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 "../include/lustre_dlm.h"
101 #include "../include/cl_object.h"
102 #include "../include/obd_class.h"
103 #include "../include/obd_support.h"
104 #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 extern struct proc_dir_entry *ldlm_ns_proc_dir;
145 static inline __u64 dru(__u64 val, __u32 shift, int round_up)
147 return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift;
150 static inline __u64 ldlm_pool_slv_max(__u32 L)
153 * Allow to have all locks for 1 client for 10 hrs.
154 * Formula is the following: limit * 10h / 1 client.
156 __u64 lim = (__u64)L * LDLM_POOL_MAX_AGE / 1;
160 static inline __u64 ldlm_pool_slv_min(__u32 L)
166 LDLM_POOL_FIRST_STAT = 0,
167 LDLM_POOL_GRANTED_STAT = LDLM_POOL_FIRST_STAT,
168 LDLM_POOL_GRANT_STAT,
169 LDLM_POOL_CANCEL_STAT,
170 LDLM_POOL_GRANT_RATE_STAT,
171 LDLM_POOL_CANCEL_RATE_STAT,
172 LDLM_POOL_GRANT_PLAN_STAT,
174 LDLM_POOL_SHRINK_REQTD_STAT,
175 LDLM_POOL_SHRINK_FREED_STAT,
176 LDLM_POOL_RECALC_STAT,
177 LDLM_POOL_TIMING_STAT,
181 static inline struct ldlm_namespace *ldlm_pl2ns(struct ldlm_pool *pl)
183 return container_of(pl, struct ldlm_namespace, ns_pool);
187 * Calculates suggested grant_step in % of available locks for passed
188 * \a period. This is later used in grant_plan calculations.
190 static inline int ldlm_pool_t2gsp(unsigned int t)
193 * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP
194 * and up to 30% for anything higher than LDLM_POOL_GSP_STEP.
196 * How this will affect execution is the following:
198 * - for thread period 1s we will have grant_step 1% which good from
199 * pov of taking some load off from server and push it out to clients.
200 * This is like that because 1% for grant_step means that server will
201 * not allow clients to get lots of locks in short period of time and
202 * keep all old locks in their caches. Clients will always have to
203 * get some locks back if they want to take some new;
205 * - for thread period 10s (which is default) we will have 23% which
206 * means that clients will have enough of room to take some new locks
207 * without getting some back. All locks from this 23% which were not
208 * taken by clients in current period will contribute in SLV growing.
209 * SLV growing means more locks cached on clients until limit or grant
212 return LDLM_POOL_MAX_GSP -
213 ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >>
214 (t >> LDLM_POOL_GSP_STEP_SHIFT));
218 * Recalculates next grant limit on passed \a pl.
220 * \pre ->pl_lock is locked.
222 static void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl)
224 int granted, grant_step, limit;
226 limit = ldlm_pool_get_limit(pl);
227 granted = atomic_read(&pl->pl_granted);
229 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
230 grant_step = ((limit - granted) * grant_step) / 100;
231 pl->pl_grant_plan = granted + grant_step;
232 limit = (limit * 5) >> 2;
233 if (pl->pl_grant_plan > limit)
234 pl->pl_grant_plan = limit;
238 * Recalculates next SLV on passed \a pl.
240 * \pre ->pl_lock is locked.
242 static void ldlm_pool_recalc_slv(struct ldlm_pool *pl)
252 slv = pl->pl_server_lock_volume;
253 grant_plan = pl->pl_grant_plan;
254 limit = ldlm_pool_get_limit(pl);
255 granted = atomic_read(&pl->pl_granted);
256 round_up = granted < limit;
258 grant_usage = max_t(int, limit - (granted - grant_plan), 1);
261 * Find out SLV change factor which is the ratio of grant usage
262 * from limit. SLV changes as fast as the ratio of grant plan
263 * consumption. The more locks from grant plan are not consumed
264 * by clients in last interval (idle time), the faster grows
265 * SLV. And the opposite, the more grant plan is over-consumed
266 * (load time) the faster drops SLV.
268 slv_factor = grant_usage << LDLM_POOL_SLV_SHIFT;
269 do_div(slv_factor, limit);
270 slv = slv * slv_factor;
271 slv = dru(slv, LDLM_POOL_SLV_SHIFT, round_up);
273 if (slv > ldlm_pool_slv_max(limit)) {
274 slv = ldlm_pool_slv_max(limit);
275 } else if (slv < ldlm_pool_slv_min(limit)) {
276 slv = ldlm_pool_slv_min(limit);
279 pl->pl_server_lock_volume = slv;
283 * Recalculates next stats on passed \a pl.
285 * \pre ->pl_lock is locked.
287 static void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
289 int grant_plan = pl->pl_grant_plan;
290 __u64 slv = pl->pl_server_lock_volume;
291 int granted = atomic_read(&pl->pl_granted);
292 int grant_rate = atomic_read(&pl->pl_grant_rate);
293 int cancel_rate = atomic_read(&pl->pl_cancel_rate);
295 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT,
297 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
299 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
301 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
303 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
308 * Sets current SLV into obd accessible via ldlm_pl2ns(pl)->ns_obd.
310 static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl)
312 struct obd_device *obd;
315 * Set new SLV in obd field for using it later without accessing the
316 * pool. This is required to avoid race between sending reply to client
317 * with new SLV and cleanup server stack in which we can't guarantee
318 * that namespace is still alive. We know only that obd is alive as
319 * long as valid export is alive.
321 obd = ldlm_pl2ns(pl)->ns_obd;
322 LASSERT(obd != NULL);
323 write_lock(&obd->obd_pool_lock);
324 obd->obd_pool_slv = pl->pl_server_lock_volume;
325 write_unlock(&obd->obd_pool_lock);
329 * Recalculates all pool fields on passed \a pl.
331 * \pre ->pl_lock is not locked.
333 static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
335 time_t recalc_interval_sec;
337 recalc_interval_sec = get_seconds() - pl->pl_recalc_time;
338 if (recalc_interval_sec < pl->pl_recalc_period)
341 spin_lock(&pl->pl_lock);
342 recalc_interval_sec = get_seconds() - pl->pl_recalc_time;
343 if (recalc_interval_sec < pl->pl_recalc_period) {
344 spin_unlock(&pl->pl_lock);
348 * Recalc SLV after last period. This should be done
349 * _before_ recalculating new grant plan.
351 ldlm_pool_recalc_slv(pl);
354 * Make sure that pool informed obd of last SLV changes.
356 ldlm_srv_pool_push_slv(pl);
359 * Update grant_plan for new period.
361 ldlm_pool_recalc_grant_plan(pl);
363 pl->pl_recalc_time = get_seconds();
364 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
365 recalc_interval_sec);
366 spin_unlock(&pl->pl_lock);
371 * This function is used on server side as main entry point for memory
372 * pressure handling. It decreases SLV on \a pl according to passed
373 * \a nr and \a gfp_mask.
375 * Our goal here is to decrease SLV such a way that clients hold \a nr
376 * locks smaller in next 10h.
378 static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
379 int nr, gfp_t gfp_mask)
384 * VM is asking how many entries may be potentially freed.
387 return atomic_read(&pl->pl_granted);
390 * Client already canceled locks but server is already in shrinker
391 * and can't cancel anything. Let's catch this race.
393 if (atomic_read(&pl->pl_granted) == 0)
396 spin_lock(&pl->pl_lock);
399 * We want shrinker to possibly cause cancellation of @nr locks from
400 * clients or grant approximately @nr locks smaller next intervals.
402 * This is why we decreased SLV by @nr. This effect will only be as
403 * long as one re-calc interval (1s these days) and this should be
404 * enough to pass this decreased SLV to all clients. On next recalc
405 * interval pool will either increase SLV if locks load is not high
406 * or will keep on same level or even decrease again, thus, shrinker
407 * decreased SLV will affect next recalc intervals and this way will
408 * make locking load lower.
410 if (nr < pl->pl_server_lock_volume) {
411 pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr;
413 limit = ldlm_pool_get_limit(pl);
414 pl->pl_server_lock_volume = ldlm_pool_slv_min(limit);
418 * Make sure that pool informed obd of last SLV changes.
420 ldlm_srv_pool_push_slv(pl);
421 spin_unlock(&pl->pl_lock);
424 * We did not really free any memory here so far, it only will be
425 * freed later may be, so that we return 0 to not confuse VM.
431 * Setup server side pool \a pl with passed \a limit.
433 static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit)
435 struct obd_device *obd;
437 obd = ldlm_pl2ns(pl)->ns_obd;
438 LASSERT(obd != NULL && obd != LP_POISON);
439 LASSERT(obd->obd_type != LP_POISON);
440 write_lock(&obd->obd_pool_lock);
441 obd->obd_pool_limit = limit;
442 write_unlock(&obd->obd_pool_lock);
444 ldlm_pool_set_limit(pl, limit);
449 * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl.
451 static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
453 struct obd_device *obd;
456 * Get new SLV and Limit from obd which is updated with coming
459 obd = ldlm_pl2ns(pl)->ns_obd;
460 LASSERT(obd != NULL);
461 read_lock(&obd->obd_pool_lock);
462 pl->pl_server_lock_volume = obd->obd_pool_slv;
463 ldlm_pool_set_limit(pl, obd->obd_pool_limit);
464 read_unlock(&obd->obd_pool_lock);
468 * Recalculates client size pool \a pl according to current SLV and Limit.
470 static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
472 time_t recalc_interval_sec;
474 recalc_interval_sec = get_seconds() - pl->pl_recalc_time;
475 if (recalc_interval_sec < pl->pl_recalc_period)
478 spin_lock(&pl->pl_lock);
480 * Check if we need to recalc lists now.
482 recalc_interval_sec = get_seconds() - pl->pl_recalc_time;
483 if (recalc_interval_sec < pl->pl_recalc_period) {
484 spin_unlock(&pl->pl_lock);
489 * Make sure that pool knows last SLV and Limit from obd.
491 ldlm_cli_pool_pop_slv(pl);
493 pl->pl_recalc_time = get_seconds();
494 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
495 recalc_interval_sec);
496 spin_unlock(&pl->pl_lock);
499 * Do not cancel locks in case lru resize is disabled for this ns.
501 if (!ns_connect_lru_resize(ldlm_pl2ns(pl)))
505 * In the time of canceling locks on client we do not need to maintain
506 * sharp timing, we only want to cancel locks asap according to new SLV.
507 * It may be called when SLV has changed much, this is why we do not
508 * take into account pl->pl_recalc_time here.
510 return ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC, LDLM_CANCEL_LRUR);
514 * This function is main entry point for memory pressure handling on client
515 * side. Main goal of this function is to cancel some number of locks on
516 * passed \a pl according to \a nr and \a gfp_mask.
518 static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
519 int nr, gfp_t gfp_mask)
521 struct ldlm_namespace *ns;
527 * Do not cancel locks in case lru resize is disabled for this ns.
529 if (!ns_connect_lru_resize(ns))
533 * Make sure that pool knows last SLV and Limit from obd.
535 ldlm_cli_pool_pop_slv(pl);
537 spin_lock(&ns->ns_lock);
538 unused = ns->ns_nr_unused;
539 spin_unlock(&ns->ns_lock);
542 return (unused / 100) * sysctl_vfs_cache_pressure;
544 return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_CANCEL_SHRINK);
547 static const struct ldlm_pool_ops ldlm_srv_pool_ops = {
548 .po_recalc = ldlm_srv_pool_recalc,
549 .po_shrink = ldlm_srv_pool_shrink,
550 .po_setup = ldlm_srv_pool_setup
553 static const struct ldlm_pool_ops ldlm_cli_pool_ops = {
554 .po_recalc = ldlm_cli_pool_recalc,
555 .po_shrink = ldlm_cli_pool_shrink
559 * Pool recalc wrapper. Will call either client or server pool recalc callback
560 * depending what pool \a pl is used.
562 int ldlm_pool_recalc(struct ldlm_pool *pl)
564 time_t recalc_interval_sec;
567 recalc_interval_sec = get_seconds() - pl->pl_recalc_time;
568 if (recalc_interval_sec <= 0)
571 spin_lock(&pl->pl_lock);
572 if (recalc_interval_sec > 0) {
574 * Update pool statistics every 1s.
576 ldlm_pool_recalc_stats(pl);
579 * Zero out all rates and speed for the last period.
581 atomic_set(&pl->pl_grant_rate, 0);
582 atomic_set(&pl->pl_cancel_rate, 0);
584 spin_unlock(&pl->pl_lock);
587 if (pl->pl_ops->po_recalc != NULL) {
588 count = pl->pl_ops->po_recalc(pl);
589 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
592 recalc_interval_sec = pl->pl_recalc_time - get_seconds() +
593 pl->pl_recalc_period;
595 return recalc_interval_sec;
599 * Pool shrink wrapper. Will call either client or server pool recalc callback
600 * depending what pool pl is used. When nr == 0, just return the number of
601 * freeable locks. Otherwise, return the number of canceled locks.
603 int ldlm_pool_shrink(struct ldlm_pool *pl, int nr,
608 if (pl->pl_ops->po_shrink != NULL) {
609 cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask);
611 lprocfs_counter_add(pl->pl_stats,
612 LDLM_POOL_SHRINK_REQTD_STAT,
614 lprocfs_counter_add(pl->pl_stats,
615 LDLM_POOL_SHRINK_FREED_STAT,
617 CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, "
618 "shrunk %d\n", pl->pl_name, nr, cancel);
623 EXPORT_SYMBOL(ldlm_pool_shrink);
626 * Pool setup wrapper. Will call either client or server pool recalc callback
627 * depending what pool \a pl is used.
629 * Sets passed \a limit into pool \a pl.
631 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
633 if (pl->pl_ops->po_setup != NULL)
634 return pl->pl_ops->po_setup(pl, limit);
637 EXPORT_SYMBOL(ldlm_pool_setup);
639 #if defined (CONFIG_PROC_FS)
640 static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
642 int granted, grant_rate, cancel_rate, grant_step;
643 int grant_speed, grant_plan, lvf;
644 struct ldlm_pool *pl = m->private;
648 spin_lock(&pl->pl_lock);
649 slv = pl->pl_server_lock_volume;
650 clv = pl->pl_client_lock_volume;
651 limit = ldlm_pool_get_limit(pl);
652 grant_plan = pl->pl_grant_plan;
653 granted = atomic_read(&pl->pl_granted);
654 grant_rate = atomic_read(&pl->pl_grant_rate);
655 cancel_rate = atomic_read(&pl->pl_cancel_rate);
656 grant_speed = grant_rate - cancel_rate;
657 lvf = atomic_read(&pl->pl_lock_volume_factor);
658 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
659 spin_unlock(&pl->pl_lock);
661 seq_printf(m, "LDLM pool state (%s):\n"
665 pl->pl_name, slv, clv, lvf);
667 if (ns_is_server(ldlm_pl2ns(pl))) {
668 seq_printf(m, " GSP: %d%%\n"
670 grant_step, grant_plan);
672 seq_printf(m, " GR: %d\n" " CR: %d\n" " GS: %d\n"
673 " G: %d\n" " L: %d\n",
674 grant_rate, cancel_rate, grant_speed,
679 LPROC_SEQ_FOPS_RO(lprocfs_pool_state);
681 static int lprocfs_grant_speed_seq_show(struct seq_file *m, void *unused)
683 struct ldlm_pool *pl = m->private;
686 spin_lock(&pl->pl_lock);
687 /* serialize with ldlm_pool_recalc */
688 grant_speed = atomic_read(&pl->pl_grant_rate) -
689 atomic_read(&pl->pl_cancel_rate);
690 spin_unlock(&pl->pl_lock);
691 return lprocfs_rd_uint(m, &grant_speed);
694 LDLM_POOL_PROC_READER_SEQ_SHOW(grant_plan, int);
695 LPROC_SEQ_FOPS_RO(lprocfs_grant_plan);
697 LDLM_POOL_PROC_READER_SEQ_SHOW(recalc_period, int);
698 LDLM_POOL_PROC_WRITER(recalc_period, int);
699 static ssize_t lprocfs_recalc_period_seq_write(struct file *file, const char *buf,
700 size_t len, loff_t *off)
702 struct seq_file *seq = file->private_data;
704 return lprocfs_wr_recalc_period(file, buf, len, seq->private);
706 LPROC_SEQ_FOPS(lprocfs_recalc_period);
708 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, u64);
709 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, atomic);
710 LPROC_SEQ_FOPS_RW_TYPE(ldlm_pool_rw, atomic);
712 LPROC_SEQ_FOPS_RO(lprocfs_grant_speed);
714 #define LDLM_POOL_ADD_VAR(name, var, ops) \
716 snprintf(var_name, MAX_STRING_SIZE, #name); \
717 pool_vars[0].data = var; \
718 pool_vars[0].fops = ops; \
719 lprocfs_add_vars(pl->pl_proc_dir, pool_vars, NULL);\
722 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
724 struct ldlm_namespace *ns = ldlm_pl2ns(pl);
725 struct proc_dir_entry *parent_ns_proc;
726 struct lprocfs_vars pool_vars[2];
727 char *var_name = NULL;
730 OBD_ALLOC(var_name, MAX_STRING_SIZE + 1);
734 parent_ns_proc = ns->ns_proc_dir_entry;
735 if (parent_ns_proc == NULL) {
736 CERROR("%s: proc entry is not initialized\n",
741 pl->pl_proc_dir = lprocfs_register("pool", parent_ns_proc,
743 if (IS_ERR(pl->pl_proc_dir)) {
744 CERROR("LProcFS failed in ldlm-pool-init\n");
745 rc = PTR_ERR(pl->pl_proc_dir);
746 pl->pl_proc_dir = NULL;
750 var_name[MAX_STRING_SIZE] = '\0';
751 memset(pool_vars, 0, sizeof(pool_vars));
752 pool_vars[0].name = var_name;
754 LDLM_POOL_ADD_VAR("server_lock_volume", &pl->pl_server_lock_volume,
755 &ldlm_pool_u64_fops);
756 LDLM_POOL_ADD_VAR("limit", &pl->pl_limit, &ldlm_pool_rw_atomic_fops);
757 LDLM_POOL_ADD_VAR("granted", &pl->pl_granted, &ldlm_pool_atomic_fops);
758 LDLM_POOL_ADD_VAR("grant_speed", pl, &lprocfs_grant_speed_fops);
759 LDLM_POOL_ADD_VAR("cancel_rate", &pl->pl_cancel_rate,
760 &ldlm_pool_atomic_fops);
761 LDLM_POOL_ADD_VAR("grant_rate", &pl->pl_grant_rate,
762 &ldlm_pool_atomic_fops);
763 LDLM_POOL_ADD_VAR("grant_plan", pl, &lprocfs_grant_plan_fops);
764 LDLM_POOL_ADD_VAR("recalc_period", pl, &lprocfs_recalc_period_fops);
765 LDLM_POOL_ADD_VAR("lock_volume_factor", &pl->pl_lock_volume_factor,
766 &ldlm_pool_rw_atomic_fops);
767 LDLM_POOL_ADD_VAR("state", pl, &lprocfs_pool_state_fops);
769 pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT -
770 LDLM_POOL_FIRST_STAT, 0);
776 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
777 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
779 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
780 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
782 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
783 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
785 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
786 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
787 "grant_rate", "locks/s");
788 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
789 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
790 "cancel_rate", "locks/s");
791 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
792 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
793 "grant_plan", "locks/s");
794 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT,
795 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
797 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT,
798 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
799 "shrink_request", "locks");
800 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT,
801 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
802 "shrink_freed", "locks");
803 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
804 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
805 "recalc_freed", "locks");
806 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
807 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
808 "recalc_timing", "sec");
809 rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
812 OBD_FREE(var_name, MAX_STRING_SIZE + 1);
816 static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
818 if (pl->pl_stats != NULL) {
819 lprocfs_free_stats(&pl->pl_stats);
822 if (pl->pl_proc_dir != NULL) {
823 lprocfs_remove(&pl->pl_proc_dir);
824 pl->pl_proc_dir = NULL;
827 #else /* !CONFIG_PROC_FS */
828 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
833 static void ldlm_pool_proc_fini(struct ldlm_pool *pl) {}
834 #endif /* CONFIG_PROC_FS */
836 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
837 int idx, ldlm_side_t client)
841 spin_lock_init(&pl->pl_lock);
842 atomic_set(&pl->pl_granted, 0);
843 pl->pl_recalc_time = get_seconds();
844 atomic_set(&pl->pl_lock_volume_factor, 1);
846 atomic_set(&pl->pl_grant_rate, 0);
847 atomic_set(&pl->pl_cancel_rate, 0);
848 pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
850 snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
851 ldlm_ns_name(ns), idx);
853 if (client == LDLM_NAMESPACE_SERVER) {
854 pl->pl_ops = &ldlm_srv_pool_ops;
855 ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L);
856 pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD;
857 pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L);
859 ldlm_pool_set_limit(pl, 1);
860 pl->pl_server_lock_volume = 0;
861 pl->pl_ops = &ldlm_cli_pool_ops;
862 pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
864 pl->pl_client_lock_volume = 0;
865 rc = ldlm_pool_proc_init(pl);
869 CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name);
873 EXPORT_SYMBOL(ldlm_pool_init);
875 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));
886 EXPORT_SYMBOL(ldlm_pool_fini);
889 * Add new taken ldlm lock \a lock into pool \a pl accounting.
891 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
894 * FLOCK locks are special in a sense that they are almost never
895 * cancelled, instead special kind of lock is used to drop them.
896 * also there is no LRU for flock locks, so no point in tracking
899 if (lock->l_resource->lr_type == LDLM_FLOCK)
902 atomic_inc(&pl->pl_granted);
903 atomic_inc(&pl->pl_grant_rate);
904 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
906 * Do not do pool recalc for client side as all locks which
907 * potentially may be canceled has already been packed into
908 * enqueue/cancel rpc. Also we do not want to run out of stack
909 * with too long call paths.
911 if (ns_is_server(ldlm_pl2ns(pl)))
912 ldlm_pool_recalc(pl);
914 EXPORT_SYMBOL(ldlm_pool_add);
917 * Remove ldlm lock \a lock from pool \a pl accounting.
919 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
922 * Filter out FLOCK locks. Read above comment in ldlm_pool_add().
924 if (lock->l_resource->lr_type == LDLM_FLOCK)
927 LASSERT(atomic_read(&pl->pl_granted) > 0);
928 atomic_dec(&pl->pl_granted);
929 atomic_inc(&pl->pl_cancel_rate);
931 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
933 if (ns_is_server(ldlm_pl2ns(pl)))
934 ldlm_pool_recalc(pl);
936 EXPORT_SYMBOL(ldlm_pool_del);
939 * Returns current \a pl SLV.
941 * \pre ->pl_lock is not locked.
943 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
946 spin_lock(&pl->pl_lock);
947 slv = pl->pl_server_lock_volume;
948 spin_unlock(&pl->pl_lock);
951 EXPORT_SYMBOL(ldlm_pool_get_slv);
954 * Sets passed \a slv to \a pl.
956 * \pre ->pl_lock is not locked.
958 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
960 spin_lock(&pl->pl_lock);
961 pl->pl_server_lock_volume = slv;
962 spin_unlock(&pl->pl_lock);
964 EXPORT_SYMBOL(ldlm_pool_set_slv);
967 * Returns current \a pl CLV.
969 * \pre ->pl_lock is not locked.
971 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
974 spin_lock(&pl->pl_lock);
975 slv = pl->pl_client_lock_volume;
976 spin_unlock(&pl->pl_lock);
979 EXPORT_SYMBOL(ldlm_pool_get_clv);
982 * Sets passed \a clv to \a pl.
984 * \pre ->pl_lock is not locked.
986 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
988 spin_lock(&pl->pl_lock);
989 pl->pl_client_lock_volume = clv;
990 spin_unlock(&pl->pl_lock);
992 EXPORT_SYMBOL(ldlm_pool_set_clv);
995 * Returns current \a pl limit.
997 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
999 return atomic_read(&pl->pl_limit);
1001 EXPORT_SYMBOL(ldlm_pool_get_limit);
1004 * Sets passed \a limit to \a pl.
1006 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1008 atomic_set(&pl->pl_limit, limit);
1010 EXPORT_SYMBOL(ldlm_pool_set_limit);
1013 * Returns current LVF from \a pl.
1015 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1017 return atomic_read(&pl->pl_lock_volume_factor);
1019 EXPORT_SYMBOL(ldlm_pool_get_lvf);
1021 static int ldlm_pool_granted(struct ldlm_pool *pl)
1023 return atomic_read(&pl->pl_granted);
1026 static struct ptlrpc_thread *ldlm_pools_thread;
1027 static struct completion ldlm_pools_comp;
1030 * count locks from all namespaces (if possible). Returns number of
1033 static unsigned long ldlm_pools_count(ldlm_side_t client, gfp_t gfp_mask)
1035 int total = 0, nr_ns;
1036 struct ldlm_namespace *ns;
1037 struct ldlm_namespace *ns_old = NULL; /* loop detection */
1040 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1043 CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n",
1044 client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1046 cookie = cl_env_reenter();
1049 * Find out how many resources we may release.
1051 for (nr_ns = ldlm_namespace_nr_read(client);
1052 nr_ns > 0; nr_ns--) {
1053 mutex_lock(ldlm_namespace_lock(client));
1054 if (list_empty(ldlm_namespace_list(client))) {
1055 mutex_unlock(ldlm_namespace_lock(client));
1056 cl_env_reexit(cookie);
1059 ns = ldlm_namespace_first_locked(client);
1062 mutex_unlock(ldlm_namespace_lock(client));
1066 if (ldlm_ns_empty(ns)) {
1067 ldlm_namespace_move_to_inactive_locked(ns, client);
1068 mutex_unlock(ldlm_namespace_lock(client));
1075 ldlm_namespace_get(ns);
1076 ldlm_namespace_move_to_active_locked(ns, client);
1077 mutex_unlock(ldlm_namespace_lock(client));
1078 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
1079 ldlm_namespace_put(ns);
1082 cl_env_reexit(cookie);
1086 static unsigned long ldlm_pools_scan(ldlm_side_t client, int nr, gfp_t gfp_mask)
1088 unsigned long freed = 0;
1090 struct ldlm_namespace *ns;
1093 if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1096 cookie = cl_env_reenter();
1099 * Shrink at least ldlm_namespace_nr_read(client) namespaces.
1101 for (tmp = nr_ns = ldlm_namespace_nr_read(client);
1103 int cancel, nr_locks;
1106 * Do not call shrink under ldlm_namespace_lock(client)
1108 mutex_lock(ldlm_namespace_lock(client));
1109 if (list_empty(ldlm_namespace_list(client))) {
1110 mutex_unlock(ldlm_namespace_lock(client));
1113 ns = ldlm_namespace_first_locked(client);
1114 ldlm_namespace_get(ns);
1115 ldlm_namespace_move_to_active_locked(ns, client);
1116 mutex_unlock(ldlm_namespace_lock(client));
1118 nr_locks = ldlm_pool_granted(&ns->ns_pool);
1120 * We use to shrink propotionally but with new shrinker API,
1121 * we lost the total number of freeable locks.
1123 cancel = 1 + min_t(int, nr_locks, nr / nr_ns);
1124 freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
1125 ldlm_namespace_put(ns);
1127 cl_env_reexit(cookie);
1129 * we only decrease the SLV in server pools shrinker, return
1130 * SHRINK_STOP to kernel to avoid needless loop. LU-1128
1132 return (client == LDLM_NAMESPACE_SERVER) ? SHRINK_STOP : freed;
1135 static unsigned long ldlm_pools_srv_count(struct shrinker *s, struct shrink_control *sc)
1137 return ldlm_pools_count(LDLM_NAMESPACE_SERVER, sc->gfp_mask);
1140 static unsigned long ldlm_pools_srv_scan(struct shrinker *s, struct shrink_control *sc)
1142 return ldlm_pools_scan(LDLM_NAMESPACE_SERVER, sc->nr_to_scan,
1146 static unsigned long ldlm_pools_cli_count(struct shrinker *s, struct shrink_control *sc)
1148 return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask);
1151 static unsigned long ldlm_pools_cli_scan(struct shrinker *s, struct shrink_control *sc)
1153 return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan,
1157 int ldlm_pools_recalc(ldlm_side_t client)
1159 __u32 nr_l = 0, nr_p = 0, l;
1160 struct ldlm_namespace *ns;
1161 struct ldlm_namespace *ns_old = NULL;
1163 int time = 50; /* seconds of sleep if no active namespaces */
1166 * No need to setup pool limit for client pools.
1168 if (client == LDLM_NAMESPACE_SERVER) {
1170 * Check all modest namespaces first.
1172 mutex_lock(ldlm_namespace_lock(client));
1173 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);