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
101 # include <lustre_dlm.h>
103 # include <liblustre.h>
106 #include <cl_object.h>
108 #include <obd_class.h>
109 #include <obd_support.h>
110 #include "ldlm_internal.h"
112 #ifdef HAVE_LRU_RESIZE_SUPPORT
115 * 50 ldlm locks for 1MB of RAM.
117 #define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_CACHE_SHIFT)) * 50)
120 * Maximal possible grant step plan in %.
122 #define LDLM_POOL_MAX_GSP (30)
125 * Minimal possible grant step plan in %.
127 #define LDLM_POOL_MIN_GSP (1)
130 * This controls the speed of reaching LDLM_POOL_MAX_GSP
131 * with increasing thread period.
133 #define LDLM_POOL_GSP_STEP_SHIFT (2)
136 * LDLM_POOL_GSP% of all locks is default GP.
138 #define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100)
141 * Max age for locks on clients.
143 #define LDLM_POOL_MAX_AGE (36000)
146 * The granularity of SLV calculation.
148 #define LDLM_POOL_SLV_SHIFT (10)
151 extern struct proc_dir_entry *ldlm_ns_proc_dir;
154 static inline __u64 dru(__u64 val, __u32 shift, int round_up)
156 return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift;
159 static inline __u64 ldlm_pool_slv_max(__u32 L)
162 * Allow to have all locks for 1 client for 10 hrs.
163 * Formula is the following: limit * 10h / 1 client.
165 __u64 lim = (__u64)L * LDLM_POOL_MAX_AGE / 1;
169 static inline __u64 ldlm_pool_slv_min(__u32 L)
175 LDLM_POOL_FIRST_STAT = 0,
176 LDLM_POOL_GRANTED_STAT = LDLM_POOL_FIRST_STAT,
177 LDLM_POOL_GRANT_STAT,
178 LDLM_POOL_CANCEL_STAT,
179 LDLM_POOL_GRANT_RATE_STAT,
180 LDLM_POOL_CANCEL_RATE_STAT,
181 LDLM_POOL_GRANT_PLAN_STAT,
183 LDLM_POOL_SHRINK_REQTD_STAT,
184 LDLM_POOL_SHRINK_FREED_STAT,
185 LDLM_POOL_RECALC_STAT,
186 LDLM_POOL_TIMING_STAT,
190 static inline struct ldlm_namespace *ldlm_pl2ns(struct ldlm_pool *pl)
192 return container_of(pl, struct ldlm_namespace, ns_pool);
196 * Calculates suggested grant_step in % of available locks for passed
197 * \a period. This is later used in grant_plan calculations.
199 static inline int ldlm_pool_t2gsp(unsigned int t)
202 * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP
203 * and up to 30% for anything higher than LDLM_POOL_GSP_STEP.
205 * How this will affect execution is the following:
207 * - for thread period 1s we will have grant_step 1% which good from
208 * pov of taking some load off from server and push it out to clients.
209 * This is like that because 1% for grant_step means that server will
210 * not allow clients to get lots of locks in short period of time and
211 * keep all old locks in their caches. Clients will always have to
212 * get some locks back if they want to take some new;
214 * - for thread period 10s (which is default) we will have 23% which
215 * means that clients will have enough of room to take some new locks
216 * without getting some back. All locks from this 23% which were not
217 * taken by clients in current period will contribute in SLV growing.
218 * SLV growing means more locks cached on clients until limit or grant
221 return LDLM_POOL_MAX_GSP -
222 ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >>
223 (t >> LDLM_POOL_GSP_STEP_SHIFT));
227 * Recalculates next grant limit on passed \a pl.
229 * \pre ->pl_lock is locked.
231 static void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl)
233 int granted, grant_step, limit;
235 limit = ldlm_pool_get_limit(pl);
236 granted = atomic_read(&pl->pl_granted);
238 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
239 grant_step = ((limit - granted) * grant_step) / 100;
240 pl->pl_grant_plan = granted + grant_step;
241 limit = (limit * 5) >> 2;
242 if (pl->pl_grant_plan > limit)
243 pl->pl_grant_plan = limit;
247 * Recalculates next SLV on passed \a pl.
249 * \pre ->pl_lock is locked.
251 static void ldlm_pool_recalc_slv(struct ldlm_pool *pl)
261 slv = pl->pl_server_lock_volume;
262 grant_plan = pl->pl_grant_plan;
263 limit = ldlm_pool_get_limit(pl);
264 granted = atomic_read(&pl->pl_granted);
265 round_up = granted < limit;
267 grant_usage = max_t(int, limit - (granted - grant_plan), 1);
270 * Find out SLV change factor which is the ratio of grant usage
271 * from limit. SLV changes as fast as the ratio of grant plan
272 * consumption. The more locks from grant plan are not consumed
273 * by clients in last interval (idle time), the faster grows
274 * SLV. And the opposite, the more grant plan is over-consumed
275 * (load time) the faster drops SLV.
277 slv_factor = (grant_usage << LDLM_POOL_SLV_SHIFT);
278 do_div(slv_factor, limit);
279 slv = slv * slv_factor;
280 slv = dru(slv, LDLM_POOL_SLV_SHIFT, round_up);
282 if (slv > ldlm_pool_slv_max(limit)) {
283 slv = ldlm_pool_slv_max(limit);
284 } else if (slv < ldlm_pool_slv_min(limit)) {
285 slv = ldlm_pool_slv_min(limit);
288 pl->pl_server_lock_volume = slv;
292 * Recalculates next stats on passed \a pl.
294 * \pre ->pl_lock is locked.
296 static void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
298 int grant_plan = pl->pl_grant_plan;
299 __u64 slv = pl->pl_server_lock_volume;
300 int granted = atomic_read(&pl->pl_granted);
301 int grant_rate = atomic_read(&pl->pl_grant_rate);
302 int cancel_rate = atomic_read(&pl->pl_cancel_rate);
304 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT,
306 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
308 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
310 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
312 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
317 * Sets current SLV into obd accessible via ldlm_pl2ns(pl)->ns_obd.
319 static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl)
321 struct obd_device *obd;
324 * Set new SLV in obd field for using it later without accessing the
325 * pool. This is required to avoid race between sending reply to client
326 * with new SLV and cleanup server stack in which we can't guarantee
327 * that namespace is still alive. We know only that obd is alive as
328 * long as valid export is alive.
330 obd = ldlm_pl2ns(pl)->ns_obd;
331 LASSERT(obd != NULL);
332 write_lock(&obd->obd_pool_lock);
333 obd->obd_pool_slv = pl->pl_server_lock_volume;
334 write_unlock(&obd->obd_pool_lock);
338 * Recalculates all pool fields on passed \a pl.
340 * \pre ->pl_lock is not locked.
342 static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
344 time_t recalc_interval_sec;
347 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
348 if (recalc_interval_sec < pl->pl_recalc_period)
351 spin_lock(&pl->pl_lock);
352 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
353 if (recalc_interval_sec < pl->pl_recalc_period) {
354 spin_unlock(&pl->pl_lock);
358 * Recalc SLV after last period. This should be done
359 * _before_ recalculating new grant plan.
361 ldlm_pool_recalc_slv(pl);
364 * Make sure that pool informed obd of last SLV changes.
366 ldlm_srv_pool_push_slv(pl);
369 * Update grant_plan for new period.
371 ldlm_pool_recalc_grant_plan(pl);
373 pl->pl_recalc_time = cfs_time_current_sec();
374 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
375 recalc_interval_sec);
376 spin_unlock(&pl->pl_lock);
381 * This function is used on server side as main entry point for memory
382 * pressure handling. It decreases SLV on \a pl according to passed
383 * \a nr and \a gfp_mask.
385 * Our goal here is to decrease SLV such a way that clients hold \a nr
386 * locks smaller in next 10h.
388 static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
389 int nr, unsigned int gfp_mask)
394 * VM is asking how many entries may be potentially freed.
397 return atomic_read(&pl->pl_granted);
400 * Client already canceled locks but server is already in shrinker
401 * and can't cancel anything. Let's catch this race.
403 if (atomic_read(&pl->pl_granted) == 0)
406 spin_lock(&pl->pl_lock);
409 * We want shrinker to possibly cause cancellation of @nr locks from
410 * clients or grant approximately @nr locks smaller next intervals.
412 * This is why we decreased SLV by @nr. This effect will only be as
413 * long as one re-calc interval (1s these days) and this should be
414 * enough to pass this decreased SLV to all clients. On next recalc
415 * interval pool will either increase SLV if locks load is not high
416 * or will keep on same level or even decrease again, thus, shrinker
417 * decreased SLV will affect next recalc intervals and this way will
418 * make locking load lower.
420 if (nr < pl->pl_server_lock_volume) {
421 pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr;
423 limit = ldlm_pool_get_limit(pl);
424 pl->pl_server_lock_volume = ldlm_pool_slv_min(limit);
428 * Make sure that pool informed obd of last SLV changes.
430 ldlm_srv_pool_push_slv(pl);
431 spin_unlock(&pl->pl_lock);
434 * We did not really free any memory here so far, it only will be
435 * freed later may be, so that we return 0 to not confuse VM.
441 * Setup server side pool \a pl with passed \a limit.
443 static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit)
445 struct obd_device *obd;
447 obd = ldlm_pl2ns(pl)->ns_obd;
448 LASSERT(obd != NULL && obd != LP_POISON);
449 LASSERT(obd->obd_type != LP_POISON);
450 write_lock(&obd->obd_pool_lock);
451 obd->obd_pool_limit = limit;
452 write_unlock(&obd->obd_pool_lock);
454 ldlm_pool_set_limit(pl, limit);
459 * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl.
461 static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
463 struct obd_device *obd;
466 * Get new SLV and Limit from obd which is updated with coming
469 obd = ldlm_pl2ns(pl)->ns_obd;
470 LASSERT(obd != NULL);
471 read_lock(&obd->obd_pool_lock);
472 pl->pl_server_lock_volume = obd->obd_pool_slv;
473 ldlm_pool_set_limit(pl, obd->obd_pool_limit);
474 read_unlock(&obd->obd_pool_lock);
478 * Recalculates client size pool \a pl according to current SLV and Limit.
480 static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
482 time_t recalc_interval_sec;
485 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
486 if (recalc_interval_sec < pl->pl_recalc_period)
489 spin_lock(&pl->pl_lock);
491 * Check if we need to recalc lists now.
493 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
494 if (recalc_interval_sec < pl->pl_recalc_period) {
495 spin_unlock(&pl->pl_lock);
500 * Make sure that pool knows last SLV and Limit from obd.
502 ldlm_cli_pool_pop_slv(pl);
504 pl->pl_recalc_time = cfs_time_current_sec();
505 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
506 recalc_interval_sec);
507 spin_unlock(&pl->pl_lock);
510 * Do not cancel locks in case lru resize is disabled for this ns.
512 if (!ns_connect_lru_resize(ldlm_pl2ns(pl)))
516 * In the time of canceling locks on client we do not need to maintain
517 * sharp timing, we only want to cancel locks asap according to new SLV.
518 * It may be called when SLV has changed much, this is why we do not
519 * take into account pl->pl_recalc_time here.
521 RETURN(ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC,
526 * This function is main entry point for memory pressure handling on client
527 * side. Main goal of this function is to cancel some number of locks on
528 * passed \a pl according to \a nr and \a gfp_mask.
530 static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
531 int nr, unsigned int gfp_mask)
533 struct ldlm_namespace *ns;
539 * Do not cancel locks in case lru resize is disabled for this ns.
541 if (!ns_connect_lru_resize(ns))
545 * Make sure that pool knows last SLV and Limit from obd.
547 ldlm_cli_pool_pop_slv(pl);
549 spin_lock(&ns->ns_lock);
550 unused = ns->ns_nr_unused;
551 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);
559 return unused - (nr ? ldlm_cancel_lru(ns, nr, LCF_ASYNC,
560 LDLM_CANCEL_SHRINK) : 0);
564 struct ldlm_pool_ops ldlm_srv_pool_ops = {
565 .po_recalc = ldlm_srv_pool_recalc,
566 .po_shrink = ldlm_srv_pool_shrink,
567 .po_setup = ldlm_srv_pool_setup
570 struct ldlm_pool_ops ldlm_cli_pool_ops = {
571 .po_recalc = ldlm_cli_pool_recalc,
572 .po_shrink = ldlm_cli_pool_shrink
576 * Pool recalc wrapper. Will call either client or server pool recalc callback
577 * depending what pool \a pl is used.
579 int ldlm_pool_recalc(struct ldlm_pool *pl)
581 time_t recalc_interval_sec;
584 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
585 if (recalc_interval_sec <= 0)
588 spin_lock(&pl->pl_lock);
589 if (recalc_interval_sec > 0) {
591 * Update pool statistics every 1s.
593 ldlm_pool_recalc_stats(pl);
596 * Zero out all rates and speed for the last period.
598 atomic_set(&pl->pl_grant_rate, 0);
599 atomic_set(&pl->pl_cancel_rate, 0);
601 spin_unlock(&pl->pl_lock);
604 if (pl->pl_ops->po_recalc != NULL) {
605 count = pl->pl_ops->po_recalc(pl);
606 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
609 recalc_interval_sec = pl->pl_recalc_time - cfs_time_current_sec() +
610 pl->pl_recalc_period;
612 return recalc_interval_sec;
616 * Pool shrink wrapper. Will call either client or server pool recalc callback
617 * depending what pool \a pl is used.
619 int ldlm_pool_shrink(struct ldlm_pool *pl, int nr,
620 unsigned int gfp_mask)
624 if (pl->pl_ops->po_shrink != NULL) {
625 cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask);
627 lprocfs_counter_add(pl->pl_stats,
628 LDLM_POOL_SHRINK_REQTD_STAT,
630 lprocfs_counter_add(pl->pl_stats,
631 LDLM_POOL_SHRINK_FREED_STAT,
633 CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, "
634 "shrunk %d\n", pl->pl_name, nr, cancel);
639 EXPORT_SYMBOL(ldlm_pool_shrink);
642 * Pool setup wrapper. Will call either client or server pool recalc callback
643 * depending what pool \a pl is used.
645 * Sets passed \a limit into pool \a pl.
647 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
649 if (pl->pl_ops->po_setup != NULL)
650 return(pl->pl_ops->po_setup(pl, limit));
653 EXPORT_SYMBOL(ldlm_pool_setup);
656 static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
658 int granted, grant_rate, cancel_rate, grant_step;
659 int grant_speed, grant_plan, lvf;
660 struct ldlm_pool *pl = m->private;
664 spin_lock(&pl->pl_lock);
665 slv = pl->pl_server_lock_volume;
666 clv = pl->pl_client_lock_volume;
667 limit = ldlm_pool_get_limit(pl);
668 grant_plan = pl->pl_grant_plan;
669 granted = atomic_read(&pl->pl_granted);
670 grant_rate = atomic_read(&pl->pl_grant_rate);
671 cancel_rate = atomic_read(&pl->pl_cancel_rate);
672 grant_speed = grant_rate - cancel_rate;
673 lvf = atomic_read(&pl->pl_lock_volume_factor);
674 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
675 spin_unlock(&pl->pl_lock);
677 seq_printf(m, "LDLM pool state (%s):\n"
681 pl->pl_name, slv, clv, lvf);
683 if (ns_is_server(ldlm_pl2ns(pl))) {
684 seq_printf(m, " GSP: %d%%\n"
686 grant_step, grant_plan);
688 seq_printf(m, " GR: %d\n" " CR: %d\n" " GS: %d\n"
689 " G: %d\n" " L: %d\n",
690 grant_rate, cancel_rate, grant_speed,
694 LPROC_SEQ_FOPS_RO(lprocfs_pool_state);
696 static int lprocfs_grant_speed_seq_show(struct seq_file *m, void *unused)
698 struct ldlm_pool *pl = m->private;
701 spin_lock(&pl->pl_lock);
702 /* serialize with ldlm_pool_recalc */
703 grant_speed = atomic_read(&pl->pl_grant_rate) -
704 atomic_read(&pl->pl_cancel_rate);
705 spin_unlock(&pl->pl_lock);
706 return lprocfs_uint_seq_show(m, &grant_speed);
709 LDLM_POOL_PROC_READER_SEQ_SHOW(grant_plan, int);
710 LPROC_SEQ_FOPS_RO(lprocfs_grant_plan);
712 LDLM_POOL_PROC_READER_SEQ_SHOW(recalc_period, int);
713 LDLM_POOL_PROC_WRITER(recalc_period, int);
714 static ssize_t lprocfs_recalc_period_seq_write(struct file *file, const char *buf,
715 size_t len, loff_t *off)
717 struct seq_file *seq = file->private_data;
719 return lprocfs_wr_recalc_period(file, buf, len, seq->private);
721 LPROC_SEQ_FOPS(lprocfs_recalc_period);
723 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, u64);
724 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, atomic);
725 LPROC_SEQ_FOPS_RW_TYPE(ldlm_pool_rw, atomic);
727 LPROC_SEQ_FOPS_RO(lprocfs_grant_speed);
729 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
731 struct ldlm_namespace *ns = ldlm_pl2ns(pl);
732 struct proc_dir_entry *parent_ns_proc;
733 struct lprocfs_seq_vars pool_vars[2];
734 char *var_name = NULL;
738 OBD_ALLOC(var_name, MAX_STRING_SIZE + 1);
742 parent_ns_proc = ns->ns_proc_dir_entry;
743 if (parent_ns_proc == NULL) {
744 CERROR("%s: proc entry is not initialized\n",
746 GOTO(out_free_name, rc = -EINVAL);
748 pl->pl_proc_dir = lprocfs_seq_register("pool", parent_ns_proc,
750 if (IS_ERR(pl->pl_proc_dir)) {
751 rc = PTR_ERR(pl->pl_proc_dir);
752 pl->pl_proc_dir = NULL;
753 CERROR("%s: cannot create 'pool' proc entry: rc = %d\n",
754 ldlm_ns_name(ns), rc);
755 GOTO(out_free_name, rc);
758 var_name[MAX_STRING_SIZE] = '\0';
759 memset(pool_vars, 0, sizeof(pool_vars));
760 pool_vars[0].name = var_name;
762 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "server_lock_volume",
763 &pl->pl_server_lock_volume, &ldlm_pool_u64_fops);
764 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "limit", &pl->pl_limit,
765 &ldlm_pool_rw_atomic_fops);
766 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "granted",
767 &pl->pl_granted, &ldlm_pool_atomic_fops);
768 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_speed", pl,
769 &lprocfs_grant_speed_fops);
770 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "cancel_rate",
771 &pl->pl_cancel_rate, &ldlm_pool_atomic_fops);
772 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_rate",
773 &pl->pl_grant_rate, &ldlm_pool_atomic_fops);
774 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "grant_plan", pl,
775 &lprocfs_grant_plan_fops);
776 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "recalc_period",
777 pl, &lprocfs_recalc_period_fops);
778 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "lock_volume_factor",
779 &pl->pl_lock_volume_factor, &ldlm_pool_rw_atomic_fops);
780 ldlm_add_var(&pool_vars[0], pl->pl_proc_dir, "state", pl,
781 &lprocfs_pool_state_fops);
783 pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT -
784 LDLM_POOL_FIRST_STAT, 0);
786 GOTO(out_free_name, rc = -ENOMEM);
788 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
789 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
791 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
792 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
794 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
795 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
797 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
798 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
799 "grant_rate", "locks/s");
800 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
801 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
802 "cancel_rate", "locks/s");
803 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
804 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
805 "grant_plan", "locks/s");
806 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT,
807 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
809 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT,
810 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
811 "shrink_request", "locks");
812 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT,
813 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
814 "shrink_freed", "locks");
815 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
816 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
817 "recalc_freed", "locks");
818 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
819 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
820 "recalc_timing", "sec");
821 rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
825 OBD_FREE(var_name, MAX_STRING_SIZE + 1);
829 static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
831 if (pl->pl_stats != NULL) {
832 lprocfs_free_stats(&pl->pl_stats);
835 if (pl->pl_proc_dir != NULL) {
836 lprocfs_remove(&pl->pl_proc_dir);
837 pl->pl_proc_dir = NULL;
840 #else /* !__KERNEL__*/
841 #define ldlm_pool_proc_init(pl) (0)
842 #define ldlm_pool_proc_fini(pl) while (0) {}
845 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
846 int idx, ldlm_side_t client)
851 spin_lock_init(&pl->pl_lock);
852 atomic_set(&pl->pl_granted, 0);
853 pl->pl_recalc_time = cfs_time_current_sec();
854 atomic_set(&pl->pl_lock_volume_factor, 1);
856 atomic_set(&pl->pl_grant_rate, 0);
857 atomic_set(&pl->pl_cancel_rate, 0);
858 pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
860 snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
861 ldlm_ns_name(ns), idx);
863 if (client == LDLM_NAMESPACE_SERVER) {
864 pl->pl_ops = &ldlm_srv_pool_ops;
865 ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L);
866 pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD;
867 pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L);
869 ldlm_pool_set_limit(pl, 1);
870 pl->pl_server_lock_volume = 0;
871 pl->pl_ops = &ldlm_cli_pool_ops;
872 pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
874 pl->pl_client_lock_volume = 0;
875 rc = ldlm_pool_proc_init(pl);
879 CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name);
883 EXPORT_SYMBOL(ldlm_pool_init);
885 void ldlm_pool_fini(struct ldlm_pool *pl)
888 ldlm_pool_proc_fini(pl);
891 * Pool should not be used after this point. We can't free it here as
892 * it lives in struct ldlm_namespace, but still interested in catching
893 * any abnormal using cases.
895 POISON(pl, 0x5a, sizeof(*pl));
898 EXPORT_SYMBOL(ldlm_pool_fini);
901 * Add new taken ldlm lock \a lock into pool \a pl accounting.
903 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
906 * FLOCK locks are special in a sense that they are almost never
907 * cancelled, instead special kind of lock is used to drop them.
908 * also there is no LRU for flock locks, so no point in tracking
911 if (lock->l_resource->lr_type == LDLM_FLOCK)
914 atomic_inc(&pl->pl_granted);
915 atomic_inc(&pl->pl_grant_rate);
916 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
918 * Do not do pool recalc for client side as all locks which
919 * potentially may be canceled has already been packed into
920 * enqueue/cancel rpc. Also we do not want to run out of stack
921 * with too long call paths.
923 if (ns_is_server(ldlm_pl2ns(pl)))
924 ldlm_pool_recalc(pl);
926 EXPORT_SYMBOL(ldlm_pool_add);
929 * Remove ldlm lock \a lock from pool \a pl accounting.
931 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
934 * Filter out FLOCK locks. Read above comment in ldlm_pool_add().
936 if (lock->l_resource->lr_type == LDLM_FLOCK)
939 LASSERT(atomic_read(&pl->pl_granted) > 0);
940 atomic_dec(&pl->pl_granted);
941 atomic_inc(&pl->pl_cancel_rate);
943 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
945 if (ns_is_server(ldlm_pl2ns(pl)))
946 ldlm_pool_recalc(pl);
948 EXPORT_SYMBOL(ldlm_pool_del);
951 * Returns current \a pl SLV.
953 * \pre ->pl_lock is not locked.
955 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
958 spin_lock(&pl->pl_lock);
959 slv = pl->pl_server_lock_volume;
960 spin_unlock(&pl->pl_lock);
963 EXPORT_SYMBOL(ldlm_pool_get_slv);
966 * Sets passed \a slv to \a pl.
968 * \pre ->pl_lock is not locked.
970 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
972 spin_lock(&pl->pl_lock);
973 pl->pl_server_lock_volume = slv;
974 spin_unlock(&pl->pl_lock);
976 EXPORT_SYMBOL(ldlm_pool_set_slv);
979 * Returns current \a pl CLV.
981 * \pre ->pl_lock is not locked.
983 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
986 spin_lock(&pl->pl_lock);
987 slv = pl->pl_client_lock_volume;
988 spin_unlock(&pl->pl_lock);
991 EXPORT_SYMBOL(ldlm_pool_get_clv);
994 * Sets passed \a clv to \a pl.
996 * \pre ->pl_lock is not locked.
998 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1000 spin_lock(&pl->pl_lock);
1001 pl->pl_client_lock_volume = clv;
1002 spin_unlock(&pl->pl_lock);
1004 EXPORT_SYMBOL(ldlm_pool_set_clv);
1007 * Returns current \a pl limit.
1009 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1011 return atomic_read(&pl->pl_limit);
1013 EXPORT_SYMBOL(ldlm_pool_get_limit);
1016 * Sets passed \a limit to \a pl.
1018 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1020 atomic_set(&pl->pl_limit, limit);
1022 EXPORT_SYMBOL(ldlm_pool_set_limit);
1025 * Returns current LVF from \a pl.
1027 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1029 return atomic_read(&pl->pl_lock_volume_factor);
1031 EXPORT_SYMBOL(ldlm_pool_get_lvf);
1034 static unsigned int ldlm_pool_granted(struct ldlm_pool *pl)
1036 return atomic_read(&pl->pl_granted);
1039 static struct ptlrpc_thread *ldlm_pools_thread;
1040 static struct shrinker *ldlm_pools_srv_shrinker;
1041 static struct shrinker *ldlm_pools_cli_shrinker;
1042 static struct completion ldlm_pools_comp;
1045 * count locks from all namespaces (if possible). Returns number of
1048 static unsigned long ldlm_pools_count(ldlm_side_t client, unsigned int gfp_mask)
1050 int total = 0, nr_ns;
1051 struct ldlm_namespace *ns;
1052 struct ldlm_namespace *ns_old = NULL; /* loop detection */
1055 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1058 CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n",
1059 client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1061 cookie = cl_env_reenter();
1064 * Find out how many resources we may release.
1066 for (nr_ns = ldlm_namespace_nr_read(client);
1067 nr_ns > 0; nr_ns--) {
1068 mutex_lock(ldlm_namespace_lock(client));
1069 if (list_empty(ldlm_namespace_list(client))) {
1070 mutex_unlock(ldlm_namespace_lock(client));
1071 cl_env_reexit(cookie);
1074 ns = ldlm_namespace_first_locked(client);
1077 mutex_unlock(ldlm_namespace_lock(client));
1081 if (ldlm_ns_empty(ns)) {
1082 ldlm_namespace_move_to_inactive_locked(ns, client);
1083 mutex_unlock(ldlm_namespace_lock(client));
1090 ldlm_namespace_get(ns);
1091 ldlm_namespace_move_to_active_locked(ns, client);
1092 mutex_unlock(ldlm_namespace_lock(client));
1093 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
1094 ldlm_namespace_put(ns);
1097 cl_env_reexit(cookie);
1101 static unsigned long ldlm_pools_scan(ldlm_side_t client, int nr,
1102 unsigned int gfp_mask)
1104 unsigned long freed = 0;
1106 struct ldlm_namespace *ns;
1109 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1112 cookie = cl_env_reenter();
1115 * Shrink at least ldlm_namespace_nr_read(client) namespaces.
1117 for (tmp = nr_ns = ldlm_namespace_nr_read(client);
1119 int cancel, nr_locks;
1122 * Do not call shrink under ldlm_namespace_lock(client)
1124 mutex_lock(ldlm_namespace_lock(client));
1125 if (list_empty(ldlm_namespace_list(client))) {
1126 mutex_unlock(ldlm_namespace_lock(client));
1129 ns = ldlm_namespace_first_locked(client);
1130 ldlm_namespace_get(ns);
1131 ldlm_namespace_move_to_active_locked(ns, client);
1132 mutex_unlock(ldlm_namespace_lock(client));
1134 nr_locks = ldlm_pool_granted(&ns->ns_pool);
1136 * We use to shrink propotionally but with new shrinker API,
1137 * we lost the total number of freeable locks.
1139 cancel = 1 + min_t(int, nr_locks, nr / nr_ns);
1140 freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
1141 ldlm_namespace_put(ns);
1143 cl_env_reexit(cookie);
1145 * we only decrease the SLV in server pools shrinker, return
1146 * SHRINK_STOP to kernel to avoid needless loop. LU-1128
1148 return (client == LDLM_NAMESPACE_SERVER) ? SHRINK_STOP : freed;
1151 #ifdef HAVE_SHRINKER_COUNT
1152 static unsigned long ldlm_pools_srv_count(struct shrinker *s,
1153 struct shrink_control *sc)
1155 return ldlm_pools_count(LDLM_NAMESPACE_SERVER, sc->gfp_mask);
1158 static unsigned long ldlm_pools_srv_scan(struct shrinker *s,
1159 struct shrink_control *sc)
1161 return ldlm_pools_scan(LDLM_NAMESPACE_SERVER, sc->nr_to_scan,
1165 static unsigned long ldlm_pools_cli_count(struct shrinker *s, struct shrink_control *sc)
1167 return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask);
1170 static unsigned long ldlm_pools_cli_scan(struct shrinker *s,
1171 struct shrink_control *sc)
1173 return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan,
1179 * Cancel \a nr locks from all namespaces (if possible). Returns number of
1180 * cached locks after shrink is finished. All namespaces are asked to
1181 * cancel approximately equal amount of locks to keep balancing.
1183 static int ldlm_pools_shrink(ldlm_side_t client, int nr,
1184 unsigned int gfp_mask)
1186 unsigned int total = 0;
1188 if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
1189 !(gfp_mask & __GFP_FS))
1192 CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n",
1193 nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1195 total = ldlm_pools_count(client, gfp_mask);
1197 if (nr == 0 || total == 0)
1200 return ldlm_pools_scan(client, nr, gfp_mask);
1203 static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1205 return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER,
1206 shrink_param(sc, nr_to_scan),
1207 shrink_param(sc, gfp_mask));
1210 static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1212 return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT,
1213 shrink_param(sc, nr_to_scan),
1214 shrink_param(sc, gfp_mask));
1217 #endif /* HAVE_SHRINKER_COUNT */
1219 int ldlm_pools_recalc(ldlm_side_t client)
1221 __u32 nr_l = 0, nr_p = 0, l;
1222 struct ldlm_namespace *ns;
1223 struct ldlm_namespace *ns_old = NULL;
1225 int time = 50; /* seconds of sleep if no active namespaces */
1228 * No need to setup pool limit for client pools.
1230 if (client == LDLM_NAMESPACE_SERVER) {
1232 * Check all modest namespaces first.
1234 mutex_lock(ldlm_namespace_lock(client));
1235 cfs_list_for_each_entry(ns, ldlm_namespace_list(client),
1238 if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
1241 l = ldlm_pool_granted(&ns->ns_pool);
1246 * Set the modest pools limit equal to their avg granted
1249 l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0);
1250 ldlm_pool_setup(&ns->ns_pool, l);
1256 * Make sure that modest namespaces did not eat more that 2/3
1259 if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) {
1260 CWARN("\"Modest\" pools eat out 2/3 of server locks "
1261 "limit (%d of %lu). This means that you have too "
1262 "many clients for this amount of server RAM. "
1263 "Upgrade server!\n", nr_l, LDLM_POOL_HOST_L);
1268 * The rest is given to greedy namespaces.
1270 cfs_list_for_each_entry(ns, ldlm_namespace_list(client),
1273 if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
1278 * In the case 2/3 locks are eaten out by
1279 * modest pools, we re-setup equal limit
1282 l = LDLM_POOL_HOST_L /
1283 ldlm_namespace_nr_read(client);
1286 * All the rest of greedy pools will have
1287 * all locks in equal parts.
1289 l = (LDLM_POOL_HOST_L - nr_l) /
1290 (ldlm_namespace_nr_read(client) -
1293 ldlm_pool_setup(&ns->ns_pool, l);
1295 mutex_unlock(ldlm_namespace_lock(client));
1299 * Recalc at least ldlm_namespace_nr(client) namespaces.
1301 for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) {
1304 * Lock the list, get first @ns in the list, getref, move it
1305 * to the tail, unlock and call pool recalc. This way we avoid
1306 * calling recalc under @ns lock what is really good as we get
1307 * rid of potential deadlock on client nodes when canceling
1308 * locks synchronously.
1310 mutex_lock(ldlm_namespace_lock(client));
1311 if (cfs_list_empty(ldlm_namespace_list(client))) {
1312 mutex_unlock(ldlm_namespace_lock(client));
1315 ns = ldlm_namespace_first_locked(client);
1317 if (ns_old == ns) { /* Full pass complete */
1318 mutex_unlock(ldlm_namespace_lock(client));
1322 /* We got an empty namespace, need to move it back to inactive
1324 * The race with parallel resource creation is fine:
1325 * - If they do namespace_get before our check, we fail the
1326 * check and they move this item to the end of the list anyway
1327 * - If we do the check and then they do namespace_get, then
1328 * we move the namespace to inactive and they will move
1329 * it back to active (synchronised by the lock, so no clash
1332 if (ldlm_ns_empty(ns)) {
1333 ldlm_namespace_move_to_inactive_locked(ns, client);
1334 mutex_unlock(ldlm_namespace_lock(client));
1341 spin_lock(&ns->ns_lock);
1343 * skip ns which is being freed, and we don't want to increase
1344 * its refcount again, not even temporarily. bz21519 & LU-499.
1346 if (ns->ns_stopping) {
1350 ldlm_namespace_get(ns);
1352 spin_unlock(&ns->ns_lock);
1354 ldlm_namespace_move_to_active_locked(ns, client);
1355 mutex_unlock(ldlm_namespace_lock(client));
1358 * After setup is done - recalc the pool.
1361 int ttime = ldlm_pool_recalc(&ns->ns_pool);
1366 ldlm_namespace_put(ns);
1371 EXPORT_SYMBOL(ldlm_pools_recalc);
1373 static int ldlm_pools_thread_main(void *arg)
1375 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
1379 thread_set_flags(thread, SVC_RUNNING);
1380 wake_up(&thread->t_ctl_waitq);
1382 CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n",
1383 "ldlm_poold", current_pid());
1386 struct l_wait_info lwi;
1389 * Recal all pools on this tick.
1391 s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
1392 c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
1395 * Wait until the next check time, or until we're
1398 lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)),
1400 l_wait_event(thread->t_ctl_waitq,
1401 thread_is_stopping(thread) ||
1402 thread_is_event(thread),
1405 if (thread_test_and_clear_flags(thread, SVC_STOPPING))
1408 thread_test_and_clear_flags(thread, SVC_EVENT);
1411 thread_set_flags(thread, SVC_STOPPED);
1412 wake_up(&thread->t_ctl_waitq);
1414 CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n",
1415 "ldlm_poold", current_pid());
1417 complete_and_exit(&ldlm_pools_comp, 0);
1420 static int ldlm_pools_thread_start(void)
1422 struct l_wait_info lwi = { 0 };
1423 struct task_struct *task;
1426 if (ldlm_pools_thread != NULL)
1429 OBD_ALLOC_PTR(ldlm_pools_thread);
1430 if (ldlm_pools_thread == NULL)
1433 init_completion(&ldlm_pools_comp);
1434 init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq);
1436 task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread,
1439 CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task));
1440 OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread));
1441 ldlm_pools_thread = NULL;
1442 RETURN(PTR_ERR(task));
1444 l_wait_event(ldlm_pools_thread->t_ctl_waitq,
1445 thread_is_running(ldlm_pools_thread), &lwi);
1449 static void ldlm_pools_thread_stop(void)
1453 if (ldlm_pools_thread == NULL) {
1458 thread_set_flags(ldlm_pools_thread, SVC_STOPPING);
1459 wake_up(&ldlm_pools_thread->t_ctl_waitq);
1462 * Make sure that pools thread is finished before freeing @thread.
1463 * This fixes possible race and oops due to accessing freed memory
1466 wait_for_completion(&ldlm_pools_comp);
1467 OBD_FREE_PTR(ldlm_pools_thread);
1468 ldlm_pools_thread = NULL;
1472 int ldlm_pools_init(void)
1475 DEF_SHRINKER_VAR(shsvar, ldlm_pools_srv_shrink,
1476 ldlm_pools_srv_count, ldlm_pools_srv_scan);
1477 DEF_SHRINKER_VAR(shcvar, ldlm_pools_cli_shrink,
1478 ldlm_pools_cli_count, ldlm_pools_cli_scan);
1481 rc = ldlm_pools_thread_start();
1483 ldlm_pools_srv_shrinker =
1484 set_shrinker(DEFAULT_SEEKS, &shsvar);
1485 ldlm_pools_cli_shrinker =
1486 set_shrinker(DEFAULT_SEEKS, &shcvar);
1490 EXPORT_SYMBOL(ldlm_pools_init);
1492 void ldlm_pools_fini(void)
1494 if (ldlm_pools_srv_shrinker != NULL) {
1495 remove_shrinker(ldlm_pools_srv_shrinker);
1496 ldlm_pools_srv_shrinker = NULL;
1498 if (ldlm_pools_cli_shrinker != NULL) {
1499 remove_shrinker(ldlm_pools_cli_shrinker);
1500 ldlm_pools_cli_shrinker = NULL;
1502 ldlm_pools_thread_stop();
1504 EXPORT_SYMBOL(ldlm_pools_fini);
1505 #endif /* __KERNEL__ */
1507 #else /* !HAVE_LRU_RESIZE_SUPPORT */
1508 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
1512 EXPORT_SYMBOL(ldlm_pool_setup);
1514 int ldlm_pool_recalc(struct ldlm_pool *pl)
1518 EXPORT_SYMBOL(ldlm_pool_recalc);
1520 int ldlm_pool_shrink(struct ldlm_pool *pl,
1521 int nr, unsigned int gfp_mask)
1525 EXPORT_SYMBOL(ldlm_pool_shrink);
1527 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
1528 int idx, ldlm_side_t client)
1532 EXPORT_SYMBOL(ldlm_pool_init);
1534 void ldlm_pool_fini(struct ldlm_pool *pl)
1538 EXPORT_SYMBOL(ldlm_pool_fini);
1540 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
1544 EXPORT_SYMBOL(ldlm_pool_add);
1546 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
1550 EXPORT_SYMBOL(ldlm_pool_del);
1552 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
1556 EXPORT_SYMBOL(ldlm_pool_get_slv);
1558 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
1562 EXPORT_SYMBOL(ldlm_pool_set_slv);
1564 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
1568 EXPORT_SYMBOL(ldlm_pool_get_clv);
1570 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1574 EXPORT_SYMBOL(ldlm_pool_set_clv);
1576 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1580 EXPORT_SYMBOL(ldlm_pool_get_limit);
1582 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1586 EXPORT_SYMBOL(ldlm_pool_set_limit);
1588 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1592 EXPORT_SYMBOL(ldlm_pool_get_lvf);
1594 int ldlm_pools_init(void)
1598 EXPORT_SYMBOL(ldlm_pools_init);
1600 void ldlm_pools_fini(void)
1604 EXPORT_SYMBOL(ldlm_pools_fini);
1606 int ldlm_pools_recalc(ldlm_side_t client)
1610 EXPORT_SYMBOL(ldlm_pools_recalc);
1611 #endif /* HAVE_LRU_RESIZE_SUPPORT */