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 struct proc_dir_entry *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;
339 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
340 if (recalc_interval_sec < pl->pl_recalc_period)
343 spin_lock(&pl->pl_lock);
344 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
345 if (recalc_interval_sec < pl->pl_recalc_period) {
346 spin_unlock(&pl->pl_lock);
350 * Recalc SLV after last period. This should be done
351 * _before_ recalculating new grant plan.
353 ldlm_pool_recalc_slv(pl);
356 * Make sure that pool informed obd of last SLV changes.
358 ldlm_srv_pool_push_slv(pl);
361 * Update grant_plan for new period.
363 ldlm_pool_recalc_grant_plan(pl);
365 pl->pl_recalc_time = cfs_time_current_sec();
366 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
367 recalc_interval_sec);
368 spin_unlock(&pl->pl_lock);
373 * This function is used on server side as main entry point for memory
374 * pressure handling. It decreases SLV on \a pl according to passed
375 * \a nr and \a gfp_mask.
377 * Our goal here is to decrease SLV such a way that clients hold \a nr
378 * locks smaller in next 10h.
380 static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
381 int nr, unsigned int gfp_mask)
386 * VM is asking how many entries may be potentially freed.
389 return atomic_read(&pl->pl_granted);
392 * Client already canceled locks but server is already in shrinker
393 * and can't cancel anything. Let's catch this race.
395 if (atomic_read(&pl->pl_granted) == 0)
398 spin_lock(&pl->pl_lock);
401 * We want shrinker to possibly cause cancellation of @nr locks from
402 * clients or grant approximately @nr locks smaller next intervals.
404 * This is why we decreased SLV by @nr. This effect will only be as
405 * long as one re-calc interval (1s these days) and this should be
406 * enough to pass this decreased SLV to all clients. On next recalc
407 * interval pool will either increase SLV if locks load is not high
408 * or will keep on same level or even decrease again, thus, shrinker
409 * decreased SLV will affect next recalc intervals and this way will
410 * make locking load lower.
412 if (nr < pl->pl_server_lock_volume) {
413 pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr;
415 limit = ldlm_pool_get_limit(pl);
416 pl->pl_server_lock_volume = ldlm_pool_slv_min(limit);
420 * Make sure that pool informed obd of last SLV changes.
422 ldlm_srv_pool_push_slv(pl);
423 spin_unlock(&pl->pl_lock);
426 * We did not really free any memory here so far, it only will be
427 * freed later may be, so that we return 0 to not confuse VM.
433 * Setup server side pool \a pl with passed \a limit.
435 static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit)
437 struct obd_device *obd;
439 obd = ldlm_pl2ns(pl)->ns_obd;
440 LASSERT(obd != NULL && obd != LP_POISON);
441 LASSERT(obd->obd_type != LP_POISON);
442 write_lock(&obd->obd_pool_lock);
443 obd->obd_pool_limit = limit;
444 write_unlock(&obd->obd_pool_lock);
446 ldlm_pool_set_limit(pl, limit);
451 * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl.
453 static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
455 struct obd_device *obd;
458 * Get new SLV and Limit from obd which is updated with coming
461 obd = ldlm_pl2ns(pl)->ns_obd;
462 LASSERT(obd != NULL);
463 read_lock(&obd->obd_pool_lock);
464 pl->pl_server_lock_volume = obd->obd_pool_slv;
465 ldlm_pool_set_limit(pl, obd->obd_pool_limit);
466 read_unlock(&obd->obd_pool_lock);
470 * Recalculates client size pool \a pl according to current SLV and Limit.
472 static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
474 time_t recalc_interval_sec;
476 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
477 if (recalc_interval_sec < pl->pl_recalc_period)
480 spin_lock(&pl->pl_lock);
482 * Check if we need to recalc lists now.
484 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
485 if (recalc_interval_sec < pl->pl_recalc_period) {
486 spin_unlock(&pl->pl_lock);
491 * Make sure that pool knows last SLV and Limit from obd.
493 ldlm_cli_pool_pop_slv(pl);
495 pl->pl_recalc_time = cfs_time_current_sec();
496 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
497 recalc_interval_sec);
498 spin_unlock(&pl->pl_lock);
501 * Do not cancel locks in case lru resize is disabled for this ns.
503 if (!ns_connect_lru_resize(ldlm_pl2ns(pl)))
507 * In the time of canceling locks on client we do not need to maintain
508 * sharp timing, we only want to cancel locks asap according to new SLV.
509 * It may be called when SLV has changed much, this is why we do not
510 * take into account pl->pl_recalc_time here.
512 return ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC, LDLM_CANCEL_LRUR);
516 * This function is main entry point for memory pressure handling on client
517 * side. Main goal of this function is to cancel some number of locks on
518 * passed \a pl according to \a nr and \a gfp_mask.
520 static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
521 int nr, unsigned int gfp_mask)
523 struct ldlm_namespace *ns;
529 * Do not cancel locks in case lru resize is disabled for this ns.
531 if (!ns_connect_lru_resize(ns))
535 * Make sure that pool knows last SLV and Limit from obd.
537 ldlm_cli_pool_pop_slv(pl);
539 spin_lock(&ns->ns_lock);
540 unused = ns->ns_nr_unused;
541 spin_unlock(&ns->ns_lock);
544 return (unused / 100) * sysctl_vfs_cache_pressure;
546 return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_CANCEL_SHRINK);
549 struct ldlm_pool_ops ldlm_srv_pool_ops = {
550 .po_recalc = ldlm_srv_pool_recalc,
551 .po_shrink = ldlm_srv_pool_shrink,
552 .po_setup = ldlm_srv_pool_setup
555 struct ldlm_pool_ops ldlm_cli_pool_ops = {
556 .po_recalc = ldlm_cli_pool_recalc,
557 .po_shrink = ldlm_cli_pool_shrink
561 * Pool recalc wrapper. Will call either client or server pool recalc callback
562 * depending what pool \a pl is used.
564 int ldlm_pool_recalc(struct ldlm_pool *pl)
566 time_t recalc_interval_sec;
569 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
570 if (recalc_interval_sec <= 0)
573 spin_lock(&pl->pl_lock);
574 if (recalc_interval_sec > 0) {
576 * Update pool statistics every 1s.
578 ldlm_pool_recalc_stats(pl);
581 * Zero out all rates and speed for the last period.
583 atomic_set(&pl->pl_grant_rate, 0);
584 atomic_set(&pl->pl_cancel_rate, 0);
586 spin_unlock(&pl->pl_lock);
589 if (pl->pl_ops->po_recalc != NULL) {
590 count = pl->pl_ops->po_recalc(pl);
591 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
594 recalc_interval_sec = pl->pl_recalc_time - cfs_time_current_sec() +
595 pl->pl_recalc_period;
597 return recalc_interval_sec;
601 * Pool shrink wrapper. Will call either client or server pool recalc callback
602 * depending what pool pl is used. When nr == 0, just return the number of
603 * freeable locks. Otherwise, return the number of canceled locks.
605 int ldlm_pool_shrink(struct ldlm_pool *pl, int nr,
606 unsigned int gfp_mask)
610 if (pl->pl_ops->po_shrink != NULL) {
611 cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask);
613 lprocfs_counter_add(pl->pl_stats,
614 LDLM_POOL_SHRINK_REQTD_STAT,
616 lprocfs_counter_add(pl->pl_stats,
617 LDLM_POOL_SHRINK_FREED_STAT,
619 CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, "
620 "shrunk %d\n", pl->pl_name, nr, cancel);
625 EXPORT_SYMBOL(ldlm_pool_shrink);
628 * Pool setup wrapper. Will call either client or server pool recalc callback
629 * depending what pool \a pl is used.
631 * Sets passed \a limit into pool \a pl.
633 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
635 if (pl->pl_ops->po_setup != NULL)
636 return(pl->pl_ops->po_setup(pl, limit));
639 EXPORT_SYMBOL(ldlm_pool_setup);
642 static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
644 int granted, grant_rate, cancel_rate, grant_step;
645 int grant_speed, grant_plan, lvf;
646 struct ldlm_pool *pl = m->private;
650 spin_lock(&pl->pl_lock);
651 slv = pl->pl_server_lock_volume;
652 clv = pl->pl_client_lock_volume;
653 limit = ldlm_pool_get_limit(pl);
654 grant_plan = pl->pl_grant_plan;
655 granted = atomic_read(&pl->pl_granted);
656 grant_rate = atomic_read(&pl->pl_grant_rate);
657 cancel_rate = atomic_read(&pl->pl_cancel_rate);
658 grant_speed = grant_rate - cancel_rate;
659 lvf = atomic_read(&pl->pl_lock_volume_factor);
660 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
661 spin_unlock(&pl->pl_lock);
663 seq_printf(m, "LDLM pool state (%s):\n"
667 pl->pl_name, slv, clv, lvf);
669 if (ns_is_server(ldlm_pl2ns(pl))) {
670 seq_printf(m, " GSP: %d%%\n"
672 grant_step, grant_plan);
674 seq_printf(m, " GR: %d\n" " CR: %d\n" " GS: %d\n"
675 " G: %d\n" " L: %d\n",
676 grant_rate, cancel_rate, grant_speed,
681 LPROC_SEQ_FOPS_RO(lprocfs_pool_state);
683 static int lprocfs_grant_speed_seq_show(struct seq_file *m, void *unused)
685 struct ldlm_pool *pl = m->private;
688 spin_lock(&pl->pl_lock);
689 /* serialize with ldlm_pool_recalc */
690 grant_speed = atomic_read(&pl->pl_grant_rate) -
691 atomic_read(&pl->pl_cancel_rate);
692 spin_unlock(&pl->pl_lock);
693 return lprocfs_rd_uint(m, &grant_speed);
696 LDLM_POOL_PROC_READER_SEQ_SHOW(grant_plan, int);
697 LPROC_SEQ_FOPS_RO(lprocfs_grant_plan);
699 LDLM_POOL_PROC_READER_SEQ_SHOW(recalc_period, int);
700 LDLM_POOL_PROC_WRITER(recalc_period, int);
701 static ssize_t lprocfs_recalc_period_seq_write(struct file *file, const char *buf,
702 size_t len, loff_t *off)
704 struct seq_file *seq = file->private_data;
706 return lprocfs_wr_recalc_period(file, buf, len, seq->private);
708 LPROC_SEQ_FOPS(lprocfs_recalc_period);
710 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, u64);
711 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, atomic);
712 LPROC_SEQ_FOPS_RW_TYPE(ldlm_pool_rw, atomic);
714 LPROC_SEQ_FOPS_RO(lprocfs_grant_speed);
716 #define LDLM_POOL_ADD_VAR(name, var, ops) \
718 snprintf(var_name, MAX_STRING_SIZE, #name); \
719 pool_vars[0].data = var; \
720 pool_vars[0].fops = ops; \
721 lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);\
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;
732 OBD_ALLOC(var_name, MAX_STRING_SIZE + 1);
736 parent_ns_proc = ns->ns_proc_dir_entry;
737 if (parent_ns_proc == NULL) {
738 CERROR("%s: proc entry is not initialized\n",
740 GOTO(out_free_name, rc = -EINVAL);
742 pl->pl_proc_dir = lprocfs_register("pool", parent_ns_proc,
744 if (IS_ERR(pl->pl_proc_dir)) {
745 CERROR("LProcFS failed in ldlm-pool-init\n");
746 rc = PTR_ERR(pl->pl_proc_dir);
747 pl->pl_proc_dir = NULL;
748 GOTO(out_free_name, rc);
751 var_name[MAX_STRING_SIZE] = '\0';
752 memset(pool_vars, 0, sizeof(pool_vars));
753 pool_vars[0].name = var_name;
755 LDLM_POOL_ADD_VAR("server_lock_volume", &pl->pl_server_lock_volume,
756 &ldlm_pool_u64_fops);
757 LDLM_POOL_ADD_VAR("limit", &pl->pl_limit, &ldlm_pool_rw_atomic_fops);
758 LDLM_POOL_ADD_VAR("granted", &pl->pl_granted, &ldlm_pool_atomic_fops);
759 LDLM_POOL_ADD_VAR("grant_speed", pl, &lprocfs_grant_speed_fops);
760 LDLM_POOL_ADD_VAR("cancel_rate", &pl->pl_cancel_rate,
761 &ldlm_pool_atomic_fops);
762 LDLM_POOL_ADD_VAR("grant_rate", &pl->pl_grant_rate,
763 &ldlm_pool_atomic_fops);
764 LDLM_POOL_ADD_VAR("grant_plan", pl, &lprocfs_grant_plan_fops);
765 LDLM_POOL_ADD_VAR("recalc_period", pl, &lprocfs_recalc_period_fops);
766 LDLM_POOL_ADD_VAR("lock_volume_factor", &pl->pl_lock_volume_factor,
767 &ldlm_pool_rw_atomic_fops);
768 LDLM_POOL_ADD_VAR("state", pl, &lprocfs_pool_state_fops);
770 pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT -
771 LDLM_POOL_FIRST_STAT, 0);
773 GOTO(out_free_name, rc = -ENOMEM);
775 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
776 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
778 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
779 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
781 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
782 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
784 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
785 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
786 "grant_rate", "locks/s");
787 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
788 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
789 "cancel_rate", "locks/s");
790 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
791 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
792 "grant_plan", "locks/s");
793 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT,
794 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
796 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT,
797 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
798 "shrink_request", "locks");
799 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT,
800 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
801 "shrink_freed", "locks");
802 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
803 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
804 "recalc_freed", "locks");
805 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
806 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
807 "recalc_timing", "sec");
808 rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
811 OBD_FREE(var_name, MAX_STRING_SIZE + 1);
815 static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
817 if (pl->pl_stats != NULL) {
818 lprocfs_free_stats(&pl->pl_stats);
821 if (pl->pl_proc_dir != NULL) {
822 lprocfs_remove(&pl->pl_proc_dir);
823 pl->pl_proc_dir = NULL;
827 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
832 static void ldlm_pool_proc_fini(struct ldlm_pool *pl) {}
835 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
836 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)
876 ldlm_pool_proc_fini(pl);
879 * Pool should not be used after this point. We can't free it here as
880 * it lives in struct ldlm_namespace, but still interested in catching
881 * any abnormal using cases.
883 POISON(pl, 0x5a, sizeof(*pl));
885 EXPORT_SYMBOL(ldlm_pool_fini);
888 * Add new taken ldlm lock \a lock into pool \a pl accounting.
890 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
893 * FLOCK locks are special in a sense that they are almost never
894 * cancelled, instead special kind of lock is used to drop them.
895 * also there is no LRU for flock locks, so no point in tracking
898 if (lock->l_resource->lr_type == LDLM_FLOCK)
901 atomic_inc(&pl->pl_granted);
902 atomic_inc(&pl->pl_grant_rate);
903 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
905 * Do not do pool recalc for client side as all locks which
906 * potentially may be canceled has already been packed into
907 * enqueue/cancel rpc. Also we do not want to run out of stack
908 * with too long call paths.
910 if (ns_is_server(ldlm_pl2ns(pl)))
911 ldlm_pool_recalc(pl);
913 EXPORT_SYMBOL(ldlm_pool_add);
916 * Remove ldlm lock \a lock from pool \a pl accounting.
918 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
921 * Filter out FLOCK locks. Read above comment in ldlm_pool_add().
923 if (lock->l_resource->lr_type == LDLM_FLOCK)
926 LASSERT(atomic_read(&pl->pl_granted) > 0);
927 atomic_dec(&pl->pl_granted);
928 atomic_inc(&pl->pl_cancel_rate);
930 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
932 if (ns_is_server(ldlm_pl2ns(pl)))
933 ldlm_pool_recalc(pl);
935 EXPORT_SYMBOL(ldlm_pool_del);
938 * Returns current \a pl SLV.
940 * \pre ->pl_lock is not locked.
942 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
945 spin_lock(&pl->pl_lock);
946 slv = pl->pl_server_lock_volume;
947 spin_unlock(&pl->pl_lock);
950 EXPORT_SYMBOL(ldlm_pool_get_slv);
953 * Sets passed \a slv to \a pl.
955 * \pre ->pl_lock is not locked.
957 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
959 spin_lock(&pl->pl_lock);
960 pl->pl_server_lock_volume = slv;
961 spin_unlock(&pl->pl_lock);
963 EXPORT_SYMBOL(ldlm_pool_set_slv);
966 * Returns current \a pl CLV.
968 * \pre ->pl_lock is not locked.
970 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
973 spin_lock(&pl->pl_lock);
974 slv = pl->pl_client_lock_volume;
975 spin_unlock(&pl->pl_lock);
978 EXPORT_SYMBOL(ldlm_pool_get_clv);
981 * Sets passed \a clv to \a pl.
983 * \pre ->pl_lock is not locked.
985 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
987 spin_lock(&pl->pl_lock);
988 pl->pl_client_lock_volume = clv;
989 spin_unlock(&pl->pl_lock);
991 EXPORT_SYMBOL(ldlm_pool_set_clv);
994 * Returns current \a pl limit.
996 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
998 return atomic_read(&pl->pl_limit);
1000 EXPORT_SYMBOL(ldlm_pool_get_limit);
1003 * Sets passed \a limit to \a pl.
1005 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1007 atomic_set(&pl->pl_limit, limit);
1009 EXPORT_SYMBOL(ldlm_pool_set_limit);
1012 * Returns current LVF from \a pl.
1014 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1016 return atomic_read(&pl->pl_lock_volume_factor);
1018 EXPORT_SYMBOL(ldlm_pool_get_lvf);
1020 static int ldlm_pool_granted(struct ldlm_pool *pl)
1022 return atomic_read(&pl->pl_granted);
1025 static struct ptlrpc_thread *ldlm_pools_thread;
1026 static struct completion ldlm_pools_comp;
1029 * count locks from all namespaces (if possible). Returns number of
1032 static unsigned long ldlm_pools_count(ldlm_side_t client, unsigned int gfp_mask)
1034 int total = 0, nr_ns;
1035 struct ldlm_namespace *ns;
1036 struct ldlm_namespace *ns_old = NULL; /* loop detection */
1039 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1042 CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n",
1043 client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1045 cookie = cl_env_reenter();
1048 * Find out how many resources we may release.
1050 for (nr_ns = ldlm_namespace_nr_read(client);
1051 nr_ns > 0; nr_ns--) {
1052 mutex_lock(ldlm_namespace_lock(client));
1053 if (list_empty(ldlm_namespace_list(client))) {
1054 mutex_unlock(ldlm_namespace_lock(client));
1055 cl_env_reexit(cookie);
1058 ns = ldlm_namespace_first_locked(client);
1061 mutex_unlock(ldlm_namespace_lock(client));
1065 if (ldlm_ns_empty(ns)) {
1066 ldlm_namespace_move_to_inactive_locked(ns, client);
1067 mutex_unlock(ldlm_namespace_lock(client));
1074 ldlm_namespace_get(ns);
1075 ldlm_namespace_move_to_active_locked(ns, client);
1076 mutex_unlock(ldlm_namespace_lock(client));
1077 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
1078 ldlm_namespace_put(ns);
1081 cl_env_reexit(cookie);
1085 static unsigned long ldlm_pools_scan(ldlm_side_t client, int nr, unsigned int gfp_mask)
1087 unsigned long freed = 0;
1089 struct ldlm_namespace *ns;
1092 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1095 cookie = cl_env_reenter();
1098 * Shrink at least ldlm_namespace_nr_read(client) namespaces.
1100 for (tmp = nr_ns = ldlm_namespace_nr_read(client);
1102 int cancel, nr_locks;
1105 * Do not call shrink under ldlm_namespace_lock(client)
1107 mutex_lock(ldlm_namespace_lock(client));
1108 if (list_empty(ldlm_namespace_list(client))) {
1109 mutex_unlock(ldlm_namespace_lock(client));
1112 ns = ldlm_namespace_first_locked(client);
1113 ldlm_namespace_get(ns);
1114 ldlm_namespace_move_to_active_locked(ns, client);
1115 mutex_unlock(ldlm_namespace_lock(client));
1117 nr_locks = ldlm_pool_granted(&ns->ns_pool);
1119 * We use to shrink propotionally but with new shrinker API,
1120 * we lost the total number of freeable locks.
1122 cancel = 1 + min_t(int, nr_locks, nr / nr_ns);
1123 freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
1124 ldlm_namespace_put(ns);
1126 cl_env_reexit(cookie);
1128 * we only decrease the SLV in server pools shrinker, return
1129 * SHRINK_STOP to kernel to avoid needless loop. LU-1128
1131 return (client == LDLM_NAMESPACE_SERVER) ? SHRINK_STOP : freed;
1134 static unsigned long ldlm_pools_srv_count(struct shrinker *s, struct shrink_control *sc)
1136 return ldlm_pools_count(LDLM_NAMESPACE_SERVER, sc->gfp_mask);
1139 static unsigned long ldlm_pools_srv_scan(struct shrinker *s, struct shrink_control *sc)
1141 return ldlm_pools_scan(LDLM_NAMESPACE_SERVER, sc->nr_to_scan,
1145 static unsigned long ldlm_pools_cli_count(struct shrinker *s, struct shrink_control *sc)
1147 return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask);
1150 static unsigned long ldlm_pools_cli_scan(struct shrinker *s, struct shrink_control *sc)
1152 return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan,
1156 int ldlm_pools_recalc(ldlm_side_t client)
1158 __u32 nr_l = 0, nr_p = 0, l;
1159 struct ldlm_namespace *ns;
1160 struct ldlm_namespace *ns_old = NULL;
1162 int time = 50; /* seconds of sleep if no active namespaces */
1165 * No need to setup pool limit for client pools.
1167 if (client == LDLM_NAMESPACE_SERVER) {
1169 * Check all modest namespaces first.
1171 mutex_lock(ldlm_namespace_lock(client));
1172 list_for_each_entry(ns, ldlm_namespace_list(client),
1175 if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
1178 l = ldlm_pool_granted(&ns->ns_pool);
1183 * Set the modest pools limit equal to their avg granted
1186 l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0);
1187 ldlm_pool_setup(&ns->ns_pool, l);
1193 * Make sure that modest namespaces did not eat more that 2/3
1196 if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) {
1197 CWARN("\"Modest\" pools eat out 2/3 of server locks "
1198 "limit (%d of %lu). This means that you have too "
1199 "many clients for this amount of server RAM. "
1200 "Upgrade server!\n", nr_l, LDLM_POOL_HOST_L);
1205 * The rest is given to greedy namespaces.
1207 list_for_each_entry(ns, ldlm_namespace_list(client),
1210 if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
1215 * In the case 2/3 locks are eaten out by
1216 * modest pools, we re-setup equal limit
1219 l = LDLM_POOL_HOST_L /
1220 ldlm_namespace_nr_read(client);
1223 * All the rest of greedy pools will have
1224 * all locks in equal parts.
1226 l = (LDLM_POOL_HOST_L - nr_l) /
1227 (ldlm_namespace_nr_read(client) -
1230 ldlm_pool_setup(&ns->ns_pool, l);
1232 mutex_unlock(ldlm_namespace_lock(client));
1236 * Recalc at least ldlm_namespace_nr_read(client) namespaces.
1238 for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) {
1241 * Lock the list, get first @ns in the list, getref, move it
1242 * to the tail, unlock and call pool recalc. This way we avoid
1243 * calling recalc under @ns lock what is really good as we get
1244 * rid of potential deadlock on client nodes when canceling
1245 * locks synchronously.
1247 mutex_lock(ldlm_namespace_lock(client));
1248 if (list_empty(ldlm_namespace_list(client))) {
1249 mutex_unlock(ldlm_namespace_lock(client));
1252 ns = ldlm_namespace_first_locked(client);
1254 if (ns_old == ns) { /* Full pass complete */
1255 mutex_unlock(ldlm_namespace_lock(client));
1259 /* We got an empty namespace, need to move it back to inactive
1261 * The race with parallel resource creation is fine:
1262 * - If they do namespace_get before our check, we fail the
1263 * check and they move this item to the end of the list anyway
1264 * - If we do the check and then they do namespace_get, then
1265 * we move the namespace to inactive and they will move
1266 * it back to active (synchronised by the lock, so no clash
1269 if (ldlm_ns_empty(ns)) {
1270 ldlm_namespace_move_to_inactive_locked(ns, client);
1271 mutex_unlock(ldlm_namespace_lock(client));
1278 spin_lock(&ns->ns_lock);
1280 * skip ns which is being freed, and we don't want to increase
1281 * its refcount again, not even temporarily. bz21519 & LU-499.
1283 if (ns->ns_stopping) {
1287 ldlm_namespace_get(ns);
1289 spin_unlock(&ns->ns_lock);
1291 ldlm_namespace_move_to_active_locked(ns, client);
1292 mutex_unlock(ldlm_namespace_lock(client));
1295 * After setup is done - recalc the pool.
1298 int ttime = ldlm_pool_recalc(&ns->ns_pool);
1303 ldlm_namespace_put(ns);
1308 EXPORT_SYMBOL(ldlm_pools_recalc);
1310 static int ldlm_pools_thread_main(void *arg)
1312 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
1315 thread_set_flags(thread, SVC_RUNNING);
1316 wake_up(&thread->t_ctl_waitq);
1318 CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n",
1319 "ldlm_poold", current_pid());
1322 struct l_wait_info lwi;
1325 * Recal all pools on this tick.
1327 s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
1328 c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
1331 * Wait until the next check time, or until we're
1334 lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)),
1336 l_wait_event(thread->t_ctl_waitq,
1337 thread_is_stopping(thread) ||
1338 thread_is_event(thread),
1341 if (thread_test_and_clear_flags(thread, SVC_STOPPING))
1344 thread_test_and_clear_flags(thread, SVC_EVENT);
1347 thread_set_flags(thread, SVC_STOPPED);
1348 wake_up(&thread->t_ctl_waitq);
1350 CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n",
1351 "ldlm_poold", current_pid());
1353 complete_and_exit(&ldlm_pools_comp, 0);
1356 static int ldlm_pools_thread_start(void)
1358 struct l_wait_info lwi = { 0 };
1359 struct task_struct *task;
1361 if (ldlm_pools_thread != NULL)
1364 OBD_ALLOC_PTR(ldlm_pools_thread);
1365 if (ldlm_pools_thread == NULL)
1368 init_completion(&ldlm_pools_comp);
1369 init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq);
1371 task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread,
1374 CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task));
1375 OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread));
1376 ldlm_pools_thread = NULL;
1377 return PTR_ERR(task);
1379 l_wait_event(ldlm_pools_thread->t_ctl_waitq,
1380 thread_is_running(ldlm_pools_thread), &lwi);
1384 static void ldlm_pools_thread_stop(void)
1386 if (ldlm_pools_thread == NULL) {
1390 thread_set_flags(ldlm_pools_thread, SVC_STOPPING);
1391 wake_up(&ldlm_pools_thread->t_ctl_waitq);
1394 * Make sure that pools thread is finished before freeing @thread.
1395 * This fixes possible race and oops due to accessing freed memory
1398 wait_for_completion(&ldlm_pools_comp);
1399 OBD_FREE_PTR(ldlm_pools_thread);
1400 ldlm_pools_thread = NULL;
1403 static struct shrinker ldlm_pools_srv_shrinker = {
1404 .count_objects = ldlm_pools_srv_count,
1405 .scan_objects = ldlm_pools_srv_scan,
1406 .seeks = DEFAULT_SEEKS,
1409 static struct shrinker ldlm_pools_cli_shrinker = {
1410 .count_objects = ldlm_pools_cli_count,
1411 .scan_objects = ldlm_pools_cli_scan,
1412 .seeks = DEFAULT_SEEKS,
1415 int ldlm_pools_init(void)
1419 rc = ldlm_pools_thread_start();
1421 register_shrinker(&ldlm_pools_srv_shrinker);
1422 register_shrinker(&ldlm_pools_cli_shrinker);
1426 EXPORT_SYMBOL(ldlm_pools_init);
1428 void ldlm_pools_fini(void)
1430 unregister_shrinker(&ldlm_pools_srv_shrinker);
1431 unregister_shrinker(&ldlm_pools_cli_shrinker);
1432 ldlm_pools_thread_stop();
1434 EXPORT_SYMBOL(ldlm_pools_fini);