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 | * | |
57 | * Client has LVF, that is, lock volume factor which regulates how much sensitive | |
58 | * client should be about last SLV from server. The higher LVF is the more locks | |
59 | * will be canceled on client. Default value for it is 1. Setting LVF to 2 means | |
60 | * that client will cancel locks 2 times faster. | |
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 | * | |
73 | * General idea of this belongs to Vitaly Fertman (vitaly@clusterfs.com). Andreas | |
74 | * Dilger (adilger@clusterfs.com) proposed few nice ideas like using LVF and many | |
75 | * cleanups. Flow definition to allow more easy understanding of the logic belongs | |
76 | * to Nikita Danilov (nikita@clusterfs.com) as well as many cleanups and fixes. | |
77 | * And design and implementation are done by Yury Umanets (umka@clusterfs.com). | |
78 | * | |
79 | * Glossary for terms used: | |
80 | * | |
81 | * pl_limit - Number of allowed locks in pool. Applies to server and client | |
82 | * side (tunable); | |
83 | * | |
84 | * pl_granted - Number of granted locks (calculated); | |
85 | * pl_grant_rate - Number of granted locks for last T (calculated); | |
86 | * pl_cancel_rate - Number of canceled locks for last T (calculated); | |
87 | * pl_grant_speed - Grant speed (GR - CR) for last T (calculated); | |
88 | * pl_grant_plan - Planned number of granted locks for next T (calculated); | |
89 | * pl_server_lock_volume - Current server lock volume (calculated); | |
90 | * | |
91 | * As it may be seen from list above, we have few possible tunables which may | |
92 | * affect behavior much. They all may be modified via proc. However, they also | |
93 | * give a possibility for constructing few pre-defined behavior policies. If | |
94 | * none of predefines is suitable for a working pattern being used, new one may | |
95 | * be "constructed" via proc tunables. | |
96 | */ | |
97 | ||
98 | #define DEBUG_SUBSYSTEM S_LDLM | |
99 | ||
100 | # include <lustre_dlm.h> | |
101 | ||
102 | #include <cl_object.h> | |
103 | ||
104 | #include <obd_class.h> | |
105 | #include <obd_support.h> | |
106 | #include "ldlm_internal.h" | |
107 | ||
108 | ||
109 | /* | |
110 | * 50 ldlm locks for 1MB of RAM. | |
111 | */ | |
112 | #define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_CACHE_SHIFT)) * 50) | |
113 | ||
114 | /* | |
115 | * Maximal possible grant step plan in %. | |
116 | */ | |
117 | #define LDLM_POOL_MAX_GSP (30) | |
118 | ||
119 | /* | |
120 | * Minimal possible grant step plan in %. | |
121 | */ | |
122 | #define LDLM_POOL_MIN_GSP (1) | |
123 | ||
124 | /* | |
125 | * This controls the speed of reaching LDLM_POOL_MAX_GSP | |
126 | * with increasing thread period. | |
127 | */ | |
128 | #define LDLM_POOL_GSP_STEP_SHIFT (2) | |
129 | ||
130 | /* | |
131 | * LDLM_POOL_GSP% of all locks is default GP. | |
132 | */ | |
133 | #define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100) | |
134 | ||
135 | /* | |
136 | * Max age for locks on clients. | |
137 | */ | |
138 | #define LDLM_POOL_MAX_AGE (36000) | |
139 | ||
140 | /* | |
141 | * The granularity of SLV calculation. | |
142 | */ | |
143 | #define LDLM_POOL_SLV_SHIFT (10) | |
144 | ||
145 | extern proc_dir_entry_t *ldlm_ns_proc_dir; | |
146 | ||
147 | static inline __u64 dru(__u64 val, __u32 shift, int round_up) | |
148 | { | |
149 | return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift; | |
150 | } | |
151 | ||
152 | static inline __u64 ldlm_pool_slv_max(__u32 L) | |
153 | { | |
154 | /* | |
155 | * Allow to have all locks for 1 client for 10 hrs. | |
156 | * Formula is the following: limit * 10h / 1 client. | |
157 | */ | |
158 | __u64 lim = (__u64)L * LDLM_POOL_MAX_AGE / 1; | |
159 | return lim; | |
160 | } | |
161 | ||
162 | static inline __u64 ldlm_pool_slv_min(__u32 L) | |
163 | { | |
164 | return 1; | |
165 | } | |
166 | ||
167 | enum { | |
168 | LDLM_POOL_FIRST_STAT = 0, | |
169 | LDLM_POOL_GRANTED_STAT = LDLM_POOL_FIRST_STAT, | |
170 | LDLM_POOL_GRANT_STAT, | |
171 | LDLM_POOL_CANCEL_STAT, | |
172 | LDLM_POOL_GRANT_RATE_STAT, | |
173 | LDLM_POOL_CANCEL_RATE_STAT, | |
174 | LDLM_POOL_GRANT_PLAN_STAT, | |
175 | LDLM_POOL_SLV_STAT, | |
176 | LDLM_POOL_SHRINK_REQTD_STAT, | |
177 | LDLM_POOL_SHRINK_FREED_STAT, | |
178 | LDLM_POOL_RECALC_STAT, | |
179 | LDLM_POOL_TIMING_STAT, | |
180 | LDLM_POOL_LAST_STAT | |
181 | }; | |
182 | ||
183 | static inline struct ldlm_namespace *ldlm_pl2ns(struct ldlm_pool *pl) | |
184 | { | |
185 | return container_of(pl, struct ldlm_namespace, ns_pool); | |
186 | } | |
187 | ||
188 | /** | |
189 | * Calculates suggested grant_step in % of available locks for passed | |
190 | * \a period. This is later used in grant_plan calculations. | |
191 | */ | |
192 | static inline int ldlm_pool_t2gsp(unsigned int t) | |
193 | { | |
194 | /* | |
195 | * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP | |
196 | * and up to 30% for anything higher than LDLM_POOL_GSP_STEP. | |
197 | * | |
198 | * How this will affect execution is the following: | |
199 | * | |
200 | * - for thread period 1s we will have grant_step 1% which good from | |
201 | * pov of taking some load off from server and push it out to clients. | |
202 | * This is like that because 1% for grant_step means that server will | |
203 | * not allow clients to get lots of locks in short period of time and | |
204 | * keep all old locks in their caches. Clients will always have to | |
205 | * get some locks back if they want to take some new; | |
206 | * | |
207 | * - for thread period 10s (which is default) we will have 23% which | |
208 | * means that clients will have enough of room to take some new locks | |
209 | * without getting some back. All locks from this 23% which were not | |
210 | * taken by clients in current period will contribute in SLV growing. | |
211 | * SLV growing means more locks cached on clients until limit or grant | |
212 | * plan is reached. | |
213 | */ | |
214 | return LDLM_POOL_MAX_GSP - | |
215 | ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >> | |
216 | (t >> LDLM_POOL_GSP_STEP_SHIFT)); | |
217 | } | |
218 | ||
219 | /** | |
220 | * Recalculates next grant limit on passed \a pl. | |
221 | * | |
222 | * \pre ->pl_lock is locked. | |
223 | */ | |
224 | static void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl) | |
225 | { | |
226 | int granted, grant_step, limit; | |
227 | ||
228 | limit = ldlm_pool_get_limit(pl); | |
229 | granted = atomic_read(&pl->pl_granted); | |
230 | ||
231 | grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period); | |
232 | grant_step = ((limit - granted) * grant_step) / 100; | |
233 | pl->pl_grant_plan = granted + grant_step; | |
234 | limit = (limit * 5) >> 2; | |
235 | if (pl->pl_grant_plan > limit) | |
236 | pl->pl_grant_plan = limit; | |
237 | } | |
238 | ||
239 | /** | |
240 | * Recalculates next SLV on passed \a pl. | |
241 | * | |
242 | * \pre ->pl_lock is locked. | |
243 | */ | |
244 | static void ldlm_pool_recalc_slv(struct ldlm_pool *pl) | |
245 | { | |
246 | int granted; | |
247 | int grant_plan; | |
248 | int round_up; | |
249 | __u64 slv; | |
250 | __u64 slv_factor; | |
251 | __u64 grant_usage; | |
252 | __u32 limit; | |
253 | ||
254 | slv = pl->pl_server_lock_volume; | |
255 | grant_plan = pl->pl_grant_plan; | |
256 | limit = ldlm_pool_get_limit(pl); | |
257 | granted = atomic_read(&pl->pl_granted); | |
258 | round_up = granted < limit; | |
259 | ||
260 | grant_usage = max_t(int, limit - (granted - grant_plan), 1); | |
261 | ||
262 | /* | |
263 | * Find out SLV change factor which is the ratio of grant usage | |
264 | * from limit. SLV changes as fast as the ratio of grant plan | |
265 | * consumption. The more locks from grant plan are not consumed | |
266 | * by clients in last interval (idle time), the faster grows | |
267 | * SLV. And the opposite, the more grant plan is over-consumed | |
268 | * (load time) the faster drops SLV. | |
269 | */ | |
270 | slv_factor = (grant_usage << LDLM_POOL_SLV_SHIFT); | |
271 | do_div(slv_factor, limit); | |
272 | slv = slv * slv_factor; | |
273 | slv = dru(slv, LDLM_POOL_SLV_SHIFT, round_up); | |
274 | ||
275 | if (slv > ldlm_pool_slv_max(limit)) { | |
276 | slv = ldlm_pool_slv_max(limit); | |
277 | } else if (slv < ldlm_pool_slv_min(limit)) { | |
278 | slv = ldlm_pool_slv_min(limit); | |
279 | } | |
280 | ||
281 | pl->pl_server_lock_volume = slv; | |
282 | } | |
283 | ||
284 | /** | |
285 | * Recalculates next stats on passed \a pl. | |
286 | * | |
287 | * \pre ->pl_lock is locked. | |
288 | */ | |
289 | static void ldlm_pool_recalc_stats(struct ldlm_pool *pl) | |
290 | { | |
291 | int grant_plan = pl->pl_grant_plan; | |
292 | __u64 slv = pl->pl_server_lock_volume; | |
293 | int granted = atomic_read(&pl->pl_granted); | |
294 | int grant_rate = atomic_read(&pl->pl_grant_rate); | |
295 | int cancel_rate = atomic_read(&pl->pl_cancel_rate); | |
296 | ||
297 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT, | |
298 | slv); | |
299 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT, | |
300 | granted); | |
301 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT, | |
302 | grant_rate); | |
303 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT, | |
304 | grant_plan); | |
305 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT, | |
306 | cancel_rate); | |
307 | } | |
308 | ||
309 | /** | |
310 | * Sets current SLV into obd accessible via ldlm_pl2ns(pl)->ns_obd. | |
311 | */ | |
312 | static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl) | |
313 | { | |
314 | struct obd_device *obd; | |
315 | ||
316 | /* | |
317 | * Set new SLV in obd field for using it later without accessing the | |
318 | * pool. This is required to avoid race between sending reply to client | |
319 | * with new SLV and cleanup server stack in which we can't guarantee | |
320 | * that namespace is still alive. We know only that obd is alive as | |
321 | * long as valid export is alive. | |
322 | */ | |
323 | obd = ldlm_pl2ns(pl)->ns_obd; | |
324 | LASSERT(obd != NULL); | |
325 | write_lock(&obd->obd_pool_lock); | |
326 | obd->obd_pool_slv = pl->pl_server_lock_volume; | |
327 | write_unlock(&obd->obd_pool_lock); | |
328 | } | |
329 | ||
330 | /** | |
331 | * Recalculates all pool fields on passed \a pl. | |
332 | * | |
333 | * \pre ->pl_lock is not locked. | |
334 | */ | |
335 | static int ldlm_srv_pool_recalc(struct ldlm_pool *pl) | |
336 | { | |
337 | time_t recalc_interval_sec; | |
d7e09d03 PT |
338 | |
339 | recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time; | |
340 | if (recalc_interval_sec < pl->pl_recalc_period) | |
0a3bdb00 | 341 | return 0; |
d7e09d03 PT |
342 | |
343 | spin_lock(&pl->pl_lock); | |
344 | recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time; | |
345 | if (recalc_interval_sec < pl->pl_recalc_period) { | |
346 | spin_unlock(&pl->pl_lock); | |
0a3bdb00 | 347 | return 0; |
d7e09d03 PT |
348 | } |
349 | /* | |
350 | * Recalc SLV after last period. This should be done | |
351 | * _before_ recalculating new grant plan. | |
352 | */ | |
353 | ldlm_pool_recalc_slv(pl); | |
354 | ||
355 | /* | |
356 | * Make sure that pool informed obd of last SLV changes. | |
357 | */ | |
358 | ldlm_srv_pool_push_slv(pl); | |
359 | ||
360 | /* | |
361 | * Update grant_plan for new period. | |
362 | */ | |
363 | ldlm_pool_recalc_grant_plan(pl); | |
364 | ||
365 | pl->pl_recalc_time = cfs_time_current_sec(); | |
366 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT, | |
367 | recalc_interval_sec); | |
368 | spin_unlock(&pl->pl_lock); | |
0a3bdb00 | 369 | return 0; |
d7e09d03 PT |
370 | } |
371 | ||
372 | /** | |
373 | * This function is used on server side as main entry point for memory | |
374 | * pressure handling. It decreases SLV on \a pl according to passed | |
375 | * \a nr and \a gfp_mask. | |
376 | * | |
377 | * Our goal here is to decrease SLV such a way that clients hold \a nr | |
378 | * locks smaller in next 10h. | |
379 | */ | |
380 | static int ldlm_srv_pool_shrink(struct ldlm_pool *pl, | |
381 | int nr, unsigned int gfp_mask) | |
382 | { | |
383 | __u32 limit; | |
384 | ||
385 | /* | |
386 | * VM is asking how many entries may be potentially freed. | |
387 | */ | |
388 | if (nr == 0) | |
389 | return atomic_read(&pl->pl_granted); | |
390 | ||
391 | /* | |
392 | * Client already canceled locks but server is already in shrinker | |
393 | * and can't cancel anything. Let's catch this race. | |
394 | */ | |
395 | if (atomic_read(&pl->pl_granted) == 0) | |
0a3bdb00 | 396 | return 0; |
d7e09d03 PT |
397 | |
398 | spin_lock(&pl->pl_lock); | |
399 | ||
400 | /* | |
401 | * We want shrinker to possibly cause cancellation of @nr locks from | |
402 | * clients or grant approximately @nr locks smaller next intervals. | |
403 | * | |
404 | * This is why we decreased SLV by @nr. This effect will only be as | |
405 | * long as one re-calc interval (1s these days) and this should be | |
406 | * enough to pass this decreased SLV to all clients. On next recalc | |
407 | * interval pool will either increase SLV if locks load is not high | |
408 | * or will keep on same level or even decrease again, thus, shrinker | |
409 | * decreased SLV will affect next recalc intervals and this way will | |
410 | * make locking load lower. | |
411 | */ | |
412 | if (nr < pl->pl_server_lock_volume) { | |
413 | pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr; | |
414 | } else { | |
415 | limit = ldlm_pool_get_limit(pl); | |
416 | pl->pl_server_lock_volume = ldlm_pool_slv_min(limit); | |
417 | } | |
418 | ||
419 | /* | |
420 | * Make sure that pool informed obd of last SLV changes. | |
421 | */ | |
422 | ldlm_srv_pool_push_slv(pl); | |
423 | spin_unlock(&pl->pl_lock); | |
424 | ||
425 | /* | |
426 | * We did not really free any memory here so far, it only will be | |
427 | * freed later may be, so that we return 0 to not confuse VM. | |
428 | */ | |
429 | return 0; | |
430 | } | |
431 | ||
432 | /** | |
433 | * Setup server side pool \a pl with passed \a limit. | |
434 | */ | |
435 | static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit) | |
436 | { | |
437 | struct obd_device *obd; | |
438 | ||
439 | obd = ldlm_pl2ns(pl)->ns_obd; | |
440 | LASSERT(obd != NULL && obd != LP_POISON); | |
441 | LASSERT(obd->obd_type != LP_POISON); | |
442 | write_lock(&obd->obd_pool_lock); | |
443 | obd->obd_pool_limit = limit; | |
444 | write_unlock(&obd->obd_pool_lock); | |
445 | ||
446 | ldlm_pool_set_limit(pl, limit); | |
447 | return 0; | |
448 | } | |
449 | ||
450 | /** | |
451 | * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl. | |
452 | */ | |
453 | static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl) | |
454 | { | |
455 | struct obd_device *obd; | |
456 | ||
457 | /* | |
458 | * Get new SLV and Limit from obd which is updated with coming | |
459 | * RPCs. | |
460 | */ | |
461 | obd = ldlm_pl2ns(pl)->ns_obd; | |
462 | LASSERT(obd != NULL); | |
463 | read_lock(&obd->obd_pool_lock); | |
464 | pl->pl_server_lock_volume = obd->obd_pool_slv; | |
465 | ldlm_pool_set_limit(pl, obd->obd_pool_limit); | |
466 | read_unlock(&obd->obd_pool_lock); | |
467 | } | |
468 | ||
469 | /** | |
470 | * Recalculates client size pool \a pl according to current SLV and Limit. | |
471 | */ | |
472 | static int ldlm_cli_pool_recalc(struct ldlm_pool *pl) | |
473 | { | |
474 | time_t recalc_interval_sec; | |
d7e09d03 PT |
475 | |
476 | recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time; | |
477 | if (recalc_interval_sec < pl->pl_recalc_period) | |
0a3bdb00 | 478 | return 0; |
d7e09d03 PT |
479 | |
480 | spin_lock(&pl->pl_lock); | |
481 | /* | |
482 | * Check if we need to recalc lists now. | |
483 | */ | |
484 | recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time; | |
485 | if (recalc_interval_sec < pl->pl_recalc_period) { | |
486 | spin_unlock(&pl->pl_lock); | |
0a3bdb00 | 487 | return 0; |
d7e09d03 PT |
488 | } |
489 | ||
490 | /* | |
491 | * Make sure that pool knows last SLV and Limit from obd. | |
492 | */ | |
493 | ldlm_cli_pool_pop_slv(pl); | |
494 | ||
495 | pl->pl_recalc_time = cfs_time_current_sec(); | |
496 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT, | |
497 | recalc_interval_sec); | |
498 | spin_unlock(&pl->pl_lock); | |
499 | ||
500 | /* | |
501 | * Do not cancel locks in case lru resize is disabled for this ns. | |
502 | */ | |
503 | if (!ns_connect_lru_resize(ldlm_pl2ns(pl))) | |
0a3bdb00 | 504 | return 0; |
d7e09d03 PT |
505 | |
506 | /* | |
507 | * In the time of canceling locks on client we do not need to maintain | |
508 | * sharp timing, we only want to cancel locks asap according to new SLV. | |
509 | * It may be called when SLV has changed much, this is why we do not | |
510 | * take into account pl->pl_recalc_time here. | |
511 | */ | |
0a3bdb00 | 512 | return ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC, LDLM_CANCEL_LRUR); |
d7e09d03 PT |
513 | } |
514 | ||
515 | /** | |
516 | * This function is main entry point for memory pressure handling on client | |
517 | * side. Main goal of this function is to cancel some number of locks on | |
518 | * passed \a pl according to \a nr and \a gfp_mask. | |
519 | */ | |
520 | static int ldlm_cli_pool_shrink(struct ldlm_pool *pl, | |
521 | int nr, unsigned int gfp_mask) | |
522 | { | |
523 | struct ldlm_namespace *ns; | |
524 | int canceled = 0, unused; | |
525 | ||
526 | ns = ldlm_pl2ns(pl); | |
527 | ||
528 | /* | |
529 | * Do not cancel locks in case lru resize is disabled for this ns. | |
530 | */ | |
531 | if (!ns_connect_lru_resize(ns)) | |
0a3bdb00 | 532 | return 0; |
d7e09d03 PT |
533 | |
534 | /* | |
535 | * Make sure that pool knows last SLV and Limit from obd. | |
536 | */ | |
537 | ldlm_cli_pool_pop_slv(pl); | |
538 | ||
539 | spin_lock(&ns->ns_lock); | |
540 | unused = ns->ns_nr_unused; | |
541 | spin_unlock(&ns->ns_lock); | |
542 | ||
543 | if (nr) { | |
544 | canceled = ldlm_cancel_lru(ns, nr, LCF_ASYNC, | |
545 | LDLM_CANCEL_SHRINK); | |
546 | } | |
547 | /* | |
548 | * Return the number of potentially reclaimable locks. | |
549 | */ | |
550 | return ((unused - canceled) / 100) * sysctl_vfs_cache_pressure; | |
551 | } | |
552 | ||
553 | struct ldlm_pool_ops ldlm_srv_pool_ops = { | |
554 | .po_recalc = ldlm_srv_pool_recalc, | |
555 | .po_shrink = ldlm_srv_pool_shrink, | |
556 | .po_setup = ldlm_srv_pool_setup | |
557 | }; | |
558 | ||
559 | struct ldlm_pool_ops ldlm_cli_pool_ops = { | |
560 | .po_recalc = ldlm_cli_pool_recalc, | |
561 | .po_shrink = ldlm_cli_pool_shrink | |
562 | }; | |
563 | ||
564 | /** | |
565 | * Pool recalc wrapper. Will call either client or server pool recalc callback | |
566 | * depending what pool \a pl is used. | |
567 | */ | |
568 | int ldlm_pool_recalc(struct ldlm_pool *pl) | |
569 | { | |
570 | time_t recalc_interval_sec; | |
571 | int count; | |
572 | ||
573 | recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time; | |
574 | if (recalc_interval_sec <= 0) | |
575 | goto recalc; | |
576 | ||
577 | spin_lock(&pl->pl_lock); | |
d7e09d03 PT |
578 | if (recalc_interval_sec > 0) { |
579 | /* | |
580 | * Update pool statistics every 1s. | |
581 | */ | |
582 | ldlm_pool_recalc_stats(pl); | |
583 | ||
584 | /* | |
585 | * Zero out all rates and speed for the last period. | |
586 | */ | |
587 | atomic_set(&pl->pl_grant_rate, 0); | |
588 | atomic_set(&pl->pl_cancel_rate, 0); | |
589 | } | |
590 | spin_unlock(&pl->pl_lock); | |
591 | ||
592 | recalc: | |
593 | if (pl->pl_ops->po_recalc != NULL) { | |
594 | count = pl->pl_ops->po_recalc(pl); | |
595 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT, | |
596 | count); | |
d7e09d03 | 597 | } |
3eface59 OD |
598 | recalc_interval_sec = pl->pl_recalc_time - cfs_time_current_sec() + |
599 | pl->pl_recalc_period; | |
d7e09d03 | 600 | |
3eface59 | 601 | return recalc_interval_sec; |
d7e09d03 | 602 | } |
d7e09d03 PT |
603 | |
604 | /** | |
605 | * Pool shrink wrapper. Will call either client or server pool recalc callback | |
606 | * depending what pool \a pl is used. | |
607 | */ | |
608 | int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, | |
609 | unsigned int gfp_mask) | |
610 | { | |
611 | int cancel = 0; | |
612 | ||
613 | if (pl->pl_ops->po_shrink != NULL) { | |
614 | cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask); | |
615 | if (nr > 0) { | |
616 | lprocfs_counter_add(pl->pl_stats, | |
617 | LDLM_POOL_SHRINK_REQTD_STAT, | |
618 | nr); | |
619 | lprocfs_counter_add(pl->pl_stats, | |
620 | LDLM_POOL_SHRINK_FREED_STAT, | |
621 | cancel); | |
622 | CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, " | |
623 | "shrunk %d\n", pl->pl_name, nr, cancel); | |
624 | } | |
625 | } | |
626 | return cancel; | |
627 | } | |
628 | EXPORT_SYMBOL(ldlm_pool_shrink); | |
629 | ||
630 | /** | |
631 | * Pool setup wrapper. Will call either client or server pool recalc callback | |
632 | * depending what pool \a pl is used. | |
633 | * | |
634 | * Sets passed \a limit into pool \a pl. | |
635 | */ | |
636 | int ldlm_pool_setup(struct ldlm_pool *pl, int limit) | |
637 | { | |
638 | if (pl->pl_ops->po_setup != NULL) | |
639 | return(pl->pl_ops->po_setup(pl, limit)); | |
640 | return 0; | |
641 | } | |
642 | EXPORT_SYMBOL(ldlm_pool_setup); | |
643 | ||
73bb1da6 | 644 | static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused) |
d7e09d03 PT |
645 | { |
646 | int granted, grant_rate, cancel_rate, grant_step; | |
73bb1da6 PT |
647 | int grant_speed, grant_plan, lvf; |
648 | struct ldlm_pool *pl = m->private; | |
d7e09d03 PT |
649 | __u64 slv, clv; |
650 | __u32 limit; | |
651 | ||
652 | spin_lock(&pl->pl_lock); | |
653 | slv = pl->pl_server_lock_volume; | |
654 | clv = pl->pl_client_lock_volume; | |
655 | limit = ldlm_pool_get_limit(pl); | |
656 | grant_plan = pl->pl_grant_plan; | |
657 | granted = atomic_read(&pl->pl_granted); | |
658 | grant_rate = atomic_read(&pl->pl_grant_rate); | |
659 | cancel_rate = atomic_read(&pl->pl_cancel_rate); | |
660 | grant_speed = grant_rate - cancel_rate; | |
661 | lvf = atomic_read(&pl->pl_lock_volume_factor); | |
662 | grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period); | |
663 | spin_unlock(&pl->pl_lock); | |
664 | ||
73bb1da6 PT |
665 | seq_printf(m, "LDLM pool state (%s):\n" |
666 | " SLV: "LPU64"\n" | |
667 | " CLV: "LPU64"\n" | |
668 | " LVF: %d\n", | |
669 | pl->pl_name, slv, clv, lvf); | |
d7e09d03 PT |
670 | |
671 | if (ns_is_server(ldlm_pl2ns(pl))) { | |
73bb1da6 PT |
672 | seq_printf(m, " GSP: %d%%\n" |
673 | " GP: %d\n", | |
674 | grant_step, grant_plan); | |
d7e09d03 | 675 | } |
73bb1da6 PT |
676 | seq_printf(m, " GR: %d\n" " CR: %d\n" " GS: %d\n" |
677 | " G: %d\n" " L: %d\n", | |
678 | grant_rate, cancel_rate, grant_speed, | |
679 | granted, limit); | |
680 | ||
681 | return 0; | |
d7e09d03 | 682 | } |
73bb1da6 | 683 | LPROC_SEQ_FOPS_RO(lprocfs_pool_state); |
d7e09d03 | 684 | |
73bb1da6 | 685 | static int lprocfs_grant_speed_seq_show(struct seq_file *m, void *unused) |
d7e09d03 | 686 | { |
73bb1da6 | 687 | struct ldlm_pool *pl = m->private; |
d7e09d03 PT |
688 | int grant_speed; |
689 | ||
690 | spin_lock(&pl->pl_lock); | |
691 | /* serialize with ldlm_pool_recalc */ | |
692 | grant_speed = atomic_read(&pl->pl_grant_rate) - | |
693 | atomic_read(&pl->pl_cancel_rate); | |
694 | spin_unlock(&pl->pl_lock); | |
73bb1da6 | 695 | return lprocfs_rd_uint(m, &grant_speed); |
d7e09d03 PT |
696 | } |
697 | ||
73bb1da6 PT |
698 | LDLM_POOL_PROC_READER_SEQ_SHOW(grant_plan, int); |
699 | LPROC_SEQ_FOPS_RO(lprocfs_grant_plan); | |
700 | ||
701 | LDLM_POOL_PROC_READER_SEQ_SHOW(recalc_period, int); | |
d7e09d03 | 702 | LDLM_POOL_PROC_WRITER(recalc_period, int); |
73bb1da6 PT |
703 | static ssize_t lprocfs_recalc_period_seq_write(struct file *file, const char *buf, |
704 | size_t len, loff_t *off) | |
705 | { | |
706 | struct seq_file *seq = file->private_data; | |
707 | ||
708 | return lprocfs_wr_recalc_period(file, buf, len, seq->private); | |
709 | } | |
710 | LPROC_SEQ_FOPS(lprocfs_recalc_period); | |
711 | ||
712 | LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, u64); | |
713 | LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, atomic); | |
714 | LPROC_SEQ_FOPS_RW_TYPE(ldlm_pool_rw, atomic); | |
715 | ||
716 | LPROC_SEQ_FOPS_RO(lprocfs_grant_speed); | |
717 | ||
718 | #define LDLM_POOL_ADD_VAR(name, var, ops) \ | |
719 | do { \ | |
720 | snprintf(var_name, MAX_STRING_SIZE, #name); \ | |
721 | pool_vars[0].data = var; \ | |
722 | pool_vars[0].fops = ops; \ | |
723 | lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);\ | |
724 | } while (0) | |
d7e09d03 PT |
725 | |
726 | static int ldlm_pool_proc_init(struct ldlm_pool *pl) | |
727 | { | |
728 | struct ldlm_namespace *ns = ldlm_pl2ns(pl); | |
729 | struct proc_dir_entry *parent_ns_proc; | |
730 | struct lprocfs_vars pool_vars[2]; | |
731 | char *var_name = NULL; | |
732 | int rc = 0; | |
d7e09d03 PT |
733 | |
734 | OBD_ALLOC(var_name, MAX_STRING_SIZE + 1); | |
735 | if (!var_name) | |
0a3bdb00 | 736 | return -ENOMEM; |
d7e09d03 | 737 | |
73bb1da6 | 738 | parent_ns_proc = ns->ns_proc_dir_entry; |
d7e09d03 PT |
739 | if (parent_ns_proc == NULL) { |
740 | CERROR("%s: proc entry is not initialized\n", | |
741 | ldlm_ns_name(ns)); | |
742 | GOTO(out_free_name, rc = -EINVAL); | |
743 | } | |
744 | pl->pl_proc_dir = lprocfs_register("pool", parent_ns_proc, | |
745 | NULL, NULL); | |
746 | if (IS_ERR(pl->pl_proc_dir)) { | |
747 | CERROR("LProcFS failed in ldlm-pool-init\n"); | |
748 | rc = PTR_ERR(pl->pl_proc_dir); | |
5907838a | 749 | pl->pl_proc_dir = NULL; |
d7e09d03 PT |
750 | GOTO(out_free_name, rc); |
751 | } | |
752 | ||
753 | var_name[MAX_STRING_SIZE] = '\0'; | |
754 | memset(pool_vars, 0, sizeof(pool_vars)); | |
755 | pool_vars[0].name = var_name; | |
756 | ||
73bb1da6 PT |
757 | LDLM_POOL_ADD_VAR("server_lock_volume", &pl->pl_server_lock_volume, |
758 | &ldlm_pool_u64_fops); | |
759 | LDLM_POOL_ADD_VAR("limit", &pl->pl_limit, &ldlm_pool_rw_atomic_fops); | |
760 | LDLM_POOL_ADD_VAR("granted", &pl->pl_granted, &ldlm_pool_atomic_fops); | |
761 | LDLM_POOL_ADD_VAR("grant_speed", pl, &lprocfs_grant_speed_fops); | |
762 | LDLM_POOL_ADD_VAR("cancel_rate", &pl->pl_cancel_rate, | |
763 | &ldlm_pool_atomic_fops); | |
764 | LDLM_POOL_ADD_VAR("grant_rate", &pl->pl_grant_rate, | |
765 | &ldlm_pool_atomic_fops); | |
766 | LDLM_POOL_ADD_VAR("grant_plan", pl, &lprocfs_grant_plan_fops); | |
767 | LDLM_POOL_ADD_VAR("recalc_period", pl, &lprocfs_recalc_period_fops); | |
768 | LDLM_POOL_ADD_VAR("lock_volume_factor", &pl->pl_lock_volume_factor, | |
769 | &ldlm_pool_rw_atomic_fops); | |
770 | LDLM_POOL_ADD_VAR("state", pl, &lprocfs_pool_state_fops); | |
d7e09d03 PT |
771 | |
772 | pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT - | |
773 | LDLM_POOL_FIRST_STAT, 0); | |
774 | if (!pl->pl_stats) | |
775 | GOTO(out_free_name, rc = -ENOMEM); | |
776 | ||
777 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT, | |
778 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
779 | "granted", "locks"); | |
780 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT, | |
781 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
782 | "grant", "locks"); | |
783 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT, | |
784 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
785 | "cancel", "locks"); | |
786 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT, | |
787 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
788 | "grant_rate", "locks/s"); | |
789 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT, | |
790 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
791 | "cancel_rate", "locks/s"); | |
792 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT, | |
793 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
794 | "grant_plan", "locks/s"); | |
795 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT, | |
796 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
797 | "slv", "slv"); | |
798 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT, | |
799 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
800 | "shrink_request", "locks"); | |
801 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT, | |
802 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
803 | "shrink_freed", "locks"); | |
804 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT, | |
805 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
806 | "recalc_freed", "locks"); | |
807 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT, | |
808 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
809 | "recalc_timing", "sec"); | |
810 | rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats); | |
811 | ||
d7e09d03 PT |
812 | out_free_name: |
813 | OBD_FREE(var_name, MAX_STRING_SIZE + 1); | |
814 | return rc; | |
815 | } | |
816 | ||
817 | static void ldlm_pool_proc_fini(struct ldlm_pool *pl) | |
818 | { | |
819 | if (pl->pl_stats != NULL) { | |
820 | lprocfs_free_stats(&pl->pl_stats); | |
821 | pl->pl_stats = NULL; | |
822 | } | |
823 | if (pl->pl_proc_dir != NULL) { | |
824 | lprocfs_remove(&pl->pl_proc_dir); | |
825 | pl->pl_proc_dir = NULL; | |
826 | } | |
827 | } | |
828 | ||
829 | int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns, | |
830 | int idx, ldlm_side_t client) | |
831 | { | |
832 | int rc; | |
d7e09d03 PT |
833 | |
834 | spin_lock_init(&pl->pl_lock); | |
835 | atomic_set(&pl->pl_granted, 0); | |
836 | pl->pl_recalc_time = cfs_time_current_sec(); | |
837 | atomic_set(&pl->pl_lock_volume_factor, 1); | |
838 | ||
839 | atomic_set(&pl->pl_grant_rate, 0); | |
840 | atomic_set(&pl->pl_cancel_rate, 0); | |
841 | pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L); | |
842 | ||
843 | snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d", | |
844 | ldlm_ns_name(ns), idx); | |
845 | ||
846 | if (client == LDLM_NAMESPACE_SERVER) { | |
847 | pl->pl_ops = &ldlm_srv_pool_ops; | |
848 | ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L); | |
849 | pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD; | |
850 | pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L); | |
851 | } else { | |
852 | ldlm_pool_set_limit(pl, 1); | |
853 | pl->pl_server_lock_volume = 0; | |
854 | pl->pl_ops = &ldlm_cli_pool_ops; | |
855 | pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD; | |
856 | } | |
857 | pl->pl_client_lock_volume = 0; | |
858 | rc = ldlm_pool_proc_init(pl); | |
859 | if (rc) | |
0a3bdb00 | 860 | return rc; |
d7e09d03 PT |
861 | |
862 | CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name); | |
863 | ||
0a3bdb00 | 864 | return rc; |
d7e09d03 PT |
865 | } |
866 | EXPORT_SYMBOL(ldlm_pool_init); | |
867 | ||
868 | void ldlm_pool_fini(struct ldlm_pool *pl) | |
869 | { | |
d7e09d03 PT |
870 | ldlm_pool_proc_fini(pl); |
871 | ||
872 | /* | |
873 | * Pool should not be used after this point. We can't free it here as | |
874 | * it lives in struct ldlm_namespace, but still interested in catching | |
875 | * any abnormal using cases. | |
876 | */ | |
877 | POISON(pl, 0x5a, sizeof(*pl)); | |
d7e09d03 PT |
878 | } |
879 | EXPORT_SYMBOL(ldlm_pool_fini); | |
880 | ||
881 | /** | |
882 | * Add new taken ldlm lock \a lock into pool \a pl accounting. | |
883 | */ | |
884 | void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock) | |
885 | { | |
886 | /* | |
887 | * FLOCK locks are special in a sense that they are almost never | |
888 | * cancelled, instead special kind of lock is used to drop them. | |
889 | * also there is no LRU for flock locks, so no point in tracking | |
890 | * them anyway. | |
891 | */ | |
892 | if (lock->l_resource->lr_type == LDLM_FLOCK) | |
893 | return; | |
894 | ||
895 | atomic_inc(&pl->pl_granted); | |
896 | atomic_inc(&pl->pl_grant_rate); | |
897 | lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT); | |
898 | /* | |
899 | * Do not do pool recalc for client side as all locks which | |
900 | * potentially may be canceled has already been packed into | |
901 | * enqueue/cancel rpc. Also we do not want to run out of stack | |
902 | * with too long call paths. | |
903 | */ | |
904 | if (ns_is_server(ldlm_pl2ns(pl))) | |
905 | ldlm_pool_recalc(pl); | |
906 | } | |
907 | EXPORT_SYMBOL(ldlm_pool_add); | |
908 | ||
909 | /** | |
910 | * Remove ldlm lock \a lock from pool \a pl accounting. | |
911 | */ | |
912 | void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock) | |
913 | { | |
914 | /* | |
915 | * Filter out FLOCK locks. Read above comment in ldlm_pool_add(). | |
916 | */ | |
917 | if (lock->l_resource->lr_type == LDLM_FLOCK) | |
918 | return; | |
919 | ||
920 | LASSERT(atomic_read(&pl->pl_granted) > 0); | |
921 | atomic_dec(&pl->pl_granted); | |
922 | atomic_inc(&pl->pl_cancel_rate); | |
923 | ||
924 | lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT); | |
925 | ||
926 | if (ns_is_server(ldlm_pl2ns(pl))) | |
927 | ldlm_pool_recalc(pl); | |
928 | } | |
929 | EXPORT_SYMBOL(ldlm_pool_del); | |
930 | ||
931 | /** | |
932 | * Returns current \a pl SLV. | |
933 | * | |
934 | * \pre ->pl_lock is not locked. | |
935 | */ | |
936 | __u64 ldlm_pool_get_slv(struct ldlm_pool *pl) | |
937 | { | |
938 | __u64 slv; | |
939 | spin_lock(&pl->pl_lock); | |
940 | slv = pl->pl_server_lock_volume; | |
941 | spin_unlock(&pl->pl_lock); | |
942 | return slv; | |
943 | } | |
944 | EXPORT_SYMBOL(ldlm_pool_get_slv); | |
945 | ||
946 | /** | |
947 | * Sets passed \a slv to \a pl. | |
948 | * | |
949 | * \pre ->pl_lock is not locked. | |
950 | */ | |
951 | void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv) | |
952 | { | |
953 | spin_lock(&pl->pl_lock); | |
954 | pl->pl_server_lock_volume = slv; | |
955 | spin_unlock(&pl->pl_lock); | |
956 | } | |
957 | EXPORT_SYMBOL(ldlm_pool_set_slv); | |
958 | ||
959 | /** | |
960 | * Returns current \a pl CLV. | |
961 | * | |
962 | * \pre ->pl_lock is not locked. | |
963 | */ | |
964 | __u64 ldlm_pool_get_clv(struct ldlm_pool *pl) | |
965 | { | |
966 | __u64 slv; | |
967 | spin_lock(&pl->pl_lock); | |
968 | slv = pl->pl_client_lock_volume; | |
969 | spin_unlock(&pl->pl_lock); | |
970 | return slv; | |
971 | } | |
972 | EXPORT_SYMBOL(ldlm_pool_get_clv); | |
973 | ||
974 | /** | |
975 | * Sets passed \a clv to \a pl. | |
976 | * | |
977 | * \pre ->pl_lock is not locked. | |
978 | */ | |
979 | void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv) | |
980 | { | |
981 | spin_lock(&pl->pl_lock); | |
982 | pl->pl_client_lock_volume = clv; | |
983 | spin_unlock(&pl->pl_lock); | |
984 | } | |
985 | EXPORT_SYMBOL(ldlm_pool_set_clv); | |
986 | ||
987 | /** | |
988 | * Returns current \a pl limit. | |
989 | */ | |
990 | __u32 ldlm_pool_get_limit(struct ldlm_pool *pl) | |
991 | { | |
992 | return atomic_read(&pl->pl_limit); | |
993 | } | |
994 | EXPORT_SYMBOL(ldlm_pool_get_limit); | |
995 | ||
996 | /** | |
997 | * Sets passed \a limit to \a pl. | |
998 | */ | |
999 | void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit) | |
1000 | { | |
1001 | atomic_set(&pl->pl_limit, limit); | |
1002 | } | |
1003 | EXPORT_SYMBOL(ldlm_pool_set_limit); | |
1004 | ||
1005 | /** | |
1006 | * Returns current LVF from \a pl. | |
1007 | */ | |
1008 | __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl) | |
1009 | { | |
1010 | return atomic_read(&pl->pl_lock_volume_factor); | |
1011 | } | |
1012 | EXPORT_SYMBOL(ldlm_pool_get_lvf); | |
1013 | ||
1014 | static int ldlm_pool_granted(struct ldlm_pool *pl) | |
1015 | { | |
1016 | return atomic_read(&pl->pl_granted); | |
1017 | } | |
1018 | ||
1019 | static struct ptlrpc_thread *ldlm_pools_thread; | |
1020 | static struct shrinker *ldlm_pools_srv_shrinker; | |
1021 | static struct shrinker *ldlm_pools_cli_shrinker; | |
1022 | static struct completion ldlm_pools_comp; | |
1023 | ||
1024 | /* | |
1025 | * Cancel \a nr locks from all namespaces (if possible). Returns number of | |
1026 | * cached locks after shrink is finished. All namespaces are asked to | |
1027 | * cancel approximately equal amount of locks to keep balancing. | |
1028 | */ | |
1029 | static int ldlm_pools_shrink(ldlm_side_t client, int nr, | |
1030 | unsigned int gfp_mask) | |
1031 | { | |
1032 | int total = 0, cached = 0, nr_ns; | |
1033 | struct ldlm_namespace *ns; | |
91a50030 | 1034 | struct ldlm_namespace *ns_old = NULL; /* loop detection */ |
d7e09d03 PT |
1035 | void *cookie; |
1036 | ||
1037 | if (client == LDLM_NAMESPACE_CLIENT && nr != 0 && | |
1038 | !(gfp_mask & __GFP_FS)) | |
1039 | return -1; | |
1040 | ||
1041 | CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n", | |
1042 | nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server"); | |
1043 | ||
1044 | cookie = cl_env_reenter(); | |
1045 | ||
1046 | /* | |
1047 | * Find out how many resources we may release. | |
1048 | */ | |
91a50030 | 1049 | for (nr_ns = ldlm_namespace_nr_read(client); |
d7e09d03 PT |
1050 | nr_ns > 0; nr_ns--) |
1051 | { | |
1052 | mutex_lock(ldlm_namespace_lock(client)); | |
1053 | if (list_empty(ldlm_namespace_list(client))) { | |
1054 | mutex_unlock(ldlm_namespace_lock(client)); | |
1055 | cl_env_reexit(cookie); | |
1056 | return 0; | |
1057 | } | |
1058 | ns = ldlm_namespace_first_locked(client); | |
91a50030 OD |
1059 | |
1060 | if (ns == ns_old) { | |
1061 | mutex_unlock(ldlm_namespace_lock(client)); | |
1062 | break; | |
1063 | } | |
1064 | ||
1065 | if (ldlm_ns_empty(ns)) { | |
1066 | ldlm_namespace_move_to_inactive_locked(ns, client); | |
1067 | mutex_unlock(ldlm_namespace_lock(client)); | |
1068 | continue; | |
1069 | } | |
1070 | ||
1071 | if (ns_old == NULL) | |
1072 | ns_old = ns; | |
1073 | ||
d7e09d03 | 1074 | ldlm_namespace_get(ns); |
91a50030 | 1075 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
1076 | mutex_unlock(ldlm_namespace_lock(client)); |
1077 | total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask); | |
1078 | ldlm_namespace_put(ns); | |
1079 | } | |
1080 | ||
1081 | if (nr == 0 || total == 0) { | |
1082 | cl_env_reexit(cookie); | |
1083 | return total; | |
1084 | } | |
1085 | ||
1086 | /* | |
1087 | * Shrink at least ldlm_namespace_nr(client) namespaces. | |
1088 | */ | |
91a50030 | 1089 | for (nr_ns = ldlm_namespace_nr_read(client) - nr_ns; |
d7e09d03 PT |
1090 | nr_ns > 0; nr_ns--) |
1091 | { | |
1092 | int cancel, nr_locks; | |
1093 | ||
1094 | /* | |
1095 | * Do not call shrink under ldlm_namespace_lock(client) | |
1096 | */ | |
1097 | mutex_lock(ldlm_namespace_lock(client)); | |
1098 | if (list_empty(ldlm_namespace_list(client))) { | |
1099 | mutex_unlock(ldlm_namespace_lock(client)); | |
1100 | /* | |
1101 | * If list is empty, we can't return any @cached > 0, | |
1102 | * that probably would cause needless shrinker | |
1103 | * call. | |
1104 | */ | |
1105 | cached = 0; | |
1106 | break; | |
1107 | } | |
1108 | ns = ldlm_namespace_first_locked(client); | |
1109 | ldlm_namespace_get(ns); | |
91a50030 | 1110 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
1111 | mutex_unlock(ldlm_namespace_lock(client)); |
1112 | ||
1113 | nr_locks = ldlm_pool_granted(&ns->ns_pool); | |
1114 | cancel = 1 + nr_locks * nr / total; | |
1115 | ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask); | |
1116 | cached += ldlm_pool_granted(&ns->ns_pool); | |
1117 | ldlm_namespace_put(ns); | |
1118 | } | |
1119 | cl_env_reexit(cookie); | |
1120 | /* we only decrease the SLV in server pools shrinker, return -1 to | |
1121 | * kernel to avoid needless loop. LU-1128 */ | |
1122 | return (client == LDLM_NAMESPACE_SERVER) ? -1 : cached; | |
1123 | } | |
1124 | ||
1125 | static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask)) | |
1126 | { | |
1127 | return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER, | |
1128 | shrink_param(sc, nr_to_scan), | |
1129 | shrink_param(sc, gfp_mask)); | |
1130 | } | |
1131 | ||
1132 | static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask)) | |
1133 | { | |
1134 | return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT, | |
1135 | shrink_param(sc, nr_to_scan), | |
1136 | shrink_param(sc, gfp_mask)); | |
1137 | } | |
1138 | ||
3eface59 | 1139 | int ldlm_pools_recalc(ldlm_side_t client) |
d7e09d03 PT |
1140 | { |
1141 | __u32 nr_l = 0, nr_p = 0, l; | |
1142 | struct ldlm_namespace *ns; | |
91a50030 | 1143 | struct ldlm_namespace *ns_old = NULL; |
d7e09d03 | 1144 | int nr, equal = 0; |
3eface59 | 1145 | int time = 50; /* seconds of sleep if no active namespaces */ |
d7e09d03 PT |
1146 | |
1147 | /* | |
1148 | * No need to setup pool limit for client pools. | |
1149 | */ | |
1150 | if (client == LDLM_NAMESPACE_SERVER) { | |
1151 | /* | |
1152 | * Check all modest namespaces first. | |
1153 | */ | |
1154 | mutex_lock(ldlm_namespace_lock(client)); | |
1155 | list_for_each_entry(ns, ldlm_namespace_list(client), | |
1156 | ns_list_chain) | |
1157 | { | |
1158 | if (ns->ns_appetite != LDLM_NAMESPACE_MODEST) | |
1159 | continue; | |
1160 | ||
1161 | l = ldlm_pool_granted(&ns->ns_pool); | |
1162 | if (l == 0) | |
1163 | l = 1; | |
1164 | ||
1165 | /* | |
1166 | * Set the modest pools limit equal to their avg granted | |
1167 | * locks + ~6%. | |
1168 | */ | |
1169 | l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0); | |
1170 | ldlm_pool_setup(&ns->ns_pool, l); | |
1171 | nr_l += l; | |
1172 | nr_p++; | |
1173 | } | |
1174 | ||
1175 | /* | |
1176 | * Make sure that modest namespaces did not eat more that 2/3 | |
1177 | * of limit. | |
1178 | */ | |
1179 | if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) { | |
1180 | CWARN("\"Modest\" pools eat out 2/3 of server locks " | |
1181 | "limit (%d of %lu). This means that you have too " | |
1182 | "many clients for this amount of server RAM. " | |
1183 | "Upgrade server!\n", nr_l, LDLM_POOL_HOST_L); | |
1184 | equal = 1; | |
1185 | } | |
1186 | ||
1187 | /* | |
1188 | * The rest is given to greedy namespaces. | |
1189 | */ | |
1190 | list_for_each_entry(ns, ldlm_namespace_list(client), | |
1191 | ns_list_chain) | |
1192 | { | |
1193 | if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY) | |
1194 | continue; | |
1195 | ||
1196 | if (equal) { | |
1197 | /* | |
1198 | * In the case 2/3 locks are eaten out by | |
1199 | * modest pools, we re-setup equal limit | |
1200 | * for _all_ pools. | |
1201 | */ | |
1202 | l = LDLM_POOL_HOST_L / | |
91a50030 | 1203 | ldlm_namespace_nr_read(client); |
d7e09d03 PT |
1204 | } else { |
1205 | /* | |
1206 | * All the rest of greedy pools will have | |
1207 | * all locks in equal parts. | |
1208 | */ | |
1209 | l = (LDLM_POOL_HOST_L - nr_l) / | |
91a50030 | 1210 | (ldlm_namespace_nr_read(client) - |
d7e09d03 PT |
1211 | nr_p); |
1212 | } | |
1213 | ldlm_pool_setup(&ns->ns_pool, l); | |
1214 | } | |
1215 | mutex_unlock(ldlm_namespace_lock(client)); | |
1216 | } | |
1217 | ||
1218 | /* | |
1219 | * Recalc at least ldlm_namespace_nr(client) namespaces. | |
1220 | */ | |
91a50030 | 1221 | for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) { |
d7e09d03 PT |
1222 | int skip; |
1223 | /* | |
1224 | * Lock the list, get first @ns in the list, getref, move it | |
1225 | * to the tail, unlock and call pool recalc. This way we avoid | |
1226 | * calling recalc under @ns lock what is really good as we get | |
1227 | * rid of potential deadlock on client nodes when canceling | |
1228 | * locks synchronously. | |
1229 | */ | |
1230 | mutex_lock(ldlm_namespace_lock(client)); | |
1231 | if (list_empty(ldlm_namespace_list(client))) { | |
1232 | mutex_unlock(ldlm_namespace_lock(client)); | |
1233 | break; | |
1234 | } | |
1235 | ns = ldlm_namespace_first_locked(client); | |
1236 | ||
91a50030 OD |
1237 | if (ns_old == ns) { /* Full pass complete */ |
1238 | mutex_unlock(ldlm_namespace_lock(client)); | |
1239 | break; | |
1240 | } | |
1241 | ||
1242 | /* We got an empty namespace, need to move it back to inactive | |
1243 | * list. | |
1244 | * The race with parallel resource creation is fine: | |
1245 | * - If they do namespace_get before our check, we fail the | |
1246 | * check and they move this item to the end of the list anyway | |
1247 | * - If we do the check and then they do namespace_get, then | |
1248 | * we move the namespace to inactive and they will move | |
1249 | * it back to active (synchronised by the lock, so no clash | |
1250 | * there). | |
1251 | */ | |
1252 | if (ldlm_ns_empty(ns)) { | |
1253 | ldlm_namespace_move_to_inactive_locked(ns, client); | |
1254 | mutex_unlock(ldlm_namespace_lock(client)); | |
1255 | continue; | |
1256 | } | |
1257 | ||
1258 | if (ns_old == NULL) | |
1259 | ns_old = ns; | |
1260 | ||
d7e09d03 PT |
1261 | spin_lock(&ns->ns_lock); |
1262 | /* | |
1263 | * skip ns which is being freed, and we don't want to increase | |
1264 | * its refcount again, not even temporarily. bz21519 & LU-499. | |
1265 | */ | |
1266 | if (ns->ns_stopping) { | |
1267 | skip = 1; | |
1268 | } else { | |
1269 | skip = 0; | |
1270 | ldlm_namespace_get(ns); | |
1271 | } | |
1272 | spin_unlock(&ns->ns_lock); | |
1273 | ||
91a50030 | 1274 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
1275 | mutex_unlock(ldlm_namespace_lock(client)); |
1276 | ||
1277 | /* | |
1278 | * After setup is done - recalc the pool. | |
1279 | */ | |
1280 | if (!skip) { | |
3eface59 OD |
1281 | int ttime = ldlm_pool_recalc(&ns->ns_pool); |
1282 | ||
1283 | if (ttime < time) | |
1284 | time = ttime; | |
1285 | ||
d7e09d03 PT |
1286 | ldlm_namespace_put(ns); |
1287 | } | |
1288 | } | |
3eface59 | 1289 | return time; |
d7e09d03 PT |
1290 | } |
1291 | EXPORT_SYMBOL(ldlm_pools_recalc); | |
1292 | ||
1293 | static int ldlm_pools_thread_main(void *arg) | |
1294 | { | |
1295 | struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg; | |
3eface59 | 1296 | int s_time, c_time; |
d7e09d03 PT |
1297 | |
1298 | thread_set_flags(thread, SVC_RUNNING); | |
1299 | wake_up(&thread->t_ctl_waitq); | |
1300 | ||
1301 | CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n", | |
1302 | "ldlm_poold", current_pid()); | |
1303 | ||
1304 | while (1) { | |
1305 | struct l_wait_info lwi; | |
1306 | ||
1307 | /* | |
1308 | * Recal all pools on this tick. | |
1309 | */ | |
3eface59 OD |
1310 | s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER); |
1311 | c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT); | |
d7e09d03 PT |
1312 | |
1313 | /* | |
1314 | * Wait until the next check time, or until we're | |
1315 | * stopped. | |
1316 | */ | |
3eface59 | 1317 | lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)), |
d7e09d03 PT |
1318 | NULL, NULL); |
1319 | l_wait_event(thread->t_ctl_waitq, | |
1320 | thread_is_stopping(thread) || | |
1321 | thread_is_event(thread), | |
1322 | &lwi); | |
1323 | ||
1324 | if (thread_test_and_clear_flags(thread, SVC_STOPPING)) | |
1325 | break; | |
1326 | else | |
1327 | thread_test_and_clear_flags(thread, SVC_EVENT); | |
1328 | } | |
1329 | ||
1330 | thread_set_flags(thread, SVC_STOPPED); | |
1331 | wake_up(&thread->t_ctl_waitq); | |
1332 | ||
1333 | CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n", | |
1334 | "ldlm_poold", current_pid()); | |
1335 | ||
1336 | complete_and_exit(&ldlm_pools_comp, 0); | |
1337 | } | |
1338 | ||
1339 | static int ldlm_pools_thread_start(void) | |
1340 | { | |
1341 | struct l_wait_info lwi = { 0 }; | |
1342 | task_t *task; | |
d7e09d03 PT |
1343 | |
1344 | if (ldlm_pools_thread != NULL) | |
0a3bdb00 | 1345 | return -EALREADY; |
d7e09d03 PT |
1346 | |
1347 | OBD_ALLOC_PTR(ldlm_pools_thread); | |
1348 | if (ldlm_pools_thread == NULL) | |
0a3bdb00 | 1349 | return -ENOMEM; |
d7e09d03 PT |
1350 | |
1351 | init_completion(&ldlm_pools_comp); | |
1352 | init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq); | |
1353 | ||
1354 | task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread, | |
1355 | "ldlm_poold"); | |
1356 | if (IS_ERR(task)) { | |
1357 | CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task)); | |
1358 | OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread)); | |
1359 | ldlm_pools_thread = NULL; | |
0a3bdb00 | 1360 | return PTR_ERR(task); |
d7e09d03 PT |
1361 | } |
1362 | l_wait_event(ldlm_pools_thread->t_ctl_waitq, | |
1363 | thread_is_running(ldlm_pools_thread), &lwi); | |
0a3bdb00 | 1364 | return 0; |
d7e09d03 PT |
1365 | } |
1366 | ||
1367 | static void ldlm_pools_thread_stop(void) | |
1368 | { | |
d7e09d03 | 1369 | if (ldlm_pools_thread == NULL) { |
d7e09d03 PT |
1370 | return; |
1371 | } | |
1372 | ||
1373 | thread_set_flags(ldlm_pools_thread, SVC_STOPPING); | |
1374 | wake_up(&ldlm_pools_thread->t_ctl_waitq); | |
1375 | ||
1376 | /* | |
1377 | * Make sure that pools thread is finished before freeing @thread. | |
1378 | * This fixes possible race and oops due to accessing freed memory | |
1379 | * in pools thread. | |
1380 | */ | |
1381 | wait_for_completion(&ldlm_pools_comp); | |
1382 | OBD_FREE_PTR(ldlm_pools_thread); | |
1383 | ldlm_pools_thread = NULL; | |
d7e09d03 PT |
1384 | } |
1385 | ||
1386 | int ldlm_pools_init(void) | |
1387 | { | |
1388 | int rc; | |
d7e09d03 PT |
1389 | |
1390 | rc = ldlm_pools_thread_start(); | |
1391 | if (rc == 0) { | |
1392 | ldlm_pools_srv_shrinker = | |
1393 | set_shrinker(DEFAULT_SEEKS, | |
1394 | ldlm_pools_srv_shrink); | |
1395 | ldlm_pools_cli_shrinker = | |
1396 | set_shrinker(DEFAULT_SEEKS, | |
1397 | ldlm_pools_cli_shrink); | |
1398 | } | |
0a3bdb00 | 1399 | return rc; |
d7e09d03 PT |
1400 | } |
1401 | EXPORT_SYMBOL(ldlm_pools_init); | |
1402 | ||
1403 | void ldlm_pools_fini(void) | |
1404 | { | |
1405 | if (ldlm_pools_srv_shrinker != NULL) { | |
1406 | remove_shrinker(ldlm_pools_srv_shrinker); | |
1407 | ldlm_pools_srv_shrinker = NULL; | |
1408 | } | |
1409 | if (ldlm_pools_cli_shrinker != NULL) { | |
1410 | remove_shrinker(ldlm_pools_cli_shrinker); | |
1411 | ldlm_pools_cli_shrinker = NULL; | |
1412 | } | |
1413 | ldlm_pools_thread_stop(); | |
1414 | } | |
1415 | EXPORT_SYMBOL(ldlm_pools_fini); |