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));
219 * Recalculates next grant limit on passed \a pl.
221 * \pre ->pl_lock is locked.
223 static void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl)
225 int granted, grant_step, limit;
227 limit = ldlm_pool_get_limit(pl);
228 granted = atomic_read(&pl->pl_granted);
230 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
231 grant_step = ((limit - granted) * grant_step) / 100;
232 pl->pl_grant_plan = granted + grant_step;
233 limit = (limit * 5) >> 2;
234 if (pl->pl_grant_plan > limit)
235 pl->pl_grant_plan = limit;
239 * Recalculates next SLV on passed \a pl.
241 * \pre ->pl_lock is locked.
243 static void ldlm_pool_recalc_slv(struct ldlm_pool *pl)
253 slv = pl->pl_server_lock_volume;
254 grant_plan = pl->pl_grant_plan;
255 limit = ldlm_pool_get_limit(pl);
256 granted = atomic_read(&pl->pl_granted);
257 round_up = granted < limit;
259 grant_usage = max_t(int, limit - (granted - grant_plan), 1);
262 * Find out SLV change factor which is the ratio of grant usage
263 * from limit. SLV changes as fast as the ratio of grant plan
264 * consumption. The more locks from grant plan are not consumed
265 * by clients in last interval (idle time), the faster grows
266 * SLV. And the opposite, the more grant plan is over-consumed
267 * (load time) the faster drops SLV.
269 slv_factor = (grant_usage << LDLM_POOL_SLV_SHIFT);
270 do_div(slv_factor, limit);
271 slv = slv * slv_factor;
272 slv = dru(slv, LDLM_POOL_SLV_SHIFT, round_up);
274 if (slv > ldlm_pool_slv_max(limit)) {
275 slv = ldlm_pool_slv_max(limit);
276 } else if (slv < ldlm_pool_slv_min(limit)) {
277 slv = ldlm_pool_slv_min(limit);
280 pl->pl_server_lock_volume = slv;
284 * Recalculates next stats on passed \a pl.
286 * \pre ->pl_lock is locked.
288 static void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
290 int grant_plan = pl->pl_grant_plan;
291 __u64 slv = pl->pl_server_lock_volume;
292 int granted = atomic_read(&pl->pl_granted);
293 int grant_rate = atomic_read(&pl->pl_grant_rate);
294 int cancel_rate = atomic_read(&pl->pl_cancel_rate);
296 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT,
298 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
300 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
302 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
304 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
309 * Sets current SLV into obd accessible via ldlm_pl2ns(pl)->ns_obd.
311 static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl)
313 struct obd_device *obd;
316 * Set new SLV in obd field for using it later without accessing the
317 * pool. This is required to avoid race between sending reply to client
318 * with new SLV and cleanup server stack in which we can't guarantee
319 * that namespace is still alive. We know only that obd is alive as
320 * long as valid export is alive.
322 obd = ldlm_pl2ns(pl)->ns_obd;
323 LASSERT(obd != NULL);
324 write_lock(&obd->obd_pool_lock);
325 obd->obd_pool_slv = pl->pl_server_lock_volume;
326 write_unlock(&obd->obd_pool_lock);
330 * Recalculates all pool fields on passed \a pl.
332 * \pre ->pl_lock is not locked.
334 static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
336 time_t recalc_interval_sec;
339 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
340 if (recalc_interval_sec < pl->pl_recalc_period)
343 spin_lock(&pl->pl_lock);
344 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
345 if (recalc_interval_sec < pl->pl_recalc_period) {
346 spin_unlock(&pl->pl_lock);
350 * Recalc SLV after last period. This should be done
351 * _before_ recalculating new grant plan.
353 ldlm_pool_recalc_slv(pl);
356 * Make sure that pool informed obd of last SLV changes.
358 ldlm_srv_pool_push_slv(pl);
361 * Update grant_plan for new period.
363 ldlm_pool_recalc_grant_plan(pl);
365 pl->pl_recalc_time = cfs_time_current_sec();
366 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
367 recalc_interval_sec);
368 spin_unlock(&pl->pl_lock);
373 * This function is used on server side as main entry point for memory
374 * pressure handling. It decreases SLV on \a pl according to passed
375 * \a nr and \a gfp_mask.
377 * Our goal here is to decrease SLV such a way that clients hold \a nr
378 * locks smaller in next 10h.
380 static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
381 int nr, gfp_t gfp_mask)
386 * VM is asking how many entries may be potentially freed.
389 return atomic_read(&pl->pl_granted);
392 * Client already canceled locks but server is already in shrinker
393 * and can't cancel anything. Let's catch this race.
395 if (atomic_read(&pl->pl_granted) == 0)
398 spin_lock(&pl->pl_lock);
401 * We want shrinker to possibly cause cancellation of @nr locks from
402 * clients or grant approximately @nr locks smaller next intervals.
404 * This is why we decreased SLV by @nr. This effect will only be as
405 * long as one re-calc interval (1s these days) and this should be
406 * enough to pass this decreased SLV to all clients. On next recalc
407 * interval pool will either increase SLV if locks load is not high
408 * or will keep on same level or even decrease again, thus, shrinker
409 * decreased SLV will affect next recalc intervals and this way will
410 * make locking load lower.
412 if (nr < pl->pl_server_lock_volume) {
413 pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr;
415 limit = ldlm_pool_get_limit(pl);
416 pl->pl_server_lock_volume = ldlm_pool_slv_min(limit);
420 * Make sure that pool informed obd of last SLV changes.
422 ldlm_srv_pool_push_slv(pl);
423 spin_unlock(&pl->pl_lock);
426 * We did not really free any memory here so far, it only will be
427 * freed later may be, so that we return 0 to not confuse VM.
433 * Setup server side pool \a pl with passed \a limit.
435 static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit)
437 struct obd_device *obd;
439 obd = ldlm_pl2ns(pl)->ns_obd;
440 LASSERT(obd != NULL && obd != LP_POISON);
441 LASSERT(obd->obd_type != LP_POISON);
442 write_lock(&obd->obd_pool_lock);
443 obd->obd_pool_limit = limit;
444 write_unlock(&obd->obd_pool_lock);
446 ldlm_pool_set_limit(pl, limit);
451 * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl.
453 static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
455 struct obd_device *obd;
458 * Get new SLV and Limit from obd which is updated with coming
461 obd = ldlm_pl2ns(pl)->ns_obd;
462 LASSERT(obd != NULL);
463 read_lock(&obd->obd_pool_lock);
464 pl->pl_server_lock_volume = obd->obd_pool_slv;
465 ldlm_pool_set_limit(pl, obd->obd_pool_limit);
466 read_unlock(&obd->obd_pool_lock);
470 * Recalculates client size pool \a pl according to current SLV and Limit.
472 static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
474 time_t recalc_interval_sec;
478 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
479 if (recalc_interval_sec < pl->pl_recalc_period)
482 spin_lock(&pl->pl_lock);
484 * Check if we need to recalc lists now.
486 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
487 if (recalc_interval_sec < pl->pl_recalc_period) {
488 spin_unlock(&pl->pl_lock);
493 * Make sure that pool knows last SLV and Limit from obd.
495 ldlm_cli_pool_pop_slv(pl);
496 spin_unlock(&pl->pl_lock);
499 * Do not cancel locks in case lru resize is disabled for this ns.
501 if (!ns_connect_lru_resize(ldlm_pl2ns(pl)))
505 * In the time of canceling locks on client we do not need to maintain
506 * sharp timing, we only want to cancel locks asap according to new SLV.
507 * It may be called when SLV has changed much, this is why we do not
508 * take into account pl->pl_recalc_time here.
510 ret = ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC,
514 spin_lock(&pl->pl_lock);
516 * Time of LRU resizing might be longer than period,
517 * so update after LRU resizing rather than before it.
519 pl->pl_recalc_time = cfs_time_current_sec();
520 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
521 recalc_interval_sec);
522 spin_unlock(&pl->pl_lock);
527 * This function is main entry point for memory pressure handling on client
528 * side. Main goal of this function is to cancel some number of locks on
529 * passed \a pl according to \a nr and \a gfp_mask.
531 static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
532 int nr, gfp_t gfp_mask)
534 struct ldlm_namespace *ns;
540 * Do not cancel locks in case lru resize is disabled for this ns.
542 if (!ns_connect_lru_resize(ns))
546 * Make sure that pool knows last SLV and Limit from obd.
548 ldlm_cli_pool_pop_slv(pl);
550 spin_lock(&ns->ns_lock);
551 unused = ns->ns_nr_unused;
552 spin_unlock(&ns->ns_lock);
555 return (unused / 100) * sysctl_vfs_cache_pressure;
557 return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_CANCEL_SHRINK);
560 struct ldlm_pool_ops ldlm_srv_pool_ops = {
561 .po_recalc = ldlm_srv_pool_recalc,
562 .po_shrink = ldlm_srv_pool_shrink,
563 .po_setup = ldlm_srv_pool_setup
566 struct ldlm_pool_ops ldlm_cli_pool_ops = {
567 .po_recalc = ldlm_cli_pool_recalc,
568 .po_shrink = ldlm_cli_pool_shrink
572 * Pool recalc wrapper. Will call either client or server pool recalc callback
573 * depending what pool \a pl is used.
575 int ldlm_pool_recalc(struct ldlm_pool *pl)
577 time_t recalc_interval_sec;
580 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
581 if (recalc_interval_sec <= 0)
584 spin_lock(&pl->pl_lock);
585 if (recalc_interval_sec > 0) {
587 * Update pool statistics every 1s.
589 ldlm_pool_recalc_stats(pl);
592 * Zero out all rates and speed for the last period.
594 atomic_set(&pl->pl_grant_rate, 0);
595 atomic_set(&pl->pl_cancel_rate, 0);
597 spin_unlock(&pl->pl_lock);
600 if (pl->pl_ops->po_recalc != NULL) {
601 count = pl->pl_ops->po_recalc(pl);
602 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
605 recalc_interval_sec = pl->pl_recalc_time - cfs_time_current_sec() +
606 pl->pl_recalc_period;
607 if (recalc_interval_sec <= 0) {
608 /* Prevent too frequent recalculation. */
609 CDEBUG(D_DLMTRACE, "Negative interval(%ld), "
610 "too short period(%ld)",
612 pl->pl_recalc_period);
613 recalc_interval_sec = 1;
616 return recalc_interval_sec;
620 * Pool shrink wrapper. Will call either client or server pool recalc callback
621 * depending what pool \a pl is used.
623 int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, gfp_t gfp_mask)
627 if (pl->pl_ops->po_shrink != NULL) {
628 cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask);
630 lprocfs_counter_add(pl->pl_stats,
631 LDLM_POOL_SHRINK_REQTD_STAT,
633 lprocfs_counter_add(pl->pl_stats,
634 LDLM_POOL_SHRINK_FREED_STAT,
636 CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, "
637 "shrunk %d\n", pl->pl_name, nr, cancel);
642 EXPORT_SYMBOL(ldlm_pool_shrink);
645 * Pool setup wrapper. Will call either client or server pool recalc callback
646 * depending what pool \a pl is used.
648 * Sets passed \a limit into pool \a pl.
650 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
652 if (pl->pl_ops->po_setup != NULL)
653 return(pl->pl_ops->po_setup(pl, limit));
656 EXPORT_SYMBOL(ldlm_pool_setup);
658 static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
660 int granted, grant_rate, cancel_rate, grant_step;
661 int grant_speed, grant_plan, lvf;
662 struct ldlm_pool *pl = m->private;
666 spin_lock(&pl->pl_lock);
667 slv = pl->pl_server_lock_volume;
668 clv = pl->pl_client_lock_volume;
669 limit = ldlm_pool_get_limit(pl);
670 grant_plan = pl->pl_grant_plan;
671 granted = atomic_read(&pl->pl_granted);
672 grant_rate = atomic_read(&pl->pl_grant_rate);
673 cancel_rate = atomic_read(&pl->pl_cancel_rate);
674 grant_speed = grant_rate - cancel_rate;
675 lvf = atomic_read(&pl->pl_lock_volume_factor);
676 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
677 spin_unlock(&pl->pl_lock);
679 seq_printf(m, "LDLM pool state (%s):\n"
683 pl->pl_name, slv, clv, lvf);
685 if (ns_is_server(ldlm_pl2ns(pl))) {
686 seq_printf(m, " GSP: %d%%\n"
688 grant_step, grant_plan);
690 seq_printf(m, " GR: %d\n" " CR: %d\n" " GS: %d\n"
691 " G: %d\n" " L: %d\n",
692 grant_rate, cancel_rate, grant_speed,
696 LPROC_SEQ_FOPS_RO(lprocfs_pool_state);
698 static int lprocfs_grant_speed_seq_show(struct seq_file *m, void *unused)
700 struct ldlm_pool *pl = m->private;
703 spin_lock(&pl->pl_lock);
704 /* serialize with ldlm_pool_recalc */
705 grant_speed = atomic_read(&pl->pl_grant_rate) -
706 atomic_read(&pl->pl_cancel_rate);
707 spin_unlock(&pl->pl_lock);
708 return lprocfs_uint_seq_show(m, &grant_speed);
711 LDLM_POOL_PROC_READER_SEQ_SHOW(grant_plan, int);
712 LPROC_SEQ_FOPS_RO(lprocfs_grant_plan);
714 LDLM_POOL_PROC_READER_SEQ_SHOW(recalc_period, int);
715 LDLM_POOL_PROC_WRITER(recalc_period, int);
716 static ssize_t lprocfs_recalc_period_seq_write(struct file *file,
717 const char __user *buf,
718 size_t len, loff_t *off)
720 struct seq_file *seq = file->private_data;
722 return lprocfs_wr_recalc_period(file, buf, len, seq->private);
724 LPROC_SEQ_FOPS(lprocfs_recalc_period);
726 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, u64);
727 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, atomic);
728 LPROC_SEQ_FOPS_RW_TYPE(ldlm_pool_rw, atomic);
730 LPROC_SEQ_FOPS_RO(lprocfs_grant_speed);
732 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
734 struct ldlm_namespace *ns = ldlm_pl2ns(pl);
735 struct proc_dir_entry *parent_ns_proc;
736 struct lprocfs_seq_vars pool_vars[2];
737 char *var_name = NULL;
741 OBD_ALLOC(var_name, MAX_STRING_SIZE + 1);
745 parent_ns_proc = ns->ns_proc_dir_entry;
746 if (parent_ns_proc == NULL) {
747 CERROR("%s: proc entry is not initialized\n",
749 GOTO(out_free_name, rc = -EINVAL);
751 pl->pl_proc_dir = lprocfs_seq_register("pool", parent_ns_proc,
753 if (IS_ERR(pl->pl_proc_dir)) {
754 rc = PTR_ERR(pl->pl_proc_dir);
755 pl->pl_proc_dir = NULL;
756 CERROR("%s: cannot create 'pool' proc entry: rc = %d\n",
757 ldlm_ns_name(ns), rc);
758 GOTO(out_free_name, rc);
761 var_name[MAX_STRING_SIZE] = '\0';
762 memset(pool_vars, 0, sizeof(pool_vars));
763 pool_vars[0].name = var_name;
765 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "server_lock_volume",
766 &pl->pl_server_lock_volume, &ldlm_pool_u64_fops);
767 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "limit", &pl->pl_limit,
768 &ldlm_pool_rw_atomic_fops);
769 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "granted",
770 &pl->pl_granted, &ldlm_pool_atomic_fops);
771 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_speed", pl,
772 &lprocfs_grant_speed_fops);
773 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "cancel_rate",
774 &pl->pl_cancel_rate, &ldlm_pool_atomic_fops);
775 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_rate",
776 &pl->pl_grant_rate, &ldlm_pool_atomic_fops);
777 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_plan", pl,
778 &lprocfs_grant_plan_fops);
779 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "recalc_period",
780 pl, &lprocfs_recalc_period_fops);
781 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "lock_volume_factor",
782 &pl->pl_lock_volume_factor, &ldlm_pool_rw_atomic_fops);
783 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "state", pl,
784 &lprocfs_pool_state_fops);
786 pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT -
787 LDLM_POOL_FIRST_STAT, 0);
789 GOTO(out_free_name, rc = -ENOMEM);
791 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
792 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
794 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
795 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
797 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
798 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
800 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
801 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
802 "grant_rate", "locks/s");
803 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
804 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
805 "cancel_rate", "locks/s");
806 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
807 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
808 "grant_plan", "locks/s");
809 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT,
810 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
812 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT,
813 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
814 "shrink_request", "locks");
815 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT,
816 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
817 "shrink_freed", "locks");
818 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
819 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
820 "recalc_freed", "locks");
821 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
822 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
823 "recalc_timing", "sec");
824 rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
828 OBD_FREE(var_name, MAX_STRING_SIZE + 1);
832 static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
834 if (pl->pl_stats != NULL) {
835 lprocfs_free_stats(&pl->pl_stats);
838 if (pl->pl_proc_dir != NULL) {
839 lprocfs_remove(&pl->pl_proc_dir);
840 pl->pl_proc_dir = NULL;
844 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
845 int idx, ldlm_side_t client)
850 spin_lock_init(&pl->pl_lock);
851 atomic_set(&pl->pl_granted, 0);
852 pl->pl_recalc_time = cfs_time_current_sec();
853 atomic_set(&pl->pl_lock_volume_factor, 1);
855 atomic_set(&pl->pl_grant_rate, 0);
856 atomic_set(&pl->pl_cancel_rate, 0);
857 pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
859 snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
860 ldlm_ns_name(ns), idx);
862 if (client == LDLM_NAMESPACE_SERVER) {
863 pl->pl_ops = &ldlm_srv_pool_ops;
864 ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L);
865 pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD;
866 pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L);
868 ldlm_pool_set_limit(pl, 1);
869 pl->pl_server_lock_volume = 0;
870 pl->pl_ops = &ldlm_cli_pool_ops;
871 pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
873 pl->pl_client_lock_volume = 0;
874 rc = ldlm_pool_proc_init(pl);
878 CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name);
882 EXPORT_SYMBOL(ldlm_pool_init);
884 void ldlm_pool_fini(struct ldlm_pool *pl)
887 ldlm_pool_proc_fini(pl);
890 * Pool should not be used after this point. We can't free it here as
891 * it lives in struct ldlm_namespace, but still interested in catching
892 * any abnormal using cases.
894 POISON(pl, 0x5a, sizeof(*pl));
897 EXPORT_SYMBOL(ldlm_pool_fini);
900 * Add new taken ldlm lock \a lock into pool \a pl accounting.
902 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
905 * FLOCK locks are special in a sense that they are almost never
906 * cancelled, instead special kind of lock is used to drop them.
907 * also there is no LRU for flock locks, so no point in tracking
910 if (lock->l_resource->lr_type == LDLM_FLOCK)
913 atomic_inc(&pl->pl_granted);
914 atomic_inc(&pl->pl_grant_rate);
915 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
917 * Do not do pool recalc for client side as all locks which
918 * potentially may be canceled has already been packed into
919 * enqueue/cancel rpc. Also we do not want to run out of stack
920 * with too long call paths.
922 if (ns_is_server(ldlm_pl2ns(pl)))
923 ldlm_pool_recalc(pl);
925 EXPORT_SYMBOL(ldlm_pool_add);
928 * Remove ldlm lock \a lock from pool \a pl accounting.
930 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
933 * Filter out FLOCK locks. Read above comment in ldlm_pool_add().
935 if (lock->l_resource->lr_type == LDLM_FLOCK)
938 LASSERT(atomic_read(&pl->pl_granted) > 0);
939 atomic_dec(&pl->pl_granted);
940 atomic_inc(&pl->pl_cancel_rate);
942 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
944 if (ns_is_server(ldlm_pl2ns(pl)))
945 ldlm_pool_recalc(pl);
947 EXPORT_SYMBOL(ldlm_pool_del);
950 * Returns current \a pl SLV.
952 * \pre ->pl_lock is not locked.
954 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
957 spin_lock(&pl->pl_lock);
958 slv = pl->pl_server_lock_volume;
959 spin_unlock(&pl->pl_lock);
962 EXPORT_SYMBOL(ldlm_pool_get_slv);
965 * Sets passed \a slv to \a pl.
967 * \pre ->pl_lock is not locked.
969 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
971 spin_lock(&pl->pl_lock);
972 pl->pl_server_lock_volume = slv;
973 spin_unlock(&pl->pl_lock);
975 EXPORT_SYMBOL(ldlm_pool_set_slv);
978 * Returns current \a pl CLV.
980 * \pre ->pl_lock is not locked.
982 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
985 spin_lock(&pl->pl_lock);
986 slv = pl->pl_client_lock_volume;
987 spin_unlock(&pl->pl_lock);
990 EXPORT_SYMBOL(ldlm_pool_get_clv);
993 * Sets passed \a clv to \a pl.
995 * \pre ->pl_lock is not locked.
997 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
999 spin_lock(&pl->pl_lock);
1000 pl->pl_client_lock_volume = clv;
1001 spin_unlock(&pl->pl_lock);
1003 EXPORT_SYMBOL(ldlm_pool_set_clv);
1006 * Returns current \a pl limit.
1008 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1010 return atomic_read(&pl->pl_limit);
1012 EXPORT_SYMBOL(ldlm_pool_get_limit);
1015 * Sets passed \a limit to \a pl.
1017 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1019 atomic_set(&pl->pl_limit, limit);
1021 EXPORT_SYMBOL(ldlm_pool_set_limit);
1024 * Returns current LVF from \a pl.
1026 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1028 return atomic_read(&pl->pl_lock_volume_factor);
1030 EXPORT_SYMBOL(ldlm_pool_get_lvf);
1032 static unsigned int ldlm_pool_granted(struct ldlm_pool *pl)
1034 return atomic_read(&pl->pl_granted);
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 int total = 0, nr_ns;
1049 struct ldlm_namespace *ns;
1050 struct ldlm_namespace *ns_old = NULL; /* loop detection */
1053 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1056 CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n",
1057 client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1059 cookie = cl_env_reenter();
1062 * Find out how many resources we may release.
1064 for (nr_ns = ldlm_namespace_nr_read(client);
1065 nr_ns > 0; nr_ns--) {
1066 mutex_lock(ldlm_namespace_lock(client));
1067 if (list_empty(ldlm_namespace_list(client))) {
1068 mutex_unlock(ldlm_namespace_lock(client));
1069 cl_env_reexit(cookie);
1072 ns = ldlm_namespace_first_locked(client);
1075 mutex_unlock(ldlm_namespace_lock(client));
1079 if (ldlm_ns_empty(ns)) {
1080 ldlm_namespace_move_to_inactive_locked(ns, client);
1081 mutex_unlock(ldlm_namespace_lock(client));
1088 ldlm_namespace_get(ns);
1089 ldlm_namespace_move_to_active_locked(ns, client);
1090 mutex_unlock(ldlm_namespace_lock(client));
1091 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
1092 ldlm_namespace_put(ns);
1095 cl_env_reexit(cookie);
1099 static unsigned long ldlm_pools_scan(ldlm_side_t client, int nr,
1102 unsigned long freed = 0;
1104 struct ldlm_namespace *ns;
1107 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1110 cookie = cl_env_reenter();
1113 * Shrink at least ldlm_namespace_nr_read(client) namespaces.
1115 for (tmp = nr_ns = ldlm_namespace_nr_read(client);
1117 int cancel, nr_locks;
1120 * Do not call shrink under ldlm_namespace_lock(client)
1122 mutex_lock(ldlm_namespace_lock(client));
1123 if (list_empty(ldlm_namespace_list(client))) {
1124 mutex_unlock(ldlm_namespace_lock(client));
1127 ns = ldlm_namespace_first_locked(client);
1128 ldlm_namespace_get(ns);
1129 ldlm_namespace_move_to_active_locked(ns, client);
1130 mutex_unlock(ldlm_namespace_lock(client));
1132 nr_locks = ldlm_pool_granted(&ns->ns_pool);
1134 * We use to shrink propotionally but with new shrinker API,
1135 * we lost the total number of freeable locks.
1137 cancel = 1 + min_t(int, nr_locks, nr / nr_ns);
1138 freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
1139 ldlm_namespace_put(ns);
1141 cl_env_reexit(cookie);
1143 * we only decrease the SLV in server pools shrinker, return
1144 * SHRINK_STOP to kernel to avoid needless loop. LU-1128
1146 return (client == LDLM_NAMESPACE_SERVER) ? SHRINK_STOP : freed;
1149 #ifdef HAVE_SHRINKER_COUNT
1150 static unsigned long ldlm_pools_srv_count(struct shrinker *s,
1151 struct shrink_control *sc)
1153 return ldlm_pools_count(LDLM_NAMESPACE_SERVER, sc->gfp_mask);
1156 static unsigned long ldlm_pools_srv_scan(struct shrinker *s,
1157 struct shrink_control *sc)
1159 return ldlm_pools_scan(LDLM_NAMESPACE_SERVER, sc->nr_to_scan,
1163 static unsigned long ldlm_pools_cli_count(struct shrinker *s, struct shrink_control *sc)
1165 return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask);
1168 static unsigned long ldlm_pools_cli_scan(struct shrinker *s,
1169 struct shrink_control *sc)
1171 return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan,
1177 * Cancel \a nr locks from all namespaces (if possible). Returns number of
1178 * cached locks after shrink is finished. All namespaces are asked to
1179 * cancel approximately equal amount of locks to keep balancing.
1181 static int ldlm_pools_shrink(ldlm_side_t client, int nr,
1184 unsigned int total = 0;
1186 if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
1187 !(gfp_mask & __GFP_FS))
1190 CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n",
1191 nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1193 total = ldlm_pools_count(client, gfp_mask);
1195 if (nr == 0 || total == 0)
1198 return ldlm_pools_scan(client, nr, gfp_mask);
1201 static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1203 return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER,
1204 shrink_param(sc, nr_to_scan),
1205 shrink_param(sc, gfp_mask));
1208 static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1210 return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT,
1211 shrink_param(sc, nr_to_scan),
1212 shrink_param(sc, gfp_mask));
1215 #endif /* HAVE_SHRINKER_COUNT */
1217 int ldlm_pools_recalc(ldlm_side_t client)
1219 __u32 nr_l = 0, nr_p = 0, l;
1220 struct ldlm_namespace *ns;
1221 struct ldlm_namespace *ns_old = NULL;
1223 int time = 50; /* seconds of sleep if no active namespaces */
1226 * No need to setup pool limit for client pools.
1228 if (client == LDLM_NAMESPACE_SERVER) {
1230 * Check all modest namespaces first.
1232 mutex_lock(ldlm_namespace_lock(client));
1233 list_for_each_entry(ns, ldlm_namespace_list(client),
1236 if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
1239 l = ldlm_pool_granted(&ns->ns_pool);
1244 * Set the modest pools limit equal to their avg granted
1247 l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0);
1248 ldlm_pool_setup(&ns->ns_pool, l);
1254 * Make sure that modest namespaces did not eat more that 2/3
1257 if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) {
1258 CWARN("\"Modest\" pools eat out 2/3 of server locks "
1259 "limit (%d of %lu). This means that you have too "
1260 "many clients for this amount of server RAM. "
1261 "Upgrade server!\n", nr_l, LDLM_POOL_HOST_L);
1266 * The rest is given to greedy namespaces.
1268 list_for_each_entry(ns, ldlm_namespace_list(client),
1271 if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
1276 * In the case 2/3 locks are eaten out by
1277 * modest pools, we re-setup equal limit
1280 l = LDLM_POOL_HOST_L /
1281 ldlm_namespace_nr_read(client);
1284 * All the rest of greedy pools will have
1285 * all locks in equal parts.
1287 l = (LDLM_POOL_HOST_L - nr_l) /
1288 (ldlm_namespace_nr_read(client) -
1291 ldlm_pool_setup(&ns->ns_pool, l);
1293 mutex_unlock(ldlm_namespace_lock(client));
1297 * Recalc at least ldlm_namespace_nr(client) namespaces.
1299 for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) {
1302 * Lock the list, get first @ns in the list, getref, move it
1303 * to the tail, unlock and call pool recalc. This way we avoid
1304 * calling recalc under @ns lock what is really good as we get
1305 * rid of potential deadlock on client nodes when canceling
1306 * locks synchronously.
1308 mutex_lock(ldlm_namespace_lock(client));
1309 if (list_empty(ldlm_namespace_list(client))) {
1310 mutex_unlock(ldlm_namespace_lock(client));
1313 ns = ldlm_namespace_first_locked(client);
1315 if (ns_old == ns) { /* Full pass complete */
1316 mutex_unlock(ldlm_namespace_lock(client));
1320 /* We got an empty namespace, need to move it back to inactive
1322 * The race with parallel resource creation is fine:
1323 * - If they do namespace_get before our check, we fail the
1324 * check and they move this item to the end of the list anyway
1325 * - If we do the check and then they do namespace_get, then
1326 * we move the namespace to inactive and they will move
1327 * it back to active (synchronised by the lock, so no clash
1330 if (ldlm_ns_empty(ns)) {
1331 ldlm_namespace_move_to_inactive_locked(ns, client);
1332 mutex_unlock(ldlm_namespace_lock(client));
1339 spin_lock(&ns->ns_lock);
1341 * skip ns which is being freed, and we don't want to increase
1342 * its refcount again, not even temporarily. bz21519 & LU-499.
1344 if (ns->ns_stopping) {
1348 ldlm_namespace_get(ns);
1350 spin_unlock(&ns->ns_lock);
1352 ldlm_namespace_move_to_active_locked(ns, client);
1353 mutex_unlock(ldlm_namespace_lock(client));
1356 * After setup is done - recalc the pool.
1359 int ttime = ldlm_pool_recalc(&ns->ns_pool);
1364 ldlm_namespace_put(ns);
1369 EXPORT_SYMBOL(ldlm_pools_recalc);
1371 static int ldlm_pools_thread_main(void *arg)
1373 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
1377 thread_set_flags(thread, SVC_RUNNING);
1378 wake_up(&thread->t_ctl_waitq);
1380 CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n",
1381 "ldlm_poold", current_pid());
1384 struct l_wait_info lwi;
1387 * Recal all pools on this tick.
1389 s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
1390 c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
1393 * Wait until the next check time, or until we're
1396 lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)),
1398 l_wait_event(thread->t_ctl_waitq,
1399 thread_is_stopping(thread) ||
1400 thread_is_event(thread),
1403 if (thread_test_and_clear_flags(thread, SVC_STOPPING))
1406 thread_test_and_clear_flags(thread, SVC_EVENT);
1409 thread_set_flags(thread, SVC_STOPPED);
1410 wake_up(&thread->t_ctl_waitq);
1412 CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n",
1413 "ldlm_poold", current_pid());
1415 complete_and_exit(&ldlm_pools_comp, 0);
1418 static int ldlm_pools_thread_start(void)
1420 struct l_wait_info lwi = { 0 };
1421 struct task_struct *task;
1424 if (ldlm_pools_thread != NULL)
1427 OBD_ALLOC_PTR(ldlm_pools_thread);
1428 if (ldlm_pools_thread == NULL)
1431 init_completion(&ldlm_pools_comp);
1432 init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq);
1434 task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread,
1437 CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task));
1438 OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread));
1439 ldlm_pools_thread = NULL;
1440 RETURN(PTR_ERR(task));
1442 l_wait_event(ldlm_pools_thread->t_ctl_waitq,
1443 thread_is_running(ldlm_pools_thread), &lwi);
1447 static void ldlm_pools_thread_stop(void)
1451 if (ldlm_pools_thread == NULL) {
1456 thread_set_flags(ldlm_pools_thread, SVC_STOPPING);
1457 wake_up(&ldlm_pools_thread->t_ctl_waitq);
1460 * Make sure that pools thread is finished before freeing @thread.
1461 * This fixes possible race and oops due to accessing freed memory
1464 wait_for_completion(&ldlm_pools_comp);
1465 OBD_FREE_PTR(ldlm_pools_thread);
1466 ldlm_pools_thread = NULL;
1470 int ldlm_pools_init(void)
1473 DEF_SHRINKER_VAR(shsvar, ldlm_pools_srv_shrink,
1474 ldlm_pools_srv_count, ldlm_pools_srv_scan);
1475 DEF_SHRINKER_VAR(shcvar, ldlm_pools_cli_shrink,
1476 ldlm_pools_cli_count, ldlm_pools_cli_scan);
1479 rc = ldlm_pools_thread_start();
1481 ldlm_pools_srv_shrinker =
1482 set_shrinker(DEFAULT_SEEKS, &shsvar);
1483 ldlm_pools_cli_shrinker =
1484 set_shrinker(DEFAULT_SEEKS, &shcvar);
1488 EXPORT_SYMBOL(ldlm_pools_init);
1490 void ldlm_pools_fini(void)
1492 if (ldlm_pools_srv_shrinker != NULL) {
1493 remove_shrinker(ldlm_pools_srv_shrinker);
1494 ldlm_pools_srv_shrinker = NULL;
1496 if (ldlm_pools_cli_shrinker != NULL) {
1497 remove_shrinker(ldlm_pools_cli_shrinker);
1498 ldlm_pools_cli_shrinker = NULL;
1500 ldlm_pools_thread_stop();
1502 EXPORT_SYMBOL(ldlm_pools_fini);
1504 #else /* !HAVE_LRU_RESIZE_SUPPORT */
1505 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
1509 EXPORT_SYMBOL(ldlm_pool_setup);
1511 int ldlm_pool_recalc(struct ldlm_pool *pl)
1515 EXPORT_SYMBOL(ldlm_pool_recalc);
1517 int ldlm_pool_shrink(struct ldlm_pool *pl,
1518 int nr, gfp_t gfp_mask)
1522 EXPORT_SYMBOL(ldlm_pool_shrink);
1524 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
1525 int idx, ldlm_side_t client)
1529 EXPORT_SYMBOL(ldlm_pool_init);
1531 void ldlm_pool_fini(struct ldlm_pool *pl)
1535 EXPORT_SYMBOL(ldlm_pool_fini);
1537 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
1541 EXPORT_SYMBOL(ldlm_pool_add);
1543 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
1547 EXPORT_SYMBOL(ldlm_pool_del);
1549 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
1553 EXPORT_SYMBOL(ldlm_pool_get_slv);
1555 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
1559 EXPORT_SYMBOL(ldlm_pool_set_slv);
1561 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
1565 EXPORT_SYMBOL(ldlm_pool_get_clv);
1567 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1571 EXPORT_SYMBOL(ldlm_pool_set_clv);
1573 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1577 EXPORT_SYMBOL(ldlm_pool_get_limit);
1579 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1583 EXPORT_SYMBOL(ldlm_pool_set_limit);
1585 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1589 EXPORT_SYMBOL(ldlm_pool_get_lvf);
1591 int ldlm_pools_init(void)
1595 EXPORT_SYMBOL(ldlm_pools_init);
1597 void ldlm_pools_fini(void)
1601 EXPORT_SYMBOL(ldlm_pools_fini);
1603 int ldlm_pools_recalc(ldlm_side_t client)
1607 EXPORT_SYMBOL(ldlm_pools_recalc);
1608 #endif /* HAVE_LRU_RESIZE_SUPPORT */