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 ((CFS_NUM_CACHEPAGES >> (20 - CFS_PAGE_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 cfs_proc_dir_entry_t *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 = cfs_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 = cfs_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 = cfs_atomic_read(&pl->pl_granted);
301 int grant_rate = cfs_atomic_read(&pl->pl_grant_rate);
302 int cancel_rate = cfs_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 cfs_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 (cfs_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;
534 int canceled = 0, unused;
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);
554 canceled = ldlm_cancel_lru(ns, nr, LCF_ASYNC,
559 * Return the number of potentially reclaimable locks.
561 return ((unused - canceled) / 100) * sysctl_vfs_cache_pressure;
563 return unused - canceled;
567 struct ldlm_pool_ops ldlm_srv_pool_ops = {
568 .po_recalc = ldlm_srv_pool_recalc,
569 .po_shrink = ldlm_srv_pool_shrink,
570 .po_setup = ldlm_srv_pool_setup
573 struct ldlm_pool_ops ldlm_cli_pool_ops = {
574 .po_recalc = ldlm_cli_pool_recalc,
575 .po_shrink = ldlm_cli_pool_shrink
579 * Pool recalc wrapper. Will call either client or server pool recalc callback
580 * depending what pool \a pl is used.
582 int ldlm_pool_recalc(struct ldlm_pool *pl)
584 time_t recalc_interval_sec;
587 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
588 if (recalc_interval_sec <= 0)
591 spin_lock(&pl->pl_lock);
592 if (recalc_interval_sec > 0) {
594 * Update pool statistics every 1s.
596 ldlm_pool_recalc_stats(pl);
599 * Zero out all rates and speed for the last period.
601 cfs_atomic_set(&pl->pl_grant_rate, 0);
602 cfs_atomic_set(&pl->pl_cancel_rate, 0);
604 spin_unlock(&pl->pl_lock);
607 if (pl->pl_ops->po_recalc != NULL) {
608 count = pl->pl_ops->po_recalc(pl);
609 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
612 recalc_interval_sec = pl->pl_recalc_time - cfs_time_current_sec() +
613 pl->pl_recalc_period;
615 return recalc_interval_sec;
619 * Pool shrink wrapper. Will call either client or server pool recalc callback
620 * depending what pool \a pl is used.
622 int ldlm_pool_shrink(struct ldlm_pool *pl, int nr,
623 unsigned int 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);
659 static int lprocfs_rd_pool_state(char *page, char **start, off_t off,
660 int count, int *eof, void *data)
662 int granted, grant_rate, cancel_rate, grant_step;
663 int nr = 0, grant_speed, grant_plan, lvf;
664 struct ldlm_pool *pl = data;
668 spin_lock(&pl->pl_lock);
669 slv = pl->pl_server_lock_volume;
670 clv = pl->pl_client_lock_volume;
671 limit = ldlm_pool_get_limit(pl);
672 grant_plan = pl->pl_grant_plan;
673 granted = cfs_atomic_read(&pl->pl_granted);
674 grant_rate = cfs_atomic_read(&pl->pl_grant_rate);
675 cancel_rate = cfs_atomic_read(&pl->pl_cancel_rate);
676 grant_speed = grant_rate - cancel_rate;
677 lvf = cfs_atomic_read(&pl->pl_lock_volume_factor);
678 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
679 spin_unlock(&pl->pl_lock);
681 nr += snprintf(page + nr, count - nr, "LDLM pool state (%s):\n",
683 nr += snprintf(page + nr, count - nr, " SLV: "LPU64"\n", slv);
684 nr += snprintf(page + nr, count - nr, " CLV: "LPU64"\n", clv);
685 nr += snprintf(page + nr, count - nr, " LVF: %d\n", lvf);
687 if (ns_is_server(ldlm_pl2ns(pl))) {
688 nr += snprintf(page + nr, count - nr, " GSP: %d%%\n",
690 nr += snprintf(page + nr, count - nr, " GP: %d\n",
693 nr += snprintf(page + nr, count - nr, " GR: %d\n",
695 nr += snprintf(page + nr, count - nr, " CR: %d\n",
697 nr += snprintf(page + nr, count - nr, " GS: %d\n",
699 nr += snprintf(page + nr, count - nr, " G: %d\n",
701 nr += snprintf(page + nr, count - nr, " L: %d\n",
706 static int lprocfs_rd_grant_speed(char *page, char **start, off_t off,
707 int count, int *eof, void *data)
709 struct ldlm_pool *pl = data;
712 spin_lock(&pl->pl_lock);
713 /* serialize with ldlm_pool_recalc */
714 grant_speed = cfs_atomic_read(&pl->pl_grant_rate) -
715 cfs_atomic_read(&pl->pl_cancel_rate);
716 spin_unlock(&pl->pl_lock);
717 return lprocfs_rd_uint(page, start, off, count, eof, &grant_speed);
720 LDLM_POOL_PROC_READER(grant_plan, int);
721 LDLM_POOL_PROC_READER(recalc_period, int);
722 LDLM_POOL_PROC_WRITER(recalc_period, int);
724 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
726 struct ldlm_namespace *ns = ldlm_pl2ns(pl);
727 struct proc_dir_entry *parent_ns_proc;
728 struct lprocfs_vars pool_vars[2];
729 char *var_name = NULL;
733 OBD_ALLOC(var_name, MAX_STRING_SIZE + 1);
737 parent_ns_proc = lprocfs_srch(ldlm_ns_proc_dir,
739 if (parent_ns_proc == NULL) {
740 CERROR("%s: proc entry is not initialized\n",
742 GOTO(out_free_name, rc = -EINVAL);
744 pl->pl_proc_dir = lprocfs_register("pool", parent_ns_proc,
746 if (IS_ERR(pl->pl_proc_dir)) {
747 CERROR("LProcFS failed in ldlm-pool-init\n");
748 rc = PTR_ERR(pl->pl_proc_dir);
749 GOTO(out_free_name, rc);
752 var_name[MAX_STRING_SIZE] = '\0';
753 memset(pool_vars, 0, sizeof(pool_vars));
754 pool_vars[0].name = var_name;
756 snprintf(var_name, MAX_STRING_SIZE, "server_lock_volume");
757 pool_vars[0].data = &pl->pl_server_lock_volume;
758 pool_vars[0].read_fptr = lprocfs_rd_u64;
759 lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
761 snprintf(var_name, MAX_STRING_SIZE, "limit");
762 pool_vars[0].data = &pl->pl_limit;
763 pool_vars[0].read_fptr = lprocfs_rd_atomic;
764 pool_vars[0].write_fptr = lprocfs_wr_atomic;
765 lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
767 snprintf(var_name, MAX_STRING_SIZE, "granted");
768 pool_vars[0].data = &pl->pl_granted;
769 pool_vars[0].read_fptr = lprocfs_rd_atomic;
770 lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
772 snprintf(var_name, MAX_STRING_SIZE, "grant_speed");
773 pool_vars[0].data = pl;
774 pool_vars[0].read_fptr = lprocfs_rd_grant_speed;
775 lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
777 snprintf(var_name, MAX_STRING_SIZE, "cancel_rate");
778 pool_vars[0].data = &pl->pl_cancel_rate;
779 pool_vars[0].read_fptr = lprocfs_rd_atomic;
780 lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
782 snprintf(var_name, MAX_STRING_SIZE, "grant_rate");
783 pool_vars[0].data = &pl->pl_grant_rate;
784 pool_vars[0].read_fptr = lprocfs_rd_atomic;
785 lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
787 snprintf(var_name, MAX_STRING_SIZE, "grant_plan");
788 pool_vars[0].data = pl;
789 pool_vars[0].read_fptr = lprocfs_rd_grant_plan;
790 lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
792 snprintf(var_name, MAX_STRING_SIZE, "recalc_period");
793 pool_vars[0].data = pl;
794 pool_vars[0].read_fptr = lprocfs_rd_recalc_period;
795 pool_vars[0].write_fptr = lprocfs_wr_recalc_period;
796 lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
798 snprintf(var_name, MAX_STRING_SIZE, "lock_volume_factor");
799 pool_vars[0].data = &pl->pl_lock_volume_factor;
800 pool_vars[0].read_fptr = lprocfs_rd_atomic;
801 pool_vars[0].write_fptr = lprocfs_wr_atomic;
802 lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
804 snprintf(var_name, MAX_STRING_SIZE, "state");
805 pool_vars[0].data = pl;
806 pool_vars[0].read_fptr = lprocfs_rd_pool_state;
807 lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
809 pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT -
810 LDLM_POOL_FIRST_STAT, 0);
812 GOTO(out_free_name, rc = -ENOMEM);
814 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
815 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
817 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
818 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
820 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
821 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
823 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
824 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
825 "grant_rate", "locks/s");
826 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
827 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
828 "cancel_rate", "locks/s");
829 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
830 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
831 "grant_plan", "locks/s");
832 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT,
833 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
835 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT,
836 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
837 "shrink_request", "locks");
838 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT,
839 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
840 "shrink_freed", "locks");
841 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
842 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
843 "recalc_freed", "locks");
844 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
845 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
846 "recalc_timing", "sec");
847 rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
851 OBD_FREE(var_name, MAX_STRING_SIZE + 1);
855 static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
857 if (pl->pl_stats != NULL) {
858 lprocfs_free_stats(&pl->pl_stats);
861 if (pl->pl_proc_dir != NULL) {
862 lprocfs_remove(&pl->pl_proc_dir);
863 pl->pl_proc_dir = NULL;
866 #else /* !__KERNEL__*/
867 #define ldlm_pool_proc_init(pl) (0)
868 #define ldlm_pool_proc_fini(pl) while (0) {}
871 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
872 int idx, ldlm_side_t client)
877 spin_lock_init(&pl->pl_lock);
878 cfs_atomic_set(&pl->pl_granted, 0);
879 pl->pl_recalc_time = cfs_time_current_sec();
880 cfs_atomic_set(&pl->pl_lock_volume_factor, 1);
882 cfs_atomic_set(&pl->pl_grant_rate, 0);
883 cfs_atomic_set(&pl->pl_cancel_rate, 0);
884 pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
886 snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
887 ldlm_ns_name(ns), idx);
889 if (client == LDLM_NAMESPACE_SERVER) {
890 pl->pl_ops = &ldlm_srv_pool_ops;
891 ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L);
892 pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD;
893 pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L);
895 ldlm_pool_set_limit(pl, 1);
896 pl->pl_server_lock_volume = 0;
897 pl->pl_ops = &ldlm_cli_pool_ops;
898 pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
900 pl->pl_client_lock_volume = 0;
901 rc = ldlm_pool_proc_init(pl);
905 CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name);
909 EXPORT_SYMBOL(ldlm_pool_init);
911 void ldlm_pool_fini(struct ldlm_pool *pl)
914 ldlm_pool_proc_fini(pl);
917 * Pool should not be used after this point. We can't free it here as
918 * it lives in struct ldlm_namespace, but still interested in catching
919 * any abnormal using cases.
921 POISON(pl, 0x5a, sizeof(*pl));
924 EXPORT_SYMBOL(ldlm_pool_fini);
927 * Add new taken ldlm lock \a lock into pool \a pl accounting.
929 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
932 * FLOCK locks are special in a sense that they are almost never
933 * cancelled, instead special kind of lock is used to drop them.
934 * also there is no LRU for flock locks, so no point in tracking
937 if (lock->l_resource->lr_type == LDLM_FLOCK)
940 cfs_atomic_inc(&pl->pl_granted);
941 cfs_atomic_inc(&pl->pl_grant_rate);
942 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
944 * Do not do pool recalc for client side as all locks which
945 * potentially may be canceled has already been packed into
946 * enqueue/cancel rpc. Also we do not want to run out of stack
947 * with too long call paths.
949 if (ns_is_server(ldlm_pl2ns(pl)))
950 ldlm_pool_recalc(pl);
952 EXPORT_SYMBOL(ldlm_pool_add);
955 * Remove ldlm lock \a lock from pool \a pl accounting.
957 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
960 * Filter out FLOCK locks. Read above comment in ldlm_pool_add().
962 if (lock->l_resource->lr_type == LDLM_FLOCK)
965 LASSERT(cfs_atomic_read(&pl->pl_granted) > 0);
966 cfs_atomic_dec(&pl->pl_granted);
967 cfs_atomic_inc(&pl->pl_cancel_rate);
969 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
971 if (ns_is_server(ldlm_pl2ns(pl)))
972 ldlm_pool_recalc(pl);
974 EXPORT_SYMBOL(ldlm_pool_del);
977 * Returns current \a pl SLV.
979 * \pre ->pl_lock is not locked.
981 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
984 spin_lock(&pl->pl_lock);
985 slv = pl->pl_server_lock_volume;
986 spin_unlock(&pl->pl_lock);
989 EXPORT_SYMBOL(ldlm_pool_get_slv);
992 * Sets passed \a slv to \a pl.
994 * \pre ->pl_lock is not locked.
996 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
998 spin_lock(&pl->pl_lock);
999 pl->pl_server_lock_volume = slv;
1000 spin_unlock(&pl->pl_lock);
1002 EXPORT_SYMBOL(ldlm_pool_set_slv);
1005 * Returns current \a pl CLV.
1007 * \pre ->pl_lock is not locked.
1009 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
1012 spin_lock(&pl->pl_lock);
1013 slv = pl->pl_client_lock_volume;
1014 spin_unlock(&pl->pl_lock);
1017 EXPORT_SYMBOL(ldlm_pool_get_clv);
1020 * Sets passed \a clv to \a pl.
1022 * \pre ->pl_lock is not locked.
1024 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1026 spin_lock(&pl->pl_lock);
1027 pl->pl_client_lock_volume = clv;
1028 spin_unlock(&pl->pl_lock);
1030 EXPORT_SYMBOL(ldlm_pool_set_clv);
1033 * Returns current \a pl limit.
1035 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1037 return cfs_atomic_read(&pl->pl_limit);
1039 EXPORT_SYMBOL(ldlm_pool_get_limit);
1042 * Sets passed \a limit to \a pl.
1044 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1046 cfs_atomic_set(&pl->pl_limit, limit);
1048 EXPORT_SYMBOL(ldlm_pool_set_limit);
1051 * Returns current LVF from \a pl.
1053 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1055 return cfs_atomic_read(&pl->pl_lock_volume_factor);
1057 EXPORT_SYMBOL(ldlm_pool_get_lvf);
1060 static int ldlm_pool_granted(struct ldlm_pool *pl)
1062 return cfs_atomic_read(&pl->pl_granted);
1065 static struct ptlrpc_thread *ldlm_pools_thread;
1066 static struct cfs_shrinker *ldlm_pools_srv_shrinker;
1067 static struct cfs_shrinker *ldlm_pools_cli_shrinker;
1068 static struct completion ldlm_pools_comp;
1071 * Cancel \a nr locks from all namespaces (if possible). Returns number of
1072 * cached locks after shrink is finished. All namespaces are asked to
1073 * cancel approximately equal amount of locks to keep balancing.
1075 static int ldlm_pools_shrink(ldlm_side_t client, int nr,
1076 unsigned int gfp_mask)
1078 int total = 0, cached = 0, nr_ns;
1079 struct ldlm_namespace *ns;
1080 struct ldlm_namespace *ns_old = NULL; /* loop detection */
1083 if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
1084 !(gfp_mask & __GFP_FS))
1087 CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n",
1088 nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1090 cookie = cl_env_reenter();
1093 * Find out how many resources we may release.
1095 for (nr_ns = ldlm_namespace_nr_read(client);
1098 mutex_lock(ldlm_namespace_lock(client));
1099 if (cfs_list_empty(ldlm_namespace_list(client))) {
1100 mutex_unlock(ldlm_namespace_lock(client));
1101 cl_env_reexit(cookie);
1104 ns = ldlm_namespace_first_locked(client);
1107 mutex_unlock(ldlm_namespace_lock(client));
1111 if (ldlm_ns_empty(ns)) {
1112 ldlm_namespace_move_to_inactive_locked(ns, client);
1113 mutex_unlock(ldlm_namespace_lock(client));
1120 ldlm_namespace_get(ns);
1121 ldlm_namespace_move_to_active_locked(ns, client);
1122 mutex_unlock(ldlm_namespace_lock(client));
1123 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
1124 ldlm_namespace_put(ns);
1127 if (nr == 0 || total == 0) {
1128 cl_env_reexit(cookie);
1133 * Shrink at least ldlm_namespace_nr(client) namespaces.
1135 for (nr_ns = ldlm_namespace_nr_read(client) - nr_ns;
1138 int cancel, nr_locks;
1141 * Do not call shrink under ldlm_namespace_lock(client)
1143 mutex_lock(ldlm_namespace_lock(client));
1144 if (cfs_list_empty(ldlm_namespace_list(client))) {
1145 mutex_unlock(ldlm_namespace_lock(client));
1147 * If list is empty, we can't return any @cached > 0,
1148 * that probably would cause needless shrinker
1154 ns = ldlm_namespace_first_locked(client);
1155 ldlm_namespace_get(ns);
1156 ldlm_namespace_move_to_active_locked(ns, client);
1157 mutex_unlock(ldlm_namespace_lock(client));
1159 nr_locks = ldlm_pool_granted(&ns->ns_pool);
1160 cancel = 1 + nr_locks * nr / total;
1161 ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
1162 cached += ldlm_pool_granted(&ns->ns_pool);
1163 ldlm_namespace_put(ns);
1165 cl_env_reexit(cookie);
1166 /* we only decrease the SLV in server pools shrinker, return -1 to
1167 * kernel to avoid needless loop. LU-1128 */
1168 return (client == LDLM_NAMESPACE_SERVER) ? -1 : cached;
1171 static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1173 return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER,
1174 shrink_param(sc, nr_to_scan),
1175 shrink_param(sc, gfp_mask));
1178 static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1180 return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT,
1181 shrink_param(sc, nr_to_scan),
1182 shrink_param(sc, gfp_mask));
1185 int ldlm_pools_recalc(ldlm_side_t client)
1187 __u32 nr_l = 0, nr_p = 0, l;
1188 struct ldlm_namespace *ns;
1189 struct ldlm_namespace *ns_old = NULL;
1191 int time = 50; /* seconds of sleep if no active namespaces */
1194 * No need to setup pool limit for client pools.
1196 if (client == LDLM_NAMESPACE_SERVER) {
1198 * Check all modest namespaces first.
1200 mutex_lock(ldlm_namespace_lock(client));
1201 cfs_list_for_each_entry(ns, ldlm_namespace_list(client),
1204 if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
1207 l = ldlm_pool_granted(&ns->ns_pool);
1212 * Set the modest pools limit equal to their avg granted
1215 l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0);
1216 ldlm_pool_setup(&ns->ns_pool, l);
1222 * Make sure that modest namespaces did not eat more that 2/3
1225 if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) {
1226 CWARN("\"Modest\" pools eat out 2/3 of server locks "
1227 "limit (%d of %lu). This means that you have too "
1228 "many clients for this amount of server RAM. "
1229 "Upgrade server!\n", nr_l, LDLM_POOL_HOST_L);
1234 * The rest is given to greedy namespaces.
1236 cfs_list_for_each_entry(ns, ldlm_namespace_list(client),
1239 if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
1244 * In the case 2/3 locks are eaten out by
1245 * modest pools, we re-setup equal limit
1248 l = LDLM_POOL_HOST_L /
1249 ldlm_namespace_nr_read(client);
1252 * All the rest of greedy pools will have
1253 * all locks in equal parts.
1255 l = (LDLM_POOL_HOST_L - nr_l) /
1256 (ldlm_namespace_nr_read(client) -
1259 ldlm_pool_setup(&ns->ns_pool, l);
1261 mutex_unlock(ldlm_namespace_lock(client));
1265 * Recalc at least ldlm_namespace_nr(client) namespaces.
1267 for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) {
1270 * Lock the list, get first @ns in the list, getref, move it
1271 * to the tail, unlock and call pool recalc. This way we avoid
1272 * calling recalc under @ns lock what is really good as we get
1273 * rid of potential deadlock on client nodes when canceling
1274 * locks synchronously.
1276 mutex_lock(ldlm_namespace_lock(client));
1277 if (cfs_list_empty(ldlm_namespace_list(client))) {
1278 mutex_unlock(ldlm_namespace_lock(client));
1281 ns = ldlm_namespace_first_locked(client);
1283 if (ns_old == ns) { /* Full pass complete */
1284 mutex_unlock(ldlm_namespace_lock(client));
1288 /* We got an empty namespace, need to move it back to inactive
1290 * The race with parallel resource creation is fine:
1291 * - If they do namespace_get before our check, we fail the
1292 * check and they move this item to the end of the list anyway
1293 * - If we do the check and then they do namespace_get, then
1294 * we move the namespace to inactive and they will move
1295 * it back to active (synchronised by the lock, so no clash
1298 if (ldlm_ns_empty(ns)) {
1299 ldlm_namespace_move_to_inactive_locked(ns, client);
1300 mutex_unlock(ldlm_namespace_lock(client));
1307 spin_lock(&ns->ns_lock);
1309 * skip ns which is being freed, and we don't want to increase
1310 * its refcount again, not even temporarily. bz21519 & LU-499.
1312 if (ns->ns_stopping) {
1316 ldlm_namespace_get(ns);
1318 spin_unlock(&ns->ns_lock);
1320 ldlm_namespace_move_to_active_locked(ns, client);
1321 mutex_unlock(ldlm_namespace_lock(client));
1324 * After setup is done - recalc the pool.
1327 int ttime = ldlm_pool_recalc(&ns->ns_pool);
1332 ldlm_namespace_put(ns);
1337 EXPORT_SYMBOL(ldlm_pools_recalc);
1339 static int ldlm_pools_thread_main(void *arg)
1341 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
1342 char *t_name = "ldlm_poold";
1346 cfs_daemonize(t_name);
1347 thread_set_flags(thread, SVC_RUNNING);
1348 cfs_waitq_signal(&thread->t_ctl_waitq);
1350 CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n",
1351 t_name, cfs_curproc_pid());
1354 struct l_wait_info lwi;
1357 * Recal all pools on this tick.
1359 s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
1360 c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
1363 * Wait until the next check time, or until we're
1366 lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)),
1368 l_wait_event(thread->t_ctl_waitq,
1369 thread_is_stopping(thread) ||
1370 thread_is_event(thread),
1373 if (thread_test_and_clear_flags(thread, SVC_STOPPING))
1376 thread_test_and_clear_flags(thread, SVC_EVENT);
1379 thread_set_flags(thread, SVC_STOPPED);
1380 cfs_waitq_signal(&thread->t_ctl_waitq);
1382 CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n",
1383 t_name, cfs_curproc_pid());
1385 complete_and_exit(&ldlm_pools_comp, 0);
1388 static int ldlm_pools_thread_start(void)
1390 struct l_wait_info lwi = { 0 };
1394 if (ldlm_pools_thread != NULL)
1397 OBD_ALLOC_PTR(ldlm_pools_thread);
1398 if (ldlm_pools_thread == NULL)
1401 init_completion(&ldlm_pools_comp);
1402 cfs_waitq_init(&ldlm_pools_thread->t_ctl_waitq);
1405 * CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1406 * just drop the VM and FILES in cfs_daemonize() right away.
1408 rc = cfs_create_thread(ldlm_pools_thread_main, ldlm_pools_thread,
1411 CERROR("Can't start pool thread, error %d\n",
1413 OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread));
1414 ldlm_pools_thread = NULL;
1417 l_wait_event(ldlm_pools_thread->t_ctl_waitq,
1418 thread_is_running(ldlm_pools_thread), &lwi);
1422 static void ldlm_pools_thread_stop(void)
1426 if (ldlm_pools_thread == NULL) {
1431 thread_set_flags(ldlm_pools_thread, SVC_STOPPING);
1432 cfs_waitq_signal(&ldlm_pools_thread->t_ctl_waitq);
1435 * Make sure that pools thread is finished before freeing @thread.
1436 * This fixes possible race and oops due to accessing freed memory
1439 wait_for_completion(&ldlm_pools_comp);
1440 OBD_FREE_PTR(ldlm_pools_thread);
1441 ldlm_pools_thread = NULL;
1445 int ldlm_pools_init(void)
1450 rc = ldlm_pools_thread_start();
1452 ldlm_pools_srv_shrinker =
1453 cfs_set_shrinker(CFS_DEFAULT_SEEKS,
1454 ldlm_pools_srv_shrink);
1455 ldlm_pools_cli_shrinker =
1456 cfs_set_shrinker(CFS_DEFAULT_SEEKS,
1457 ldlm_pools_cli_shrink);
1461 EXPORT_SYMBOL(ldlm_pools_init);
1463 void ldlm_pools_fini(void)
1465 if (ldlm_pools_srv_shrinker != NULL) {
1466 cfs_remove_shrinker(ldlm_pools_srv_shrinker);
1467 ldlm_pools_srv_shrinker = NULL;
1469 if (ldlm_pools_cli_shrinker != NULL) {
1470 cfs_remove_shrinker(ldlm_pools_cli_shrinker);
1471 ldlm_pools_cli_shrinker = NULL;
1473 ldlm_pools_thread_stop();
1475 EXPORT_SYMBOL(ldlm_pools_fini);
1476 #endif /* __KERNEL__ */
1478 #else /* !HAVE_LRU_RESIZE_SUPPORT */
1479 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
1483 EXPORT_SYMBOL(ldlm_pool_setup);
1485 int ldlm_pool_recalc(struct ldlm_pool *pl)
1489 EXPORT_SYMBOL(ldlm_pool_recalc);
1491 int ldlm_pool_shrink(struct ldlm_pool *pl,
1492 int nr, unsigned int gfp_mask)
1496 EXPORT_SYMBOL(ldlm_pool_shrink);
1498 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
1499 int idx, ldlm_side_t client)
1503 EXPORT_SYMBOL(ldlm_pool_init);
1505 void ldlm_pool_fini(struct ldlm_pool *pl)
1509 EXPORT_SYMBOL(ldlm_pool_fini);
1511 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
1515 EXPORT_SYMBOL(ldlm_pool_add);
1517 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
1521 EXPORT_SYMBOL(ldlm_pool_del);
1523 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
1527 EXPORT_SYMBOL(ldlm_pool_get_slv);
1529 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
1533 EXPORT_SYMBOL(ldlm_pool_set_slv);
1535 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
1539 EXPORT_SYMBOL(ldlm_pool_get_clv);
1541 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1545 EXPORT_SYMBOL(ldlm_pool_set_clv);
1547 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1551 EXPORT_SYMBOL(ldlm_pool_get_limit);
1553 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1557 EXPORT_SYMBOL(ldlm_pool_set_limit);
1559 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1563 EXPORT_SYMBOL(ldlm_pool_get_lvf);
1565 int ldlm_pools_init(void)
1569 EXPORT_SYMBOL(ldlm_pools_init);
1571 void ldlm_pools_fini(void)
1575 EXPORT_SYMBOL(ldlm_pools_fini);
1577 int ldlm_pools_recalc(ldlm_side_t client)
1581 EXPORT_SYMBOL(ldlm_pools_recalc);
1582 #endif /* HAVE_LRU_RESIZE_SUPPORT */