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>
102 #include <cl_object.h>
104 #include <obd_class.h>
105 #include <obd_support.h>
106 #include "ldlm_internal.h"
110 * 50 ldlm locks for 1MB of RAM.
112 #define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_CACHE_SHIFT)) * 50)
115 * Maximal possible grant step plan in %.
117 #define LDLM_POOL_MAX_GSP (30)
120 * Minimal possible grant step plan in %.
122 #define LDLM_POOL_MIN_GSP (1)
125 * This controls the speed of reaching LDLM_POOL_MAX_GSP
126 * with increasing thread period.
128 #define LDLM_POOL_GSP_STEP_SHIFT (2)
131 * LDLM_POOL_GSP% of all locks is default GP.
133 #define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100)
136 * Max age for locks on clients.
138 #define LDLM_POOL_MAX_AGE (36000)
141 * The granularity of SLV calculation.
143 #define LDLM_POOL_SLV_SHIFT (10)
145 extern proc_dir_entry_t *ldlm_ns_proc_dir;
147 static inline __u64 dru(__u64 val, __u32 shift, int round_up)
149 return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift;
152 static inline __u64 ldlm_pool_slv_max(__u32 L)
155 * Allow to have all locks for 1 client for 10 hrs.
156 * Formula is the following: limit * 10h / 1 client.
158 __u64 lim = (__u64)L * LDLM_POOL_MAX_AGE / 1;
162 static inline __u64 ldlm_pool_slv_min(__u32 L)
168 LDLM_POOL_FIRST_STAT = 0,
169 LDLM_POOL_GRANTED_STAT = LDLM_POOL_FIRST_STAT,
170 LDLM_POOL_GRANT_STAT,
171 LDLM_POOL_CANCEL_STAT,
172 LDLM_POOL_GRANT_RATE_STAT,
173 LDLM_POOL_CANCEL_RATE_STAT,
174 LDLM_POOL_GRANT_PLAN_STAT,
176 LDLM_POOL_SHRINK_REQTD_STAT,
177 LDLM_POOL_SHRINK_FREED_STAT,
178 LDLM_POOL_RECALC_STAT,
179 LDLM_POOL_TIMING_STAT,
183 static inline struct ldlm_namespace *ldlm_pl2ns(struct ldlm_pool *pl)
185 return container_of(pl, struct ldlm_namespace, ns_pool);
189 * Calculates suggested grant_step in % of available locks for passed
190 * \a period. This is later used in grant_plan calculations.
192 static inline int ldlm_pool_t2gsp(unsigned int t)
195 * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP
196 * and up to 30% for anything higher than LDLM_POOL_GSP_STEP.
198 * How this will affect execution is the following:
200 * - for thread period 1s we will have grant_step 1% which good from
201 * pov of taking some load off from server and push it out to clients.
202 * This is like that because 1% for grant_step means that server will
203 * not allow clients to get lots of locks in short period of time and
204 * keep all old locks in their caches. Clients will always have to
205 * get some locks back if they want to take some new;
207 * - for thread period 10s (which is default) we will have 23% which
208 * means that clients will have enough of room to take some new locks
209 * without getting some back. All locks from this 23% which were not
210 * taken by clients in current period will contribute in SLV growing.
211 * SLV growing means more locks cached on clients until limit or grant
214 return LDLM_POOL_MAX_GSP -
215 ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >>
216 (t >> LDLM_POOL_GSP_STEP_SHIFT));
220 * Recalculates next grant limit on passed \a pl.
222 * \pre ->pl_lock is locked.
224 static void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl)
226 int granted, grant_step, limit;
228 limit = ldlm_pool_get_limit(pl);
229 granted = atomic_read(&pl->pl_granted);
231 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
232 grant_step = ((limit - granted) * grant_step) / 100;
233 pl->pl_grant_plan = granted + grant_step;
234 limit = (limit * 5) >> 2;
235 if (pl->pl_grant_plan > limit)
236 pl->pl_grant_plan = limit;
240 * Recalculates next SLV on passed \a pl.
242 * \pre ->pl_lock is locked.
244 static void ldlm_pool_recalc_slv(struct ldlm_pool *pl)
254 slv = pl->pl_server_lock_volume;
255 grant_plan = pl->pl_grant_plan;
256 limit = ldlm_pool_get_limit(pl);
257 granted = atomic_read(&pl->pl_granted);
258 round_up = granted < limit;
260 grant_usage = max_t(int, limit - (granted - grant_plan), 1);
263 * Find out SLV change factor which is the ratio of grant usage
264 * from limit. SLV changes as fast as the ratio of grant plan
265 * consumption. The more locks from grant plan are not consumed
266 * by clients in last interval (idle time), the faster grows
267 * SLV. And the opposite, the more grant plan is over-consumed
268 * (load time) the faster drops SLV.
270 slv_factor = (grant_usage << LDLM_POOL_SLV_SHIFT);
271 do_div(slv_factor, limit);
272 slv = slv * slv_factor;
273 slv = dru(slv, LDLM_POOL_SLV_SHIFT, round_up);
275 if (slv > ldlm_pool_slv_max(limit)) {
276 slv = ldlm_pool_slv_max(limit);
277 } else if (slv < ldlm_pool_slv_min(limit)) {
278 slv = ldlm_pool_slv_min(limit);
281 pl->pl_server_lock_volume = slv;
285 * Recalculates next stats on passed \a pl.
287 * \pre ->pl_lock is locked.
289 static void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
291 int grant_plan = pl->pl_grant_plan;
292 __u64 slv = pl->pl_server_lock_volume;
293 int granted = atomic_read(&pl->pl_granted);
294 int grant_rate = atomic_read(&pl->pl_grant_rate);
295 int cancel_rate = atomic_read(&pl->pl_cancel_rate);
297 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT,
299 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
301 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
303 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
305 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
310 * Sets current SLV into obd accessible via ldlm_pl2ns(pl)->ns_obd.
312 static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl)
314 struct obd_device *obd;
317 * Set new SLV in obd field for using it later without accessing the
318 * pool. This is required to avoid race between sending reply to client
319 * with new SLV and cleanup server stack in which we can't guarantee
320 * that namespace is still alive. We know only that obd is alive as
321 * long as valid export is alive.
323 obd = ldlm_pl2ns(pl)->ns_obd;
324 LASSERT(obd != NULL);
325 write_lock(&obd->obd_pool_lock);
326 obd->obd_pool_slv = pl->pl_server_lock_volume;
327 write_unlock(&obd->obd_pool_lock);
331 * Recalculates all pool fields on passed \a pl.
333 * \pre ->pl_lock is not locked.
335 static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
337 time_t recalc_interval_sec;
340 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
341 if (recalc_interval_sec < pl->pl_recalc_period)
344 spin_lock(&pl->pl_lock);
345 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
346 if (recalc_interval_sec < pl->pl_recalc_period) {
347 spin_unlock(&pl->pl_lock);
351 * Recalc SLV after last period. This should be done
352 * _before_ recalculating new grant plan.
354 ldlm_pool_recalc_slv(pl);
357 * Make sure that pool informed obd of last SLV changes.
359 ldlm_srv_pool_push_slv(pl);
362 * Update grant_plan for new period.
364 ldlm_pool_recalc_grant_plan(pl);
366 pl->pl_recalc_time = cfs_time_current_sec();
367 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
368 recalc_interval_sec);
369 spin_unlock(&pl->pl_lock);
374 * This function is used on server side as main entry point for memory
375 * pressure handling. It decreases SLV on \a pl according to passed
376 * \a nr and \a gfp_mask.
378 * Our goal here is to decrease SLV such a way that clients hold \a nr
379 * locks smaller in next 10h.
381 static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
382 int nr, unsigned int gfp_mask)
387 * VM is asking how many entries may be potentially freed.
390 return atomic_read(&pl->pl_granted);
393 * Client already canceled locks but server is already in shrinker
394 * and can't cancel anything. Let's catch this race.
396 if (atomic_read(&pl->pl_granted) == 0)
399 spin_lock(&pl->pl_lock);
402 * We want shrinker to possibly cause cancellation of @nr locks from
403 * clients or grant approximately @nr locks smaller next intervals.
405 * This is why we decreased SLV by @nr. This effect will only be as
406 * long as one re-calc interval (1s these days) and this should be
407 * enough to pass this decreased SLV to all clients. On next recalc
408 * interval pool will either increase SLV if locks load is not high
409 * or will keep on same level or even decrease again, thus, shrinker
410 * decreased SLV will affect next recalc intervals and this way will
411 * make locking load lower.
413 if (nr < pl->pl_server_lock_volume) {
414 pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr;
416 limit = ldlm_pool_get_limit(pl);
417 pl->pl_server_lock_volume = ldlm_pool_slv_min(limit);
421 * Make sure that pool informed obd of last SLV changes.
423 ldlm_srv_pool_push_slv(pl);
424 spin_unlock(&pl->pl_lock);
427 * We did not really free any memory here so far, it only will be
428 * freed later may be, so that we return 0 to not confuse VM.
434 * Setup server side pool \a pl with passed \a limit.
436 static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit)
438 struct obd_device *obd;
440 obd = ldlm_pl2ns(pl)->ns_obd;
441 LASSERT(obd != NULL && obd != LP_POISON);
442 LASSERT(obd->obd_type != LP_POISON);
443 write_lock(&obd->obd_pool_lock);
444 obd->obd_pool_limit = limit;
445 write_unlock(&obd->obd_pool_lock);
447 ldlm_pool_set_limit(pl, limit);
452 * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl.
454 static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
456 struct obd_device *obd;
459 * Get new SLV and Limit from obd which is updated with coming
462 obd = ldlm_pl2ns(pl)->ns_obd;
463 LASSERT(obd != NULL);
464 read_lock(&obd->obd_pool_lock);
465 pl->pl_server_lock_volume = obd->obd_pool_slv;
466 ldlm_pool_set_limit(pl, obd->obd_pool_limit);
467 read_unlock(&obd->obd_pool_lock);
471 * Recalculates client size pool \a pl according to current SLV and Limit.
473 static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
475 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);
497 pl->pl_recalc_time = cfs_time_current_sec();
498 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
499 recalc_interval_sec);
500 spin_unlock(&pl->pl_lock);
503 * Do not cancel locks in case lru resize is disabled for this ns.
505 if (!ns_connect_lru_resize(ldlm_pl2ns(pl)))
509 * In the time of canceling locks on client we do not need to maintain
510 * sharp timing, we only want to cancel locks asap according to new SLV.
511 * It may be called when SLV has changed much, this is why we do not
512 * take into account pl->pl_recalc_time here.
514 RETURN(ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC,
519 * This function is main entry point for memory pressure handling on client
520 * side. Main goal of this function is to cancel some number of locks on
521 * passed \a pl according to \a nr and \a gfp_mask.
523 static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
524 int nr, unsigned int gfp_mask)
526 struct ldlm_namespace *ns;
527 int canceled = 0, unused;
532 * Do not cancel locks in case lru resize is disabled for this ns.
534 if (!ns_connect_lru_resize(ns))
538 * Make sure that pool knows last SLV and Limit from obd.
540 ldlm_cli_pool_pop_slv(pl);
542 spin_lock(&ns->ns_lock);
543 unused = ns->ns_nr_unused;
544 spin_unlock(&ns->ns_lock);
547 canceled = ldlm_cancel_lru(ns, nr, LCF_ASYNC,
551 * Return the number of potentially reclaimable locks.
553 return ((unused - canceled) / 100) * sysctl_vfs_cache_pressure;
556 struct ldlm_pool_ops ldlm_srv_pool_ops = {
557 .po_recalc = ldlm_srv_pool_recalc,
558 .po_shrink = ldlm_srv_pool_shrink,
559 .po_setup = ldlm_srv_pool_setup
562 struct ldlm_pool_ops ldlm_cli_pool_ops = {
563 .po_recalc = ldlm_cli_pool_recalc,
564 .po_shrink = ldlm_cli_pool_shrink
568 * Pool recalc wrapper. Will call either client or server pool recalc callback
569 * depending what pool \a pl is used.
571 int ldlm_pool_recalc(struct ldlm_pool *pl)
573 time_t recalc_interval_sec;
576 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
577 if (recalc_interval_sec <= 0)
580 spin_lock(&pl->pl_lock);
581 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
582 if (recalc_interval_sec > 0) {
584 * Update pool statistics every 1s.
586 ldlm_pool_recalc_stats(pl);
589 * Zero out all rates and speed for the last period.
591 atomic_set(&pl->pl_grant_rate, 0);
592 atomic_set(&pl->pl_cancel_rate, 0);
594 spin_unlock(&pl->pl_lock);
597 if (pl->pl_ops->po_recalc != NULL) {
598 count = pl->pl_ops->po_recalc(pl);
599 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
606 EXPORT_SYMBOL(ldlm_pool_recalc);
609 * Pool shrink wrapper. Will call either client or server pool recalc callback
610 * depending what pool \a pl is used.
612 int ldlm_pool_shrink(struct ldlm_pool *pl, int nr,
613 unsigned int gfp_mask)
617 if (pl->pl_ops->po_shrink != NULL) {
618 cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask);
620 lprocfs_counter_add(pl->pl_stats,
621 LDLM_POOL_SHRINK_REQTD_STAT,
623 lprocfs_counter_add(pl->pl_stats,
624 LDLM_POOL_SHRINK_FREED_STAT,
626 CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, "
627 "shrunk %d\n", pl->pl_name, nr, cancel);
632 EXPORT_SYMBOL(ldlm_pool_shrink);
635 * Pool setup wrapper. Will call either client or server pool recalc callback
636 * depending what pool \a pl is used.
638 * Sets passed \a limit into pool \a pl.
640 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
642 if (pl->pl_ops->po_setup != NULL)
643 return(pl->pl_ops->po_setup(pl, limit));
646 EXPORT_SYMBOL(ldlm_pool_setup);
648 static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
650 int granted, grant_rate, cancel_rate, grant_step;
651 int grant_speed, grant_plan, lvf;
652 struct ldlm_pool *pl = m->private;
656 spin_lock(&pl->pl_lock);
657 slv = pl->pl_server_lock_volume;
658 clv = pl->pl_client_lock_volume;
659 limit = ldlm_pool_get_limit(pl);
660 grant_plan = pl->pl_grant_plan;
661 granted = atomic_read(&pl->pl_granted);
662 grant_rate = atomic_read(&pl->pl_grant_rate);
663 cancel_rate = atomic_read(&pl->pl_cancel_rate);
664 grant_speed = grant_rate - cancel_rate;
665 lvf = atomic_read(&pl->pl_lock_volume_factor);
666 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
667 spin_unlock(&pl->pl_lock);
669 seq_printf(m, "LDLM pool state (%s):\n"
673 pl->pl_name, slv, clv, lvf);
675 if (ns_is_server(ldlm_pl2ns(pl))) {
676 seq_printf(m, " GSP: %d%%\n"
678 grant_step, grant_plan);
680 seq_printf(m, " GR: %d\n" " CR: %d\n" " GS: %d\n"
681 " G: %d\n" " L: %d\n",
682 grant_rate, cancel_rate, grant_speed,
687 LPROC_SEQ_FOPS_RO(lprocfs_pool_state);
689 static int lprocfs_grant_speed_seq_show(struct seq_file *m, void *unused)
691 struct ldlm_pool *pl = m->private;
694 spin_lock(&pl->pl_lock);
695 /* serialize with ldlm_pool_recalc */
696 grant_speed = atomic_read(&pl->pl_grant_rate) -
697 atomic_read(&pl->pl_cancel_rate);
698 spin_unlock(&pl->pl_lock);
699 return lprocfs_rd_uint(m, &grant_speed);
702 LDLM_POOL_PROC_READER_SEQ_SHOW(grant_plan, int);
703 LPROC_SEQ_FOPS_RO(lprocfs_grant_plan);
705 LDLM_POOL_PROC_READER_SEQ_SHOW(recalc_period, int);
706 LDLM_POOL_PROC_WRITER(recalc_period, int);
707 static ssize_t lprocfs_recalc_period_seq_write(struct file *file, const char *buf,
708 size_t len, loff_t *off)
710 struct seq_file *seq = file->private_data;
712 return lprocfs_wr_recalc_period(file, buf, len, seq->private);
714 LPROC_SEQ_FOPS(lprocfs_recalc_period);
716 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, u64);
717 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, atomic);
718 LPROC_SEQ_FOPS_RW_TYPE(ldlm_pool_rw, atomic);
720 LPROC_SEQ_FOPS_RO(lprocfs_grant_speed);
722 #define LDLM_POOL_ADD_VAR(name, var, ops) \
724 snprintf(var_name, MAX_STRING_SIZE, #name); \
725 pool_vars[0].data = var; \
726 pool_vars[0].fops = ops; \
727 lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);\
730 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
732 struct ldlm_namespace *ns = ldlm_pl2ns(pl);
733 struct proc_dir_entry *parent_ns_proc;
734 struct lprocfs_vars pool_vars[2];
735 char *var_name = NULL;
739 OBD_ALLOC(var_name, MAX_STRING_SIZE + 1);
743 parent_ns_proc = ns->ns_proc_dir_entry;
744 if (parent_ns_proc == NULL) {
745 CERROR("%s: proc entry is not initialized\n",
747 GOTO(out_free_name, rc = -EINVAL);
749 pl->pl_proc_dir = lprocfs_register("pool", parent_ns_proc,
751 if (IS_ERR(pl->pl_proc_dir)) {
752 CERROR("LProcFS failed in ldlm-pool-init\n");
753 rc = PTR_ERR(pl->pl_proc_dir);
754 GOTO(out_free_name, rc);
757 var_name[MAX_STRING_SIZE] = '\0';
758 memset(pool_vars, 0, sizeof(pool_vars));
759 pool_vars[0].name = var_name;
761 LDLM_POOL_ADD_VAR("server_lock_volume", &pl->pl_server_lock_volume,
762 &ldlm_pool_u64_fops);
763 LDLM_POOL_ADD_VAR("limit", &pl->pl_limit, &ldlm_pool_rw_atomic_fops);
764 LDLM_POOL_ADD_VAR("granted", &pl->pl_granted, &ldlm_pool_atomic_fops);
765 LDLM_POOL_ADD_VAR("grant_speed", pl, &lprocfs_grant_speed_fops);
766 LDLM_POOL_ADD_VAR("cancel_rate", &pl->pl_cancel_rate,
767 &ldlm_pool_atomic_fops);
768 LDLM_POOL_ADD_VAR("grant_rate", &pl->pl_grant_rate,
769 &ldlm_pool_atomic_fops);
770 LDLM_POOL_ADD_VAR("grant_plan", pl, &lprocfs_grant_plan_fops);
771 LDLM_POOL_ADD_VAR("recalc_period", pl, &lprocfs_recalc_period_fops);
772 LDLM_POOL_ADD_VAR("lock_volume_factor", &pl->pl_lock_volume_factor,
773 &ldlm_pool_rw_atomic_fops);
774 LDLM_POOL_ADD_VAR("state", pl, &lprocfs_pool_state_fops);
776 pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT -
777 LDLM_POOL_FIRST_STAT, 0);
779 GOTO(out_free_name, rc = -ENOMEM);
781 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
782 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
784 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
785 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
787 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
788 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
790 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
791 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
792 "grant_rate", "locks/s");
793 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
794 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
795 "cancel_rate", "locks/s");
796 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
797 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
798 "grant_plan", "locks/s");
799 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT,
800 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
802 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT,
803 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
804 "shrink_request", "locks");
805 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT,
806 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
807 "shrink_freed", "locks");
808 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
809 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
810 "recalc_freed", "locks");
811 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
812 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
813 "recalc_timing", "sec");
814 rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
818 OBD_FREE(var_name, MAX_STRING_SIZE + 1);
822 static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
824 if (pl->pl_stats != NULL) {
825 lprocfs_free_stats(&pl->pl_stats);
828 if (pl->pl_proc_dir != NULL) {
829 lprocfs_remove(&pl->pl_proc_dir);
830 pl->pl_proc_dir = NULL;
834 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
835 int idx, ldlm_side_t client)
840 spin_lock_init(&pl->pl_lock);
841 atomic_set(&pl->pl_granted, 0);
842 pl->pl_recalc_time = cfs_time_current_sec();
843 atomic_set(&pl->pl_lock_volume_factor, 1);
845 atomic_set(&pl->pl_grant_rate, 0);
846 atomic_set(&pl->pl_cancel_rate, 0);
847 pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
849 snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
850 ldlm_ns_name(ns), idx);
852 if (client == LDLM_NAMESPACE_SERVER) {
853 pl->pl_ops = &ldlm_srv_pool_ops;
854 ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L);
855 pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD;
856 pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L);
858 ldlm_pool_set_limit(pl, 1);
859 pl->pl_server_lock_volume = 0;
860 pl->pl_ops = &ldlm_cli_pool_ops;
861 pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
863 pl->pl_client_lock_volume = 0;
864 rc = ldlm_pool_proc_init(pl);
868 CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name);
872 EXPORT_SYMBOL(ldlm_pool_init);
874 void ldlm_pool_fini(struct ldlm_pool *pl)
877 ldlm_pool_proc_fini(pl);
880 * Pool should not be used after this point. We can't free it here as
881 * it lives in struct ldlm_namespace, but still interested in catching
882 * any abnormal using cases.
884 POISON(pl, 0x5a, sizeof(*pl));
887 EXPORT_SYMBOL(ldlm_pool_fini);
890 * Add new taken ldlm lock \a lock into pool \a pl accounting.
892 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
895 * FLOCK locks are special in a sense that they are almost never
896 * cancelled, instead special kind of lock is used to drop them.
897 * also there is no LRU for flock locks, so no point in tracking
900 if (lock->l_resource->lr_type == LDLM_FLOCK)
903 atomic_inc(&pl->pl_granted);
904 atomic_inc(&pl->pl_grant_rate);
905 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
907 * Do not do pool recalc for client side as all locks which
908 * potentially may be canceled has already been packed into
909 * enqueue/cancel rpc. Also we do not want to run out of stack
910 * with too long call paths.
912 if (ns_is_server(ldlm_pl2ns(pl)))
913 ldlm_pool_recalc(pl);
915 EXPORT_SYMBOL(ldlm_pool_add);
918 * Remove ldlm lock \a lock from pool \a pl accounting.
920 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
923 * Filter out FLOCK locks. Read above comment in ldlm_pool_add().
925 if (lock->l_resource->lr_type == LDLM_FLOCK)
928 LASSERT(atomic_read(&pl->pl_granted) > 0);
929 atomic_dec(&pl->pl_granted);
930 atomic_inc(&pl->pl_cancel_rate);
932 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
934 if (ns_is_server(ldlm_pl2ns(pl)))
935 ldlm_pool_recalc(pl);
937 EXPORT_SYMBOL(ldlm_pool_del);
940 * Returns current \a pl SLV.
942 * \pre ->pl_lock is not locked.
944 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
947 spin_lock(&pl->pl_lock);
948 slv = pl->pl_server_lock_volume;
949 spin_unlock(&pl->pl_lock);
952 EXPORT_SYMBOL(ldlm_pool_get_slv);
955 * Sets passed \a slv to \a pl.
957 * \pre ->pl_lock is not locked.
959 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
961 spin_lock(&pl->pl_lock);
962 pl->pl_server_lock_volume = slv;
963 spin_unlock(&pl->pl_lock);
965 EXPORT_SYMBOL(ldlm_pool_set_slv);
968 * Returns current \a pl CLV.
970 * \pre ->pl_lock is not locked.
972 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
975 spin_lock(&pl->pl_lock);
976 slv = pl->pl_client_lock_volume;
977 spin_unlock(&pl->pl_lock);
980 EXPORT_SYMBOL(ldlm_pool_get_clv);
983 * Sets passed \a clv to \a pl.
985 * \pre ->pl_lock is not locked.
987 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
989 spin_lock(&pl->pl_lock);
990 pl->pl_client_lock_volume = clv;
991 spin_unlock(&pl->pl_lock);
993 EXPORT_SYMBOL(ldlm_pool_set_clv);
996 * Returns current \a pl limit.
998 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1000 return atomic_read(&pl->pl_limit);
1002 EXPORT_SYMBOL(ldlm_pool_get_limit);
1005 * Sets passed \a limit to \a pl.
1007 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1009 atomic_set(&pl->pl_limit, limit);
1011 EXPORT_SYMBOL(ldlm_pool_set_limit);
1014 * Returns current LVF from \a pl.
1016 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1018 return atomic_read(&pl->pl_lock_volume_factor);
1020 EXPORT_SYMBOL(ldlm_pool_get_lvf);
1022 static int ldlm_pool_granted(struct ldlm_pool *pl)
1024 return atomic_read(&pl->pl_granted);
1027 static struct ptlrpc_thread *ldlm_pools_thread;
1028 static struct shrinker *ldlm_pools_srv_shrinker;
1029 static struct shrinker *ldlm_pools_cli_shrinker;
1030 static struct completion ldlm_pools_comp;
1033 * Cancel \a nr locks from all namespaces (if possible). Returns number of
1034 * cached locks after shrink is finished. All namespaces are asked to
1035 * cancel approximately equal amount of locks to keep balancing.
1037 static int ldlm_pools_shrink(ldlm_side_t client, int nr,
1038 unsigned int gfp_mask)
1040 int total = 0, cached = 0, nr_ns;
1041 struct ldlm_namespace *ns;
1044 if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
1045 !(gfp_mask & __GFP_FS))
1048 CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n",
1049 nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1051 cookie = cl_env_reenter();
1054 * Find out how many resources we may release.
1056 for (nr_ns = atomic_read(ldlm_namespace_nr(client));
1059 mutex_lock(ldlm_namespace_lock(client));
1060 if (list_empty(ldlm_namespace_list(client))) {
1061 mutex_unlock(ldlm_namespace_lock(client));
1062 cl_env_reexit(cookie);
1065 ns = ldlm_namespace_first_locked(client);
1066 ldlm_namespace_get(ns);
1067 ldlm_namespace_move_locked(ns, client);
1068 mutex_unlock(ldlm_namespace_lock(client));
1069 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
1070 ldlm_namespace_put(ns);
1073 if (nr == 0 || total == 0) {
1074 cl_env_reexit(cookie);
1079 * Shrink at least ldlm_namespace_nr(client) namespaces.
1081 for (nr_ns = atomic_read(ldlm_namespace_nr(client));
1084 int cancel, nr_locks;
1087 * Do not call shrink under ldlm_namespace_lock(client)
1089 mutex_lock(ldlm_namespace_lock(client));
1090 if (list_empty(ldlm_namespace_list(client))) {
1091 mutex_unlock(ldlm_namespace_lock(client));
1093 * If list is empty, we can't return any @cached > 0,
1094 * that probably would cause needless shrinker
1100 ns = ldlm_namespace_first_locked(client);
1101 ldlm_namespace_get(ns);
1102 ldlm_namespace_move_locked(ns, client);
1103 mutex_unlock(ldlm_namespace_lock(client));
1105 nr_locks = ldlm_pool_granted(&ns->ns_pool);
1106 cancel = 1 + nr_locks * nr / total;
1107 ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
1108 cached += ldlm_pool_granted(&ns->ns_pool);
1109 ldlm_namespace_put(ns);
1111 cl_env_reexit(cookie);
1112 /* we only decrease the SLV in server pools shrinker, return -1 to
1113 * kernel to avoid needless loop. LU-1128 */
1114 return (client == LDLM_NAMESPACE_SERVER) ? -1 : cached;
1117 static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1119 return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER,
1120 shrink_param(sc, nr_to_scan),
1121 shrink_param(sc, gfp_mask));
1124 static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1126 return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT,
1127 shrink_param(sc, nr_to_scan),
1128 shrink_param(sc, gfp_mask));
1131 void ldlm_pools_recalc(ldlm_side_t client)
1133 __u32 nr_l = 0, nr_p = 0, l;
1134 struct ldlm_namespace *ns;
1138 * No need to setup pool limit for client pools.
1140 if (client == LDLM_NAMESPACE_SERVER) {
1142 * Check all modest namespaces first.
1144 mutex_lock(ldlm_namespace_lock(client));
1145 list_for_each_entry(ns, ldlm_namespace_list(client),
1148 if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
1151 l = ldlm_pool_granted(&ns->ns_pool);
1156 * Set the modest pools limit equal to their avg granted
1159 l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0);
1160 ldlm_pool_setup(&ns->ns_pool, l);
1166 * Make sure that modest namespaces did not eat more that 2/3
1169 if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) {
1170 CWARN("\"Modest\" pools eat out 2/3 of server locks "
1171 "limit (%d of %lu). This means that you have too "
1172 "many clients for this amount of server RAM. "
1173 "Upgrade server!\n", nr_l, LDLM_POOL_HOST_L);
1178 * The rest is given to greedy namespaces.
1180 list_for_each_entry(ns, ldlm_namespace_list(client),
1183 if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
1188 * In the case 2/3 locks are eaten out by
1189 * modest pools, we re-setup equal limit
1192 l = LDLM_POOL_HOST_L /
1194 ldlm_namespace_nr(client));
1197 * All the rest of greedy pools will have
1198 * all locks in equal parts.
1200 l = (LDLM_POOL_HOST_L - nr_l) /
1202 ldlm_namespace_nr(client)) -
1205 ldlm_pool_setup(&ns->ns_pool, l);
1207 mutex_unlock(ldlm_namespace_lock(client));
1211 * Recalc at least ldlm_namespace_nr(client) namespaces.
1213 for (nr = atomic_read(ldlm_namespace_nr(client)); nr > 0; nr--) {
1216 * Lock the list, get first @ns in the list, getref, move it
1217 * to the tail, unlock and call pool recalc. This way we avoid
1218 * calling recalc under @ns lock what is really good as we get
1219 * rid of potential deadlock on client nodes when canceling
1220 * locks synchronously.
1222 mutex_lock(ldlm_namespace_lock(client));
1223 if (list_empty(ldlm_namespace_list(client))) {
1224 mutex_unlock(ldlm_namespace_lock(client));
1227 ns = ldlm_namespace_first_locked(client);
1229 spin_lock(&ns->ns_lock);
1231 * skip ns which is being freed, and we don't want to increase
1232 * its refcount again, not even temporarily. bz21519 & LU-499.
1234 if (ns->ns_stopping) {
1238 ldlm_namespace_get(ns);
1240 spin_unlock(&ns->ns_lock);
1242 ldlm_namespace_move_locked(ns, client);
1243 mutex_unlock(ldlm_namespace_lock(client));
1246 * After setup is done - recalc the pool.
1249 ldlm_pool_recalc(&ns->ns_pool);
1250 ldlm_namespace_put(ns);
1254 EXPORT_SYMBOL(ldlm_pools_recalc);
1256 static int ldlm_pools_thread_main(void *arg)
1258 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
1261 thread_set_flags(thread, SVC_RUNNING);
1262 wake_up(&thread->t_ctl_waitq);
1264 CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n",
1265 "ldlm_poold", current_pid());
1268 struct l_wait_info lwi;
1271 * Recal all pools on this tick.
1273 ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
1274 ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
1277 * Wait until the next check time, or until we're
1280 lwi = LWI_TIMEOUT(cfs_time_seconds(LDLM_POOLS_THREAD_PERIOD),
1282 l_wait_event(thread->t_ctl_waitq,
1283 thread_is_stopping(thread) ||
1284 thread_is_event(thread),
1287 if (thread_test_and_clear_flags(thread, SVC_STOPPING))
1290 thread_test_and_clear_flags(thread, SVC_EVENT);
1293 thread_set_flags(thread, SVC_STOPPED);
1294 wake_up(&thread->t_ctl_waitq);
1296 CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n",
1297 "ldlm_poold", current_pid());
1299 complete_and_exit(&ldlm_pools_comp, 0);
1302 static int ldlm_pools_thread_start(void)
1304 struct l_wait_info lwi = { 0 };
1308 if (ldlm_pools_thread != NULL)
1311 OBD_ALLOC_PTR(ldlm_pools_thread);
1312 if (ldlm_pools_thread == NULL)
1315 init_completion(&ldlm_pools_comp);
1316 init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq);
1318 task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread,
1321 CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task));
1322 OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread));
1323 ldlm_pools_thread = NULL;
1324 RETURN(PTR_ERR(task));
1326 l_wait_event(ldlm_pools_thread->t_ctl_waitq,
1327 thread_is_running(ldlm_pools_thread), &lwi);
1331 static void ldlm_pools_thread_stop(void)
1335 if (ldlm_pools_thread == NULL) {
1340 thread_set_flags(ldlm_pools_thread, SVC_STOPPING);
1341 wake_up(&ldlm_pools_thread->t_ctl_waitq);
1344 * Make sure that pools thread is finished before freeing @thread.
1345 * This fixes possible race and oops due to accessing freed memory
1348 wait_for_completion(&ldlm_pools_comp);
1349 OBD_FREE_PTR(ldlm_pools_thread);
1350 ldlm_pools_thread = NULL;
1354 int ldlm_pools_init(void)
1359 rc = ldlm_pools_thread_start();
1361 ldlm_pools_srv_shrinker =
1362 set_shrinker(DEFAULT_SEEKS,
1363 ldlm_pools_srv_shrink);
1364 ldlm_pools_cli_shrinker =
1365 set_shrinker(DEFAULT_SEEKS,
1366 ldlm_pools_cli_shrink);
1370 EXPORT_SYMBOL(ldlm_pools_init);
1372 void ldlm_pools_fini(void)
1374 if (ldlm_pools_srv_shrinker != NULL) {
1375 remove_shrinker(ldlm_pools_srv_shrinker);
1376 ldlm_pools_srv_shrinker = NULL;
1378 if (ldlm_pools_cli_shrinker != NULL) {
1379 remove_shrinker(ldlm_pools_cli_shrinker);
1380 ldlm_pools_cli_shrinker = NULL;
1382 ldlm_pools_thread_stop();
1384 EXPORT_SYMBOL(ldlm_pools_fini);