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; | |
338 | ENTRY; | |
339 | ||
340 | recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time; | |
341 | if (recalc_interval_sec < pl->pl_recalc_period) | |
342 | RETURN(0); | |
343 | ||
344 | spin_lock(&pl->pl_lock); | |
345 | recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time; | |
346 | if (recalc_interval_sec < pl->pl_recalc_period) { | |
347 | spin_unlock(&pl->pl_lock); | |
348 | RETURN(0); | |
349 | } | |
350 | /* | |
351 | * Recalc SLV after last period. This should be done | |
352 | * _before_ recalculating new grant plan. | |
353 | */ | |
354 | ldlm_pool_recalc_slv(pl); | |
355 | ||
356 | /* | |
357 | * Make sure that pool informed obd of last SLV changes. | |
358 | */ | |
359 | ldlm_srv_pool_push_slv(pl); | |
360 | ||
361 | /* | |
362 | * Update grant_plan for new period. | |
363 | */ | |
364 | ldlm_pool_recalc_grant_plan(pl); | |
365 | ||
366 | pl->pl_recalc_time = cfs_time_current_sec(); | |
367 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT, | |
368 | recalc_interval_sec); | |
369 | spin_unlock(&pl->pl_lock); | |
370 | RETURN(0); | |
371 | } | |
372 | ||
373 | /** | |
374 | * This function is used on server side as main entry point for memory | |
375 | * pressure handling. It decreases SLV on \a pl according to passed | |
376 | * \a nr and \a gfp_mask. | |
377 | * | |
378 | * Our goal here is to decrease SLV such a way that clients hold \a nr | |
379 | * locks smaller in next 10h. | |
380 | */ | |
381 | static int ldlm_srv_pool_shrink(struct ldlm_pool *pl, | |
382 | int nr, unsigned int gfp_mask) | |
383 | { | |
384 | __u32 limit; | |
385 | ||
386 | /* | |
387 | * VM is asking how many entries may be potentially freed. | |
388 | */ | |
389 | if (nr == 0) | |
390 | return atomic_read(&pl->pl_granted); | |
391 | ||
392 | /* | |
393 | * Client already canceled locks but server is already in shrinker | |
394 | * and can't cancel anything. Let's catch this race. | |
395 | */ | |
396 | if (atomic_read(&pl->pl_granted) == 0) | |
397 | RETURN(0); | |
398 | ||
399 | spin_lock(&pl->pl_lock); | |
400 | ||
401 | /* | |
402 | * We want shrinker to possibly cause cancellation of @nr locks from | |
403 | * clients or grant approximately @nr locks smaller next intervals. | |
404 | * | |
405 | * This is why we decreased SLV by @nr. This effect will only be as | |
406 | * long as one re-calc interval (1s these days) and this should be | |
407 | * enough to pass this decreased SLV to all clients. On next recalc | |
408 | * interval pool will either increase SLV if locks load is not high | |
409 | * or will keep on same level or even decrease again, thus, shrinker | |
410 | * decreased SLV will affect next recalc intervals and this way will | |
411 | * make locking load lower. | |
412 | */ | |
413 | if (nr < pl->pl_server_lock_volume) { | |
414 | pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr; | |
415 | } else { | |
416 | limit = ldlm_pool_get_limit(pl); | |
417 | pl->pl_server_lock_volume = ldlm_pool_slv_min(limit); | |
418 | } | |
419 | ||
420 | /* | |
421 | * Make sure that pool informed obd of last SLV changes. | |
422 | */ | |
423 | ldlm_srv_pool_push_slv(pl); | |
424 | spin_unlock(&pl->pl_lock); | |
425 | ||
426 | /* | |
427 | * We did not really free any memory here so far, it only will be | |
428 | * freed later may be, so that we return 0 to not confuse VM. | |
429 | */ | |
430 | return 0; | |
431 | } | |
432 | ||
433 | /** | |
434 | * Setup server side pool \a pl with passed \a limit. | |
435 | */ | |
436 | static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit) | |
437 | { | |
438 | struct obd_device *obd; | |
439 | ||
440 | obd = ldlm_pl2ns(pl)->ns_obd; | |
441 | LASSERT(obd != NULL && obd != LP_POISON); | |
442 | LASSERT(obd->obd_type != LP_POISON); | |
443 | write_lock(&obd->obd_pool_lock); | |
444 | obd->obd_pool_limit = limit; | |
445 | write_unlock(&obd->obd_pool_lock); | |
446 | ||
447 | ldlm_pool_set_limit(pl, limit); | |
448 | return 0; | |
449 | } | |
450 | ||
451 | /** | |
452 | * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl. | |
453 | */ | |
454 | static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl) | |
455 | { | |
456 | struct obd_device *obd; | |
457 | ||
458 | /* | |
459 | * Get new SLV and Limit from obd which is updated with coming | |
460 | * RPCs. | |
461 | */ | |
462 | obd = ldlm_pl2ns(pl)->ns_obd; | |
463 | LASSERT(obd != NULL); | |
464 | read_lock(&obd->obd_pool_lock); | |
465 | pl->pl_server_lock_volume = obd->obd_pool_slv; | |
466 | ldlm_pool_set_limit(pl, obd->obd_pool_limit); | |
467 | read_unlock(&obd->obd_pool_lock); | |
468 | } | |
469 | ||
470 | /** | |
471 | * Recalculates client size pool \a pl according to current SLV and Limit. | |
472 | */ | |
473 | static int ldlm_cli_pool_recalc(struct ldlm_pool *pl) | |
474 | { | |
475 | time_t recalc_interval_sec; | |
476 | ENTRY; | |
477 | ||
478 | recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time; | |
479 | if (recalc_interval_sec < pl->pl_recalc_period) | |
480 | RETURN(0); | |
481 | ||
482 | spin_lock(&pl->pl_lock); | |
483 | /* | |
484 | * Check if we need to recalc lists now. | |
485 | */ | |
486 | recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time; | |
487 | if (recalc_interval_sec < pl->pl_recalc_period) { | |
488 | spin_unlock(&pl->pl_lock); | |
489 | RETURN(0); | |
490 | } | |
491 | ||
492 | /* | |
493 | * Make sure that pool knows last SLV and Limit from obd. | |
494 | */ | |
495 | ldlm_cli_pool_pop_slv(pl); | |
496 | ||
497 | pl->pl_recalc_time = cfs_time_current_sec(); | |
498 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT, | |
499 | recalc_interval_sec); | |
500 | spin_unlock(&pl->pl_lock); | |
501 | ||
502 | /* | |
503 | * Do not cancel locks in case lru resize is disabled for this ns. | |
504 | */ | |
505 | if (!ns_connect_lru_resize(ldlm_pl2ns(pl))) | |
506 | RETURN(0); | |
507 | ||
508 | /* | |
509 | * In the time of canceling locks on client we do not need to maintain | |
510 | * sharp timing, we only want to cancel locks asap according to new SLV. | |
511 | * It may be called when SLV has changed much, this is why we do not | |
512 | * take into account pl->pl_recalc_time here. | |
513 | */ | |
514 | RETURN(ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC, | |
515 | LDLM_CANCEL_LRUR)); | |
516 | } | |
517 | ||
518 | /** | |
519 | * This function is main entry point for memory pressure handling on client | |
520 | * side. Main goal of this function is to cancel some number of locks on | |
521 | * passed \a pl according to \a nr and \a gfp_mask. | |
522 | */ | |
523 | static int ldlm_cli_pool_shrink(struct ldlm_pool *pl, | |
524 | int nr, unsigned int gfp_mask) | |
525 | { | |
526 | struct ldlm_namespace *ns; | |
527 | int canceled = 0, unused; | |
528 | ||
529 | ns = ldlm_pl2ns(pl); | |
530 | ||
531 | /* | |
532 | * Do not cancel locks in case lru resize is disabled for this ns. | |
533 | */ | |
534 | if (!ns_connect_lru_resize(ns)) | |
535 | RETURN(0); | |
536 | ||
537 | /* | |
538 | * Make sure that pool knows last SLV and Limit from obd. | |
539 | */ | |
540 | ldlm_cli_pool_pop_slv(pl); | |
541 | ||
542 | spin_lock(&ns->ns_lock); | |
543 | unused = ns->ns_nr_unused; | |
544 | spin_unlock(&ns->ns_lock); | |
545 | ||
546 | if (nr) { | |
547 | canceled = ldlm_cancel_lru(ns, nr, LCF_ASYNC, | |
548 | LDLM_CANCEL_SHRINK); | |
549 | } | |
550 | /* | |
551 | * Return the number of potentially reclaimable locks. | |
552 | */ | |
553 | return ((unused - canceled) / 100) * sysctl_vfs_cache_pressure; | |
554 | } | |
555 | ||
556 | struct ldlm_pool_ops ldlm_srv_pool_ops = { | |
557 | .po_recalc = ldlm_srv_pool_recalc, | |
558 | .po_shrink = ldlm_srv_pool_shrink, | |
559 | .po_setup = ldlm_srv_pool_setup | |
560 | }; | |
561 | ||
562 | struct ldlm_pool_ops ldlm_cli_pool_ops = { | |
563 | .po_recalc = ldlm_cli_pool_recalc, | |
564 | .po_shrink = ldlm_cli_pool_shrink | |
565 | }; | |
566 | ||
567 | /** | |
568 | * Pool recalc wrapper. Will call either client or server pool recalc callback | |
569 | * depending what pool \a pl is used. | |
570 | */ | |
571 | int ldlm_pool_recalc(struct ldlm_pool *pl) | |
572 | { | |
573 | time_t recalc_interval_sec; | |
574 | int count; | |
575 | ||
576 | recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time; | |
577 | if (recalc_interval_sec <= 0) | |
578 | goto recalc; | |
579 | ||
580 | spin_lock(&pl->pl_lock); | |
581 | recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time; | |
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); | |
601 | return count; | |
602 | } | |
603 | ||
604 | return 0; | |
605 | } | |
606 | EXPORT_SYMBOL(ldlm_pool_recalc); | |
607 | ||
608 | /** | |
609 | * Pool shrink wrapper. Will call either client or server pool recalc callback | |
610 | * depending what pool \a pl is used. | |
611 | */ | |
612 | int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, | |
613 | unsigned int gfp_mask) | |
614 | { | |
615 | int cancel = 0; | |
616 | ||
617 | if (pl->pl_ops->po_shrink != NULL) { | |
618 | cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask); | |
619 | if (nr > 0) { | |
620 | lprocfs_counter_add(pl->pl_stats, | |
621 | LDLM_POOL_SHRINK_REQTD_STAT, | |
622 | nr); | |
623 | lprocfs_counter_add(pl->pl_stats, | |
624 | LDLM_POOL_SHRINK_FREED_STAT, | |
625 | cancel); | |
626 | CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, " | |
627 | "shrunk %d\n", pl->pl_name, nr, cancel); | |
628 | } | |
629 | } | |
630 | return cancel; | |
631 | } | |
632 | EXPORT_SYMBOL(ldlm_pool_shrink); | |
633 | ||
634 | /** | |
635 | * Pool setup wrapper. Will call either client or server pool recalc callback | |
636 | * depending what pool \a pl is used. | |
637 | * | |
638 | * Sets passed \a limit into pool \a pl. | |
639 | */ | |
640 | int ldlm_pool_setup(struct ldlm_pool *pl, int limit) | |
641 | { | |
642 | if (pl->pl_ops->po_setup != NULL) | |
643 | return(pl->pl_ops->po_setup(pl, limit)); | |
644 | return 0; | |
645 | } | |
646 | EXPORT_SYMBOL(ldlm_pool_setup); | |
647 | ||
648 | static int lprocfs_rd_pool_state(char *page, char **start, off_t off, | |
649 | int count, int *eof, void *data) | |
650 | { | |
651 | int granted, grant_rate, cancel_rate, grant_step; | |
652 | int nr = 0, grant_speed, grant_plan, lvf; | |
653 | struct ldlm_pool *pl = data; | |
654 | __u64 slv, clv; | |
655 | __u32 limit; | |
656 | ||
657 | spin_lock(&pl->pl_lock); | |
658 | slv = pl->pl_server_lock_volume; | |
659 | clv = pl->pl_client_lock_volume; | |
660 | limit = ldlm_pool_get_limit(pl); | |
661 | grant_plan = pl->pl_grant_plan; | |
662 | granted = atomic_read(&pl->pl_granted); | |
663 | grant_rate = atomic_read(&pl->pl_grant_rate); | |
664 | cancel_rate = atomic_read(&pl->pl_cancel_rate); | |
665 | grant_speed = grant_rate - cancel_rate; | |
666 | lvf = atomic_read(&pl->pl_lock_volume_factor); | |
667 | grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period); | |
668 | spin_unlock(&pl->pl_lock); | |
669 | ||
670 | nr += snprintf(page + nr, count - nr, "LDLM pool state (%s):\n", | |
671 | pl->pl_name); | |
672 | nr += snprintf(page + nr, count - nr, " SLV: "LPU64"\n", slv); | |
673 | nr += snprintf(page + nr, count - nr, " CLV: "LPU64"\n", clv); | |
674 | nr += snprintf(page + nr, count - nr, " LVF: %d\n", lvf); | |
675 | ||
676 | if (ns_is_server(ldlm_pl2ns(pl))) { | |
677 | nr += snprintf(page + nr, count - nr, " GSP: %d%%\n", | |
678 | grant_step); | |
679 | nr += snprintf(page + nr, count - nr, " GP: %d\n", | |
680 | grant_plan); | |
681 | } | |
682 | nr += snprintf(page + nr, count - nr, " GR: %d\n", | |
683 | grant_rate); | |
684 | nr += snprintf(page + nr, count - nr, " CR: %d\n", | |
685 | cancel_rate); | |
686 | nr += snprintf(page + nr, count - nr, " GS: %d\n", | |
687 | grant_speed); | |
688 | nr += snprintf(page + nr, count - nr, " G: %d\n", | |
689 | granted); | |
690 | nr += snprintf(page + nr, count - nr, " L: %d\n", | |
691 | limit); | |
692 | return nr; | |
693 | } | |
694 | ||
695 | static int lprocfs_rd_grant_speed(char *page, char **start, off_t off, | |
696 | int count, int *eof, void *data) | |
697 | { | |
698 | struct ldlm_pool *pl = data; | |
699 | int grant_speed; | |
700 | ||
701 | spin_lock(&pl->pl_lock); | |
702 | /* serialize with ldlm_pool_recalc */ | |
703 | grant_speed = atomic_read(&pl->pl_grant_rate) - | |
704 | atomic_read(&pl->pl_cancel_rate); | |
705 | spin_unlock(&pl->pl_lock); | |
706 | return lprocfs_rd_uint(page, start, off, count, eof, &grant_speed); | |
707 | } | |
708 | ||
709 | LDLM_POOL_PROC_READER(grant_plan, int); | |
710 | LDLM_POOL_PROC_READER(recalc_period, int); | |
711 | LDLM_POOL_PROC_WRITER(recalc_period, int); | |
712 | ||
713 | static int ldlm_pool_proc_init(struct ldlm_pool *pl) | |
714 | { | |
715 | struct ldlm_namespace *ns = ldlm_pl2ns(pl); | |
716 | struct proc_dir_entry *parent_ns_proc; | |
717 | struct lprocfs_vars pool_vars[2]; | |
718 | char *var_name = NULL; | |
719 | int rc = 0; | |
720 | ENTRY; | |
721 | ||
722 | OBD_ALLOC(var_name, MAX_STRING_SIZE + 1); | |
723 | if (!var_name) | |
724 | RETURN(-ENOMEM); | |
725 | ||
726 | parent_ns_proc = lprocfs_srch(ldlm_ns_proc_dir, | |
727 | ldlm_ns_name(ns)); | |
728 | if (parent_ns_proc == NULL) { | |
729 | CERROR("%s: proc entry is not initialized\n", | |
730 | ldlm_ns_name(ns)); | |
731 | GOTO(out_free_name, rc = -EINVAL); | |
732 | } | |
733 | pl->pl_proc_dir = lprocfs_register("pool", parent_ns_proc, | |
734 | NULL, NULL); | |
735 | if (IS_ERR(pl->pl_proc_dir)) { | |
736 | CERROR("LProcFS failed in ldlm-pool-init\n"); | |
737 | rc = PTR_ERR(pl->pl_proc_dir); | |
738 | GOTO(out_free_name, rc); | |
739 | } | |
740 | ||
741 | var_name[MAX_STRING_SIZE] = '\0'; | |
742 | memset(pool_vars, 0, sizeof(pool_vars)); | |
743 | pool_vars[0].name = var_name; | |
744 | ||
745 | snprintf(var_name, MAX_STRING_SIZE, "server_lock_volume"); | |
746 | pool_vars[0].data = &pl->pl_server_lock_volume; | |
747 | pool_vars[0].read_fptr = lprocfs_rd_u64; | |
748 | lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0); | |
749 | ||
750 | snprintf(var_name, MAX_STRING_SIZE, "limit"); | |
751 | pool_vars[0].data = &pl->pl_limit; | |
752 | pool_vars[0].read_fptr = lprocfs_rd_atomic; | |
753 | pool_vars[0].write_fptr = lprocfs_wr_atomic; | |
754 | lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0); | |
755 | ||
756 | snprintf(var_name, MAX_STRING_SIZE, "granted"); | |
757 | pool_vars[0].data = &pl->pl_granted; | |
758 | pool_vars[0].read_fptr = lprocfs_rd_atomic; | |
759 | lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0); | |
760 | ||
761 | snprintf(var_name, MAX_STRING_SIZE, "grant_speed"); | |
762 | pool_vars[0].data = pl; | |
763 | pool_vars[0].read_fptr = lprocfs_rd_grant_speed; | |
764 | lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0); | |
765 | ||
766 | snprintf(var_name, MAX_STRING_SIZE, "cancel_rate"); | |
767 | pool_vars[0].data = &pl->pl_cancel_rate; | |
768 | pool_vars[0].read_fptr = lprocfs_rd_atomic; | |
769 | lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0); | |
770 | ||
771 | snprintf(var_name, MAX_STRING_SIZE, "grant_rate"); | |
772 | pool_vars[0].data = &pl->pl_grant_rate; | |
773 | pool_vars[0].read_fptr = lprocfs_rd_atomic; | |
774 | lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0); | |
775 | ||
776 | snprintf(var_name, MAX_STRING_SIZE, "grant_plan"); | |
777 | pool_vars[0].data = pl; | |
778 | pool_vars[0].read_fptr = lprocfs_rd_grant_plan; | |
779 | lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0); | |
780 | ||
781 | snprintf(var_name, MAX_STRING_SIZE, "recalc_period"); | |
782 | pool_vars[0].data = pl; | |
783 | pool_vars[0].read_fptr = lprocfs_rd_recalc_period; | |
784 | pool_vars[0].write_fptr = lprocfs_wr_recalc_period; | |
785 | lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0); | |
786 | ||
787 | snprintf(var_name, MAX_STRING_SIZE, "lock_volume_factor"); | |
788 | pool_vars[0].data = &pl->pl_lock_volume_factor; | |
789 | pool_vars[0].read_fptr = lprocfs_rd_atomic; | |
790 | pool_vars[0].write_fptr = lprocfs_wr_atomic; | |
791 | lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0); | |
792 | ||
793 | snprintf(var_name, MAX_STRING_SIZE, "state"); | |
794 | pool_vars[0].data = pl; | |
795 | pool_vars[0].read_fptr = lprocfs_rd_pool_state; | |
796 | lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0); | |
797 | ||
798 | pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT - | |
799 | LDLM_POOL_FIRST_STAT, 0); | |
800 | if (!pl->pl_stats) | |
801 | GOTO(out_free_name, rc = -ENOMEM); | |
802 | ||
803 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT, | |
804 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
805 | "granted", "locks"); | |
806 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT, | |
807 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
808 | "grant", "locks"); | |
809 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT, | |
810 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
811 | "cancel", "locks"); | |
812 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT, | |
813 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
814 | "grant_rate", "locks/s"); | |
815 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT, | |
816 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
817 | "cancel_rate", "locks/s"); | |
818 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT, | |
819 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
820 | "grant_plan", "locks/s"); | |
821 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT, | |
822 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
823 | "slv", "slv"); | |
824 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT, | |
825 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
826 | "shrink_request", "locks"); | |
827 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT, | |
828 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
829 | "shrink_freed", "locks"); | |
830 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT, | |
831 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
832 | "recalc_freed", "locks"); | |
833 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT, | |
834 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
835 | "recalc_timing", "sec"); | |
836 | rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats); | |
837 | ||
838 | EXIT; | |
839 | out_free_name: | |
840 | OBD_FREE(var_name, MAX_STRING_SIZE + 1); | |
841 | return rc; | |
842 | } | |
843 | ||
844 | static void ldlm_pool_proc_fini(struct ldlm_pool *pl) | |
845 | { | |
846 | if (pl->pl_stats != NULL) { | |
847 | lprocfs_free_stats(&pl->pl_stats); | |
848 | pl->pl_stats = NULL; | |
849 | } | |
850 | if (pl->pl_proc_dir != NULL) { | |
851 | lprocfs_remove(&pl->pl_proc_dir); | |
852 | pl->pl_proc_dir = NULL; | |
853 | } | |
854 | } | |
855 | ||
856 | int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns, | |
857 | int idx, ldlm_side_t client) | |
858 | { | |
859 | int rc; | |
860 | ENTRY; | |
861 | ||
862 | spin_lock_init(&pl->pl_lock); | |
863 | atomic_set(&pl->pl_granted, 0); | |
864 | pl->pl_recalc_time = cfs_time_current_sec(); | |
865 | atomic_set(&pl->pl_lock_volume_factor, 1); | |
866 | ||
867 | atomic_set(&pl->pl_grant_rate, 0); | |
868 | atomic_set(&pl->pl_cancel_rate, 0); | |
869 | pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L); | |
870 | ||
871 | snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d", | |
872 | ldlm_ns_name(ns), idx); | |
873 | ||
874 | if (client == LDLM_NAMESPACE_SERVER) { | |
875 | pl->pl_ops = &ldlm_srv_pool_ops; | |
876 | ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L); | |
877 | pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD; | |
878 | pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L); | |
879 | } else { | |
880 | ldlm_pool_set_limit(pl, 1); | |
881 | pl->pl_server_lock_volume = 0; | |
882 | pl->pl_ops = &ldlm_cli_pool_ops; | |
883 | pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD; | |
884 | } | |
885 | pl->pl_client_lock_volume = 0; | |
886 | rc = ldlm_pool_proc_init(pl); | |
887 | if (rc) | |
888 | RETURN(rc); | |
889 | ||
890 | CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name); | |
891 | ||
892 | RETURN(rc); | |
893 | } | |
894 | EXPORT_SYMBOL(ldlm_pool_init); | |
895 | ||
896 | void ldlm_pool_fini(struct ldlm_pool *pl) | |
897 | { | |
898 | ENTRY; | |
899 | ldlm_pool_proc_fini(pl); | |
900 | ||
901 | /* | |
902 | * Pool should not be used after this point. We can't free it here as | |
903 | * it lives in struct ldlm_namespace, but still interested in catching | |
904 | * any abnormal using cases. | |
905 | */ | |
906 | POISON(pl, 0x5a, sizeof(*pl)); | |
907 | EXIT; | |
908 | } | |
909 | EXPORT_SYMBOL(ldlm_pool_fini); | |
910 | ||
911 | /** | |
912 | * Add new taken ldlm lock \a lock into pool \a pl accounting. | |
913 | */ | |
914 | void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock) | |
915 | { | |
916 | /* | |
917 | * FLOCK locks are special in a sense that they are almost never | |
918 | * cancelled, instead special kind of lock is used to drop them. | |
919 | * also there is no LRU for flock locks, so no point in tracking | |
920 | * them anyway. | |
921 | */ | |
922 | if (lock->l_resource->lr_type == LDLM_FLOCK) | |
923 | return; | |
924 | ||
925 | atomic_inc(&pl->pl_granted); | |
926 | atomic_inc(&pl->pl_grant_rate); | |
927 | lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT); | |
928 | /* | |
929 | * Do not do pool recalc for client side as all locks which | |
930 | * potentially may be canceled has already been packed into | |
931 | * enqueue/cancel rpc. Also we do not want to run out of stack | |
932 | * with too long call paths. | |
933 | */ | |
934 | if (ns_is_server(ldlm_pl2ns(pl))) | |
935 | ldlm_pool_recalc(pl); | |
936 | } | |
937 | EXPORT_SYMBOL(ldlm_pool_add); | |
938 | ||
939 | /** | |
940 | * Remove ldlm lock \a lock from pool \a pl accounting. | |
941 | */ | |
942 | void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock) | |
943 | { | |
944 | /* | |
945 | * Filter out FLOCK locks. Read above comment in ldlm_pool_add(). | |
946 | */ | |
947 | if (lock->l_resource->lr_type == LDLM_FLOCK) | |
948 | return; | |
949 | ||
950 | LASSERT(atomic_read(&pl->pl_granted) > 0); | |
951 | atomic_dec(&pl->pl_granted); | |
952 | atomic_inc(&pl->pl_cancel_rate); | |
953 | ||
954 | lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT); | |
955 | ||
956 | if (ns_is_server(ldlm_pl2ns(pl))) | |
957 | ldlm_pool_recalc(pl); | |
958 | } | |
959 | EXPORT_SYMBOL(ldlm_pool_del); | |
960 | ||
961 | /** | |
962 | * Returns current \a pl SLV. | |
963 | * | |
964 | * \pre ->pl_lock is not locked. | |
965 | */ | |
966 | __u64 ldlm_pool_get_slv(struct ldlm_pool *pl) | |
967 | { | |
968 | __u64 slv; | |
969 | spin_lock(&pl->pl_lock); | |
970 | slv = pl->pl_server_lock_volume; | |
971 | spin_unlock(&pl->pl_lock); | |
972 | return slv; | |
973 | } | |
974 | EXPORT_SYMBOL(ldlm_pool_get_slv); | |
975 | ||
976 | /** | |
977 | * Sets passed \a slv to \a pl. | |
978 | * | |
979 | * \pre ->pl_lock is not locked. | |
980 | */ | |
981 | void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv) | |
982 | { | |
983 | spin_lock(&pl->pl_lock); | |
984 | pl->pl_server_lock_volume = slv; | |
985 | spin_unlock(&pl->pl_lock); | |
986 | } | |
987 | EXPORT_SYMBOL(ldlm_pool_set_slv); | |
988 | ||
989 | /** | |
990 | * Returns current \a pl CLV. | |
991 | * | |
992 | * \pre ->pl_lock is not locked. | |
993 | */ | |
994 | __u64 ldlm_pool_get_clv(struct ldlm_pool *pl) | |
995 | { | |
996 | __u64 slv; | |
997 | spin_lock(&pl->pl_lock); | |
998 | slv = pl->pl_client_lock_volume; | |
999 | spin_unlock(&pl->pl_lock); | |
1000 | return slv; | |
1001 | } | |
1002 | EXPORT_SYMBOL(ldlm_pool_get_clv); | |
1003 | ||
1004 | /** | |
1005 | * Sets passed \a clv to \a pl. | |
1006 | * | |
1007 | * \pre ->pl_lock is not locked. | |
1008 | */ | |
1009 | void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv) | |
1010 | { | |
1011 | spin_lock(&pl->pl_lock); | |
1012 | pl->pl_client_lock_volume = clv; | |
1013 | spin_unlock(&pl->pl_lock); | |
1014 | } | |
1015 | EXPORT_SYMBOL(ldlm_pool_set_clv); | |
1016 | ||
1017 | /** | |
1018 | * Returns current \a pl limit. | |
1019 | */ | |
1020 | __u32 ldlm_pool_get_limit(struct ldlm_pool *pl) | |
1021 | { | |
1022 | return atomic_read(&pl->pl_limit); | |
1023 | } | |
1024 | EXPORT_SYMBOL(ldlm_pool_get_limit); | |
1025 | ||
1026 | /** | |
1027 | * Sets passed \a limit to \a pl. | |
1028 | */ | |
1029 | void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit) | |
1030 | { | |
1031 | atomic_set(&pl->pl_limit, limit); | |
1032 | } | |
1033 | EXPORT_SYMBOL(ldlm_pool_set_limit); | |
1034 | ||
1035 | /** | |
1036 | * Returns current LVF from \a pl. | |
1037 | */ | |
1038 | __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl) | |
1039 | { | |
1040 | return atomic_read(&pl->pl_lock_volume_factor); | |
1041 | } | |
1042 | EXPORT_SYMBOL(ldlm_pool_get_lvf); | |
1043 | ||
1044 | static int ldlm_pool_granted(struct ldlm_pool *pl) | |
1045 | { | |
1046 | return atomic_read(&pl->pl_granted); | |
1047 | } | |
1048 | ||
1049 | static struct ptlrpc_thread *ldlm_pools_thread; | |
1050 | static struct shrinker *ldlm_pools_srv_shrinker; | |
1051 | static struct shrinker *ldlm_pools_cli_shrinker; | |
1052 | static struct completion ldlm_pools_comp; | |
1053 | ||
1054 | /* | |
1055 | * Cancel \a nr locks from all namespaces (if possible). Returns number of | |
1056 | * cached locks after shrink is finished. All namespaces are asked to | |
1057 | * cancel approximately equal amount of locks to keep balancing. | |
1058 | */ | |
1059 | static int ldlm_pools_shrink(ldlm_side_t client, int nr, | |
1060 | unsigned int gfp_mask) | |
1061 | { | |
1062 | int total = 0, cached = 0, nr_ns; | |
1063 | struct ldlm_namespace *ns; | |
1064 | void *cookie; | |
1065 | ||
1066 | if (client == LDLM_NAMESPACE_CLIENT && nr != 0 && | |
1067 | !(gfp_mask & __GFP_FS)) | |
1068 | return -1; | |
1069 | ||
1070 | CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n", | |
1071 | nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server"); | |
1072 | ||
1073 | cookie = cl_env_reenter(); | |
1074 | ||
1075 | /* | |
1076 | * Find out how many resources we may release. | |
1077 | */ | |
1078 | for (nr_ns = atomic_read(ldlm_namespace_nr(client)); | |
1079 | nr_ns > 0; nr_ns--) | |
1080 | { | |
1081 | mutex_lock(ldlm_namespace_lock(client)); | |
1082 | if (list_empty(ldlm_namespace_list(client))) { | |
1083 | mutex_unlock(ldlm_namespace_lock(client)); | |
1084 | cl_env_reexit(cookie); | |
1085 | return 0; | |
1086 | } | |
1087 | ns = ldlm_namespace_first_locked(client); | |
1088 | ldlm_namespace_get(ns); | |
1089 | ldlm_namespace_move_locked(ns, client); | |
1090 | mutex_unlock(ldlm_namespace_lock(client)); | |
1091 | total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask); | |
1092 | ldlm_namespace_put(ns); | |
1093 | } | |
1094 | ||
1095 | if (nr == 0 || total == 0) { | |
1096 | cl_env_reexit(cookie); | |
1097 | return total; | |
1098 | } | |
1099 | ||
1100 | /* | |
1101 | * Shrink at least ldlm_namespace_nr(client) namespaces. | |
1102 | */ | |
1103 | for (nr_ns = atomic_read(ldlm_namespace_nr(client)); | |
1104 | nr_ns > 0; nr_ns--) | |
1105 | { | |
1106 | int cancel, nr_locks; | |
1107 | ||
1108 | /* | |
1109 | * Do not call shrink under ldlm_namespace_lock(client) | |
1110 | */ | |
1111 | mutex_lock(ldlm_namespace_lock(client)); | |
1112 | if (list_empty(ldlm_namespace_list(client))) { | |
1113 | mutex_unlock(ldlm_namespace_lock(client)); | |
1114 | /* | |
1115 | * If list is empty, we can't return any @cached > 0, | |
1116 | * that probably would cause needless shrinker | |
1117 | * call. | |
1118 | */ | |
1119 | cached = 0; | |
1120 | break; | |
1121 | } | |
1122 | ns = ldlm_namespace_first_locked(client); | |
1123 | ldlm_namespace_get(ns); | |
1124 | ldlm_namespace_move_locked(ns, client); | |
1125 | mutex_unlock(ldlm_namespace_lock(client)); | |
1126 | ||
1127 | nr_locks = ldlm_pool_granted(&ns->ns_pool); | |
1128 | cancel = 1 + nr_locks * nr / total; | |
1129 | ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask); | |
1130 | cached += ldlm_pool_granted(&ns->ns_pool); | |
1131 | ldlm_namespace_put(ns); | |
1132 | } | |
1133 | cl_env_reexit(cookie); | |
1134 | /* we only decrease the SLV in server pools shrinker, return -1 to | |
1135 | * kernel to avoid needless loop. LU-1128 */ | |
1136 | return (client == LDLM_NAMESPACE_SERVER) ? -1 : cached; | |
1137 | } | |
1138 | ||
1139 | static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask)) | |
1140 | { | |
1141 | return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER, | |
1142 | shrink_param(sc, nr_to_scan), | |
1143 | shrink_param(sc, gfp_mask)); | |
1144 | } | |
1145 | ||
1146 | static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask)) | |
1147 | { | |
1148 | return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT, | |
1149 | shrink_param(sc, nr_to_scan), | |
1150 | shrink_param(sc, gfp_mask)); | |
1151 | } | |
1152 | ||
1153 | void ldlm_pools_recalc(ldlm_side_t client) | |
1154 | { | |
1155 | __u32 nr_l = 0, nr_p = 0, l; | |
1156 | struct ldlm_namespace *ns; | |
1157 | int nr, equal = 0; | |
1158 | ||
1159 | /* | |
1160 | * No need to setup pool limit for client pools. | |
1161 | */ | |
1162 | if (client == LDLM_NAMESPACE_SERVER) { | |
1163 | /* | |
1164 | * Check all modest namespaces first. | |
1165 | */ | |
1166 | mutex_lock(ldlm_namespace_lock(client)); | |
1167 | list_for_each_entry(ns, ldlm_namespace_list(client), | |
1168 | ns_list_chain) | |
1169 | { | |
1170 | if (ns->ns_appetite != LDLM_NAMESPACE_MODEST) | |
1171 | continue; | |
1172 | ||
1173 | l = ldlm_pool_granted(&ns->ns_pool); | |
1174 | if (l == 0) | |
1175 | l = 1; | |
1176 | ||
1177 | /* | |
1178 | * Set the modest pools limit equal to their avg granted | |
1179 | * locks + ~6%. | |
1180 | */ | |
1181 | l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0); | |
1182 | ldlm_pool_setup(&ns->ns_pool, l); | |
1183 | nr_l += l; | |
1184 | nr_p++; | |
1185 | } | |
1186 | ||
1187 | /* | |
1188 | * Make sure that modest namespaces did not eat more that 2/3 | |
1189 | * of limit. | |
1190 | */ | |
1191 | if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) { | |
1192 | CWARN("\"Modest\" pools eat out 2/3 of server locks " | |
1193 | "limit (%d of %lu). This means that you have too " | |
1194 | "many clients for this amount of server RAM. " | |
1195 | "Upgrade server!\n", nr_l, LDLM_POOL_HOST_L); | |
1196 | equal = 1; | |
1197 | } | |
1198 | ||
1199 | /* | |
1200 | * The rest is given to greedy namespaces. | |
1201 | */ | |
1202 | list_for_each_entry(ns, ldlm_namespace_list(client), | |
1203 | ns_list_chain) | |
1204 | { | |
1205 | if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY) | |
1206 | continue; | |
1207 | ||
1208 | if (equal) { | |
1209 | /* | |
1210 | * In the case 2/3 locks are eaten out by | |
1211 | * modest pools, we re-setup equal limit | |
1212 | * for _all_ pools. | |
1213 | */ | |
1214 | l = LDLM_POOL_HOST_L / | |
1215 | atomic_read( | |
1216 | ldlm_namespace_nr(client)); | |
1217 | } else { | |
1218 | /* | |
1219 | * All the rest of greedy pools will have | |
1220 | * all locks in equal parts. | |
1221 | */ | |
1222 | l = (LDLM_POOL_HOST_L - nr_l) / | |
1223 | (atomic_read( | |
1224 | ldlm_namespace_nr(client)) - | |
1225 | nr_p); | |
1226 | } | |
1227 | ldlm_pool_setup(&ns->ns_pool, l); | |
1228 | } | |
1229 | mutex_unlock(ldlm_namespace_lock(client)); | |
1230 | } | |
1231 | ||
1232 | /* | |
1233 | * Recalc at least ldlm_namespace_nr(client) namespaces. | |
1234 | */ | |
1235 | for (nr = atomic_read(ldlm_namespace_nr(client)); nr > 0; nr--) { | |
1236 | int skip; | |
1237 | /* | |
1238 | * Lock the list, get first @ns in the list, getref, move it | |
1239 | * to the tail, unlock and call pool recalc. This way we avoid | |
1240 | * calling recalc under @ns lock what is really good as we get | |
1241 | * rid of potential deadlock on client nodes when canceling | |
1242 | * locks synchronously. | |
1243 | */ | |
1244 | mutex_lock(ldlm_namespace_lock(client)); | |
1245 | if (list_empty(ldlm_namespace_list(client))) { | |
1246 | mutex_unlock(ldlm_namespace_lock(client)); | |
1247 | break; | |
1248 | } | |
1249 | ns = ldlm_namespace_first_locked(client); | |
1250 | ||
1251 | spin_lock(&ns->ns_lock); | |
1252 | /* | |
1253 | * skip ns which is being freed, and we don't want to increase | |
1254 | * its refcount again, not even temporarily. bz21519 & LU-499. | |
1255 | */ | |
1256 | if (ns->ns_stopping) { | |
1257 | skip = 1; | |
1258 | } else { | |
1259 | skip = 0; | |
1260 | ldlm_namespace_get(ns); | |
1261 | } | |
1262 | spin_unlock(&ns->ns_lock); | |
1263 | ||
1264 | ldlm_namespace_move_locked(ns, client); | |
1265 | mutex_unlock(ldlm_namespace_lock(client)); | |
1266 | ||
1267 | /* | |
1268 | * After setup is done - recalc the pool. | |
1269 | */ | |
1270 | if (!skip) { | |
1271 | ldlm_pool_recalc(&ns->ns_pool); | |
1272 | ldlm_namespace_put(ns); | |
1273 | } | |
1274 | } | |
1275 | } | |
1276 | EXPORT_SYMBOL(ldlm_pools_recalc); | |
1277 | ||
1278 | static int ldlm_pools_thread_main(void *arg) | |
1279 | { | |
1280 | struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg; | |
1281 | ENTRY; | |
1282 | ||
1283 | thread_set_flags(thread, SVC_RUNNING); | |
1284 | wake_up(&thread->t_ctl_waitq); | |
1285 | ||
1286 | CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n", | |
1287 | "ldlm_poold", current_pid()); | |
1288 | ||
1289 | while (1) { | |
1290 | struct l_wait_info lwi; | |
1291 | ||
1292 | /* | |
1293 | * Recal all pools on this tick. | |
1294 | */ | |
1295 | ldlm_pools_recalc(LDLM_NAMESPACE_SERVER); | |
1296 | ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT); | |
1297 | ||
1298 | /* | |
1299 | * Wait until the next check time, or until we're | |
1300 | * stopped. | |
1301 | */ | |
1302 | lwi = LWI_TIMEOUT(cfs_time_seconds(LDLM_POOLS_THREAD_PERIOD), | |
1303 | NULL, NULL); | |
1304 | l_wait_event(thread->t_ctl_waitq, | |
1305 | thread_is_stopping(thread) || | |
1306 | thread_is_event(thread), | |
1307 | &lwi); | |
1308 | ||
1309 | if (thread_test_and_clear_flags(thread, SVC_STOPPING)) | |
1310 | break; | |
1311 | else | |
1312 | thread_test_and_clear_flags(thread, SVC_EVENT); | |
1313 | } | |
1314 | ||
1315 | thread_set_flags(thread, SVC_STOPPED); | |
1316 | wake_up(&thread->t_ctl_waitq); | |
1317 | ||
1318 | CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n", | |
1319 | "ldlm_poold", current_pid()); | |
1320 | ||
1321 | complete_and_exit(&ldlm_pools_comp, 0); | |
1322 | } | |
1323 | ||
1324 | static int ldlm_pools_thread_start(void) | |
1325 | { | |
1326 | struct l_wait_info lwi = { 0 }; | |
1327 | task_t *task; | |
1328 | ENTRY; | |
1329 | ||
1330 | if (ldlm_pools_thread != NULL) | |
1331 | RETURN(-EALREADY); | |
1332 | ||
1333 | OBD_ALLOC_PTR(ldlm_pools_thread); | |
1334 | if (ldlm_pools_thread == NULL) | |
1335 | RETURN(-ENOMEM); | |
1336 | ||
1337 | init_completion(&ldlm_pools_comp); | |
1338 | init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq); | |
1339 | ||
1340 | task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread, | |
1341 | "ldlm_poold"); | |
1342 | if (IS_ERR(task)) { | |
1343 | CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task)); | |
1344 | OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread)); | |
1345 | ldlm_pools_thread = NULL; | |
1346 | RETURN(PTR_ERR(task)); | |
1347 | } | |
1348 | l_wait_event(ldlm_pools_thread->t_ctl_waitq, | |
1349 | thread_is_running(ldlm_pools_thread), &lwi); | |
1350 | RETURN(0); | |
1351 | } | |
1352 | ||
1353 | static void ldlm_pools_thread_stop(void) | |
1354 | { | |
1355 | ENTRY; | |
1356 | ||
1357 | if (ldlm_pools_thread == NULL) { | |
1358 | EXIT; | |
1359 | return; | |
1360 | } | |
1361 | ||
1362 | thread_set_flags(ldlm_pools_thread, SVC_STOPPING); | |
1363 | wake_up(&ldlm_pools_thread->t_ctl_waitq); | |
1364 | ||
1365 | /* | |
1366 | * Make sure that pools thread is finished before freeing @thread. | |
1367 | * This fixes possible race and oops due to accessing freed memory | |
1368 | * in pools thread. | |
1369 | */ | |
1370 | wait_for_completion(&ldlm_pools_comp); | |
1371 | OBD_FREE_PTR(ldlm_pools_thread); | |
1372 | ldlm_pools_thread = NULL; | |
1373 | EXIT; | |
1374 | } | |
1375 | ||
1376 | int ldlm_pools_init(void) | |
1377 | { | |
1378 | int rc; | |
1379 | ENTRY; | |
1380 | ||
1381 | rc = ldlm_pools_thread_start(); | |
1382 | if (rc == 0) { | |
1383 | ldlm_pools_srv_shrinker = | |
1384 | set_shrinker(DEFAULT_SEEKS, | |
1385 | ldlm_pools_srv_shrink); | |
1386 | ldlm_pools_cli_shrinker = | |
1387 | set_shrinker(DEFAULT_SEEKS, | |
1388 | ldlm_pools_cli_shrink); | |
1389 | } | |
1390 | RETURN(rc); | |
1391 | } | |
1392 | EXPORT_SYMBOL(ldlm_pools_init); | |
1393 | ||
1394 | void ldlm_pools_fini(void) | |
1395 | { | |
1396 | if (ldlm_pools_srv_shrinker != NULL) { | |
1397 | remove_shrinker(ldlm_pools_srv_shrinker); | |
1398 | ldlm_pools_srv_shrinker = NULL; | |
1399 | } | |
1400 | if (ldlm_pools_cli_shrinker != NULL) { | |
1401 | remove_shrinker(ldlm_pools_cli_shrinker); | |
1402 | ldlm_pools_cli_shrinker = NULL; | |
1403 | } | |
1404 | ldlm_pools_thread_stop(); | |
1405 | } | |
1406 | EXPORT_SYMBOL(ldlm_pools_fini); |