Commit | Line | Data |
---|---|---|
d7e09d03 PT |
1 | /* |
2 | * GPL HEADER START | |
3 | * | |
4 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
5 | * | |
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. | |
9 | * | |
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). | |
15 | * | |
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 | |
19 | * | |
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 | |
22 | * have any questions. | |
23 | * | |
24 | * GPL HEADER END | |
25 | */ | |
26 | /* | |
27 | * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. | |
28 | * Use is subject to license terms. | |
29 | * | |
30 | * Copyright (c) 2010, 2012, Intel Corporation. | |
31 | */ | |
32 | /* | |
33 | * This file is part of Lustre, http://www.lustre.org/ | |
34 | * Lustre is a trademark of Sun Microsystems, Inc. | |
35 | * | |
36 | * lustre/ldlm/ldlm_pool.c | |
37 | * | |
38 | * Author: Yury Umanets <umka@clusterfs.com> | |
39 | */ | |
40 | ||
41 | /* | |
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. | |
50 | * | |
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. | |
56 | * | |
e7ddc48c AR |
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. | |
d7e09d03 PT |
61 | * |
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); | |
67 | * | |
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. | |
72 | * | |
e7ddc48c AR |
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). | |
d7e09d03 PT |
79 | * |
80 | * Glossary for terms used: | |
81 | * | |
82 | * pl_limit - Number of allowed locks in pool. Applies to server and client | |
83 | * side (tunable); | |
84 | * | |
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); | |
91 | * | |
92 | * As it may be seen from list above, we have few possible tunables which may | |
f2825e03 | 93 | * affect behavior much. They all may be modified via sysfs. However, they also |
d7e09d03 PT |
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 | |
f2825e03 | 96 | * be "constructed" via sysfs tunables. |
d7e09d03 PT |
97 | */ |
98 | ||
99 | #define DEBUG_SUBSYSTEM S_LDLM | |
100 | ||
e27db149 GKH |
101 | #include "../include/lustre_dlm.h" |
102 | #include "../include/cl_object.h" | |
103 | #include "../include/obd_class.h" | |
104 | #include "../include/obd_support.h" | |
d7e09d03 PT |
105 | #include "ldlm_internal.h" |
106 | ||
107 | ||
108 | /* | |
109 | * 50 ldlm locks for 1MB of RAM. | |
110 | */ | |
111 | #define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_CACHE_SHIFT)) * 50) | |
112 | ||
113 | /* | |
114 | * Maximal possible grant step plan in %. | |
115 | */ | |
116 | #define LDLM_POOL_MAX_GSP (30) | |
117 | ||
118 | /* | |
119 | * Minimal possible grant step plan in %. | |
120 | */ | |
121 | #define LDLM_POOL_MIN_GSP (1) | |
122 | ||
123 | /* | |
124 | * This controls the speed of reaching LDLM_POOL_MAX_GSP | |
125 | * with increasing thread period. | |
126 | */ | |
127 | #define LDLM_POOL_GSP_STEP_SHIFT (2) | |
128 | ||
129 | /* | |
130 | * LDLM_POOL_GSP% of all locks is default GP. | |
131 | */ | |
132 | #define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100) | |
133 | ||
134 | /* | |
135 | * Max age for locks on clients. | |
136 | */ | |
137 | #define LDLM_POOL_MAX_AGE (36000) | |
138 | ||
139 | /* | |
140 | * The granularity of SLV calculation. | |
141 | */ | |
142 | #define LDLM_POOL_SLV_SHIFT (10) | |
143 | ||
d7e09d03 PT |
144 | static inline __u64 dru(__u64 val, __u32 shift, int round_up) |
145 | { | |
146 | return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift; | |
147 | } | |
148 | ||
149 | static inline __u64 ldlm_pool_slv_max(__u32 L) | |
150 | { | |
151 | /* | |
152 | * Allow to have all locks for 1 client for 10 hrs. | |
153 | * Formula is the following: limit * 10h / 1 client. | |
154 | */ | |
155 | __u64 lim = (__u64)L * LDLM_POOL_MAX_AGE / 1; | |
156 | return lim; | |
157 | } | |
158 | ||
159 | static inline __u64 ldlm_pool_slv_min(__u32 L) | |
160 | { | |
161 | return 1; | |
162 | } | |
163 | ||
164 | enum { | |
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, | |
172 | LDLM_POOL_SLV_STAT, | |
173 | LDLM_POOL_SHRINK_REQTD_STAT, | |
174 | LDLM_POOL_SHRINK_FREED_STAT, | |
175 | LDLM_POOL_RECALC_STAT, | |
176 | LDLM_POOL_TIMING_STAT, | |
177 | LDLM_POOL_LAST_STAT | |
178 | }; | |
179 | ||
180 | static inline struct ldlm_namespace *ldlm_pl2ns(struct ldlm_pool *pl) | |
181 | { | |
182 | return container_of(pl, struct ldlm_namespace, ns_pool); | |
183 | } | |
184 | ||
185 | /** | |
186 | * Calculates suggested grant_step in % of available locks for passed | |
187 | * \a period. This is later used in grant_plan calculations. | |
188 | */ | |
189 | static inline int ldlm_pool_t2gsp(unsigned int t) | |
190 | { | |
191 | /* | |
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. | |
194 | * | |
195 | * How this will affect execution is the following: | |
196 | * | |
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; | |
203 | * | |
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 | |
209 | * plan is reached. | |
210 | */ | |
211 | return LDLM_POOL_MAX_GSP - | |
212 | ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >> | |
213 | (t >> LDLM_POOL_GSP_STEP_SHIFT)); | |
214 | } | |
215 | ||
216 | /** | |
217 | * Recalculates next grant limit on passed \a pl. | |
218 | * | |
219 | * \pre ->pl_lock is locked. | |
220 | */ | |
221 | static void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl) | |
222 | { | |
223 | int granted, grant_step, limit; | |
224 | ||
225 | limit = ldlm_pool_get_limit(pl); | |
226 | granted = atomic_read(&pl->pl_granted); | |
227 | ||
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; | |
234 | } | |
235 | ||
236 | /** | |
237 | * Recalculates next SLV on passed \a pl. | |
238 | * | |
239 | * \pre ->pl_lock is locked. | |
240 | */ | |
241 | static void ldlm_pool_recalc_slv(struct ldlm_pool *pl) | |
242 | { | |
243 | int granted; | |
244 | int grant_plan; | |
245 | int round_up; | |
246 | __u64 slv; | |
247 | __u64 slv_factor; | |
248 | __u64 grant_usage; | |
249 | __u32 limit; | |
250 | ||
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; | |
256 | ||
257 | grant_usage = max_t(int, limit - (granted - grant_plan), 1); | |
258 | ||
259 | /* | |
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. | |
266 | */ | |
1d06bb4e | 267 | slv_factor = grant_usage << LDLM_POOL_SLV_SHIFT; |
d7e09d03 PT |
268 | do_div(slv_factor, limit); |
269 | slv = slv * slv_factor; | |
270 | slv = dru(slv, LDLM_POOL_SLV_SHIFT, round_up); | |
271 | ||
8d2ff65d | 272 | if (slv > ldlm_pool_slv_max(limit)) |
d7e09d03 | 273 | slv = ldlm_pool_slv_max(limit); |
8d2ff65d | 274 | else if (slv < ldlm_pool_slv_min(limit)) |
d7e09d03 | 275 | slv = ldlm_pool_slv_min(limit); |
d7e09d03 PT |
276 | |
277 | pl->pl_server_lock_volume = slv; | |
278 | } | |
279 | ||
280 | /** | |
281 | * Recalculates next stats on passed \a pl. | |
282 | * | |
283 | * \pre ->pl_lock is locked. | |
284 | */ | |
285 | static void ldlm_pool_recalc_stats(struct ldlm_pool *pl) | |
286 | { | |
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); | |
292 | ||
293 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT, | |
294 | slv); | |
295 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT, | |
296 | granted); | |
297 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT, | |
298 | grant_rate); | |
299 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT, | |
300 | grant_plan); | |
301 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT, | |
302 | cancel_rate); | |
303 | } | |
304 | ||
305 | /** | |
306 | * Sets current SLV into obd accessible via ldlm_pl2ns(pl)->ns_obd. | |
307 | */ | |
308 | static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl) | |
309 | { | |
310 | struct obd_device *obd; | |
311 | ||
312 | /* | |
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. | |
318 | */ | |
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); | |
324 | } | |
325 | ||
326 | /** | |
327 | * Recalculates all pool fields on passed \a pl. | |
328 | * | |
329 | * \pre ->pl_lock is not locked. | |
330 | */ | |
331 | static int ldlm_srv_pool_recalc(struct ldlm_pool *pl) | |
332 | { | |
333 | time_t recalc_interval_sec; | |
d7e09d03 | 334 | |
7264b8a5 | 335 | recalc_interval_sec = get_seconds() - pl->pl_recalc_time; |
d7e09d03 | 336 | if (recalc_interval_sec < pl->pl_recalc_period) |
0a3bdb00 | 337 | return 0; |
d7e09d03 PT |
338 | |
339 | spin_lock(&pl->pl_lock); | |
7264b8a5 | 340 | recalc_interval_sec = get_seconds() - pl->pl_recalc_time; |
d7e09d03 PT |
341 | if (recalc_interval_sec < pl->pl_recalc_period) { |
342 | spin_unlock(&pl->pl_lock); | |
0a3bdb00 | 343 | return 0; |
d7e09d03 PT |
344 | } |
345 | /* | |
346 | * Recalc SLV after last period. This should be done | |
347 | * _before_ recalculating new grant plan. | |
348 | */ | |
349 | ldlm_pool_recalc_slv(pl); | |
350 | ||
351 | /* | |
352 | * Make sure that pool informed obd of last SLV changes. | |
353 | */ | |
354 | ldlm_srv_pool_push_slv(pl); | |
355 | ||
356 | /* | |
357 | * Update grant_plan for new period. | |
358 | */ | |
359 | ldlm_pool_recalc_grant_plan(pl); | |
360 | ||
7264b8a5 | 361 | pl->pl_recalc_time = get_seconds(); |
d7e09d03 PT |
362 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT, |
363 | recalc_interval_sec); | |
364 | spin_unlock(&pl->pl_lock); | |
0a3bdb00 | 365 | return 0; |
d7e09d03 PT |
366 | } |
367 | ||
368 | /** | |
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. | |
372 | * | |
373 | * Our goal here is to decrease SLV such a way that clients hold \a nr | |
374 | * locks smaller in next 10h. | |
375 | */ | |
376 | static int ldlm_srv_pool_shrink(struct ldlm_pool *pl, | |
5802572e | 377 | int nr, gfp_t gfp_mask) |
d7e09d03 PT |
378 | { |
379 | __u32 limit; | |
380 | ||
381 | /* | |
382 | * VM is asking how many entries may be potentially freed. | |
383 | */ | |
384 | if (nr == 0) | |
385 | return atomic_read(&pl->pl_granted); | |
386 | ||
387 | /* | |
388 | * Client already canceled locks but server is already in shrinker | |
389 | * and can't cancel anything. Let's catch this race. | |
390 | */ | |
391 | if (atomic_read(&pl->pl_granted) == 0) | |
0a3bdb00 | 392 | return 0; |
d7e09d03 PT |
393 | |
394 | spin_lock(&pl->pl_lock); | |
395 | ||
396 | /* | |
397 | * We want shrinker to possibly cause cancellation of @nr locks from | |
398 | * clients or grant approximately @nr locks smaller next intervals. | |
399 | * | |
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. | |
407 | */ | |
408 | if (nr < pl->pl_server_lock_volume) { | |
409 | pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr; | |
410 | } else { | |
411 | limit = ldlm_pool_get_limit(pl); | |
412 | pl->pl_server_lock_volume = ldlm_pool_slv_min(limit); | |
413 | } | |
414 | ||
415 | /* | |
416 | * Make sure that pool informed obd of last SLV changes. | |
417 | */ | |
418 | ldlm_srv_pool_push_slv(pl); | |
419 | spin_unlock(&pl->pl_lock); | |
420 | ||
421 | /* | |
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. | |
424 | */ | |
425 | return 0; | |
426 | } | |
427 | ||
428 | /** | |
429 | * Setup server side pool \a pl with passed \a limit. | |
430 | */ | |
431 | static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit) | |
432 | { | |
433 | struct obd_device *obd; | |
434 | ||
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); | |
441 | ||
442 | ldlm_pool_set_limit(pl, limit); | |
443 | return 0; | |
444 | } | |
445 | ||
446 | /** | |
447 | * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl. | |
448 | */ | |
449 | static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl) | |
450 | { | |
451 | struct obd_device *obd; | |
452 | ||
453 | /* | |
454 | * Get new SLV and Limit from obd which is updated with coming | |
455 | * RPCs. | |
456 | */ | |
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); | |
463 | } | |
464 | ||
465 | /** | |
466 | * Recalculates client size pool \a pl according to current SLV and Limit. | |
467 | */ | |
468 | static int ldlm_cli_pool_recalc(struct ldlm_pool *pl) | |
469 | { | |
470 | time_t recalc_interval_sec; | |
4d2c7b30 | 471 | int ret; |
d7e09d03 | 472 | |
7264b8a5 | 473 | recalc_interval_sec = get_seconds() - pl->pl_recalc_time; |
d7e09d03 | 474 | if (recalc_interval_sec < pl->pl_recalc_period) |
0a3bdb00 | 475 | return 0; |
d7e09d03 PT |
476 | |
477 | spin_lock(&pl->pl_lock); | |
478 | /* | |
479 | * Check if we need to recalc lists now. | |
480 | */ | |
7264b8a5 | 481 | recalc_interval_sec = get_seconds() - pl->pl_recalc_time; |
d7e09d03 PT |
482 | if (recalc_interval_sec < pl->pl_recalc_period) { |
483 | spin_unlock(&pl->pl_lock); | |
0a3bdb00 | 484 | return 0; |
d7e09d03 PT |
485 | } |
486 | ||
487 | /* | |
488 | * Make sure that pool knows last SLV and Limit from obd. | |
489 | */ | |
490 | ldlm_cli_pool_pop_slv(pl); | |
491 | ||
d7e09d03 PT |
492 | spin_unlock(&pl->pl_lock); |
493 | ||
494 | /* | |
495 | * Do not cancel locks in case lru resize is disabled for this ns. | |
496 | */ | |
4d2c7b30 LX |
497 | if (!ns_connect_lru_resize(ldlm_pl2ns(pl))) { |
498 | ret = 0; | |
499 | goto out; | |
500 | } | |
d7e09d03 PT |
501 | |
502 | /* | |
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. | |
507 | */ | |
4d2c7b30 LX |
508 | ret = ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC, LDLM_CANCEL_LRUR); |
509 | ||
510 | out: | |
511 | spin_lock(&pl->pl_lock); | |
512 | /* | |
513 | * Time of LRU resizing might be longer than period, | |
514 | * so update after LRU resizing rather than before it. | |
515 | */ | |
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); | |
520 | return ret; | |
d7e09d03 PT |
521 | } |
522 | ||
523 | /** | |
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. | |
527 | */ | |
528 | static int ldlm_cli_pool_shrink(struct ldlm_pool *pl, | |
5802572e | 529 | int nr, gfp_t gfp_mask) |
d7e09d03 PT |
530 | { |
531 | struct ldlm_namespace *ns; | |
cbc3769e | 532 | int unused; |
d7e09d03 PT |
533 | |
534 | ns = ldlm_pl2ns(pl); | |
535 | ||
536 | /* | |
537 | * Do not cancel locks in case lru resize is disabled for this ns. | |
538 | */ | |
539 | if (!ns_connect_lru_resize(ns)) | |
0a3bdb00 | 540 | return 0; |
d7e09d03 PT |
541 | |
542 | /* | |
543 | * Make sure that pool knows last SLV and Limit from obd. | |
544 | */ | |
545 | ldlm_cli_pool_pop_slv(pl); | |
546 | ||
547 | spin_lock(&ns->ns_lock); | |
548 | unused = ns->ns_nr_unused; | |
549 | spin_unlock(&ns->ns_lock); | |
550 | ||
cbc3769e PT |
551 | if (nr == 0) |
552 | return (unused / 100) * sysctl_vfs_cache_pressure; | |
553 | else | |
554 | return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_CANCEL_SHRINK); | |
d7e09d03 PT |
555 | } |
556 | ||
b9c98cfa | 557 | static const struct ldlm_pool_ops ldlm_srv_pool_ops = { |
d7e09d03 PT |
558 | .po_recalc = ldlm_srv_pool_recalc, |
559 | .po_shrink = ldlm_srv_pool_shrink, | |
560 | .po_setup = ldlm_srv_pool_setup | |
561 | }; | |
562 | ||
b9c98cfa | 563 | static const struct ldlm_pool_ops ldlm_cli_pool_ops = { |
d7e09d03 PT |
564 | .po_recalc = ldlm_cli_pool_recalc, |
565 | .po_shrink = ldlm_cli_pool_shrink | |
566 | }; | |
567 | ||
568 | /** | |
569 | * Pool recalc wrapper. Will call either client or server pool recalc callback | |
570 | * depending what pool \a pl is used. | |
571 | */ | |
572 | int ldlm_pool_recalc(struct ldlm_pool *pl) | |
573 | { | |
574 | time_t recalc_interval_sec; | |
575 | int count; | |
576 | ||
7264b8a5 | 577 | recalc_interval_sec = get_seconds() - pl->pl_recalc_time; |
d7e09d03 PT |
578 | if (recalc_interval_sec <= 0) |
579 | goto recalc; | |
580 | ||
581 | spin_lock(&pl->pl_lock); | |
d7e09d03 PT |
582 | if (recalc_interval_sec > 0) { |
583 | /* | |
584 | * Update pool statistics every 1s. | |
585 | */ | |
586 | ldlm_pool_recalc_stats(pl); | |
587 | ||
588 | /* | |
589 | * Zero out all rates and speed for the last period. | |
590 | */ | |
591 | atomic_set(&pl->pl_grant_rate, 0); | |
592 | atomic_set(&pl->pl_cancel_rate, 0); | |
593 | } | |
594 | spin_unlock(&pl->pl_lock); | |
595 | ||
596 | recalc: | |
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, | |
600 | count); | |
d7e09d03 | 601 | } |
7264b8a5 | 602 | recalc_interval_sec = pl->pl_recalc_time - get_seconds() + |
3eface59 | 603 | pl->pl_recalc_period; |
4d2c7b30 LX |
604 | if (recalc_interval_sec <= 0) { |
605 | /* Prevent too frequent recalculation. */ | |
606 | CDEBUG(D_DLMTRACE, "Negative interval(%ld), " | |
607 | "too short period(%ld)", | |
608 | recalc_interval_sec, | |
609 | pl->pl_recalc_period); | |
610 | recalc_interval_sec = 1; | |
611 | } | |
d7e09d03 | 612 | |
3eface59 | 613 | return recalc_interval_sec; |
d7e09d03 | 614 | } |
d7e09d03 | 615 | |
cbc3769e | 616 | /* |
d7e09d03 | 617 | * Pool shrink wrapper. Will call either client or server pool recalc callback |
cbc3769e PT |
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. | |
d7e09d03 PT |
620 | */ |
621 | int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, | |
5802572e | 622 | gfp_t gfp_mask) |
d7e09d03 PT |
623 | { |
624 | int cancel = 0; | |
625 | ||
626 | if (pl->pl_ops->po_shrink != NULL) { | |
627 | cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask); | |
628 | if (nr > 0) { | |
629 | lprocfs_counter_add(pl->pl_stats, | |
630 | LDLM_POOL_SHRINK_REQTD_STAT, | |
631 | nr); | |
632 | lprocfs_counter_add(pl->pl_stats, | |
633 | LDLM_POOL_SHRINK_FREED_STAT, | |
634 | cancel); | |
2d00bd17 JP |
635 | CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, shrunk %d\n", |
636 | pl->pl_name, nr, cancel); | |
d7e09d03 PT |
637 | } |
638 | } | |
639 | return cancel; | |
640 | } | |
641 | EXPORT_SYMBOL(ldlm_pool_shrink); | |
642 | ||
643 | /** | |
644 | * Pool setup wrapper. Will call either client or server pool recalc callback | |
645 | * depending what pool \a pl is used. | |
646 | * | |
647 | * Sets passed \a limit into pool \a pl. | |
648 | */ | |
649 | int ldlm_pool_setup(struct ldlm_pool *pl, int limit) | |
650 | { | |
651 | if (pl->pl_ops->po_setup != NULL) | |
e8291974 | 652 | return pl->pl_ops->po_setup(pl, limit); |
d7e09d03 PT |
653 | return 0; |
654 | } | |
655 | EXPORT_SYMBOL(ldlm_pool_setup); | |
656 | ||
73bb1da6 | 657 | static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused) |
d7e09d03 PT |
658 | { |
659 | int granted, grant_rate, cancel_rate, grant_step; | |
73bb1da6 PT |
660 | int grant_speed, grant_plan, lvf; |
661 | struct ldlm_pool *pl = m->private; | |
d7e09d03 PT |
662 | __u64 slv, clv; |
663 | __u32 limit; | |
664 | ||
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); | |
677 | ||
73bb1da6 | 678 | seq_printf(m, "LDLM pool state (%s):\n" |
b0f5aad5 GKH |
679 | " SLV: %llu\n" |
680 | " CLV: %llu\n" | |
73bb1da6 PT |
681 | " LVF: %d\n", |
682 | pl->pl_name, slv, clv, lvf); | |
d7e09d03 PT |
683 | |
684 | if (ns_is_server(ldlm_pl2ns(pl))) { | |
73bb1da6 PT |
685 | seq_printf(m, " GSP: %d%%\n" |
686 | " GP: %d\n", | |
687 | grant_step, grant_plan); | |
d7e09d03 | 688 | } |
2c2b7c05 HM |
689 | seq_printf(m, " GR: %d\n CR: %d\n GS: %d\n" |
690 | " G: %d\n L: %d\n", | |
73bb1da6 PT |
691 | grant_rate, cancel_rate, grant_speed, |
692 | granted, limit); | |
693 | ||
694 | return 0; | |
d7e09d03 | 695 | } |
73bb1da6 | 696 | LPROC_SEQ_FOPS_RO(lprocfs_pool_state); |
d7e09d03 | 697 | |
24b8c88a OD |
698 | static ssize_t grant_speed_show(struct kobject *kobj, struct attribute *attr, |
699 | char *buf) | |
d7e09d03 | 700 | { |
24b8c88a OD |
701 | struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool, |
702 | pl_kobj); | |
703 | ||
d7e09d03 PT |
704 | int grant_speed; |
705 | ||
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); | |
24b8c88a | 711 | return sprintf(buf, "%d\n", grant_speed); |
d7e09d03 | 712 | } |
24b8c88a | 713 | LUSTRE_RO_ATTR(grant_speed); |
d7e09d03 | 714 | |
24b8c88a OD |
715 | LDLM_POOL_SYSFS_READER_SHOW(grant_plan, int); |
716 | LUSTRE_RO_ATTR(grant_plan); | |
73bb1da6 | 717 | |
24b8c88a OD |
718 | LDLM_POOL_SYSFS_READER_SHOW(recalc_period, int); |
719 | LDLM_POOL_SYSFS_WRITER_STORE(recalc_period, int); | |
720 | LUSTRE_RW_ATTR(recalc_period); | |
73bb1da6 | 721 | |
24b8c88a OD |
722 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(server_lock_volume, u64); |
723 | LUSTRE_RO_ATTR(server_lock_volume); | |
724 | ||
725 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(limit, atomic); | |
726 | LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(limit, atomic); | |
727 | LUSTRE_RW_ATTR(limit); | |
728 | ||
729 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(granted, atomic); | |
730 | LUSTRE_RO_ATTR(granted); | |
731 | ||
732 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(cancel_rate, atomic); | |
733 | LUSTRE_RO_ATTR(cancel_rate); | |
73bb1da6 | 734 | |
24b8c88a OD |
735 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(grant_rate, atomic); |
736 | LUSTRE_RO_ATTR(grant_rate); | |
73bb1da6 | 737 | |
24b8c88a OD |
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); | |
73bb1da6 PT |
741 | |
742 | #define LDLM_POOL_ADD_VAR(name, var, ops) \ | |
743 | do { \ | |
744 | snprintf(var_name, MAX_STRING_SIZE, #name); \ | |
745 | pool_vars[0].data = var; \ | |
746 | pool_vars[0].fops = ops; \ | |
700815d4 | 747 | ldebugfs_add_vars(pl->pl_debugfs_entry, pool_vars, NULL);\ |
73bb1da6 | 748 | } while (0) |
d7e09d03 | 749 | |
f2825e03 OD |
750 | /* These are for pools in /sys/fs/lustre/ldlm/namespaces/.../pool */ |
751 | static struct attribute *ldlm_pl_attrs[] = { | |
24b8c88a OD |
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, | |
f2825e03 OD |
761 | NULL, |
762 | }; | |
763 | ||
764 | static void ldlm_pl_release(struct kobject *kobj) | |
765 | { | |
766 | struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool, | |
767 | pl_kobj); | |
768 | complete(&pl->pl_kobj_unregister); | |
769 | } | |
770 | ||
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, | |
775 | }; | |
776 | ||
777 | static int ldlm_pool_sysfs_init(struct ldlm_pool *pl) | |
778 | { | |
779 | struct ldlm_namespace *ns = ldlm_pl2ns(pl); | |
780 | int err; | |
781 | ||
782 | init_completion(&pl->pl_kobj_unregister); | |
783 | err = kobject_init_and_add(&pl->pl_kobj, &ldlm_pl_ktype, &ns->ns_kobj, | |
784 | "pool"); | |
785 | ||
786 | return err; | |
787 | } | |
788 | ||
700815d4 | 789 | static int ldlm_pool_debugfs_init(struct ldlm_pool *pl) |
d7e09d03 PT |
790 | { |
791 | struct ldlm_namespace *ns = ldlm_pl2ns(pl); | |
700815d4 | 792 | struct dentry *debugfs_ns_parent; |
d7e09d03 PT |
793 | struct lprocfs_vars pool_vars[2]; |
794 | char *var_name = NULL; | |
795 | int rc = 0; | |
d7e09d03 | 796 | |
352f7891 | 797 | var_name = kzalloc(MAX_STRING_SIZE + 1, GFP_NOFS); |
d7e09d03 | 798 | if (!var_name) |
0a3bdb00 | 799 | return -ENOMEM; |
d7e09d03 | 800 | |
700815d4 DE |
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", | |
d7e09d03 | 804 | ldlm_ns_name(ns)); |
d1c0d446 JL |
805 | rc = -EINVAL; |
806 | goto out_free_name; | |
d7e09d03 | 807 | } |
700815d4 DE |
808 | pl->pl_debugfs_entry = ldebugfs_register("pool", debugfs_ns_parent, |
809 | NULL, NULL); | |
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; | |
d1c0d446 | 814 | goto out_free_name; |
d7e09d03 PT |
815 | } |
816 | ||
817 | var_name[MAX_STRING_SIZE] = '\0'; | |
818 | memset(pool_vars, 0, sizeof(pool_vars)); | |
819 | pool_vars[0].name = var_name; | |
820 | ||
700815d4 | 821 | LDLM_POOL_ADD_VAR(state, pl, &lprocfs_pool_state_fops); |
d7e09d03 PT |
822 | |
823 | pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT - | |
824 | LDLM_POOL_FIRST_STAT, 0); | |
d1c0d446 JL |
825 | if (!pl->pl_stats) { |
826 | rc = -ENOMEM; | |
827 | goto out_free_name; | |
828 | } | |
d7e09d03 PT |
829 | |
830 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT, | |
831 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
832 | "granted", "locks"); | |
833 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT, | |
834 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
835 | "grant", "locks"); | |
836 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT, | |
837 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
838 | "cancel", "locks"); | |
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, | |
850 | "slv", "slv"); | |
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"); | |
700815d4 DE |
863 | rc = ldebugfs_register_stats(pl->pl_debugfs_entry, "stats", |
864 | pl->pl_stats); | |
d7e09d03 | 865 | |
d7e09d03 | 866 | out_free_name: |
352f7891 | 867 | kfree(var_name); |
d7e09d03 PT |
868 | return rc; |
869 | } | |
870 | ||
f2825e03 OD |
871 | static void ldlm_pool_sysfs_fini(struct ldlm_pool *pl) |
872 | { | |
873 | kobject_put(&pl->pl_kobj); | |
874 | wait_for_completion(&pl->pl_kobj_unregister); | |
875 | } | |
876 | ||
700815d4 | 877 | static void ldlm_pool_debugfs_fini(struct ldlm_pool *pl) |
d7e09d03 PT |
878 | { |
879 | if (pl->pl_stats != NULL) { | |
880 | lprocfs_free_stats(&pl->pl_stats); | |
881 | pl->pl_stats = NULL; | |
882 | } | |
700815d4 DE |
883 | if (pl->pl_debugfs_entry != NULL) { |
884 | ldebugfs_remove(&pl->pl_debugfs_entry); | |
885 | pl->pl_debugfs_entry = NULL; | |
d7e09d03 PT |
886 | } |
887 | } | |
888 | ||
889 | int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns, | |
890 | int idx, ldlm_side_t client) | |
891 | { | |
892 | int rc; | |
d7e09d03 PT |
893 | |
894 | spin_lock_init(&pl->pl_lock); | |
895 | atomic_set(&pl->pl_granted, 0); | |
7264b8a5 | 896 | pl->pl_recalc_time = get_seconds(); |
d7e09d03 PT |
897 | atomic_set(&pl->pl_lock_volume_factor, 1); |
898 | ||
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); | |
902 | ||
903 | snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d", | |
904 | ldlm_ns_name(ns), idx); | |
905 | ||
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); | |
911 | } else { | |
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; | |
916 | } | |
917 | pl->pl_client_lock_volume = 0; | |
700815d4 | 918 | rc = ldlm_pool_debugfs_init(pl); |
d7e09d03 | 919 | if (rc) |
0a3bdb00 | 920 | return rc; |
d7e09d03 | 921 | |
f2825e03 OD |
922 | rc = ldlm_pool_sysfs_init(pl); |
923 | if (rc) | |
924 | return rc; | |
925 | ||
d7e09d03 PT |
926 | CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name); |
927 | ||
0a3bdb00 | 928 | return rc; |
d7e09d03 PT |
929 | } |
930 | EXPORT_SYMBOL(ldlm_pool_init); | |
931 | ||
932 | void ldlm_pool_fini(struct ldlm_pool *pl) | |
933 | { | |
f2825e03 | 934 | ldlm_pool_sysfs_fini(pl); |
700815d4 | 935 | ldlm_pool_debugfs_fini(pl); |
d7e09d03 PT |
936 | |
937 | /* | |
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. | |
941 | */ | |
942 | POISON(pl, 0x5a, sizeof(*pl)); | |
d7e09d03 PT |
943 | } |
944 | EXPORT_SYMBOL(ldlm_pool_fini); | |
945 | ||
946 | /** | |
947 | * Add new taken ldlm lock \a lock into pool \a pl accounting. | |
948 | */ | |
949 | void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock) | |
950 | { | |
951 | /* | |
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 | |
955 | * them anyway. | |
956 | */ | |
957 | if (lock->l_resource->lr_type == LDLM_FLOCK) | |
958 | return; | |
959 | ||
960 | atomic_inc(&pl->pl_granted); | |
961 | atomic_inc(&pl->pl_grant_rate); | |
962 | lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT); | |
963 | /* | |
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. | |
968 | */ | |
969 | if (ns_is_server(ldlm_pl2ns(pl))) | |
970 | ldlm_pool_recalc(pl); | |
971 | } | |
972 | EXPORT_SYMBOL(ldlm_pool_add); | |
973 | ||
974 | /** | |
975 | * Remove ldlm lock \a lock from pool \a pl accounting. | |
976 | */ | |
977 | void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock) | |
978 | { | |
979 | /* | |
980 | * Filter out FLOCK locks. Read above comment in ldlm_pool_add(). | |
981 | */ | |
982 | if (lock->l_resource->lr_type == LDLM_FLOCK) | |
983 | return; | |
984 | ||
985 | LASSERT(atomic_read(&pl->pl_granted) > 0); | |
986 | atomic_dec(&pl->pl_granted); | |
987 | atomic_inc(&pl->pl_cancel_rate); | |
988 | ||
989 | lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT); | |
990 | ||
991 | if (ns_is_server(ldlm_pl2ns(pl))) | |
992 | ldlm_pool_recalc(pl); | |
993 | } | |
994 | EXPORT_SYMBOL(ldlm_pool_del); | |
995 | ||
996 | /** | |
997 | * Returns current \a pl SLV. | |
998 | * | |
999 | * \pre ->pl_lock is not locked. | |
1000 | */ | |
1001 | __u64 ldlm_pool_get_slv(struct ldlm_pool *pl) | |
1002 | { | |
1003 | __u64 slv; | |
902f3bb1 | 1004 | |
d7e09d03 PT |
1005 | spin_lock(&pl->pl_lock); |
1006 | slv = pl->pl_server_lock_volume; | |
1007 | spin_unlock(&pl->pl_lock); | |
1008 | return slv; | |
1009 | } | |
1010 | EXPORT_SYMBOL(ldlm_pool_get_slv); | |
1011 | ||
1012 | /** | |
1013 | * Sets passed \a slv to \a pl. | |
1014 | * | |
1015 | * \pre ->pl_lock is not locked. | |
1016 | */ | |
1017 | void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv) | |
1018 | { | |
1019 | spin_lock(&pl->pl_lock); | |
1020 | pl->pl_server_lock_volume = slv; | |
1021 | spin_unlock(&pl->pl_lock); | |
1022 | } | |
1023 | EXPORT_SYMBOL(ldlm_pool_set_slv); | |
1024 | ||
1025 | /** | |
1026 | * Returns current \a pl CLV. | |
1027 | * | |
1028 | * \pre ->pl_lock is not locked. | |
1029 | */ | |
1030 | __u64 ldlm_pool_get_clv(struct ldlm_pool *pl) | |
1031 | { | |
1032 | __u64 slv; | |
902f3bb1 | 1033 | |
d7e09d03 PT |
1034 | spin_lock(&pl->pl_lock); |
1035 | slv = pl->pl_client_lock_volume; | |
1036 | spin_unlock(&pl->pl_lock); | |
1037 | return slv; | |
1038 | } | |
1039 | EXPORT_SYMBOL(ldlm_pool_get_clv); | |
1040 | ||
1041 | /** | |
1042 | * Sets passed \a clv to \a pl. | |
1043 | * | |
1044 | * \pre ->pl_lock is not locked. | |
1045 | */ | |
1046 | void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv) | |
1047 | { | |
1048 | spin_lock(&pl->pl_lock); | |
1049 | pl->pl_client_lock_volume = clv; | |
1050 | spin_unlock(&pl->pl_lock); | |
1051 | } | |
1052 | EXPORT_SYMBOL(ldlm_pool_set_clv); | |
1053 | ||
1054 | /** | |
1055 | * Returns current \a pl limit. | |
1056 | */ | |
1057 | __u32 ldlm_pool_get_limit(struct ldlm_pool *pl) | |
1058 | { | |
1059 | return atomic_read(&pl->pl_limit); | |
1060 | } | |
1061 | EXPORT_SYMBOL(ldlm_pool_get_limit); | |
1062 | ||
1063 | /** | |
1064 | * Sets passed \a limit to \a pl. | |
1065 | */ | |
1066 | void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit) | |
1067 | { | |
1068 | atomic_set(&pl->pl_limit, limit); | |
1069 | } | |
1070 | EXPORT_SYMBOL(ldlm_pool_set_limit); | |
1071 | ||
1072 | /** | |
1073 | * Returns current LVF from \a pl. | |
1074 | */ | |
1075 | __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl) | |
1076 | { | |
1077 | return atomic_read(&pl->pl_lock_volume_factor); | |
1078 | } | |
1079 | EXPORT_SYMBOL(ldlm_pool_get_lvf); | |
1080 | ||
1081 | static int ldlm_pool_granted(struct ldlm_pool *pl) | |
1082 | { | |
1083 | return atomic_read(&pl->pl_granted); | |
1084 | } | |
1085 | ||
1086 | static struct ptlrpc_thread *ldlm_pools_thread; | |
d7e09d03 PT |
1087 | static struct completion ldlm_pools_comp; |
1088 | ||
1089 | /* | |
cbc3769e PT |
1090 | * count locks from all namespaces (if possible). Returns number of |
1091 | * cached locks. | |
d7e09d03 | 1092 | */ |
5802572e | 1093 | static unsigned long ldlm_pools_count(ldlm_side_t client, gfp_t gfp_mask) |
d7e09d03 | 1094 | { |
cbc3769e | 1095 | int total = 0, nr_ns; |
d7e09d03 | 1096 | struct ldlm_namespace *ns; |
91a50030 | 1097 | struct ldlm_namespace *ns_old = NULL; /* loop detection */ |
d7e09d03 PT |
1098 | void *cookie; |
1099 | ||
cbc3769e PT |
1100 | if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS)) |
1101 | return 0; | |
d7e09d03 | 1102 | |
cbc3769e PT |
1103 | CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n", |
1104 | client == LDLM_NAMESPACE_CLIENT ? "client" : "server"); | |
d7e09d03 PT |
1105 | |
1106 | cookie = cl_env_reenter(); | |
1107 | ||
1108 | /* | |
1109 | * Find out how many resources we may release. | |
1110 | */ | |
91a50030 | 1111 | for (nr_ns = ldlm_namespace_nr_read(client); |
cbc3769e | 1112 | nr_ns > 0; nr_ns--) { |
d7e09d03 PT |
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); | |
1117 | return 0; | |
1118 | } | |
1119 | ns = ldlm_namespace_first_locked(client); | |
91a50030 OD |
1120 | |
1121 | if (ns == ns_old) { | |
1122 | mutex_unlock(ldlm_namespace_lock(client)); | |
1123 | break; | |
1124 | } | |
1125 | ||
1126 | if (ldlm_ns_empty(ns)) { | |
1127 | ldlm_namespace_move_to_inactive_locked(ns, client); | |
1128 | mutex_unlock(ldlm_namespace_lock(client)); | |
1129 | continue; | |
1130 | } | |
1131 | ||
1132 | if (ns_old == NULL) | |
1133 | ns_old = ns; | |
1134 | ||
d7e09d03 | 1135 | ldlm_namespace_get(ns); |
91a50030 | 1136 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
1137 | mutex_unlock(ldlm_namespace_lock(client)); |
1138 | total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask); | |
1139 | ldlm_namespace_put(ns); | |
1140 | } | |
1141 | ||
cbc3769e PT |
1142 | cl_env_reexit(cookie); |
1143 | return total; | |
1144 | } | |
1145 | ||
5802572e | 1146 | static unsigned long ldlm_pools_scan(ldlm_side_t client, int nr, gfp_t gfp_mask) |
cbc3769e PT |
1147 | { |
1148 | unsigned long freed = 0; | |
1149 | int tmp, nr_ns; | |
1150 | struct ldlm_namespace *ns; | |
1151 | void *cookie; | |
1152 | ||
1153 | if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS)) | |
1154 | return -1; | |
1155 | ||
1156 | cookie = cl_env_reenter(); | |
d7e09d03 PT |
1157 | |
1158 | /* | |
cbc3769e | 1159 | * Shrink at least ldlm_namespace_nr_read(client) namespaces. |
d7e09d03 | 1160 | */ |
cbc3769e PT |
1161 | for (tmp = nr_ns = ldlm_namespace_nr_read(client); |
1162 | tmp > 0; tmp--) { | |
d7e09d03 PT |
1163 | int cancel, nr_locks; |
1164 | ||
1165 | /* | |
1166 | * Do not call shrink under ldlm_namespace_lock(client) | |
1167 | */ | |
1168 | mutex_lock(ldlm_namespace_lock(client)); | |
1169 | if (list_empty(ldlm_namespace_list(client))) { | |
1170 | mutex_unlock(ldlm_namespace_lock(client)); | |
d7e09d03 PT |
1171 | break; |
1172 | } | |
1173 | ns = ldlm_namespace_first_locked(client); | |
1174 | ldlm_namespace_get(ns); | |
91a50030 | 1175 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
1176 | mutex_unlock(ldlm_namespace_lock(client)); |
1177 | ||
1178 | nr_locks = ldlm_pool_granted(&ns->ns_pool); | |
cbc3769e PT |
1179 | /* |
1180 | * We use to shrink propotionally but with new shrinker API, | |
1181 | * we lost the total number of freeable locks. | |
1182 | */ | |
1183 | cancel = 1 + min_t(int, nr_locks, nr / nr_ns); | |
1184 | freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask); | |
d7e09d03 PT |
1185 | ldlm_namespace_put(ns); |
1186 | } | |
1187 | cl_env_reexit(cookie); | |
cbc3769e PT |
1188 | /* |
1189 | * we only decrease the SLV in server pools shrinker, return | |
1190 | * SHRINK_STOP to kernel to avoid needless loop. LU-1128 | |
1191 | */ | |
1192 | return (client == LDLM_NAMESPACE_SERVER) ? SHRINK_STOP : freed; | |
1193 | } | |
1194 | ||
e7ddc48c AR |
1195 | static unsigned long ldlm_pools_srv_count(struct shrinker *s, |
1196 | struct shrink_control *sc) | |
cbc3769e PT |
1197 | { |
1198 | return ldlm_pools_count(LDLM_NAMESPACE_SERVER, sc->gfp_mask); | |
d7e09d03 PT |
1199 | } |
1200 | ||
e7ddc48c AR |
1201 | static unsigned long ldlm_pools_srv_scan(struct shrinker *s, |
1202 | struct shrink_control *sc) | |
d7e09d03 | 1203 | { |
cbc3769e PT |
1204 | return ldlm_pools_scan(LDLM_NAMESPACE_SERVER, sc->nr_to_scan, |
1205 | sc->gfp_mask); | |
d7e09d03 PT |
1206 | } |
1207 | ||
e7ddc48c AR |
1208 | static unsigned long ldlm_pools_cli_count(struct shrinker *s, |
1209 | struct shrink_control *sc) | |
d7e09d03 | 1210 | { |
cbc3769e PT |
1211 | return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask); |
1212 | } | |
1213 | ||
e7ddc48c AR |
1214 | static unsigned long ldlm_pools_cli_scan(struct shrinker *s, |
1215 | struct shrink_control *sc) | |
cbc3769e PT |
1216 | { |
1217 | return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan, | |
1218 | sc->gfp_mask); | |
d7e09d03 PT |
1219 | } |
1220 | ||
3eface59 | 1221 | int ldlm_pools_recalc(ldlm_side_t client) |
d7e09d03 PT |
1222 | { |
1223 | __u32 nr_l = 0, nr_p = 0, l; | |
1224 | struct ldlm_namespace *ns; | |
91a50030 | 1225 | struct ldlm_namespace *ns_old = NULL; |
d7e09d03 | 1226 | int nr, equal = 0; |
3eface59 | 1227 | int time = 50; /* seconds of sleep if no active namespaces */ |
d7e09d03 PT |
1228 | |
1229 | /* | |
1230 | * No need to setup pool limit for client pools. | |
1231 | */ | |
1232 | if (client == LDLM_NAMESPACE_SERVER) { | |
1233 | /* | |
1234 | * Check all modest namespaces first. | |
1235 | */ | |
1236 | mutex_lock(ldlm_namespace_lock(client)); | |
1237 | list_for_each_entry(ns, ldlm_namespace_list(client), | |
9d0b2b7a | 1238 | ns_list_chain) { |
d7e09d03 PT |
1239 | if (ns->ns_appetite != LDLM_NAMESPACE_MODEST) |
1240 | continue; | |
1241 | ||
1242 | l = ldlm_pool_granted(&ns->ns_pool); | |
1243 | if (l == 0) | |
1244 | l = 1; | |
1245 | ||
1246 | /* | |
1247 | * Set the modest pools limit equal to their avg granted | |
1248 | * locks + ~6%. | |
1249 | */ | |
1250 | l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0); | |
1251 | ldlm_pool_setup(&ns->ns_pool, l); | |
1252 | nr_l += l; | |
1253 | nr_p++; | |
1254 | } | |
1255 | ||
1256 | /* | |
1257 | * Make sure that modest namespaces did not eat more that 2/3 | |
1258 | * of limit. | |
1259 | */ | |
1260 | if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) { | |
2d00bd17 JP |
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); | |
d7e09d03 PT |
1263 | equal = 1; |
1264 | } | |
1265 | ||
1266 | /* | |
1267 | * The rest is given to greedy namespaces. | |
1268 | */ | |
1269 | list_for_each_entry(ns, ldlm_namespace_list(client), | |
e9e2fa69 | 1270 | ns_list_chain) { |
d7e09d03 PT |
1271 | if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY) |
1272 | continue; | |
1273 | ||
1274 | if (equal) { | |
1275 | /* | |
1276 | * In the case 2/3 locks are eaten out by | |
1277 | * modest pools, we re-setup equal limit | |
1278 | * for _all_ pools. | |
1279 | */ | |
1280 | l = LDLM_POOL_HOST_L / | |
91a50030 | 1281 | ldlm_namespace_nr_read(client); |
d7e09d03 PT |
1282 | } else { |
1283 | /* | |
1284 | * All the rest of greedy pools will have | |
1285 | * all locks in equal parts. | |
1286 | */ | |
1287 | l = (LDLM_POOL_HOST_L - nr_l) / | |
91a50030 | 1288 | (ldlm_namespace_nr_read(client) - |
d7e09d03 PT |
1289 | nr_p); |
1290 | } | |
1291 | ldlm_pool_setup(&ns->ns_pool, l); | |
1292 | } | |
1293 | mutex_unlock(ldlm_namespace_lock(client)); | |
1294 | } | |
1295 | ||
1296 | /* | |
cbc3769e | 1297 | * Recalc at least ldlm_namespace_nr_read(client) namespaces. |
d7e09d03 | 1298 | */ |
91a50030 | 1299 | for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) { |
d7e09d03 PT |
1300 | int skip; |
1301 | /* | |
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. | |
1307 | */ | |
1308 | mutex_lock(ldlm_namespace_lock(client)); | |
1309 | if (list_empty(ldlm_namespace_list(client))) { | |
1310 | mutex_unlock(ldlm_namespace_lock(client)); | |
1311 | break; | |
1312 | } | |
1313 | ns = ldlm_namespace_first_locked(client); | |
1314 | ||
91a50030 OD |
1315 | if (ns_old == ns) { /* Full pass complete */ |
1316 | mutex_unlock(ldlm_namespace_lock(client)); | |
1317 | break; | |
1318 | } | |
1319 | ||
1320 | /* We got an empty namespace, need to move it back to inactive | |
1321 | * list. | |
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 | |
1328 | * there). | |
1329 | */ | |
1330 | if (ldlm_ns_empty(ns)) { | |
1331 | ldlm_namespace_move_to_inactive_locked(ns, client); | |
1332 | mutex_unlock(ldlm_namespace_lock(client)); | |
1333 | continue; | |
1334 | } | |
1335 | ||
1336 | if (ns_old == NULL) | |
1337 | ns_old = ns; | |
1338 | ||
d7e09d03 PT |
1339 | spin_lock(&ns->ns_lock); |
1340 | /* | |
1341 | * skip ns which is being freed, and we don't want to increase | |
1342 | * its refcount again, not even temporarily. bz21519 & LU-499. | |
1343 | */ | |
1344 | if (ns->ns_stopping) { | |
1345 | skip = 1; | |
1346 | } else { | |
1347 | skip = 0; | |
1348 | ldlm_namespace_get(ns); | |
1349 | } | |
1350 | spin_unlock(&ns->ns_lock); | |
1351 | ||
91a50030 | 1352 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
1353 | mutex_unlock(ldlm_namespace_lock(client)); |
1354 | ||
1355 | /* | |
1356 | * After setup is done - recalc the pool. | |
1357 | */ | |
1358 | if (!skip) { | |
3eface59 OD |
1359 | int ttime = ldlm_pool_recalc(&ns->ns_pool); |
1360 | ||
1361 | if (ttime < time) | |
1362 | time = ttime; | |
1363 | ||
d7e09d03 PT |
1364 | ldlm_namespace_put(ns); |
1365 | } | |
1366 | } | |
3eface59 | 1367 | return time; |
d7e09d03 PT |
1368 | } |
1369 | EXPORT_SYMBOL(ldlm_pools_recalc); | |
1370 | ||
1371 | static int ldlm_pools_thread_main(void *arg) | |
1372 | { | |
1373 | struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg; | |
3eface59 | 1374 | int s_time, c_time; |
d7e09d03 PT |
1375 | |
1376 | thread_set_flags(thread, SVC_RUNNING); | |
1377 | wake_up(&thread->t_ctl_waitq); | |
1378 | ||
1379 | CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n", | |
1380 | "ldlm_poold", current_pid()); | |
1381 | ||
1382 | while (1) { | |
1383 | struct l_wait_info lwi; | |
1384 | ||
1385 | /* | |
1386 | * Recal all pools on this tick. | |
1387 | */ | |
3eface59 OD |
1388 | s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER); |
1389 | c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT); | |
d7e09d03 PT |
1390 | |
1391 | /* | |
1392 | * Wait until the next check time, or until we're | |
1393 | * stopped. | |
1394 | */ | |
3eface59 | 1395 | lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)), |
d7e09d03 PT |
1396 | NULL, NULL); |
1397 | l_wait_event(thread->t_ctl_waitq, | |
1398 | thread_is_stopping(thread) || | |
1399 | thread_is_event(thread), | |
1400 | &lwi); | |
1401 | ||
1402 | if (thread_test_and_clear_flags(thread, SVC_STOPPING)) | |
1403 | break; | |
71e8dd9a | 1404 | thread_test_and_clear_flags(thread, SVC_EVENT); |
d7e09d03 PT |
1405 | } |
1406 | ||
1407 | thread_set_flags(thread, SVC_STOPPED); | |
1408 | wake_up(&thread->t_ctl_waitq); | |
1409 | ||
1410 | CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n", | |
1411 | "ldlm_poold", current_pid()); | |
1412 | ||
1413 | complete_and_exit(&ldlm_pools_comp, 0); | |
1414 | } | |
1415 | ||
1416 | static int ldlm_pools_thread_start(void) | |
1417 | { | |
1418 | struct l_wait_info lwi = { 0 }; | |
68b636b6 | 1419 | struct task_struct *task; |
d7e09d03 PT |
1420 | |
1421 | if (ldlm_pools_thread != NULL) | |
0a3bdb00 | 1422 | return -EALREADY; |
d7e09d03 | 1423 | |
352f7891 | 1424 | ldlm_pools_thread = kzalloc(sizeof(*ldlm_pools_thread), GFP_NOFS); |
94e67761 | 1425 | if (!ldlm_pools_thread) |
0a3bdb00 | 1426 | return -ENOMEM; |
d7e09d03 PT |
1427 | |
1428 | init_completion(&ldlm_pools_comp); | |
1429 | init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq); | |
1430 | ||
1431 | task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread, | |
1432 | "ldlm_poold"); | |
1433 | if (IS_ERR(task)) { | |
1434 | CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task)); | |
352f7891 | 1435 | kfree(ldlm_pools_thread); |
d7e09d03 | 1436 | ldlm_pools_thread = NULL; |
0a3bdb00 | 1437 | return PTR_ERR(task); |
d7e09d03 PT |
1438 | } |
1439 | l_wait_event(ldlm_pools_thread->t_ctl_waitq, | |
1440 | thread_is_running(ldlm_pools_thread), &lwi); | |
0a3bdb00 | 1441 | return 0; |
d7e09d03 PT |
1442 | } |
1443 | ||
1444 | static void ldlm_pools_thread_stop(void) | |
1445 | { | |
8d2ff65d | 1446 | if (ldlm_pools_thread == NULL) |
d7e09d03 | 1447 | return; |
d7e09d03 PT |
1448 | |
1449 | thread_set_flags(ldlm_pools_thread, SVC_STOPPING); | |
1450 | wake_up(&ldlm_pools_thread->t_ctl_waitq); | |
1451 | ||
1452 | /* | |
1453 | * Make sure that pools thread is finished before freeing @thread. | |
1454 | * This fixes possible race and oops due to accessing freed memory | |
1455 | * in pools thread. | |
1456 | */ | |
1457 | wait_for_completion(&ldlm_pools_comp); | |
352f7891 | 1458 | kfree(ldlm_pools_thread); |
d7e09d03 | 1459 | ldlm_pools_thread = NULL; |
d7e09d03 PT |
1460 | } |
1461 | ||
cbc3769e PT |
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, | |
1466 | }; | |
1467 | ||
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, | |
1472 | }; | |
1473 | ||
d7e09d03 PT |
1474 | int ldlm_pools_init(void) |
1475 | { | |
1476 | int rc; | |
d7e09d03 PT |
1477 | |
1478 | rc = ldlm_pools_thread_start(); | |
1479 | if (rc == 0) { | |
cbc3769e PT |
1480 | register_shrinker(&ldlm_pools_srv_shrinker); |
1481 | register_shrinker(&ldlm_pools_cli_shrinker); | |
d7e09d03 | 1482 | } |
0a3bdb00 | 1483 | return rc; |
d7e09d03 PT |
1484 | } |
1485 | EXPORT_SYMBOL(ldlm_pools_init); | |
1486 | ||
1487 | void ldlm_pools_fini(void) | |
1488 | { | |
cbc3769e PT |
1489 | unregister_shrinker(&ldlm_pools_srv_shrinker); |
1490 | unregister_shrinker(&ldlm_pools_cli_shrinker); | |
d7e09d03 PT |
1491 | ldlm_pools_thread_stop(); |
1492 | } | |
1493 | EXPORT_SYMBOL(ldlm_pools_fini); |