 * sharp timing, we only want to cancel locks asap according to new SLV.
509 * It may be called when SLV has changed much, this is why we do not
510 * take into account pl->pl_recalc_time here.
512 ret = ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC,
516 spin_lock(&pl->pl_lock);
518 * Time of LRU resizing might be longer than period,
519 * so update after LRU resizing rather than before it.
521 pl->pl_recalc_time = ktime_get_real_seconds();
522 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
523 recalc_interval_sec);
524 spin_unlock(&pl->pl_lock);
529 * This function is main entry point for memory pressure handling on client
530 * side. Main goal of this function is to cancel some number of locks on
531 * passed \a pl according to \a nr and \a gfp_mask.
533 static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
534 int nr, gfp_t gfp_mask)
536 struct ldlm_namespace *ns;
542 * Do not cancel locks in case lru resize is disabled for this ns.
544 if (!ns_connect_lru_resize(ns))
548 * Make sure that pool knows last SLV and Limit from obd.
550 ldlm_cli_pool_pop_slv(pl);
552 spin_lock(&ns->ns_lock);
553 unused = ns->ns_nr_unused;
554 spin_unlock(&ns->ns_lock);
557 return (unused / 100) * sysctl_vfs_cache_pressure;
559 return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_LRU_FLAG_SHRINK);
562 static struct ldlm_pool_ops ldlm_srv_pool_ops = {
563 .po_recalc = ldlm_srv_pool_recalc,
564 .po_shrink = ldlm_srv_pool_shrink,
565 .po_setup = ldlm_srv_pool_setup
568 static struct ldlm_pool_ops ldlm_cli_pool_ops = {
569 .po_recalc = ldlm_cli_pool_recalc,
570 .po_shrink = ldlm_cli_pool_shrink
574 * Pool recalc wrapper. Will call either client or server pool recalc callback
575 * depending what pool \a pl is used.
577 int ldlm_pool_recalc(struct ldlm_pool *pl)
579 time64_t recalc_interval_sec;
582 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time;
583 if (recalc_interval_sec > 0) {
584 spin_lock(&pl->pl_lock);
585 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time;
587 if (recalc_interval_sec > 0) {
589 * Update pool statistics every 1s.
591 ldlm_pool_recalc_stats(pl);
594 * Zero out all rates and speed for the last period.
596 atomic_set(&pl->pl_grant_rate, 0);
597 atomic_set(&pl->pl_cancel_rate, 0);
599 spin_unlock(&pl->pl_lock);
602 if (pl->pl_ops->po_recalc != NULL) {
603 count = pl->pl_ops->po_recalc(pl);
604 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
608 recalc_interval_sec = pl->pl_recalc_time - ktime_get_real_seconds() +
609 pl->pl_recalc_period;
610 if (recalc_interval_sec <= 0) {
611 /* DEBUG: should be re-removed after LU-4536 is fixed */
612 CDEBUG(D_DLMTRACE, "%s: Negative interval(%lld), too short period(%lld)\n",
613 pl->pl_name, recalc_interval_sec,
614 (s64)pl->pl_recalc_period);
616 /* Prevent too frequent recalculation. */
617 recalc_interval_sec = 1;
620 return recalc_interval_sec;
624 * Pool shrink wrapper. Will call either client or server pool recalc callback
625 * depending what pool \a pl is used.
627 int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, gfp_t gfp_mask)
631 if (pl->pl_ops->po_shrink != NULL) {
632 cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask);
634 lprocfs_counter_add(pl->pl_stats,
635 LDLM_POOL_SHRINK_REQTD_STAT,
637 lprocfs_counter_add(pl->pl_stats,
638 LDLM_POOL_SHRINK_FREED_STAT,
640 CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, "
641 "shrunk %d\n", pl->pl_name, nr, cancel);
648 * Pool setup wrapper. Will call either client or server pool recalc callback
649 * depending what pool \a pl is used.
651 * Sets passed \a limit into pool \a pl.
653 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
655 if (pl->pl_ops->po_setup != NULL)
656 return(pl->pl_ops->po_setup(pl, limit));
660 static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
662 int granted, grant_rate, cancel_rate, grant_step;
663 int grant_speed, grant_plan, lvf;
664 struct ldlm_pool *pl = m->private;
668 spin_lock(&pl->pl_lock);
669 slv = pl->pl_server_lock_volume;
670 clv = pl->pl_client_lock_volume;
671 limit = ldlm_pool_get_limit(pl);
672 grant_plan = pl->pl_grant_plan;
673 granted = ldlm_pool_granted(pl);
674 grant_rate = atomic_read(&pl->pl_grant_rate);
675 cancel_rate = atomic_read(&pl->pl_cancel_rate);
676 grant_speed = grant_rate - cancel_rate;
677 lvf = atomic_read(&pl->pl_lock_volume_factor);
678 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
679 spin_unlock(&pl->pl_lock);
681 seq_printf(m, "LDLM pool state (%s):\n"
685 pl->pl_name, slv, clv, lvf);
687 if (ns_is_server(ldlm_pl2ns(pl))) {
688 seq_printf(m, " GSP: %d%%\n", grant_step);
689 seq_printf(m, " GP: %d\n", grant_plan);
692 seq_printf(m, " GR: %d\n CR: %d\n GS: %d\n G: %d\n L: %d\n",
693 grant_rate, cancel_rate, grant_speed,
697 LPROC_SEQ_FOPS_RO(lprocfs_pool_state);
699 static ssize_t grant_speed_show(struct kobject *kobj, struct attribute *attr,
702 struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool,
706 spin_lock(&pl->pl_lock);
707 /* serialize with ldlm_pool_recalc */
708 grant_speed = atomic_read(&pl->pl_grant_rate) -
709 atomic_read(&pl->pl_cancel_rate);
710 spin_unlock(&pl->pl_lock);
711 return sprintf(buf, "%d\n", grant_speed);
713 LUSTRE_RO_ATTR(grant_speed);
715 LDLM_POOL_SYSFS_READER_SHOW(grant_plan, int);
716 LUSTRE_RO_ATTR(grant_plan);
718 LDLM_POOL_SYSFS_READER_SHOW(recalc_period, int);
719 LDLM_POOL_SYSFS_WRITER_STORE(recalc_period, int);
720 LUSTRE_RW_ATTR(recalc_period);
722 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(server_lock_volume, u64);
723 LUSTRE_RO_ATTR(server_lock_volume);
725 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(limit, atomic);
726 LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(limit, atomic);
727 LUSTRE_RW_ATTR(limit);
729 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(granted, atomic);
730 LUSTRE_RO_ATTR(granted);
732 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(cancel_rate, atomic);
733 LUSTRE_RO_ATTR(cancel_rate);
735 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(grant_rate, atomic);
736 LUSTRE_RO_ATTR(grant_rate);
738 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(lock_volume_factor, atomic);
739 LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(lock_volume_factor, atomic);
740 LUSTRE_RW_ATTR(lock_volume_factor);
742 /* These are for pools in /sys/fs/lustre/ldlm/namespaces/.../pool */
743 static struct attribute *ldlm_pl_attrs[] = {
744 &lustre_attr_grant_speed.attr,
745 &lustre_attr_grant_plan.attr,
746 &lustre_attr_recalc_period.attr,
747 &lustre_attr_server_lock_volume.attr,
748 &lustre_attr_limit.attr,
749 &lustre_attr_granted.attr,
750 &lustre_attr_cancel_rate.attr,
751 &lustre_attr_grant_rate.attr,
752 &lustre_attr_lock_volume_factor.attr,
756 static void ldlm_pl_release(struct kobject *kobj)
758 struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool,
760 complete(&pl->pl_kobj_unregister);
763 static struct kobj_type ldlm_pl_ktype = {
764 .default_attrs = ldlm_pl_attrs,
765 .sysfs_ops = &lustre_sysfs_ops,
766 .release = ldlm_pl_release,
769 static int ldlm_pool_sysfs_init(struct ldlm_pool *pl)
771 struct ldlm_namespace *ns = ldlm_pl2ns(pl);
774 init_completion(&pl->pl_kobj_unregister);
775 err = kobject_init_and_add(&pl->pl_kobj, &ldlm_pl_ktype, &ns->ns_kobj,
781 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
783 struct ldlm_namespace *ns = ldlm_pl2ns(pl);
784 struct proc_dir_entry *parent_ns_proc;
785 struct lprocfs_vars pool_vars[2];
786 char *var_name = NULL;
790 OBD_ALLOC(var_name, MAX_STRING_SIZE + 1);
794 parent_ns_proc = ns->ns_proc_dir_entry;
795 if (parent_ns_proc == NULL) {
796 CERROR("%s: proc entry is not initialized\n",
798 GOTO(out_free_name, rc = -EINVAL);
800 pl->pl_proc_dir = lprocfs_register("pool", parent_ns_proc,
802 if (IS_ERR(pl->pl_proc_dir)) {
803 rc = PTR_ERR(pl->pl_proc_dir);
804 pl->pl_proc_dir = NULL;
805 CERROR("%s: cannot create 'pool' proc entry: rc = %d\n",
806 ldlm_ns_name(ns), rc);
807 GOTO(out_free_name, rc);
810 var_name[MAX_STRING_SIZE] = '\0';
811 memset(pool_vars, 0, sizeof(pool_vars));
812 pool_vars[0].name = var_name;
814 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "state", pl,
815 &lprocfs_pool_state_fops);
817 pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT -
818 LDLM_POOL_FIRST_STAT, 0);
820 GOTO(out_free_name, rc = -ENOMEM);
822 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
823 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
825 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
826 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
828 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
829 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
831 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
832 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
833 "grant_rate", "locks/s");
834 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
835 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
836 "cancel_rate", "locks/s");
837 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
838 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
839 "grant_plan", "locks/s");
840 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT,
841 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
843 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT,
844 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
845 "shrink_request", "locks");
846 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT,
847 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
848 "shrink_freed", "locks");
849 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
850 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
851 "recalc_freed", "locks");
852 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
853 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
854 "recalc_timing", "sec");
855 rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
859 OBD_FREE(var_name, MAX_STRING_SIZE + 1);
863 static void ldlm_pool_sysfs_fini(struct ldlm_pool *pl)
865 kobject_put(&pl->pl_kobj);
866 wait_for_completion(&pl->pl_kobj_unregister);
869 static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
871 if (pl->pl_stats != NULL) {
872 lprocfs_free_stats(&pl->pl_stats);
875 if (pl->pl_proc_dir != NULL) {
876 lprocfs_remove(&pl->pl_proc_dir);
877 pl->pl_proc_dir = NULL;
881 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
882 int idx, enum ldlm_side client)
887 spin_lock_init(&pl->pl_lock);
888 atomic_set(&pl->pl_granted, 0);
889 pl->pl_recalc_time = ktime_get_real_seconds();
890 atomic_set(&pl->pl_lock_volume_factor, 1);
892 atomic_set(&pl->pl_grant_rate, 0);
893 atomic_set(&pl->pl_cancel_rate, 0);
894 pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
896 snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
897 ldlm_ns_name(ns), idx);
899 if (client == LDLM_NAMESPACE_SERVER) {
900 pl->pl_ops = &ldlm_srv_pool_ops;
901 ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L);
902 pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD;
903 pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L);
905 ldlm_pool_set_limit(pl, 1);
906 pl->pl_server_lock_volume = 0;
907 pl->pl_ops = &ldlm_cli_pool_ops;
908 pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
910 pl->pl_client_lock_volume = 0;
911 rc = ldlm_pool_proc_init(pl);
915 rc = ldlm_pool_sysfs_init(pl);
919 CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name);
924 void ldlm_pool_fini(struct ldlm_pool *pl)
927 ldlm_pool_sysfs_fini(pl);
928 ldlm_pool_proc_fini(pl);
931 * Pool should not be used after this point. We can't free it here as
932 * it lives in struct ldlm_namespace, but still interested in catching
933 * any abnormal using cases.
935 POISON(pl, 0x5a, sizeof(*pl));
940 * Add new taken ldlm lock \a lock into pool \a pl accounting.
942 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
945 * FLOCK locks are special in a sense that they are almost never
946 * cancelled, instead special kind of lock is used to drop them.
947 * also there is no LRU for flock locks, so no point in tracking
950 * PLAIN locks are used by config and quota, the quantity is small
951 * and usually they are not in LRU.
953 if (lock->l_resource->lr_type == LDLM_FLOCK ||
954 lock->l_resource->lr_type == LDLM_PLAIN)
957 ldlm_reclaim_add(lock);
959 atomic_inc(&pl->pl_granted);
960 atomic_inc(&pl->pl_grant_rate);
961 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
963 * Do not do pool recalc for client side as all locks which
964 * potentially may be canceled has already been packed into
965 * enqueue/cancel rpc. Also we do not want to run out of stack
966 * with too long call paths.
968 if (ns_is_server(ldlm_pl2ns(pl)))
969 ldlm_pool_recalc(pl);
973 * Remove ldlm lock \a lock from pool \a pl accounting.
975 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
978 * Filter out FLOCK & PLAIN locks. Read above comment in
981 if (lock->l_resource->lr_type == LDLM_FLOCK ||
982 lock->l_resource->lr_type == LDLM_PLAIN)
985 ldlm_reclaim_del(lock);
987 LASSERT(atomic_read(&pl->pl_granted) > 0);
988 atomic_dec(&pl->pl_granted);
989 atomic_inc(&pl->pl_cancel_rate);
991 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
993 if (ns_is_server(ldlm_pl2ns(pl)))
994 ldlm_pool_recalc(pl);
998 * Returns current \a pl SLV.
1000 * \pre ->pl_lock is not locked.
1002 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
1005 spin_lock(&pl->pl_lock);
1006 slv = pl->pl_server_lock_volume;
1007 spin_unlock(&pl->pl_lock);
1012 * Sets passed \a slv to \a pl.
1014 * \pre ->pl_lock is not locked.
1016 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
1018 spin_lock(&pl->pl_lock);
1019 pl->pl_server_lock_volume = slv;
1020 spin_unlock(&pl->pl_lock);
1024 * Returns current \a pl CLV.
1026 * \pre ->pl_lock is not locked.
1028 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
1031 spin_lock(&pl->pl_lock);
1032 slv = pl->pl_client_lock_volume;
1033 spin_unlock(&pl->pl_lock);
1038 * Sets passed \a clv to \a pl.
1040 * \pre ->pl_lock is not locked.
1042 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1044 spin_lock(&pl->pl_lock);
1045 pl->pl_client_lock_volume = clv;
1046 spin_unlock(&pl->pl_lock);
1050 * Returns current \a pl limit.
1052 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1054 return atomic_read(&pl->pl_limit);
1058 * Sets passed \a limit to \a pl.
1060 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1062 atomic_set(&pl->pl_limit, limit);
1066 * Returns current LVF from \a pl.
1068 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1070 return atomic_read(&pl->pl_lock_volume_factor);
1073 static struct ptlrpc_thread *ldlm_pools_thread;
1074 static struct shrinker *ldlm_pools_srv_shrinker;
1075 static struct shrinker *ldlm_pools_cli_shrinker;
1076 static struct completion ldlm_pools_comp;
1079 * count locks from all namespaces (if possible). Returns number of
1082 static unsigned long ldlm_pools_count(enum ldlm_side client, gfp_t gfp_mask)
1084 unsigned long total = 0;
1086 struct ldlm_namespace *ns;
1087 struct ldlm_namespace *ns_old = NULL; /* loop detection */
1089 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1092 CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n",
1093 client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1096 * Find out how many resources we may release.
1098 for (nr_ns = ldlm_namespace_nr_read(client);
1099 nr_ns > 0; nr_ns--) {
1100 mutex_lock(ldlm_namespace_lock(client));
1101 if (list_empty(ldlm_namespace_list(client))) {
1102 mutex_unlock(ldlm_namespace_lock(client));
1105 ns = ldlm_namespace_first_locked(client);
1108 mutex_unlock(ldlm_namespace_lock(client));
1112 if (ldlm_ns_empty(ns)) {
1113 ldlm_namespace_move_to_inactive_locked(ns, client);
1114 mutex_unlock(ldlm_namespace_lock(client));
1121 ldlm_namespace_get(ns);
1122 ldlm_namespace_move_to_active_locked(ns, client);
1123 mutex_unlock(ldlm_namespace_lock(client));
1124 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
1125 ldlm_namespace_put(ns);
1131 static unsigned long ldlm_pools_scan(enum ldlm_side client, int nr,
1134 unsigned long freed = 0;
1136 struct ldlm_namespace *ns;
1138 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1142 * Shrink at least ldlm_namespace_nr_read(client) namespaces.
1144 for (tmp = nr_ns = ldlm_namespace_nr_read(client);
1146 int cancel, nr_locks;
1149 * Do not call shrink under ldlm_namespace_lock(client)
1151 mutex_lock(ldlm_namespace_lock(client));
1152 if (list_empty(ldlm_namespace_list(client))) {
1153 mutex_unlock(ldlm_namespace_lock(client));
1156 ns = ldlm_namespace_first_locked(client);
1157 ldlm_namespace_get(ns);
1158 ldlm_namespace_move_to_active_locked(ns, client);
1159 mutex_unlock(ldlm_namespace_lock(client));
1161 nr_locks = ldlm_pool_granted(&ns->ns_pool);
1163 * We use to shrink propotionally but with new shrinker API,
1164 * we lost the total number of freeable locks.
1166 cancel = 1 + min_t(int, nr_locks, nr / nr_ns);
1167 freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
1168 ldlm_namespace_put(ns);
1171 * we only decrease the SLV in server pools shrinker, return
1172 * SHRINK_STOP to kernel to avoid needless loop. LU-1128
1174 return (client == LDLM_NAMESPACE_SERVER) ? SHRINK_STOP : freed;
1177 #ifdef HAVE_SHRINKER_COUNT
1178 static unsigned long ldlm_pools_srv_count(struct shrinker *s,
1179 struct shrink_control *sc)
1181 return ldlm_pools_count(LDLM_NAMESPACE_SERVER, sc->gfp_mask);
1184 static unsigned long ldlm_pools_srv_scan(struct shrinker *s,
1185 struct shrink_control *sc)
1187 return ldlm_pools_scan(LDLM_NAMESPACE_SERVER, sc->nr_to_scan,
1191 static unsigned long ldlm_pools_cli_count(struct shrinker *s, struct shrink_control *sc)
1193 return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask);
1196 static unsigned long ldlm_pools_cli_scan(struct shrinker *s,
1197 struct shrink_control *sc)
1199 return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan,
1205 * Cancel \a nr locks from all namespaces (if possible). Returns number of
1206 * cached locks after shrink is finished. All namespaces are asked to
1207 * cancel approximately equal amount of locks to keep balancing.
1209 static int ldlm_pools_shrink(enum ldlm_side client, int nr, gfp_t gfp_mask)
1211 unsigned long total = 0;
1213 if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
1214 !(gfp_mask & __GFP_FS))
1217 CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n",
1218 nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1220 total = ldlm_pools_count(client, gfp_mask);
1222 if (nr == 0 || total == 0)
1225 return ldlm_pools_scan(client, nr, gfp_mask);
1228 static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1230 return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER,
1231 shrink_param(sc, nr_to_scan),
1232 shrink_param(sc, gfp_mask));
1235 static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1237 return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT,
1238 shrink_param(sc, nr_to_scan),
1239 shrink_param(sc, gfp_mask));
1242 #endif /* HAVE_SHRINKER_COUNT */
1244 int ldlm_pools_recalc(enum ldlm_side client)
1246 unsigned long nr_l = 0, nr_p = 0, l;
1247 struct ldlm_namespace *ns;
1248 struct ldlm_namespace *ns_old = NULL;
1250 /* seconds of sleep if no active namespaces */
1251 int time = client ? LDLM_POOL_CLI_DEF_RECALC_PERIOD :
1252 LDLM_POOL_SRV_DEF_RECALC_PERIOD;
1255 * No need to setup pool limit for client pools.
1257 if (client == LDLM_NAMESPACE_SERVER) {
1259 * Check all modest namespaces first.
1261 mutex_lock(ldlm_namespace_lock(client));
1262 list_for_each_entry(ns, ldlm_namespace_list(client),
1265 if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
1268 l = ldlm_pool_granted(&ns->ns_pool);
1273 * Set the modest pools limit equal to their avg granted
1276 l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0);
1277 ldlm_pool_setup(&ns->ns_pool, l);
1283 * Make sure that modest namespaces did not eat more that 2/3
1286 if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) {
1287 CWARN("\"Modest\" pools eat out 2/3 of server locks "
1288 "limit (%lu of %lu). This means that you have too "
1289 "many clients for this amount of server RAM. "
1290 "Upgrade server!\n", nr_l, LDLM_POOL_HOST_L);
1295 * The rest is given to greedy namespaces.
1297 list_for_each_entry(ns, ldlm_namespace_list(client),
1300 if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
1305 * In the case 2/3 locks are eaten out by
1306 * modest pools, we re-setup equal limit
1309 l = LDLM_POOL_HOST_L /
1310 ldlm_namespace_nr_read(client);
1313 * All the rest of greedy pools will have
1314 * all locks in equal parts.
1316 l = (LDLM_POOL_HOST_L - nr_l) /
1317 (ldlm_namespace_nr_read(client) -
1320 ldlm_pool_setup(&ns->ns_pool, l);
1322 mutex_unlock(ldlm_namespace_lock(client));
1326 * Recalc at least ldlm_namespace_nr(client) namespaces.
1328 for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) {
1331 * Lock the list, get first @ns in the list, getref, move it
1332 * to the tail, unlock and call pool recalc. This way we avoid
1333 * calling recalc under @ns lock what is really good as we get
1334 * rid of potential deadlock on client nodes when canceling
1335 * locks synchronously.
1337 mutex_lock(ldlm_namespace_lock(client));
1338 if (list_empty(ldlm_namespace_list(client))) {
1339 mutex_unlock(ldlm_namespace_lock(client));
1342 ns = ldlm_namespace_first_locked(client);
1344 if (ns_old == ns) { /* Full pass complete */
1345 mutex_unlock(ldlm_namespace_lock(client));
1349 /* We got an empty namespace, need to move it back to inactive
1351 * The race with parallel resource creation is fine:
1352 * - If they do namespace_get before our check, we fail the
1353 * check and they move this item to the end of the list anyway
1354 * - If we do the check and then they do namespace_get, then
1355 * we move the namespace to inactive and they will move
1356 * it back to active (synchronised by the lock, so no clash
1359 if (ldlm_ns_empty(ns)) {
1360 ldlm_namespace_move_to_inactive_locked(ns, client);
1361 mutex_unlock(ldlm_namespace_lock(client));
1368 spin_lock(&ns->ns_lock);
1370 * skip ns which is being freed, and we don't want to increase
1371 * its refcount again, not even temporarily. bz21519 & LU-499.
1373 if (ns->ns_stopping) {
1377 ldlm_namespace_get(ns);
1379 spin_unlock(&ns->ns_lock);
1381 ldlm_namespace_move_to_active_locked(ns, client);
1382 mutex_unlock(ldlm_namespace_lock(client));
1385 * After setup is done - recalc the pool.
1388 int ttime = ldlm_pool_recalc(&ns->ns_pool);
1393 ldlm_namespace_put(ns);
1397 /* Wake up the blocking threads from time to time. */
1398 ldlm_bl_thread_wakeup();
1403 static int ldlm_pools_thread_main(void *arg)
1405 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
1409 thread_set_flags(thread, SVC_RUNNING);
1410 wake_up(&thread->t_ctl_waitq);
1412 CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n",
1413 "ldlm_poold", current_pid());
1416 struct l_wait_info lwi;
1419 * Recal all pools on this tick.
1421 s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
1422 c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
1425 * Wait until the next check time, or until we're
1428 lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)),
1430 l_wait_event(thread->t_ctl_waitq,
1431 thread_is_stopping(thread) ||
1432 thread_is_event(thread),
1435 if (thread_test_and_clear_flags(thread, SVC_STOPPING))
1438 thread_test_and_clear_flags(thread, SVC_EVENT);
1441 thread_set_flags(thread, SVC_STOPPED);
1442 wake_up(&thread->t_ctl_waitq);
1444 CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n",
1445 "ldlm_poold", current_pid());
1447 complete_and_exit(&ldlm_pools_comp, 0);
1450 static int ldlm_pools_thread_start(void)
1452 struct l_wait_info lwi = { 0 };
1453 struct task_struct *task;
1456 if (ldlm_pools_thread != NULL)
1459 OBD_ALLOC_PTR(ldlm_pools_thread);
1460 if (ldlm_pools_thread == NULL)
1463 init_completion(&ldlm_pools_comp);
1464 init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq);
1466 task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread,
1469 CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task));
1470 OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread));
1471 ldlm_pools_thread = NULL;
1472 RETURN(PTR_ERR(task));
1474 l_wait_event(ldlm_pools_thread->t_ctl_waitq,
1475 thread_is_running(ldlm_pools_thread), &lwi);
1479 static void ldlm_pools_thread_stop(void)
1483 if (ldlm_pools_thread == NULL) {
1488 thread_set_flags(ldlm_pools_thread, SVC_STOPPING);
1489 wake_up(&ldlm_pools_thread->t_ctl_waitq);
1492 * Make sure that pools thread is finished before freeing @thread.
1493 * This fixes possible race and oops due to accessing freed memory
1496 wait_for_completion(&ldlm_pools_comp);
1497 OBD_FREE_PTR(ldlm_pools_thread);
1498 ldlm_pools_thread = NULL;
1502 int ldlm_pools_init(void)
1505 DEF_SHRINKER_VAR(shsvar, ldlm_pools_srv_shrink,
1506 ldlm_pools_srv_count, ldlm_pools_srv_scan);
1507 DEF_SHRINKER_VAR(shcvar, ldlm_pools_cli_shrink,
1508 ldlm_pools_cli_count, ldlm_pools_cli_scan);
1511 rc = ldlm_pools_thread_start();
1513 ldlm_pools_srv_shrinker =
1514 set_shrinker(DEFAULT_SEEKS, &shsvar);
1515 ldlm_pools_cli_shrinker =
1516 set_shrinker(DEFAULT_SEEKS, &shcvar);
1521 void ldlm_pools_fini(void)
1523 if (ldlm_pools_srv_shrinker != NULL) {
1524 remove_shrinker(ldlm_pools_srv_shrinker);
1525 ldlm_pools_srv_shrinker = NULL;
1527 if (ldlm_pools_cli_shrinker != NULL) {
1528 remove_shrinker(ldlm_pools_cli_shrinker);
1529 ldlm_pools_cli_shrinker = NULL;
1531 ldlm_pools_thread_stop();
1534 #else /* !HAVE_LRU_RESIZE_SUPPORT */
1535 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
1540 int ldlm_pool_recalc(struct ldlm_pool *pl)
1545 int ldlm_pool_shrink(struct ldlm_pool *pl,
1546 int nr, gfp_t gfp_mask)
1551 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
1552 int idx, enum ldlm_side client)
1557 void ldlm_pool_fini(struct ldlm_pool *pl)
1562 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
1567 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
1572 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
1577 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
1582 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
1587 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1592 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1597 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1602 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1607 int ldlm_pools_init(void)
1612 void ldlm_pools_fini(void)
1617 int ldlm_pools_recalc(enum ldlm_side client)
1621 #endif /* HAVE_LRU_RESIZE_SUPPORT */