1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2010, 2012, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/ldlm/ldlm_pool.c
37  *
38  * Author: Yury Umanets <umka@clusterfs.com>
39  */
40 
41 /*
42  * Idea of this code is rather simple. Each second, for each server namespace
43  * we have SLV - server lock volume which is calculated on current number of
44  * granted locks, grant speed for past period, etc - that is, locking load.
45  * This SLV number may be thought as a flow definition for simplicity. It is
46  * sent to clients with each occasion to let them know what is current load
47  * situation on the server. By default, at the beginning, SLV on server is
48  * set max value which is calculated as the following: allow to one client
49  * have all locks of limit ->pl_limit for 10h.
50  *
51  * Next, on clients, number of cached locks is not limited artificially in any
52  * way as it was before. Instead, client calculates CLV, that is, client lock
53  * volume for each lock and compares it with last SLV from the server. CLV is
54  * calculated as the number of locks in LRU * lock live time in seconds. If
55  * CLV > SLV - lock is canceled.
56  *
57  * Client has LVF, that is, lock volume factor which regulates how much
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.
61  *
62  * Locks on a client will be canceled more intensively in these cases:
63  * (1) if SLV is smaller, that is, load is higher on the server;
64  * (2) client has a lot of locks (the more locks are held by client, the bigger
65  *     chances that some of them should be canceled);
66  * (3) client has old locks (taken some time ago);
67  *
68  * Thus, according to flow paradigm that we use for better understanding SLV,
69  * CLV is the volume of particle in flow described by SLV. According to this,
70  * if flow is getting thinner, more and more particles become outside of it and
71  * as particles are locks, they should be canceled.
72  *
73  * General idea of this belongs to Vitaly Fertman (vitaly@clusterfs.com).
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).
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
93  * affect behavior much. They all may be modified via proc. However, they also
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
96  * be "constructed" via proc tunables.
97  */
98 
99 #define DEBUG_SUBSYSTEM S_LDLM
100 
101 #include "../include/lustre_dlm.h"
102 #include "../include/cl_object.h"
103 #include "../include/obd_class.h"
104 #include "../include/obd_support.h"
105 #include "ldlm_internal.h"
106 
107 
108 /*
109  * 50 ldlm locks for 1MB of RAM.
110  */
111 #define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_CACHE_SHIFT)) * 50)
112 
113 /*
114  * Maximal possible grant step plan in %.
115  */
116 #define LDLM_POOL_MAX_GSP (30)
117 
118 /*
119  * Minimal possible grant step plan in %.
120  */
121 #define LDLM_POOL_MIN_GSP (1)
122 
123 /*
124  * This controls the speed of reaching LDLM_POOL_MAX_GSP
125  * with increasing thread period.
126  */
127 #define LDLM_POOL_GSP_STEP_SHIFT (2)
128 
129 /*
130  * LDLM_POOL_GSP% of all locks is default GP.
131  */
132 #define LDLM_POOL_GP(L)   (((L) * LDLM_POOL_MAX_GSP) / 100)
133 
134 /*
135  * Max age for locks on clients.
136  */
137 #define LDLM_POOL_MAX_AGE (36000)
138 
139 /*
140  * The granularity of SLV calculation.
141  */
142 #define LDLM_POOL_SLV_SHIFT (10)
143 
dru(__u64 val,__u32 shift,int round_up)144 static inline __u64 dru(__u64 val, __u32 shift, int round_up)
145 {
146 	return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift;
147 }
148 
ldlm_pool_slv_max(__u32 L)149 static inline __u64 ldlm_pool_slv_max(__u32 L)
150 {
151 	/*
152 	 * Allow to have all locks for 1 client for 10 hrs.
153 	 * Formula is the following: limit * 10h / 1 client.
154 	 */
155 	__u64 lim = (__u64)L *  LDLM_POOL_MAX_AGE / 1;
156 	return lim;
157 }
158 
ldlm_pool_slv_min(__u32 L)159 static inline __u64 ldlm_pool_slv_min(__u32 L)
160 {
161 	return 1;
162 }
163 
164 enum {
165 	LDLM_POOL_FIRST_STAT = 0,
166 	LDLM_POOL_GRANTED_STAT = LDLM_POOL_FIRST_STAT,
167 	LDLM_POOL_GRANT_STAT,
168 	LDLM_POOL_CANCEL_STAT,
169 	LDLM_POOL_GRANT_RATE_STAT,
170 	LDLM_POOL_CANCEL_RATE_STAT,
171 	LDLM_POOL_GRANT_PLAN_STAT,
172 	LDLM_POOL_SLV_STAT,
173 	LDLM_POOL_SHRINK_REQTD_STAT,
174 	LDLM_POOL_SHRINK_FREED_STAT,
175 	LDLM_POOL_RECALC_STAT,
176 	LDLM_POOL_TIMING_STAT,
177 	LDLM_POOL_LAST_STAT
178 };
179 
ldlm_pl2ns(struct ldlm_pool * pl)180 static inline struct ldlm_namespace *ldlm_pl2ns(struct ldlm_pool *pl)
181 {
182 	return container_of(pl, struct ldlm_namespace, ns_pool);
183 }
184 
185 /**
186  * Calculates suggested grant_step in % of available locks for passed
187  * \a period. This is later used in grant_plan calculations.
188  */
ldlm_pool_t2gsp(unsigned int t)189 static inline int ldlm_pool_t2gsp(unsigned int t)
190 {
191 	/*
192 	 * This yields 1% grant step for anything below LDLM_POOL_GSP_STEP
193 	 * and up to 30% for anything higher than LDLM_POOL_GSP_STEP.
194 	 *
195 	 * How this will affect execution is the following:
196 	 *
197 	 * - for thread period 1s we will have grant_step 1% which good from
198 	 * pov of taking some load off from server and push it out to clients.
199 	 * This is like that because 1% for grant_step means that server will
200 	 * not allow clients to get lots of locks in short period of time and
201 	 * keep all old locks in their caches. Clients will always have to
202 	 * get some locks back if they want to take some new;
203 	 *
204 	 * - for thread period 10s (which is default) we will have 23% which
205 	 * means that clients will have enough of room to take some new locks
206 	 * without getting some back. All locks from this 23% which were not
207 	 * taken by clients in current period will contribute in SLV growing.
208 	 * SLV growing means more locks cached on clients until limit or grant
209 	 * plan is reached.
210 	 */
211 	return LDLM_POOL_MAX_GSP -
212 		((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >>
213 		 (t >> LDLM_POOL_GSP_STEP_SHIFT));
214 }
215 
216 /**
217  * Recalculates next grant limit on passed \a pl.
218  *
219  * \pre ->pl_lock is locked.
220  */
ldlm_pool_recalc_grant_plan(struct ldlm_pool * pl)221 static void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl)
222 {
223 	int granted, grant_step, limit;
224 
225 	limit = ldlm_pool_get_limit(pl);
226 	granted = atomic_read(&pl->pl_granted);
227 
228 	grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
229 	grant_step = ((limit - granted) * grant_step) / 100;
230 	pl->pl_grant_plan = granted + grant_step;
231 	limit = (limit * 5) >> 2;
232 	if (pl->pl_grant_plan > limit)
233 		pl->pl_grant_plan = limit;
234 }
235 
236 /**
237  * Recalculates next SLV on passed \a pl.
238  *
239  * \pre ->pl_lock is locked.
240  */
ldlm_pool_recalc_slv(struct ldlm_pool * pl)241 static void ldlm_pool_recalc_slv(struct ldlm_pool *pl)
242 {
243 	int granted;
244 	int grant_plan;
245 	int round_up;
246 	__u64 slv;
247 	__u64 slv_factor;
248 	__u64 grant_usage;
249 	__u32 limit;
250 
251 	slv = pl->pl_server_lock_volume;
252 	grant_plan = pl->pl_grant_plan;
253 	limit = ldlm_pool_get_limit(pl);
254 	granted = atomic_read(&pl->pl_granted);
255 	round_up = granted < limit;
256 
257 	grant_usage = max_t(int, limit - (granted - grant_plan), 1);
258 
259 	/*
260 	 * Find out SLV change factor which is the ratio of grant usage
261 	 * from limit. SLV changes as fast as the ratio of grant plan
262 	 * consumption. The more locks from grant plan are not consumed
263 	 * by clients in last interval (idle time), the faster grows
264 	 * SLV. And the opposite, the more grant plan is over-consumed
265 	 * (load time) the faster drops SLV.
266 	 */
267 	slv_factor = grant_usage << LDLM_POOL_SLV_SHIFT;
268 	do_div(slv_factor, limit);
269 	slv = slv * slv_factor;
270 	slv = dru(slv, LDLM_POOL_SLV_SHIFT, round_up);
271 
272 	if (slv > ldlm_pool_slv_max(limit))
273 		slv = ldlm_pool_slv_max(limit);
274 	else if (slv < ldlm_pool_slv_min(limit))
275 		slv = ldlm_pool_slv_min(limit);
276 
277 	pl->pl_server_lock_volume = slv;
278 }
279 
280 /**
281  * Recalculates next stats on passed \a pl.
282  *
283  * \pre ->pl_lock is locked.
284  */
ldlm_pool_recalc_stats(struct ldlm_pool * pl)285 static void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
286 {
287 	int grant_plan = pl->pl_grant_plan;
288 	__u64 slv = pl->pl_server_lock_volume;
289 	int granted = atomic_read(&pl->pl_granted);
290 	int grant_rate = atomic_read(&pl->pl_grant_rate);
291 	int cancel_rate = atomic_read(&pl->pl_cancel_rate);
292 
293 	lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT,
294 			    slv);
295 	lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
296 			    granted);
297 	lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
298 			    grant_rate);
299 	lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
300 			    grant_plan);
301 	lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
302 			    cancel_rate);
303 }
304 
305 /**
306  * Sets current SLV into obd accessible via ldlm_pl2ns(pl)->ns_obd.
307  */
ldlm_srv_pool_push_slv(struct ldlm_pool * pl)308 static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl)
309 {
310 	struct obd_device *obd;
311 
312 	/*
313 	 * Set new SLV in obd field for using it later without accessing the
314 	 * pool. This is required to avoid race between sending reply to client
315 	 * with new SLV and cleanup server stack in which we can't guarantee
316 	 * that namespace is still alive. We know only that obd is alive as
317 	 * long as valid export is alive.
318 	 */
319 	obd = ldlm_pl2ns(pl)->ns_obd;
320 	LASSERT(obd != NULL);
321 	write_lock(&obd->obd_pool_lock);
322 	obd->obd_pool_slv = pl->pl_server_lock_volume;
323 	write_unlock(&obd->obd_pool_lock);
324 }
325 
326 /**
327  * Recalculates all pool fields on passed \a pl.
328  *
329  * \pre ->pl_lock is not locked.
330  */
ldlm_srv_pool_recalc(struct ldlm_pool * pl)331 static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
332 {
333 	time_t recalc_interval_sec;
334 
335 	recalc_interval_sec = get_seconds() - pl->pl_recalc_time;
336 	if (recalc_interval_sec < pl->pl_recalc_period)
337 		return 0;
338 
339 	spin_lock(&pl->pl_lock);
340 	recalc_interval_sec = get_seconds() - pl->pl_recalc_time;
341 	if (recalc_interval_sec < pl->pl_recalc_period) {
342 		spin_unlock(&pl->pl_lock);
343 		return 0;
344 	}
345 	/*
346 	 * Recalc SLV after last period. This should be done
347 	 * _before_ recalculating new grant plan.
348 	 */
349 	ldlm_pool_recalc_slv(pl);
350 
351 	/*
352 	 * Make sure that pool informed obd of last SLV changes.
353 	 */
354 	ldlm_srv_pool_push_slv(pl);
355 
356 	/*
357 	 * Update grant_plan for new period.
358 	 */
359 	ldlm_pool_recalc_grant_plan(pl);
360 
361 	pl->pl_recalc_time = get_seconds();
362 	lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
363 			    recalc_interval_sec);
364 	spin_unlock(&pl->pl_lock);
365 	return 0;
366 }
367 
368 /**
369  * This function is used on server side as main entry point for memory
370  * pressure handling. It decreases SLV on \a pl according to passed
371  * \a nr and \a gfp_mask.
372  *
373  * Our goal here is to decrease SLV such a way that clients hold \a nr
374  * locks smaller in next 10h.
375  */
ldlm_srv_pool_shrink(struct ldlm_pool * pl,int nr,gfp_t gfp_mask)376 static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
377 				int nr, gfp_t gfp_mask)
378 {
379 	__u32 limit;
380 
381 	/*
382 	 * VM is asking how many entries may be potentially freed.
383 	 */
384 	if (nr == 0)
385 		return atomic_read(&pl->pl_granted);
386 
387 	/*
388 	 * Client already canceled locks but server is already in shrinker
389 	 * and can't cancel anything. Let's catch this race.
390 	 */
391 	if (atomic_read(&pl->pl_granted) == 0)
392 		return 0;
393 
394 	spin_lock(&pl->pl_lock);
395 
396 	/*
397 	 * We want shrinker to possibly cause cancellation of @nr locks from
398 	 * clients or grant approximately @nr locks smaller next intervals.
399 	 *
400 	 * This is why we decreased SLV by @nr. This effect will only be as
401 	 * long as one re-calc interval (1s these days) and this should be
402 	 * enough to pass this decreased SLV to all clients. On next recalc
403 	 * interval pool will either increase SLV if locks load is not high
404 	 * or will keep on same level or even decrease again, thus, shrinker
405 	 * decreased SLV will affect next recalc intervals and this way will
406 	 * make locking load lower.
407 	 */
408 	if (nr < pl->pl_server_lock_volume) {
409 		pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr;
410 	} else {
411 		limit = ldlm_pool_get_limit(pl);
412 		pl->pl_server_lock_volume = ldlm_pool_slv_min(limit);
413 	}
414 
415 	/*
416 	 * Make sure that pool informed obd of last SLV changes.
417 	 */
418 	ldlm_srv_pool_push_slv(pl);
419 	spin_unlock(&pl->pl_lock);
420 
421 	/*
422 	 * We did not really free any memory here so far, it only will be
423 	 * freed later may be, so that we return 0 to not confuse VM.
424 	 */
425 	return 0;
426 }
427 
428 /**
429  * Setup server side pool \a pl with passed \a limit.
430  */
ldlm_srv_pool_setup(struct ldlm_pool * pl,int limit)431 static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit)
432 {
433 	struct obd_device *obd;
434 
435 	obd = ldlm_pl2ns(pl)->ns_obd;
436 	LASSERT(obd != NULL && obd != LP_POISON);
437 	LASSERT(obd->obd_type != LP_POISON);
438 	write_lock(&obd->obd_pool_lock);
439 	obd->obd_pool_limit = limit;
440 	write_unlock(&obd->obd_pool_lock);
441 
442 	ldlm_pool_set_limit(pl, limit);
443 	return 0;
444 }
445 
446 /**
447  * Sets SLV and Limit from ldlm_pl2ns(pl)->ns_obd tp passed \a pl.
448  */
ldlm_cli_pool_pop_slv(struct ldlm_pool * pl)449 static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
450 {
451 	struct obd_device *obd;
452 
453 	/*
454 	 * Get new SLV and Limit from obd which is updated with coming
455 	 * RPCs.
456 	 */
457 	obd = ldlm_pl2ns(pl)->ns_obd;
458 	LASSERT(obd != NULL);
459 	read_lock(&obd->obd_pool_lock);
460 	pl->pl_server_lock_volume = obd->obd_pool_slv;
461 	ldlm_pool_set_limit(pl, obd->obd_pool_limit);
462 	read_unlock(&obd->obd_pool_lock);
463 }
464 
465 /**
466  * Recalculates client size pool \a pl according to current SLV and Limit.
467  */
ldlm_cli_pool_recalc(struct ldlm_pool * pl)468 static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
469 {
470 	time_t recalc_interval_sec;
471 	int ret;
472 
473 	recalc_interval_sec = get_seconds() - pl->pl_recalc_time;
474 	if (recalc_interval_sec < pl->pl_recalc_period)
475 		return 0;
476 
477 	spin_lock(&pl->pl_lock);
478 	/*
479 	 * Check if we need to recalc lists now.
480 	 */
481 	recalc_interval_sec = get_seconds() - pl->pl_recalc_time;
482 	if (recalc_interval_sec < pl->pl_recalc_period) {
483 		spin_unlock(&pl->pl_lock);
484 		return 0;
485 	}
486 
487 	/*
488 	 * Make sure that pool knows last SLV and Limit from obd.
489 	 */
490 	ldlm_cli_pool_pop_slv(pl);
491 
492 	spin_unlock(&pl->pl_lock);
493 
494 	/*
495 	 * Do not cancel locks in case lru resize is disabled for this ns.
496 	 */
497 	if (!ns_connect_lru_resize(ldlm_pl2ns(pl))) {
498 		ret = 0;
499 		goto out;
500 	}
501 
502 	/*
503 	 * In the time of canceling locks on client we do not need to maintain
504 	 * sharp timing, we only want to cancel locks asap according to new SLV.
505 	 * It may be called when SLV has changed much, this is why we do not
506 	 * take into account pl->pl_recalc_time here.
507 	 */
508 	ret = ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC, LDLM_CANCEL_LRUR);
509 
510 out:
511 	spin_lock(&pl->pl_lock);
512 	/*
513 	 * Time of LRU resizing might be longer than period,
514 	 * so update after LRU resizing rather than before it.
515 	 */
516 	pl->pl_recalc_time = get_seconds();
517 	lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
518 			    recalc_interval_sec);
519 	spin_unlock(&pl->pl_lock);
520 	return ret;
521 }
522 
523 /**
524  * This function is main entry point for memory pressure handling on client
525  * side.  Main goal of this function is to cancel some number of locks on
526  * passed \a pl according to \a nr and \a gfp_mask.
527  */
ldlm_cli_pool_shrink(struct ldlm_pool * pl,int nr,gfp_t gfp_mask)528 static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
529 				int nr, gfp_t gfp_mask)
530 {
531 	struct ldlm_namespace *ns;
532 	int unused;
533 
534 	ns = ldlm_pl2ns(pl);
535 
536 	/*
537 	 * Do not cancel locks in case lru resize is disabled for this ns.
538 	 */
539 	if (!ns_connect_lru_resize(ns))
540 		return 0;
541 
542 	/*
543 	 * Make sure that pool knows last SLV and Limit from obd.
544 	 */
545 	ldlm_cli_pool_pop_slv(pl);
546 
547 	spin_lock(&ns->ns_lock);
548 	unused = ns->ns_nr_unused;
549 	spin_unlock(&ns->ns_lock);
550 
551 	if (nr == 0)
552 		return (unused / 100) * sysctl_vfs_cache_pressure;
553 	else
554 		return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_CANCEL_SHRINK);
555 }
556 
557 static const struct ldlm_pool_ops ldlm_srv_pool_ops = {
558 	.po_recalc = ldlm_srv_pool_recalc,
559 	.po_shrink = ldlm_srv_pool_shrink,
560 	.po_setup  = ldlm_srv_pool_setup
561 };
562 
563 static const struct ldlm_pool_ops ldlm_cli_pool_ops = {
564 	.po_recalc = ldlm_cli_pool_recalc,
565 	.po_shrink = ldlm_cli_pool_shrink
566 };
567 
568 /**
569  * Pool recalc wrapper. Will call either client or server pool recalc callback
570  * depending what pool \a pl is used.
571  */
ldlm_pool_recalc(struct ldlm_pool * pl)572 int ldlm_pool_recalc(struct ldlm_pool *pl)
573 {
574 	time_t recalc_interval_sec;
575 	int count;
576 
577 	recalc_interval_sec = get_seconds() - pl->pl_recalc_time;
578 	if (recalc_interval_sec <= 0)
579 		goto recalc;
580 
581 	spin_lock(&pl->pl_lock);
582 	if (recalc_interval_sec > 0) {
583 		/*
584 		 * Update pool statistics every 1s.
585 		 */
586 		ldlm_pool_recalc_stats(pl);
587 
588 		/*
589 		 * Zero out all rates and speed for the last period.
590 		 */
591 		atomic_set(&pl->pl_grant_rate, 0);
592 		atomic_set(&pl->pl_cancel_rate, 0);
593 	}
594 	spin_unlock(&pl->pl_lock);
595 
596  recalc:
597 	if (pl->pl_ops->po_recalc != NULL) {
598 		count = pl->pl_ops->po_recalc(pl);
599 		lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
600 				    count);
601 	}
602 	recalc_interval_sec = pl->pl_recalc_time - get_seconds() +
603 			      pl->pl_recalc_period;
604 	if (recalc_interval_sec <= 0) {
605 		/* Prevent too frequent recalculation. */
606 		CDEBUG(D_DLMTRACE, "Negative interval(%ld), "
607 		       "too short period(%ld)",
608 		       recalc_interval_sec,
609 		       pl->pl_recalc_period);
610 		recalc_interval_sec = 1;
611 	}
612 
613 	return recalc_interval_sec;
614 }
615 
616 /*
617  * Pool shrink wrapper. Will call either client or server pool recalc callback
618  * depending what pool pl is used. When nr == 0, just return the number of
619  * freeable locks. Otherwise, return the number of canceled locks.
620  */
ldlm_pool_shrink(struct ldlm_pool * pl,int nr,gfp_t gfp_mask)621 int ldlm_pool_shrink(struct ldlm_pool *pl, int nr,
622 		     gfp_t gfp_mask)
623 {
624 	int cancel = 0;
625 
626 	if (pl->pl_ops->po_shrink != NULL) {
627 		cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask);
628 		if (nr > 0) {
629 			lprocfs_counter_add(pl->pl_stats,
630 					    LDLM_POOL_SHRINK_REQTD_STAT,
631 					    nr);
632 			lprocfs_counter_add(pl->pl_stats,
633 					    LDLM_POOL_SHRINK_FREED_STAT,
634 					    cancel);
635 			CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, shrunk %d\n",
636 			       pl->pl_name, nr, cancel);
637 		}
638 	}
639 	return cancel;
640 }
641 EXPORT_SYMBOL(ldlm_pool_shrink);
642 
643 /**
644  * Pool setup wrapper. Will call either client or server pool recalc callback
645  * depending what pool \a pl is used.
646  *
647  * Sets passed \a limit into pool \a pl.
648  */
ldlm_pool_setup(struct ldlm_pool * pl,int limit)649 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
650 {
651 	if (pl->pl_ops->po_setup != NULL)
652 		return pl->pl_ops->po_setup(pl, limit);
653 	return 0;
654 }
655 EXPORT_SYMBOL(ldlm_pool_setup);
656 
657 #if defined(CONFIG_PROC_FS)
lprocfs_pool_state_seq_show(struct seq_file * m,void * unused)658 static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
659 {
660 	int granted, grant_rate, cancel_rate, grant_step;
661 	int grant_speed, grant_plan, lvf;
662 	struct ldlm_pool *pl = m->private;
663 	__u64 slv, clv;
664 	__u32 limit;
665 
666 	spin_lock(&pl->pl_lock);
667 	slv = pl->pl_server_lock_volume;
668 	clv = pl->pl_client_lock_volume;
669 	limit = ldlm_pool_get_limit(pl);
670 	grant_plan = pl->pl_grant_plan;
671 	granted = atomic_read(&pl->pl_granted);
672 	grant_rate = atomic_read(&pl->pl_grant_rate);
673 	cancel_rate = atomic_read(&pl->pl_cancel_rate);
674 	grant_speed = grant_rate - cancel_rate;
675 	lvf = atomic_read(&pl->pl_lock_volume_factor);
676 	grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
677 	spin_unlock(&pl->pl_lock);
678 
679 	seq_printf(m, "LDLM pool state (%s):\n"
680 		      "  SLV: %llu\n"
681 		      "  CLV: %llu\n"
682 		      "  LVF: %d\n",
683 		      pl->pl_name, slv, clv, lvf);
684 
685 	if (ns_is_server(ldlm_pl2ns(pl))) {
686 		seq_printf(m, "  GSP: %d%%\n"
687 			      "  GP:  %d\n",
688 			      grant_step, grant_plan);
689 	}
690 	seq_printf(m, "  GR:  %d\n  CR:  %d\n  GS:  %d\n"
691 		      "  G:   %d\n  L:   %d\n",
692 		      grant_rate, cancel_rate, grant_speed,
693 		      granted, limit);
694 
695 	return 0;
696 }
697 LPROC_SEQ_FOPS_RO(lprocfs_pool_state);
698 
lprocfs_grant_speed_seq_show(struct seq_file * m,void * unused)699 static int lprocfs_grant_speed_seq_show(struct seq_file *m, void *unused)
700 {
701 	struct ldlm_pool *pl = m->private;
702 	int	       grant_speed;
703 
704 	spin_lock(&pl->pl_lock);
705 	/* serialize with ldlm_pool_recalc */
706 	grant_speed = atomic_read(&pl->pl_grant_rate) -
707 			atomic_read(&pl->pl_cancel_rate);
708 	spin_unlock(&pl->pl_lock);
709 	return lprocfs_rd_uint(m, &grant_speed);
710 }
711 
712 LDLM_POOL_PROC_READER_SEQ_SHOW(grant_plan, int);
713 LPROC_SEQ_FOPS_RO(lprocfs_grant_plan);
714 
715 LDLM_POOL_PROC_READER_SEQ_SHOW(recalc_period, int);
716 LDLM_POOL_PROC_WRITER(recalc_period, int);
lprocfs_recalc_period_seq_write(struct file * file,const char __user * buf,size_t len,loff_t * off)717 static ssize_t lprocfs_recalc_period_seq_write(struct file *file,
718 					       const char __user *buf,
719 					       size_t len, loff_t *off)
720 {
721 	struct seq_file *seq = file->private_data;
722 
723 	return lprocfs_wr_recalc_period(file, buf, len, seq->private);
724 }
725 LPROC_SEQ_FOPS(lprocfs_recalc_period);
726 
727 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, u64);
728 LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, atomic);
729 LPROC_SEQ_FOPS_RW_TYPE(ldlm_pool_rw, atomic);
730 
731 LPROC_SEQ_FOPS_RO(lprocfs_grant_speed);
732 
733 #define LDLM_POOL_ADD_VAR(name, var, ops)			\
734 	do {							\
735 		snprintf(var_name, MAX_STRING_SIZE, #name);	\
736 		pool_vars[0].data = var;			\
737 		pool_vars[0].fops = ops;			\
738 		lprocfs_add_vars(pl->pl_proc_dir, pool_vars, NULL);\
739 	} while (0)
740 
ldlm_pool_proc_init(struct ldlm_pool * pl)741 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
742 {
743 	struct ldlm_namespace *ns = ldlm_pl2ns(pl);
744 	struct proc_dir_entry *parent_ns_proc;
745 	struct lprocfs_vars pool_vars[2];
746 	char *var_name = NULL;
747 	int rc = 0;
748 
749 	OBD_ALLOC(var_name, MAX_STRING_SIZE + 1);
750 	if (!var_name)
751 		return -ENOMEM;
752 
753 	parent_ns_proc = ns->ns_proc_dir_entry;
754 	if (parent_ns_proc == NULL) {
755 		CERROR("%s: proc entry is not initialized\n",
756 		       ldlm_ns_name(ns));
757 		rc = -EINVAL;
758 		goto out_free_name;
759 	}
760 	pl->pl_proc_dir = lprocfs_register("pool", parent_ns_proc,
761 					   NULL, NULL);
762 	if (IS_ERR(pl->pl_proc_dir)) {
763 		CERROR("LProcFS failed in ldlm-pool-init\n");
764 		rc = PTR_ERR(pl->pl_proc_dir);
765 		pl->pl_proc_dir = NULL;
766 		goto out_free_name;
767 	}
768 
769 	var_name[MAX_STRING_SIZE] = '\0';
770 	memset(pool_vars, 0, sizeof(pool_vars));
771 	pool_vars[0].name = var_name;
772 
773 	LDLM_POOL_ADD_VAR("server_lock_volume", &pl->pl_server_lock_volume,
774 			  &ldlm_pool_u64_fops);
775 	LDLM_POOL_ADD_VAR("limit", &pl->pl_limit, &ldlm_pool_rw_atomic_fops);
776 	LDLM_POOL_ADD_VAR("granted", &pl->pl_granted, &ldlm_pool_atomic_fops);
777 	LDLM_POOL_ADD_VAR("grant_speed", pl, &lprocfs_grant_speed_fops);
778 	LDLM_POOL_ADD_VAR("cancel_rate", &pl->pl_cancel_rate,
779 			  &ldlm_pool_atomic_fops);
780 	LDLM_POOL_ADD_VAR("grant_rate", &pl->pl_grant_rate,
781 			  &ldlm_pool_atomic_fops);
782 	LDLM_POOL_ADD_VAR("grant_plan", pl, &lprocfs_grant_plan_fops);
783 	LDLM_POOL_ADD_VAR("recalc_period", pl, &lprocfs_recalc_period_fops);
784 	LDLM_POOL_ADD_VAR("lock_volume_factor", &pl->pl_lock_volume_factor,
785 			  &ldlm_pool_rw_atomic_fops);
786 	LDLM_POOL_ADD_VAR("state", pl, &lprocfs_pool_state_fops);
787 
788 	pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT -
789 					   LDLM_POOL_FIRST_STAT, 0);
790 	if (!pl->pl_stats) {
791 		rc = -ENOMEM;
792 		goto out_free_name;
793 	}
794 
795 	lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
796 			     LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
797 			     "granted", "locks");
798 	lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
799 			     LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
800 			     "grant", "locks");
801 	lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
802 			     LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
803 			     "cancel", "locks");
804 	lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
805 			     LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
806 			     "grant_rate", "locks/s");
807 	lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
808 			     LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
809 			     "cancel_rate", "locks/s");
810 	lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
811 			     LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
812 			     "grant_plan", "locks/s");
813 	lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT,
814 			     LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
815 			     "slv", "slv");
816 	lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT,
817 			     LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
818 			     "shrink_request", "locks");
819 	lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT,
820 			     LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
821 			     "shrink_freed", "locks");
822 	lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
823 			     LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
824 			     "recalc_freed", "locks");
825 	lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
826 			     LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
827 			     "recalc_timing", "sec");
828 	rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
829 
830 out_free_name:
831 	OBD_FREE(var_name, MAX_STRING_SIZE + 1);
832 	return rc;
833 }
834 
ldlm_pool_proc_fini(struct ldlm_pool * pl)835 static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
836 {
837 	if (pl->pl_stats != NULL) {
838 		lprocfs_free_stats(&pl->pl_stats);
839 		pl->pl_stats = NULL;
840 	}
841 	if (pl->pl_proc_dir != NULL) {
842 		lprocfs_remove(&pl->pl_proc_dir);
843 		pl->pl_proc_dir = NULL;
844 	}
845 }
846 #else /* !CONFIG_PROC_FS */
ldlm_pool_proc_init(struct ldlm_pool * pl)847 static int ldlm_pool_proc_init(struct ldlm_pool *pl)
848 {
849 	return 0;
850 }
851 
ldlm_pool_proc_fini(struct ldlm_pool * pl)852 static void ldlm_pool_proc_fini(struct ldlm_pool *pl) {}
853 #endif /* CONFIG_PROC_FS */
854 
ldlm_pool_init(struct ldlm_pool * pl,struct ldlm_namespace * ns,int idx,ldlm_side_t client)855 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
856 		   int idx, ldlm_side_t client)
857 {
858 	int rc;
859 
860 	spin_lock_init(&pl->pl_lock);
861 	atomic_set(&pl->pl_granted, 0);
862 	pl->pl_recalc_time = get_seconds();
863 	atomic_set(&pl->pl_lock_volume_factor, 1);
864 
865 	atomic_set(&pl->pl_grant_rate, 0);
866 	atomic_set(&pl->pl_cancel_rate, 0);
867 	pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
868 
869 	snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
870 		 ldlm_ns_name(ns), idx);
871 
872 	if (client == LDLM_NAMESPACE_SERVER) {
873 		pl->pl_ops = &ldlm_srv_pool_ops;
874 		ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L);
875 		pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD;
876 		pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L);
877 	} else {
878 		ldlm_pool_set_limit(pl, 1);
879 		pl->pl_server_lock_volume = 0;
880 		pl->pl_ops = &ldlm_cli_pool_ops;
881 		pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
882 	}
883 	pl->pl_client_lock_volume = 0;
884 	rc = ldlm_pool_proc_init(pl);
885 	if (rc)
886 		return rc;
887 
888 	CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name);
889 
890 	return rc;
891 }
892 EXPORT_SYMBOL(ldlm_pool_init);
893 
ldlm_pool_fini(struct ldlm_pool * pl)894 void ldlm_pool_fini(struct ldlm_pool *pl)
895 {
896 	ldlm_pool_proc_fini(pl);
897 
898 	/*
899 	 * Pool should not be used after this point. We can't free it here as
900 	 * it lives in struct ldlm_namespace, but still interested in catching
901 	 * any abnormal using cases.
902 	 */
903 	POISON(pl, 0x5a, sizeof(*pl));
904 }
905 EXPORT_SYMBOL(ldlm_pool_fini);
906 
907 /**
908  * Add new taken ldlm lock \a lock into pool \a pl accounting.
909  */
ldlm_pool_add(struct ldlm_pool * pl,struct ldlm_lock * lock)910 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
911 {
912 	/*
913 	 * FLOCK locks are special in a sense that they are almost never
914 	 * cancelled, instead special kind of lock is used to drop them.
915 	 * also there is no LRU for flock locks, so no point in tracking
916 	 * them anyway.
917 	 */
918 	if (lock->l_resource->lr_type == LDLM_FLOCK)
919 		return;
920 
921 	atomic_inc(&pl->pl_granted);
922 	atomic_inc(&pl->pl_grant_rate);
923 	lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
924 	/*
925 	 * Do not do pool recalc for client side as all locks which
926 	 * potentially may be canceled has already been packed into
927 	 * enqueue/cancel rpc. Also we do not want to run out of stack
928 	 * with too long call paths.
929 	 */
930 	if (ns_is_server(ldlm_pl2ns(pl)))
931 		ldlm_pool_recalc(pl);
932 }
933 EXPORT_SYMBOL(ldlm_pool_add);
934 
935 /**
936  * Remove ldlm lock \a lock from pool \a pl accounting.
937  */
ldlm_pool_del(struct ldlm_pool * pl,struct ldlm_lock * lock)938 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
939 {
940 	/*
941 	 * Filter out FLOCK locks. Read above comment in ldlm_pool_add().
942 	 */
943 	if (lock->l_resource->lr_type == LDLM_FLOCK)
944 		return;
945 
946 	LASSERT(atomic_read(&pl->pl_granted) > 0);
947 	atomic_dec(&pl->pl_granted);
948 	atomic_inc(&pl->pl_cancel_rate);
949 
950 	lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
951 
952 	if (ns_is_server(ldlm_pl2ns(pl)))
953 		ldlm_pool_recalc(pl);
954 }
955 EXPORT_SYMBOL(ldlm_pool_del);
956 
957 /**
958  * Returns current \a pl SLV.
959  *
960  * \pre ->pl_lock is not locked.
961  */
ldlm_pool_get_slv(struct ldlm_pool * pl)962 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
963 {
964 	__u64 slv;
965 
966 	spin_lock(&pl->pl_lock);
967 	slv = pl->pl_server_lock_volume;
968 	spin_unlock(&pl->pl_lock);
969 	return slv;
970 }
971 EXPORT_SYMBOL(ldlm_pool_get_slv);
972 
973 /**
974  * Sets passed \a slv to \a pl.
975  *
976  * \pre ->pl_lock is not locked.
977  */
ldlm_pool_set_slv(struct ldlm_pool * pl,__u64 slv)978 void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
979 {
980 	spin_lock(&pl->pl_lock);
981 	pl->pl_server_lock_volume = slv;
982 	spin_unlock(&pl->pl_lock);
983 }
984 EXPORT_SYMBOL(ldlm_pool_set_slv);
985 
986 /**
987  * Returns current \a pl CLV.
988  *
989  * \pre ->pl_lock is not locked.
990  */
ldlm_pool_get_clv(struct ldlm_pool * pl)991 __u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
992 {
993 	__u64 slv;
994 
995 	spin_lock(&pl->pl_lock);
996 	slv = pl->pl_client_lock_volume;
997 	spin_unlock(&pl->pl_lock);
998 	return slv;
999 }
1000 EXPORT_SYMBOL(ldlm_pool_get_clv);
1001 
1002 /**
1003  * Sets passed \a clv to \a pl.
1004  *
1005  * \pre ->pl_lock is not locked.
1006  */
ldlm_pool_set_clv(struct ldlm_pool * pl,__u64 clv)1007 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
1008 {
1009 	spin_lock(&pl->pl_lock);
1010 	pl->pl_client_lock_volume = clv;
1011 	spin_unlock(&pl->pl_lock);
1012 }
1013 EXPORT_SYMBOL(ldlm_pool_set_clv);
1014 
1015 /**
1016  * Returns current \a pl limit.
1017  */
ldlm_pool_get_limit(struct ldlm_pool * pl)1018 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
1019 {
1020 	return atomic_read(&pl->pl_limit);
1021 }
1022 EXPORT_SYMBOL(ldlm_pool_get_limit);
1023 
1024 /**
1025  * Sets passed \a limit to \a pl.
1026  */
ldlm_pool_set_limit(struct ldlm_pool * pl,__u32 limit)1027 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1028 {
1029 	atomic_set(&pl->pl_limit, limit);
1030 }
1031 EXPORT_SYMBOL(ldlm_pool_set_limit);
1032 
1033 /**
1034  * Returns current LVF from \a pl.
1035  */
ldlm_pool_get_lvf(struct ldlm_pool * pl)1036 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1037 {
1038 	return atomic_read(&pl->pl_lock_volume_factor);
1039 }
1040 EXPORT_SYMBOL(ldlm_pool_get_lvf);
1041 
ldlm_pool_granted(struct ldlm_pool * pl)1042 static int ldlm_pool_granted(struct ldlm_pool *pl)
1043 {
1044 	return atomic_read(&pl->pl_granted);
1045 }
1046 
1047 static struct ptlrpc_thread *ldlm_pools_thread;
1048 static struct completion ldlm_pools_comp;
1049 
1050 /*
1051  * count locks from all namespaces (if possible). Returns number of
1052  * cached locks.
1053  */
ldlm_pools_count(ldlm_side_t client,gfp_t gfp_mask)1054 static unsigned long ldlm_pools_count(ldlm_side_t client, gfp_t gfp_mask)
1055 {
1056 	int total = 0, nr_ns;
1057 	struct ldlm_namespace *ns;
1058 	struct ldlm_namespace *ns_old = NULL; /* loop detection */
1059 	void *cookie;
1060 
1061 	if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1062 		return 0;
1063 
1064 	CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n",
1065 	       client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1066 
1067 	cookie = cl_env_reenter();
1068 
1069 	/*
1070 	 * Find out how many resources we may release.
1071 	 */
1072 	for (nr_ns = ldlm_namespace_nr_read(client);
1073 	     nr_ns > 0; nr_ns--) {
1074 		mutex_lock(ldlm_namespace_lock(client));
1075 		if (list_empty(ldlm_namespace_list(client))) {
1076 			mutex_unlock(ldlm_namespace_lock(client));
1077 			cl_env_reexit(cookie);
1078 			return 0;
1079 		}
1080 		ns = ldlm_namespace_first_locked(client);
1081 
1082 		if (ns == ns_old) {
1083 			mutex_unlock(ldlm_namespace_lock(client));
1084 			break;
1085 		}
1086 
1087 		if (ldlm_ns_empty(ns)) {
1088 			ldlm_namespace_move_to_inactive_locked(ns, client);
1089 			mutex_unlock(ldlm_namespace_lock(client));
1090 			continue;
1091 		}
1092 
1093 		if (ns_old == NULL)
1094 			ns_old = ns;
1095 
1096 		ldlm_namespace_get(ns);
1097 		ldlm_namespace_move_to_active_locked(ns, client);
1098 		mutex_unlock(ldlm_namespace_lock(client));
1099 		total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
1100 		ldlm_namespace_put(ns);
1101 	}
1102 
1103 	cl_env_reexit(cookie);
1104 	return total;
1105 }
1106 
ldlm_pools_scan(ldlm_side_t client,int nr,gfp_t gfp_mask)1107 static unsigned long ldlm_pools_scan(ldlm_side_t client, int nr, gfp_t gfp_mask)
1108 {
1109 	unsigned long freed = 0;
1110 	int tmp, nr_ns;
1111 	struct ldlm_namespace *ns;
1112 	void *cookie;
1113 
1114 	if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
1115 		return -1;
1116 
1117 	cookie = cl_env_reenter();
1118 
1119 	/*
1120 	 * Shrink at least ldlm_namespace_nr_read(client) namespaces.
1121 	 */
1122 	for (tmp = nr_ns = ldlm_namespace_nr_read(client);
1123 	     tmp > 0; tmp--) {
1124 		int cancel, nr_locks;
1125 
1126 		/*
1127 		 * Do not call shrink under ldlm_namespace_lock(client)
1128 		 */
1129 		mutex_lock(ldlm_namespace_lock(client));
1130 		if (list_empty(ldlm_namespace_list(client))) {
1131 			mutex_unlock(ldlm_namespace_lock(client));
1132 			break;
1133 		}
1134 		ns = ldlm_namespace_first_locked(client);
1135 		ldlm_namespace_get(ns);
1136 		ldlm_namespace_move_to_active_locked(ns, client);
1137 		mutex_unlock(ldlm_namespace_lock(client));
1138 
1139 		nr_locks = ldlm_pool_granted(&ns->ns_pool);
1140 		/*
1141 		 * We use to shrink propotionally but with new shrinker API,
1142 		 * we lost the total number of freeable locks.
1143 		 */
1144 		cancel = 1 + min_t(int, nr_locks, nr / nr_ns);
1145 		freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
1146 		ldlm_namespace_put(ns);
1147 	}
1148 	cl_env_reexit(cookie);
1149 	/*
1150 	 * we only decrease the SLV in server pools shrinker, return
1151 	 * SHRINK_STOP to kernel to avoid needless loop. LU-1128
1152 	 */
1153 	return (client == LDLM_NAMESPACE_SERVER) ? SHRINK_STOP : freed;
1154 }
1155 
ldlm_pools_srv_count(struct shrinker * s,struct shrink_control * sc)1156 static unsigned long ldlm_pools_srv_count(struct shrinker *s,
1157 					  struct shrink_control *sc)
1158 {
1159 	return ldlm_pools_count(LDLM_NAMESPACE_SERVER, sc->gfp_mask);
1160 }
1161 
ldlm_pools_srv_scan(struct shrinker * s,struct shrink_control * sc)1162 static unsigned long ldlm_pools_srv_scan(struct shrinker *s,
1163 					 struct shrink_control *sc)
1164 {
1165 	return ldlm_pools_scan(LDLM_NAMESPACE_SERVER, sc->nr_to_scan,
1166 			       sc->gfp_mask);
1167 }
1168 
ldlm_pools_cli_count(struct shrinker * s,struct shrink_control * sc)1169 static unsigned long ldlm_pools_cli_count(struct shrinker *s,
1170 					  struct shrink_control *sc)
1171 {
1172 	return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask);
1173 }
1174 
ldlm_pools_cli_scan(struct shrinker * s,struct shrink_control * sc)1175 static unsigned long ldlm_pools_cli_scan(struct shrinker *s,
1176 					 struct shrink_control *sc)
1177 {
1178 	return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan,
1179 			       sc->gfp_mask);
1180 }
1181 
ldlm_pools_recalc(ldlm_side_t client)1182 int ldlm_pools_recalc(ldlm_side_t client)
1183 {
1184 	__u32 nr_l = 0, nr_p = 0, l;
1185 	struct ldlm_namespace *ns;
1186 	struct ldlm_namespace *ns_old = NULL;
1187 	int nr, equal = 0;
1188 	int time = 50; /* seconds of sleep if no active namespaces */
1189 
1190 	/*
1191 	 * No need to setup pool limit for client pools.
1192 	 */
1193 	if (client == LDLM_NAMESPACE_SERVER) {
1194 		/*
1195 		 * Check all modest namespaces first.
1196 		 */
1197 		mutex_lock(ldlm_namespace_lock(client));
1198 		list_for_each_entry(ns, ldlm_namespace_list(client),
1199 					ns_list_chain) {
1200 			if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
1201 				continue;
1202 
1203 			l = ldlm_pool_granted(&ns->ns_pool);
1204 			if (l == 0)
1205 				l = 1;
1206 
1207 			/*
1208 			 * Set the modest pools limit equal to their avg granted
1209 			 * locks + ~6%.
1210 			 */
1211 			l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0);
1212 			ldlm_pool_setup(&ns->ns_pool, l);
1213 			nr_l += l;
1214 			nr_p++;
1215 		}
1216 
1217 		/*
1218 		 * Make sure that modest namespaces did not eat more that 2/3
1219 		 * of limit.
1220 		 */
1221 		if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) {
1222 			CWARN("\"Modest\" pools eat out 2/3 of server locks limit (%d of %lu). This means that you have too many clients for this amount of server RAM. Upgrade server!\n",
1223 			      nr_l, LDLM_POOL_HOST_L);
1224 			equal = 1;
1225 		}
1226 
1227 		/*
1228 		 * The rest is given to greedy namespaces.
1229 		 */
1230 		list_for_each_entry(ns, ldlm_namespace_list(client),
1231 				    ns_list_chain) {
1232 			if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
1233 				continue;
1234 
1235 			if (equal) {
1236 				/*
1237 				 * In the case 2/3 locks are eaten out by
1238 				 * modest pools, we re-setup equal limit
1239 				 * for _all_ pools.
1240 				 */
1241 				l = LDLM_POOL_HOST_L /
1242 					ldlm_namespace_nr_read(client);
1243 			} else {
1244 				/*
1245 				 * All the rest of greedy pools will have
1246 				 * all locks in equal parts.
1247 				 */
1248 				l = (LDLM_POOL_HOST_L - nr_l) /
1249 					(ldlm_namespace_nr_read(client) -
1250 					 nr_p);
1251 			}
1252 			ldlm_pool_setup(&ns->ns_pool, l);
1253 		}
1254 		mutex_unlock(ldlm_namespace_lock(client));
1255 	}
1256 
1257 	/*
1258 	 * Recalc at least ldlm_namespace_nr_read(client) namespaces.
1259 	 */
1260 	for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) {
1261 		int     skip;
1262 		/*
1263 		 * Lock the list, get first @ns in the list, getref, move it
1264 		 * to the tail, unlock and call pool recalc. This way we avoid
1265 		 * calling recalc under @ns lock what is really good as we get
1266 		 * rid of potential deadlock on client nodes when canceling
1267 		 * locks synchronously.
1268 		 */
1269 		mutex_lock(ldlm_namespace_lock(client));
1270 		if (list_empty(ldlm_namespace_list(client))) {
1271 			mutex_unlock(ldlm_namespace_lock(client));
1272 			break;
1273 		}
1274 		ns = ldlm_namespace_first_locked(client);
1275 
1276 		if (ns_old == ns) { /* Full pass complete */
1277 			mutex_unlock(ldlm_namespace_lock(client));
1278 			break;
1279 		}
1280 
1281 		/* We got an empty namespace, need to move it back to inactive
1282 		 * list.
1283 		 * The race with parallel resource creation is fine:
1284 		 * - If they do namespace_get before our check, we fail the
1285 		 *   check and they move this item to the end of the list anyway
1286 		 * - If we do the check and then they do namespace_get, then
1287 		 *   we move the namespace to inactive and they will move
1288 		 *   it back to active (synchronised by the lock, so no clash
1289 		 *   there).
1290 		 */
1291 		if (ldlm_ns_empty(ns)) {
1292 			ldlm_namespace_move_to_inactive_locked(ns, client);
1293 			mutex_unlock(ldlm_namespace_lock(client));
1294 			continue;
1295 		}
1296 
1297 		if (ns_old == NULL)
1298 			ns_old = ns;
1299 
1300 		spin_lock(&ns->ns_lock);
1301 		/*
1302 		 * skip ns which is being freed, and we don't want to increase
1303 		 * its refcount again, not even temporarily. bz21519 & LU-499.
1304 		 */
1305 		if (ns->ns_stopping) {
1306 			skip = 1;
1307 		} else {
1308 			skip = 0;
1309 			ldlm_namespace_get(ns);
1310 		}
1311 		spin_unlock(&ns->ns_lock);
1312 
1313 		ldlm_namespace_move_to_active_locked(ns, client);
1314 		mutex_unlock(ldlm_namespace_lock(client));
1315 
1316 		/*
1317 		 * After setup is done - recalc the pool.
1318 		 */
1319 		if (!skip) {
1320 			int ttime = ldlm_pool_recalc(&ns->ns_pool);
1321 
1322 			if (ttime < time)
1323 				time = ttime;
1324 
1325 			ldlm_namespace_put(ns);
1326 		}
1327 	}
1328 	return time;
1329 }
1330 EXPORT_SYMBOL(ldlm_pools_recalc);
1331 
ldlm_pools_thread_main(void * arg)1332 static int ldlm_pools_thread_main(void *arg)
1333 {
1334 	struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
1335 	int s_time, c_time;
1336 
1337 	thread_set_flags(thread, SVC_RUNNING);
1338 	wake_up(&thread->t_ctl_waitq);
1339 
1340 	CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n",
1341 		"ldlm_poold", current_pid());
1342 
1343 	while (1) {
1344 		struct l_wait_info lwi;
1345 
1346 		/*
1347 		 * Recal all pools on this tick.
1348 		 */
1349 		s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
1350 		c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
1351 
1352 		/*
1353 		 * Wait until the next check time, or until we're
1354 		 * stopped.
1355 		 */
1356 		lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)),
1357 				  NULL, NULL);
1358 		l_wait_event(thread->t_ctl_waitq,
1359 			     thread_is_stopping(thread) ||
1360 			     thread_is_event(thread),
1361 			     &lwi);
1362 
1363 		if (thread_test_and_clear_flags(thread, SVC_STOPPING))
1364 			break;
1365 		else
1366 			thread_test_and_clear_flags(thread, SVC_EVENT);
1367 	}
1368 
1369 	thread_set_flags(thread, SVC_STOPPED);
1370 	wake_up(&thread->t_ctl_waitq);
1371 
1372 	CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n",
1373 		"ldlm_poold", current_pid());
1374 
1375 	complete_and_exit(&ldlm_pools_comp, 0);
1376 }
1377 
ldlm_pools_thread_start(void)1378 static int ldlm_pools_thread_start(void)
1379 {
1380 	struct l_wait_info lwi = { 0 };
1381 	struct task_struct *task;
1382 
1383 	if (ldlm_pools_thread != NULL)
1384 		return -EALREADY;
1385 
1386 	OBD_ALLOC_PTR(ldlm_pools_thread);
1387 	if (ldlm_pools_thread == NULL)
1388 		return -ENOMEM;
1389 
1390 	init_completion(&ldlm_pools_comp);
1391 	init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq);
1392 
1393 	task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread,
1394 			   "ldlm_poold");
1395 	if (IS_ERR(task)) {
1396 		CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task));
1397 		OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread));
1398 		ldlm_pools_thread = NULL;
1399 		return PTR_ERR(task);
1400 	}
1401 	l_wait_event(ldlm_pools_thread->t_ctl_waitq,
1402 		     thread_is_running(ldlm_pools_thread), &lwi);
1403 	return 0;
1404 }
1405 
ldlm_pools_thread_stop(void)1406 static void ldlm_pools_thread_stop(void)
1407 {
1408 	if (ldlm_pools_thread == NULL)
1409 		return;
1410 
1411 	thread_set_flags(ldlm_pools_thread, SVC_STOPPING);
1412 	wake_up(&ldlm_pools_thread->t_ctl_waitq);
1413 
1414 	/*
1415 	 * Make sure that pools thread is finished before freeing @thread.
1416 	 * This fixes possible race and oops due to accessing freed memory
1417 	 * in pools thread.
1418 	 */
1419 	wait_for_completion(&ldlm_pools_comp);
1420 	OBD_FREE_PTR(ldlm_pools_thread);
1421 	ldlm_pools_thread = NULL;
1422 }
1423 
1424 static struct shrinker ldlm_pools_srv_shrinker = {
1425 	.count_objects	= ldlm_pools_srv_count,
1426 	.scan_objects	= ldlm_pools_srv_scan,
1427 	.seeks		= DEFAULT_SEEKS,
1428 };
1429 
1430 static struct shrinker ldlm_pools_cli_shrinker = {
1431 	.count_objects	= ldlm_pools_cli_count,
1432 	.scan_objects	= ldlm_pools_cli_scan,
1433 	.seeks		= DEFAULT_SEEKS,
1434 };
1435 
ldlm_pools_init(void)1436 int ldlm_pools_init(void)
1437 {
1438 	int rc;
1439 
1440 	rc = ldlm_pools_thread_start();
1441 	if (rc == 0) {
1442 		register_shrinker(&ldlm_pools_srv_shrinker);
1443 		register_shrinker(&ldlm_pools_cli_shrinker);
1444 	}
1445 	return rc;
1446 }
1447 EXPORT_SYMBOL(ldlm_pools_init);
1448 
ldlm_pools_fini(void)1449 void ldlm_pools_fini(void)
1450 {
1451 	unregister_shrinker(&ldlm_pools_srv_shrinker);
1452 	unregister_shrinker(&ldlm_pools_cli_shrinker);
1453 	ldlm_pools_thread_stop();
1454 }
1455 EXPORT_SYMBOL(ldlm_pools_fini);
1456