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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_pool.c
38 * Author: Yury Umanets <umka@clusterfs.com>
42 * Idea of this code is rather simple. Each second, for each server namespace
43 * we have SLV - server lock volume which is calculated on current number of
44 * granted locks, grant speed for past period, etc - that is, locking load.
45 * This SLV number may be thought as a flow definition for simplicity. It is
46 * sent to clients with each occasion to let them know what is current load
47 * situation on the server. By default, at the beginning, SLV on server is
48 * set max value which is calculated as the following: allow to one client
49 * have all locks of limit ->pl_limit for 10h.
51 * Next, on clients, number of cached locks is not limited artificially in any
52 * way as it was before. Instead, client calculates CLV, that is, client lock
53 * volume for each lock and compares it with last SLV from the server. CLV is
54 * calculated as the number of locks in LRU * lock live time in seconds. If
55 * CLV > SLV - lock is canceled.
57 * Client has LVF, that is, lock volume factor which regulates how much sensitive
58 * client should be about last SLV from server. The higher LVF is the more locks
59 * will be canceled on client. Default value for it is 1. Setting LVF to 2 means
60 * that client will cancel locks 2 times faster.
62 * Locks on a client will be canceled more intensively in these cases:
63 * (1) if SLV is smaller, that is, load is higher on the server;
64 * (2) client has a lot of locks (the more locks are held by client, the bigger
65 * chances that some of them should be canceled);
66 * (3) client has old locks (taken some time ago);
68 * Thus, according to flow paradigm that we use for better understanding SLV,
69 * CLV is the volume of particle in flow described by SLV. According to this,
70 * if flow is getting thinner, more and more particles become outside of it and
71 * as particles are locks, they should be canceled.
73 * General idea of this belongs to Vitaly Fertman (vitaly@clusterfs.com). Andreas
74 * Dilger (adilger@clusterfs.com) proposed few nice ideas like using LVF and many
75 * cleanups. Flow definition to allow more easy understanding of the logic belongs
76 * to Nikita Danilov (nikita@clusterfs.com) as well as many cleanups and fixes.
77 * And design and implementation are done by Yury Umanets (umka@clusterfs.com).
79 * Glossary for terms used:
81 * pl_limit - Number of allowed locks in pool. Applies to server and client
84 * pl_granted - Number of granted locks (calculated);
85 * pl_grant_rate - Number of granted locks for last T (calculated);
86 * pl_cancel_rate - Number of canceled locks for last T (calculated);
87 * pl_grant_speed - Grant speed (GR - CR) for last T (calculated);
88 * pl_grant_plan - Planned number of granted locks for next T (calculated);
89 * pl_server_lock_volume - Current server lock volume (calculated);
91 * As it may be seen from list above, we have few possible tunables which may
92 * affect behavior much. They all may be modified via proc. However, they also
93 * give a possibility for constructing few pre-defined behavior policies. If
94 * none of predefines is suitable for a working pattern being used, new one may
95 * be "constructed" via proc tunables.
98 #define DEBUG_SUBSYSTEM S_LDLM
100 #include <linux/kthread.h>
101 #include <lustre_dlm.h>
102 #include <cl_object.h>
103 #include <obd_class.h>
104 #include <obd_support.h>
105 #include "ldlm_internal.h"
107 #ifdef HAVE_LRU_RESIZE_SUPPORT
110 * 50 ldlm locks for 1MB of RAM.
112 #define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_CACHE_SHIFT)) * 50)
115 * Maximal possible grant step plan in %.
117 #define LDLM_POOL_MAX_GSP (30)
120 * Minimal possible grant step plan in %.
122 #define LDLM_POOL_MIN_GSP (1)
125 * This controls the speed of reaching LDLM_POOL_MAX_GSP
126 * with increasing thread period.
128 #define LDLM_POOL_GSP_STEP_SHIFT (2)
131 * LDLM_POOL_GSP% of all locks is default GP.
133 #define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100)
136 * Max age for locks on clients.
138 #define LDLM_POOL_MAX_AGE (36000)
141 * The granularity of SLV calculation.
143 #define LDLM_POOL_SLV_SHIFT (10)
145 extern struct proc_dir_entry *ldlm_ns_proc_dir;
147 static inline __u64 dru(__u64 val, __u32 shift, int round_up)
149 return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift;
152 static inline __u64 ldlm_pool_slv_max(__u32 L)
155 * Allow to have all locks for 1 client for 10 hrs.
156 * Formula is the following: limit * 10h / 1 client.
158 __u64 lim = (__u64)L * LDLM_POOL_MAX_AGE / 1;
162 static inline __u64 ldlm_pool_slv_min(__u32 L)
168 LDLM_POOL_FIRST_STAT = 0,
169 LDLM_POOL_GRANTED_STAT = LDLM_POOL_FIRST_STAT,
170 LDLM_POOL_GRANT_STAT,
171 LDLM_POOL_CANCEL_STAT,
172 LDLM_POOL_GRANT_RATE_STAT,
173 LDLM_POOL_CANCEL_RATE_STAT,
174 LDLM_POOL_GRANT_PLAN_STAT,
176 LDLM_POOL_SHRINK_REQTD_STAT,
177 LDLM_POOL_SHRINK_FREED_STAT,
178 LDLM_POOL_RECALC_STAT,
179 LDLM_POOL_TIMING_STAT,
183 static inline struct ldlm_namespace *ldlm_pl2ns(struct ldlm_pool *pl)
185 return container_of(pl, struct ldlm_namespace, ns_pool);
189 * Calculates suggested grant_step in % of available locks for passed
190 * \a period. This is later used in grant_plan calculations.
192 static inline int ldlm_pool_t2gsp(unsigned int t)
195 * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP
196 * and up to 30% for anything higher than LDLM_POOL_GSP_STEP.
198 * How this will affect execution is the following:
200 * - for thread period 1s we will have grant_step 1% which good from
201 * pov of taking some load off from server and push it out to clients.
202 * This is like that because 1% for grant_step means that server will
203 * not allow clients to get lots of locks in short period of time and
204 * keep all old locks in their caches. Clients will always have to
205 * get some locks back if they want to take some new;
207 * - for thread period 10s (which is default) we will have 23% which
208 * means that clients will have enough of room to take some new locks
209 * without getting some back. All locks from this 23% which were not
210 * taken by clients in current period will contribute in SLV growing.
211 * SLV growing means more locks cached on clients until limit or grant
214 return LDLM_POOL_MAX_GSP -
215 ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >>
216 (t >> LDLM_POOL_GSP_STEP_SHIFT));
219 static inline int ldlm_pool_granted(struct ldlm_pool *pl)
221 return atomic_read(&pl->pl_granted);
225 * Recalculates next grant limit on passed \a pl.
227 * \pre ->pl_lock is locked.
229 static void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl)
231 int granted, grant_step, limit;
233 limit = ldlm_pool_get_limit(pl);
234 granted = ldlm_pool_granted(pl);
236 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
237 grant_step = ((limit - granted) * grant_step) / 100;
238 pl->pl_grant_plan = granted + grant_step;
239 limit = (limit * 5) >> 2;
240 if (pl->pl_grant_plan > limit)
241 pl->pl_grant_plan = limit;
245 * Recalculates next SLV on passed \a pl.
247 * \pre ->pl_lock is locked.
249 static void ldlm_pool_recalc_slv(struct ldlm_pool *pl)
259 slv = pl->pl_server_lock_volume;
260 grant_plan = pl->pl_grant_plan;
261 limit = ldlm_pool_get_limit(pl);
262 granted = ldlm_pool_granted(pl);
263 round_up = granted < limit;
265 grant_usage = max_t(int, limit - (granted - grant_plan), 1);
268 * Find out SLV change factor which is the ratio of grant usage
269 * from limit. SLV changes as fast as the ratio of grant plan
270 * consumption. The more locks from grant plan are not consumed
271 * by clients in last interval (idle time), the faster grows
272 * SLV. And the opposite, the more grant plan is over-consumed
273 * (load time) the faster drops SLV.
275 slv_factor = (grant_usage << LDLM_POOL_SLV_SHIFT);
276 do_div(slv_factor, limit);
277 slv = slv * slv_factor;
278 slv = dru(slv, LDLM_POOL_SLV_SHIFT, round_up);
280 if (slv > ldlm_pool_slv_max(limit)) {
281 slv = ldlm_pool_slv_max(limit);
282 } else if (slv < ldlm_pool_slv_min(limit)) {
283 slv = ldlm_pool_slv_min(limit);
286 pl->pl_server_lock_volume = slv;
290 * Recalculates next stats on passed \a pl.
292 * \pre ->pl_lock is locked.
294 static void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
296 int grant_plan = pl->pl_grant_plan;
297 __u64 slv = pl->pl_server_lock_volume;
298 int granted = ldlm_pool_granted(pl);
299 int grant_rate = atomic_read(&pl->pl_grant_rate);
300 int cancel_rate = atomic_read(&pl->pl_cancel_rate);
302 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT,
304 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
306 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
308 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
310 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
315 * Sets current SLV into obd accessible via ldlm_pl2ns(pl)->ns_obd.
317 static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl)
319 struct obd_device *obd;
322 * Set new SLV in obd field for using it later without accessing the
323 * pool. This is required to avoid race between sending reply to client
324 * with new SLV and cleanup server stack in which we can't guarantee
325 * that namespace is still alive. We know only that obd is alive as
326 * long as valid export is alive.
328 obd = ldlm_pl2ns(pl)->ns_obd;
329 LASSERT(obd != NULL);
330 write_lock(&obd->obd_pool_lock);
331 obd->obd_pool_slv = pl->pl_server_lock_volume;
332 write_unlock(&obd->obd_pool_lock);
336 * Recalculates all pool fields on passed \a pl.
338 * \pre ->pl_lock is not locked.
340 static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
342 time_t recalc_interval_sec;
345 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
346 if (recalc_interval_sec < pl->pl_recalc_period)
349 spin_lock(&pl->pl_lock);
350 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
351 if (recalc_interval_sec < pl->pl_recalc_period) {
352 spin_unlock(&pl->pl_lock);
356 * Recalc SLV after last period. This should be done
357 * _before_ recalculating new grant plan.
359 ldlm_pool_recalc_slv(pl);
362 * Make sure that pool informed obd of last SLV changes.
364 ldlm_srv_pool_push_slv(pl);
367 * Update grant_plan for new period.
369 ldlm_pool_recalc_grant_plan(pl);
371 pl->pl_recalc_time = cfs_time_current_sec();
372 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
373 recalc_interval_sec);
374 spin_unlock(&pl->pl_lock);
379 * This function is used on server side as main entry point for memory
380 * pressure handling. It decreases SLV on \a pl according to passed
381 * \a nr and \a gfp_mask.
383 * Our goal here is to decrease SLV such a way that clients hold \a nr
384 * locks smaller in next 10h.
386 static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
387 int nr, gfp_t gfp_mask)
392 * VM is asking how many entries may be potentially freed.
395 return ldlm_pool_granted(pl);
398 * Client already canceled locks but server is already in shrinker
399 * and can't cancel anything. Let's catch this race.
401 if (ldlm_pool_granted(pl) == 0)
404 spin_lock(&pl->pl_lock);
407 * We want shrinker to possibly cause cancellation of @nr locks from
408 * clients or grant approximately @nr locks smaller next intervals.
410 * This is why we decreased SLV by @nr. This effect will only be as
411 * long as one re-calc interval (1s these days) and this should be
412 * enough to pass this decreased SLV to all clients. On next recalc
413 * interval pool will either increase SLV if locks load is not high
414 * or will keep on same level or even decrease again, thus, shrinker
415 * decreased SLV will affect next recalc intervals and this way will
416 * make locking load lower.
418 if (nr < pl->pl_server_lock_volume) {
419 pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr;
421 limit = ldlm_pool_get_limit(pl);
422 pl->pl_server_lock_volume = ldlm_pool_slv_min(limit);
426 * Make sure that pool informed obd of last SLV changes.
428 ldlm_srv_pool_push_slv(pl);
429 spin_unlock(&pl->pl_lock);
432 * We did not really free any memory here so far, it only will be
433 * freed later may be, so that we return 0 to not confuse VM.
439 * Setup server side pool \a pl with passed \a limit.
441 static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit)
443 struct obd_device *obd;
445 obd = ldlm_pl2ns(pl)->ns_obd;
446 LASSERT(obd != NULL && obd != LP_POISON);
447 LASSERT(obd->obd_type != LP_POISON);
448 write_lock(&obd->obd_pool_lock);
449 obd->obd_pool_limit = limit;
450 write_unlock(&obd->obd_pool_lock);
452 ldlm_pool_set_limit(pl, limit);
457 * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl.
459 static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
461 struct obd_device *obd;
464 * Get new SLV and Limit from obd which is updated with coming
467 obd = ldlm_pl2ns(pl)->ns_obd;
468 LASSERT(obd != NULL);
469 read_lock(&obd->obd_pool_lock);
470 pl->pl_server_lock_volume = obd->obd_pool_slv;
471 ldlm_pool_set_limit(pl, obd->obd_pool_limit);
472 read_unlock(&obd->obd_pool_lock);
476 * Recalculates client size pool \a pl according to current SLV and Limit.
478 static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
480 time_t recalc_interval_sec;
484 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
485 if (recalc_interval_sec < pl->pl_recalc_period)
488 spin_lock(&pl->pl_lock);
490 * Check if we need to recalc lists now.
492 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
493 if (recalc_interval_sec < pl->pl_recalc_period) {
494 spin_unlock(&pl->pl_lock);
499 * Make sure that pool knows last SLV and Limit from obd.
501 ldlm_cli_pool_pop_slv(pl);
502 spin_unlock(&pl->pl_lock);
505 * Do not cancel locks in case lru resize is disabled for this ns.
507 if (!ns_connect_lru_resize(ldlm_pl2ns(pl)))
511 * In the time of canceling locks on client we do not need to maintain
512 * sharp timing, we only want to cancel locks asap according to new SLV.
513 * It may be called when SLV has changed much, this is why we do not
514 * take into account pl->pl_recalc_time here.
516 ret = ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC,
520 spin_lock(&pl->pl_lock);
522 * Time of LRU resizing might be longer than period,
523 * so update after LRU resizing rather than before it.
525 pl->pl_recalc_time = cfs_time_current_sec();
526 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
527 recalc_interval_sec);
528 spin_unlock(&pl->pl_lock);
533 * This function is main entry point for memory pressure handling on client
534 * side. Main goal of this function is to cancel some number of locks on
535 * passed \a pl according to \a nr and \a gfp_mask.
537 static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
538 int nr, gfp_t gfp_mask)
540 struct ldlm_namespace *ns;
546 * Do not cancel locks in case lru resize is disabled for this ns.
548 if (!ns_connect_lru_resize(ns))
552 * Make sure that pool knows last SLV and Limit from obd.
554 ldlm_cli_pool_pop_slv(pl);
556 spin_lock(&ns->ns_lock);
557 unused = ns->ns_nr_unused;
558 spin_unlock(&ns->ns_lock);
561 return (unused / 100) * sysctl_vfs_cache_pressure;
563 return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_LRU_FLAG_SHRINK);
566 static struct ldlm_pool_ops ldlm_srv_pool_ops = {
567 .po_recalc = ldlm_srv_pool_recalc,
568 .po_shrink = ldlm_srv_pool_shrink,
569 .po_setup = ldlm_srv_pool_setup
572 static struct ldlm_pool_ops ldlm_cli_pool_ops = {
573 .po_recalc = ldlm_cli_pool_recalc,
574 .po_shrink = ldlm_cli_pool_shrink
578 * Pool recalc wrapper. Will call either client or server pool recalc callback
579 * depending what pool \a pl is used.
581 int ldlm_pool_recalc(struct ldlm_pool *pl)
583 time_t recalc_interval_sec;
586 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
587 if (recalc_interval_sec > 0) {
588 spin_lock(&pl->pl_lock);
589 recalc_interval_sec = cfs_time_current_sec() -
592 if (recalc_interval_sec > 0) {
594 * Update pool statistics every 1s.
596 ldlm_pool_recalc_stats(pl);
599 * Zero out all rates and speed for the last period.
601 atomic_set(&pl->pl_grant_rate, 0);
602 atomic_set(&pl->pl_cancel_rate, 0);
604 spin_unlock(&pl->pl_lock);
607 if (pl->pl_ops->po_recalc != NULL) {
608 count = pl->pl_ops->po_recalc(pl);
609 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
613 recalc_interval_sec = pl->pl_recalc_time - cfs_time_current_sec() +
614 pl->pl_recalc_period;
615 if (recalc_interval_sec <= 0) {
616 /* DEBUG: should be re-removed after LU-4536 is fixed */
617 CDEBUG(D_DLMTRACE, "%s: Negative interval(%ld), "
618 "too short period(%ld)\n",
619 pl->pl_name, recalc_interval_sec,
620 pl->pl_recalc_period);
622 /* Prevent too frequent recalculation. */
623 recalc_interval_sec = 1;
626 return recalc_interval_sec;
630 * Pool shrink wrapper. Will call either client or server pool recalc callback
631 * depending what pool \a pl is used.
633 int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, gfp_t gfp_mask)
637 if (pl->pl_ops->po_shrink != NULL) {
638 cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask);
640 lprocfs_counter_add(pl->pl_stats,
641 LDLM_POOL_SHRINK_REQTD_STAT,
643 lprocfs_counter_add(pl->pl_stats,
644 LDLM_POOL_SHRINK_FREED_STAT,
646 CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, "
647 "shrunk %d\n", pl->pl_name, nr, cancel);
654 * Pool setup wrapper. Will call either client or server pool recalc callback
655 * depending what pool \a pl is used.
657 * Sets passed \a limit into pool \a pl.
659 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
661 if (pl->pl_ops->po_setup != NULL)
662 return(pl->pl_ops->po_setup(pl, limit));
666 static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
668 int granted, grant_rate, cancel_rate, grant_step;
669 int grant_speed, grant_plan, lvf;
670 struct ldlm_pool *pl = m->private;
674 spin_lock(&pl->pl_lock);
675 slv = pl->pl_server_lock_volume;
676 clv = pl->pl_client_lock_volume;
677 limit = ldlm_pool_get_limit(pl);
678 grant_plan = pl->pl_grant_plan;
679 granted = ldlm_pool_granted(pl);
680 grant_rate = atomic_read(&pl->pl_grant_rate);
681 cancel_rate = atomic_read(&pl->pl_cancel_rate);
682 grant_speed = grant_rate - cancel_rate;
683 lvf = atomic_read(&pl->pl_lock_volume_factor);
684 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
685 spin_unlock(&pl->pl_lock);
687 seq_printf(m, "LDLM pool state (%s):\n"
691 pl->pl_name, slv, clv, lvf);
693 if (ns_is_server(ldlm_pl2ns(pl))) {
694 seq_printf(m, " GSP: %d%%\n"
696 grant_step, grant_plan);
698 seq_printf(m, " GR: %d\n" " CR: %d\n" " GS: %d\n"
699 " G: %d\n" " L: %d\n",
700 grant_rate, cancel_rate, grant_speed,
704 LPROC_SEQ_FOPS_RO(lprocfs_pool_state);
706 static int lprocfs_grant_speed_seq_show(struct seq_file *m, void *unused)
708 struct ldlm_pool *pl = m->private;
711 spin_lock(&pl->pl_lock);
712 /* serialize with ldlm_pool_recalc */
713 grant_speed = atomic_read(&pl->pl_grant_rate) -
714 atomic_read(&pl->pl_cancel_rate);
715 spin_unlock(&pl->pl_lock);
716 return lprocfs_uint_seq_show(m, &grant_speed);
719 LDLM_POOL_PROC_READER_SEQ_SHOW(grant_plan, int);
720 LPROC_SEQ_FOPS_RO(lprocfs_grant_plan);
722 LDLM_POOL_PROC_READER_SEQ_SHOW(recalc_period, int);
723 LDLM_POOL_PROC_WRITER(recalc_period, int);
724 static ssize_t lprocfs_recalc_period_seq_write(struct file *file,
725 const char __user *buf,
726 size_t len, loff_t *off)
728 struct seq_file *seq = file->private_data;
730 return lprocfs_wr_recalc_period(file, buf, len, seq->private);
732 LPROC_SEQ_FOPS(lprocfs_recalc_period);
734 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, u64);
735 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, atomic);
736 LPROC_SEQ_FOPS_RW_TYPE(ldlm_pool_rw, atomic);
738 LPROC_SEQ_FOPS_RO(lprocfs_grant_speed);
740 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
742 struct ldlm_namespace *ns = ldlm_pl2ns(pl);
743 struct proc_dir_entry *parent_ns_proc;
744 struct lprocfs_vars pool_vars[2];
745 char *var_name = NULL;
749 OBD_ALLOC(var_name, MAX_STRING_SIZE + 1);
753 parent_ns_proc = ns->ns_proc_dir_entry;
754 if (parent_ns_proc == NULL) {
755 CERROR("%s: proc entry is not initialized\n",
757 GOTO(out_free_name, rc = -EINVAL);
759 pl->pl_proc_dir = lprocfs_register("pool", parent_ns_proc,
761 if (IS_ERR(pl->pl_proc_dir)) {
762 rc = PTR_ERR(pl->pl_proc_dir);
763 pl->pl_proc_dir = NULL;
764 CERROR("%s: cannot create 'pool' proc entry: rc = %d\n",
765 ldlm_ns_name(ns), rc);
766 GOTO(out_free_name, rc);
769 var_name[MAX_STRING_SIZE] = '\0';
770 memset(pool_vars, 0, sizeof(pool_vars));
771 pool_vars[0].name = var_name;
773 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "server_lock_volume",
774 &pl->pl_server_lock_volume, &ldlm_pool_u64_fops);
775 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "limit", &pl->pl_limit,
776 &ldlm_pool_rw_atomic_fops);
777 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "granted",
778 &pl->pl_granted, &ldlm_pool_atomic_fops);
779 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_speed", pl,
780 &lprocfs_grant_speed_fops);
781 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "cancel_rate",
782 &pl->pl_cancel_rate, &ldlm_pool_atomic_fops);
783 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_rate",
784 &pl->pl_grant_rate, &ldlm_pool_atomic_fops);
785 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_plan", pl,
786 &lprocfs_grant_plan_fops);
787 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "recalc_period",
788 pl, &lprocfs_recalc_period_fops);
789 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "lock_volume_factor",
790 &pl->pl_lock_volume_factor, &ldlm_pool_rw_atomic_fops);
791 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "state", pl,
792 &lprocfs_pool_state_fops);
794 pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT -
795 LDLM_POOL_FIRST_STAT, 0);
797 GOTO(out_free_name, rc = -ENOMEM);
799 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
800 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
802 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
803 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
805 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
806 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
808 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
809 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
810 "grant_rate", "locks/s");
811 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
812 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
813 "cancel_rate", "locks/s");
814 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
815 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
816 "grant_plan", "locks/s");
817 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT,
818 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
820 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT,
821 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
822 "shrink_request", "locks");
823 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT,
824 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
825 "shrink_freed", "locks");
826 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
827 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
828 "recalc_freed", "locks");
829 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
830 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
831 "recalc_timing", "sec");
832 rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
836 OBD_FREE(var_name, MAX_STRING_SIZE + 1);
840 static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
842 if (pl->pl_stats != NULL) {
843 lprocfs_free_stats(&pl->pl_stats);
846 if (pl->pl_proc_dir != NULL) {
847 lprocfs_remove(&pl->pl_proc_dir);
848 pl->pl_proc_dir = NULL;
852 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
853 int idx, enum ldlm_side client)
858 spin_lock_init(&pl->pl_lock);
859 atomic_set(&pl->pl_granted, 0);
860 pl->pl_recalc_time = cfs_time_current_sec();
861 atomic_set(&pl->pl_lock_volume_factor, 1);
863 atomic_set(&pl->pl_grant_rate, 0);
864 atomic_set(&pl->pl_cancel_rate, 0);
865 pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
867 snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
868 ldlm_ns_name(ns), idx);
870 if (client == LDLM_NAMESPACE_SERVER) {
871 pl->pl_ops = &ldlm_srv_pool_ops;
872 ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L);
873 pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD;
874 pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L);
876 ldlm_pool_set_limit(pl, 1);
877 pl->pl_server_lock_volume = 0;
878 pl->pl_ops = &ldlm_cli_pool_ops;
879 pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
881 pl->pl_client_lock_volume = 0;
882 rc = ldlm_pool_proc_init(pl);
886 CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name);
891 void ldlm_pool_fini(struct ldlm_pool *pl)
894 ldlm_pool_proc_fini(pl);
897 * Pool should not be used after this point. We can't free it here as
898 * it lives in struct ldlm_namespace, but still interested in catching
899 * any abnormal using cases.
901 POISON(pl, 0x5a, sizeof(*pl));
906 * Add new taken ldlm lock \a lock into pool \a pl accounting.
908 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
911 * FLOCK locks are special in a sense that they are almost never
912 * cancelled, instead special kind of lock is used to drop them.
913 * also there is no LRU for flock locks, so no point in tracking
916 * PLAIN locks are used by config and quota, the quantity is small
917 * and usually they are not in LRU.
919 if (lock->l_resource->lr_type == LDLM_FLOCK ||
920 lock->l_resource->lr_type == LDLM_PLAIN)
923 ldlm_reclaim_add(lock);
925 atomic_inc(&pl->pl_granted);
926 atomic_inc(&pl->pl_grant_rate);
927 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
929 * Do not do pool recalc for client side as all locks which
930 * potentially may be canceled has already been packed into
931 * enqueue/cancel rpc. Also we do not want to run out of stack
932 * with too long call paths.
934 if (ns_is_server(ldlm_pl2ns(pl)))
935 ldlm_pool_recalc(pl);
939 * Remove ldlm lock \a lock from pool \a pl accounting.
941 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
944 * Filter out FLOCK & PLAIN locks. Read above comment in
947 if (lock->l_resource->lr_type == LDLM_FLOCK ||
948 lock->l_resource->lr_type == LDLM_PLAIN)
951 ldlm_reclaim_del(lock);
953 LASSERT(atomic_read(&pl->pl_granted) > 0);
954 atomic_dec(&pl->pl_granted);
955 atomic_inc(&pl->pl_cancel_rate);
957 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
959 if (ns_is_server(ldlm_pl2ns(pl)))
960 ldlm_pool_recalc(pl);
964 * Returns current \a pl SLV.
966 * \pre ->pl_lock is not locked.
968 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
971 spin_lock(&pl->pl_lock);
972 slv = pl->pl_server_lock_volume;
973 spin_unlock(&pl->pl_lock);
978 * Sets passed \a slv to \a pl.
980 * \pre ->pl_lock is not locked.
982 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
984 spin_lock(&pl->pl_lock);
985 pl->pl_server_lock_volume = slv;
986 spin_unlock(&pl->pl_lock);
990 * Returns current \a pl CLV.
992 * \pre ->pl_lock is not locked.
994 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
997 spin_lock(&pl->pl_lock);
998 slv = pl->pl_client_lock_volume;
999 spin_unlock(&pl->pl_lock);
1004 * Sets passed \a clv to \a pl.
1006 * \pre ->pl_lock is not locked.
1008 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1010 spin_lock(&pl->pl_lock);
1011 pl->pl_client_lock_volume = clv;
1012 spin_unlock(&pl->pl_lock);
1016 * Returns current \a pl limit.
1018 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1020 return atomic_read(&pl->pl_limit);
1024 * Sets passed \a limit to \a pl.
1026 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1028 atomic_set(&pl->pl_limit, limit);
1032 * Returns current LVF from \a pl.
1034 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1036 return atomic_read(&pl->pl_lock_volume_factor);
1039 static struct ptlrpc_thread *ldlm_pools_thread;
1040 static struct shrinker *ldlm_pools_srv_shrinker;
1041 static struct shrinker *ldlm_pools_cli_shrinker;
1042 static struct completion ldlm_pools_comp;
1045 * count locks from all namespaces (if possible). Returns number of
1048 static unsigned long ldlm_pools_count(enum ldlm_side client, gfp_t gfp_mask)
1050 unsigned long total = 0;
1052 struct ldlm_namespace *ns;
1053 struct ldlm_namespace *ns_old = NULL; /* loop detection */
1056 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1059 CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n",
1060 client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1062 cookie = cl_env_reenter();
1065 * Find out how many resources we may release.
1067 for (nr_ns = ldlm_namespace_nr_read(client);
1068 nr_ns > 0; nr_ns--) {
1069 mutex_lock(ldlm_namespace_lock(client));
1070 if (list_empty(ldlm_namespace_list(client))) {
1071 mutex_unlock(ldlm_namespace_lock(client));
1072 cl_env_reexit(cookie);
1075 ns = ldlm_namespace_first_locked(client);
1078 mutex_unlock(ldlm_namespace_lock(client));
1082 if (ldlm_ns_empty(ns)) {
1083 ldlm_namespace_move_to_inactive_locked(ns, client);
1084 mutex_unlock(ldlm_namespace_lock(client));
1091 ldlm_namespace_get(ns);
1092 ldlm_namespace_move_to_active_locked(ns, client);
1093 mutex_unlock(ldlm_namespace_lock(client));
1094 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
1095 ldlm_namespace_put(ns);
1098 cl_env_reexit(cookie);
1102 static unsigned long ldlm_pools_scan(enum ldlm_side client, int nr,
1105 unsigned long freed = 0;
1107 struct ldlm_namespace *ns;
1110 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1113 cookie = cl_env_reenter();
1116 * Shrink at least ldlm_namespace_nr_read(client) namespaces.
1118 for (tmp = nr_ns = ldlm_namespace_nr_read(client);
1120 int cancel, nr_locks;
1123 * Do not call shrink under ldlm_namespace_lock(client)
1125 mutex_lock(ldlm_namespace_lock(client));
1126 if (list_empty(ldlm_namespace_list(client))) {
1127 mutex_unlock(ldlm_namespace_lock(client));
1130 ns = ldlm_namespace_first_locked(client);
1131 ldlm_namespace_get(ns);
1132 ldlm_namespace_move_to_active_locked(ns, client);
1133 mutex_unlock(ldlm_namespace_lock(client));
1135 nr_locks = ldlm_pool_granted(&ns->ns_pool);
1137 * We use to shrink propotionally but with new shrinker API,
1138 * we lost the total number of freeable locks.
1140 cancel = 1 + min_t(int, nr_locks, nr / nr_ns);
1141 freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
1142 ldlm_namespace_put(ns);
1144 cl_env_reexit(cookie);
1146 * we only decrease the SLV in server pools shrinker, return
1147 * SHRINK_STOP to kernel to avoid needless loop. LU-1128
1149 return (client == LDLM_NAMESPACE_SERVER) ? SHRINK_STOP : freed;
1152 #ifdef HAVE_SHRINKER_COUNT
1153 static unsigned long ldlm_pools_srv_count(struct shrinker *s,
1154 struct shrink_control *sc)
1156 return ldlm_pools_count(LDLM_NAMESPACE_SERVER, sc->gfp_mask);
1159 static unsigned long ldlm_pools_srv_scan(struct shrinker *s,
1160 struct shrink_control *sc)
1162 return ldlm_pools_scan(LDLM_NAMESPACE_SERVER, sc->nr_to_scan,
1166 static unsigned long ldlm_pools_cli_count(struct shrinker *s, struct shrink_control *sc)
1168 return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask);
1171 static unsigned long ldlm_pools_cli_scan(struct shrinker *s,
1172 struct shrink_control *sc)
1174 return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan,
1180 * Cancel \a nr locks from all namespaces (if possible). Returns number of
1181 * cached locks after shrink is finished. All namespaces are asked to
1182 * cancel approximately equal amount of locks to keep balancing.
1184 static int ldlm_pools_shrink(enum ldlm_side client, int nr, gfp_t gfp_mask)
1186 unsigned long total = 0;
1188 if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
1189 !(gfp_mask & __GFP_FS))
1192 CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n",
1193 nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1195 total = ldlm_pools_count(client, gfp_mask);
1197 if (nr == 0 || total == 0)
1200 return ldlm_pools_scan(client, nr, gfp_mask);
1203 static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1205 return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER,
1206 shrink_param(sc, nr_to_scan),
1207 shrink_param(sc, gfp_mask));
1210 static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1212 return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT,
1213 shrink_param(sc, nr_to_scan),
1214 shrink_param(sc, gfp_mask));
1217 #endif /* HAVE_SHRINKER_COUNT */
1219 int ldlm_pools_recalc(enum ldlm_side client)
1221 unsigned long nr_l = 0, nr_p = 0, l;
1222 struct ldlm_namespace *ns;
1223 struct ldlm_namespace *ns_old = NULL;
1225 /* seconds of sleep if no active namespaces */
1226 int time = client ? LDLM_POOL_CLI_DEF_RECALC_PERIOD :
1227 LDLM_POOL_SRV_DEF_RECALC_PERIOD;
1230 * No need to setup pool limit for client pools.
1232 if (client == LDLM_NAMESPACE_SERVER) {
1234 * Check all modest namespaces first.
1236 mutex_lock(ldlm_namespace_lock(client));
1237 list_for_each_entry(ns, ldlm_namespace_list(client),
1240 if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
1243 l = ldlm_pool_granted(&ns->ns_pool);
1248 * Set the modest pools limit equal to their avg granted
1251 l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0);
1252 ldlm_pool_setup(&ns->ns_pool, l);
1258 * Make sure that modest namespaces did not eat more that 2/3
1261 if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) {
1262 CWARN("\"Modest\" pools eat out 2/3 of server locks "
1263 "limit (%lu of %lu). This means that you have too "
1264 "many clients for this amount of server RAM. "
1265 "Upgrade server!\n", nr_l, LDLM_POOL_HOST_L);
1270 * The rest is given to greedy namespaces.
1272 list_for_each_entry(ns, ldlm_namespace_list(client),
1275 if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
1280 * In the case 2/3 locks are eaten out by
1281 * modest pools, we re-setup equal limit
1284 l = LDLM_POOL_HOST_L /
1285 ldlm_namespace_nr_read(client);
1288 * All the rest of greedy pools will have
1289 * all locks in equal parts.
1291 l = (LDLM_POOL_HOST_L - nr_l) /
1292 (ldlm_namespace_nr_read(client) -
1295 ldlm_pool_setup(&ns->ns_pool, l);
1297 mutex_unlock(ldlm_namespace_lock(client));
1301 * Recalc at least ldlm_namespace_nr(client) namespaces.
1303 for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) {
1306 * Lock the list, get first @ns in the list, getref, move it
1307 * to the tail, unlock and call pool recalc. This way we avoid
1308 * calling recalc under @ns lock what is really good as we get
1309 * rid of potential deadlock on client nodes when canceling
1310 * locks synchronously.
1312 mutex_lock(ldlm_namespace_lock(client));
1313 if (list_empty(ldlm_namespace_list(client))) {
1314 mutex_unlock(ldlm_namespace_lock(client));
1317 ns = ldlm_namespace_first_locked(client);
1319 if (ns_old == ns) { /* Full pass complete */
1320 mutex_unlock(ldlm_namespace_lock(client));
1324 /* We got an empty namespace, need to move it back to inactive
1326 * The race with parallel resource creation is fine:
1327 * - If they do namespace_get before our check, we fail the
1328 * check and they move this item to the end of the list anyway
1329 * - If we do the check and then they do namespace_get, then
1330 * we move the namespace to inactive and they will move
1331 * it back to active (synchronised by the lock, so no clash
1334 if (ldlm_ns_empty(ns)) {
1335 ldlm_namespace_move_to_inactive_locked(ns, client);
1336 mutex_unlock(ldlm_namespace_lock(client));
1343 spin_lock(&ns->ns_lock);
1345 * skip ns which is being freed, and we don't want to increase
1346 * its refcount again, not even temporarily. bz21519 & LU-499.
1348 if (ns->ns_stopping) {
1352 ldlm_namespace_get(ns);
1354 spin_unlock(&ns->ns_lock);
1356 ldlm_namespace_move_to_active_locked(ns, client);
1357 mutex_unlock(ldlm_namespace_lock(client));
1360 * After setup is done - recalc the pool.
1363 int ttime = ldlm_pool_recalc(&ns->ns_pool);
1368 ldlm_namespace_put(ns);
1372 /* Wake up the blocking threads from time to time. */
1373 ldlm_bl_thread_wakeup();
1378 static int ldlm_pools_thread_main(void *arg)
1380 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
1384 thread_set_flags(thread, SVC_RUNNING);
1385 wake_up(&thread->t_ctl_waitq);
1387 CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n",
1388 "ldlm_poold", current_pid());
1391 struct l_wait_info lwi;
1394 * Recal all pools on this tick.
1396 s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
1397 c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
1400 * Wait until the next check time, or until we're
1403 lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)),
1405 l_wait_event(thread->t_ctl_waitq,
1406 thread_is_stopping(thread) ||
1407 thread_is_event(thread),
1410 if (thread_test_and_clear_flags(thread, SVC_STOPPING))
1413 thread_test_and_clear_flags(thread, SVC_EVENT);
1416 thread_set_flags(thread, SVC_STOPPED);
1417 wake_up(&thread->t_ctl_waitq);
1419 CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n",
1420 "ldlm_poold", current_pid());
1422 complete_and_exit(&ldlm_pools_comp, 0);
1425 static int ldlm_pools_thread_start(void)
1427 struct l_wait_info lwi = { 0 };
1428 struct task_struct *task;
1431 if (ldlm_pools_thread != NULL)
1434 OBD_ALLOC_PTR(ldlm_pools_thread);
1435 if (ldlm_pools_thread == NULL)
1438 init_completion(&ldlm_pools_comp);
1439 init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq);
1441 task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread,
1444 CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task));
1445 OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread));
1446 ldlm_pools_thread = NULL;
1447 RETURN(PTR_ERR(task));
1449 l_wait_event(ldlm_pools_thread->t_ctl_waitq,
1450 thread_is_running(ldlm_pools_thread), &lwi);
1454 static void ldlm_pools_thread_stop(void)
1458 if (ldlm_pools_thread == NULL) {
1463 thread_set_flags(ldlm_pools_thread, SVC_STOPPING);
1464 wake_up(&ldlm_pools_thread->t_ctl_waitq);
1467 * Make sure that pools thread is finished before freeing @thread.
1468 * This fixes possible race and oops due to accessing freed memory
1471 wait_for_completion(&ldlm_pools_comp);
1472 OBD_FREE_PTR(ldlm_pools_thread);
1473 ldlm_pools_thread = NULL;
1477 int ldlm_pools_init(void)
1480 DEF_SHRINKER_VAR(shsvar, ldlm_pools_srv_shrink,
1481 ldlm_pools_srv_count, ldlm_pools_srv_scan);
1482 DEF_SHRINKER_VAR(shcvar, ldlm_pools_cli_shrink,
1483 ldlm_pools_cli_count, ldlm_pools_cli_scan);
1486 rc = ldlm_pools_thread_start();
1488 ldlm_pools_srv_shrinker =
1489 set_shrinker(DEFAULT_SEEKS, &shsvar);
1490 ldlm_pools_cli_shrinker =
1491 set_shrinker(DEFAULT_SEEKS, &shcvar);
1496 void ldlm_pools_fini(void)
1498 if (ldlm_pools_srv_shrinker != NULL) {
1499 remove_shrinker(ldlm_pools_srv_shrinker);
1500 ldlm_pools_srv_shrinker = NULL;
1502 if (ldlm_pools_cli_shrinker != NULL) {
1503 remove_shrinker(ldlm_pools_cli_shrinker);
1504 ldlm_pools_cli_shrinker = NULL;
1506 ldlm_pools_thread_stop();
1509 #else /* !HAVE_LRU_RESIZE_SUPPORT */
1510 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
1515 int ldlm_pool_recalc(struct ldlm_pool *pl)
1520 int ldlm_pool_shrink(struct ldlm_pool *pl,
1521 int nr, gfp_t gfp_mask)
1526 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
1527 int idx, enum ldlm_side client)
1532 void ldlm_pool_fini(struct ldlm_pool *pl)
1537 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
1542 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
1547 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
1552 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
1557 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
1562 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1567 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1572 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1577 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1582 int ldlm_pools_init(void)
1587 void ldlm_pools_fini(void)
1592 int ldlm_pools_recalc(enum ldlm_side client)
1596 #endif /* HAVE_LRU_RESIZE_SUPPORT */