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
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"
109 * 50 ldlm locks for 1MB of RAM.
111 #define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_CACHE_SHIFT)) * 50)
114 * Maximal possible grant step plan in %.
116 #define LDLM_POOL_MAX_GSP (30)
119 * Minimal possible grant step plan in %.
121 #define LDLM_POOL_MIN_GSP (1)
124 * This controls the speed of reaching LDLM_POOL_MAX_GSP
125 * with increasing thread period.
127 #define LDLM_POOL_GSP_STEP_SHIFT (2)
130 * LDLM_POOL_GSP% of all locks is default GP.
132 #define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100)
135 * Max age for locks on clients.
137 #define LDLM_POOL_MAX_AGE (36000)
140 * The granularity of SLV calculation.
142 #define LDLM_POOL_SLV_SHIFT (10)
144 static inline __u64
dru(__u64 val
, __u32 shift
, int round_up
)
146 return (val
+ (round_up
? (1 << shift
) - 1 : 0)) >> shift
;
149 static inline __u64
ldlm_pool_slv_max(__u32 L
)
152 * Allow to have all locks for 1 client for 10 hrs.
153 * Formula is the following: limit * 10h / 1 client.
155 __u64 lim
= (__u64
)L
* LDLM_POOL_MAX_AGE
/ 1;
159 static inline __u64
ldlm_pool_slv_min(__u32 L
)
165 LDLM_POOL_FIRST_STAT
= 0,
166 LDLM_POOL_GRANTED_STAT
= LDLM_POOL_FIRST_STAT
,
167 LDLM_POOL_GRANT_STAT
,
168 LDLM_POOL_CANCEL_STAT
,
169 LDLM_POOL_GRANT_RATE_STAT
,
170 LDLM_POOL_CANCEL_RATE_STAT
,
171 LDLM_POOL_GRANT_PLAN_STAT
,
173 LDLM_POOL_SHRINK_REQTD_STAT
,
174 LDLM_POOL_SHRINK_FREED_STAT
,
175 LDLM_POOL_RECALC_STAT
,
176 LDLM_POOL_TIMING_STAT
,
180 static inline struct ldlm_namespace
*ldlm_pl2ns(struct ldlm_pool
*pl
)
182 return container_of(pl
, struct ldlm_namespace
, ns_pool
);
186 * Calculates suggested grant_step in % of available locks for passed
187 * \a period. This is later used in grant_plan calculations.
189 static inline int ldlm_pool_t2gsp(unsigned int t
)
192 * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP
193 * and up to 30% for anything higher than LDLM_POOL_GSP_STEP.
195 * How this will affect execution is the following:
197 * - for thread period 1s we will have grant_step 1% which good from
198 * pov of taking some load off from server and push it out to clients.
199 * This is like that because 1% for grant_step means that server will
200 * not allow clients to get lots of locks in short period of time and
201 * keep all old locks in their caches. Clients will always have to
202 * get some locks back if they want to take some new;
204 * - for thread period 10s (which is default) we will have 23% which
205 * means that clients will have enough of room to take some new locks
206 * without getting some back. All locks from this 23% which were not
207 * taken by clients in current period will contribute in SLV growing.
208 * SLV growing means more locks cached on clients until limit or grant
211 return LDLM_POOL_MAX_GSP
-
212 ((LDLM_POOL_MAX_GSP
- LDLM_POOL_MIN_GSP
) >>
213 (t
>> LDLM_POOL_GSP_STEP_SHIFT
));
217 * Recalculates next grant limit on passed \a pl.
219 * \pre ->pl_lock is locked.
221 static void ldlm_pool_recalc_grant_plan(struct ldlm_pool
*pl
)
223 int granted
, grant_step
, limit
;
225 limit
= ldlm_pool_get_limit(pl
);
226 granted
= atomic_read(&pl
->pl_granted
);
228 grant_step
= ldlm_pool_t2gsp(pl
->pl_recalc_period
);
229 grant_step
= ((limit
- granted
) * grant_step
) / 100;
230 pl
->pl_grant_plan
= granted
+ grant_step
;
231 limit
= (limit
* 5) >> 2;
232 if (pl
->pl_grant_plan
> limit
)
233 pl
->pl_grant_plan
= limit
;
237 * Recalculates next SLV on passed \a pl.
239 * \pre ->pl_lock is locked.
241 static void ldlm_pool_recalc_slv(struct ldlm_pool
*pl
)
251 slv
= pl
->pl_server_lock_volume
;
252 grant_plan
= pl
->pl_grant_plan
;
253 limit
= ldlm_pool_get_limit(pl
);
254 granted
= atomic_read(&pl
->pl_granted
);
255 round_up
= granted
< limit
;
257 grant_usage
= max_t(int, limit
- (granted
- grant_plan
), 1);
260 * Find out SLV change factor which is the ratio of grant usage
261 * from limit. SLV changes as fast as the ratio of grant plan
262 * consumption. The more locks from grant plan are not consumed
263 * by clients in last interval (idle time), the faster grows
264 * SLV. And the opposite, the more grant plan is over-consumed
265 * (load time) the faster drops SLV.
267 slv_factor
= grant_usage
<< LDLM_POOL_SLV_SHIFT
;
268 do_div(slv_factor
, limit
);
269 slv
= slv
* slv_factor
;
270 slv
= dru(slv
, LDLM_POOL_SLV_SHIFT
, round_up
);
272 if (slv
> ldlm_pool_slv_max(limit
))
273 slv
= ldlm_pool_slv_max(limit
);
274 else if (slv
< ldlm_pool_slv_min(limit
))
275 slv
= ldlm_pool_slv_min(limit
);
277 pl
->pl_server_lock_volume
= slv
;
281 * Recalculates next stats on passed \a pl.
283 * \pre ->pl_lock is locked.
285 static void ldlm_pool_recalc_stats(struct ldlm_pool
*pl
)
287 int grant_plan
= pl
->pl_grant_plan
;
288 __u64 slv
= pl
->pl_server_lock_volume
;
289 int granted
= atomic_read(&pl
->pl_granted
);
290 int grant_rate
= atomic_read(&pl
->pl_grant_rate
);
291 int cancel_rate
= atomic_read(&pl
->pl_cancel_rate
);
293 lprocfs_counter_add(pl
->pl_stats
, LDLM_POOL_SLV_STAT
,
295 lprocfs_counter_add(pl
->pl_stats
, LDLM_POOL_GRANTED_STAT
,
297 lprocfs_counter_add(pl
->pl_stats
, LDLM_POOL_GRANT_RATE_STAT
,
299 lprocfs_counter_add(pl
->pl_stats
, LDLM_POOL_GRANT_PLAN_STAT
,
301 lprocfs_counter_add(pl
->pl_stats
, LDLM_POOL_CANCEL_RATE_STAT
,
306 * Sets current SLV into obd accessible via ldlm_pl2ns(pl)->ns_obd.
308 static void ldlm_srv_pool_push_slv(struct ldlm_pool
*pl
)
310 struct obd_device
*obd
;
313 * Set new SLV in obd field for using it later without accessing the
314 * pool. This is required to avoid race between sending reply to client
315 * with new SLV and cleanup server stack in which we can't guarantee
316 * that namespace is still alive. We know only that obd is alive as
317 * long as valid export is alive.
319 obd
= ldlm_pl2ns(pl
)->ns_obd
;
320 LASSERT(obd
!= NULL
);
321 write_lock(&obd
->obd_pool_lock
);
322 obd
->obd_pool_slv
= pl
->pl_server_lock_volume
;
323 write_unlock(&obd
->obd_pool_lock
);
327 * Recalculates all pool fields on passed \a pl.
329 * \pre ->pl_lock is not locked.
331 static int ldlm_srv_pool_recalc(struct ldlm_pool
*pl
)
333 time_t recalc_interval_sec
;
335 recalc_interval_sec
= get_seconds() - pl
->pl_recalc_time
;
336 if (recalc_interval_sec
< pl
->pl_recalc_period
)
339 spin_lock(&pl
->pl_lock
);
340 recalc_interval_sec
= get_seconds() - pl
->pl_recalc_time
;
341 if (recalc_interval_sec
< pl
->pl_recalc_period
) {
342 spin_unlock(&pl
->pl_lock
);
346 * Recalc SLV after last period. This should be done
347 * _before_ recalculating new grant plan.
349 ldlm_pool_recalc_slv(pl
);
352 * Make sure that pool informed obd of last SLV changes.
354 ldlm_srv_pool_push_slv(pl
);
357 * Update grant_plan for new period.
359 ldlm_pool_recalc_grant_plan(pl
);
361 pl
->pl_recalc_time
= get_seconds();
362 lprocfs_counter_add(pl
->pl_stats
, LDLM_POOL_TIMING_STAT
,
363 recalc_interval_sec
);
364 spin_unlock(&pl
->pl_lock
);
369 * This function is used on server side as main entry point for memory
370 * pressure handling. It decreases SLV on \a pl according to passed
371 * \a nr and \a gfp_mask.
373 * Our goal here is to decrease SLV such a way that clients hold \a nr
374 * locks smaller in next 10h.
376 static int ldlm_srv_pool_shrink(struct ldlm_pool
*pl
,
377 int nr
, gfp_t gfp_mask
)
382 * VM is asking how many entries may be potentially freed.
385 return atomic_read(&pl
->pl_granted
);
388 * Client already canceled locks but server is already in shrinker
389 * and can't cancel anything. Let's catch this race.
391 if (atomic_read(&pl
->pl_granted
) == 0)
394 spin_lock(&pl
->pl_lock
);
397 * We want shrinker to possibly cause cancellation of @nr locks from
398 * clients or grant approximately @nr locks smaller next intervals.
400 * This is why we decreased SLV by @nr. This effect will only be as
401 * long as one re-calc interval (1s these days) and this should be
402 * enough to pass this decreased SLV to all clients. On next recalc
403 * interval pool will either increase SLV if locks load is not high
404 * or will keep on same level or even decrease again, thus, shrinker
405 * decreased SLV will affect next recalc intervals and this way will
406 * make locking load lower.
408 if (nr
< pl
->pl_server_lock_volume
) {
409 pl
->pl_server_lock_volume
= pl
->pl_server_lock_volume
- nr
;
411 limit
= ldlm_pool_get_limit(pl
);
412 pl
->pl_server_lock_volume
= ldlm_pool_slv_min(limit
);
416 * Make sure that pool informed obd of last SLV changes.
418 ldlm_srv_pool_push_slv(pl
);
419 spin_unlock(&pl
->pl_lock
);
422 * We did not really free any memory here so far, it only will be
423 * freed later may be, so that we return 0 to not confuse VM.
429 * Setup server side pool \a pl with passed \a limit.
431 static int ldlm_srv_pool_setup(struct ldlm_pool
*pl
, int limit
)
433 struct obd_device
*obd
;
435 obd
= ldlm_pl2ns(pl
)->ns_obd
;
436 LASSERT(obd
!= NULL
&& obd
!= LP_POISON
);
437 LASSERT(obd
->obd_type
!= LP_POISON
);
438 write_lock(&obd
->obd_pool_lock
);
439 obd
->obd_pool_limit
= limit
;
440 write_unlock(&obd
->obd_pool_lock
);
442 ldlm_pool_set_limit(pl
, limit
);
447 * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl.
449 static void ldlm_cli_pool_pop_slv(struct ldlm_pool
*pl
)
451 struct obd_device
*obd
;
454 * Get new SLV and Limit from obd which is updated with coming
457 obd
= ldlm_pl2ns(pl
)->ns_obd
;
458 LASSERT(obd
!= NULL
);
459 read_lock(&obd
->obd_pool_lock
);
460 pl
->pl_server_lock_volume
= obd
->obd_pool_slv
;
461 ldlm_pool_set_limit(pl
, obd
->obd_pool_limit
);
462 read_unlock(&obd
->obd_pool_lock
);
466 * Recalculates client size pool \a pl according to current SLV and Limit.
468 static int ldlm_cli_pool_recalc(struct ldlm_pool
*pl
)
470 time_t recalc_interval_sec
;
473 recalc_interval_sec
= get_seconds() - pl
->pl_recalc_time
;
474 if (recalc_interval_sec
< pl
->pl_recalc_period
)
477 spin_lock(&pl
->pl_lock
);
479 * Check if we need to recalc lists now.
481 recalc_interval_sec
= get_seconds() - pl
->pl_recalc_time
;
482 if (recalc_interval_sec
< pl
->pl_recalc_period
) {
483 spin_unlock(&pl
->pl_lock
);
488 * Make sure that pool knows last SLV and Limit from obd.
490 ldlm_cli_pool_pop_slv(pl
);
492 spin_unlock(&pl
->pl_lock
);
495 * Do not cancel locks in case lru resize is disabled for this ns.
497 if (!ns_connect_lru_resize(ldlm_pl2ns(pl
))) {
503 * In the time of canceling locks on client we do not need to maintain
504 * sharp timing, we only want to cancel locks asap according to new SLV.
505 * It may be called when SLV has changed much, this is why we do not
506 * take into account pl->pl_recalc_time here.
508 ret
= ldlm_cancel_lru(ldlm_pl2ns(pl
), 0, LCF_ASYNC
, LDLM_CANCEL_LRUR
);
511 spin_lock(&pl
->pl_lock
);
513 * Time of LRU resizing might be longer than period,
514 * so update after LRU resizing rather than before it.
516 pl
->pl_recalc_time
= get_seconds();
517 lprocfs_counter_add(pl
->pl_stats
, LDLM_POOL_TIMING_STAT
,
518 recalc_interval_sec
);
519 spin_unlock(&pl
->pl_lock
);
524 * This function is main entry point for memory pressure handling on client
525 * side. Main goal of this function is to cancel some number of locks on
526 * passed \a pl according to \a nr and \a gfp_mask.
528 static int ldlm_cli_pool_shrink(struct ldlm_pool
*pl
,
529 int nr
, gfp_t gfp_mask
)
531 struct ldlm_namespace
*ns
;
537 * Do not cancel locks in case lru resize is disabled for this ns.
539 if (!ns_connect_lru_resize(ns
))
543 * Make sure that pool knows last SLV and Limit from obd.
545 ldlm_cli_pool_pop_slv(pl
);
547 spin_lock(&ns
->ns_lock
);
548 unused
= ns
->ns_nr_unused
;
549 spin_unlock(&ns
->ns_lock
);
552 return (unused
/ 100) * sysctl_vfs_cache_pressure
;
554 return ldlm_cancel_lru(ns
, nr
, LCF_ASYNC
, LDLM_CANCEL_SHRINK
);
557 static const struct ldlm_pool_ops ldlm_srv_pool_ops
= {
558 .po_recalc
= ldlm_srv_pool_recalc
,
559 .po_shrink
= ldlm_srv_pool_shrink
,
560 .po_setup
= ldlm_srv_pool_setup
563 static const struct ldlm_pool_ops ldlm_cli_pool_ops
= {
564 .po_recalc
= ldlm_cli_pool_recalc
,
565 .po_shrink
= ldlm_cli_pool_shrink
569 * Pool recalc wrapper. Will call either client or server pool recalc callback
570 * depending what pool \a pl is used.
572 int ldlm_pool_recalc(struct ldlm_pool
*pl
)
574 time_t recalc_interval_sec
;
577 recalc_interval_sec
= get_seconds() - pl
->pl_recalc_time
;
578 if (recalc_interval_sec
<= 0)
581 spin_lock(&pl
->pl_lock
);
582 if (recalc_interval_sec
> 0) {
584 * Update pool statistics every 1s.
586 ldlm_pool_recalc_stats(pl
);
589 * Zero out all rates and speed for the last period.
591 atomic_set(&pl
->pl_grant_rate
, 0);
592 atomic_set(&pl
->pl_cancel_rate
, 0);
594 spin_unlock(&pl
->pl_lock
);
597 if (pl
->pl_ops
->po_recalc
!= NULL
) {
598 count
= pl
->pl_ops
->po_recalc(pl
);
599 lprocfs_counter_add(pl
->pl_stats
, LDLM_POOL_RECALC_STAT
,
602 recalc_interval_sec
= pl
->pl_recalc_time
- get_seconds() +
603 pl
->pl_recalc_period
;
604 if (recalc_interval_sec
<= 0) {
605 /* Prevent too frequent recalculation. */
606 CDEBUG(D_DLMTRACE
, "Negative interval(%ld), "
607 "too short period(%ld)",
609 pl
->pl_recalc_period
);
610 recalc_interval_sec
= 1;
613 return recalc_interval_sec
;
617 * Pool shrink wrapper. Will call either client or server pool recalc callback
618 * depending what pool pl is used. When nr == 0, just return the number of
619 * freeable locks. Otherwise, return the number of canceled locks.
621 int ldlm_pool_shrink(struct ldlm_pool
*pl
, int nr
,
626 if (pl
->pl_ops
->po_shrink
!= NULL
) {
627 cancel
= pl
->pl_ops
->po_shrink(pl
, nr
, gfp_mask
);
629 lprocfs_counter_add(pl
->pl_stats
,
630 LDLM_POOL_SHRINK_REQTD_STAT
,
632 lprocfs_counter_add(pl
->pl_stats
,
633 LDLM_POOL_SHRINK_FREED_STAT
,
635 CDEBUG(D_DLMTRACE
, "%s: request to shrink %d locks, shrunk %d\n",
636 pl
->pl_name
, nr
, cancel
);
641 EXPORT_SYMBOL(ldlm_pool_shrink
);
644 * Pool setup wrapper. Will call either client or server pool recalc callback
645 * depending what pool \a pl is used.
647 * Sets passed \a limit into pool \a pl.
649 int ldlm_pool_setup(struct ldlm_pool
*pl
, int limit
)
651 if (pl
->pl_ops
->po_setup
!= NULL
)
652 return pl
->pl_ops
->po_setup(pl
, limit
);
655 EXPORT_SYMBOL(ldlm_pool_setup
);
657 static int lprocfs_pool_state_seq_show(struct seq_file
*m
, void *unused
)
659 int granted
, grant_rate
, cancel_rate
, grant_step
;
660 int grant_speed
, grant_plan
, lvf
;
661 struct ldlm_pool
*pl
= m
->private;
665 spin_lock(&pl
->pl_lock
);
666 slv
= pl
->pl_server_lock_volume
;
667 clv
= pl
->pl_client_lock_volume
;
668 limit
= ldlm_pool_get_limit(pl
);
669 grant_plan
= pl
->pl_grant_plan
;
670 granted
= atomic_read(&pl
->pl_granted
);
671 grant_rate
= atomic_read(&pl
->pl_grant_rate
);
672 cancel_rate
= atomic_read(&pl
->pl_cancel_rate
);
673 grant_speed
= grant_rate
- cancel_rate
;
674 lvf
= atomic_read(&pl
->pl_lock_volume_factor
);
675 grant_step
= ldlm_pool_t2gsp(pl
->pl_recalc_period
);
676 spin_unlock(&pl
->pl_lock
);
678 seq_printf(m
, "LDLM pool state (%s):\n"
682 pl
->pl_name
, slv
, clv
, lvf
);
684 if (ns_is_server(ldlm_pl2ns(pl
))) {
685 seq_printf(m
, " GSP: %d%%\n"
687 grant_step
, grant_plan
);
689 seq_printf(m
, " GR: %d\n CR: %d\n GS: %d\n"
691 grant_rate
, cancel_rate
, grant_speed
,
696 LPROC_SEQ_FOPS_RO(lprocfs_pool_state
);
698 static ssize_t
grant_speed_show(struct kobject
*kobj
, struct attribute
*attr
,
701 struct ldlm_pool
*pl
= container_of(kobj
, struct ldlm_pool
,
706 spin_lock(&pl
->pl_lock
);
707 /* serialize with ldlm_pool_recalc */
708 grant_speed
= atomic_read(&pl
->pl_grant_rate
) -
709 atomic_read(&pl
->pl_cancel_rate
);
710 spin_unlock(&pl
->pl_lock
);
711 return sprintf(buf
, "%d\n", grant_speed
);
713 LUSTRE_RO_ATTR(grant_speed
);
715 LDLM_POOL_SYSFS_READER_SHOW(grant_plan
, int);
716 LUSTRE_RO_ATTR(grant_plan
);
718 LDLM_POOL_SYSFS_READER_SHOW(recalc_period
, int);
719 LDLM_POOL_SYSFS_WRITER_STORE(recalc_period
, int);
720 LUSTRE_RW_ATTR(recalc_period
);
722 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(server_lock_volume
, u64
);
723 LUSTRE_RO_ATTR(server_lock_volume
);
725 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(limit
, atomic
);
726 LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(limit
, atomic
);
727 LUSTRE_RW_ATTR(limit
);
729 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(granted
, atomic
);
730 LUSTRE_RO_ATTR(granted
);
732 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(cancel_rate
, atomic
);
733 LUSTRE_RO_ATTR(cancel_rate
);
735 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(grant_rate
, atomic
);
736 LUSTRE_RO_ATTR(grant_rate
);
738 LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(lock_volume_factor
, atomic
);
739 LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(lock_volume_factor
, atomic
);
740 LUSTRE_RW_ATTR(lock_volume_factor
);
742 #define LDLM_POOL_ADD_VAR(name, var, ops) \
744 snprintf(var_name, MAX_STRING_SIZE, #name); \
745 pool_vars[0].data = var; \
746 pool_vars[0].fops = ops; \
747 ldebugfs_add_vars(pl->pl_debugfs_entry, pool_vars, NULL);\
750 /* These are for pools in /sys/fs/lustre/ldlm/namespaces/.../pool */
751 static struct attribute
*ldlm_pl_attrs
[] = {
752 &lustre_attr_grant_speed
.attr
,
753 &lustre_attr_grant_plan
.attr
,
754 &lustre_attr_recalc_period
.attr
,
755 &lustre_attr_server_lock_volume
.attr
,
756 &lustre_attr_limit
.attr
,
757 &lustre_attr_granted
.attr
,
758 &lustre_attr_cancel_rate
.attr
,
759 &lustre_attr_grant_rate
.attr
,
760 &lustre_attr_lock_volume_factor
.attr
,
764 static void ldlm_pl_release(struct kobject
*kobj
)
766 struct ldlm_pool
*pl
= container_of(kobj
, struct ldlm_pool
,
768 complete(&pl
->pl_kobj_unregister
);
771 static struct kobj_type ldlm_pl_ktype
= {
772 .default_attrs
= ldlm_pl_attrs
,
773 .sysfs_ops
= &lustre_sysfs_ops
,
774 .release
= ldlm_pl_release
,
777 static int ldlm_pool_sysfs_init(struct ldlm_pool
*pl
)
779 struct ldlm_namespace
*ns
= ldlm_pl2ns(pl
);
782 init_completion(&pl
->pl_kobj_unregister
);
783 err
= kobject_init_and_add(&pl
->pl_kobj
, &ldlm_pl_ktype
, &ns
->ns_kobj
,
789 static int ldlm_pool_debugfs_init(struct ldlm_pool
*pl
)
791 struct ldlm_namespace
*ns
= ldlm_pl2ns(pl
);
792 struct dentry
*debugfs_ns_parent
;
793 struct lprocfs_vars pool_vars
[2];
794 char *var_name
= NULL
;
797 var_name
= kzalloc(MAX_STRING_SIZE
+ 1, GFP_NOFS
);
801 debugfs_ns_parent
= ns
->ns_debugfs_entry
;
802 if (IS_ERR_OR_NULL(debugfs_ns_parent
)) {
803 CERROR("%s: debugfs entry is not initialized\n",
808 pl
->pl_debugfs_entry
= ldebugfs_register("pool", debugfs_ns_parent
,
810 if (IS_ERR(pl
->pl_debugfs_entry
)) {
811 CERROR("LdebugFS failed in ldlm-pool-init\n");
812 rc
= PTR_ERR(pl
->pl_debugfs_entry
);
813 pl
->pl_debugfs_entry
= NULL
;
817 var_name
[MAX_STRING_SIZE
] = '\0';
818 memset(pool_vars
, 0, sizeof(pool_vars
));
819 pool_vars
[0].name
= var_name
;
821 LDLM_POOL_ADD_VAR(state
, pl
, &lprocfs_pool_state_fops
);
823 pl
->pl_stats
= lprocfs_alloc_stats(LDLM_POOL_LAST_STAT
-
824 LDLM_POOL_FIRST_STAT
, 0);
830 lprocfs_counter_init(pl
->pl_stats
, LDLM_POOL_GRANTED_STAT
,
831 LPROCFS_CNTR_AVGMINMAX
| LPROCFS_CNTR_STDDEV
,
833 lprocfs_counter_init(pl
->pl_stats
, LDLM_POOL_GRANT_STAT
,
834 LPROCFS_CNTR_AVGMINMAX
| LPROCFS_CNTR_STDDEV
,
836 lprocfs_counter_init(pl
->pl_stats
, LDLM_POOL_CANCEL_STAT
,
837 LPROCFS_CNTR_AVGMINMAX
| LPROCFS_CNTR_STDDEV
,
839 lprocfs_counter_init(pl
->pl_stats
, LDLM_POOL_GRANT_RATE_STAT
,
840 LPROCFS_CNTR_AVGMINMAX
| LPROCFS_CNTR_STDDEV
,
841 "grant_rate", "locks/s");
842 lprocfs_counter_init(pl
->pl_stats
, LDLM_POOL_CANCEL_RATE_STAT
,
843 LPROCFS_CNTR_AVGMINMAX
| LPROCFS_CNTR_STDDEV
,
844 "cancel_rate", "locks/s");
845 lprocfs_counter_init(pl
->pl_stats
, LDLM_POOL_GRANT_PLAN_STAT
,
846 LPROCFS_CNTR_AVGMINMAX
| LPROCFS_CNTR_STDDEV
,
847 "grant_plan", "locks/s");
848 lprocfs_counter_init(pl
->pl_stats
, LDLM_POOL_SLV_STAT
,
849 LPROCFS_CNTR_AVGMINMAX
| LPROCFS_CNTR_STDDEV
,
851 lprocfs_counter_init(pl
->pl_stats
, LDLM_POOL_SHRINK_REQTD_STAT
,
852 LPROCFS_CNTR_AVGMINMAX
| LPROCFS_CNTR_STDDEV
,
853 "shrink_request", "locks");
854 lprocfs_counter_init(pl
->pl_stats
, LDLM_POOL_SHRINK_FREED_STAT
,
855 LPROCFS_CNTR_AVGMINMAX
| LPROCFS_CNTR_STDDEV
,
856 "shrink_freed", "locks");
857 lprocfs_counter_init(pl
->pl_stats
, LDLM_POOL_RECALC_STAT
,
858 LPROCFS_CNTR_AVGMINMAX
| LPROCFS_CNTR_STDDEV
,
859 "recalc_freed", "locks");
860 lprocfs_counter_init(pl
->pl_stats
, LDLM_POOL_TIMING_STAT
,
861 LPROCFS_CNTR_AVGMINMAX
| LPROCFS_CNTR_STDDEV
,
862 "recalc_timing", "sec");
863 rc
= ldebugfs_register_stats(pl
->pl_debugfs_entry
, "stats",
871 static void ldlm_pool_sysfs_fini(struct ldlm_pool
*pl
)
873 kobject_put(&pl
->pl_kobj
);
874 wait_for_completion(&pl
->pl_kobj_unregister
);
877 static void ldlm_pool_debugfs_fini(struct ldlm_pool
*pl
)
879 if (pl
->pl_stats
!= NULL
) {
880 lprocfs_free_stats(&pl
->pl_stats
);
883 if (pl
->pl_debugfs_entry
!= NULL
) {
884 ldebugfs_remove(&pl
->pl_debugfs_entry
);
885 pl
->pl_debugfs_entry
= NULL
;
889 int ldlm_pool_init(struct ldlm_pool
*pl
, struct ldlm_namespace
*ns
,
890 int idx
, ldlm_side_t client
)
894 spin_lock_init(&pl
->pl_lock
);
895 atomic_set(&pl
->pl_granted
, 0);
896 pl
->pl_recalc_time
= get_seconds();
897 atomic_set(&pl
->pl_lock_volume_factor
, 1);
899 atomic_set(&pl
->pl_grant_rate
, 0);
900 atomic_set(&pl
->pl_cancel_rate
, 0);
901 pl
->pl_grant_plan
= LDLM_POOL_GP(LDLM_POOL_HOST_L
);
903 snprintf(pl
->pl_name
, sizeof(pl
->pl_name
), "ldlm-pool-%s-%d",
904 ldlm_ns_name(ns
), idx
);
906 if (client
== LDLM_NAMESPACE_SERVER
) {
907 pl
->pl_ops
= &ldlm_srv_pool_ops
;
908 ldlm_pool_set_limit(pl
, LDLM_POOL_HOST_L
);
909 pl
->pl_recalc_period
= LDLM_POOL_SRV_DEF_RECALC_PERIOD
;
910 pl
->pl_server_lock_volume
= ldlm_pool_slv_max(LDLM_POOL_HOST_L
);
912 ldlm_pool_set_limit(pl
, 1);
913 pl
->pl_server_lock_volume
= 0;
914 pl
->pl_ops
= &ldlm_cli_pool_ops
;
915 pl
->pl_recalc_period
= LDLM_POOL_CLI_DEF_RECALC_PERIOD
;
917 pl
->pl_client_lock_volume
= 0;
918 rc
= ldlm_pool_debugfs_init(pl
);
922 rc
= ldlm_pool_sysfs_init(pl
);
926 CDEBUG(D_DLMTRACE
, "Lock pool %s is initialized\n", pl
->pl_name
);
930 EXPORT_SYMBOL(ldlm_pool_init
);
932 void ldlm_pool_fini(struct ldlm_pool
*pl
)
934 ldlm_pool_sysfs_fini(pl
);
935 ldlm_pool_debugfs_fini(pl
);
938 * Pool should not be used after this point. We can't free it here as
939 * it lives in struct ldlm_namespace, but still interested in catching
940 * any abnormal using cases.
942 POISON(pl
, 0x5a, sizeof(*pl
));
944 EXPORT_SYMBOL(ldlm_pool_fini
);
947 * Add new taken ldlm lock \a lock into pool \a pl accounting.
949 void ldlm_pool_add(struct ldlm_pool
*pl
, struct ldlm_lock
*lock
)
952 * FLOCK locks are special in a sense that they are almost never
953 * cancelled, instead special kind of lock is used to drop them.
954 * also there is no LRU for flock locks, so no point in tracking
957 if (lock
->l_resource
->lr_type
== LDLM_FLOCK
)
960 atomic_inc(&pl
->pl_granted
);
961 atomic_inc(&pl
->pl_grant_rate
);
962 lprocfs_counter_incr(pl
->pl_stats
, LDLM_POOL_GRANT_STAT
);
964 * Do not do pool recalc for client side as all locks which
965 * potentially may be canceled has already been packed into
966 * enqueue/cancel rpc. Also we do not want to run out of stack
967 * with too long call paths.
969 if (ns_is_server(ldlm_pl2ns(pl
)))
970 ldlm_pool_recalc(pl
);
972 EXPORT_SYMBOL(ldlm_pool_add
);
975 * Remove ldlm lock \a lock from pool \a pl accounting.
977 void ldlm_pool_del(struct ldlm_pool
*pl
, struct ldlm_lock
*lock
)
980 * Filter out FLOCK locks. Read above comment in ldlm_pool_add().
982 if (lock
->l_resource
->lr_type
== LDLM_FLOCK
)
985 LASSERT(atomic_read(&pl
->pl_granted
) > 0);
986 atomic_dec(&pl
->pl_granted
);
987 atomic_inc(&pl
->pl_cancel_rate
);
989 lprocfs_counter_incr(pl
->pl_stats
, LDLM_POOL_CANCEL_STAT
);
991 if (ns_is_server(ldlm_pl2ns(pl
)))
992 ldlm_pool_recalc(pl
);
994 EXPORT_SYMBOL(ldlm_pool_del
);
997 * Returns current \a pl SLV.
999 * \pre ->pl_lock is not locked.
1001 __u64
ldlm_pool_get_slv(struct ldlm_pool
*pl
)
1005 spin_lock(&pl
->pl_lock
);
1006 slv
= pl
->pl_server_lock_volume
;
1007 spin_unlock(&pl
->pl_lock
);
1010 EXPORT_SYMBOL(ldlm_pool_get_slv
);
1013 * Sets passed \a slv to \a pl.
1015 * \pre ->pl_lock is not locked.
1017 void ldlm_pool_set_slv(struct ldlm_pool
*pl
, __u64 slv
)
1019 spin_lock(&pl
->pl_lock
);
1020 pl
->pl_server_lock_volume
= slv
;
1021 spin_unlock(&pl
->pl_lock
);
1023 EXPORT_SYMBOL(ldlm_pool_set_slv
);
1026 * Returns current \a pl CLV.
1028 * \pre ->pl_lock is not locked.
1030 __u64
ldlm_pool_get_clv(struct ldlm_pool
*pl
)
1034 spin_lock(&pl
->pl_lock
);
1035 slv
= pl
->pl_client_lock_volume
;
1036 spin_unlock(&pl
->pl_lock
);
1039 EXPORT_SYMBOL(ldlm_pool_get_clv
);
1042 * Sets passed \a clv to \a pl.
1044 * \pre ->pl_lock is not locked.
1046 void ldlm_pool_set_clv(struct ldlm_pool
*pl
, __u64 clv
)
1048 spin_lock(&pl
->pl_lock
);
1049 pl
->pl_client_lock_volume
= clv
;
1050 spin_unlock(&pl
->pl_lock
);
1052 EXPORT_SYMBOL(ldlm_pool_set_clv
);
1055 * Returns current \a pl limit.
1057 __u32
ldlm_pool_get_limit(struct ldlm_pool
*pl
)
1059 return atomic_read(&pl
->pl_limit
);
1061 EXPORT_SYMBOL(ldlm_pool_get_limit
);
1064 * Sets passed \a limit to \a pl.
1066 void ldlm_pool_set_limit(struct ldlm_pool
*pl
, __u32 limit
)
1068 atomic_set(&pl
->pl_limit
, limit
);
1070 EXPORT_SYMBOL(ldlm_pool_set_limit
);
1073 * Returns current LVF from \a pl.
1075 __u32
ldlm_pool_get_lvf(struct ldlm_pool
*pl
)
1077 return atomic_read(&pl
->pl_lock_volume_factor
);
1079 EXPORT_SYMBOL(ldlm_pool_get_lvf
);
1081 static int ldlm_pool_granted(struct ldlm_pool
*pl
)
1083 return atomic_read(&pl
->pl_granted
);
1086 static struct ptlrpc_thread
*ldlm_pools_thread
;
1087 static struct completion ldlm_pools_comp
;
1090 * count locks from all namespaces (if possible). Returns number of
1093 static unsigned long ldlm_pools_count(ldlm_side_t client
, gfp_t gfp_mask
)
1095 int total
= 0, nr_ns
;
1096 struct ldlm_namespace
*ns
;
1097 struct ldlm_namespace
*ns_old
= NULL
; /* loop detection */
1100 if (client
== LDLM_NAMESPACE_CLIENT
&& !(gfp_mask
& __GFP_FS
))
1103 CDEBUG(D_DLMTRACE
, "Request to count %s locks from all pools\n",
1104 client
== LDLM_NAMESPACE_CLIENT
? "client" : "server");
1106 cookie
= cl_env_reenter();
1109 * Find out how many resources we may release.
1111 for (nr_ns
= ldlm_namespace_nr_read(client
);
1112 nr_ns
> 0; nr_ns
--) {
1113 mutex_lock(ldlm_namespace_lock(client
));
1114 if (list_empty(ldlm_namespace_list(client
))) {
1115 mutex_unlock(ldlm_namespace_lock(client
));
1116 cl_env_reexit(cookie
);
1119 ns
= ldlm_namespace_first_locked(client
);
1122 mutex_unlock(ldlm_namespace_lock(client
));
1126 if (ldlm_ns_empty(ns
)) {
1127 ldlm_namespace_move_to_inactive_locked(ns
, client
);
1128 mutex_unlock(ldlm_namespace_lock(client
));
1135 ldlm_namespace_get(ns
);
1136 ldlm_namespace_move_to_active_locked(ns
, client
);
1137 mutex_unlock(ldlm_namespace_lock(client
));
1138 total
+= ldlm_pool_shrink(&ns
->ns_pool
, 0, gfp_mask
);
1139 ldlm_namespace_put(ns
);
1142 cl_env_reexit(cookie
);
1146 static unsigned long ldlm_pools_scan(ldlm_side_t client
, int nr
, gfp_t gfp_mask
)
1148 unsigned long freed
= 0;
1150 struct ldlm_namespace
*ns
;
1153 if (client
== LDLM_NAMESPACE_CLIENT
&& !(gfp_mask
& __GFP_FS
))
1156 cookie
= cl_env_reenter();
1159 * Shrink at least ldlm_namespace_nr_read(client) namespaces.
1161 for (tmp
= nr_ns
= ldlm_namespace_nr_read(client
);
1163 int cancel
, nr_locks
;
1166 * Do not call shrink under ldlm_namespace_lock(client)
1168 mutex_lock(ldlm_namespace_lock(client
));
1169 if (list_empty(ldlm_namespace_list(client
))) {
1170 mutex_unlock(ldlm_namespace_lock(client
));
1173 ns
= ldlm_namespace_first_locked(client
);
1174 ldlm_namespace_get(ns
);
1175 ldlm_namespace_move_to_active_locked(ns
, client
);
1176 mutex_unlock(ldlm_namespace_lock(client
));
1178 nr_locks
= ldlm_pool_granted(&ns
->ns_pool
);
1180 * We use to shrink propotionally but with new shrinker API,
1181 * we lost the total number of freeable locks.
1183 cancel
= 1 + min_t(int, nr_locks
, nr
/ nr_ns
);
1184 freed
+= ldlm_pool_shrink(&ns
->ns_pool
, cancel
, gfp_mask
);
1185 ldlm_namespace_put(ns
);
1187 cl_env_reexit(cookie
);
1189 * we only decrease the SLV in server pools shrinker, return
1190 * SHRINK_STOP to kernel to avoid needless loop. LU-1128
1192 return (client
== LDLM_NAMESPACE_SERVER
) ? SHRINK_STOP
: freed
;
1195 static unsigned long ldlm_pools_srv_count(struct shrinker
*s
,
1196 struct shrink_control
*sc
)
1198 return ldlm_pools_count(LDLM_NAMESPACE_SERVER
, sc
->gfp_mask
);
1201 static unsigned long ldlm_pools_srv_scan(struct shrinker
*s
,
1202 struct shrink_control
*sc
)
1204 return ldlm_pools_scan(LDLM_NAMESPACE_SERVER
, sc
->nr_to_scan
,
1208 static unsigned long ldlm_pools_cli_count(struct shrinker
*s
,
1209 struct shrink_control
*sc
)
1211 return ldlm_pools_count(LDLM_NAMESPACE_CLIENT
, sc
->gfp_mask
);
1214 static unsigned long ldlm_pools_cli_scan(struct shrinker
*s
,
1215 struct shrink_control
*sc
)
1217 return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT
, sc
->nr_to_scan
,
1221 int ldlm_pools_recalc(ldlm_side_t client
)
1223 __u32 nr_l
= 0, nr_p
= 0, l
;
1224 struct ldlm_namespace
*ns
;
1225 struct ldlm_namespace
*ns_old
= NULL
;
1227 int time
= 50; /* seconds of sleep if no active namespaces */
1230 * No need to setup pool limit for client pools.
1232 if (client
== LDLM_NAMESPACE_SERVER
) {
1234 * Check all modest namespaces first.
1236 mutex_lock(ldlm_namespace_lock(client
));
1237 list_for_each_entry(ns
, ldlm_namespace_list(client
),
1239 if (ns
->ns_appetite
!= LDLM_NAMESPACE_MODEST
)
1242 l
= ldlm_pool_granted(&ns
->ns_pool
);
1247 * Set the modest pools limit equal to their avg granted
1250 l
+= dru(l
, LDLM_POOLS_MODEST_MARGIN_SHIFT
, 0);
1251 ldlm_pool_setup(&ns
->ns_pool
, l
);
1257 * Make sure that modest namespaces did not eat more that 2/3
1260 if (nr_l
>= 2 * (LDLM_POOL_HOST_L
/ 3)) {
1261 CWARN("\"Modest\" pools eat out 2/3 of server locks limit (%d of %lu). This means that you have too many clients for this amount of server RAM. Upgrade server!\n",
1262 nr_l
, LDLM_POOL_HOST_L
);
1267 * The rest is given to greedy namespaces.
1269 list_for_each_entry(ns
, ldlm_namespace_list(client
),
1271 if (!equal
&& ns
->ns_appetite
!= LDLM_NAMESPACE_GREEDY
)
1276 * In the case 2/3 locks are eaten out by
1277 * modest pools, we re-setup equal limit
1280 l
= LDLM_POOL_HOST_L
/
1281 ldlm_namespace_nr_read(client
);
1284 * All the rest of greedy pools will have
1285 * all locks in equal parts.
1287 l
= (LDLM_POOL_HOST_L
- nr_l
) /
1288 (ldlm_namespace_nr_read(client
) -
1291 ldlm_pool_setup(&ns
->ns_pool
, l
);
1293 mutex_unlock(ldlm_namespace_lock(client
));
1297 * Recalc at least ldlm_namespace_nr_read(client) namespaces.
1299 for (nr
= ldlm_namespace_nr_read(client
); nr
> 0; nr
--) {
1302 * Lock the list, get first @ns in the list, getref, move it
1303 * to the tail, unlock and call pool recalc. This way we avoid
1304 * calling recalc under @ns lock what is really good as we get
1305 * rid of potential deadlock on client nodes when canceling
1306 * locks synchronously.
1308 mutex_lock(ldlm_namespace_lock(client
));
1309 if (list_empty(ldlm_namespace_list(client
))) {
1310 mutex_unlock(ldlm_namespace_lock(client
));
1313 ns
= ldlm_namespace_first_locked(client
);
1315 if (ns_old
== ns
) { /* Full pass complete */
1316 mutex_unlock(ldlm_namespace_lock(client
));
1320 /* We got an empty namespace, need to move it back to inactive
1322 * The race with parallel resource creation is fine:
1323 * - If they do namespace_get before our check, we fail the
1324 * check and they move this item to the end of the list anyway
1325 * - If we do the check and then they do namespace_get, then
1326 * we move the namespace to inactive and they will move
1327 * it back to active (synchronised by the lock, so no clash
1330 if (ldlm_ns_empty(ns
)) {
1331 ldlm_namespace_move_to_inactive_locked(ns
, client
);
1332 mutex_unlock(ldlm_namespace_lock(client
));
1339 spin_lock(&ns
->ns_lock
);
1341 * skip ns which is being freed, and we don't want to increase
1342 * its refcount again, not even temporarily. bz21519 & LU-499.
1344 if (ns
->ns_stopping
) {
1348 ldlm_namespace_get(ns
);
1350 spin_unlock(&ns
->ns_lock
);
1352 ldlm_namespace_move_to_active_locked(ns
, client
);
1353 mutex_unlock(ldlm_namespace_lock(client
));
1356 * After setup is done - recalc the pool.
1359 int ttime
= ldlm_pool_recalc(&ns
->ns_pool
);
1364 ldlm_namespace_put(ns
);
1369 EXPORT_SYMBOL(ldlm_pools_recalc
);
1371 static int ldlm_pools_thread_main(void *arg
)
1373 struct ptlrpc_thread
*thread
= (struct ptlrpc_thread
*)arg
;
1376 thread_set_flags(thread
, SVC_RUNNING
);
1377 wake_up(&thread
->t_ctl_waitq
);
1379 CDEBUG(D_DLMTRACE
, "%s: pool thread starting, process %d\n",
1380 "ldlm_poold", current_pid());
1383 struct l_wait_info lwi
;
1386 * Recal all pools on this tick.
1388 s_time
= ldlm_pools_recalc(LDLM_NAMESPACE_SERVER
);
1389 c_time
= ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT
);
1392 * Wait until the next check time, or until we're
1395 lwi
= LWI_TIMEOUT(cfs_time_seconds(min(s_time
, c_time
)),
1397 l_wait_event(thread
->t_ctl_waitq
,
1398 thread_is_stopping(thread
) ||
1399 thread_is_event(thread
),
1402 if (thread_test_and_clear_flags(thread
, SVC_STOPPING
))
1404 thread_test_and_clear_flags(thread
, SVC_EVENT
);
1407 thread_set_flags(thread
, SVC_STOPPED
);
1408 wake_up(&thread
->t_ctl_waitq
);
1410 CDEBUG(D_DLMTRACE
, "%s: pool thread exiting, process %d\n",
1411 "ldlm_poold", current_pid());
1413 complete_and_exit(&ldlm_pools_comp
, 0);
1416 static int ldlm_pools_thread_start(void)
1418 struct l_wait_info lwi
= { 0 };
1419 struct task_struct
*task
;
1421 if (ldlm_pools_thread
!= NULL
)
1424 ldlm_pools_thread
= kzalloc(sizeof(*ldlm_pools_thread
), GFP_NOFS
);
1425 if (ldlm_pools_thread
== NULL
)
1428 init_completion(&ldlm_pools_comp
);
1429 init_waitqueue_head(&ldlm_pools_thread
->t_ctl_waitq
);
1431 task
= kthread_run(ldlm_pools_thread_main
, ldlm_pools_thread
,
1434 CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task
));
1435 kfree(ldlm_pools_thread
);
1436 ldlm_pools_thread
= NULL
;
1437 return PTR_ERR(task
);
1439 l_wait_event(ldlm_pools_thread
->t_ctl_waitq
,
1440 thread_is_running(ldlm_pools_thread
), &lwi
);
1444 static void ldlm_pools_thread_stop(void)
1446 if (ldlm_pools_thread
== NULL
)
1449 thread_set_flags(ldlm_pools_thread
, SVC_STOPPING
);
1450 wake_up(&ldlm_pools_thread
->t_ctl_waitq
);
1453 * Make sure that pools thread is finished before freeing @thread.
1454 * This fixes possible race and oops due to accessing freed memory
1457 wait_for_completion(&ldlm_pools_comp
);
1458 kfree(ldlm_pools_thread
);
1459 ldlm_pools_thread
= NULL
;
1462 static struct shrinker ldlm_pools_srv_shrinker
= {
1463 .count_objects
= ldlm_pools_srv_count
,
1464 .scan_objects
= ldlm_pools_srv_scan
,
1465 .seeks
= DEFAULT_SEEKS
,
1468 static struct shrinker ldlm_pools_cli_shrinker
= {
1469 .count_objects
= ldlm_pools_cli_count
,
1470 .scan_objects
= ldlm_pools_cli_scan
,
1471 .seeks
= DEFAULT_SEEKS
,
1474 int ldlm_pools_init(void)
1478 rc
= ldlm_pools_thread_start();
1480 register_shrinker(&ldlm_pools_srv_shrinker
);
1481 register_shrinker(&ldlm_pools_cli_shrinker
);
1485 EXPORT_SYMBOL(ldlm_pools_init
);
1487 void ldlm_pools_fini(void)
1489 unregister_shrinker(&ldlm_pools_srv_shrinker
);
1490 unregister_shrinker(&ldlm_pools_cli_shrinker
);
1491 ldlm_pools_thread_stop();
1493 EXPORT_SYMBOL(ldlm_pools_fini
);