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, 2012, 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 <lustre_dlm.h>
101 #include <cl_object.h>
102 #include <obd_class.h>
103 #include <obd_support.h>
104 #include "ldlm_internal.h"
106 #ifdef HAVE_LRU_RESIZE_SUPPORT
109 * 50 ldlm locks for 1MB of RAM.
111 #define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_CACHE_SHIFT)) * 50)
114 * Maximal possible grant step plan in %.
116 #define LDLM_POOL_MAX_GSP (30)
119 * Minimal possible grant step plan in %.
121 #define LDLM_POOL_MIN_GSP (1)
124 * This controls the speed of reaching LDLM_POOL_MAX_GSP
125 * with increasing thread period.
127 #define LDLM_POOL_GSP_STEP_SHIFT (2)
130 * LDLM_POOL_GSP% of all locks is default GP.
132 #define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100)
135 * Max age for locks on clients.
137 #define LDLM_POOL_MAX_AGE (36000)
140 * The granularity of SLV calculation.
142 #define LDLM_POOL_SLV_SHIFT (10)
144 extern struct proc_dir_entry *ldlm_ns_proc_dir;
146 static inline __u64 dru(__u64 val, __u32 shift, int round_up)
148 return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift;
151 static inline __u64 ldlm_pool_slv_max(__u32 L)
154 * Allow to have all locks for 1 client for 10 hrs.
155 * Formula is the following: limit * 10h / 1 client.
157 __u64 lim = (__u64)L * LDLM_POOL_MAX_AGE / 1;
161 static inline __u64 ldlm_pool_slv_min(__u32 L)
167 LDLM_POOL_FIRST_STAT = 0,
168 LDLM_POOL_GRANTED_STAT = LDLM_POOL_FIRST_STAT,
169 LDLM_POOL_GRANT_STAT,
170 LDLM_POOL_CANCEL_STAT,
171 LDLM_POOL_GRANT_RATE_STAT,
172 LDLM_POOL_CANCEL_RATE_STAT,
173 LDLM_POOL_GRANT_PLAN_STAT,
175 LDLM_POOL_SHRINK_REQTD_STAT,
176 LDLM_POOL_SHRINK_FREED_STAT,
177 LDLM_POOL_RECALC_STAT,
178 LDLM_POOL_TIMING_STAT,
182 static inline struct ldlm_namespace *ldlm_pl2ns(struct ldlm_pool *pl)
184 return container_of(pl, struct ldlm_namespace, ns_pool);
188 * Calculates suggested grant_step in % of available locks for passed
189 * \a period. This is later used in grant_plan calculations.
191 static inline int ldlm_pool_t2gsp(unsigned int t)
194 * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP
195 * and up to 30% for anything higher than LDLM_POOL_GSP_STEP.
197 * How this will affect execution is the following:
199 * - for thread period 1s we will have grant_step 1% which good from
200 * pov of taking some load off from server and push it out to clients.
201 * This is like that because 1% for grant_step means that server will
202 * not allow clients to get lots of locks in short period of time and
203 * keep all old locks in their caches. Clients will always have to
204 * get some locks back if they want to take some new;
206 * - for thread period 10s (which is default) we will have 23% which
207 * means that clients will have enough of room to take some new locks
208 * without getting some back. All locks from this 23% which were not
209 * taken by clients in current period will contribute in SLV growing.
210 * SLV growing means more locks cached on clients until limit or grant
213 return LDLM_POOL_MAX_GSP -
214 ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >>
215 (t >> LDLM_POOL_GSP_STEP_SHIFT));
218 static inline int ldlm_pool_granted(struct ldlm_pool *pl)
220 return atomic_read(&pl->pl_granted);
224 * Recalculates next grant limit on passed \a pl.
226 * \pre ->pl_lock is locked.
228 static void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl)
230 int granted, grant_step, limit;
232 limit = ldlm_pool_get_limit(pl);
233 granted = ldlm_pool_granted(pl);
235 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
236 grant_step = ((limit - granted) * grant_step) / 100;
237 pl->pl_grant_plan = granted + grant_step;
238 limit = (limit * 5) >> 2;
239 if (pl->pl_grant_plan > limit)
240 pl->pl_grant_plan = limit;
244 * Recalculates next SLV on passed \a pl.
246 * \pre ->pl_lock is locked.
248 static void ldlm_pool_recalc_slv(struct ldlm_pool *pl)
258 slv = pl->pl_server_lock_volume;
259 grant_plan = pl->pl_grant_plan;
260 limit = ldlm_pool_get_limit(pl);
261 granted = ldlm_pool_granted(pl);
262 round_up = granted < limit;
264 grant_usage = max_t(int, limit - (granted - grant_plan), 1);
267 * Find out SLV change factor which is the ratio of grant usage
268 * from limit. SLV changes as fast as the ratio of grant plan
269 * consumption. The more locks from grant plan are not consumed
270 * by clients in last interval (idle time), the faster grows
271 * SLV. And the opposite, the more grant plan is over-consumed
272 * (load time) the faster drops SLV.
274 slv_factor = (grant_usage << LDLM_POOL_SLV_SHIFT);
275 do_div(slv_factor, limit);
276 slv = slv * slv_factor;
277 slv = dru(slv, LDLM_POOL_SLV_SHIFT, round_up);
279 if (slv > ldlm_pool_slv_max(limit)) {
280 slv = ldlm_pool_slv_max(limit);
281 } else if (slv < ldlm_pool_slv_min(limit)) {
282 slv = ldlm_pool_slv_min(limit);
285 pl->pl_server_lock_volume = slv;
289 * Recalculates next stats on passed \a pl.
291 * \pre ->pl_lock is locked.
293 static void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
295 int grant_plan = pl->pl_grant_plan;
296 __u64 slv = pl->pl_server_lock_volume;
297 int granted = ldlm_pool_granted(pl);
298 int grant_rate = atomic_read(&pl->pl_grant_rate);
299 int cancel_rate = atomic_read(&pl->pl_cancel_rate);
301 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT,
303 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
305 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
307 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
309 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
314 * Sets current SLV into obd accessible via ldlm_pl2ns(pl)->ns_obd.
316 static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl)
318 struct obd_device *obd;
321 * Set new SLV in obd field for using it later without accessing the
322 * pool. This is required to avoid race between sending reply to client
323 * with new SLV and cleanup server stack in which we can't guarantee
324 * that namespace is still alive. We know only that obd is alive as
325 * long as valid export is alive.
327 obd = ldlm_pl2ns(pl)->ns_obd;
328 LASSERT(obd != NULL);
329 write_lock(&obd->obd_pool_lock);
330 obd->obd_pool_slv = pl->pl_server_lock_volume;
331 write_unlock(&obd->obd_pool_lock);
335 * Recalculates all pool fields on passed \a pl.
337 * \pre ->pl_lock is not locked.
339 static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
341 time_t recalc_interval_sec;
344 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
345 if (recalc_interval_sec < pl->pl_recalc_period)
348 spin_lock(&pl->pl_lock);
349 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
350 if (recalc_interval_sec < pl->pl_recalc_period) {
351 spin_unlock(&pl->pl_lock);
355 * Recalc SLV after last period. This should be done
356 * _before_ recalculating new grant plan.
358 ldlm_pool_recalc_slv(pl);
361 * Make sure that pool informed obd of last SLV changes.
363 ldlm_srv_pool_push_slv(pl);
366 * Update grant_plan for new period.
368 ldlm_pool_recalc_grant_plan(pl);
370 pl->pl_recalc_time = cfs_time_current_sec();
371 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
372 recalc_interval_sec);
373 spin_unlock(&pl->pl_lock);
378 * This function is used on server side as main entry point for memory
379 * pressure handling. It decreases SLV on \a pl according to passed
380 * \a nr and \a gfp_mask.
382 * Our goal here is to decrease SLV such a way that clients hold \a nr
383 * locks smaller in next 10h.
385 static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
386 int nr, gfp_t gfp_mask)
391 * VM is asking how many entries may be potentially freed.
394 return ldlm_pool_granted(pl);
397 * Client already canceled locks but server is already in shrinker
398 * and can't cancel anything. Let's catch this race.
400 if (ldlm_pool_granted(pl) == 0)
403 spin_lock(&pl->pl_lock);
406 * We want shrinker to possibly cause cancellation of @nr locks from
407 * clients or grant approximately @nr locks smaller next intervals.
409 * This is why we decreased SLV by @nr. This effect will only be as
410 * long as one re-calc interval (1s these days) and this should be
411 * enough to pass this decreased SLV to all clients. On next recalc
412 * interval pool will either increase SLV if locks load is not high
413 * or will keep on same level or even decrease again, thus, shrinker
414 * decreased SLV will affect next recalc intervals and this way will
415 * make locking load lower.
417 if (nr < pl->pl_server_lock_volume) {
418 pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr;
420 limit = ldlm_pool_get_limit(pl);
421 pl->pl_server_lock_volume = ldlm_pool_slv_min(limit);
425 * Make sure that pool informed obd of last SLV changes.
427 ldlm_srv_pool_push_slv(pl);
428 spin_unlock(&pl->pl_lock);
431 * We did not really free any memory here so far, it only will be
432 * freed later may be, so that we return 0 to not confuse VM.
438 * Setup server side pool \a pl with passed \a limit.
440 static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit)
442 struct obd_device *obd;
444 obd = ldlm_pl2ns(pl)->ns_obd;
445 LASSERT(obd != NULL && obd != LP_POISON);
446 LASSERT(obd->obd_type != LP_POISON);
447 write_lock(&obd->obd_pool_lock);
448 obd->obd_pool_limit = limit;
449 write_unlock(&obd->obd_pool_lock);
451 ldlm_pool_set_limit(pl, limit);
456 * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl.
458 static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
460 struct obd_device *obd;
463 * Get new SLV and Limit from obd which is updated with coming
466 obd = ldlm_pl2ns(pl)->ns_obd;
467 LASSERT(obd != NULL);
468 read_lock(&obd->obd_pool_lock);
469 pl->pl_server_lock_volume = obd->obd_pool_slv;
470 ldlm_pool_set_limit(pl, obd->obd_pool_limit);
471 read_unlock(&obd->obd_pool_lock);
475 * Recalculates client size pool \a pl according to current SLV and Limit.
477 static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
479 time_t recalc_interval_sec;
483 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
484 if (recalc_interval_sec < pl->pl_recalc_period)
487 spin_lock(&pl->pl_lock);
489 * Check if we need to recalc lists now.
491 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
492 if (recalc_interval_sec < pl->pl_recalc_period) {
493 spin_unlock(&pl->pl_lock);
498 * Make sure that pool knows last SLV and Limit from obd.
500 ldlm_cli_pool_pop_slv(pl);
501 spin_unlock(&pl->pl_lock);
504 * Do not cancel locks in case lru resize is disabled for this ns.
506 if (!ns_connect_lru_resize(ldlm_pl2ns(pl)))
510 * In the time of canceling locks on client we do not need to maintain
511 * sharp timing, we only want to cancel locks asap according to new SLV.
512 * It may be called when SLV has changed much, this is why we do not
513 * take into account pl->pl_recalc_time here.
515 ret = ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC,
519 spin_lock(&pl->pl_lock);
521 * Time of LRU resizing might be longer than period,
522 * so update after LRU resizing rather than before it.
524 pl->pl_recalc_time = cfs_time_current_sec();
525 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
526 recalc_interval_sec);
527 spin_unlock(&pl->pl_lock);
532 * This function is main entry point for memory pressure handling on client
533 * side. Main goal of this function is to cancel some number of locks on
534 * passed \a pl according to \a nr and \a gfp_mask.
536 static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
537 int nr, gfp_t gfp_mask)
539 struct ldlm_namespace *ns;
545 * Do not cancel locks in case lru resize is disabled for this ns.
547 if (!ns_connect_lru_resize(ns))
551 * Make sure that pool knows last SLV and Limit from obd.
553 ldlm_cli_pool_pop_slv(pl);
555 spin_lock(&ns->ns_lock);
556 unused = ns->ns_nr_unused;
557 spin_unlock(&ns->ns_lock);
560 return (unused / 100) * sysctl_vfs_cache_pressure;
562 return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_CANCEL_SHRINK);
565 struct ldlm_pool_ops ldlm_srv_pool_ops = {
566 .po_recalc = ldlm_srv_pool_recalc,
567 .po_shrink = ldlm_srv_pool_shrink,
568 .po_setup = ldlm_srv_pool_setup
571 struct ldlm_pool_ops ldlm_cli_pool_ops = {
572 .po_recalc = ldlm_cli_pool_recalc,
573 .po_shrink = ldlm_cli_pool_shrink
577 * Pool recalc wrapper. Will call either client or server pool recalc callback
578 * depending what pool \a pl is used.
580 int ldlm_pool_recalc(struct ldlm_pool *pl)
582 time_t recalc_interval_sec;
585 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
586 if (recalc_interval_sec <= 0)
589 spin_lock(&pl->pl_lock);
590 if (recalc_interval_sec > 0) {
592 * Update pool statistics every 1s.
594 ldlm_pool_recalc_stats(pl);
597 * Zero out all rates and speed for the last period.
599 atomic_set(&pl->pl_grant_rate, 0);
600 atomic_set(&pl->pl_cancel_rate, 0);
602 spin_unlock(&pl->pl_lock);
605 if (pl->pl_ops->po_recalc != NULL) {
606 count = pl->pl_ops->po_recalc(pl);
607 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
610 recalc_interval_sec = pl->pl_recalc_time - cfs_time_current_sec() +
611 pl->pl_recalc_period;
612 if (recalc_interval_sec <= 0) {
613 /* Prevent too frequent recalculation. */
614 CDEBUG(D_DLMTRACE, "Negative interval(%ld), "
615 "too short period(%ld)",
617 pl->pl_recalc_period);
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);
647 EXPORT_SYMBOL(ldlm_pool_shrink);
650 * Pool setup wrapper. Will call either client or server pool recalc callback
651 * depending what pool \a pl is used.
653 * Sets passed \a limit into pool \a pl.
655 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
657 if (pl->pl_ops->po_setup != NULL)
658 return(pl->pl_ops->po_setup(pl, limit));
661 EXPORT_SYMBOL(ldlm_pool_setup);
663 static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
665 int granted, grant_rate, cancel_rate, grant_step;
666 int grant_speed, grant_plan, lvf;
667 struct ldlm_pool *pl = m->private;
671 spin_lock(&pl->pl_lock);
672 slv = pl->pl_server_lock_volume;
673 clv = pl->pl_client_lock_volume;
674 limit = ldlm_pool_get_limit(pl);
675 grant_plan = pl->pl_grant_plan;
676 granted = ldlm_pool_granted(pl);
677 grant_rate = atomic_read(&pl->pl_grant_rate);
678 cancel_rate = atomic_read(&pl->pl_cancel_rate);
679 grant_speed = grant_rate - cancel_rate;
680 lvf = atomic_read(&pl->pl_lock_volume_factor);
681 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
682 spin_unlock(&pl->pl_lock);
684 seq_printf(m, "LDLM pool state (%s):\n"
688 pl->pl_name, slv, clv, lvf);
690 if (ns_is_server(ldlm_pl2ns(pl))) {
691 seq_printf(m, " GSP: %d%%\n"
693 grant_step, grant_plan);
695 seq_printf(m, " GR: %d\n" " CR: %d\n" " GS: %d\n"
696 " G: %d\n" " L: %d\n",
697 grant_rate, cancel_rate, grant_speed,
701 LPROC_SEQ_FOPS_RO(lprocfs_pool_state);
703 static int lprocfs_grant_speed_seq_show(struct seq_file *m, void *unused)
705 struct ldlm_pool *pl = m->private;
708 spin_lock(&pl->pl_lock);
709 /* serialize with ldlm_pool_recalc */
710 grant_speed = atomic_read(&pl->pl_grant_rate) -
711 atomic_read(&pl->pl_cancel_rate);
712 spin_unlock(&pl->pl_lock);
713 return lprocfs_uint_seq_show(m, &grant_speed);
716 LDLM_POOL_PROC_READER_SEQ_SHOW(grant_plan, int);
717 LPROC_SEQ_FOPS_RO(lprocfs_grant_plan);
719 LDLM_POOL_PROC_READER_SEQ_SHOW(recalc_period, int);
720 LDLM_POOL_PROC_WRITER(recalc_period, int);
721 static ssize_t lprocfs_recalc_period_seq_write(struct file *file,
722 const char __user *buf,
723 size_t len, loff_t *off)
725 struct seq_file *seq = file->private_data;
727 return lprocfs_wr_recalc_period(file, buf, len, seq->private);
729 LPROC_SEQ_FOPS(lprocfs_recalc_period);
731 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, u64);
732 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, atomic);
733 LPROC_SEQ_FOPS_RW_TYPE(ldlm_pool_rw, atomic);
735 LPROC_SEQ_FOPS_RO(lprocfs_grant_speed);
737 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
739 struct ldlm_namespace *ns = ldlm_pl2ns(pl);
740 struct proc_dir_entry *parent_ns_proc;
741 struct lprocfs_seq_vars pool_vars[2];
742 char *var_name = NULL;
746 OBD_ALLOC(var_name, MAX_STRING_SIZE + 1);
750 parent_ns_proc = ns->ns_proc_dir_entry;
751 if (parent_ns_proc == NULL) {
752 CERROR("%s: proc entry is not initialized\n",
754 GOTO(out_free_name, rc = -EINVAL);
756 pl->pl_proc_dir = lprocfs_seq_register("pool", parent_ns_proc,
758 if (IS_ERR(pl->pl_proc_dir)) {
759 rc = PTR_ERR(pl->pl_proc_dir);
760 pl->pl_proc_dir = NULL;
761 CERROR("%s: cannot create 'pool' proc entry: rc = %d\n",
762 ldlm_ns_name(ns), rc);
763 GOTO(out_free_name, rc);
766 var_name[MAX_STRING_SIZE] = '\0';
767 memset(pool_vars, 0, sizeof(pool_vars));
768 pool_vars[0].name = var_name;
770 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "server_lock_volume",
771 &pl->pl_server_lock_volume, &ldlm_pool_u64_fops);
772 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "limit", &pl->pl_limit,
773 &ldlm_pool_rw_atomic_fops);
774 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "granted",
775 &pl->pl_granted, &ldlm_pool_atomic_fops);
776 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_speed", pl,
777 &lprocfs_grant_speed_fops);
778 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "cancel_rate",
779 &pl->pl_cancel_rate, &ldlm_pool_atomic_fops);
780 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_rate",
781 &pl->pl_grant_rate, &ldlm_pool_atomic_fops);
782 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_plan", pl,
783 &lprocfs_grant_plan_fops);
784 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "recalc_period",
785 pl, &lprocfs_recalc_period_fops);
786 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "lock_volume_factor",
787 &pl->pl_lock_volume_factor, &ldlm_pool_rw_atomic_fops);
788 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "state", pl,
789 &lprocfs_pool_state_fops);
791 pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT -
792 LDLM_POOL_FIRST_STAT, 0);
794 GOTO(out_free_name, rc = -ENOMEM);
796 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
797 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
799 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
800 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
802 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
803 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
805 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
806 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
807 "grant_rate", "locks/s");
808 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
809 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
810 "cancel_rate", "locks/s");
811 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
812 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
813 "grant_plan", "locks/s");
814 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT,
815 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
817 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT,
818 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
819 "shrink_request", "locks");
820 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT,
821 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
822 "shrink_freed", "locks");
823 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
824 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
825 "recalc_freed", "locks");
826 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
827 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
828 "recalc_timing", "sec");
829 rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
833 OBD_FREE(var_name, MAX_STRING_SIZE + 1);
837 static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
839 if (pl->pl_stats != NULL) {
840 lprocfs_free_stats(&pl->pl_stats);
843 if (pl->pl_proc_dir != NULL) {
844 lprocfs_remove(&pl->pl_proc_dir);
845 pl->pl_proc_dir = NULL;
849 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
850 int idx, ldlm_side_t client)
855 spin_lock_init(&pl->pl_lock);
856 atomic_set(&pl->pl_granted, 0);
857 pl->pl_recalc_time = cfs_time_current_sec();
858 atomic_set(&pl->pl_lock_volume_factor, 1);
860 atomic_set(&pl->pl_grant_rate, 0);
861 atomic_set(&pl->pl_cancel_rate, 0);
862 pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
864 snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
865 ldlm_ns_name(ns), idx);
867 if (client == LDLM_NAMESPACE_SERVER) {
868 pl->pl_ops = &ldlm_srv_pool_ops;
869 ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L);
870 pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD;
871 pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L);
873 ldlm_pool_set_limit(pl, 1);
874 pl->pl_server_lock_volume = 0;
875 pl->pl_ops = &ldlm_cli_pool_ops;
876 pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
878 pl->pl_client_lock_volume = 0;
879 rc = ldlm_pool_proc_init(pl);
883 CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name);
887 EXPORT_SYMBOL(ldlm_pool_init);
889 void ldlm_pool_fini(struct ldlm_pool *pl)
892 ldlm_pool_proc_fini(pl);
895 * Pool should not be used after this point. We can't free it here as
896 * it lives in struct ldlm_namespace, but still interested in catching
897 * any abnormal using cases.
899 POISON(pl, 0x5a, sizeof(*pl));
902 EXPORT_SYMBOL(ldlm_pool_fini);
905 * Add new taken ldlm lock \a lock into pool \a pl accounting.
907 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
910 * FLOCK locks are special in a sense that they are almost never
911 * cancelled, instead special kind of lock is used to drop them.
912 * also there is no LRU for flock locks, so no point in tracking
915 if (lock->l_resource->lr_type == LDLM_FLOCK)
918 atomic_inc(&pl->pl_granted);
919 atomic_inc(&pl->pl_grant_rate);
920 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
922 * Do not do pool recalc for client side as all locks which
923 * potentially may be canceled has already been packed into
924 * enqueue/cancel rpc. Also we do not want to run out of stack
925 * with too long call paths.
927 if (ns_is_server(ldlm_pl2ns(pl)))
928 ldlm_pool_recalc(pl);
930 EXPORT_SYMBOL(ldlm_pool_add);
933 * Remove ldlm lock \a lock from pool \a pl accounting.
935 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
938 * Filter out FLOCK locks. Read above comment in ldlm_pool_add().
940 if (lock->l_resource->lr_type == LDLM_FLOCK)
943 LASSERT(atomic_read(&pl->pl_granted) > 0);
944 atomic_dec(&pl->pl_granted);
945 atomic_inc(&pl->pl_cancel_rate);
947 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
949 if (ns_is_server(ldlm_pl2ns(pl)))
950 ldlm_pool_recalc(pl);
952 EXPORT_SYMBOL(ldlm_pool_del);
955 * Returns current \a pl SLV.
957 * \pre ->pl_lock is not locked.
959 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
962 spin_lock(&pl->pl_lock);
963 slv = pl->pl_server_lock_volume;
964 spin_unlock(&pl->pl_lock);
967 EXPORT_SYMBOL(ldlm_pool_get_slv);
970 * Sets passed \a slv to \a pl.
972 * \pre ->pl_lock is not locked.
974 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
976 spin_lock(&pl->pl_lock);
977 pl->pl_server_lock_volume = slv;
978 spin_unlock(&pl->pl_lock);
980 EXPORT_SYMBOL(ldlm_pool_set_slv);
983 * Returns current \a pl CLV.
985 * \pre ->pl_lock is not locked.
987 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
990 spin_lock(&pl->pl_lock);
991 slv = pl->pl_client_lock_volume;
992 spin_unlock(&pl->pl_lock);
995 EXPORT_SYMBOL(ldlm_pool_get_clv);
998 * Sets passed \a clv to \a pl.
1000 * \pre ->pl_lock is not locked.
1002 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1004 spin_lock(&pl->pl_lock);
1005 pl->pl_client_lock_volume = clv;
1006 spin_unlock(&pl->pl_lock);
1008 EXPORT_SYMBOL(ldlm_pool_set_clv);
1011 * Returns current \a pl limit.
1013 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1015 return atomic_read(&pl->pl_limit);
1017 EXPORT_SYMBOL(ldlm_pool_get_limit);
1020 * Sets passed \a limit to \a pl.
1022 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1024 atomic_set(&pl->pl_limit, limit);
1026 EXPORT_SYMBOL(ldlm_pool_set_limit);
1029 * Returns current LVF from \a pl.
1031 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1033 return atomic_read(&pl->pl_lock_volume_factor);
1035 EXPORT_SYMBOL(ldlm_pool_get_lvf);
1037 static struct ptlrpc_thread *ldlm_pools_thread;
1038 static struct shrinker *ldlm_pools_srv_shrinker;
1039 static struct shrinker *ldlm_pools_cli_shrinker;
1040 static struct completion ldlm_pools_comp;
1043 * count locks from all namespaces (if possible). Returns number of
1046 static unsigned long ldlm_pools_count(ldlm_side_t client, gfp_t gfp_mask)
1048 unsigned long total = 0;
1050 struct ldlm_namespace *ns;
1051 struct ldlm_namespace *ns_old = NULL; /* loop detection */
1054 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1057 CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n",
1058 client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1060 cookie = cl_env_reenter();
1063 * Find out how many resources we may release.
1065 for (nr_ns = ldlm_namespace_nr_read(client);
1066 nr_ns > 0; nr_ns--) {
1067 mutex_lock(ldlm_namespace_lock(client));
1068 if (list_empty(ldlm_namespace_list(client))) {
1069 mutex_unlock(ldlm_namespace_lock(client));
1070 cl_env_reexit(cookie);
1073 ns = ldlm_namespace_first_locked(client);
1076 mutex_unlock(ldlm_namespace_lock(client));
1080 if (ldlm_ns_empty(ns)) {
1081 ldlm_namespace_move_to_inactive_locked(ns, client);
1082 mutex_unlock(ldlm_namespace_lock(client));
1089 ldlm_namespace_get(ns);
1090 ldlm_namespace_move_to_active_locked(ns, client);
1091 mutex_unlock(ldlm_namespace_lock(client));
1092 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
1093 ldlm_namespace_put(ns);
1096 cl_env_reexit(cookie);
1100 static unsigned long ldlm_pools_scan(ldlm_side_t client, int nr,
1103 unsigned long freed = 0;
1105 struct ldlm_namespace *ns;
1108 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1111 cookie = cl_env_reenter();
1114 * Shrink at least ldlm_namespace_nr_read(client) namespaces.
1116 for (tmp = nr_ns = ldlm_namespace_nr_read(client);
1118 int cancel, nr_locks;
1121 * Do not call shrink under ldlm_namespace_lock(client)
1123 mutex_lock(ldlm_namespace_lock(client));
1124 if (list_empty(ldlm_namespace_list(client))) {
1125 mutex_unlock(ldlm_namespace_lock(client));
1128 ns = ldlm_namespace_first_locked(client);
1129 ldlm_namespace_get(ns);
1130 ldlm_namespace_move_to_active_locked(ns, client);
1131 mutex_unlock(ldlm_namespace_lock(client));
1133 nr_locks = ldlm_pool_granted(&ns->ns_pool);
1135 * We use to shrink propotionally but with new shrinker API,
1136 * we lost the total number of freeable locks.
1138 cancel = 1 + min_t(int, nr_locks, nr / nr_ns);
1139 freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
1140 ldlm_namespace_put(ns);
1142 cl_env_reexit(cookie);
1144 * we only decrease the SLV in server pools shrinker, return
1145 * SHRINK_STOP to kernel to avoid needless loop. LU-1128
1147 return (client == LDLM_NAMESPACE_SERVER) ? SHRINK_STOP : freed;
1150 #ifdef HAVE_SHRINKER_COUNT
1151 static unsigned long ldlm_pools_srv_count(struct shrinker *s,
1152 struct shrink_control *sc)
1154 return ldlm_pools_count(LDLM_NAMESPACE_SERVER, sc->gfp_mask);
1157 static unsigned long ldlm_pools_srv_scan(struct shrinker *s,
1158 struct shrink_control *sc)
1160 return ldlm_pools_scan(LDLM_NAMESPACE_SERVER, sc->nr_to_scan,
1164 static unsigned long ldlm_pools_cli_count(struct shrinker *s, struct shrink_control *sc)
1166 return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask);
1169 static unsigned long ldlm_pools_cli_scan(struct shrinker *s,
1170 struct shrink_control *sc)
1172 return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan,
1178 * Cancel \a nr locks from all namespaces (if possible). Returns number of
1179 * cached locks after shrink is finished. All namespaces are asked to
1180 * cancel approximately equal amount of locks to keep balancing.
1182 static int ldlm_pools_shrink(ldlm_side_t client, int nr,
1185 unsigned long total = 0;
1187 if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
1188 !(gfp_mask & __GFP_FS))
1191 CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n",
1192 nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1194 total = ldlm_pools_count(client, gfp_mask);
1196 if (nr == 0 || total == 0)
1199 return ldlm_pools_scan(client, nr, gfp_mask);
1202 static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1204 return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER,
1205 shrink_param(sc, nr_to_scan),
1206 shrink_param(sc, gfp_mask));
1209 static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1211 return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT,
1212 shrink_param(sc, nr_to_scan),
1213 shrink_param(sc, gfp_mask));
1216 #endif /* HAVE_SHRINKER_COUNT */
1218 int ldlm_pools_recalc(ldlm_side_t client)
1220 unsigned long nr_l = 0, nr_p = 0, l;
1221 struct ldlm_namespace *ns;
1222 struct ldlm_namespace *ns_old = NULL;
1224 int time = 50; /* seconds of sleep if no active namespaces */
1227 * No need to setup pool limit for client pools.
1229 if (client == LDLM_NAMESPACE_SERVER) {
1231 * Check all modest namespaces first.
1233 mutex_lock(ldlm_namespace_lock(client));
1234 list_for_each_entry(ns, ldlm_namespace_list(client),
1237 if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
1240 l = ldlm_pool_granted(&ns->ns_pool);
1245 * Set the modest pools limit equal to their avg granted
1248 l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0);
1249 ldlm_pool_setup(&ns->ns_pool, l);
1255 * Make sure that modest namespaces did not eat more that 2/3
1258 if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) {
1259 CWARN("\"Modest\" pools eat out 2/3 of server locks "
1260 "limit (%lu of %lu). This means that you have too "
1261 "many clients for this amount of server RAM. "
1262 "Upgrade server!\n", nr_l, LDLM_POOL_HOST_L);
1267 * The rest is given to greedy namespaces.
1269 list_for_each_entry(ns, ldlm_namespace_list(client),
1272 if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
1277 * In the case 2/3 locks are eaten out by
1278 * modest pools, we re-setup equal limit
1281 l = LDLM_POOL_HOST_L /
1282 ldlm_namespace_nr_read(client);
1285 * All the rest of greedy pools will have
1286 * all locks in equal parts.
1288 l = (LDLM_POOL_HOST_L - nr_l) /
1289 (ldlm_namespace_nr_read(client) -
1292 ldlm_pool_setup(&ns->ns_pool, l);
1294 mutex_unlock(ldlm_namespace_lock(client));
1298 * Recalc at least ldlm_namespace_nr(client) namespaces.
1300 for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) {
1303 * Lock the list, get first @ns in the list, getref, move it
1304 * to the tail, unlock and call pool recalc. This way we avoid
1305 * calling recalc under @ns lock what is really good as we get
1306 * rid of potential deadlock on client nodes when canceling
1307 * locks synchronously.
1309 mutex_lock(ldlm_namespace_lock(client));
1310 if (list_empty(ldlm_namespace_list(client))) {
1311 mutex_unlock(ldlm_namespace_lock(client));
1314 ns = ldlm_namespace_first_locked(client);
1316 if (ns_old == ns) { /* Full pass complete */
1317 mutex_unlock(ldlm_namespace_lock(client));
1321 /* We got an empty namespace, need to move it back to inactive
1323 * The race with parallel resource creation is fine:
1324 * - If they do namespace_get before our check, we fail the
1325 * check and they move this item to the end of the list anyway
1326 * - If we do the check and then they do namespace_get, then
1327 * we move the namespace to inactive and they will move
1328 * it back to active (synchronised by the lock, so no clash
1331 if (ldlm_ns_empty(ns)) {
1332 ldlm_namespace_move_to_inactive_locked(ns, client);
1333 mutex_unlock(ldlm_namespace_lock(client));
1340 spin_lock(&ns->ns_lock);
1342 * skip ns which is being freed, and we don't want to increase
1343 * its refcount again, not even temporarily. bz21519 & LU-499.
1345 if (ns->ns_stopping) {
1349 ldlm_namespace_get(ns);
1351 spin_unlock(&ns->ns_lock);
1353 ldlm_namespace_move_to_active_locked(ns, client);
1354 mutex_unlock(ldlm_namespace_lock(client));
1357 * After setup is done - recalc the pool.
1360 int ttime = ldlm_pool_recalc(&ns->ns_pool);
1365 ldlm_namespace_put(ns);
1370 EXPORT_SYMBOL(ldlm_pools_recalc);
1372 static int ldlm_pools_thread_main(void *arg)
1374 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
1378 thread_set_flags(thread, SVC_RUNNING);
1379 wake_up(&thread->t_ctl_waitq);
1381 CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n",
1382 "ldlm_poold", current_pid());
1385 struct l_wait_info lwi;
1388 * Recal all pools on this tick.
1390 s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
1391 c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
1394 * Wait until the next check time, or until we're
1397 lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)),
1399 l_wait_event(thread->t_ctl_waitq,
1400 thread_is_stopping(thread) ||
1401 thread_is_event(thread),
1404 if (thread_test_and_clear_flags(thread, SVC_STOPPING))
1407 thread_test_and_clear_flags(thread, SVC_EVENT);
1410 thread_set_flags(thread, SVC_STOPPED);
1411 wake_up(&thread->t_ctl_waitq);
1413 CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n",
1414 "ldlm_poold", current_pid());
1416 complete_and_exit(&ldlm_pools_comp, 0);
1419 static int ldlm_pools_thread_start(void)
1421 struct l_wait_info lwi = { 0 };
1422 struct task_struct *task;
1425 if (ldlm_pools_thread != NULL)
1428 OBD_ALLOC_PTR(ldlm_pools_thread);
1429 if (ldlm_pools_thread == NULL)
1432 init_completion(&ldlm_pools_comp);
1433 init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq);
1435 task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread,
1438 CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task));
1439 OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread));
1440 ldlm_pools_thread = NULL;
1441 RETURN(PTR_ERR(task));
1443 l_wait_event(ldlm_pools_thread->t_ctl_waitq,
1444 thread_is_running(ldlm_pools_thread), &lwi);
1448 static void ldlm_pools_thread_stop(void)
1452 if (ldlm_pools_thread == NULL) {
1457 thread_set_flags(ldlm_pools_thread, SVC_STOPPING);
1458 wake_up(&ldlm_pools_thread->t_ctl_waitq);
1461 * Make sure that pools thread is finished before freeing @thread.
1462 * This fixes possible race and oops due to accessing freed memory
1465 wait_for_completion(&ldlm_pools_comp);
1466 OBD_FREE_PTR(ldlm_pools_thread);
1467 ldlm_pools_thread = NULL;
1471 int ldlm_pools_init(void)
1474 DEF_SHRINKER_VAR(shsvar, ldlm_pools_srv_shrink,
1475 ldlm_pools_srv_count, ldlm_pools_srv_scan);
1476 DEF_SHRINKER_VAR(shcvar, ldlm_pools_cli_shrink,
1477 ldlm_pools_cli_count, ldlm_pools_cli_scan);
1480 rc = ldlm_pools_thread_start();
1482 ldlm_pools_srv_shrinker =
1483 set_shrinker(DEFAULT_SEEKS, &shsvar);
1484 ldlm_pools_cli_shrinker =
1485 set_shrinker(DEFAULT_SEEKS, &shcvar);
1489 EXPORT_SYMBOL(ldlm_pools_init);
1491 void ldlm_pools_fini(void)
1493 if (ldlm_pools_srv_shrinker != NULL) {
1494 remove_shrinker(ldlm_pools_srv_shrinker);
1495 ldlm_pools_srv_shrinker = NULL;
1497 if (ldlm_pools_cli_shrinker != NULL) {
1498 remove_shrinker(ldlm_pools_cli_shrinker);
1499 ldlm_pools_cli_shrinker = NULL;
1501 ldlm_pools_thread_stop();
1503 EXPORT_SYMBOL(ldlm_pools_fini);
1505 #else /* !HAVE_LRU_RESIZE_SUPPORT */
1506 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
1510 EXPORT_SYMBOL(ldlm_pool_setup);
1512 int ldlm_pool_recalc(struct ldlm_pool *pl)
1516 EXPORT_SYMBOL(ldlm_pool_recalc);
1518 int ldlm_pool_shrink(struct ldlm_pool *pl,
1519 int nr, gfp_t gfp_mask)
1523 EXPORT_SYMBOL(ldlm_pool_shrink);
1525 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
1526 int idx, ldlm_side_t client)
1530 EXPORT_SYMBOL(ldlm_pool_init);
1532 void ldlm_pool_fini(struct ldlm_pool *pl)
1536 EXPORT_SYMBOL(ldlm_pool_fini);
1538 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
1542 EXPORT_SYMBOL(ldlm_pool_add);
1544 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
1548 EXPORT_SYMBOL(ldlm_pool_del);
1550 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
1554 EXPORT_SYMBOL(ldlm_pool_get_slv);
1556 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
1560 EXPORT_SYMBOL(ldlm_pool_set_slv);
1562 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
1566 EXPORT_SYMBOL(ldlm_pool_get_clv);
1568 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1572 EXPORT_SYMBOL(ldlm_pool_set_clv);
1574 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1578 EXPORT_SYMBOL(ldlm_pool_get_limit);
1580 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1584 EXPORT_SYMBOL(ldlm_pool_set_limit);
1586 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1590 EXPORT_SYMBOL(ldlm_pool_get_lvf);
1592 int ldlm_pools_init(void)
1596 EXPORT_SYMBOL(ldlm_pools_init);
1598 void ldlm_pools_fini(void)
1602 EXPORT_SYMBOL(ldlm_pools_fini);
1604 int ldlm_pools_recalc(ldlm_side_t client)
1608 EXPORT_SYMBOL(ldlm_pools_recalc);
1609 #endif /* HAVE_LRU_RESIZE_SUPPORT */