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, 2014, 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 static 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 static 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) {
587 spin_lock(&pl->pl_lock);
588 recalc_interval_sec = cfs_time_current_sec() -
591 if (recalc_interval_sec > 0) {
593 * Update pool statistics every 1s.
595 ldlm_pool_recalc_stats(pl);
598 * Zero out all rates and speed for the last period.
600 atomic_set(&pl->pl_grant_rate, 0);
601 atomic_set(&pl->pl_cancel_rate, 0);
603 spin_unlock(&pl->pl_lock);
606 if (pl->pl_ops->po_recalc != NULL) {
607 count = pl->pl_ops->po_recalc(pl);
608 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
612 recalc_interval_sec = pl->pl_recalc_time - cfs_time_current_sec() +
613 pl->pl_recalc_period;
614 if (recalc_interval_sec <= 0) {
615 /* DEBUG: should be re-removed after LU-4536 is fixed */
616 CDEBUG(D_DLMTRACE, "%s: Negative interval(%ld), "
617 "too short period(%ld)\n",
618 pl->pl_name, recalc_interval_sec,
619 pl->pl_recalc_period);
621 /* Prevent too frequent recalculation. */
622 recalc_interval_sec = 1;
625 return recalc_interval_sec;
629 * Pool shrink wrapper. Will call either client or server pool recalc callback
630 * depending what pool \a pl is used.
632 int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, gfp_t gfp_mask)
636 if (pl->pl_ops->po_shrink != NULL) {
637 cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask);
639 lprocfs_counter_add(pl->pl_stats,
640 LDLM_POOL_SHRINK_REQTD_STAT,
642 lprocfs_counter_add(pl->pl_stats,
643 LDLM_POOL_SHRINK_FREED_STAT,
645 CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, "
646 "shrunk %d\n", pl->pl_name, nr, cancel);
651 EXPORT_SYMBOL(ldlm_pool_shrink);
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));
665 EXPORT_SYMBOL(ldlm_pool_setup);
667 static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
669 int granted, grant_rate, cancel_rate, grant_step;
670 int grant_speed, grant_plan, lvf;
671 struct ldlm_pool *pl = m->private;
675 spin_lock(&pl->pl_lock);
676 slv = pl->pl_server_lock_volume;
677 clv = pl->pl_client_lock_volume;
678 limit = ldlm_pool_get_limit(pl);
679 grant_plan = pl->pl_grant_plan;
680 granted = ldlm_pool_granted(pl);
681 grant_rate = atomic_read(&pl->pl_grant_rate);
682 cancel_rate = atomic_read(&pl->pl_cancel_rate);
683 grant_speed = grant_rate - cancel_rate;
684 lvf = atomic_read(&pl->pl_lock_volume_factor);
685 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
686 spin_unlock(&pl->pl_lock);
688 seq_printf(m, "LDLM pool state (%s):\n"
692 pl->pl_name, slv, clv, lvf);
694 if (ns_is_server(ldlm_pl2ns(pl))) {
695 seq_printf(m, " GSP: %d%%\n"
697 grant_step, grant_plan);
699 seq_printf(m, " GR: %d\n" " CR: %d\n" " GS: %d\n"
700 " G: %d\n" " L: %d\n",
701 grant_rate, cancel_rate, grant_speed,
705 LPROC_SEQ_FOPS_RO(lprocfs_pool_state);
707 static int lprocfs_grant_speed_seq_show(struct seq_file *m, void *unused)
709 struct ldlm_pool *pl = m->private;
712 spin_lock(&pl->pl_lock);
713 /* serialize with ldlm_pool_recalc */
714 grant_speed = atomic_read(&pl->pl_grant_rate) -
715 atomic_read(&pl->pl_cancel_rate);
716 spin_unlock(&pl->pl_lock);
717 return lprocfs_uint_seq_show(m, &grant_speed);
720 LDLM_POOL_PROC_READER_SEQ_SHOW(grant_plan, int);
721 LPROC_SEQ_FOPS_RO(lprocfs_grant_plan);
723 LDLM_POOL_PROC_READER_SEQ_SHOW(recalc_period, int);
724 LDLM_POOL_PROC_WRITER(recalc_period, int);
725 static ssize_t lprocfs_recalc_period_seq_write(struct file *file,
726 const char __user *buf,
727 size_t len, loff_t *off)
729 struct seq_file *seq = file->private_data;
731 return lprocfs_wr_recalc_period(file, buf, len, seq->private);
733 LPROC_SEQ_FOPS(lprocfs_recalc_period);
735 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, u64);
736 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, atomic);
737 LPROC_SEQ_FOPS_RW_TYPE(ldlm_pool_rw, atomic);
739 LPROC_SEQ_FOPS_RO(lprocfs_grant_speed);
741 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
743 struct ldlm_namespace *ns = ldlm_pl2ns(pl);
744 struct proc_dir_entry *parent_ns_proc;
745 struct lprocfs_vars pool_vars[2];
746 char *var_name = NULL;
750 OBD_ALLOC(var_name, MAX_STRING_SIZE + 1);
754 parent_ns_proc = ns->ns_proc_dir_entry;
755 if (parent_ns_proc == NULL) {
756 CERROR("%s: proc entry is not initialized\n",
758 GOTO(out_free_name, rc = -EINVAL);
760 pl->pl_proc_dir = lprocfs_register("pool", parent_ns_proc,
762 if (IS_ERR(pl->pl_proc_dir)) {
763 rc = PTR_ERR(pl->pl_proc_dir);
764 pl->pl_proc_dir = NULL;
765 CERROR("%s: cannot create 'pool' proc entry: rc = %d\n",
766 ldlm_ns_name(ns), rc);
767 GOTO(out_free_name, rc);
770 var_name[MAX_STRING_SIZE] = '\0';
771 memset(pool_vars, 0, sizeof(pool_vars));
772 pool_vars[0].name = var_name;
774 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "server_lock_volume",
775 &pl->pl_server_lock_volume, &ldlm_pool_u64_fops);
776 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "limit", &pl->pl_limit,
777 &ldlm_pool_rw_atomic_fops);
778 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "granted",
779 &pl->pl_granted, &ldlm_pool_atomic_fops);
780 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_speed", pl,
781 &lprocfs_grant_speed_fops);
782 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "cancel_rate",
783 &pl->pl_cancel_rate, &ldlm_pool_atomic_fops);
784 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_rate",
785 &pl->pl_grant_rate, &ldlm_pool_atomic_fops);
786 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_plan", pl,
787 &lprocfs_grant_plan_fops);
788 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "recalc_period",
789 pl, &lprocfs_recalc_period_fops);
790 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "lock_volume_factor",
791 &pl->pl_lock_volume_factor, &ldlm_pool_rw_atomic_fops);
792 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "state", pl,
793 &lprocfs_pool_state_fops);
795 pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT -
796 LDLM_POOL_FIRST_STAT, 0);
798 GOTO(out_free_name, rc = -ENOMEM);
800 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
801 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
803 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
804 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
806 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
807 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
809 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
810 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
811 "grant_rate", "locks/s");
812 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
813 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
814 "cancel_rate", "locks/s");
815 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
816 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
817 "grant_plan", "locks/s");
818 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT,
819 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
821 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT,
822 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
823 "shrink_request", "locks");
824 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT,
825 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
826 "shrink_freed", "locks");
827 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
828 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
829 "recalc_freed", "locks");
830 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
831 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
832 "recalc_timing", "sec");
833 rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
837 OBD_FREE(var_name, MAX_STRING_SIZE + 1);
841 static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
843 if (pl->pl_stats != NULL) {
844 lprocfs_free_stats(&pl->pl_stats);
847 if (pl->pl_proc_dir != NULL) {
848 lprocfs_remove(&pl->pl_proc_dir);
849 pl->pl_proc_dir = NULL;
853 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
854 int idx, ldlm_side_t client)
859 spin_lock_init(&pl->pl_lock);
860 atomic_set(&pl->pl_granted, 0);
861 pl->pl_recalc_time = cfs_time_current_sec();
862 atomic_set(&pl->pl_lock_volume_factor, 1);
864 atomic_set(&pl->pl_grant_rate, 0);
865 atomic_set(&pl->pl_cancel_rate, 0);
866 pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
868 snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
869 ldlm_ns_name(ns), idx);
871 if (client == LDLM_NAMESPACE_SERVER) {
872 pl->pl_ops = &ldlm_srv_pool_ops;
873 ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L);
874 pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD;
875 pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L);
877 ldlm_pool_set_limit(pl, 1);
878 pl->pl_server_lock_volume = 0;
879 pl->pl_ops = &ldlm_cli_pool_ops;
880 pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
882 pl->pl_client_lock_volume = 0;
883 rc = ldlm_pool_proc_init(pl);
887 CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name);
891 EXPORT_SYMBOL(ldlm_pool_init);
893 void ldlm_pool_fini(struct ldlm_pool *pl)
896 ldlm_pool_proc_fini(pl);
899 * Pool should not be used after this point. We can't free it here as
900 * it lives in struct ldlm_namespace, but still interested in catching
901 * any abnormal using cases.
903 POISON(pl, 0x5a, sizeof(*pl));
906 EXPORT_SYMBOL(ldlm_pool_fini);
909 * Add new taken ldlm lock \a lock into pool \a pl accounting.
911 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
914 * FLOCK locks are special in a sense that they are almost never
915 * cancelled, instead special kind of lock is used to drop them.
916 * also there is no LRU for flock locks, so no point in tracking
919 if (lock->l_resource->lr_type == LDLM_FLOCK)
922 atomic_inc(&pl->pl_granted);
923 atomic_inc(&pl->pl_grant_rate);
924 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
926 * Do not do pool recalc for client side as all locks which
927 * potentially may be canceled has already been packed into
928 * enqueue/cancel rpc. Also we do not want to run out of stack
929 * with too long call paths.
931 if (ns_is_server(ldlm_pl2ns(pl)))
932 ldlm_pool_recalc(pl);
934 EXPORT_SYMBOL(ldlm_pool_add);
937 * Remove ldlm lock \a lock from pool \a pl accounting.
939 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
942 * Filter out FLOCK locks. Read above comment in ldlm_pool_add().
944 if (lock->l_resource->lr_type == LDLM_FLOCK)
947 LASSERT(atomic_read(&pl->pl_granted) > 0);
948 atomic_dec(&pl->pl_granted);
949 atomic_inc(&pl->pl_cancel_rate);
951 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
953 if (ns_is_server(ldlm_pl2ns(pl)))
954 ldlm_pool_recalc(pl);
956 EXPORT_SYMBOL(ldlm_pool_del);
959 * Returns current \a pl SLV.
961 * \pre ->pl_lock is not locked.
963 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
966 spin_lock(&pl->pl_lock);
967 slv = pl->pl_server_lock_volume;
968 spin_unlock(&pl->pl_lock);
971 EXPORT_SYMBOL(ldlm_pool_get_slv);
974 * Sets passed \a slv to \a pl.
976 * \pre ->pl_lock is not locked.
978 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
980 spin_lock(&pl->pl_lock);
981 pl->pl_server_lock_volume = slv;
982 spin_unlock(&pl->pl_lock);
984 EXPORT_SYMBOL(ldlm_pool_set_slv);
987 * Returns current \a pl CLV.
989 * \pre ->pl_lock is not locked.
991 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
994 spin_lock(&pl->pl_lock);
995 slv = pl->pl_client_lock_volume;
996 spin_unlock(&pl->pl_lock);
999 EXPORT_SYMBOL(ldlm_pool_get_clv);
1002 * Sets passed \a clv to \a pl.
1004 * \pre ->pl_lock is not locked.
1006 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1008 spin_lock(&pl->pl_lock);
1009 pl->pl_client_lock_volume = clv;
1010 spin_unlock(&pl->pl_lock);
1012 EXPORT_SYMBOL(ldlm_pool_set_clv);
1015 * Returns current \a pl limit.
1017 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1019 return atomic_read(&pl->pl_limit);
1021 EXPORT_SYMBOL(ldlm_pool_get_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);
1030 EXPORT_SYMBOL(ldlm_pool_set_limit);
1033 * Returns current LVF from \a pl.
1035 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1037 return atomic_read(&pl->pl_lock_volume_factor);
1039 EXPORT_SYMBOL(ldlm_pool_get_lvf);
1041 static struct ptlrpc_thread *ldlm_pools_thread;
1042 static struct shrinker *ldlm_pools_srv_shrinker;
1043 static struct shrinker *ldlm_pools_cli_shrinker;
1044 static struct completion ldlm_pools_comp;
1047 * count locks from all namespaces (if possible). Returns number of
1050 static unsigned long ldlm_pools_count(ldlm_side_t client, gfp_t gfp_mask)
1052 unsigned long total = 0;
1054 struct ldlm_namespace *ns;
1055 struct ldlm_namespace *ns_old = NULL; /* loop detection */
1058 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1061 CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n",
1062 client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1064 cookie = cl_env_reenter();
1067 * Find out how many resources we may release.
1069 for (nr_ns = ldlm_namespace_nr_read(client);
1070 nr_ns > 0; nr_ns--) {
1071 mutex_lock(ldlm_namespace_lock(client));
1072 if (list_empty(ldlm_namespace_list(client))) {
1073 mutex_unlock(ldlm_namespace_lock(client));
1074 cl_env_reexit(cookie);
1077 ns = ldlm_namespace_first_locked(client);
1080 mutex_unlock(ldlm_namespace_lock(client));
1084 if (ldlm_ns_empty(ns)) {
1085 ldlm_namespace_move_to_inactive_locked(ns, client);
1086 mutex_unlock(ldlm_namespace_lock(client));
1093 ldlm_namespace_get(ns);
1094 ldlm_namespace_move_to_active_locked(ns, client);
1095 mutex_unlock(ldlm_namespace_lock(client));
1096 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
1097 ldlm_namespace_put(ns);
1100 cl_env_reexit(cookie);
1104 static unsigned long ldlm_pools_scan(ldlm_side_t client, int nr,
1107 unsigned long freed = 0;
1109 struct ldlm_namespace *ns;
1112 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1115 cookie = cl_env_reenter();
1118 * Shrink at least ldlm_namespace_nr_read(client) namespaces.
1120 for (tmp = nr_ns = ldlm_namespace_nr_read(client);
1122 int cancel, nr_locks;
1125 * Do not call shrink under ldlm_namespace_lock(client)
1127 mutex_lock(ldlm_namespace_lock(client));
1128 if (list_empty(ldlm_namespace_list(client))) {
1129 mutex_unlock(ldlm_namespace_lock(client));
1132 ns = ldlm_namespace_first_locked(client);
1133 ldlm_namespace_get(ns);
1134 ldlm_namespace_move_to_active_locked(ns, client);
1135 mutex_unlock(ldlm_namespace_lock(client));
1137 nr_locks = ldlm_pool_granted(&ns->ns_pool);
1139 * We use to shrink propotionally but with new shrinker API,
1140 * we lost the total number of freeable locks.
1142 cancel = 1 + min_t(int, nr_locks, nr / nr_ns);
1143 freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
1144 ldlm_namespace_put(ns);
1146 cl_env_reexit(cookie);
1148 * we only decrease the SLV in server pools shrinker, return
1149 * SHRINK_STOP to kernel to avoid needless loop. LU-1128
1151 return (client == LDLM_NAMESPACE_SERVER) ? SHRINK_STOP : freed;
1154 #ifdef HAVE_SHRINKER_COUNT
1155 static unsigned long ldlm_pools_srv_count(struct shrinker *s,
1156 struct shrink_control *sc)
1158 return ldlm_pools_count(LDLM_NAMESPACE_SERVER, sc->gfp_mask);
1161 static unsigned long ldlm_pools_srv_scan(struct shrinker *s,
1162 struct shrink_control *sc)
1164 return ldlm_pools_scan(LDLM_NAMESPACE_SERVER, sc->nr_to_scan,
1168 static unsigned long ldlm_pools_cli_count(struct shrinker *s, struct shrink_control *sc)
1170 return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask);
1173 static unsigned long ldlm_pools_cli_scan(struct shrinker *s,
1174 struct shrink_control *sc)
1176 return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan,
1182 * Cancel \a nr locks from all namespaces (if possible). Returns number of
1183 * cached locks after shrink is finished. All namespaces are asked to
1184 * cancel approximately equal amount of locks to keep balancing.
1186 static int ldlm_pools_shrink(ldlm_side_t client, int nr,
1189 unsigned long total = 0;
1191 if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
1192 !(gfp_mask & __GFP_FS))
1195 CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n",
1196 nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1198 total = ldlm_pools_count(client, gfp_mask);
1200 if (nr == 0 || total == 0)
1203 return ldlm_pools_scan(client, nr, gfp_mask);
1206 static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1208 return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER,
1209 shrink_param(sc, nr_to_scan),
1210 shrink_param(sc, gfp_mask));
1213 static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1215 return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT,
1216 shrink_param(sc, nr_to_scan),
1217 shrink_param(sc, gfp_mask));
1220 #endif /* HAVE_SHRINKER_COUNT */
1222 int ldlm_pools_recalc(ldlm_side_t client)
1224 unsigned long nr_l = 0, nr_p = 0, l;
1225 struct ldlm_namespace *ns;
1226 struct ldlm_namespace *ns_old = NULL;
1228 int time = 50; /* seconds of sleep if no active namespaces */
1231 * No need to setup pool limit for client pools.
1233 if (client == LDLM_NAMESPACE_SERVER) {
1235 * Check all modest namespaces first.
1237 mutex_lock(ldlm_namespace_lock(client));
1238 list_for_each_entry(ns, ldlm_namespace_list(client),
1241 if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
1244 l = ldlm_pool_granted(&ns->ns_pool);
1249 * Set the modest pools limit equal to their avg granted
1252 l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0);
1253 ldlm_pool_setup(&ns->ns_pool, l);
1259 * Make sure that modest namespaces did not eat more that 2/3
1262 if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) {
1263 CWARN("\"Modest\" pools eat out 2/3 of server locks "
1264 "limit (%lu of %lu). This means that you have too "
1265 "many clients for this amount of server RAM. "
1266 "Upgrade server!\n", nr_l, LDLM_POOL_HOST_L);
1271 * The rest is given to greedy namespaces.
1273 list_for_each_entry(ns, ldlm_namespace_list(client),
1276 if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
1281 * In the case 2/3 locks are eaten out by
1282 * modest pools, we re-setup equal limit
1285 l = LDLM_POOL_HOST_L /
1286 ldlm_namespace_nr_read(client);
1289 * All the rest of greedy pools will have
1290 * all locks in equal parts.
1292 l = (LDLM_POOL_HOST_L - nr_l) /
1293 (ldlm_namespace_nr_read(client) -
1296 ldlm_pool_setup(&ns->ns_pool, l);
1298 mutex_unlock(ldlm_namespace_lock(client));
1302 * Recalc at least ldlm_namespace_nr(client) namespaces.
1304 for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) {
1307 * Lock the list, get first @ns in the list, getref, move it
1308 * to the tail, unlock and call pool recalc. This way we avoid
1309 * calling recalc under @ns lock what is really good as we get
1310 * rid of potential deadlock on client nodes when canceling
1311 * locks synchronously.
1313 mutex_lock(ldlm_namespace_lock(client));
1314 if (list_empty(ldlm_namespace_list(client))) {
1315 mutex_unlock(ldlm_namespace_lock(client));
1318 ns = ldlm_namespace_first_locked(client);
1320 if (ns_old == ns) { /* Full pass complete */
1321 mutex_unlock(ldlm_namespace_lock(client));
1325 /* We got an empty namespace, need to move it back to inactive
1327 * The race with parallel resource creation is fine:
1328 * - If they do namespace_get before our check, we fail the
1329 * check and they move this item to the end of the list anyway
1330 * - If we do the check and then they do namespace_get, then
1331 * we move the namespace to inactive and they will move
1332 * it back to active (synchronised by the lock, so no clash
1335 if (ldlm_ns_empty(ns)) {
1336 ldlm_namespace_move_to_inactive_locked(ns, client);
1337 mutex_unlock(ldlm_namespace_lock(client));
1344 spin_lock(&ns->ns_lock);
1346 * skip ns which is being freed, and we don't want to increase
1347 * its refcount again, not even temporarily. bz21519 & LU-499.
1349 if (ns->ns_stopping) {
1353 ldlm_namespace_get(ns);
1355 spin_unlock(&ns->ns_lock);
1357 ldlm_namespace_move_to_active_locked(ns, client);
1358 mutex_unlock(ldlm_namespace_lock(client));
1361 * After setup is done - recalc the pool.
1364 int ttime = ldlm_pool_recalc(&ns->ns_pool);
1369 ldlm_namespace_put(ns);
1374 EXPORT_SYMBOL(ldlm_pools_recalc);
1376 static int ldlm_pools_thread_main(void *arg)
1378 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
1382 thread_set_flags(thread, SVC_RUNNING);
1383 wake_up(&thread->t_ctl_waitq);
1385 CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n",
1386 "ldlm_poold", current_pid());
1389 struct l_wait_info lwi;
1392 * Recal all pools on this tick.
1394 s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
1395 c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
1398 * Wait until the next check time, or until we're
1401 lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)),
1403 l_wait_event(thread->t_ctl_waitq,
1404 thread_is_stopping(thread) ||
1405 thread_is_event(thread),
1408 if (thread_test_and_clear_flags(thread, SVC_STOPPING))
1411 thread_test_and_clear_flags(thread, SVC_EVENT);
1414 thread_set_flags(thread, SVC_STOPPED);
1415 wake_up(&thread->t_ctl_waitq);
1417 CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n",
1418 "ldlm_poold", current_pid());
1420 complete_and_exit(&ldlm_pools_comp, 0);
1423 static int ldlm_pools_thread_start(void)
1425 struct l_wait_info lwi = { 0 };
1426 struct task_struct *task;
1429 if (ldlm_pools_thread != NULL)
1432 OBD_ALLOC_PTR(ldlm_pools_thread);
1433 if (ldlm_pools_thread == NULL)
1436 init_completion(&ldlm_pools_comp);
1437 init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq);
1439 task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread,
1442 CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task));
1443 OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread));
1444 ldlm_pools_thread = NULL;
1445 RETURN(PTR_ERR(task));
1447 l_wait_event(ldlm_pools_thread->t_ctl_waitq,
1448 thread_is_running(ldlm_pools_thread), &lwi);
1452 static void ldlm_pools_thread_stop(void)
1456 if (ldlm_pools_thread == NULL) {
1461 thread_set_flags(ldlm_pools_thread, SVC_STOPPING);
1462 wake_up(&ldlm_pools_thread->t_ctl_waitq);
1465 * Make sure that pools thread is finished before freeing @thread.
1466 * This fixes possible race and oops due to accessing freed memory
1469 wait_for_completion(&ldlm_pools_comp);
1470 OBD_FREE_PTR(ldlm_pools_thread);
1471 ldlm_pools_thread = NULL;
1475 int ldlm_pools_init(void)
1478 DEF_SHRINKER_VAR(shsvar, ldlm_pools_srv_shrink,
1479 ldlm_pools_srv_count, ldlm_pools_srv_scan);
1480 DEF_SHRINKER_VAR(shcvar, ldlm_pools_cli_shrink,
1481 ldlm_pools_cli_count, ldlm_pools_cli_scan);
1484 rc = ldlm_pools_thread_start();
1486 ldlm_pools_srv_shrinker =
1487 set_shrinker(DEFAULT_SEEKS, &shsvar);
1488 ldlm_pools_cli_shrinker =
1489 set_shrinker(DEFAULT_SEEKS, &shcvar);
1493 EXPORT_SYMBOL(ldlm_pools_init);
1495 void ldlm_pools_fini(void)
1497 if (ldlm_pools_srv_shrinker != NULL) {
1498 remove_shrinker(ldlm_pools_srv_shrinker);
1499 ldlm_pools_srv_shrinker = NULL;
1501 if (ldlm_pools_cli_shrinker != NULL) {
1502 remove_shrinker(ldlm_pools_cli_shrinker);
1503 ldlm_pools_cli_shrinker = NULL;
1505 ldlm_pools_thread_stop();
1507 EXPORT_SYMBOL(ldlm_pools_fini);
1509 #else /* !HAVE_LRU_RESIZE_SUPPORT */
1510 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
1514 EXPORT_SYMBOL(ldlm_pool_setup);
1516 int ldlm_pool_recalc(struct ldlm_pool *pl)
1520 EXPORT_SYMBOL(ldlm_pool_recalc);
1522 int ldlm_pool_shrink(struct ldlm_pool *pl,
1523 int nr, gfp_t gfp_mask)
1527 EXPORT_SYMBOL(ldlm_pool_shrink);
1529 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
1530 int idx, ldlm_side_t client)
1534 EXPORT_SYMBOL(ldlm_pool_init);
1536 void ldlm_pool_fini(struct ldlm_pool *pl)
1540 EXPORT_SYMBOL(ldlm_pool_fini);
1542 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
1546 EXPORT_SYMBOL(ldlm_pool_add);
1548 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
1552 EXPORT_SYMBOL(ldlm_pool_del);
1554 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
1558 EXPORT_SYMBOL(ldlm_pool_get_slv);
1560 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
1564 EXPORT_SYMBOL(ldlm_pool_set_slv);
1566 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
1570 EXPORT_SYMBOL(ldlm_pool_get_clv);
1572 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1576 EXPORT_SYMBOL(ldlm_pool_set_clv);
1578 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1582 EXPORT_SYMBOL(ldlm_pool_get_limit);
1584 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1588 EXPORT_SYMBOL(ldlm_pool_set_limit);
1590 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1594 EXPORT_SYMBOL(ldlm_pool_get_lvf);
1596 int ldlm_pools_init(void)
1600 EXPORT_SYMBOL(ldlm_pools_init);
1602 void ldlm_pools_fini(void)
1606 EXPORT_SYMBOL(ldlm_pools_fini);
1608 int ldlm_pools_recalc(ldlm_side_t client)
1612 EXPORT_SYMBOL(ldlm_pools_recalc);
1613 #endif /* HAVE_LRU_RESIZE_SUPPORT */