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 | */ | |
09cbfeaf | 110 | #define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_SHIFT)) * 50) |
d7e09d03 PT |
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 | read_lock(&obd->obd_pool_lock); |
250 | pl->pl_server_lock_volume = obd->obd_pool_slv; | |
f7ec22b5 | 251 | atomic_set(&pl->pl_limit, obd->obd_pool_limit); |
d7e09d03 PT |
252 | read_unlock(&obd->obd_pool_lock); |
253 | } | |
254 | ||
255 | /** | |
256 | * Recalculates client size pool \a pl according to current SLV and Limit. | |
257 | */ | |
258 | static int ldlm_cli_pool_recalc(struct ldlm_pool *pl) | |
259 | { | |
8f83409c | 260 | time64_t recalc_interval_sec; |
4d2c7b30 | 261 | int ret; |
d7e09d03 | 262 | |
8f83409c | 263 | recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time; |
d7e09d03 | 264 | if (recalc_interval_sec < pl->pl_recalc_period) |
0a3bdb00 | 265 | return 0; |
d7e09d03 PT |
266 | |
267 | spin_lock(&pl->pl_lock); | |
268 | /* | |
269 | * Check if we need to recalc lists now. | |
270 | */ | |
8f83409c | 271 | recalc_interval_sec = ktime_get_real_seconds() - pl->pl_recalc_time; |
d7e09d03 PT |
272 | if (recalc_interval_sec < pl->pl_recalc_period) { |
273 | spin_unlock(&pl->pl_lock); | |
0a3bdb00 | 274 | return 0; |
d7e09d03 PT |
275 | } |
276 | ||
277 | /* | |
278 | * Make sure that pool knows last SLV and Limit from obd. | |
279 | */ | |
280 | ldlm_cli_pool_pop_slv(pl); | |
281 | ||
d7e09d03 PT |
282 | spin_unlock(&pl->pl_lock); |
283 | ||
284 | /* | |
285 | * Do not cancel locks in case lru resize is disabled for this ns. | |
286 | */ | |
7c37abe0 SB |
287 | if (!ns_connect_lru_resize(container_of(pl, struct ldlm_namespace, |
288 | ns_pool))) { | |
4d2c7b30 LX |
289 | ret = 0; |
290 | goto out; | |
291 | } | |
d7e09d03 PT |
292 | |
293 | /* | |
294 | * In the time of canceling locks on client we do not need to maintain | |
295 | * sharp timing, we only want to cancel locks asap according to new SLV. | |
296 | * It may be called when SLV has changed much, this is why we do not | |
297 | * take into account pl->pl_recalc_time here. | |
298 | */ | |
7c37abe0 SB |
299 | ret = ldlm_cancel_lru(container_of(pl, struct ldlm_namespace, ns_pool), |
300 | 0, LCF_ASYNC, LDLM_CANCEL_LRUR); | |
4d2c7b30 LX |
301 | |
302 | out: | |
303 | spin_lock(&pl->pl_lock); | |
304 | /* | |
305 | * Time of LRU resizing might be longer than period, | |
306 | * so update after LRU resizing rather than before it. | |
307 | */ | |
8f83409c | 308 | pl->pl_recalc_time = ktime_get_real_seconds(); |
4d2c7b30 LX |
309 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT, |
310 | recalc_interval_sec); | |
311 | spin_unlock(&pl->pl_lock); | |
312 | return ret; | |
d7e09d03 PT |
313 | } |
314 | ||
315 | /** | |
316 | * This function is main entry point for memory pressure handling on client | |
317 | * side. Main goal of this function is to cancel some number of locks on | |
318 | * passed \a pl according to \a nr and \a gfp_mask. | |
319 | */ | |
320 | static int ldlm_cli_pool_shrink(struct ldlm_pool *pl, | |
5802572e | 321 | int nr, gfp_t gfp_mask) |
d7e09d03 PT |
322 | { |
323 | struct ldlm_namespace *ns; | |
cbc3769e | 324 | int unused; |
d7e09d03 | 325 | |
7c37abe0 | 326 | ns = container_of(pl, struct ldlm_namespace, ns_pool); |
d7e09d03 PT |
327 | |
328 | /* | |
329 | * Do not cancel locks in case lru resize is disabled for this ns. | |
330 | */ | |
331 | if (!ns_connect_lru_resize(ns)) | |
0a3bdb00 | 332 | return 0; |
d7e09d03 PT |
333 | |
334 | /* | |
335 | * Make sure that pool knows last SLV and Limit from obd. | |
336 | */ | |
337 | ldlm_cli_pool_pop_slv(pl); | |
338 | ||
339 | spin_lock(&ns->ns_lock); | |
340 | unused = ns->ns_nr_unused; | |
341 | spin_unlock(&ns->ns_lock); | |
342 | ||
cbc3769e PT |
343 | if (nr == 0) |
344 | return (unused / 100) * sysctl_vfs_cache_pressure; | |
345 | else | |
346 | return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_CANCEL_SHRINK); | |
d7e09d03 PT |
347 | } |
348 | ||
b9c98cfa | 349 | static const struct ldlm_pool_ops ldlm_cli_pool_ops = { |
d7e09d03 PT |
350 | .po_recalc = ldlm_cli_pool_recalc, |
351 | .po_shrink = ldlm_cli_pool_shrink | |
352 | }; | |
353 | ||
354 | /** | |
355 | * Pool recalc wrapper. Will call either client or server pool recalc callback | |
356 | * depending what pool \a pl is used. | |
357 | */ | |
58c6d133 | 358 | static int ldlm_pool_recalc(struct ldlm_pool *pl) |
d7e09d03 | 359 | { |
8f83409c | 360 | u32 recalc_interval_sec; |
d7e09d03 PT |
361 | int count; |
362 | ||
8f83409c | 363 | recalc_interval_sec = ktime_get_seconds() - pl->pl_recalc_time; |
d7e09d03 PT |
364 | if (recalc_interval_sec <= 0) |
365 | goto recalc; | |
366 | ||
367 | spin_lock(&pl->pl_lock); | |
d7e09d03 PT |
368 | if (recalc_interval_sec > 0) { |
369 | /* | |
370 | * Update pool statistics every 1s. | |
371 | */ | |
372 | ldlm_pool_recalc_stats(pl); | |
373 | ||
374 | /* | |
375 | * Zero out all rates and speed for the last period. | |
376 | */ | |
377 | atomic_set(&pl->pl_grant_rate, 0); | |
378 | atomic_set(&pl->pl_cancel_rate, 0); | |
379 | } | |
380 | spin_unlock(&pl->pl_lock); | |
381 | ||
382 | recalc: | |
44b53f18 | 383 | if (pl->pl_ops->po_recalc) { |
d7e09d03 PT |
384 | count = pl->pl_ops->po_recalc(pl); |
385 | lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT, | |
386 | count); | |
d7e09d03 | 387 | } |
8f83409c | 388 | recalc_interval_sec = pl->pl_recalc_time - ktime_get_seconds() + |
3eface59 | 389 | pl->pl_recalc_period; |
4d2c7b30 LX |
390 | if (recalc_interval_sec <= 0) { |
391 | /* Prevent too frequent recalculation. */ | |
8f83409c AB |
392 | CDEBUG(D_DLMTRACE, |
393 | "Negative interval(%d), too short period(%lld)", | |
4d2c7b30 | 394 | recalc_interval_sec, |
8f83409c | 395 | (s64)pl->pl_recalc_period); |
4d2c7b30 LX |
396 | recalc_interval_sec = 1; |
397 | } | |
d7e09d03 | 398 | |
3eface59 | 399 | return recalc_interval_sec; |
d7e09d03 | 400 | } |
d7e09d03 | 401 | |
cbc3769e | 402 | /* |
d7e09d03 | 403 | * Pool shrink wrapper. Will call either client or server pool recalc callback |
cbc3769e PT |
404 | * depending what pool pl is used. When nr == 0, just return the number of |
405 | * freeable locks. Otherwise, return the number of canceled locks. | |
d7e09d03 | 406 | */ |
58c6d133 | 407 | static int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, gfp_t gfp_mask) |
d7e09d03 PT |
408 | { |
409 | int cancel = 0; | |
410 | ||
44b53f18 | 411 | if (pl->pl_ops->po_shrink) { |
d7e09d03 PT |
412 | cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask); |
413 | if (nr > 0) { | |
414 | lprocfs_counter_add(pl->pl_stats, | |
415 | LDLM_POOL_SHRINK_REQTD_STAT, | |
416 | nr); | |
417 | lprocfs_counter_add(pl->pl_stats, | |
418 | LDLM_POOL_SHRINK_FREED_STAT, | |
419 | cancel); | |
2d00bd17 JP |
420 | CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, shrunk %d\n", |
421 | pl->pl_name, nr, cancel); | |
d7e09d03 PT |
422 | } |
423 | } | |
424 | return cancel; | |
425 | } | |
d7e09d03 | 426 | |
73bb1da6 | 427 | static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused) |
d7e09d03 | 428 | { |
71570b98 OD |
429 | int granted, grant_rate, cancel_rate; |
430 | int grant_speed, lvf; | |
73bb1da6 | 431 | struct ldlm_pool *pl = m->private; |
d7e09d03 PT |
432 | __u64 slv, clv; |
433 | __u32 limit; | |
434 | ||
435 | spin_lock(&pl->pl_lock); | |
436 | slv = pl->pl_server_lock_volume; | |
437 | clv = pl->pl_client_lock_volume; | |
946d6f95 | 438 | limit = atomic_read(&pl->pl_limit); |
d7e09d03 PT |
439 | granted = atomic_read(&pl->pl_granted); |
440 | grant_rate = atomic_read(&pl->pl_grant_rate); | |
441 | cancel_rate = atomic_read(&pl->pl_cancel_rate); | |
442 | grant_speed = grant_rate - cancel_rate; | |
443 | lvf = atomic_read(&pl->pl_lock_volume_factor); | |
d7e09d03 PT |
444 | spin_unlock(&pl->pl_lock); |
445 | ||
73bb1da6 | 446 | seq_printf(m, "LDLM pool state (%s):\n" |
b0f5aad5 GKH |
447 | " SLV: %llu\n" |
448 | " CLV: %llu\n" | |
73bb1da6 PT |
449 | " LVF: %d\n", |
450 | pl->pl_name, slv, clv, lvf); | |
d7e09d03 | 451 | |
2c2b7c05 HM |
452 | seq_printf(m, " GR: %d\n CR: %d\n GS: %d\n" |
453 | " G: %d\n L: %d\n", | |
73bb1da6 PT |
454 | grant_rate, cancel_rate, grant_speed, |
455 | granted, limit); | |
456 | ||
457 | return 0; | |
d7e09d03 | 458 | } |
c9f6bb96 | 459 | |
73bb1da6 | 460 | LPROC_SEQ_FOPS_RO(lprocfs_pool_state); |
d7e09d03 | 461 | |
24b8c88a OD |
462 | static ssize_t grant_speed_show(struct kobject *kobj, struct attribute *attr, |
463 | char *buf) | |
d7e09d03 | 464 | { |
24b8c88a OD |
465 | struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool, |
466 | pl_kobj); | |
467 | ||
d7e09d03 PT |
468 | int grant_speed; |
469 | ||
470 | spin_lock(&pl->pl_lock); | |
471 | /* serialize with ldlm_pool_recalc */ | |
472 | grant_speed = atomic_read(&pl->pl_grant_rate) - | |
473 | atomic_read(&pl->pl_cancel_rate); | |
474 | spin_unlock(&pl->pl_lock); | |
24b8c88a | 475 | return sprintf(buf, "%d\n", grant_speed); |
d7e09d03 | 476 | } |
24b8c88a | 477 | LUSTRE_RO_ATTR(grant_speed); |
d7e09d03 | 478 | |
24b8c88a OD |
479 | LDLM_POOL_SYSFS_READER_SHOW(grant_plan, int); |
480 | LUSTRE_RO_ATTR(grant_plan); | |
73bb1da6 | 481 | |
24b8c88a OD |
482 | LDLM_POOL_SYSFS_READER_SHOW(recalc_period, int); |
483 | LDLM_POOL_SYSFS_WRITER_STORE(recalc_period, int); | |
484 | LUSTRE_RW_ATTR(recalc_period); | |
73bb1da6 | 485 | |
24b8c88a OD |
486 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(server_lock_volume, u64); |
487 | LUSTRE_RO_ATTR(server_lock_volume); | |
488 | ||
489 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(limit, atomic); | |
490 | LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(limit, atomic); | |
491 | LUSTRE_RW_ATTR(limit); | |
492 | ||
493 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(granted, atomic); | |
494 | LUSTRE_RO_ATTR(granted); | |
495 | ||
496 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(cancel_rate, atomic); | |
497 | LUSTRE_RO_ATTR(cancel_rate); | |
73bb1da6 | 498 | |
24b8c88a OD |
499 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(grant_rate, atomic); |
500 | LUSTRE_RO_ATTR(grant_rate); | |
73bb1da6 | 501 | |
24b8c88a OD |
502 | LDLM_POOL_SYSFS_READER_NOLOCK_SHOW(lock_volume_factor, atomic); |
503 | LDLM_POOL_SYSFS_WRITER_NOLOCK_STORE(lock_volume_factor, atomic); | |
504 | LUSTRE_RW_ATTR(lock_volume_factor); | |
73bb1da6 PT |
505 | |
506 | #define LDLM_POOL_ADD_VAR(name, var, ops) \ | |
507 | do { \ | |
508 | snprintf(var_name, MAX_STRING_SIZE, #name); \ | |
509 | pool_vars[0].data = var; \ | |
510 | pool_vars[0].fops = ops; \ | |
700815d4 | 511 | ldebugfs_add_vars(pl->pl_debugfs_entry, pool_vars, NULL);\ |
73bb1da6 | 512 | } while (0) |
d7e09d03 | 513 | |
f2825e03 OD |
514 | /* These are for pools in /sys/fs/lustre/ldlm/namespaces/.../pool */ |
515 | static struct attribute *ldlm_pl_attrs[] = { | |
24b8c88a OD |
516 | &lustre_attr_grant_speed.attr, |
517 | &lustre_attr_grant_plan.attr, | |
518 | &lustre_attr_recalc_period.attr, | |
519 | &lustre_attr_server_lock_volume.attr, | |
520 | &lustre_attr_limit.attr, | |
521 | &lustre_attr_granted.attr, | |
522 | &lustre_attr_cancel_rate.attr, | |
523 | &lustre_attr_grant_rate.attr, | |
524 | &lustre_attr_lock_volume_factor.attr, | |
f2825e03 OD |
525 | NULL, |
526 | }; | |
527 | ||
528 | static void ldlm_pl_release(struct kobject *kobj) | |
529 | { | |
530 | struct ldlm_pool *pl = container_of(kobj, struct ldlm_pool, | |
531 | pl_kobj); | |
532 | complete(&pl->pl_kobj_unregister); | |
533 | } | |
534 | ||
535 | static struct kobj_type ldlm_pl_ktype = { | |
536 | .default_attrs = ldlm_pl_attrs, | |
537 | .sysfs_ops = &lustre_sysfs_ops, | |
538 | .release = ldlm_pl_release, | |
539 | }; | |
540 | ||
541 | static int ldlm_pool_sysfs_init(struct ldlm_pool *pl) | |
542 | { | |
7c37abe0 SB |
543 | struct ldlm_namespace *ns = container_of(pl, struct ldlm_namespace, |
544 | ns_pool); | |
f2825e03 OD |
545 | int err; |
546 | ||
547 | init_completion(&pl->pl_kobj_unregister); | |
548 | err = kobject_init_and_add(&pl->pl_kobj, &ldlm_pl_ktype, &ns->ns_kobj, | |
549 | "pool"); | |
550 | ||
551 | return err; | |
552 | } | |
553 | ||
700815d4 | 554 | static int ldlm_pool_debugfs_init(struct ldlm_pool *pl) |
d7e09d03 | 555 | { |
7c37abe0 SB |
556 | struct ldlm_namespace *ns = container_of(pl, struct ldlm_namespace, |
557 | ns_pool); | |
700815d4 | 558 | struct dentry *debugfs_ns_parent; |
d7e09d03 PT |
559 | struct lprocfs_vars pool_vars[2]; |
560 | char *var_name = NULL; | |
561 | int rc = 0; | |
d7e09d03 | 562 | |
352f7891 | 563 | var_name = kzalloc(MAX_STRING_SIZE + 1, GFP_NOFS); |
d7e09d03 | 564 | if (!var_name) |
0a3bdb00 | 565 | return -ENOMEM; |
d7e09d03 | 566 | |
700815d4 DE |
567 | debugfs_ns_parent = ns->ns_debugfs_entry; |
568 | if (IS_ERR_OR_NULL(debugfs_ns_parent)) { | |
569 | CERROR("%s: debugfs entry is not initialized\n", | |
d7e09d03 | 570 | ldlm_ns_name(ns)); |
d1c0d446 JL |
571 | rc = -EINVAL; |
572 | goto out_free_name; | |
d7e09d03 | 573 | } |
700815d4 DE |
574 | pl->pl_debugfs_entry = ldebugfs_register("pool", debugfs_ns_parent, |
575 | NULL, NULL); | |
576 | if (IS_ERR(pl->pl_debugfs_entry)) { | |
577 | CERROR("LdebugFS failed in ldlm-pool-init\n"); | |
578 | rc = PTR_ERR(pl->pl_debugfs_entry); | |
579 | pl->pl_debugfs_entry = NULL; | |
d1c0d446 | 580 | goto out_free_name; |
d7e09d03 PT |
581 | } |
582 | ||
583 | var_name[MAX_STRING_SIZE] = '\0'; | |
584 | memset(pool_vars, 0, sizeof(pool_vars)); | |
585 | pool_vars[0].name = var_name; | |
586 | ||
700815d4 | 587 | LDLM_POOL_ADD_VAR(state, pl, &lprocfs_pool_state_fops); |
d7e09d03 PT |
588 | |
589 | pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT - | |
590 | LDLM_POOL_FIRST_STAT, 0); | |
d1c0d446 JL |
591 | if (!pl->pl_stats) { |
592 | rc = -ENOMEM; | |
593 | goto out_free_name; | |
594 | } | |
d7e09d03 PT |
595 | |
596 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT, | |
597 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
598 | "granted", "locks"); | |
599 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT, | |
600 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
601 | "grant", "locks"); | |
602 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT, | |
603 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
604 | "cancel", "locks"); | |
605 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT, | |
606 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
607 | "grant_rate", "locks/s"); | |
608 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT, | |
609 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
610 | "cancel_rate", "locks/s"); | |
611 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT, | |
612 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
613 | "grant_plan", "locks/s"); | |
614 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT, | |
615 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
616 | "slv", "slv"); | |
617 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT, | |
618 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
619 | "shrink_request", "locks"); | |
620 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT, | |
621 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
622 | "shrink_freed", "locks"); | |
623 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT, | |
624 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
625 | "recalc_freed", "locks"); | |
626 | lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT, | |
627 | LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, | |
628 | "recalc_timing", "sec"); | |
700815d4 DE |
629 | rc = ldebugfs_register_stats(pl->pl_debugfs_entry, "stats", |
630 | pl->pl_stats); | |
d7e09d03 | 631 | |
d7e09d03 | 632 | out_free_name: |
352f7891 | 633 | kfree(var_name); |
d7e09d03 PT |
634 | return rc; |
635 | } | |
636 | ||
f2825e03 OD |
637 | static void ldlm_pool_sysfs_fini(struct ldlm_pool *pl) |
638 | { | |
639 | kobject_put(&pl->pl_kobj); | |
640 | wait_for_completion(&pl->pl_kobj_unregister); | |
641 | } | |
642 | ||
700815d4 | 643 | static void ldlm_pool_debugfs_fini(struct ldlm_pool *pl) |
d7e09d03 | 644 | { |
44b53f18 | 645 | if (pl->pl_stats) { |
d7e09d03 PT |
646 | lprocfs_free_stats(&pl->pl_stats); |
647 | pl->pl_stats = NULL; | |
648 | } | |
44b53f18 | 649 | if (pl->pl_debugfs_entry) { |
700815d4 DE |
650 | ldebugfs_remove(&pl->pl_debugfs_entry); |
651 | pl->pl_debugfs_entry = NULL; | |
d7e09d03 PT |
652 | } |
653 | } | |
654 | ||
655 | int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns, | |
656 | int idx, ldlm_side_t client) | |
657 | { | |
658 | int rc; | |
d7e09d03 PT |
659 | |
660 | spin_lock_init(&pl->pl_lock); | |
661 | atomic_set(&pl->pl_granted, 0); | |
8f83409c | 662 | pl->pl_recalc_time = ktime_get_seconds(); |
d7e09d03 PT |
663 | atomic_set(&pl->pl_lock_volume_factor, 1); |
664 | ||
665 | atomic_set(&pl->pl_grant_rate, 0); | |
666 | atomic_set(&pl->pl_cancel_rate, 0); | |
667 | pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L); | |
668 | ||
669 | snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d", | |
670 | ldlm_ns_name(ns), idx); | |
671 | ||
f7ec22b5 | 672 | atomic_set(&pl->pl_limit, 1); |
00f9d12b OD |
673 | pl->pl_server_lock_volume = 0; |
674 | pl->pl_ops = &ldlm_cli_pool_ops; | |
675 | pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD; | |
d7e09d03 | 676 | pl->pl_client_lock_volume = 0; |
700815d4 | 677 | rc = ldlm_pool_debugfs_init(pl); |
d7e09d03 | 678 | if (rc) |
0a3bdb00 | 679 | return rc; |
d7e09d03 | 680 | |
f2825e03 OD |
681 | rc = ldlm_pool_sysfs_init(pl); |
682 | if (rc) | |
683 | return rc; | |
684 | ||
d7e09d03 PT |
685 | CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name); |
686 | ||
0a3bdb00 | 687 | return rc; |
d7e09d03 PT |
688 | } |
689 | EXPORT_SYMBOL(ldlm_pool_init); | |
690 | ||
691 | void ldlm_pool_fini(struct ldlm_pool *pl) | |
692 | { | |
f2825e03 | 693 | ldlm_pool_sysfs_fini(pl); |
700815d4 | 694 | ldlm_pool_debugfs_fini(pl); |
d7e09d03 PT |
695 | |
696 | /* | |
697 | * Pool should not be used after this point. We can't free it here as | |
698 | * it lives in struct ldlm_namespace, but still interested in catching | |
699 | * any abnormal using cases. | |
700 | */ | |
701 | POISON(pl, 0x5a, sizeof(*pl)); | |
d7e09d03 PT |
702 | } |
703 | EXPORT_SYMBOL(ldlm_pool_fini); | |
704 | ||
705 | /** | |
706 | * Add new taken ldlm lock \a lock into pool \a pl accounting. | |
707 | */ | |
708 | void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock) | |
709 | { | |
710 | /* | |
711 | * FLOCK locks are special in a sense that they are almost never | |
712 | * cancelled, instead special kind of lock is used to drop them. | |
713 | * also there is no LRU for flock locks, so no point in tracking | |
714 | * them anyway. | |
715 | */ | |
716 | if (lock->l_resource->lr_type == LDLM_FLOCK) | |
717 | return; | |
718 | ||
719 | atomic_inc(&pl->pl_granted); | |
720 | atomic_inc(&pl->pl_grant_rate); | |
721 | lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT); | |
722 | /* | |
723 | * Do not do pool recalc for client side as all locks which | |
724 | * potentially may be canceled has already been packed into | |
725 | * enqueue/cancel rpc. Also we do not want to run out of stack | |
726 | * with too long call paths. | |
727 | */ | |
d7e09d03 PT |
728 | } |
729 | EXPORT_SYMBOL(ldlm_pool_add); | |
730 | ||
731 | /** | |
732 | * Remove ldlm lock \a lock from pool \a pl accounting. | |
733 | */ | |
734 | void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock) | |
735 | { | |
736 | /* | |
737 | * Filter out FLOCK locks. Read above comment in ldlm_pool_add(). | |
738 | */ | |
739 | if (lock->l_resource->lr_type == LDLM_FLOCK) | |
740 | return; | |
741 | ||
742 | LASSERT(atomic_read(&pl->pl_granted) > 0); | |
743 | atomic_dec(&pl->pl_granted); | |
744 | atomic_inc(&pl->pl_cancel_rate); | |
745 | ||
746 | lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT); | |
d7e09d03 PT |
747 | } |
748 | EXPORT_SYMBOL(ldlm_pool_del); | |
749 | ||
750 | /** | |
751 | * Returns current \a pl SLV. | |
752 | * | |
753 | * \pre ->pl_lock is not locked. | |
754 | */ | |
755 | __u64 ldlm_pool_get_slv(struct ldlm_pool *pl) | |
756 | { | |
757 | __u64 slv; | |
902f3bb1 | 758 | |
d7e09d03 PT |
759 | spin_lock(&pl->pl_lock); |
760 | slv = pl->pl_server_lock_volume; | |
761 | spin_unlock(&pl->pl_lock); | |
762 | return slv; | |
763 | } | |
d7e09d03 | 764 | |
d7e09d03 PT |
765 | /** |
766 | * Sets passed \a clv to \a pl. | |
767 | * | |
768 | * \pre ->pl_lock is not locked. | |
769 | */ | |
770 | void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv) | |
771 | { | |
772 | spin_lock(&pl->pl_lock); | |
773 | pl->pl_client_lock_volume = clv; | |
774 | spin_unlock(&pl->pl_lock); | |
775 | } | |
d7e09d03 PT |
776 | |
777 | /** | |
778 | * Returns current LVF from \a pl. | |
779 | */ | |
780 | __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl) | |
781 | { | |
782 | return atomic_read(&pl->pl_lock_volume_factor); | |
783 | } | |
d7e09d03 PT |
784 | |
785 | static int ldlm_pool_granted(struct ldlm_pool *pl) | |
786 | { | |
787 | return atomic_read(&pl->pl_granted); | |
788 | } | |
789 | ||
790 | static struct ptlrpc_thread *ldlm_pools_thread; | |
d7e09d03 PT |
791 | static struct completion ldlm_pools_comp; |
792 | ||
793 | /* | |
cbc3769e PT |
794 | * count locks from all namespaces (if possible). Returns number of |
795 | * cached locks. | |
d7e09d03 | 796 | */ |
5802572e | 797 | static unsigned long ldlm_pools_count(ldlm_side_t client, gfp_t gfp_mask) |
d7e09d03 | 798 | { |
cbc3769e | 799 | int total = 0, nr_ns; |
d7e09d03 | 800 | struct ldlm_namespace *ns; |
91a50030 | 801 | struct ldlm_namespace *ns_old = NULL; /* loop detection */ |
d7e09d03 PT |
802 | void *cookie; |
803 | ||
cbc3769e PT |
804 | if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS)) |
805 | return 0; | |
d7e09d03 | 806 | |
cbc3769e PT |
807 | CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n", |
808 | client == LDLM_NAMESPACE_CLIENT ? "client" : "server"); | |
d7e09d03 PT |
809 | |
810 | cookie = cl_env_reenter(); | |
811 | ||
812 | /* | |
813 | * Find out how many resources we may release. | |
814 | */ | |
91a50030 | 815 | for (nr_ns = ldlm_namespace_nr_read(client); |
cbc3769e | 816 | nr_ns > 0; nr_ns--) { |
d7e09d03 PT |
817 | mutex_lock(ldlm_namespace_lock(client)); |
818 | if (list_empty(ldlm_namespace_list(client))) { | |
819 | mutex_unlock(ldlm_namespace_lock(client)); | |
820 | cl_env_reexit(cookie); | |
821 | return 0; | |
822 | } | |
823 | ns = ldlm_namespace_first_locked(client); | |
91a50030 OD |
824 | |
825 | if (ns == ns_old) { | |
826 | mutex_unlock(ldlm_namespace_lock(client)); | |
827 | break; | |
828 | } | |
829 | ||
830 | if (ldlm_ns_empty(ns)) { | |
831 | ldlm_namespace_move_to_inactive_locked(ns, client); | |
832 | mutex_unlock(ldlm_namespace_lock(client)); | |
833 | continue; | |
834 | } | |
835 | ||
44b53f18 | 836 | if (!ns_old) |
91a50030 OD |
837 | ns_old = ns; |
838 | ||
d7e09d03 | 839 | ldlm_namespace_get(ns); |
91a50030 | 840 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
841 | mutex_unlock(ldlm_namespace_lock(client)); |
842 | total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask); | |
843 | ldlm_namespace_put(ns); | |
844 | } | |
845 | ||
cbc3769e PT |
846 | cl_env_reexit(cookie); |
847 | return total; | |
848 | } | |
849 | ||
5802572e | 850 | static unsigned long ldlm_pools_scan(ldlm_side_t client, int nr, gfp_t gfp_mask) |
cbc3769e PT |
851 | { |
852 | unsigned long freed = 0; | |
853 | int tmp, nr_ns; | |
854 | struct ldlm_namespace *ns; | |
855 | void *cookie; | |
856 | ||
857 | if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS)) | |
858 | return -1; | |
859 | ||
860 | cookie = cl_env_reenter(); | |
d7e09d03 PT |
861 | |
862 | /* | |
cbc3769e | 863 | * Shrink at least ldlm_namespace_nr_read(client) namespaces. |
d7e09d03 | 864 | */ |
cbc3769e PT |
865 | for (tmp = nr_ns = ldlm_namespace_nr_read(client); |
866 | tmp > 0; tmp--) { | |
d7e09d03 PT |
867 | int cancel, nr_locks; |
868 | ||
869 | /* | |
870 | * Do not call shrink under ldlm_namespace_lock(client) | |
871 | */ | |
872 | mutex_lock(ldlm_namespace_lock(client)); | |
873 | if (list_empty(ldlm_namespace_list(client))) { | |
874 | mutex_unlock(ldlm_namespace_lock(client)); | |
d7e09d03 PT |
875 | break; |
876 | } | |
877 | ns = ldlm_namespace_first_locked(client); | |
878 | ldlm_namespace_get(ns); | |
91a50030 | 879 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
880 | mutex_unlock(ldlm_namespace_lock(client)); |
881 | ||
882 | nr_locks = ldlm_pool_granted(&ns->ns_pool); | |
cbc3769e PT |
883 | /* |
884 | * We use to shrink propotionally but with new shrinker API, | |
885 | * we lost the total number of freeable locks. | |
886 | */ | |
887 | cancel = 1 + min_t(int, nr_locks, nr / nr_ns); | |
888 | freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask); | |
d7e09d03 PT |
889 | ldlm_namespace_put(ns); |
890 | } | |
891 | cl_env_reexit(cookie); | |
cbc3769e PT |
892 | /* |
893 | * we only decrease the SLV in server pools shrinker, return | |
894 | * SHRINK_STOP to kernel to avoid needless loop. LU-1128 | |
895 | */ | |
00f9d12b | 896 | return freed; |
d7e09d03 PT |
897 | } |
898 | ||
e7ddc48c AR |
899 | static unsigned long ldlm_pools_cli_count(struct shrinker *s, |
900 | struct shrink_control *sc) | |
d7e09d03 | 901 | { |
cbc3769e PT |
902 | return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask); |
903 | } | |
904 | ||
e7ddc48c AR |
905 | static unsigned long ldlm_pools_cli_scan(struct shrinker *s, |
906 | struct shrink_control *sc) | |
cbc3769e PT |
907 | { |
908 | return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan, | |
909 | sc->gfp_mask); | |
d7e09d03 PT |
910 | } |
911 | ||
00f9d12b | 912 | static int ldlm_pools_recalc(ldlm_side_t client) |
d7e09d03 | 913 | { |
d7e09d03 | 914 | struct ldlm_namespace *ns; |
91a50030 | 915 | struct ldlm_namespace *ns_old = NULL; |
00f9d12b | 916 | int nr; |
3eface59 | 917 | int time = 50; /* seconds of sleep if no active namespaces */ |
d7e09d03 | 918 | |
d7e09d03 | 919 | /* |
cbc3769e | 920 | * Recalc at least ldlm_namespace_nr_read(client) namespaces. |
d7e09d03 | 921 | */ |
91a50030 | 922 | for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) { |
d7e09d03 PT |
923 | int skip; |
924 | /* | |
925 | * Lock the list, get first @ns in the list, getref, move it | |
926 | * to the tail, unlock and call pool recalc. This way we avoid | |
927 | * calling recalc under @ns lock what is really good as we get | |
928 | * rid of potential deadlock on client nodes when canceling | |
929 | * locks synchronously. | |
930 | */ | |
931 | mutex_lock(ldlm_namespace_lock(client)); | |
932 | if (list_empty(ldlm_namespace_list(client))) { | |
933 | mutex_unlock(ldlm_namespace_lock(client)); | |
934 | break; | |
935 | } | |
936 | ns = ldlm_namespace_first_locked(client); | |
937 | ||
91a50030 OD |
938 | if (ns_old == ns) { /* Full pass complete */ |
939 | mutex_unlock(ldlm_namespace_lock(client)); | |
940 | break; | |
941 | } | |
942 | ||
943 | /* We got an empty namespace, need to move it back to inactive | |
944 | * list. | |
945 | * The race with parallel resource creation is fine: | |
946 | * - If they do namespace_get before our check, we fail the | |
947 | * check and they move this item to the end of the list anyway | |
948 | * - If we do the check and then they do namespace_get, then | |
949 | * we move the namespace to inactive and they will move | |
950 | * it back to active (synchronised by the lock, so no clash | |
951 | * there). | |
952 | */ | |
953 | if (ldlm_ns_empty(ns)) { | |
954 | ldlm_namespace_move_to_inactive_locked(ns, client); | |
955 | mutex_unlock(ldlm_namespace_lock(client)); | |
956 | continue; | |
957 | } | |
958 | ||
44b53f18 | 959 | if (!ns_old) |
91a50030 OD |
960 | ns_old = ns; |
961 | ||
d7e09d03 PT |
962 | spin_lock(&ns->ns_lock); |
963 | /* | |
964 | * skip ns which is being freed, and we don't want to increase | |
965 | * its refcount again, not even temporarily. bz21519 & LU-499. | |
966 | */ | |
967 | if (ns->ns_stopping) { | |
968 | skip = 1; | |
969 | } else { | |
970 | skip = 0; | |
971 | ldlm_namespace_get(ns); | |
972 | } | |
973 | spin_unlock(&ns->ns_lock); | |
974 | ||
91a50030 | 975 | ldlm_namespace_move_to_active_locked(ns, client); |
d7e09d03 PT |
976 | mutex_unlock(ldlm_namespace_lock(client)); |
977 | ||
978 | /* | |
979 | * After setup is done - recalc the pool. | |
980 | */ | |
981 | if (!skip) { | |
3eface59 OD |
982 | int ttime = ldlm_pool_recalc(&ns->ns_pool); |
983 | ||
984 | if (ttime < time) | |
985 | time = ttime; | |
986 | ||
d7e09d03 PT |
987 | ldlm_namespace_put(ns); |
988 | } | |
989 | } | |
3eface59 | 990 | return time; |
d7e09d03 | 991 | } |
d7e09d03 PT |
992 | |
993 | static int ldlm_pools_thread_main(void *arg) | |
994 | { | |
995 | struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg; | |
00f9d12b | 996 | int c_time; |
d7e09d03 PT |
997 | |
998 | thread_set_flags(thread, SVC_RUNNING); | |
999 | wake_up(&thread->t_ctl_waitq); | |
1000 | ||
1001 | CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n", | |
1002 | "ldlm_poold", current_pid()); | |
1003 | ||
1004 | while (1) { | |
1005 | struct l_wait_info lwi; | |
1006 | ||
1007 | /* | |
1008 | * Recal all pools on this tick. | |
1009 | */ | |
3eface59 | 1010 | c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT); |
d7e09d03 PT |
1011 | |
1012 | /* | |
1013 | * Wait until the next check time, or until we're | |
1014 | * stopped. | |
1015 | */ | |
00f9d12b | 1016 | lwi = LWI_TIMEOUT(cfs_time_seconds(c_time), |
d7e09d03 PT |
1017 | NULL, NULL); |
1018 | l_wait_event(thread->t_ctl_waitq, | |
1019 | thread_is_stopping(thread) || | |
1020 | thread_is_event(thread), | |
1021 | &lwi); | |
1022 | ||
1023 | if (thread_test_and_clear_flags(thread, SVC_STOPPING)) | |
1024 | break; | |
71e8dd9a | 1025 | thread_test_and_clear_flags(thread, SVC_EVENT); |
d7e09d03 PT |
1026 | } |
1027 | ||
1028 | thread_set_flags(thread, SVC_STOPPED); | |
1029 | wake_up(&thread->t_ctl_waitq); | |
1030 | ||
1031 | CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n", | |
1032 | "ldlm_poold", current_pid()); | |
1033 | ||
1034 | complete_and_exit(&ldlm_pools_comp, 0); | |
1035 | } | |
1036 | ||
1037 | static int ldlm_pools_thread_start(void) | |
1038 | { | |
1039 | struct l_wait_info lwi = { 0 }; | |
68b636b6 | 1040 | struct task_struct *task; |
d7e09d03 | 1041 | |
44b53f18 | 1042 | if (ldlm_pools_thread) |
0a3bdb00 | 1043 | return -EALREADY; |
d7e09d03 | 1044 | |
352f7891 | 1045 | ldlm_pools_thread = kzalloc(sizeof(*ldlm_pools_thread), GFP_NOFS); |
94e67761 | 1046 | if (!ldlm_pools_thread) |
0a3bdb00 | 1047 | return -ENOMEM; |
d7e09d03 PT |
1048 | |
1049 | init_completion(&ldlm_pools_comp); | |
1050 | init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq); | |
1051 | ||
1052 | task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread, | |
1053 | "ldlm_poold"); | |
1054 | if (IS_ERR(task)) { | |
1055 | CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task)); | |
352f7891 | 1056 | kfree(ldlm_pools_thread); |
d7e09d03 | 1057 | ldlm_pools_thread = NULL; |
0a3bdb00 | 1058 | return PTR_ERR(task); |
d7e09d03 PT |
1059 | } |
1060 | l_wait_event(ldlm_pools_thread->t_ctl_waitq, | |
1061 | thread_is_running(ldlm_pools_thread), &lwi); | |
0a3bdb00 | 1062 | return 0; |
d7e09d03 PT |
1063 | } |
1064 | ||
1065 | static void ldlm_pools_thread_stop(void) | |
1066 | { | |
44b53f18 | 1067 | if (!ldlm_pools_thread) |
d7e09d03 | 1068 | return; |
d7e09d03 PT |
1069 | |
1070 | thread_set_flags(ldlm_pools_thread, SVC_STOPPING); | |
1071 | wake_up(&ldlm_pools_thread->t_ctl_waitq); | |
1072 | ||
1073 | /* | |
1074 | * Make sure that pools thread is finished before freeing @thread. | |
1075 | * This fixes possible race and oops due to accessing freed memory | |
1076 | * in pools thread. | |
1077 | */ | |
1078 | wait_for_completion(&ldlm_pools_comp); | |
352f7891 | 1079 | kfree(ldlm_pools_thread); |
d7e09d03 | 1080 | ldlm_pools_thread = NULL; |
d7e09d03 PT |
1081 | } |
1082 | ||
cbc3769e PT |
1083 | static struct shrinker ldlm_pools_cli_shrinker = { |
1084 | .count_objects = ldlm_pools_cli_count, | |
1085 | .scan_objects = ldlm_pools_cli_scan, | |
1086 | .seeks = DEFAULT_SEEKS, | |
1087 | }; | |
1088 | ||
d7e09d03 PT |
1089 | int ldlm_pools_init(void) |
1090 | { | |
1091 | int rc; | |
d7e09d03 PT |
1092 | |
1093 | rc = ldlm_pools_thread_start(); | |
00f9d12b | 1094 | if (rc == 0) |
cbc3769e | 1095 | register_shrinker(&ldlm_pools_cli_shrinker); |
00f9d12b | 1096 | |
0a3bdb00 | 1097 | return rc; |
d7e09d03 PT |
1098 | } |
1099 | EXPORT_SYMBOL(ldlm_pools_init); | |
1100 | ||
1101 | void ldlm_pools_fini(void) | |
1102 | { | |
00f9d12b | 1103 | if (ldlm_pools_thread) |
faa7a4e3 | 1104 | unregister_shrinker(&ldlm_pools_cli_shrinker); |
00f9d12b | 1105 | |
d7e09d03 PT |
1106 | ldlm_pools_thread_stop(); |
1107 | } | |
1108 | EXPORT_SYMBOL(ldlm_pools_fini); |