]>
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 | * | |
1dc563a6 | 30 | * Copyright (c) 2010, 2015, Intel Corporation. |
d7e09d03 PT |
31 | */ |
32 | /* | |
33 | * This file is part of Lustre, http://www.lustre.org/ | |
34 | * Lustre is a trademark of Sun Microsystems, Inc. | |
35 | * | |
36 | * lustre/ldlm/ldlm_pool.c | |
37 | * | |
38 | * Author: Yury Umanets <umka@clusterfs.com> | |
39 | */ | |
40 | ||
41 | /* | |
42 | * Idea of this code is rather simple. Each second, for each server namespace | |
43 | * we have SLV - server lock volume which is calculated on current number of | |
44 | * granted locks, grant speed for past period, etc - that is, locking load. | |
45 | * This SLV number may be thought as a flow definition for simplicity. It is | |
46 | * sent to clients with each occasion to let them know what is current load | |
47 | * situation on the server. By default, at the beginning, SLV on server is | |
48 | * set max value which is calculated as the following: allow to one client | |
49 | * have all locks of limit ->pl_limit for 10h. | |
50 | * | |
51 | * Next, on clients, number of cached locks is not limited artificially in any | |
52 | * way as it was before. Instead, client calculates CLV, that is, client lock | |
53 | * volume for each lock and compares it with last SLV from the server. CLV is | |
54 | * calculated as the number of locks in LRU * lock live time in seconds. If | |
55 | * CLV > SLV - lock is canceled. | |
56 | * | |
e7ddc48c AR |
57 | * Client has LVF, that is, lock volume factor which regulates how much |
58 | * sensitive client should be about last SLV from server. The higher LVF is the | |
59 | * more locks will be canceled on client. Default value for it is 1. Setting LVF | |
60 | * to 2 means that client will cancel locks 2 times faster. | |
d7e09d03 PT |
61 | * |
62 | * Locks on a client will be canceled more intensively in these cases: | |
63 | * (1) if SLV is smaller, that is, load is higher on the server; | |
64 | * (2) client has a lot of locks (the more locks are held by client, the bigger | |
65 | * chances that some of them should be canceled); | |
66 | * (3) client has old locks (taken some time ago); | |
67 | * | |
68 | * Thus, according to flow paradigm that we use for better understanding SLV, | |
69 | * CLV is the volume of particle in flow described by SLV. According to this, | |
70 | * if flow is getting thinner, more and more particles become outside of it and | |
71 | * as particles are locks, they should be canceled. | |
72 | * | |
e7ddc48c AR |
73 | * General idea of this belongs to Vitaly Fertman (vitaly@clusterfs.com). |
74 | * Andreas Dilger (adilger@clusterfs.com) proposed few nice ideas like using | |
75 | * LVF and many cleanups. Flow definition to allow more easy understanding of | |
76 | * the logic belongs to Nikita Danilov (nikita@clusterfs.com) as well as many | |
77 | * cleanups and fixes. And design and implementation are done by Yury Umanets | |
78 | * (umka@clusterfs.com). | |
d7e09d03 PT |
79 | * |
80 | * Glossary for terms used: | |
81 | * | |
82 | * pl_limit - Number of allowed locks in pool. Applies to server and client | |
83 | * side (tunable); | |
84 | * | |
85 | * pl_granted - Number of granted locks (calculated); | |
86 | * pl_grant_rate - Number of granted locks for last T (calculated); | |
87 | * pl_cancel_rate - Number of canceled locks for last T (calculated); | |
88 | * pl_grant_speed - Grant speed (GR - CR) for last T (calculated); | |
89 | * pl_grant_plan - Planned number of granted locks for next T (calculated); | |
90 | * pl_server_lock_volume - Current server lock volume (calculated); | |
91 | * | |
92 | * As it may be seen from list above, we have few possible tunables which may | |
f2825e03 | 93 | * affect behavior much. They all may be modified via sysfs. However, they also |
d7e09d03 PT |
94 | * give a possibility for constructing few pre-defined behavior policies. If |
95 | * none of predefines is suitable for a working pattern being used, new one may | |
f2825e03 | 96 | * be "constructed" via sysfs tunables. |
d7e09d03 PT |
97 | */ |
98 | ||
99 | #define DEBUG_SUBSYSTEM S_LDLM | |
100 | ||
e27db149 GKH |
101 | #include "../include/lustre_dlm.h" |
102 | #include "../include/cl_object.h" | |
103 | #include "../include/obd_class.h" | |
104 | #include "../include/obd_support.h" | |
d7e09d03 PT |
105 | #include "ldlm_internal.h" |
106 | ||
d7e09d03 PT |
107 | /* |
108 | * 50 ldlm locks for 1MB of RAM. | |
109 | */ | |
110 | #define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_CACHE_SHIFT)) * 50) | |
111 | ||
112 | /* | |
113 | * Maximal possible grant step plan in %. | |
114 | */ | |
115 | #define LDLM_POOL_MAX_GSP (30) | |
116 | ||
117 | /* | |
118 | * Minimal possible grant step plan in %. | |
119 | */ | |
120 | #define LDLM_POOL_MIN_GSP (1) | |
121 | ||
122 | /* | |
123 | * This controls the speed of reaching LDLM_POOL_MAX_GSP | |
124 | * with increasing thread period. | |
125 | */ | |
126 | #define LDLM_POOL_GSP_STEP_SHIFT (2) | |
127 | ||
128 | /* | |
129 | * LDLM_POOL_GSP% of all locks is default GP. | |
130 | */ | |
131 | #define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100) | |
132 | ||
133 | /* | |
134 | * Max age for locks on clients. | |
135 | */ | |
136 | #define LDLM_POOL_MAX_AGE (36000) | |
137 | ||
138 | /* | |
139 | * The granularity of SLV calculation. | |
140 | */ | |
141 | #define LDLM_POOL_SLV_SHIFT (10) | |
142 | ||
d7e09d03 PT |
143 | static inline __u64 dru(__u64 val, __u32 shift, int round_up) |
144 | { | |
145 | return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift; | |
146 | } | |
147 | ||
148 | static inline __u64 ldlm_pool_slv_max(__u32 L) | |
149 | { | |
150 | /* | |
151 | * Allow to have all locks for 1 client for 10 hrs. | |
152 | * Formula is the following: limit * 10h / 1 client. | |
153 | */ | |
154 | __u64 lim = (__u64)L * LDLM_POOL_MAX_AGE / 1; | |
155 | return lim; | |
156 | } | |
157 | ||
158 | static inline __u64 ldlm_pool_slv_min(__u32 L) | |
159 | { | |
160 | return 1; | |
161 | } | |
162 | ||
163 | enum { | |
164 | LDLM_POOL_FIRST_STAT = 0, | |
165 | LDLM_POOL_GRANTED_STAT = LDLM_POOL_FIRST_STAT, | |
166 | LDLM_POOL_GRANT_STAT, | |
167 | LDLM_POOL_CANCEL_STAT, | |
168 | LDLM_POOL_GRANT_RATE_STAT, | |
169 | LDLM_POOL_CANCEL_RATE_STAT, | |
170 | LDLM_POOL_GRANT_PLAN_STAT, | |
171 | LDLM_POOL_SLV_STAT, | |
172 | LDLM_POOL_SHRINK_REQTD_STAT, | |
173 | LDLM_POOL_SHRINK_FREED_STAT, | |
174 | LDLM_POOL_RECALC_STAT, | |
175 | LDLM_POOL_TIMING_STAT, | |
176 | LDLM_POOL_LAST_STAT | |
177 | }; | |
178 | ||
d7e09d03 PT |
179 | /** |
180 | * Calculates suggested grant_step in % of available locks for passed | |
181 | * \a period. This is later used in grant_plan calculations. | |
182 | */ | |
183 | static inline int ldlm_pool_t2gsp(unsigned int t) | |
184 | { | |
185 | /* | |
186 | * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP | |
187 | * and up to 30% for anything higher than LDLM_POOL_GSP_STEP. | |
188 | * | |
189 | * How this will affect execution is the following: | |
190 | * | |
191 | * - for thread period 1s we will have grant_step 1% which good from | |
192 | * pov of taking some load off from server and push it out to clients. | |
193 | * This is like that because 1% for grant_step means that server will | |
194 | * not allow clients to get lots of locks in short period of time and | |
195 | * keep all old locks in their caches. Clients will always have to | |
196 | * get some locks back if they want to take some new; | |
197 | * | |
198 | * - for thread period 10s (which is default) we will have 23% which | |
199 | * means that clients will have enough of room to take some new locks | |
200 | * without getting some back. All locks from this 23% which were not | |
201 | * taken by clients in current period will contribute in SLV growing. | |
202 | * SLV growing means more locks cached on clients until limit or grant | |
203 | * plan is reached. | |
204 | */ | |
205 | return LDLM_POOL_MAX_GSP - | |
206 | ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >> | |
207 | (t >> LDLM_POOL_GSP_STEP_SHIFT)); | |
208 | } | |
209 | ||
d7e09d03 PT |
210 | /** |
211 | * Recalculates next stats on passed \a pl. | |
212 | * | |
213 | * \pre ->pl_lock is locked. | |
214 | */ | |
215 | static void ldlm_pool_recalc_stats(struct ldlm_pool *pl) | |
216 | { | |
217 | int grant_plan = pl->pl_grant_plan; | |
218 | __u64 slv = pl->pl_server_lock_volume; | |
219 | int granted = atomic_read(&pl->pl_granted); | |
220 | int grant_rate = atomic_read(&pl->pl_grant_rate); | |
221 | int cancel_rate = atomic_read(&pl->pl_cancel_rate); | |
222 | ||
223 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT, | |
224 | slv); | |
225 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT, | |
226 | granted); | |
227 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT, | |
228 | grant_rate); | |
229 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT, | |
230 | grant_plan); | |
231 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT, | |
232 | cancel_rate); | |
233 | } | |
234 | ||
d7e09d03 | 235 | /** |
7c37abe0 SB |
236 | * Sets SLV and Limit from container_of(pl, struct ldlm_namespace, |
237 | * ns_pool)->ns_obd tp passed \a pl. | |
d7e09d03 PT |
238 | */ |
239 | static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl) | |
240 | { | |
241 | struct obd_device *obd; | |
242 | ||
243 | /* | |
244 | * Get new SLV and Limit from obd which is updated with coming | |
245 | * RPCs. | |
246 | */ | |
7c37abe0 SB |
247 | obd = container_of(pl, struct ldlm_namespace, |
248 | ns_pool)->ns_obd; | |
d7e09d03 PT |
249 | LASSERT(obd != NULL); |
250 | read_lock(&obd->obd_pool_lock); | |
251 | pl->pl_server_lock_volume = obd->obd_pool_slv; | |
f7ec22b5 | 252 | atomic_set(&pl->pl_limit, obd->obd_pool_limit); |
d7e09d03 PT |
253 | read_unlock(&obd->obd_pool_lock); |
254 | } | |
255 | ||
256 | /** | |
257 | * Recalculates client size pool \a pl according to current SLV and Limit. | |
258 | */ | |
259 | static int ldlm_cli_pool_recalc(struct ldlm_pool *pl) | |
260 | { | |
8f83409c | 261 | time64_t recalc_interval_sec; |
4d2c7b30 | 262 | int ret; |
d7e09d03 | 263 | |
8f83409c | 264 | recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time; |
d7e09d03 | 265 | if (recalc_interval_sec < pl->pl_recalc_period) |
0a3bdb00 | 266 | return 0; |
d7e09d03 PT |
267 | |
268 | spin_lock(&pl->pl_lock); | |
269 | /* | |
270 | * Check if we need to recalc lists now. | |
271 | */ | |
8f83409c | 272 | recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time; |
d7e09d03 PT |
273 | if (recalc_interval_sec < pl->pl_recalc_period) { |
274 | spin_unlock(&pl->pl_lock); | |
0a3bdb00 | 275 | return 0; |
d7e09d03 PT |
276 | } |
277 | ||
278 | /* | |
279 | * Make sure that pool knows last SLV and Limit from obd. | |
280 | */ | |
281 | ldlm_cli_pool_pop_slv(pl); | |
282 | ||
d7e09d03 PT |
283 | spin_unlock(&pl->pl_lock); |
284 | ||
285 | /* | |
286 | * Do not cancel locks in case lru resize is disabled for this ns. | |
287 | */ | |
7c37abe0 SB |
288 | if (!ns_connect_lru_resize(container_of(pl, struct ldlm_namespace, |
289 | ns_pool))) { | |
4d2c7b30 LX |
290 | ret = 0; |
291 | goto out; | |
292 | } | |
d7e09d03 PT |
293 | |
294 | /* | |
295 | * In the time of canceling locks on client we do not need to maintain | |
296 | * sharp timing, we only want to cancel locks asap according to new SLV. | |
297 | * It may be called when SLV has changed much, this is why we do not | |
298 | * take into account pl->pl_recalc_time here. | |
299 | */ | |
7c37abe0 SB |
300 | ret = ldlm_cancel_lru(container_of(pl, struct ldlm_namespace, ns_pool), |
301 | 0, LCF_ASYNC, LDLM_CANCEL_LRUR); | |
4d2c7b30 LX |
302 | |
303 | out: | |
304 | spin_lock(&pl->pl_lock); | |
305 | /* | |
306 | * Time of LRU resizing might be longer than period, | |
307 | * so update after LRU resizing rather than before it. | |
308 | */ | |
8f83409c | 309 | pl->pl_recalc_time = ktime_get_real_seconds(); |
4d2c7b30 LX |
310 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT, |
311 | recalc_interval_sec); | |
312 | spin_unlock(&pl->pl_lock); | |
313 | return ret; | |
d7e09d03 PT |
314 | } |
315 | ||
316 | /** | |
317 | * This function is main entry point for memory pressure handling on client | |
318 | * side. Main goal of this function is to cancel some number of locks on | |
319 | * passed \a pl according to \a nr and \a gfp_mask. | |
320 | */ | |
321 | static int ldlm_cli_pool_shrink(struct ldlm_pool *pl, | |
5802572e | 322 | int nr, gfp_t gfp_mask) |
d7e09d03 PT |
323 | { |
324 | struct ldlm_namespace *ns; | |
cbc3769e | 325 | int unused; |
d7e09d03 | 326 | |
7c37abe0 | 327 | ns = container_of(pl, struct ldlm_namespace, ns_pool); |
d7e09d03 PT |
328 | |
329 | /* | |
330 | * Do not cancel locks in case lru resize is disabled for this ns. | |
331 | */ | |
332 | if (!ns_connect_lru_resize(ns)) | |
0a3bdb00 | 333 | return 0; |
d7e09d03 PT |
334 | |
335 | /* | |
336 | * Make sure that pool knows last SLV and Limit from obd. | |
337 | */ | |
338 | ldlm_cli_pool_pop_slv(pl); | |
339 | ||
340 | spin_lock(&ns->ns_lock); | |
341 | unused = ns->ns_nr_unused; | |
342 | spin_unlock(&ns->ns_lock); | |
343 | ||
cbc3769e PT |
344 | if (nr == 0) |
345 | return (unused / 100) * sysctl_vfs_cache_pressure; | |
346 | else | |
347 | return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_CANCEL_SHRINK); | |
d7e09d03 PT |
348 | } |
349 | ||
b9c98cfa | 350 | static const struct ldlm_pool_ops ldlm_cli_pool_ops = { |
d7e09d03 PT |
351 | .po_recalc = ldlm_cli_pool_recalc, |
352 | .po_shrink = ldlm_cli_pool_shrink | |
353 | }; | |
354 | ||
355 | /** | |
356 | * Pool recalc wrapper. Will call either client or server pool recalc callback | |
357 | * depending what pool \a pl is used. | |
358 | */ | |
58c6d133 | 359 | static int ldlm_pool_recalc(struct ldlm_pool *pl) |
d7e09d03 | 360 | { |
8f83409c | 361 | u32 recalc_interval_sec; |
d7e09d03 PT |
362 | int count; |
363 | ||
8f83409c | 364 | recalc_interval_sec = ktime_get_seconds() - pl->pl_recalc_time; |
d7e09d03 PT |
365 | if (recalc_interval_sec <= 0) |
366 | goto recalc; | |
367 | ||
368 | spin_lock(&pl->pl_lock); | |
d7e09d03 PT |
369 | if (recalc_interval_sec > 0) { |
370 | /* | |
371 | * Update pool statistics every 1s. | |
372 | */ | |
373 | ldlm_pool_recalc_stats(pl); | |
374 | ||
375 | /* | |
376 | * Zero out all rates and speed for the last period. | |
377 | */ | |
378 | atomic_set(&pl->pl_grant_rate, 0); | |
379 | atomic_set(&pl->pl_cancel_rate, 0); | |
380 | } | |
381 | spin_unlock(&pl->pl_lock); | |
382 | ||
383 | recalc: | |
384 | if (pl->pl_ops->po_recalc != NULL) { | |
385 | count = pl->pl_ops->po_recalc(pl); | |
386 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT, | |
387 | count); | |
d7e09d03 | 388 | } |
8f83409c | 389 | recalc_interval_sec = pl->pl_recalc_time - ktime_get_seconds() + |
3eface59 | 390 | pl->pl_recalc_period; |
4d2c7b30 LX |
391 | if (recalc_interval_sec <= 0) { |
392 | /* Prevent too frequent recalculation. */ | |
8f83409c AB |
393 | CDEBUG(D_DLMTRACE, |
394 | "Negative interval(%d), too short period(%lld)", | |
4d2c7b30 | 395 | recalc_interval_sec, |
8f83409c | 396 | (s64)pl->pl_recalc_period); |
4d2c7b30 LX |
397 | recalc_interval_sec = 1; |
398 | } | |
d7e09d03 | 399 | |
3eface59 | 400 | return recalc_interval_sec; |
d7e09d03 | 401 | } |
d7e09d03 | 402 | |
cbc3769e | 403 | /* |
d7e09d03 | 404 | * Pool shrink wrapper. Will call either client or server pool recalc callback |
cbc3769e PT |
405 | * depending what pool pl is used. When nr == 0, just return the number of |
406 | * freeable locks. Otherwise, return the number of canceled locks. | |
d7e09d03 | 407 | */ |
58c6d133 | 408 | static int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, gfp_t gfp_mask) |
d7e09d03 PT |
409 | { |
410 | int cancel = 0; | |
411 | ||
412 | if (pl->pl_ops->po_shrink != NULL) { | |
413 | cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask); | |
414 | if (nr > 0) { | |
415 | lprocfs_counter_add(pl->pl_stats, | |
416 | LDLM_POOL_SHRINK_REQTD_STAT, | |
417 | nr); | |
418 | lprocfs_counter_add(pl->pl_stats, | |
419 | LDLM_POOL_SHRINK_FREED_STAT, | |
420 | cancel); | |
2d00bd17 JP |
421 | CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, shrunk %d\n", |
422 | pl->pl_name, nr, cancel); | |
d7e09d03 PT |
423 | } |
424 | } | |
425 | return cancel; | |
426 | } | |
d7e09d03 | 427 | |
73bb1da6 | 428 | static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused) |
d7e09d03 | 429 | { |
71570b98 OD |
430 | int granted, grant_rate, cancel_rate; |
431 | int grant_speed, lvf; | |
73bb1da6 | 432 | struct ldlm_pool *pl = m->private; |
d7e09d03 PT |
433 | __u64 slv, clv; |
434 | __u32 limit; | |
435 | ||
436 | spin_lock(&pl->pl_lock); | |
437 | slv = pl->pl_server_lock_volume; | |
438 | clv = pl->pl_client_lock_volume; | |
946d6f95 | 439 | limit = atomic_read(&pl->pl_limit); |
d7e09d03 PT |
440 | granted = atomic_read(&pl->pl_granted); |
441 | grant_rate = atomic_read(&pl->pl_grant_rate); | |
442 | cancel_rate = atomic_read(&pl->pl_cancel_rate); | |
443 | grant_speed = grant_rate - cancel_rate; | |
444 | lvf = atomic_read(&pl->pl_lock_volume_factor); | |
d7e09d03 PT |
445 | spin_unlock(&pl->pl_lock); |
446 | ||
73bb1da6 | 447 | seq_printf(m, "LDLM pool state (%s):\n" |
b0f5aad5 GKH |
448 | " SLV: %llu\n" |
449 | " CLV: %llu\n" | |
73bb1da6 PT |
450 | " LVF: %d\n", |
451 | pl->pl_name, slv, clv, lvf); | |
d7e09d03 | 452 | |
2c2b7c05 HM |
453 | seq_printf(m, " GR: %d\n CR: %d\n GS: %d\n" |
454 | " G: %d\n L: %d\n", | |
73bb1da6 PT |
455 | grant_rate, cancel_rate, grant_speed, |
456 | granted, limit); | |
457 | ||
458 | return 0; | |
d7e09d03 | 459 | } |
c9f6bb96 | 460 | |
73bb1da6 | 461 | LPROC_SEQ_FOPS_RO(lprocfs_pool_state); |
d7e09d03 | 462 | |
24b8c88a OD |
463 | static ssize_t grant_speed_show(struct kobject *kobj, struct attribute *attr, |
464 | char *buf) | |
d7e09d03 | 465 | { |
24b8c88a OD |
466 | struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool, |
467 | pl_kobj); | |
468 | ||
d7e09d03 PT |
469 | int grant_speed; |
470 | ||
471 | spin_lock(&pl->pl_lock); | |
472 | /* serialize with ldlm_pool_recalc */ | |
473 | grant_speed = atomic_read(&pl->pl_grant_rate) - | |
474 | atomic_read(&pl->pl_cancel_rate); | |
475 | spin_unlock(&pl->pl_lock); | |
24b8c88a | 476 | return sprintf(buf, "%d\n", grant_speed); |
d7e09d03 | 477 | } |
24b8c88a | 478 | LUSTRE_RO_ATTR(grant_speed); |
d7e09d03 | 479 | |
24b8c88a OD |
480 | LDLM_POOL_SYSFS_READER_SHOW(grant_plan, int); |
481 | LUSTRE_RO_ATTR(grant_plan); | |
73bb1da6 | 482 | |
24b8c88a OD |
483 | LDLM_POOL_SYSFS_READER_SHOW(recalc_period, int); |
484 | LDLM_POOL_SYSFS_WRITER_STORE(recalc_period, int); | |
485 | LUSTRE_RW_ATTR(recalc_period); | |
73bb1da6 | 486 | |
24b8c88a OD |
487 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(server_lock_volume, u64); |
488 | LUSTRE_RO_ATTR(server_lock_volume); | |
489 | ||
490 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(limit, atomic); | |
491 | LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(limit, atomic); | |
492 | LUSTRE_RW_ATTR(limit); | |
493 | ||
494 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(granted, atomic); | |
495 | LUSTRE_RO_ATTR(granted); | |
496 | ||
497 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(cancel_rate, atomic); | |
498 | LUSTRE_RO_ATTR(cancel_rate); | |
73bb1da6 | 499 | |
24b8c88a OD |
500 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(grant_rate, atomic); |
501 | LUSTRE_RO_ATTR(grant_rate); | |
73bb1da6 | 502 | |
24b8c88a OD |
503 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(lock_volume_factor, atomic); |
504 | LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(lock_volume_factor, atomic); | |
505 | LUSTRE_RW_ATTR(lock_volume_factor); | |
73bb1da6 PT |
506 | |
507 | #define LDLM_POOL_ADD_VAR(name, var, ops) \ | |
508 | do { \ | |
509 | snprintf(var_name, MAX_STRING_SIZE, #name); \ | |
510 | pool_vars[0].data = var; \ | |
511 | pool_vars[0].fops = ops; \ | |
700815d4 | 512 | ldebugfs_add_vars(pl->pl_debugfs_entry, pool_vars, NULL);\ |
73bb1da6 | 513 | } while (0) |
d7e09d03 | 514 | |
f2825e03 OD |
515 | /* These are for pools in /sys/fs/lustre/ldlm/namespaces/.../pool */ |
516 | static struct attribute *ldlm_pl_attrs[] = { | |
24b8c88a OD |
517 | &lustre_attr_grant_speed.attr, |
518 | &lustre_attr_grant_plan.attr, | |
519 | &lustre_attr_recalc_period.attr, | |
520 | &lustre_attr_server_lock_volume.attr, | |
521 | &lustre_attr_limit.attr, | |
522 | &lustre_attr_granted.attr, | |
523 | &lustre_attr_cancel_rate.attr, | |
524 | &lustre_attr_grant_rate.attr, | |
525 | &lustre_attr_lock_volume_factor.attr, | |
f2825e03 OD |
526 | NULL, |
527 | }; | |
528 | ||
529 | static void ldlm_pl_release(struct kobject *kobj) | |
530 | { | |
531 | struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool, | |
532 | pl_kobj); | |
533 | complete(&pl->pl_kobj_unregister); | |
534 | } | |
535 | ||
536 | static struct kobj_type ldlm_pl_ktype = { | |
537 | .default_attrs = ldlm_pl_attrs, | |
538 | .sysfs_ops = &lustre_sysfs_ops, | |
539 | .release = ldlm_pl_release, | |
540 | }; | |
541 | ||
542 | static int ldlm_pool_sysfs_init(struct ldlm_pool *pl) | |
543 | { | |
7c37abe0 SB |
544 | struct ldlm_namespace *ns = container_of(pl, struct ldlm_namespace, |
545 | ns_pool); | |
f2825e03 OD |
546 | int err; |
547 | ||
548 | init_completion(&pl->pl_kobj_unregister); | |
549 | err = kobject_init_and_add(&pl->pl_kobj, &ldlm_pl_ktype, &ns->ns_kobj, | |
550 | "pool"); | |
551 | ||
552 | return err; | |
553 | } | |
554 | ||
700815d4 | 555 | static int ldlm_pool_debugfs_init(struct ldlm_pool *pl) |
d7e09d03 | 556 | { |
7c37abe0 SB |
557 | struct ldlm_namespace *ns = container_of(pl, struct ldlm_namespace, |
558 | ns_pool); | |
700815d4 | 559 | struct dentry *debugfs_ns_parent; |
d7e09d03 PT |
560 | struct lprocfs_vars pool_vars[2]; |
561 | char *var_name = NULL; | |
562 | int rc = 0; | |
d7e09d03 | 563 | |
352f7891 | 564 | var_name = kzalloc(MAX_STRING_SIZE + 1, GFP_NOFS); |
d7e09d03 | 565 | if (!var_name) |
0a3bdb00 | 566 | return -ENOMEM; |
d7e09d03 | 567 | |
700815d4 DE |
568 | debugfs_ns_parent = ns->ns_debugfs_entry; |
569 | if (IS_ERR_OR_NULL(debugfs_ns_parent)) { | |
570 | CERROR("%s: debugfs entry is not initialized\n", | |
d7e09d03 | 571 | ldlm_ns_name(ns)); |
d1c0d446 JL |
572 | rc = -EINVAL; |
573 | goto out_free_name; | |
d7e09d03 | 574 | } |
700815d4 DE |
575 | pl->pl_debugfs_entry = ldebugfs_register("pool", debugfs_ns_parent, |
576 | NULL, NULL); | |
577 | if (IS_ERR(pl->pl_debugfs_entry)) { | |
578 | CERROR("LdebugFS failed in ldlm-pool-init\n"); | |
579 | rc = PTR_ERR(pl->pl_debugfs_entry); | |
580 | pl->pl_debugfs_entry = NULL; | |
d1c0d446 | 581 | goto out_free_name; |
d7e09d03 PT |
582 | } |
583 | ||
584 | var_name[MAX_STRING_SIZE] = '\0'; | |
585 | memset(pool_vars, 0, sizeof(pool_vars)); | |
586 | pool_vars[0].name = var_name; | |
587 | ||
700815d4 | 588 | LDLM_POOL_ADD_VAR(state, pl, &lprocfs_pool_state_fops); |
d7e09d03 PT |
589 | |
590 | pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT - | |
591 | LDLM_POOL_FIRST_STAT, 0); | |
d1c0d446 JL |
592 | if (!pl->pl_stats) { |
593 | rc = -ENOMEM; | |
594 | goto out_free_name; | |
595 | } | |
d7e09d03 PT |
596 | |
597 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT, | |
598 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
599 | "granted", "locks"); | |
600 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT, | |
601 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
602 | "grant", "locks"); | |
603 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT, | |
604 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
605 | "cancel", "locks"); | |
606 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT, | |
607 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
608 | "grant_rate", "locks/s"); | |
609 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT, | |
610 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
611 | "cancel_rate", "locks/s"); | |
612 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT, | |
613 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
614 | "grant_plan", "locks/s"); | |
615 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT, | |
616 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
617 | "slv", "slv"); | |
618 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT, | |
619 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
620 | "shrink_request", "locks"); | |
621 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT, | |
622 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
623 | "shrink_freed", "locks"); | |
624 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT, | |
625 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
626 | "recalc_freed", "locks"); | |
627 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT, | |
628 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
629 | "recalc_timing", "sec"); | |
700815d4 DE |
630 | rc = ldebugfs_register_stats(pl->pl_debugfs_entry, "stats", |
631 | pl->pl_stats); | |
d7e09d03 | 632 | |
d7e09d03 | 633 | out_free_name: |
352f7891 | 634 | kfree(var_name); |
d7e09d03 PT |
635 | return rc; |
636 | } | |
637 | ||
f2825e03 OD |
638 | static void ldlm_pool_sysfs_fini(struct ldlm_pool *pl) |
639 | { | |
640 | kobject_put(&pl->pl_kobj); | |
641 | wait_for_completion(&pl->pl_kobj_unregister); | |
642 | } | |
643 | ||
700815d4 | 644 | static void ldlm_pool_debugfs_fini(struct ldlm_pool *pl) |
d7e09d03 PT |
645 | { |
646 | if (pl->pl_stats != NULL) { | |
647 | lprocfs_free_stats(&pl->pl_stats); | |
648 | pl->pl_stats = NULL; | |
649 | } | |
700815d4 DE |
650 | if (pl->pl_debugfs_entry != NULL) { |
651 | ldebugfs_remove(&pl->pl_debugfs_entry); | |
652 | pl->pl_debugfs_entry = NULL; | |
d7e09d03 PT |
653 | } |
654 | } | |
655 | ||
656 | int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns, | |
657 | int idx, ldlm_side_t client) | |
658 | { | |
659 | int rc; | |
d7e09d03 PT |
660 | |
661 | spin_lock_init(&pl->pl_lock); | |
662 | atomic_set(&pl->pl_granted, 0); | |
8f83409c | 663 | pl->pl_recalc_time = ktime_get_seconds(); |
d7e09d03 PT |
664 | atomic_set(&pl->pl_lock_volume_factor, 1); |
665 | ||
666 | atomic_set(&pl->pl_grant_rate, 0); | |
667 | atomic_set(&pl->pl_cancel_rate, 0); | |
668 | pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L); | |
669 | ||
670 | snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d", | |
671 | ldlm_ns_name(ns), idx); | |
672 | ||
f7ec22b5 | 673 | atomic_set(&pl->pl_limit, 1); |
00f9d12b OD |
674 | pl->pl_server_lock_volume = 0; |
675 | pl->pl_ops = &ldlm_cli_pool_ops; | |
676 | pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD; | |
d7e09d03 | 677 | pl->pl_client_lock_volume = 0; |
700815d4 | 678 | rc = ldlm_pool_debugfs_init(pl); |
d7e09d03 | 679 | if (rc) |
0a3bdb00 | 680 | return rc; |
d7e09d03 | 681 | |
f2825e03 OD |
682 | rc = ldlm_pool_sysfs_init(pl); |
683 | if (rc) | |
684 | return rc; | |
685 | ||
d7e09d03 PT |
686 | CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name); |
687 | ||
0a3bdb00 | 688 | return rc; |
d7e09d03 PT |
689 | } |
690 | EXPORT_SYMBOL(ldlm_pool_init); | |
691 | ||
692 | void ldlm_pool_fini(struct ldlm_pool *pl) | |
693 | { | |
f2825e03 | 694 | ldlm_pool_sysfs_fini(pl); |
700815d4 | 695 | ldlm_pool_debugfs_fini(pl); |
d7e09d03 PT |
696 | |
697 | /* | |
698 | * Pool should not be used after this point. We can't free it here as | |
699 | * it lives in struct ldlm_namespace, but still interested in catching | |
700 | * any abnormal using cases. | |
701 | */ | |
702 | POISON(pl, 0x5a, sizeof(*pl)); | |
d7e09d03 PT |
703 | } |
704 | EXPORT_SYMBOL(ldlm_pool_fini); | |
705 | ||
706 | /** | |
707 | * Add new taken ldlm lock \a lock into pool \a pl accounting. | |
708 | */ | |
709 | void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock) | |
710 | { | |
711 | /* | |
712 | * FLOCK locks are special in a sense that they are almost never | |
713 | * cancelled, instead special kind of lock is used to drop them. | |
714 | * also there is no LRU for flock locks, so no point in tracking | |
715 | * them anyway. | |
716 | */ | |
717 | if (lock->l_resource->lr_type == LDLM_FLOCK) | |
718 | return; | |
719 | ||
720 | atomic_inc(&pl->pl_granted); | |
721 | atomic_inc(&pl->pl_grant_rate); | |
722 | lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT); | |
723 | /* | |
724 | * Do not do pool recalc for client side as all locks which | |
725 | * potentially may be canceled has already been packed into | |
726 | * enqueue/cancel rpc. Also we do not want to run out of stack | |
727 | * with too long call paths. | |
728 | */ | |
d7e09d03 PT |
729 | } |
730 | EXPORT_SYMBOL(ldlm_pool_add); | |
731 | ||
732 | /** | |
733 | * Remove ldlm lock \a lock from pool \a pl accounting. | |
734 | */ | |
735 | void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock) | |
736 | { | |
737 | /* | |
738 | * Filter out FLOCK locks. Read above comment in ldlm_pool_add(). | |
739 | */ | |
740 | if (lock->l_resource->lr_type == LDLM_FLOCK) | |
741 | return; | |
742 | ||
743 | LASSERT(atomic_read(&pl->pl_granted) > 0); | |
744 | atomic_dec(&pl->pl_granted); | |
745 | atomic_inc(&pl->pl_cancel_rate); | |
746 | ||
747 | lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT); | |
d7e09d03 PT |
748 | } |
749 | EXPORT_SYMBOL(ldlm_pool_del); | |
750 | ||
751 | /** | |
752 | * Returns current \a pl SLV. | |
753 | * | |
754 | * \pre ->pl_lock is not locked. | |
755 | */ | |
756 | __u64 ldlm_pool_get_slv(struct ldlm_pool *pl) | |
757 | { | |
758 | __u64 slv; | |
902f3bb1 | 759 | |
d7e09d03 PT |
760 | spin_lock(&pl->pl_lock); |
761 | slv = pl->pl_server_lock_volume; | |
762 | spin_unlock(&pl->pl_lock); | |
763 | return slv; | |
764 | } | |
d7e09d03 | 765 | |
d7e09d03 PT |
766 | /** |
767 | * Sets passed \a clv to \a pl. | |
768 | * | |
769 | * \pre ->pl_lock is not locked. | |
770 | */ | |
771 | void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv) | |
772 | { | |
773 | spin_lock(&pl->pl_lock); | |
774 | pl->pl_client_lock_volume = clv; | |
775 | spin_unlock(&pl->pl_lock); | |
776 | } | |
d7e09d03 PT |
777 | |
778 | /** | |
779 | * Returns current LVF from \a pl. | |
780 | */ | |
781 | __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl) | |
782 | { | |
783 | return atomic_read(&pl->pl_lock_volume_factor); | |
784 | } | |
d7e09d03 PT |
785 | |
786 | static int ldlm_pool_granted(struct ldlm_pool *pl) | |
787 | { | |
788 | return atomic_read(&pl->pl_granted); | |
789 | } | |
790 | ||
791 | static struct ptlrpc_thread *ldlm_pools_thread; | |
d7e09d03 PT |
792 | static struct completion ldlm_pools_comp; |
793 | ||
794 | /* | |
cbc3769e PT |
795 | * count locks from all namespaces (if possible). Returns number of |
796 | * cached locks. | |
d7e09d03 | 797 | */ |
5802572e | 798 | static unsigned long ldlm_pools_count(ldlm_side_t client, gfp_t gfp_mask) |
d7e09d03 | 799 | { |
cbc3769e | 800 | int total = 0, nr_ns; |
d7e09d03 | 801 | struct ldlm_namespace *ns; |
91a50030 | 802 | struct ldlm_namespace *ns_old = NULL; /* loop detection */ |
d7e09d03 PT |
803 | void *cookie; |
804 | ||
cbc3769e PT |
805 | if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS)) |
806 | return 0; | |
d7e09d03 | 807 | |
cbc3769e PT |
808 | CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n", |
809 | client == LDLM_NAMESPACE_CLIENT ? "client" : "server"); | |
d7e09d03 PT |
810 | |
811 | cookie = cl_env_reenter(); | |
812 | ||
813 | /* | |
814 | * Find out how many resources we may release. | |
815 | */ | |
91a50030 | 816 | for (nr_ns = ldlm_namespace_nr_read(client); |
cbc3769e | 817 | nr_ns > 0; nr_ns--) { |
d7e09d03 PT |
818 | mutex_lock(ldlm_namespace_lock(client)); |
819 | if (list_empty(ldlm_namespace_list(client))) { | |
820 | mutex_unlock(ldlm_namespace_lock(client)); | |
821 | cl_env_reexit(cookie); | |
822 | return 0; | |
823 | } | |
824 | ns = ldlm_namespace_first_locked(client); | |
91a50030 OD |
825 | |
826 | if (ns == ns_old) { | |
827 | mutex_unlock(ldlm_namespace_lock(client)); | |
828 | break; | |
829 | } | |
830 | ||
831 | if (ldlm_ns_empty(ns)) { | |
832 | ldlm_namespace_move_to_inactive_locked(ns, client); | |
833 | mutex_unlock(ldlm_namespace_lock(client)); | |
834 | continue; | |
835 | } | |
836 | ||
837 | if (ns_old == NULL) | |
838 | ns_old = ns; | |
839 | ||
d7e09d03 | 840 | ldlm_namespace_get(ns); |
91a50030 | 841 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
842 | mutex_unlock(ldlm_namespace_lock(client)); |
843 | total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask); | |
844 | ldlm_namespace_put(ns); | |
845 | } | |
846 | ||
cbc3769e PT |
847 | cl_env_reexit(cookie); |
848 | return total; | |
849 | } | |
850 | ||
5802572e | 851 | static unsigned long ldlm_pools_scan(ldlm_side_t client, int nr, gfp_t gfp_mask) |
cbc3769e PT |
852 | { |
853 | unsigned long freed = 0; | |
854 | int tmp, nr_ns; | |
855 | struct ldlm_namespace *ns; | |
856 | void *cookie; | |
857 | ||
858 | if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS)) | |
859 | return -1; | |
860 | ||
861 | cookie = cl_env_reenter(); | |
d7e09d03 PT |
862 | |
863 | /* | |
cbc3769e | 864 | * Shrink at least ldlm_namespace_nr_read(client) namespaces. |
d7e09d03 | 865 | */ |
cbc3769e PT |
866 | for (tmp = nr_ns = ldlm_namespace_nr_read(client); |
867 | tmp > 0; tmp--) { | |
d7e09d03 PT |
868 | int cancel, nr_locks; |
869 | ||
870 | /* | |
871 | * Do not call shrink under ldlm_namespace_lock(client) | |
872 | */ | |
873 | mutex_lock(ldlm_namespace_lock(client)); | |
874 | if (list_empty(ldlm_namespace_list(client))) { | |
875 | mutex_unlock(ldlm_namespace_lock(client)); | |
d7e09d03 PT |
876 | break; |
877 | } | |
878 | ns = ldlm_namespace_first_locked(client); | |
879 | ldlm_namespace_get(ns); | |
91a50030 | 880 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
881 | mutex_unlock(ldlm_namespace_lock(client)); |
882 | ||
883 | nr_locks = ldlm_pool_granted(&ns->ns_pool); | |
cbc3769e PT |
884 | /* |
885 | * We use to shrink propotionally but with new shrinker API, | |
886 | * we lost the total number of freeable locks. | |
887 | */ | |
888 | cancel = 1 + min_t(int, nr_locks, nr / nr_ns); | |
889 | freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask); | |
d7e09d03 PT |
890 | ldlm_namespace_put(ns); |
891 | } | |
892 | cl_env_reexit(cookie); | |
cbc3769e PT |
893 | /* |
894 | * we only decrease the SLV in server pools shrinker, return | |
895 | * SHRINK_STOP to kernel to avoid needless loop. LU-1128 | |
896 | */ | |
00f9d12b | 897 | return freed; |
d7e09d03 PT |
898 | } |
899 | ||
e7ddc48c AR |
900 | static unsigned long ldlm_pools_cli_count(struct shrinker *s, |
901 | struct shrink_control *sc) | |
d7e09d03 | 902 | { |
cbc3769e PT |
903 | return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask); |
904 | } | |
905 | ||
e7ddc48c AR |
906 | static unsigned long ldlm_pools_cli_scan(struct shrinker *s, |
907 | struct shrink_control *sc) | |
cbc3769e PT |
908 | { |
909 | return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan, | |
910 | sc->gfp_mask); | |
d7e09d03 PT |
911 | } |
912 | ||
00f9d12b | 913 | static int ldlm_pools_recalc(ldlm_side_t client) |
d7e09d03 | 914 | { |
d7e09d03 | 915 | struct ldlm_namespace *ns; |
91a50030 | 916 | struct ldlm_namespace *ns_old = NULL; |
00f9d12b | 917 | int nr; |
3eface59 | 918 | int time = 50; /* seconds of sleep if no active namespaces */ |
d7e09d03 | 919 | |
d7e09d03 | 920 | /* |
cbc3769e | 921 | * Recalc at least ldlm_namespace_nr_read(client) namespaces. |
d7e09d03 | 922 | */ |
91a50030 | 923 | for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) { |
d7e09d03 PT |
924 | int skip; |
925 | /* | |
926 | * Lock the list, get first @ns in the list, getref, move it | |
927 | * to the tail, unlock and call pool recalc. This way we avoid | |
928 | * calling recalc under @ns lock what is really good as we get | |
929 | * rid of potential deadlock on client nodes when canceling | |
930 | * locks synchronously. | |
931 | */ | |
932 | mutex_lock(ldlm_namespace_lock(client)); | |
933 | if (list_empty(ldlm_namespace_list(client))) { | |
934 | mutex_unlock(ldlm_namespace_lock(client)); | |
935 | break; | |
936 | } | |
937 | ns = ldlm_namespace_first_locked(client); | |
938 | ||
91a50030 OD |
939 | if (ns_old == ns) { /* Full pass complete */ |
940 | mutex_unlock(ldlm_namespace_lock(client)); | |
941 | break; | |
942 | } | |
943 | ||
944 | /* We got an empty namespace, need to move it back to inactive | |
945 | * list. | |
946 | * The race with parallel resource creation is fine: | |
947 | * - If they do namespace_get before our check, we fail the | |
948 | * check and they move this item to the end of the list anyway | |
949 | * - If we do the check and then they do namespace_get, then | |
950 | * we move the namespace to inactive and they will move | |
951 | * it back to active (synchronised by the lock, so no clash | |
952 | * there). | |
953 | */ | |
954 | if (ldlm_ns_empty(ns)) { | |
955 | ldlm_namespace_move_to_inactive_locked(ns, client); | |
956 | mutex_unlock(ldlm_namespace_lock(client)); | |
957 | continue; | |
958 | } | |
959 | ||
960 | if (ns_old == NULL) | |
961 | ns_old = ns; | |
962 | ||
d7e09d03 PT |
963 | spin_lock(&ns->ns_lock); |
964 | /* | |
965 | * skip ns which is being freed, and we don't want to increase | |
966 | * its refcount again, not even temporarily. bz21519 & LU-499. | |
967 | */ | |
968 | if (ns->ns_stopping) { | |
969 | skip = 1; | |
970 | } else { | |
971 | skip = 0; | |
972 | ldlm_namespace_get(ns); | |
973 | } | |
974 | spin_unlock(&ns->ns_lock); | |
975 | ||
91a50030 | 976 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
977 | mutex_unlock(ldlm_namespace_lock(client)); |
978 | ||
979 | /* | |
980 | * After setup is done - recalc the pool. | |
981 | */ | |
982 | if (!skip) { | |
3eface59 OD |
983 | int ttime = ldlm_pool_recalc(&ns->ns_pool); |
984 | ||
985 | if (ttime < time) | |
986 | time = ttime; | |
987 | ||
d7e09d03 PT |
988 | ldlm_namespace_put(ns); |
989 | } | |
990 | } | |
3eface59 | 991 | return time; |
d7e09d03 | 992 | } |
d7e09d03 PT |
993 | |
994 | static int ldlm_pools_thread_main(void *arg) | |
995 | { | |
996 | struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg; | |
00f9d12b | 997 | int c_time; |
d7e09d03 PT |
998 | |
999 | thread_set_flags(thread, SVC_RUNNING); | |
1000 | wake_up(&thread->t_ctl_waitq); | |
1001 | ||
1002 | CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n", | |
1003 | "ldlm_poold", current_pid()); | |
1004 | ||
1005 | while (1) { | |
1006 | struct l_wait_info lwi; | |
1007 | ||
1008 | /* | |
1009 | * Recal all pools on this tick. | |
1010 | */ | |
3eface59 | 1011 | c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT); |
d7e09d03 PT |
1012 | |
1013 | /* | |
1014 | * Wait until the next check time, or until we're | |
1015 | * stopped. | |
1016 | */ | |
00f9d12b | 1017 | lwi = LWI_TIMEOUT(cfs_time_seconds(c_time), |
d7e09d03 PT |
1018 | NULL, NULL); |
1019 | l_wait_event(thread->t_ctl_waitq, | |
1020 | thread_is_stopping(thread) || | |
1021 | thread_is_event(thread), | |
1022 | &lwi); | |
1023 | ||
1024 | if (thread_test_and_clear_flags(thread, SVC_STOPPING)) | |
1025 | break; | |
71e8dd9a | 1026 | thread_test_and_clear_flags(thread, SVC_EVENT); |
d7e09d03 PT |
1027 | } |
1028 | ||
1029 | thread_set_flags(thread, SVC_STOPPED); | |
1030 | wake_up(&thread->t_ctl_waitq); | |
1031 | ||
1032 | CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n", | |
1033 | "ldlm_poold", current_pid()); | |
1034 | ||
1035 | complete_and_exit(&ldlm_pools_comp, 0); | |
1036 | } | |
1037 | ||
1038 | static int ldlm_pools_thread_start(void) | |
1039 | { | |
1040 | struct l_wait_info lwi = { 0 }; | |
68b636b6 | 1041 | struct task_struct *task; |
d7e09d03 PT |
1042 | |
1043 | if (ldlm_pools_thread != NULL) | |
0a3bdb00 | 1044 | return -EALREADY; |
d7e09d03 | 1045 | |
352f7891 | 1046 | ldlm_pools_thread = kzalloc(sizeof(*ldlm_pools_thread), GFP_NOFS); |
94e67761 | 1047 | if (!ldlm_pools_thread) |
0a3bdb00 | 1048 | return -ENOMEM; |
d7e09d03 PT |
1049 | |
1050 | init_completion(&ldlm_pools_comp); | |
1051 | init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq); | |
1052 | ||
1053 | task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread, | |
1054 | "ldlm_poold"); | |
1055 | if (IS_ERR(task)) { | |
1056 | CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task)); | |
352f7891 | 1057 | kfree(ldlm_pools_thread); |
d7e09d03 | 1058 | ldlm_pools_thread = NULL; |
0a3bdb00 | 1059 | return PTR_ERR(task); |
d7e09d03 PT |
1060 | } |
1061 | l_wait_event(ldlm_pools_thread->t_ctl_waitq, | |
1062 | thread_is_running(ldlm_pools_thread), &lwi); | |
0a3bdb00 | 1063 | return 0; |
d7e09d03 PT |
1064 | } |
1065 | ||
1066 | static void ldlm_pools_thread_stop(void) | |
1067 | { | |
8d2ff65d | 1068 | if (ldlm_pools_thread == NULL) |
d7e09d03 | 1069 | return; |
d7e09d03 PT |
1070 | |
1071 | thread_set_flags(ldlm_pools_thread, SVC_STOPPING); | |
1072 | wake_up(&ldlm_pools_thread->t_ctl_waitq); | |
1073 | ||
1074 | /* | |
1075 | * Make sure that pools thread is finished before freeing @thread. | |
1076 | * This fixes possible race and oops due to accessing freed memory | |
1077 | * in pools thread. | |
1078 | */ | |
1079 | wait_for_completion(&ldlm_pools_comp); | |
352f7891 | 1080 | kfree(ldlm_pools_thread); |
d7e09d03 | 1081 | ldlm_pools_thread = NULL; |
d7e09d03 PT |
1082 | } |
1083 | ||
cbc3769e PT |
1084 | static struct shrinker ldlm_pools_cli_shrinker = { |
1085 | .count_objects = ldlm_pools_cli_count, | |
1086 | .scan_objects = ldlm_pools_cli_scan, | |
1087 | .seeks = DEFAULT_SEEKS, | |
1088 | }; | |
1089 | ||
d7e09d03 PT |
1090 | int ldlm_pools_init(void) |
1091 | { | |
1092 | int rc; | |
d7e09d03 PT |
1093 | |
1094 | rc = ldlm_pools_thread_start(); | |
00f9d12b | 1095 | if (rc == 0) |
cbc3769e | 1096 | register_shrinker(&ldlm_pools_cli_shrinker); |
00f9d12b | 1097 | |
0a3bdb00 | 1098 | return rc; |
d7e09d03 PT |
1099 | } |
1100 | EXPORT_SYMBOL(ldlm_pools_init); | |
1101 | ||
1102 | void ldlm_pools_fini(void) | |
1103 | { | |
00f9d12b | 1104 | if (ldlm_pools_thread) |
faa7a4e3 | 1105 | unregister_shrinker(&ldlm_pools_cli_shrinker); |
00f9d12b | 1106 | |
d7e09d03 PT |
1107 | ldlm_pools_thread_stop(); |
1108 | } | |
1109 | EXPORT_SYMBOL(ldlm_pools_fini); |