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 | |
6a5b99a4 | 18 | * http://www.gnu.org/licenses/gpl-2.0.html |
d7e09d03 | 19 | * |
d7e09d03 PT |
20 | * GPL HEADER END |
21 | */ | |
22 | /* | |
23 | * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. | |
24 | * Use is subject to license terms. | |
25 | * | |
1dc563a6 | 26 | * Copyright (c) 2010, 2015, Intel Corporation. |
d7e09d03 PT |
27 | */ |
28 | /* | |
29 | * This file is part of Lustre, http://www.lustre.org/ | |
30 | * Lustre is a trademark of Sun Microsystems, Inc. | |
31 | * | |
32 | * lustre/ldlm/ldlm_pool.c | |
33 | * | |
34 | * Author: Yury Umanets <umka@clusterfs.com> | |
35 | */ | |
36 | ||
37 | /* | |
38 | * Idea of this code is rather simple. Each second, for each server namespace | |
39 | * we have SLV - server lock volume which is calculated on current number of | |
40 | * granted locks, grant speed for past period, etc - that is, locking load. | |
41 | * This SLV number may be thought as a flow definition for simplicity. It is | |
42 | * sent to clients with each occasion to let them know what is current load | |
43 | * situation on the server. By default, at the beginning, SLV on server is | |
44 | * set max value which is calculated as the following: allow to one client | |
45 | * have all locks of limit ->pl_limit for 10h. | |
46 | * | |
47 | * Next, on clients, number of cached locks is not limited artificially in any | |
48 | * way as it was before. Instead, client calculates CLV, that is, client lock | |
49 | * volume for each lock and compares it with last SLV from the server. CLV is | |
50 | * calculated as the number of locks in LRU * lock live time in seconds. If | |
51 | * CLV > SLV - lock is canceled. | |
52 | * | |
e7ddc48c AR |
53 | * Client has LVF, that is, lock volume factor which regulates how much |
54 | * sensitive client should be about last SLV from server. The higher LVF is the | |
55 | * more locks will be canceled on client. Default value for it is 1. Setting LVF | |
56 | * to 2 means that client will cancel locks 2 times faster. | |
d7e09d03 PT |
57 | * |
58 | * Locks on a client will be canceled more intensively in these cases: | |
59 | * (1) if SLV is smaller, that is, load is higher on the server; | |
60 | * (2) client has a lot of locks (the more locks are held by client, the bigger | |
61 | * chances that some of them should be canceled); | |
62 | * (3) client has old locks (taken some time ago); | |
63 | * | |
64 | * Thus, according to flow paradigm that we use for better understanding SLV, | |
65 | * CLV is the volume of particle in flow described by SLV. According to this, | |
66 | * if flow is getting thinner, more and more particles become outside of it and | |
67 | * as particles are locks, they should be canceled. | |
68 | * | |
e7ddc48c AR |
69 | * General idea of this belongs to Vitaly Fertman (vitaly@clusterfs.com). |
70 | * Andreas Dilger (adilger@clusterfs.com) proposed few nice ideas like using | |
71 | * LVF and many cleanups. Flow definition to allow more easy understanding of | |
72 | * the logic belongs to Nikita Danilov (nikita@clusterfs.com) as well as many | |
73 | * cleanups and fixes. And design and implementation are done by Yury Umanets | |
74 | * (umka@clusterfs.com). | |
d7e09d03 PT |
75 | * |
76 | * Glossary for terms used: | |
77 | * | |
78 | * pl_limit - Number of allowed locks in pool. Applies to server and client | |
79 | * side (tunable); | |
80 | * | |
81 | * pl_granted - Number of granted locks (calculated); | |
82 | * pl_grant_rate - Number of granted locks for last T (calculated); | |
83 | * pl_cancel_rate - Number of canceled locks for last T (calculated); | |
84 | * pl_grant_speed - Grant speed (GR - CR) for last T (calculated); | |
85 | * pl_grant_plan - Planned number of granted locks for next T (calculated); | |
86 | * pl_server_lock_volume - Current server lock volume (calculated); | |
87 | * | |
88 | * As it may be seen from list above, we have few possible tunables which may | |
f2825e03 | 89 | * affect behavior much. They all may be modified via sysfs. However, they also |
d7e09d03 PT |
90 | * give a possibility for constructing few pre-defined behavior policies. If |
91 | * none of predefines is suitable for a working pattern being used, new one may | |
f2825e03 | 92 | * be "constructed" via sysfs tunables. |
d7e09d03 PT |
93 | */ |
94 | ||
95 | #define DEBUG_SUBSYSTEM S_LDLM | |
96 | ||
e27db149 GKH |
97 | #include "../include/lustre_dlm.h" |
98 | #include "../include/cl_object.h" | |
99 | #include "../include/obd_class.h" | |
100 | #include "../include/obd_support.h" | |
d7e09d03 PT |
101 | #include "ldlm_internal.h" |
102 | ||
d7e09d03 PT |
103 | /* |
104 | * 50 ldlm locks for 1MB of RAM. | |
105 | */ | |
09cbfeaf | 106 | #define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_SHIFT)) * 50) |
d7e09d03 PT |
107 | |
108 | /* | |
109 | * Maximal possible grant step plan in %. | |
110 | */ | |
111 | #define LDLM_POOL_MAX_GSP (30) | |
112 | ||
113 | /* | |
114 | * Minimal possible grant step plan in %. | |
115 | */ | |
116 | #define LDLM_POOL_MIN_GSP (1) | |
117 | ||
118 | /* | |
119 | * This controls the speed of reaching LDLM_POOL_MAX_GSP | |
120 | * with increasing thread period. | |
121 | */ | |
122 | #define LDLM_POOL_GSP_STEP_SHIFT (2) | |
123 | ||
124 | /* | |
125 | * LDLM_POOL_GSP% of all locks is default GP. | |
126 | */ | |
127 | #define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100) | |
128 | ||
129 | /* | |
130 | * Max age for locks on clients. | |
131 | */ | |
132 | #define LDLM_POOL_MAX_AGE (36000) | |
133 | ||
134 | /* | |
135 | * The granularity of SLV calculation. | |
136 | */ | |
137 | #define LDLM_POOL_SLV_SHIFT (10) | |
138 | ||
d7e09d03 PT |
139 | static inline __u64 dru(__u64 val, __u32 shift, int round_up) |
140 | { | |
141 | return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift; | |
142 | } | |
143 | ||
144 | static inline __u64 ldlm_pool_slv_max(__u32 L) | |
145 | { | |
146 | /* | |
147 | * Allow to have all locks for 1 client for 10 hrs. | |
148 | * Formula is the following: limit * 10h / 1 client. | |
149 | */ | |
150 | __u64 lim = (__u64)L * LDLM_POOL_MAX_AGE / 1; | |
151 | return lim; | |
152 | } | |
153 | ||
154 | static inline __u64 ldlm_pool_slv_min(__u32 L) | |
155 | { | |
156 | return 1; | |
157 | } | |
158 | ||
159 | enum { | |
160 | LDLM_POOL_FIRST_STAT = 0, | |
161 | LDLM_POOL_GRANTED_STAT = LDLM_POOL_FIRST_STAT, | |
162 | LDLM_POOL_GRANT_STAT, | |
163 | LDLM_POOL_CANCEL_STAT, | |
164 | LDLM_POOL_GRANT_RATE_STAT, | |
165 | LDLM_POOL_CANCEL_RATE_STAT, | |
166 | LDLM_POOL_GRANT_PLAN_STAT, | |
167 | LDLM_POOL_SLV_STAT, | |
168 | LDLM_POOL_SHRINK_REQTD_STAT, | |
169 | LDLM_POOL_SHRINK_FREED_STAT, | |
170 | LDLM_POOL_RECALC_STAT, | |
171 | LDLM_POOL_TIMING_STAT, | |
172 | LDLM_POOL_LAST_STAT | |
173 | }; | |
174 | ||
d7e09d03 PT |
175 | /** |
176 | * Calculates suggested grant_step in % of available locks for passed | |
177 | * \a period. This is later used in grant_plan calculations. | |
178 | */ | |
179 | static inline int ldlm_pool_t2gsp(unsigned int t) | |
180 | { | |
181 | /* | |
182 | * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP | |
183 | * and up to 30% for anything higher than LDLM_POOL_GSP_STEP. | |
184 | * | |
185 | * How this will affect execution is the following: | |
186 | * | |
187 | * - for thread period 1s we will have grant_step 1% which good from | |
188 | * pov of taking some load off from server and push it out to clients. | |
189 | * This is like that because 1% for grant_step means that server will | |
190 | * not allow clients to get lots of locks in short period of time and | |
191 | * keep all old locks in their caches. Clients will always have to | |
192 | * get some locks back if they want to take some new; | |
193 | * | |
194 | * - for thread period 10s (which is default) we will have 23% which | |
195 | * means that clients will have enough of room to take some new locks | |
196 | * without getting some back. All locks from this 23% which were not | |
197 | * taken by clients in current period will contribute in SLV growing. | |
198 | * SLV growing means more locks cached on clients until limit or grant | |
199 | * plan is reached. | |
200 | */ | |
201 | return LDLM_POOL_MAX_GSP - | |
202 | ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >> | |
203 | (t >> LDLM_POOL_GSP_STEP_SHIFT)); | |
204 | } | |
205 | ||
d7e09d03 PT |
206 | /** |
207 | * Recalculates next stats on passed \a pl. | |
208 | * | |
209 | * \pre ->pl_lock is locked. | |
210 | */ | |
211 | static void ldlm_pool_recalc_stats(struct ldlm_pool *pl) | |
212 | { | |
213 | int grant_plan = pl->pl_grant_plan; | |
214 | __u64 slv = pl->pl_server_lock_volume; | |
215 | int granted = atomic_read(&pl->pl_granted); | |
216 | int grant_rate = atomic_read(&pl->pl_grant_rate); | |
217 | int cancel_rate = atomic_read(&pl->pl_cancel_rate); | |
218 | ||
219 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT, | |
220 | slv); | |
221 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT, | |
222 | granted); | |
223 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT, | |
224 | grant_rate); | |
225 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT, | |
226 | grant_plan); | |
227 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT, | |
228 | cancel_rate); | |
229 | } | |
230 | ||
d7e09d03 | 231 | /** |
7c37abe0 SB |
232 | * Sets SLV and Limit from container_of(pl, struct ldlm_namespace, |
233 | * ns_pool)->ns_obd tp passed \a pl. | |
d7e09d03 PT |
234 | */ |
235 | static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl) | |
236 | { | |
237 | struct obd_device *obd; | |
238 | ||
239 | /* | |
240 | * Get new SLV and Limit from obd which is updated with coming | |
241 | * RPCs. | |
242 | */ | |
7c37abe0 SB |
243 | obd = container_of(pl, struct ldlm_namespace, |
244 | ns_pool)->ns_obd; | |
d7e09d03 PT |
245 | read_lock(&obd->obd_pool_lock); |
246 | pl->pl_server_lock_volume = obd->obd_pool_slv; | |
f7ec22b5 | 247 | atomic_set(&pl->pl_limit, obd->obd_pool_limit); |
d7e09d03 PT |
248 | read_unlock(&obd->obd_pool_lock); |
249 | } | |
250 | ||
251 | /** | |
252 | * Recalculates client size pool \a pl according to current SLV and Limit. | |
253 | */ | |
254 | static int ldlm_cli_pool_recalc(struct ldlm_pool *pl) | |
255 | { | |
8f83409c | 256 | time64_t recalc_interval_sec; |
4d2c7b30 | 257 | int ret; |
d7e09d03 | 258 | |
8f83409c | 259 | recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time; |
d7e09d03 | 260 | if (recalc_interval_sec < pl->pl_recalc_period) |
0a3bdb00 | 261 | return 0; |
d7e09d03 PT |
262 | |
263 | spin_lock(&pl->pl_lock); | |
264 | /* | |
265 | * Check if we need to recalc lists now. | |
266 | */ | |
8f83409c | 267 | recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time; |
d7e09d03 PT |
268 | if (recalc_interval_sec < pl->pl_recalc_period) { |
269 | spin_unlock(&pl->pl_lock); | |
0a3bdb00 | 270 | return 0; |
d7e09d03 PT |
271 | } |
272 | ||
273 | /* | |
274 | * Make sure that pool knows last SLV and Limit from obd. | |
275 | */ | |
276 | ldlm_cli_pool_pop_slv(pl); | |
277 | ||
d7e09d03 PT |
278 | spin_unlock(&pl->pl_lock); |
279 | ||
280 | /* | |
281 | * Do not cancel locks in case lru resize is disabled for this ns. | |
282 | */ | |
7c37abe0 SB |
283 | if (!ns_connect_lru_resize(container_of(pl, struct ldlm_namespace, |
284 | ns_pool))) { | |
4d2c7b30 LX |
285 | ret = 0; |
286 | goto out; | |
287 | } | |
d7e09d03 PT |
288 | |
289 | /* | |
290 | * In the time of canceling locks on client we do not need to maintain | |
291 | * sharp timing, we only want to cancel locks asap according to new SLV. | |
292 | * It may be called when SLV has changed much, this is why we do not | |
293 | * take into account pl->pl_recalc_time here. | |
294 | */ | |
7c37abe0 SB |
295 | ret = ldlm_cancel_lru(container_of(pl, struct ldlm_namespace, ns_pool), |
296 | 0, LCF_ASYNC, LDLM_CANCEL_LRUR); | |
4d2c7b30 LX |
297 | |
298 | out: | |
299 | spin_lock(&pl->pl_lock); | |
300 | /* | |
301 | * Time of LRU resizing might be longer than period, | |
302 | * so update after LRU resizing rather than before it. | |
303 | */ | |
8f83409c | 304 | pl->pl_recalc_time = ktime_get_real_seconds(); |
4d2c7b30 LX |
305 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT, |
306 | recalc_interval_sec); | |
307 | spin_unlock(&pl->pl_lock); | |
308 | return ret; | |
d7e09d03 PT |
309 | } |
310 | ||
311 | /** | |
312 | * This function is main entry point for memory pressure handling on client | |
313 | * side. Main goal of this function is to cancel some number of locks on | |
314 | * passed \a pl according to \a nr and \a gfp_mask. | |
315 | */ | |
316 | static int ldlm_cli_pool_shrink(struct ldlm_pool *pl, | |
5802572e | 317 | int nr, gfp_t gfp_mask) |
d7e09d03 PT |
318 | { |
319 | struct ldlm_namespace *ns; | |
cbc3769e | 320 | int unused; |
d7e09d03 | 321 | |
7c37abe0 | 322 | ns = container_of(pl, struct ldlm_namespace, ns_pool); |
d7e09d03 PT |
323 | |
324 | /* | |
325 | * Do not cancel locks in case lru resize is disabled for this ns. | |
326 | */ | |
327 | if (!ns_connect_lru_resize(ns)) | |
0a3bdb00 | 328 | return 0; |
d7e09d03 PT |
329 | |
330 | /* | |
331 | * Make sure that pool knows last SLV and Limit from obd. | |
332 | */ | |
333 | ldlm_cli_pool_pop_slv(pl); | |
334 | ||
335 | spin_lock(&ns->ns_lock); | |
336 | unused = ns->ns_nr_unused; | |
337 | spin_unlock(&ns->ns_lock); | |
338 | ||
cbc3769e PT |
339 | if (nr == 0) |
340 | return (unused / 100) * sysctl_vfs_cache_pressure; | |
341 | else | |
342 | return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_CANCEL_SHRINK); | |
d7e09d03 PT |
343 | } |
344 | ||
b9c98cfa | 345 | static const struct ldlm_pool_ops ldlm_cli_pool_ops = { |
d7e09d03 PT |
346 | .po_recalc = ldlm_cli_pool_recalc, |
347 | .po_shrink = ldlm_cli_pool_shrink | |
348 | }; | |
349 | ||
350 | /** | |
351 | * Pool recalc wrapper. Will call either client or server pool recalc callback | |
352 | * depending what pool \a pl is used. | |
353 | */ | |
58c6d133 | 354 | static int ldlm_pool_recalc(struct ldlm_pool *pl) |
d7e09d03 | 355 | { |
8f83409c | 356 | u32 recalc_interval_sec; |
d7e09d03 PT |
357 | int count; |
358 | ||
8f83409c | 359 | recalc_interval_sec = ktime_get_seconds() - pl->pl_recalc_time; |
d7e09d03 PT |
360 | if (recalc_interval_sec <= 0) |
361 | goto recalc; | |
362 | ||
363 | spin_lock(&pl->pl_lock); | |
d7e09d03 PT |
364 | if (recalc_interval_sec > 0) { |
365 | /* | |
366 | * Update pool statistics every 1s. | |
367 | */ | |
368 | ldlm_pool_recalc_stats(pl); | |
369 | ||
370 | /* | |
371 | * Zero out all rates and speed for the last period. | |
372 | */ | |
373 | atomic_set(&pl->pl_grant_rate, 0); | |
374 | atomic_set(&pl->pl_cancel_rate, 0); | |
375 | } | |
376 | spin_unlock(&pl->pl_lock); | |
377 | ||
378 | recalc: | |
44b53f18 | 379 | if (pl->pl_ops->po_recalc) { |
d7e09d03 PT |
380 | count = pl->pl_ops->po_recalc(pl); |
381 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT, | |
382 | count); | |
d7e09d03 | 383 | } |
8f83409c | 384 | recalc_interval_sec = pl->pl_recalc_time - ktime_get_seconds() + |
3eface59 | 385 | pl->pl_recalc_period; |
4d2c7b30 LX |
386 | if (recalc_interval_sec <= 0) { |
387 | /* Prevent too frequent recalculation. */ | |
8f83409c AB |
388 | CDEBUG(D_DLMTRACE, |
389 | "Negative interval(%d), too short period(%lld)", | |
4d2c7b30 | 390 | recalc_interval_sec, |
8f83409c | 391 | (s64)pl->pl_recalc_period); |
4d2c7b30 LX |
392 | recalc_interval_sec = 1; |
393 | } | |
d7e09d03 | 394 | |
3eface59 | 395 | return recalc_interval_sec; |
d7e09d03 | 396 | } |
d7e09d03 | 397 | |
cbc3769e | 398 | /* |
d7e09d03 | 399 | * Pool shrink wrapper. Will call either client or server pool recalc callback |
cbc3769e PT |
400 | * depending what pool pl is used. When nr == 0, just return the number of |
401 | * freeable locks. Otherwise, return the number of canceled locks. | |
d7e09d03 | 402 | */ |
58c6d133 | 403 | static int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, gfp_t gfp_mask) |
d7e09d03 PT |
404 | { |
405 | int cancel = 0; | |
406 | ||
44b53f18 | 407 | if (pl->pl_ops->po_shrink) { |
d7e09d03 PT |
408 | cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask); |
409 | if (nr > 0) { | |
410 | lprocfs_counter_add(pl->pl_stats, | |
411 | LDLM_POOL_SHRINK_REQTD_STAT, | |
412 | nr); | |
413 | lprocfs_counter_add(pl->pl_stats, | |
414 | LDLM_POOL_SHRINK_FREED_STAT, | |
415 | cancel); | |
2d00bd17 JP |
416 | CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, shrunk %d\n", |
417 | pl->pl_name, nr, cancel); | |
d7e09d03 PT |
418 | } |
419 | } | |
420 | return cancel; | |
421 | } | |
d7e09d03 | 422 | |
73bb1da6 | 423 | static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused) |
d7e09d03 | 424 | { |
71570b98 OD |
425 | int granted, grant_rate, cancel_rate; |
426 | int grant_speed, lvf; | |
73bb1da6 | 427 | struct ldlm_pool *pl = m->private; |
d7e09d03 PT |
428 | __u64 slv, clv; |
429 | __u32 limit; | |
430 | ||
431 | spin_lock(&pl->pl_lock); | |
432 | slv = pl->pl_server_lock_volume; | |
433 | clv = pl->pl_client_lock_volume; | |
946d6f95 | 434 | limit = atomic_read(&pl->pl_limit); |
d7e09d03 PT |
435 | granted = atomic_read(&pl->pl_granted); |
436 | grant_rate = atomic_read(&pl->pl_grant_rate); | |
437 | cancel_rate = atomic_read(&pl->pl_cancel_rate); | |
438 | grant_speed = grant_rate - cancel_rate; | |
439 | lvf = atomic_read(&pl->pl_lock_volume_factor); | |
d7e09d03 PT |
440 | spin_unlock(&pl->pl_lock); |
441 | ||
73bb1da6 | 442 | seq_printf(m, "LDLM pool state (%s):\n" |
b0f5aad5 GKH |
443 | " SLV: %llu\n" |
444 | " CLV: %llu\n" | |
73bb1da6 PT |
445 | " LVF: %d\n", |
446 | pl->pl_name, slv, clv, lvf); | |
d7e09d03 | 447 | |
2c2b7c05 HM |
448 | seq_printf(m, " GR: %d\n CR: %d\n GS: %d\n" |
449 | " G: %d\n L: %d\n", | |
73bb1da6 PT |
450 | grant_rate, cancel_rate, grant_speed, |
451 | granted, limit); | |
452 | ||
453 | return 0; | |
d7e09d03 | 454 | } |
c9f6bb96 | 455 | |
73bb1da6 | 456 | LPROC_SEQ_FOPS_RO(lprocfs_pool_state); |
d7e09d03 | 457 | |
24b8c88a OD |
458 | static ssize_t grant_speed_show(struct kobject *kobj, struct attribute *attr, |
459 | char *buf) | |
d7e09d03 | 460 | { |
24b8c88a OD |
461 | struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool, |
462 | pl_kobj); | |
463 | ||
d7e09d03 PT |
464 | int grant_speed; |
465 | ||
466 | spin_lock(&pl->pl_lock); | |
467 | /* serialize with ldlm_pool_recalc */ | |
468 | grant_speed = atomic_read(&pl->pl_grant_rate) - | |
469 | atomic_read(&pl->pl_cancel_rate); | |
470 | spin_unlock(&pl->pl_lock); | |
24b8c88a | 471 | return sprintf(buf, "%d\n", grant_speed); |
d7e09d03 | 472 | } |
24b8c88a | 473 | LUSTRE_RO_ATTR(grant_speed); |
d7e09d03 | 474 | |
24b8c88a OD |
475 | LDLM_POOL_SYSFS_READER_SHOW(grant_plan, int); |
476 | LUSTRE_RO_ATTR(grant_plan); | |
73bb1da6 | 477 | |
24b8c88a OD |
478 | LDLM_POOL_SYSFS_READER_SHOW(recalc_period, int); |
479 | LDLM_POOL_SYSFS_WRITER_STORE(recalc_period, int); | |
480 | LUSTRE_RW_ATTR(recalc_period); | |
73bb1da6 | 481 | |
24b8c88a OD |
482 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(server_lock_volume, u64); |
483 | LUSTRE_RO_ATTR(server_lock_volume); | |
484 | ||
485 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(limit, atomic); | |
486 | LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(limit, atomic); | |
487 | LUSTRE_RW_ATTR(limit); | |
488 | ||
489 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(granted, atomic); | |
490 | LUSTRE_RO_ATTR(granted); | |
491 | ||
492 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(cancel_rate, atomic); | |
493 | LUSTRE_RO_ATTR(cancel_rate); | |
73bb1da6 | 494 | |
24b8c88a OD |
495 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(grant_rate, atomic); |
496 | LUSTRE_RO_ATTR(grant_rate); | |
73bb1da6 | 497 | |
24b8c88a OD |
498 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(lock_volume_factor, atomic); |
499 | LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(lock_volume_factor, atomic); | |
500 | LUSTRE_RW_ATTR(lock_volume_factor); | |
73bb1da6 PT |
501 | |
502 | #define LDLM_POOL_ADD_VAR(name, var, ops) \ | |
503 | do { \ | |
504 | snprintf(var_name, MAX_STRING_SIZE, #name); \ | |
505 | pool_vars[0].data = var; \ | |
506 | pool_vars[0].fops = ops; \ | |
700815d4 | 507 | ldebugfs_add_vars(pl->pl_debugfs_entry, pool_vars, NULL);\ |
73bb1da6 | 508 | } while (0) |
d7e09d03 | 509 | |
f2825e03 OD |
510 | /* These are for pools in /sys/fs/lustre/ldlm/namespaces/.../pool */ |
511 | static struct attribute *ldlm_pl_attrs[] = { | |
24b8c88a OD |
512 | &lustre_attr_grant_speed.attr, |
513 | &lustre_attr_grant_plan.attr, | |
514 | &lustre_attr_recalc_period.attr, | |
515 | &lustre_attr_server_lock_volume.attr, | |
516 | &lustre_attr_limit.attr, | |
517 | &lustre_attr_granted.attr, | |
518 | &lustre_attr_cancel_rate.attr, | |
519 | &lustre_attr_grant_rate.attr, | |
520 | &lustre_attr_lock_volume_factor.attr, | |
f2825e03 OD |
521 | NULL, |
522 | }; | |
523 | ||
524 | static void ldlm_pl_release(struct kobject *kobj) | |
525 | { | |
526 | struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool, | |
527 | pl_kobj); | |
528 | complete(&pl->pl_kobj_unregister); | |
529 | } | |
530 | ||
531 | static struct kobj_type ldlm_pl_ktype = { | |
532 | .default_attrs = ldlm_pl_attrs, | |
533 | .sysfs_ops = &lustre_sysfs_ops, | |
534 | .release = ldlm_pl_release, | |
535 | }; | |
536 | ||
537 | static int ldlm_pool_sysfs_init(struct ldlm_pool *pl) | |
538 | { | |
7c37abe0 SB |
539 | struct ldlm_namespace *ns = container_of(pl, struct ldlm_namespace, |
540 | ns_pool); | |
f2825e03 OD |
541 | int err; |
542 | ||
543 | init_completion(&pl->pl_kobj_unregister); | |
544 | err = kobject_init_and_add(&pl->pl_kobj, &ldlm_pl_ktype, &ns->ns_kobj, | |
545 | "pool"); | |
546 | ||
547 | return err; | |
548 | } | |
549 | ||
700815d4 | 550 | static int ldlm_pool_debugfs_init(struct ldlm_pool *pl) |
d7e09d03 | 551 | { |
7c37abe0 SB |
552 | struct ldlm_namespace *ns = container_of(pl, struct ldlm_namespace, |
553 | ns_pool); | |
700815d4 | 554 | struct dentry *debugfs_ns_parent; |
d7e09d03 PT |
555 | struct lprocfs_vars pool_vars[2]; |
556 | char *var_name = NULL; | |
557 | int rc = 0; | |
d7e09d03 | 558 | |
352f7891 | 559 | var_name = kzalloc(MAX_STRING_SIZE + 1, GFP_NOFS); |
d7e09d03 | 560 | if (!var_name) |
0a3bdb00 | 561 | return -ENOMEM; |
d7e09d03 | 562 | |
700815d4 DE |
563 | debugfs_ns_parent = ns->ns_debugfs_entry; |
564 | if (IS_ERR_OR_NULL(debugfs_ns_parent)) { | |
565 | CERROR("%s: debugfs entry is not initialized\n", | |
d7e09d03 | 566 | ldlm_ns_name(ns)); |
d1c0d446 JL |
567 | rc = -EINVAL; |
568 | goto out_free_name; | |
d7e09d03 | 569 | } |
700815d4 DE |
570 | pl->pl_debugfs_entry = ldebugfs_register("pool", debugfs_ns_parent, |
571 | NULL, NULL); | |
572 | if (IS_ERR(pl->pl_debugfs_entry)) { | |
573 | CERROR("LdebugFS failed in ldlm-pool-init\n"); | |
574 | rc = PTR_ERR(pl->pl_debugfs_entry); | |
575 | pl->pl_debugfs_entry = NULL; | |
d1c0d446 | 576 | goto out_free_name; |
d7e09d03 PT |
577 | } |
578 | ||
579 | var_name[MAX_STRING_SIZE] = '\0'; | |
580 | memset(pool_vars, 0, sizeof(pool_vars)); | |
581 | pool_vars[0].name = var_name; | |
582 | ||
700815d4 | 583 | LDLM_POOL_ADD_VAR(state, pl, &lprocfs_pool_state_fops); |
d7e09d03 PT |
584 | |
585 | pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT - | |
586 | LDLM_POOL_FIRST_STAT, 0); | |
d1c0d446 JL |
587 | if (!pl->pl_stats) { |
588 | rc = -ENOMEM; | |
589 | goto out_free_name; | |
590 | } | |
d7e09d03 PT |
591 | |
592 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT, | |
593 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
594 | "granted", "locks"); | |
595 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT, | |
596 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
597 | "grant", "locks"); | |
598 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT, | |
599 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
600 | "cancel", "locks"); | |
601 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT, | |
602 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
603 | "grant_rate", "locks/s"); | |
604 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT, | |
605 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
606 | "cancel_rate", "locks/s"); | |
607 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT, | |
608 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
609 | "grant_plan", "locks/s"); | |
610 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT, | |
611 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
612 | "slv", "slv"); | |
613 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT, | |
614 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
615 | "shrink_request", "locks"); | |
616 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT, | |
617 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
618 | "shrink_freed", "locks"); | |
619 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT, | |
620 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
621 | "recalc_freed", "locks"); | |
622 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT, | |
623 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
624 | "recalc_timing", "sec"); | |
700815d4 DE |
625 | rc = ldebugfs_register_stats(pl->pl_debugfs_entry, "stats", |
626 | pl->pl_stats); | |
d7e09d03 | 627 | |
d7e09d03 | 628 | out_free_name: |
352f7891 | 629 | kfree(var_name); |
d7e09d03 PT |
630 | return rc; |
631 | } | |
632 | ||
f2825e03 OD |
633 | static void ldlm_pool_sysfs_fini(struct ldlm_pool *pl) |
634 | { | |
635 | kobject_put(&pl->pl_kobj); | |
636 | wait_for_completion(&pl->pl_kobj_unregister); | |
637 | } | |
638 | ||
700815d4 | 639 | static void ldlm_pool_debugfs_fini(struct ldlm_pool *pl) |
d7e09d03 | 640 | { |
44b53f18 | 641 | if (pl->pl_stats) { |
d7e09d03 PT |
642 | lprocfs_free_stats(&pl->pl_stats); |
643 | pl->pl_stats = NULL; | |
644 | } | |
44b53f18 | 645 | if (pl->pl_debugfs_entry) { |
700815d4 DE |
646 | ldebugfs_remove(&pl->pl_debugfs_entry); |
647 | pl->pl_debugfs_entry = NULL; | |
d7e09d03 PT |
648 | } |
649 | } | |
650 | ||
651 | int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns, | |
652 | int idx, ldlm_side_t client) | |
653 | { | |
654 | int rc; | |
d7e09d03 PT |
655 | |
656 | spin_lock_init(&pl->pl_lock); | |
657 | atomic_set(&pl->pl_granted, 0); | |
8f83409c | 658 | pl->pl_recalc_time = ktime_get_seconds(); |
d7e09d03 PT |
659 | atomic_set(&pl->pl_lock_volume_factor, 1); |
660 | ||
661 | atomic_set(&pl->pl_grant_rate, 0); | |
662 | atomic_set(&pl->pl_cancel_rate, 0); | |
663 | pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L); | |
664 | ||
665 | snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d", | |
666 | ldlm_ns_name(ns), idx); | |
667 | ||
f7ec22b5 | 668 | atomic_set(&pl->pl_limit, 1); |
00f9d12b OD |
669 | pl->pl_server_lock_volume = 0; |
670 | pl->pl_ops = &ldlm_cli_pool_ops; | |
671 | pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD; | |
d7e09d03 | 672 | pl->pl_client_lock_volume = 0; |
700815d4 | 673 | rc = ldlm_pool_debugfs_init(pl); |
d7e09d03 | 674 | if (rc) |
0a3bdb00 | 675 | return rc; |
d7e09d03 | 676 | |
f2825e03 OD |
677 | rc = ldlm_pool_sysfs_init(pl); |
678 | if (rc) | |
679 | return rc; | |
680 | ||
d7e09d03 PT |
681 | CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name); |
682 | ||
0a3bdb00 | 683 | return rc; |
d7e09d03 PT |
684 | } |
685 | EXPORT_SYMBOL(ldlm_pool_init); | |
686 | ||
687 | void ldlm_pool_fini(struct ldlm_pool *pl) | |
688 | { | |
f2825e03 | 689 | ldlm_pool_sysfs_fini(pl); |
700815d4 | 690 | ldlm_pool_debugfs_fini(pl); |
d7e09d03 PT |
691 | |
692 | /* | |
693 | * Pool should not be used after this point. We can't free it here as | |
694 | * it lives in struct ldlm_namespace, but still interested in catching | |
695 | * any abnormal using cases. | |
696 | */ | |
697 | POISON(pl, 0x5a, sizeof(*pl)); | |
d7e09d03 PT |
698 | } |
699 | EXPORT_SYMBOL(ldlm_pool_fini); | |
700 | ||
701 | /** | |
702 | * Add new taken ldlm lock \a lock into pool \a pl accounting. | |
703 | */ | |
704 | void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock) | |
705 | { | |
706 | /* | |
707 | * FLOCK locks are special in a sense that they are almost never | |
708 | * cancelled, instead special kind of lock is used to drop them. | |
709 | * also there is no LRU for flock locks, so no point in tracking | |
710 | * them anyway. | |
711 | */ | |
712 | if (lock->l_resource->lr_type == LDLM_FLOCK) | |
713 | return; | |
714 | ||
715 | atomic_inc(&pl->pl_granted); | |
716 | atomic_inc(&pl->pl_grant_rate); | |
717 | lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT); | |
718 | /* | |
719 | * Do not do pool recalc for client side as all locks which | |
720 | * potentially may be canceled has already been packed into | |
721 | * enqueue/cancel rpc. Also we do not want to run out of stack | |
722 | * with too long call paths. | |
723 | */ | |
d7e09d03 PT |
724 | } |
725 | EXPORT_SYMBOL(ldlm_pool_add); | |
726 | ||
727 | /** | |
728 | * Remove ldlm lock \a lock from pool \a pl accounting. | |
729 | */ | |
730 | void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock) | |
731 | { | |
732 | /* | |
733 | * Filter out FLOCK locks. Read above comment in ldlm_pool_add(). | |
734 | */ | |
735 | if (lock->l_resource->lr_type == LDLM_FLOCK) | |
736 | return; | |
737 | ||
738 | LASSERT(atomic_read(&pl->pl_granted) > 0); | |
739 | atomic_dec(&pl->pl_granted); | |
740 | atomic_inc(&pl->pl_cancel_rate); | |
741 | ||
742 | lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT); | |
d7e09d03 PT |
743 | } |
744 | EXPORT_SYMBOL(ldlm_pool_del); | |
745 | ||
746 | /** | |
747 | * Returns current \a pl SLV. | |
748 | * | |
749 | * \pre ->pl_lock is not locked. | |
750 | */ | |
751 | __u64 ldlm_pool_get_slv(struct ldlm_pool *pl) | |
752 | { | |
753 | __u64 slv; | |
902f3bb1 | 754 | |
d7e09d03 PT |
755 | spin_lock(&pl->pl_lock); |
756 | slv = pl->pl_server_lock_volume; | |
757 | spin_unlock(&pl->pl_lock); | |
758 | return slv; | |
759 | } | |
d7e09d03 | 760 | |
d7e09d03 PT |
761 | /** |
762 | * Sets passed \a clv to \a pl. | |
763 | * | |
764 | * \pre ->pl_lock is not locked. | |
765 | */ | |
766 | void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv) | |
767 | { | |
768 | spin_lock(&pl->pl_lock); | |
769 | pl->pl_client_lock_volume = clv; | |
770 | spin_unlock(&pl->pl_lock); | |
771 | } | |
d7e09d03 PT |
772 | |
773 | /** | |
774 | * Returns current LVF from \a pl. | |
775 | */ | |
776 | __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl) | |
777 | { | |
778 | return atomic_read(&pl->pl_lock_volume_factor); | |
779 | } | |
d7e09d03 PT |
780 | |
781 | static int ldlm_pool_granted(struct ldlm_pool *pl) | |
782 | { | |
783 | return atomic_read(&pl->pl_granted); | |
784 | } | |
785 | ||
786 | static struct ptlrpc_thread *ldlm_pools_thread; | |
d7e09d03 PT |
787 | static struct completion ldlm_pools_comp; |
788 | ||
789 | /* | |
cbc3769e PT |
790 | * count locks from all namespaces (if possible). Returns number of |
791 | * cached locks. | |
d7e09d03 | 792 | */ |
5802572e | 793 | static unsigned long ldlm_pools_count(ldlm_side_t client, gfp_t gfp_mask) |
d7e09d03 | 794 | { |
cbc3769e | 795 | int total = 0, nr_ns; |
d7e09d03 | 796 | struct ldlm_namespace *ns; |
91a50030 | 797 | struct ldlm_namespace *ns_old = NULL; /* loop detection */ |
d7e09d03 PT |
798 | void *cookie; |
799 | ||
cbc3769e PT |
800 | if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS)) |
801 | return 0; | |
d7e09d03 | 802 | |
cbc3769e PT |
803 | CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n", |
804 | client == LDLM_NAMESPACE_CLIENT ? "client" : "server"); | |
d7e09d03 PT |
805 | |
806 | cookie = cl_env_reenter(); | |
807 | ||
808 | /* | |
809 | * Find out how many resources we may release. | |
810 | */ | |
91a50030 | 811 | for (nr_ns = ldlm_namespace_nr_read(client); |
cbc3769e | 812 | nr_ns > 0; nr_ns--) { |
d7e09d03 PT |
813 | mutex_lock(ldlm_namespace_lock(client)); |
814 | if (list_empty(ldlm_namespace_list(client))) { | |
815 | mutex_unlock(ldlm_namespace_lock(client)); | |
816 | cl_env_reexit(cookie); | |
817 | return 0; | |
818 | } | |
819 | ns = ldlm_namespace_first_locked(client); | |
91a50030 OD |
820 | |
821 | if (ns == ns_old) { | |
822 | mutex_unlock(ldlm_namespace_lock(client)); | |
823 | break; | |
824 | } | |
825 | ||
826 | if (ldlm_ns_empty(ns)) { | |
827 | ldlm_namespace_move_to_inactive_locked(ns, client); | |
828 | mutex_unlock(ldlm_namespace_lock(client)); | |
829 | continue; | |
830 | } | |
831 | ||
44b53f18 | 832 | if (!ns_old) |
91a50030 OD |
833 | ns_old = ns; |
834 | ||
d7e09d03 | 835 | ldlm_namespace_get(ns); |
91a50030 | 836 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
837 | mutex_unlock(ldlm_namespace_lock(client)); |
838 | total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask); | |
839 | ldlm_namespace_put(ns); | |
840 | } | |
841 | ||
cbc3769e PT |
842 | cl_env_reexit(cookie); |
843 | return total; | |
844 | } | |
845 | ||
5802572e | 846 | static unsigned long ldlm_pools_scan(ldlm_side_t client, int nr, gfp_t gfp_mask) |
cbc3769e PT |
847 | { |
848 | unsigned long freed = 0; | |
849 | int tmp, nr_ns; | |
850 | struct ldlm_namespace *ns; | |
851 | void *cookie; | |
852 | ||
853 | if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS)) | |
854 | return -1; | |
855 | ||
856 | cookie = cl_env_reenter(); | |
d7e09d03 PT |
857 | |
858 | /* | |
cbc3769e | 859 | * Shrink at least ldlm_namespace_nr_read(client) namespaces. |
d7e09d03 | 860 | */ |
cbc3769e PT |
861 | for (tmp = nr_ns = ldlm_namespace_nr_read(client); |
862 | tmp > 0; tmp--) { | |
d7e09d03 PT |
863 | int cancel, nr_locks; |
864 | ||
865 | /* | |
866 | * Do not call shrink under ldlm_namespace_lock(client) | |
867 | */ | |
868 | mutex_lock(ldlm_namespace_lock(client)); | |
869 | if (list_empty(ldlm_namespace_list(client))) { | |
870 | mutex_unlock(ldlm_namespace_lock(client)); | |
d7e09d03 PT |
871 | break; |
872 | } | |
873 | ns = ldlm_namespace_first_locked(client); | |
874 | ldlm_namespace_get(ns); | |
91a50030 | 875 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
876 | mutex_unlock(ldlm_namespace_lock(client)); |
877 | ||
878 | nr_locks = ldlm_pool_granted(&ns->ns_pool); | |
cbc3769e PT |
879 | /* |
880 | * We use to shrink propotionally but with new shrinker API, | |
881 | * we lost the total number of freeable locks. | |
882 | */ | |
883 | cancel = 1 + min_t(int, nr_locks, nr / nr_ns); | |
884 | freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask); | |
d7e09d03 PT |
885 | ldlm_namespace_put(ns); |
886 | } | |
887 | cl_env_reexit(cookie); | |
cbc3769e PT |
888 | /* |
889 | * we only decrease the SLV in server pools shrinker, return | |
890 | * SHRINK_STOP to kernel to avoid needless loop. LU-1128 | |
891 | */ | |
00f9d12b | 892 | return freed; |
d7e09d03 PT |
893 | } |
894 | ||
e7ddc48c AR |
895 | static unsigned long ldlm_pools_cli_count(struct shrinker *s, |
896 | struct shrink_control *sc) | |
d7e09d03 | 897 | { |
cbc3769e PT |
898 | return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask); |
899 | } | |
900 | ||
e7ddc48c AR |
901 | static unsigned long ldlm_pools_cli_scan(struct shrinker *s, |
902 | struct shrink_control *sc) | |
cbc3769e PT |
903 | { |
904 | return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan, | |
905 | sc->gfp_mask); | |
d7e09d03 PT |
906 | } |
907 | ||
00f9d12b | 908 | static int ldlm_pools_recalc(ldlm_side_t client) |
d7e09d03 | 909 | { |
d7e09d03 | 910 | struct ldlm_namespace *ns; |
91a50030 | 911 | struct ldlm_namespace *ns_old = NULL; |
00f9d12b | 912 | int nr; |
3eface59 | 913 | int time = 50; /* seconds of sleep if no active namespaces */ |
d7e09d03 | 914 | |
d7e09d03 | 915 | /* |
cbc3769e | 916 | * Recalc at least ldlm_namespace_nr_read(client) namespaces. |
d7e09d03 | 917 | */ |
91a50030 | 918 | for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) { |
d7e09d03 PT |
919 | int skip; |
920 | /* | |
921 | * Lock the list, get first @ns in the list, getref, move it | |
922 | * to the tail, unlock and call pool recalc. This way we avoid | |
923 | * calling recalc under @ns lock what is really good as we get | |
924 | * rid of potential deadlock on client nodes when canceling | |
925 | * locks synchronously. | |
926 | */ | |
927 | mutex_lock(ldlm_namespace_lock(client)); | |
928 | if (list_empty(ldlm_namespace_list(client))) { | |
929 | mutex_unlock(ldlm_namespace_lock(client)); | |
930 | break; | |
931 | } | |
932 | ns = ldlm_namespace_first_locked(client); | |
933 | ||
91a50030 OD |
934 | if (ns_old == ns) { /* Full pass complete */ |
935 | mutex_unlock(ldlm_namespace_lock(client)); | |
936 | break; | |
937 | } | |
938 | ||
939 | /* We got an empty namespace, need to move it back to inactive | |
940 | * list. | |
941 | * The race with parallel resource creation is fine: | |
942 | * - If they do namespace_get before our check, we fail the | |
943 | * check and they move this item to the end of the list anyway | |
944 | * - If we do the check and then they do namespace_get, then | |
945 | * we move the namespace to inactive and they will move | |
946 | * it back to active (synchronised by the lock, so no clash | |
947 | * there). | |
948 | */ | |
949 | if (ldlm_ns_empty(ns)) { | |
950 | ldlm_namespace_move_to_inactive_locked(ns, client); | |
951 | mutex_unlock(ldlm_namespace_lock(client)); | |
952 | continue; | |
953 | } | |
954 | ||
44b53f18 | 955 | if (!ns_old) |
91a50030 OD |
956 | ns_old = ns; |
957 | ||
d7e09d03 PT |
958 | spin_lock(&ns->ns_lock); |
959 | /* | |
960 | * skip ns which is being freed, and we don't want to increase | |
961 | * its refcount again, not even temporarily. bz21519 & LU-499. | |
962 | */ | |
963 | if (ns->ns_stopping) { | |
964 | skip = 1; | |
965 | } else { | |
966 | skip = 0; | |
967 | ldlm_namespace_get(ns); | |
968 | } | |
969 | spin_unlock(&ns->ns_lock); | |
970 | ||
91a50030 | 971 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
972 | mutex_unlock(ldlm_namespace_lock(client)); |
973 | ||
974 | /* | |
975 | * After setup is done - recalc the pool. | |
976 | */ | |
977 | if (!skip) { | |
3eface59 OD |
978 | int ttime = ldlm_pool_recalc(&ns->ns_pool); |
979 | ||
980 | if (ttime < time) | |
981 | time = ttime; | |
982 | ||
d7e09d03 PT |
983 | ldlm_namespace_put(ns); |
984 | } | |
985 | } | |
3eface59 | 986 | return time; |
d7e09d03 | 987 | } |
d7e09d03 PT |
988 | |
989 | static int ldlm_pools_thread_main(void *arg) | |
990 | { | |
991 | struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg; | |
00f9d12b | 992 | int c_time; |
d7e09d03 PT |
993 | |
994 | thread_set_flags(thread, SVC_RUNNING); | |
995 | wake_up(&thread->t_ctl_waitq); | |
996 | ||
997 | CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n", | |
24c198e9 | 998 | "ldlm_poold", current_pid()); |
d7e09d03 PT |
999 | |
1000 | while (1) { | |
1001 | struct l_wait_info lwi; | |
1002 | ||
1003 | /* | |
1004 | * Recal all pools on this tick. | |
1005 | */ | |
3eface59 | 1006 | c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT); |
d7e09d03 PT |
1007 | |
1008 | /* | |
1009 | * Wait until the next check time, or until we're | |
1010 | * stopped. | |
1011 | */ | |
00f9d12b | 1012 | lwi = LWI_TIMEOUT(cfs_time_seconds(c_time), |
d7e09d03 PT |
1013 | NULL, NULL); |
1014 | l_wait_event(thread->t_ctl_waitq, | |
1015 | thread_is_stopping(thread) || | |
1016 | thread_is_event(thread), | |
1017 | &lwi); | |
1018 | ||
1019 | if (thread_test_and_clear_flags(thread, SVC_STOPPING)) | |
1020 | break; | |
71e8dd9a | 1021 | thread_test_and_clear_flags(thread, SVC_EVENT); |
d7e09d03 PT |
1022 | } |
1023 | ||
1024 | thread_set_flags(thread, SVC_STOPPED); | |
1025 | wake_up(&thread->t_ctl_waitq); | |
1026 | ||
1027 | CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n", | |
24c198e9 | 1028 | "ldlm_poold", current_pid()); |
d7e09d03 PT |
1029 | |
1030 | complete_and_exit(&ldlm_pools_comp, 0); | |
1031 | } | |
1032 | ||
1033 | static int ldlm_pools_thread_start(void) | |
1034 | { | |
1035 | struct l_wait_info lwi = { 0 }; | |
68b636b6 | 1036 | struct task_struct *task; |
d7e09d03 | 1037 | |
44b53f18 | 1038 | if (ldlm_pools_thread) |
0a3bdb00 | 1039 | return -EALREADY; |
d7e09d03 | 1040 | |
352f7891 | 1041 | ldlm_pools_thread = kzalloc(sizeof(*ldlm_pools_thread), GFP_NOFS); |
94e67761 | 1042 | if (!ldlm_pools_thread) |
0a3bdb00 | 1043 | return -ENOMEM; |
d7e09d03 PT |
1044 | |
1045 | init_completion(&ldlm_pools_comp); | |
1046 | init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq); | |
1047 | ||
1048 | task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread, | |
1049 | "ldlm_poold"); | |
1050 | if (IS_ERR(task)) { | |
1051 | CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task)); | |
352f7891 | 1052 | kfree(ldlm_pools_thread); |
d7e09d03 | 1053 | ldlm_pools_thread = NULL; |
0a3bdb00 | 1054 | return PTR_ERR(task); |
d7e09d03 PT |
1055 | } |
1056 | l_wait_event(ldlm_pools_thread->t_ctl_waitq, | |
1057 | thread_is_running(ldlm_pools_thread), &lwi); | |
0a3bdb00 | 1058 | return 0; |
d7e09d03 PT |
1059 | } |
1060 | ||
1061 | static void ldlm_pools_thread_stop(void) | |
1062 | { | |
44b53f18 | 1063 | if (!ldlm_pools_thread) |
d7e09d03 | 1064 | return; |
d7e09d03 PT |
1065 | |
1066 | thread_set_flags(ldlm_pools_thread, SVC_STOPPING); | |
1067 | wake_up(&ldlm_pools_thread->t_ctl_waitq); | |
1068 | ||
1069 | /* | |
1070 | * Make sure that pools thread is finished before freeing @thread. | |
1071 | * This fixes possible race and oops due to accessing freed memory | |
1072 | * in pools thread. | |
1073 | */ | |
1074 | wait_for_completion(&ldlm_pools_comp); | |
352f7891 | 1075 | kfree(ldlm_pools_thread); |
d7e09d03 | 1076 | ldlm_pools_thread = NULL; |
d7e09d03 PT |
1077 | } |
1078 | ||
cbc3769e PT |
1079 | static struct shrinker ldlm_pools_cli_shrinker = { |
1080 | .count_objects = ldlm_pools_cli_count, | |
1081 | .scan_objects = ldlm_pools_cli_scan, | |
1082 | .seeks = DEFAULT_SEEKS, | |
1083 | }; | |
1084 | ||
d7e09d03 PT |
1085 | int ldlm_pools_init(void) |
1086 | { | |
1087 | int rc; | |
d7e09d03 PT |
1088 | |
1089 | rc = ldlm_pools_thread_start(); | |
00f9d12b | 1090 | if (rc == 0) |
cbc3769e | 1091 | register_shrinker(&ldlm_pools_cli_shrinker); |
00f9d12b | 1092 | |
0a3bdb00 | 1093 | return rc; |
d7e09d03 PT |
1094 | } |
1095 | EXPORT_SYMBOL(ldlm_pools_init); | |
1096 | ||
1097 | void ldlm_pools_fini(void) | |
1098 | { | |
00f9d12b | 1099 | if (ldlm_pools_thread) |
faa7a4e3 | 1100 | unregister_shrinker(&ldlm_pools_cli_shrinker); |
00f9d12b | 1101 | |
d7e09d03 PT |
1102 | ldlm_pools_thread_stop(); |
1103 | } | |
1104 | EXPORT_SYMBOL(ldlm_pools_fini); |