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, 2017, 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 <libcfs/linux/linux-mem.h>
98 #include <lustre_dlm.h>
99 #include <cl_object.h>
100 #include <obd_class.h>
101 #include <obd_support.h>
102 #include "ldlm_internal.h"
104 #ifdef HAVE_LRU_RESIZE_SUPPORT
107 * 50 ldlm locks for 1MB of RAM.
109 #define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_SHIFT)) * 50)
112 * Maximal possible grant step plan in %.
114 #define LDLM_POOL_MAX_GSP (30)
117 * Minimal possible grant step plan in %.
119 #define LDLM_POOL_MIN_GSP (1)
122 * This controls the speed of reaching LDLM_POOL_MAX_GSP
123 * with increasing thread period.
125 #define LDLM_POOL_GSP_STEP_SHIFT (2)
128 * LDLM_POOL_GSP% of all locks is default GP.
130 #define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100)
133 * Max age for locks on clients.
135 #define LDLM_POOL_MAX_AGE (36000)
138 * The granularity of SLV calculation.
140 #define LDLM_POOL_SLV_SHIFT (10)
142 extern struct proc_dir_entry *ldlm_ns_proc_dir;
144 static inline __u64 dru(__u64 val, __u32 shift, int round_up)
146 return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift;
149 static inline __u64 ldlm_pool_slv_max(__u32 L)
152 * Allow to have all locks for 1 client for 10 hrs.
153 * Formula is the following: limit * 10h / 1 client.
155 __u64 lim = (__u64)L * LDLM_POOL_MAX_AGE / 1;
159 static inline __u64 ldlm_pool_slv_min(__u32 L)
165 LDLM_POOL_FIRST_STAT = 0,
166 LDLM_POOL_GRANTED_STAT = LDLM_POOL_FIRST_STAT,
167 LDLM_POOL_GRANT_STAT,
168 LDLM_POOL_CANCEL_STAT,
169 LDLM_POOL_GRANT_RATE_STAT,
170 LDLM_POOL_CANCEL_RATE_STAT,
171 LDLM_POOL_GRANT_PLAN_STAT,
173 LDLM_POOL_SHRINK_REQTD_STAT,
174 LDLM_POOL_SHRINK_FREED_STAT,
175 LDLM_POOL_RECALC_STAT,
176 LDLM_POOL_TIMING_STAT,
180 static inline struct ldlm_namespace *ldlm_pl2ns(struct ldlm_pool *pl)
182 return container_of(pl, struct ldlm_namespace, ns_pool);
186 * Calculates suggested grant_step in % of available locks for passed
187 * \a period. This is later used in grant_plan calculations.
189 static inline int ldlm_pool_t2gsp(unsigned int t)
192 * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP
193 * and up to 30% for anything higher than LDLM_POOL_GSP_STEP.
195 * How this will affect execution is the following:
197 * - for thread period 1s we will have grant_step 1% which good from
198 * pov of taking some load off from server and push it out to clients.
199 * This is like that because 1% for grant_step means that server will
200 * not allow clients to get lots of locks in short period of time and
201 * keep all old locks in their caches. Clients will always have to
202 * get some locks back if they want to take some new;
204 * - for thread period 10s (which is default) we will have 23% which
205 * means that clients will have enough of room to take some new locks
206 * without getting some back. All locks from this 23% which were not
207 * taken by clients in current period will contribute in SLV growing.
208 * SLV growing means more locks cached on clients until limit or grant
211 return LDLM_POOL_MAX_GSP -
212 ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >>
213 (t >> LDLM_POOL_GSP_STEP_SHIFT));
216 static inline int ldlm_pool_granted(struct ldlm_pool *pl)
218 return atomic_read(&pl->pl_granted);
222 * Recalculates next grant limit on passed \a pl.
224 * \pre ->pl_lock is locked.
226 static void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl)
228 int granted, grant_step, limit;
230 limit = ldlm_pool_get_limit(pl);
231 granted = ldlm_pool_granted(pl);
233 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
234 grant_step = ((limit - granted) * grant_step) / 100;
235 pl->pl_grant_plan = granted + grant_step;
236 limit = (limit * 5) >> 2;
237 if (pl->pl_grant_plan > limit)
238 pl->pl_grant_plan = limit;
242 * Recalculates next SLV on passed \a pl.
244 * \pre ->pl_lock is locked.
246 static void ldlm_pool_recalc_slv(struct ldlm_pool *pl)
256 slv = pl->pl_server_lock_volume;
257 grant_plan = pl->pl_grant_plan;
258 limit = ldlm_pool_get_limit(pl);
259 granted = ldlm_pool_granted(pl);
260 round_up = granted < limit;
262 grant_usage = max_t(int, limit - (granted - grant_plan), 1);
265 * Find out SLV change factor which is the ratio of grant usage
266 * from limit. SLV changes as fast as the ratio of grant plan
267 * consumption. The more locks from grant plan are not consumed
268 * by clients in last interval (idle time), the faster grows
269 * SLV. And the opposite, the more grant plan is over-consumed
270 * (load time) the faster drops SLV.
272 slv_factor = (grant_usage << LDLM_POOL_SLV_SHIFT);
273 do_div(slv_factor, limit);
274 slv = slv * slv_factor;
275 slv = dru(slv, LDLM_POOL_SLV_SHIFT, round_up);
277 if (slv > ldlm_pool_slv_max(limit)) {
278 slv = ldlm_pool_slv_max(limit);
279 } else if (slv < ldlm_pool_slv_min(limit)) {
280 slv = ldlm_pool_slv_min(limit);
283 pl->pl_server_lock_volume = slv;
287 * Recalculates next stats on passed \a pl.
289 * \pre ->pl_lock is locked.
291 static void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
293 int grant_plan = pl->pl_grant_plan;
294 __u64 slv = pl->pl_server_lock_volume;
295 int granted = ldlm_pool_granted(pl);
296 int grant_rate = atomic_read(&pl->pl_grant_rate);
297 int cancel_rate = atomic_read(&pl->pl_cancel_rate);
299 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT,
301 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
303 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
305 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
307 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
312 * Sets current SLV into obd accessible via ldlm_pl2ns(pl)->ns_obd.
314 static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl)
316 struct obd_device *obd;
319 * Set new SLV in obd field for using it later without accessing the
320 * pool. This is required to avoid race between sending reply to client
321 * with new SLV and cleanup server stack in which we can't guarantee
322 * that namespace is still alive. We know only that obd is alive as
323 * long as valid export is alive.
325 obd = ldlm_pl2ns(pl)->ns_obd;
326 LASSERT(obd != NULL);
327 write_lock(&obd->obd_pool_lock);
328 obd->obd_pool_slv = pl->pl_server_lock_volume;
329 write_unlock(&obd->obd_pool_lock);
333 * Recalculates all pool fields on passed \a pl.
335 * \pre ->pl_lock is not locked.
337 static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
339 time64_t recalc_interval_sec;
342 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time;
343 if (recalc_interval_sec < pl->pl_recalc_period)
346 spin_lock(&pl->pl_lock);
347 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time;
348 if (recalc_interval_sec < pl->pl_recalc_period) {
349 spin_unlock(&pl->pl_lock);
353 * Recalc SLV after last period. This should be done
354 * _before_ recalculating new grant plan.
356 ldlm_pool_recalc_slv(pl);
359 * Make sure that pool informed obd of last SLV changes.
361 ldlm_srv_pool_push_slv(pl);
364 * Update grant_plan for new period.
366 ldlm_pool_recalc_grant_plan(pl);
368 pl->pl_recalc_time = ktime_get_real_seconds();
369 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
370 recalc_interval_sec);
371 spin_unlock(&pl->pl_lock);
376 * This function is used on server side as main entry point for memory
377 * pressure handling. It decreases SLV on \a pl according to passed
378 * \a nr and \a gfp_mask.
380 * Our goal here is to decrease SLV such a way that clients hold \a nr
381 * locks smaller in next 10h.
383 static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
384 int nr, gfp_t gfp_mask)
389 * VM is asking how many entries may be potentially freed.
392 return ldlm_pool_granted(pl);
395 * Client already canceled locks but server is already in shrinker
396 * and can't cancel anything. Let's catch this race.
398 if (ldlm_pool_granted(pl) == 0)
401 spin_lock(&pl->pl_lock);
404 * We want shrinker to possibly cause cancellation of @nr locks from
405 * clients or grant approximately @nr locks smaller next intervals.
407 * This is why we decreased SLV by @nr. This effect will only be as
408 * long as one re-calc interval (1s these days) and this should be
409 * enough to pass this decreased SLV to all clients. On next recalc
410 * interval pool will either increase SLV if locks load is not high
411 * or will keep on same level or even decrease again, thus, shrinker
412 * decreased SLV will affect next recalc intervals and this way will
413 * make locking load lower.
415 if (nr < pl->pl_server_lock_volume) {
416 pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr;
418 limit = ldlm_pool_get_limit(pl);
419 pl->pl_server_lock_volume = ldlm_pool_slv_min(limit);
423 * Make sure that pool informed obd of last SLV changes.
425 ldlm_srv_pool_push_slv(pl);
426 spin_unlock(&pl->pl_lock);
429 * We did not really free any memory here so far, it only will be
430 * freed later may be, so that we return 0 to not confuse VM.
436 * Setup server side pool \a pl with passed \a limit.
438 static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit)
440 struct obd_device *obd;
442 obd = ldlm_pl2ns(pl)->ns_obd;
443 LASSERT(obd != NULL && obd != LP_POISON);
444 LASSERT(obd->obd_type != LP_POISON);
445 write_lock(&obd->obd_pool_lock);
446 obd->obd_pool_limit = limit;
447 write_unlock(&obd->obd_pool_lock);
449 ldlm_pool_set_limit(pl, limit);
454 * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl.
456 static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
458 struct obd_device *obd;
461 * Get new SLV and Limit from obd which is updated with coming
464 obd = ldlm_pl2ns(pl)->ns_obd;
465 LASSERT(obd != NULL);
466 read_lock(&obd->obd_pool_lock);
467 pl->pl_server_lock_volume = obd->obd_pool_slv;
468 ldlm_pool_set_limit(pl, obd->obd_pool_limit);
469 read_unlock(&obd->obd_pool_lock);
473 * Recalculates client size pool \a pl according to current SLV and Limit.
475 static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
477 time64_t recalc_interval_sec;
481 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time;
482 if (recalc_interval_sec < pl->pl_recalc_period)
485 spin_lock(&pl->pl_lock);
487 * Check if we need to recalc lists now.
489 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time;
490 if (recalc_interval_sec < pl->pl_recalc_period) {
491 spin_unlock(&pl->pl_lock);
496 * Make sure that pool knows last SLV and Limit from obd.
498 ldlm_cli_pool_pop_slv(pl);
499 spin_unlock(&pl->pl_lock);
502 * Do not cancel locks in case lru resize is disabled for this ns.
504 if (!ns_connect_lru_resize(ldlm_pl2ns(pl)))
508 * In the time of canceling locks on client we do not need to maintain
509 * sharp timing, we only want to cancel locks asap according to new SLV.
510 * It may be called when SLV has changed much, this is why we do not
511 * take into account pl->pl_recalc_time here.
513 ret = ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC,
517 spin_lock(&pl->pl_lock);
519 * Time of LRU resizing might be longer than period,
520 * so update after LRU resizing rather than before it.
522 pl->pl_recalc_time = ktime_get_real_seconds();
523 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
524 recalc_interval_sec);
525 spin_unlock(&pl->pl_lock);
530 * This function is main entry point for memory pressure handling on client
531 * side. Main goal of this function is to cancel some number of locks on
532 * passed \a pl according to \a nr and \a gfp_mask.
534 static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
535 int nr, gfp_t gfp_mask)
537 struct ldlm_namespace *ns;
543 * Do not cancel locks in case lru resize is disabled for this ns.
545 if (!ns_connect_lru_resize(ns))
549 * Make sure that pool knows last SLV and Limit from obd.
551 ldlm_cli_pool_pop_slv(pl);
553 spin_lock(&ns->ns_lock);
554 unused = ns->ns_nr_unused;
555 spin_unlock(&ns->ns_lock);
558 return (unused / 100) * sysctl_vfs_cache_pressure;
560 return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_LRU_FLAG_SHRINK);
563 static struct ldlm_pool_ops ldlm_srv_pool_ops = {
564 .po_recalc = ldlm_srv_pool_recalc,
565 .po_shrink = ldlm_srv_pool_shrink,
566 .po_setup = ldlm_srv_pool_setup
569 static struct ldlm_pool_ops ldlm_cli_pool_ops = {
570 .po_recalc = ldlm_cli_pool_recalc,
571 .po_shrink = ldlm_cli_pool_shrink
575 * Pool recalc wrapper. Will call either client or server pool recalc callback
576 * depending what pool \a pl is used.
578 int ldlm_pool_recalc(struct ldlm_pool *pl)
580 time64_t recalc_interval_sec;
583 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time;
584 if (recalc_interval_sec > 0) {
585 spin_lock(&pl->pl_lock);
586 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time;
588 if (recalc_interval_sec > 0) {
590 * Update pool statistics every 1s.
592 ldlm_pool_recalc_stats(pl);
595 * Zero out all rates and speed for the last period.
597 atomic_set(&pl->pl_grant_rate, 0);
598 atomic_set(&pl->pl_cancel_rate, 0);
600 spin_unlock(&pl->pl_lock);
603 if (pl->pl_ops->po_recalc != NULL) {
604 count = pl->pl_ops->po_recalc(pl);
605 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
609 recalc_interval_sec = pl->pl_recalc_time - ktime_get_real_seconds() +
610 pl->pl_recalc_period;
611 if (recalc_interval_sec <= 0) {
612 /* DEBUG: should be re-removed after LU-4536 is fixed */
613 CDEBUG(D_DLMTRACE, "%s: Negative interval(%lld), too short period(%lld)\n",
614 pl->pl_name, recalc_interval_sec,
615 (s64)pl->pl_recalc_period);
617 /* Prevent too frequent recalculation. */
618 recalc_interval_sec = 1;
621 return recalc_interval_sec;
625 * Pool shrink wrapper. Will call either client or server pool recalc callback
626 * depending what pool \a pl is used.
628 int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, gfp_t gfp_mask)
632 if (pl->pl_ops->po_shrink != NULL) {
633 cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask);
635 lprocfs_counter_add(pl->pl_stats,
636 LDLM_POOL_SHRINK_REQTD_STAT,
638 lprocfs_counter_add(pl->pl_stats,
639 LDLM_POOL_SHRINK_FREED_STAT,
641 CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, "
642 "shrunk %d\n", pl->pl_name, nr, cancel);
649 * Pool setup wrapper. Will call either client or server pool recalc callback
650 * depending what pool \a pl is used.
652 * Sets passed \a limit into pool \a pl.
654 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
656 if (pl->pl_ops->po_setup != NULL)
657 return(pl->pl_ops->po_setup(pl, limit));
661 static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
663 int granted, grant_rate, cancel_rate, grant_step;
664 int grant_speed, grant_plan, lvf;
665 struct ldlm_pool *pl = m->private;
669 spin_lock(&pl->pl_lock);
670 slv = pl->pl_server_lock_volume;
671 clv = pl->pl_client_lock_volume;
672 limit = ldlm_pool_get_limit(pl);
673 grant_plan = pl->pl_grant_plan;
674 granted = ldlm_pool_granted(pl);
675 grant_rate = atomic_read(&pl->pl_grant_rate);
676 cancel_rate = atomic_read(&pl->pl_cancel_rate);
677 grant_speed = grant_rate - cancel_rate;
678 lvf = atomic_read(&pl->pl_lock_volume_factor);
679 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
680 spin_unlock(&pl->pl_lock);
682 seq_printf(m, "LDLM pool state (%s):\n"
686 pl->pl_name, slv, clv, lvf);
688 if (ns_is_server(ldlm_pl2ns(pl))) {
689 seq_printf(m, " GSP: %d%%\n", grant_step);
690 seq_printf(m, " GP: %d\n", grant_plan);
693 seq_printf(m, " GR: %d\n CR: %d\n GS: %d\n G: %d\n L: %d\n",
694 grant_rate, cancel_rate, grant_speed,
698 LPROC_SEQ_FOPS_RO(lprocfs_pool_state);
700 static ssize_t grant_speed_show(struct kobject *kobj, struct attribute *attr,
703 struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool,
707 spin_lock(&pl->pl_lock);
708 /* serialize with ldlm_pool_recalc */
709 grant_speed = atomic_read(&pl->pl_grant_rate) -
710 atomic_read(&pl->pl_cancel_rate);
711 spin_unlock(&pl->pl_lock);
712 return sprintf(buf, "%d\n", grant_speed);
714 LUSTRE_RO_ATTR(grant_speed);
716 LDLM_POOL_SYSFS_READER_SHOW(grant_plan, int);
717 LUSTRE_RO_ATTR(grant_plan);
719 LDLM_POOL_SYSFS_READER_SHOW(recalc_period, int);
720 LDLM_POOL_SYSFS_WRITER_STORE(recalc_period, int);
721 LUSTRE_RW_ATTR(recalc_period);
723 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(server_lock_volume, u64);
724 LUSTRE_RO_ATTR(server_lock_volume);
726 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(limit, atomic);
727 LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(limit, atomic);
728 LUSTRE_RW_ATTR(limit);
730 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(granted, atomic);
731 LUSTRE_RO_ATTR(granted);
733 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(cancel_rate, atomic);
734 LUSTRE_RO_ATTR(cancel_rate);
736 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(grant_rate, atomic);
737 LUSTRE_RO_ATTR(grant_rate);
739 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(lock_volume_factor, atomic);
740 LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(lock_volume_factor, atomic);
741 LUSTRE_RW_ATTR(lock_volume_factor);
743 /* These are for pools in /sys/fs/lustre/ldlm/namespaces/.../pool */
744 static struct attribute *ldlm_pl_attrs[] = {
745 &lustre_attr_grant_speed.attr,
746 &lustre_attr_grant_plan.attr,
747 &lustre_attr_recalc_period.attr,
748 &lustre_attr_server_lock_volume.attr,
749 &lustre_attr_limit.attr,
750 &lustre_attr_granted.attr,
751 &lustre_attr_cancel_rate.attr,
752 &lustre_attr_grant_rate.attr,
753 &lustre_attr_lock_volume_factor.attr,
757 static void ldlm_pl_release(struct kobject *kobj)
759 struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool,
761 complete(&pl->pl_kobj_unregister);
764 static struct kobj_type ldlm_pl_ktype = {
765 .default_attrs = ldlm_pl_attrs,
766 .sysfs_ops = &lustre_sysfs_ops,
767 .release = ldlm_pl_release,
770 static int ldlm_pool_sysfs_init(struct ldlm_pool *pl)
772 struct ldlm_namespace *ns = ldlm_pl2ns(pl);
775 init_completion(&pl->pl_kobj_unregister);
776 err = kobject_init_and_add(&pl->pl_kobj, &ldlm_pl_ktype, &ns->ns_kobj,
782 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
784 struct ldlm_namespace *ns = ldlm_pl2ns(pl);
785 struct proc_dir_entry *parent_ns_proc;
786 struct lprocfs_vars pool_vars[2];
787 char *var_name = NULL;
791 OBD_ALLOC(var_name, MAX_STRING_SIZE + 1);
795 parent_ns_proc = ns->ns_proc_dir_entry;
796 if (parent_ns_proc == NULL) {
797 CERROR("%s: proc entry is not initialized\n",
799 GOTO(out_free_name, rc = -EINVAL);
801 pl->pl_proc_dir = lprocfs_register("pool", parent_ns_proc,
803 if (IS_ERR(pl->pl_proc_dir)) {
804 rc = PTR_ERR(pl->pl_proc_dir);
805 pl->pl_proc_dir = NULL;
806 CERROR("%s: cannot create 'pool' proc entry: rc = %d\n",
807 ldlm_ns_name(ns), rc);
808 GOTO(out_free_name, rc);
811 var_name[MAX_STRING_SIZE] = '\0';
812 memset(pool_vars, 0, sizeof(pool_vars));
813 pool_vars[0].name = var_name;
815 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "state", pl,
816 &lprocfs_pool_state_fops);
818 pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT -
819 LDLM_POOL_FIRST_STAT, 0);
821 GOTO(out_free_name, rc = -ENOMEM);
823 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
824 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
826 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
827 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
829 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
830 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
832 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
833 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
834 "grant_rate", "locks/s");
835 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
836 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
837 "cancel_rate", "locks/s");
838 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
839 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
840 "grant_plan", "locks/s");
841 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT,
842 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
844 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT,
845 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
846 "shrink_request", "locks");
847 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT,
848 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
849 "shrink_freed", "locks");
850 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
851 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
852 "recalc_freed", "locks");
853 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
854 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
855 "recalc_timing", "sec");
856 rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
860 OBD_FREE(var_name, MAX_STRING_SIZE + 1);
864 static void ldlm_pool_sysfs_fini(struct ldlm_pool *pl)
866 kobject_put(&pl->pl_kobj);
867 wait_for_completion(&pl->pl_kobj_unregister);
870 static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
872 if (pl->pl_stats != NULL) {
873 lprocfs_free_stats(&pl->pl_stats);
876 if (pl->pl_proc_dir != NULL) {
877 lprocfs_remove(&pl->pl_proc_dir);
878 pl->pl_proc_dir = NULL;
882 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
883 int idx, enum ldlm_side client)
888 spin_lock_init(&pl->pl_lock);
889 atomic_set(&pl->pl_granted, 0);
890 pl->pl_recalc_time = ktime_get_real_seconds();
891 atomic_set(&pl->pl_lock_volume_factor, 1);
893 atomic_set(&pl->pl_grant_rate, 0);
894 atomic_set(&pl->pl_cancel_rate, 0);
895 pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
897 snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
898 ldlm_ns_name(ns), idx);
900 if (client == LDLM_NAMESPACE_SERVER) {
901 pl->pl_ops = &ldlm_srv_pool_ops;
902 ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L);
903 pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD;
904 pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L);
906 ldlm_pool_set_limit(pl, 1);
907 pl->pl_server_lock_volume = 0;
908 pl->pl_ops = &ldlm_cli_pool_ops;
909 pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
911 pl->pl_client_lock_volume = 0;
912 rc = ldlm_pool_proc_init(pl);
916 rc = ldlm_pool_sysfs_init(pl);
920 CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name);
925 void ldlm_pool_fini(struct ldlm_pool *pl)
928 ldlm_pool_sysfs_fini(pl);
929 ldlm_pool_proc_fini(pl);
932 * Pool should not be used after this point. We can't free it here as
933 * it lives in struct ldlm_namespace, but still interested in catching
934 * any abnormal using cases.
936 POISON(pl, 0x5a, sizeof(*pl));
941 * Add new taken ldlm lock \a lock into pool \a pl accounting.
943 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
946 * FLOCK locks are special in a sense that they are almost never
947 * cancelled, instead special kind of lock is used to drop them.
948 * also there is no LRU for flock locks, so no point in tracking
951 * PLAIN locks are used by config and quota, the quantity is small
952 * and usually they are not in LRU.
954 if (lock->l_resource->lr_type == LDLM_FLOCK ||
955 lock->l_resource->lr_type == LDLM_PLAIN)
958 ldlm_reclaim_add(lock);
960 atomic_inc(&pl->pl_granted);
961 atomic_inc(&pl->pl_grant_rate);
962 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
964 * Do not do pool recalc for client side as all locks which
965 * potentially may be canceled has already been packed into
966 * enqueue/cancel rpc. Also we do not want to run out of stack
967 * with too long call paths.
969 if (ns_is_server(ldlm_pl2ns(pl)))
970 ldlm_pool_recalc(pl);
974 * Remove ldlm lock \a lock from pool \a pl accounting.
976 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
979 * Filter out FLOCK & PLAIN locks. Read above comment in
982 if (lock->l_resource->lr_type == LDLM_FLOCK ||
983 lock->l_resource->lr_type == LDLM_PLAIN)
986 ldlm_reclaim_del(lock);
988 LASSERT(atomic_read(&pl->pl_granted) > 0);
989 atomic_dec(&pl->pl_granted);
990 atomic_inc(&pl->pl_cancel_rate);
992 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
994 if (ns_is_server(ldlm_pl2ns(pl)))
995 ldlm_pool_recalc(pl);
999 * Returns current \a pl SLV.
1001 * \pre ->pl_lock is not locked.
1003 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
1006 spin_lock(&pl->pl_lock);
1007 slv = pl->pl_server_lock_volume;
1008 spin_unlock(&pl->pl_lock);
1013 * Sets passed \a slv to \a pl.
1015 * \pre ->pl_lock is not locked.
1017 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
1019 spin_lock(&pl->pl_lock);
1020 pl->pl_server_lock_volume = slv;
1021 spin_unlock(&pl->pl_lock);
1025 * Returns current \a pl CLV.
1027 * \pre ->pl_lock is not locked.
1029 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
1032 spin_lock(&pl->pl_lock);
1033 slv = pl->pl_client_lock_volume;
1034 spin_unlock(&pl->pl_lock);
1039 * Sets passed \a clv to \a pl.
1041 * \pre ->pl_lock is not locked.
1043 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1045 spin_lock(&pl->pl_lock);
1046 pl->pl_client_lock_volume = clv;
1047 spin_unlock(&pl->pl_lock);
1051 * Returns current \a pl limit.
1053 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1055 return atomic_read(&pl->pl_limit);
1059 * Sets passed \a limit to \a pl.
1061 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1063 atomic_set(&pl->pl_limit, limit);
1067 * Returns current LVF from \a pl.
1069 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1071 return atomic_read(&pl->pl_lock_volume_factor);
1074 static struct ptlrpc_thread *ldlm_pools_thread;
1075 static struct shrinker *ldlm_pools_srv_shrinker;
1076 static struct shrinker *ldlm_pools_cli_shrinker;
1077 static struct completion ldlm_pools_comp;
1080 * count locks from all namespaces (if possible). Returns number of
1083 static unsigned long ldlm_pools_count(enum ldlm_side client, gfp_t gfp_mask)
1085 unsigned long total = 0;
1087 struct ldlm_namespace *ns;
1088 struct ldlm_namespace *ns_old = NULL; /* loop detection */
1090 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1093 CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n",
1094 client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1097 * Find out how many resources we may release.
1099 for (nr_ns = ldlm_namespace_nr_read(client);
1100 nr_ns > 0; nr_ns--) {
1101 mutex_lock(ldlm_namespace_lock(client));
1102 if (list_empty(ldlm_namespace_list(client))) {
1103 mutex_unlock(ldlm_namespace_lock(client));
1106 ns = ldlm_namespace_first_locked(client);
1109 mutex_unlock(ldlm_namespace_lock(client));
1113 if (ldlm_ns_empty(ns)) {
1114 ldlm_namespace_move_to_inactive_locked(ns, client);
1115 mutex_unlock(ldlm_namespace_lock(client));
1122 ldlm_namespace_get(ns);
1123 ldlm_namespace_move_to_active_locked(ns, client);
1124 mutex_unlock(ldlm_namespace_lock(client));
1125 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
1126 ldlm_namespace_put(ns);
1132 static unsigned long ldlm_pools_scan(enum ldlm_side client, int nr,
1135 unsigned long freed = 0;
1137 struct ldlm_namespace *ns;
1139 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1143 * Shrink at least ldlm_namespace_nr_read(client) namespaces.
1145 for (tmp = nr_ns = ldlm_namespace_nr_read(client);
1147 int cancel, nr_locks;
1150 * Do not call shrink under ldlm_namespace_lock(client)
1152 mutex_lock(ldlm_namespace_lock(client));
1153 if (list_empty(ldlm_namespace_list(client))) {
1154 mutex_unlock(ldlm_namespace_lock(client));
1157 ns = ldlm_namespace_first_locked(client);
1158 ldlm_namespace_get(ns);
1159 ldlm_namespace_move_to_active_locked(ns, client);
1160 mutex_unlock(ldlm_namespace_lock(client));
1162 nr_locks = ldlm_pool_granted(&ns->ns_pool);
1164 * We use to shrink propotionally but with new shrinker API,
1165 * we lost the total number of freeable locks.
1167 cancel = 1 + min_t(int, nr_locks, nr / nr_ns);
1168 freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
1169 ldlm_namespace_put(ns);
1172 * we only decrease the SLV in server pools shrinker, return
1173 * SHRINK_STOP to kernel to avoid needless loop. LU-1128
1175 return (client == LDLM_NAMESPACE_SERVER) ? SHRINK_STOP : freed;
1178 #ifdef HAVE_SHRINKER_COUNT
1179 static unsigned long ldlm_pools_srv_count(struct shrinker *s,
1180 struct shrink_control *sc)
1182 return ldlm_pools_count(LDLM_NAMESPACE_SERVER, sc->gfp_mask);
1185 static unsigned long ldlm_pools_srv_scan(struct shrinker *s,
1186 struct shrink_control *sc)
1188 return ldlm_pools_scan(LDLM_NAMESPACE_SERVER, sc->nr_to_scan,
1192 static unsigned long ldlm_pools_cli_count(struct shrinker *s, struct shrink_control *sc)
1194 return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask);
1197 static unsigned long ldlm_pools_cli_scan(struct shrinker *s,
1198 struct shrink_control *sc)
1200 return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan,
1206 * Cancel \a nr locks from all namespaces (if possible). Returns number of
1207 * cached locks after shrink is finished. All namespaces are asked to
1208 * cancel approximately equal amount of locks to keep balancing.
1210 static int ldlm_pools_shrink(enum ldlm_side client, int nr, gfp_t gfp_mask)
1212 unsigned long total = 0;
1214 if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
1215 !(gfp_mask & __GFP_FS))
1218 CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n",
1219 nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1221 total = ldlm_pools_count(client, gfp_mask);
1223 if (nr == 0 || total == 0)
1226 return ldlm_pools_scan(client, nr, gfp_mask);
1229 static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1231 return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER,
1232 shrink_param(sc, nr_to_scan),
1233 shrink_param(sc, gfp_mask));
1236 static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1238 return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT,
1239 shrink_param(sc, nr_to_scan),
1240 shrink_param(sc, gfp_mask));
1243 #endif /* HAVE_SHRINKER_COUNT */
1245 int ldlm_pools_recalc(enum ldlm_side client)
1247 unsigned long nr_l = 0, nr_p = 0, l;
1248 struct ldlm_namespace *ns;
1249 struct ldlm_namespace *ns_old = NULL;
1251 /* seconds of sleep if no active namespaces */
1252 int time = client ? LDLM_POOL_CLI_DEF_RECALC_PERIOD :
1253 LDLM_POOL_SRV_DEF_RECALC_PERIOD;
1256 * No need to setup pool limit for client pools.
1258 if (client == LDLM_NAMESPACE_SERVER) {
1260 * Check all modest namespaces first.
1262 mutex_lock(ldlm_namespace_lock(client));
1263 list_for_each_entry(ns, ldlm_namespace_list(client),
1266 if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
1269 l = ldlm_pool_granted(&ns->ns_pool);
1274 * Set the modest pools limit equal to their avg granted
1277 l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0);
1278 ldlm_pool_setup(&ns->ns_pool, l);
1284 * Make sure that modest namespaces did not eat more that 2/3
1287 if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) {
1288 CWARN("\"Modest\" pools eat out 2/3 of server locks "
1289 "limit (%lu of %lu). This means that you have too "
1290 "many clients for this amount of server RAM. "
1291 "Upgrade server!\n", nr_l, LDLM_POOL_HOST_L);
1296 * The rest is given to greedy namespaces.
1298 list_for_each_entry(ns, ldlm_namespace_list(client),
1301 if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
1306 * In the case 2/3 locks are eaten out by
1307 * modest pools, we re-setup equal limit
1310 l = LDLM_POOL_HOST_L /
1311 ldlm_namespace_nr_read(client);
1314 * All the rest of greedy pools will have
1315 * all locks in equal parts.
1317 l = (LDLM_POOL_HOST_L - nr_l) /
1318 (ldlm_namespace_nr_read(client) -
1321 ldlm_pool_setup(&ns->ns_pool, l);
1323 mutex_unlock(ldlm_namespace_lock(client));
1327 * Recalc at least ldlm_namespace_nr(client) namespaces.
1329 for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) {
1332 * Lock the list, get first @ns in the list, getref, move it
1333 * to the tail, unlock and call pool recalc. This way we avoid
1334 * calling recalc under @ns lock what is really good as we get
1335 * rid of potential deadlock on client nodes when canceling
1336 * locks synchronously.
1338 mutex_lock(ldlm_namespace_lock(client));
1339 if (list_empty(ldlm_namespace_list(client))) {
1340 mutex_unlock(ldlm_namespace_lock(client));
1343 ns = ldlm_namespace_first_locked(client);
1345 if (ns_old == ns) { /* Full pass complete */
1346 mutex_unlock(ldlm_namespace_lock(client));
1350 /* We got an empty namespace, need to move it back to inactive
1352 * The race with parallel resource creation is fine:
1353 * - If they do namespace_get before our check, we fail the
1354 * check and they move this item to the end of the list anyway
1355 * - If we do the check and then they do namespace_get, then
1356 * we move the namespace to inactive and they will move
1357 * it back to active (synchronised by the lock, so no clash
1360 if (ldlm_ns_empty(ns)) {
1361 ldlm_namespace_move_to_inactive_locked(ns, client);
1362 mutex_unlock(ldlm_namespace_lock(client));
1369 spin_lock(&ns->ns_lock);
1371 * skip ns which is being freed, and we don't want to increase
1372 * its refcount again, not even temporarily. bz21519 & LU-499.
1374 if (ns->ns_stopping) {
1378 ldlm_namespace_get(ns);
1380 spin_unlock(&ns->ns_lock);
1382 ldlm_namespace_move_to_active_locked(ns, client);
1383 mutex_unlock(ldlm_namespace_lock(client));
1386 * After setup is done - recalc the pool.
1389 int ttime = ldlm_pool_recalc(&ns->ns_pool);
1394 ldlm_namespace_put(ns);
1398 /* Wake up the blocking threads from time to time. */
1399 ldlm_bl_thread_wakeup();
1404 static int ldlm_pools_thread_main(void *arg)
1406 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
1410 thread_set_flags(thread, SVC_RUNNING);
1411 wake_up(&thread->t_ctl_waitq);
1413 CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n",
1414 "ldlm_poold", current_pid());
1417 struct l_wait_info lwi;
1420 * Recal all pools on this tick.
1422 s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
1423 c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
1426 * Wait until the next check time, or until we're
1429 lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)),
1431 l_wait_event(thread->t_ctl_waitq,
1432 thread_is_stopping(thread) ||
1433 thread_is_event(thread),
1436 if (thread_test_and_clear_flags(thread, SVC_STOPPING))
1439 thread_test_and_clear_flags(thread, SVC_EVENT);
1442 thread_set_flags(thread, SVC_STOPPED);
1443 wake_up(&thread->t_ctl_waitq);
1445 CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n",
1446 "ldlm_poold", current_pid());
1448 complete_and_exit(&ldlm_pools_comp, 0);
1451 static int ldlm_pools_thread_start(void)
1453 struct l_wait_info lwi = { 0 };
1454 struct task_struct *task;
1457 if (ldlm_pools_thread != NULL)
1460 OBD_ALLOC_PTR(ldlm_pools_thread);
1461 if (ldlm_pools_thread == NULL)
1464 init_completion(&ldlm_pools_comp);
1465 init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq);
1467 task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread,
1470 CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task));
1471 OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread));
1472 ldlm_pools_thread = NULL;
1473 RETURN(PTR_ERR(task));
1475 l_wait_event(ldlm_pools_thread->t_ctl_waitq,
1476 thread_is_running(ldlm_pools_thread), &lwi);
1480 static void ldlm_pools_thread_stop(void)
1484 if (ldlm_pools_thread == NULL) {
1489 thread_set_flags(ldlm_pools_thread, SVC_STOPPING);
1490 wake_up(&ldlm_pools_thread->t_ctl_waitq);
1493 * Make sure that pools thread is finished before freeing @thread.
1494 * This fixes possible race and oops due to accessing freed memory
1497 wait_for_completion(&ldlm_pools_comp);
1498 OBD_FREE_PTR(ldlm_pools_thread);
1499 ldlm_pools_thread = NULL;
1503 int ldlm_pools_init(void)
1506 DEF_SHRINKER_VAR(shsvar, ldlm_pools_srv_shrink,
1507 ldlm_pools_srv_count, ldlm_pools_srv_scan);
1508 DEF_SHRINKER_VAR(shcvar, ldlm_pools_cli_shrink,
1509 ldlm_pools_cli_count, ldlm_pools_cli_scan);
1512 rc = ldlm_pools_thread_start();
1514 ldlm_pools_srv_shrinker =
1515 set_shrinker(DEFAULT_SEEKS, &shsvar);
1516 ldlm_pools_cli_shrinker =
1517 set_shrinker(DEFAULT_SEEKS, &shcvar);
1522 void ldlm_pools_fini(void)
1524 if (ldlm_pools_srv_shrinker != NULL) {
1525 remove_shrinker(ldlm_pools_srv_shrinker);
1526 ldlm_pools_srv_shrinker = NULL;
1528 if (ldlm_pools_cli_shrinker != NULL) {
1529 remove_shrinker(ldlm_pools_cli_shrinker);
1530 ldlm_pools_cli_shrinker = NULL;
1532 ldlm_pools_thread_stop();
1535 #else /* !HAVE_LRU_RESIZE_SUPPORT */
1536 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
1541 int ldlm_pool_recalc(struct ldlm_pool *pl)
1546 int ldlm_pool_shrink(struct ldlm_pool *pl,
1547 int nr, gfp_t gfp_mask)
1552 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
1553 int idx, enum ldlm_side client)
1558 void ldlm_pool_fini(struct ldlm_pool *pl)
1563 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
1568 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
1573 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
1578 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
1583 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
1588 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1593 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1598 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1603 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1608 int ldlm_pools_init(void)
1613 void ldlm_pools_fini(void)
1618 int ldlm_pools_recalc(enum ldlm_side client)
1622 #endif /* HAVE_LRU_RESIZE_SUPPORT */