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, 2015, 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
58 * sensitive client should be about last SLV from server. The higher LVF is the
59 * more locks will be canceled on client. Default value for it is 1. Setting LVF
60 * to 2 means 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).
74 * Andreas Dilger (adilger@clusterfs.com) proposed few nice ideas like using
75 * LVF and many cleanups. Flow definition to allow more easy understanding of
76 * the logic belongs to Nikita Danilov (nikita@clusterfs.com) as well as many
77 * cleanups and fixes. And design and implementation are done by Yury Umanets
78 * (umka@clusterfs.com).
80 * Glossary for terms used:
82 * pl_limit - Number of allowed locks in pool. Applies to server and client
85 * pl_granted - Number of granted locks (calculated);
86 * pl_grant_rate - Number of granted locks for last T (calculated);
87 * pl_cancel_rate - Number of canceled locks for last T (calculated);
88 * pl_grant_speed - Grant speed (GR - CR) for last T (calculated);
89 * pl_grant_plan - Planned number of granted locks for next T (calculated);
90 * pl_server_lock_volume - Current server lock volume (calculated);
92 * As it may be seen from list above, we have few possible tunables which may
93 * affect behavior much. They all may be modified via sysfs. However, they also
94 * give a possibility for constructing few pre-defined behavior policies. If
95 * none of predefines is suitable for a working pattern being used, new one may
96 * be "constructed" via sysfs tunables.
99 #define DEBUG_SUBSYSTEM S_LDLM
101 #include "../include/lustre_dlm.h"
102 #include "../include/cl_object.h"
103 #include "../include/obd_class.h"
104 #include "../include/obd_support.h"
105 #include "ldlm_internal.h"
108 * 50 ldlm locks for 1MB of RAM.
110 #define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_CACHE_SHIFT)) * 50)
113 * Maximal possible grant step plan in %.
115 #define LDLM_POOL_MAX_GSP (30)
118 * Minimal possible grant step plan in %.
120 #define LDLM_POOL_MIN_GSP (1)
123 * This controls the speed of reaching LDLM_POOL_MAX_GSP
124 * with increasing thread period.
126 #define LDLM_POOL_GSP_STEP_SHIFT (2)
129 * LDLM_POOL_GSP% of all locks is default GP.
131 #define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100)
134 * Max age for locks on clients.
136 #define LDLM_POOL_MAX_AGE (36000)
139 * The granularity of SLV calculation.
141 #define LDLM_POOL_SLV_SHIFT (10)
143 static inline __u64 dru(__u64 val, __u32 shift, int round_up)
145 return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift;
148 static inline __u64 ldlm_pool_slv_max(__u32 L)
151 * Allow to have all locks for 1 client for 10 hrs.
152 * Formula is the following: limit * 10h / 1 client.
154 __u64 lim = (__u64)L * LDLM_POOL_MAX_AGE / 1;
158 static inline __u64 ldlm_pool_slv_min(__u32 L)
164 LDLM_POOL_FIRST_STAT = 0,
165 LDLM_POOL_GRANTED_STAT = LDLM_POOL_FIRST_STAT,
166 LDLM_POOL_GRANT_STAT,
167 LDLM_POOL_CANCEL_STAT,
168 LDLM_POOL_GRANT_RATE_STAT,
169 LDLM_POOL_CANCEL_RATE_STAT,
170 LDLM_POOL_GRANT_PLAN_STAT,
172 LDLM_POOL_SHRINK_REQTD_STAT,
173 LDLM_POOL_SHRINK_FREED_STAT,
174 LDLM_POOL_RECALC_STAT,
175 LDLM_POOL_TIMING_STAT,
180 * Calculates suggested grant_step in % of available locks for passed
181 * \a period. This is later used in grant_plan calculations.
183 static inline int ldlm_pool_t2gsp(unsigned int t)
186 * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP
187 * and up to 30% for anything higher than LDLM_POOL_GSP_STEP.
189 * How this will affect execution is the following:
191 * - for thread period 1s we will have grant_step 1% which good from
192 * pov of taking some load off from server and push it out to clients.
193 * This is like that because 1% for grant_step means that server will
194 * not allow clients to get lots of locks in short period of time and
195 * keep all old locks in their caches. Clients will always have to
196 * get some locks back if they want to take some new;
198 * - for thread period 10s (which is default) we will have 23% which
199 * means that clients will have enough of room to take some new locks
200 * without getting some back. All locks from this 23% which were not
201 * taken by clients in current period will contribute in SLV growing.
202 * SLV growing means more locks cached on clients until limit or grant
205 return LDLM_POOL_MAX_GSP -
206 ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >>
207 (t >> LDLM_POOL_GSP_STEP_SHIFT));
211 * Recalculates next stats on passed \a pl.
213 * \pre ->pl_lock is locked.
215 static void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
217 int grant_plan = pl->pl_grant_plan;
218 __u64 slv = pl->pl_server_lock_volume;
219 int granted = atomic_read(&pl->pl_granted);
220 int grant_rate = atomic_read(&pl->pl_grant_rate);
221 int cancel_rate = atomic_read(&pl->pl_cancel_rate);
223 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT,
225 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
227 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
229 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
231 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
236 * Sets SLV and Limit from container_of(pl, struct ldlm_namespace,
237 * ns_pool)->ns_obd tp passed \a pl.
239 static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
241 struct obd_device *obd;
244 * Get new SLV and Limit from obd which is updated with coming
247 obd = container_of(pl, struct ldlm_namespace,
249 read_lock(&obd->obd_pool_lock);
250 pl->pl_server_lock_volume = obd->obd_pool_slv;
251 atomic_set(&pl->pl_limit, obd->obd_pool_limit);
252 read_unlock(&obd->obd_pool_lock);
256 * Recalculates client size pool \a pl according to current SLV and Limit.
258 static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
260 time64_t recalc_interval_sec;
263 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time;
264 if (recalc_interval_sec < pl->pl_recalc_period)
267 spin_lock(&pl->pl_lock);
269 * Check if we need to recalc lists now.
271 recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time;
272 if (recalc_interval_sec < pl->pl_recalc_period) {
273 spin_unlock(&pl->pl_lock);
278 * Make sure that pool knows last SLV and Limit from obd.
280 ldlm_cli_pool_pop_slv(pl);
282 spin_unlock(&pl->pl_lock);
285 * Do not cancel locks in case lru resize is disabled for this ns.
287 if (!ns_connect_lru_resize(container_of(pl, struct ldlm_namespace,
294 * In the time of canceling locks on client we do not need to maintain
295 * sharp timing, we only want to cancel locks asap according to new SLV.
296 * It may be called when SLV has changed much, this is why we do not
297 * take into account pl->pl_recalc_time here.
299 ret = ldlm_cancel_lru(container_of(pl, struct ldlm_namespace, ns_pool),
300 0, LCF_ASYNC, LDLM_CANCEL_LRUR);
303 spin_lock(&pl->pl_lock);
305 * Time of LRU resizing might be longer than period,
306 * so update after LRU resizing rather than before it.
308 pl->pl_recalc_time = ktime_get_real_seconds();
309 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
310 recalc_interval_sec);
311 spin_unlock(&pl->pl_lock);
316 * This function is main entry point for memory pressure handling on client
317 * side. Main goal of this function is to cancel some number of locks on
318 * passed \a pl according to \a nr and \a gfp_mask.
320 static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
321 int nr, gfp_t gfp_mask)
323 struct ldlm_namespace *ns;
326 ns = container_of(pl, struct ldlm_namespace, ns_pool);
329 * Do not cancel locks in case lru resize is disabled for this ns.
331 if (!ns_connect_lru_resize(ns))
335 * Make sure that pool knows last SLV and Limit from obd.
337 ldlm_cli_pool_pop_slv(pl);
339 spin_lock(&ns->ns_lock);
340 unused = ns->ns_nr_unused;
341 spin_unlock(&ns->ns_lock);
344 return (unused / 100) * sysctl_vfs_cache_pressure;
346 return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_CANCEL_SHRINK);
349 static const struct ldlm_pool_ops ldlm_cli_pool_ops = {
350 .po_recalc = ldlm_cli_pool_recalc,
351 .po_shrink = ldlm_cli_pool_shrink
355 * Pool recalc wrapper. Will call either client or server pool recalc callback
356 * depending what pool \a pl is used.
358 static int ldlm_pool_recalc(struct ldlm_pool *pl)
360 u32 recalc_interval_sec;
363 recalc_interval_sec = ktime_get_seconds() - pl->pl_recalc_time;
364 if (recalc_interval_sec <= 0)
367 spin_lock(&pl->pl_lock);
368 if (recalc_interval_sec > 0) {
370 * Update pool statistics every 1s.
372 ldlm_pool_recalc_stats(pl);
375 * Zero out all rates and speed for the last period.
377 atomic_set(&pl->pl_grant_rate, 0);
378 atomic_set(&pl->pl_cancel_rate, 0);
380 spin_unlock(&pl->pl_lock);
383 if (pl->pl_ops->po_recalc) {
384 count = pl->pl_ops->po_recalc(pl);
385 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
388 recalc_interval_sec = pl->pl_recalc_time - ktime_get_seconds() +
389 pl->pl_recalc_period;
390 if (recalc_interval_sec <= 0) {
391 /* Prevent too frequent recalculation. */
393 "Negative interval(%d), too short period(%lld)",
395 (s64)pl->pl_recalc_period);
396 recalc_interval_sec = 1;
399 return recalc_interval_sec;
403 * Pool shrink wrapper. Will call either client or server pool recalc callback
404 * depending what pool pl is used. When nr == 0, just return the number of
405 * freeable locks. Otherwise, return the number of canceled locks.
407 static int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, gfp_t gfp_mask)
411 if (pl->pl_ops->po_shrink) {
412 cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask);
414 lprocfs_counter_add(pl->pl_stats,
415 LDLM_POOL_SHRINK_REQTD_STAT,
417 lprocfs_counter_add(pl->pl_stats,
418 LDLM_POOL_SHRINK_FREED_STAT,
420 CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, shrunk %d\n",
421 pl->pl_name, nr, cancel);
427 static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
429 int granted, grant_rate, cancel_rate;
430 int grant_speed, lvf;
431 struct ldlm_pool *pl = m->private;
435 spin_lock(&pl->pl_lock);
436 slv = pl->pl_server_lock_volume;
437 clv = pl->pl_client_lock_volume;
438 limit = atomic_read(&pl->pl_limit);
439 granted = atomic_read(&pl->pl_granted);
440 grant_rate = atomic_read(&pl->pl_grant_rate);
441 cancel_rate = atomic_read(&pl->pl_cancel_rate);
442 grant_speed = grant_rate - cancel_rate;
443 lvf = atomic_read(&pl->pl_lock_volume_factor);
444 spin_unlock(&pl->pl_lock);
446 seq_printf(m, "LDLM pool state (%s):\n"
450 pl->pl_name, slv, clv, lvf);
452 seq_printf(m, " GR: %d\n CR: %d\n GS: %d\n"
454 grant_rate, cancel_rate, grant_speed,
460 LPROC_SEQ_FOPS_RO(lprocfs_pool_state);
462 static ssize_t grant_speed_show(struct kobject *kobj, struct attribute *attr,
465 struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool,
470 spin_lock(&pl->pl_lock);
471 /* serialize with ldlm_pool_recalc */
472 grant_speed = atomic_read(&pl->pl_grant_rate) -
473 atomic_read(&pl->pl_cancel_rate);
474 spin_unlock(&pl->pl_lock);
475 return sprintf(buf, "%d\n", grant_speed);
477 LUSTRE_RO_ATTR(grant_speed);
479 LDLM_POOL_SYSFS_READER_SHOW(grant_plan, int);
480 LUSTRE_RO_ATTR(grant_plan);
482 LDLM_POOL_SYSFS_READER_SHOW(recalc_period, int);
483 LDLM_POOL_SYSFS_WRITER_STORE(recalc_period, int);
484 LUSTRE_RW_ATTR(recalc_period);
486 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(server_lock_volume, u64);
487 LUSTRE_RO_ATTR(server_lock_volume);
489 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(limit, atomic);
490 LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(limit, atomic);
491 LUSTRE_RW_ATTR(limit);
493 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(granted, atomic);
494 LUSTRE_RO_ATTR(granted);
496 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(cancel_rate, atomic);
497 LUSTRE_RO_ATTR(cancel_rate);
499 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(grant_rate, atomic);
500 LUSTRE_RO_ATTR(grant_rate);
502 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(lock_volume_factor, atomic);
503 LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(lock_volume_factor, atomic);
504 LUSTRE_RW_ATTR(lock_volume_factor);
506 #define LDLM_POOL_ADD_VAR(name, var, ops) \
508 snprintf(var_name, MAX_STRING_SIZE, #name); \
509 pool_vars[0].data = var; \
510 pool_vars[0].fops = ops; \
511 ldebugfs_add_vars(pl->pl_debugfs_entry, pool_vars, NULL);\
514 /* These are for pools in /sys/fs/lustre/ldlm/namespaces/.../pool */
515 static struct attribute *ldlm_pl_attrs[] = {
516 &lustre_attr_grant_speed.attr,
517 &lustre_attr_grant_plan.attr,
518 &lustre_attr_recalc_period.attr,
519 &lustre_attr_server_lock_volume.attr,
520 &lustre_attr_limit.attr,
521 &lustre_attr_granted.attr,
522 &lustre_attr_cancel_rate.attr,
523 &lustre_attr_grant_rate.attr,
524 &lustre_attr_lock_volume_factor.attr,
528 static void ldlm_pl_release(struct kobject *kobj)
530 struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool,
532 complete(&pl->pl_kobj_unregister);
535 static struct kobj_type ldlm_pl_ktype = {
536 .default_attrs = ldlm_pl_attrs,
537 .sysfs_ops = &lustre_sysfs_ops,
538 .release = ldlm_pl_release,
541 static int ldlm_pool_sysfs_init(struct ldlm_pool *pl)
543 struct ldlm_namespace *ns = container_of(pl, struct ldlm_namespace,
547 init_completion(&pl->pl_kobj_unregister);
548 err = kobject_init_and_add(&pl->pl_kobj, &ldlm_pl_ktype, &ns->ns_kobj,
554 static int ldlm_pool_debugfs_init(struct ldlm_pool *pl)
556 struct ldlm_namespace *ns = container_of(pl, struct ldlm_namespace,
558 struct dentry *debugfs_ns_parent;
559 struct lprocfs_vars pool_vars[2];
560 char *var_name = NULL;
563 var_name = kzalloc(MAX_STRING_SIZE + 1, GFP_NOFS);
567 debugfs_ns_parent = ns->ns_debugfs_entry;
568 if (IS_ERR_OR_NULL(debugfs_ns_parent)) {
569 CERROR("%s: debugfs entry is not initialized\n",
574 pl->pl_debugfs_entry = ldebugfs_register("pool", debugfs_ns_parent,
576 if (IS_ERR(pl->pl_debugfs_entry)) {
577 CERROR("LdebugFS failed in ldlm-pool-init\n");
578 rc = PTR_ERR(pl->pl_debugfs_entry);
579 pl->pl_debugfs_entry = NULL;
583 var_name[MAX_STRING_SIZE] = '\0';
584 memset(pool_vars, 0, sizeof(pool_vars));
585 pool_vars[0].name = var_name;
587 LDLM_POOL_ADD_VAR(state, pl, &lprocfs_pool_state_fops);
589 pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT -
590 LDLM_POOL_FIRST_STAT, 0);
596 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
597 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
599 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
600 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
602 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
603 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
605 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
606 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
607 "grant_rate", "locks/s");
608 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
609 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
610 "cancel_rate", "locks/s");
611 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
612 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
613 "grant_plan", "locks/s");
614 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT,
615 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
617 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT,
618 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
619 "shrink_request", "locks");
620 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT,
621 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
622 "shrink_freed", "locks");
623 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
624 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
625 "recalc_freed", "locks");
626 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
627 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
628 "recalc_timing", "sec");
629 rc = ldebugfs_register_stats(pl->pl_debugfs_entry, "stats",
637 static void ldlm_pool_sysfs_fini(struct ldlm_pool *pl)
639 kobject_put(&pl->pl_kobj);
640 wait_for_completion(&pl->pl_kobj_unregister);
643 static void ldlm_pool_debugfs_fini(struct ldlm_pool *pl)
646 lprocfs_free_stats(&pl->pl_stats);
649 if (pl->pl_debugfs_entry) {
650 ldebugfs_remove(&pl->pl_debugfs_entry);
651 pl->pl_debugfs_entry = NULL;
655 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
656 int idx, ldlm_side_t client)
660 spin_lock_init(&pl->pl_lock);
661 atomic_set(&pl->pl_granted, 0);
662 pl->pl_recalc_time = ktime_get_seconds();
663 atomic_set(&pl->pl_lock_volume_factor, 1);
665 atomic_set(&pl->pl_grant_rate, 0);
666 atomic_set(&pl->pl_cancel_rate, 0);
667 pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
669 snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
670 ldlm_ns_name(ns), idx);
672 atomic_set(&pl->pl_limit, 1);
673 pl->pl_server_lock_volume = 0;
674 pl->pl_ops = &ldlm_cli_pool_ops;
675 pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
676 pl->pl_client_lock_volume = 0;
677 rc = ldlm_pool_debugfs_init(pl);
681 rc = ldlm_pool_sysfs_init(pl);
685 CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name);
689 EXPORT_SYMBOL(ldlm_pool_init);
691 void ldlm_pool_fini(struct ldlm_pool *pl)
693 ldlm_pool_sysfs_fini(pl);
694 ldlm_pool_debugfs_fini(pl);
697 * Pool should not be used after this point. We can't free it here as
698 * it lives in struct ldlm_namespace, but still interested in catching
699 * any abnormal using cases.
701 POISON(pl, 0x5a, sizeof(*pl));
703 EXPORT_SYMBOL(ldlm_pool_fini);
706 * Add new taken ldlm lock \a lock into pool \a pl accounting.
708 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
711 * FLOCK locks are special in a sense that they are almost never
712 * cancelled, instead special kind of lock is used to drop them.
713 * also there is no LRU for flock locks, so no point in tracking
716 if (lock->l_resource->lr_type == LDLM_FLOCK)
719 atomic_inc(&pl->pl_granted);
720 atomic_inc(&pl->pl_grant_rate);
721 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
723 * Do not do pool recalc for client side as all locks which
724 * potentially may be canceled has already been packed into
725 * enqueue/cancel rpc. Also we do not want to run out of stack
726 * with too long call paths.
729 EXPORT_SYMBOL(ldlm_pool_add);
732 * Remove ldlm lock \a lock from pool \a pl accounting.
734 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
737 * Filter out FLOCK locks. Read above comment in ldlm_pool_add().
739 if (lock->l_resource->lr_type == LDLM_FLOCK)
742 LASSERT(atomic_read(&pl->pl_granted) > 0);
743 atomic_dec(&pl->pl_granted);
744 atomic_inc(&pl->pl_cancel_rate);
746 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
748 EXPORT_SYMBOL(ldlm_pool_del);
751 * Returns current \a pl SLV.
753 * \pre ->pl_lock is not locked.
755 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
759 spin_lock(&pl->pl_lock);
760 slv = pl->pl_server_lock_volume;
761 spin_unlock(&pl->pl_lock);
766 * Sets passed \a clv to \a pl.
768 * \pre ->pl_lock is not locked.
770 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
772 spin_lock(&pl->pl_lock);
773 pl->pl_client_lock_volume = clv;
774 spin_unlock(&pl->pl_lock);
778 * Returns current LVF from \a pl.
780 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
782 return atomic_read(&pl->pl_lock_volume_factor);
785 static int ldlm_pool_granted(struct ldlm_pool *pl)
787 return atomic_read(&pl->pl_granted);
790 static struct ptlrpc_thread *ldlm_pools_thread;
791 static struct completion ldlm_pools_comp;
794 * count locks from all namespaces (if possible). Returns number of
797 static unsigned long ldlm_pools_count(ldlm_side_t client, gfp_t gfp_mask)
799 int total = 0, nr_ns;
800 struct ldlm_namespace *ns;
801 struct ldlm_namespace *ns_old = NULL; /* loop detection */
804 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
807 CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n",
808 client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
810 cookie = cl_env_reenter();
813 * Find out how many resources we may release.
815 for (nr_ns = ldlm_namespace_nr_read(client);
816 nr_ns > 0; nr_ns--) {
817 mutex_lock(ldlm_namespace_lock(client));
818 if (list_empty(ldlm_namespace_list(client))) {
819 mutex_unlock(ldlm_namespace_lock(client));
820 cl_env_reexit(cookie);
823 ns = ldlm_namespace_first_locked(client);
826 mutex_unlock(ldlm_namespace_lock(client));
830 if (ldlm_ns_empty(ns)) {
831 ldlm_namespace_move_to_inactive_locked(ns, client);
832 mutex_unlock(ldlm_namespace_lock(client));
839 ldlm_namespace_get(ns);
840 ldlm_namespace_move_to_active_locked(ns, client);
841 mutex_unlock(ldlm_namespace_lock(client));
842 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
843 ldlm_namespace_put(ns);
846 cl_env_reexit(cookie);
850 static unsigned long ldlm_pools_scan(ldlm_side_t client, int nr, gfp_t gfp_mask)
852 unsigned long freed = 0;
854 struct ldlm_namespace *ns;
857 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
860 cookie = cl_env_reenter();
863 * Shrink at least ldlm_namespace_nr_read(client) namespaces.
865 for (tmp = nr_ns = ldlm_namespace_nr_read(client);
867 int cancel, nr_locks;
870 * Do not call shrink under ldlm_namespace_lock(client)
872 mutex_lock(ldlm_namespace_lock(client));
873 if (list_empty(ldlm_namespace_list(client))) {
874 mutex_unlock(ldlm_namespace_lock(client));
877 ns = ldlm_namespace_first_locked(client);
878 ldlm_namespace_get(ns);
879 ldlm_namespace_move_to_active_locked(ns, client);
880 mutex_unlock(ldlm_namespace_lock(client));
882 nr_locks = ldlm_pool_granted(&ns->ns_pool);
884 * We use to shrink propotionally but with new shrinker API,
885 * we lost the total number of freeable locks.
887 cancel = 1 + min_t(int, nr_locks, nr / nr_ns);
888 freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
889 ldlm_namespace_put(ns);
891 cl_env_reexit(cookie);
893 * we only decrease the SLV in server pools shrinker, return
894 * SHRINK_STOP to kernel to avoid needless loop. LU-1128
899 static unsigned long ldlm_pools_cli_count(struct shrinker *s,
900 struct shrink_control *sc)
902 return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask);
905 static unsigned long ldlm_pools_cli_scan(struct shrinker *s,
906 struct shrink_control *sc)
908 return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan,
912 static int ldlm_pools_recalc(ldlm_side_t client)
914 struct ldlm_namespace *ns;
915 struct ldlm_namespace *ns_old = NULL;
917 int time = 50; /* seconds of sleep if no active namespaces */
920 * Recalc at least ldlm_namespace_nr_read(client) namespaces.
922 for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) {
925 * Lock the list, get first @ns in the list, getref, move it
926 * to the tail, unlock and call pool recalc. This way we avoid
927 * calling recalc under @ns lock what is really good as we get
928 * rid of potential deadlock on client nodes when canceling
929 * locks synchronously.
931 mutex_lock(ldlm_namespace_lock(client));
932 if (list_empty(ldlm_namespace_list(client))) {
933 mutex_unlock(ldlm_namespace_lock(client));
936 ns = ldlm_namespace_first_locked(client);
938 if (ns_old == ns) { /* Full pass complete */
939 mutex_unlock(ldlm_namespace_lock(client));
943 /* We got an empty namespace, need to move it back to inactive
945 * The race with parallel resource creation is fine:
946 * - If they do namespace_get before our check, we fail the
947 * check and they move this item to the end of the list anyway
948 * - If we do the check and then they do namespace_get, then
949 * we move the namespace to inactive and they will move
950 * it back to active (synchronised by the lock, so no clash
953 if (ldlm_ns_empty(ns)) {
954 ldlm_namespace_move_to_inactive_locked(ns, client);
955 mutex_unlock(ldlm_namespace_lock(client));
962 spin_lock(&ns->ns_lock);
964 * skip ns which is being freed, and we don't want to increase
965 * its refcount again, not even temporarily. bz21519 & LU-499.
967 if (ns->ns_stopping) {
971 ldlm_namespace_get(ns);
973 spin_unlock(&ns->ns_lock);
975 ldlm_namespace_move_to_active_locked(ns, client);
976 mutex_unlock(ldlm_namespace_lock(client));
979 * After setup is done - recalc the pool.
982 int ttime = ldlm_pool_recalc(&ns->ns_pool);
987 ldlm_namespace_put(ns);
993 static int ldlm_pools_thread_main(void *arg)
995 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
998 thread_set_flags(thread, SVC_RUNNING);
999 wake_up(&thread->t_ctl_waitq);
1001 CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n",
1002 "ldlm_poold", current_pid());
1005 struct l_wait_info lwi;
1008 * Recal all pools on this tick.
1010 c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
1013 * Wait until the next check time, or until we're
1016 lwi = LWI_TIMEOUT(cfs_time_seconds(c_time),
1018 l_wait_event(thread->t_ctl_waitq,
1019 thread_is_stopping(thread) ||
1020 thread_is_event(thread),
1023 if (thread_test_and_clear_flags(thread, SVC_STOPPING))
1025 thread_test_and_clear_flags(thread, SVC_EVENT);
1028 thread_set_flags(thread, SVC_STOPPED);
1029 wake_up(&thread->t_ctl_waitq);
1031 CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n",
1032 "ldlm_poold", current_pid());
1034 complete_and_exit(&ldlm_pools_comp, 0);
1037 static int ldlm_pools_thread_start(void)
1039 struct l_wait_info lwi = { 0 };
1040 struct task_struct *task;
1042 if (ldlm_pools_thread)
1045 ldlm_pools_thread = kzalloc(sizeof(*ldlm_pools_thread), GFP_NOFS);
1046 if (!ldlm_pools_thread)
1049 init_completion(&ldlm_pools_comp);
1050 init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq);
1052 task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread,
1055 CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task));
1056 kfree(ldlm_pools_thread);
1057 ldlm_pools_thread = NULL;
1058 return PTR_ERR(task);
1060 l_wait_event(ldlm_pools_thread->t_ctl_waitq,
1061 thread_is_running(ldlm_pools_thread), &lwi);
1065 static void ldlm_pools_thread_stop(void)
1067 if (!ldlm_pools_thread)
1070 thread_set_flags(ldlm_pools_thread, SVC_STOPPING);
1071 wake_up(&ldlm_pools_thread->t_ctl_waitq);
1074 * Make sure that pools thread is finished before freeing @thread.
1075 * This fixes possible race and oops due to accessing freed memory
1078 wait_for_completion(&ldlm_pools_comp);
1079 kfree(ldlm_pools_thread);
1080 ldlm_pools_thread = NULL;
1083 static struct shrinker ldlm_pools_cli_shrinker = {
1084 .count_objects = ldlm_pools_cli_count,
1085 .scan_objects = ldlm_pools_cli_scan,
1086 .seeks = DEFAULT_SEEKS,
1089 int ldlm_pools_init(void)
1093 rc = ldlm_pools_thread_start();
1095 register_shrinker(&ldlm_pools_cli_shrinker);
1099 EXPORT_SYMBOL(ldlm_pools_init);
1101 void ldlm_pools_fini(void)
1103 if (ldlm_pools_thread)
1104 unregister_shrinker(&ldlm_pools_cli_shrinker);
1106 ldlm_pools_thread_stop();
1108 EXPORT_SYMBOL(ldlm_pools_fini);