4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
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
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * libcfs/libcfs/workitem.c
38 * Author: Isaac Huang <isaac@clusterfs.com>
39 * Liang Zhen <zhen.liang@sun.com>
42 #define DEBUG_SUBSYSTEM S_LNET
44 #include <libcfs/libcfs.h>
46 #define CFS_WS_NAME_LEN 16
48 typedef struct cfs_wi_sched {
49 cfs_list_t ws_list; /* chain on global list */
51 /** serialised workitems */
53 /** where schedulers sleep */
56 /** concurrent workitems */
58 /** rescheduled running-workitems, a workitem can be rescheduled
59 * while running in wi_action(), but we don't to execute it again
60 * unless it returns from wi_action(), so we put it on ws_rerunq
61 * while rescheduling, and move it to runq after it returns
64 /** CPT-table for this scheduler */
65 struct cfs_cpt_table *ws_cptab;
66 /** CPT id for affinity */
68 /** number of scheduled workitems */
70 /** started scheduler thread, protected by cfs_wi_data::wi_glock */
71 unsigned int ws_nthreads:30;
72 /** shutting down, protected by cfs_wi_data::wi_glock */
73 unsigned int ws_stopping:1;
74 /** serialize starting thread, protected by cfs_wi_data::wi_glock */
75 unsigned int ws_starting:1;
77 char ws_name[CFS_WS_NAME_LEN];
80 struct cfs_workitem_data {
83 /** list of all schedulers */
85 /** WI module is initialized */
87 /** shutting down the whole WI module */
93 cfs_wi_sched_lock(cfs_wi_sched_t *sched)
95 spin_lock(&sched->ws_lock);
99 cfs_wi_sched_unlock(cfs_wi_sched_t *sched)
101 spin_unlock(&sched->ws_lock);
105 cfs_wi_sched_cansleep(cfs_wi_sched_t *sched)
107 cfs_wi_sched_lock(sched);
108 if (sched->ws_stopping) {
109 cfs_wi_sched_unlock(sched);
113 if (!cfs_list_empty(&sched->ws_runq)) {
114 cfs_wi_sched_unlock(sched);
117 cfs_wi_sched_unlock(sched);
121 #else /* !__KERNEL__ */
124 cfs_wi_sched_lock(cfs_wi_sched_t *sched)
126 spin_lock(&cfs_wi_data.wi_glock);
130 cfs_wi_sched_unlock(cfs_wi_sched_t *sched)
132 spin_unlock(&cfs_wi_data.wi_glock);
135 #endif /* __KERNEL__ */
138 * 0. it only works when called from wi->wi_action.
139 * 1. when it returns no one shall try to schedule the workitem.
142 cfs_wi_exit(struct cfs_wi_sched *sched, cfs_workitem_t *wi)
144 LASSERT(!cfs_in_interrupt()); /* because we use plain spinlock */
145 LASSERT(!sched->ws_stopping);
147 cfs_wi_sched_lock(sched);
150 LASSERT(wi->wi_running);
152 if (wi->wi_scheduled) { /* cancel pending schedules */
153 LASSERT(!cfs_list_empty(&wi->wi_list));
154 cfs_list_del_init(&wi->wi_list);
156 LASSERT(sched->ws_nscheduled > 0);
157 sched->ws_nscheduled--;
160 LASSERT(cfs_list_empty(&wi->wi_list));
162 wi->wi_scheduled = 1; /* LBUG future schedule attempts */
163 cfs_wi_sched_unlock(sched);
167 EXPORT_SYMBOL(cfs_wi_exit);
170 * cancel schedule request of workitem \a wi
173 cfs_wi_deschedule(struct cfs_wi_sched *sched, cfs_workitem_t *wi)
177 LASSERT(!cfs_in_interrupt()); /* because we use plain spinlock */
178 LASSERT(!sched->ws_stopping);
181 * return 0 if it's running already, otherwise return 1, which
182 * means the workitem will not be scheduled and will not have
183 * any race with wi_action.
185 cfs_wi_sched_lock(sched);
187 rc = !(wi->wi_running);
189 if (wi->wi_scheduled) { /* cancel pending schedules */
190 LASSERT(!cfs_list_empty(&wi->wi_list));
191 cfs_list_del_init(&wi->wi_list);
193 LASSERT(sched->ws_nscheduled > 0);
194 sched->ws_nscheduled--;
196 wi->wi_scheduled = 0;
199 LASSERT (cfs_list_empty(&wi->wi_list));
201 cfs_wi_sched_unlock(sched);
204 EXPORT_SYMBOL(cfs_wi_deschedule);
207 * Workitem scheduled with (serial == 1) is strictly serialised not only with
208 * itself, but also with others scheduled this way.
210 * Now there's only one static serialised queue, but in the future more might
211 * be added, and even dynamic creation of serialised queues might be supported.
214 cfs_wi_schedule(struct cfs_wi_sched *sched, cfs_workitem_t *wi)
216 LASSERT(!cfs_in_interrupt()); /* because we use plain spinlock */
217 LASSERT(!sched->ws_stopping);
219 cfs_wi_sched_lock(sched);
221 if (!wi->wi_scheduled) {
222 LASSERT (cfs_list_empty(&wi->wi_list));
224 wi->wi_scheduled = 1;
225 sched->ws_nscheduled++;
226 if (!wi->wi_running) {
227 cfs_list_add_tail(&wi->wi_list, &sched->ws_runq);
229 cfs_waitq_signal(&sched->ws_waitq);
232 cfs_list_add(&wi->wi_list, &sched->ws_rerunq);
236 LASSERT (!cfs_list_empty(&wi->wi_list));
237 cfs_wi_sched_unlock(sched);
240 EXPORT_SYMBOL(cfs_wi_schedule);
245 cfs_wi_scheduler (void *arg)
247 struct cfs_wi_sched *sched = (cfs_wi_sched_t *)arg;
250 if (sched->ws_cptab != NULL && sched->ws_cpt >= 0) {
251 snprintf(name, sizeof(name), "%s_%02d_%02d",
252 sched->ws_name, sched->ws_cpt, sched->ws_nthreads);
254 snprintf(name, sizeof(name), "%s_%02d",
255 sched->ws_name, sched->ws_nthreads);
261 /* CPT affinity scheduler? */
262 if (sched->ws_cptab != NULL)
263 cfs_cpt_bind(sched->ws_cptab, sched->ws_cpt);
265 spin_lock(&cfs_wi_data.wi_glock);
267 LASSERT(sched->ws_starting == 1);
268 sched->ws_starting--;
269 sched->ws_nthreads++;
271 spin_unlock(&cfs_wi_data.wi_glock);
273 cfs_wi_sched_lock(sched);
275 while (!sched->ws_stopping) {
280 while (!cfs_list_empty(&sched->ws_runq) &&
281 nloops < CFS_WI_RESCHED) {
282 wi = cfs_list_entry(sched->ws_runq.next,
283 cfs_workitem_t, wi_list);
284 LASSERT(wi->wi_scheduled && !wi->wi_running);
286 cfs_list_del_init(&wi->wi_list);
288 LASSERT(sched->ws_nscheduled > 0);
289 sched->ws_nscheduled--;
292 wi->wi_scheduled = 0;
295 cfs_wi_sched_unlock(sched);
298 rc = (*wi->wi_action) (wi);
300 cfs_wi_sched_lock(sched);
301 if (rc != 0) /* WI should be dead, even be freed! */
305 if (cfs_list_empty(&wi->wi_list))
308 LASSERT(wi->wi_scheduled);
309 /* wi is rescheduled, should be on rerunq now, we
310 * move it to runq so it can run action now */
311 cfs_list_move_tail(&wi->wi_list, &sched->ws_runq);
314 if (!cfs_list_empty(&sched->ws_runq)) {
315 cfs_wi_sched_unlock(sched);
316 /* don't sleep because some workitems still
317 * expect me to come back soon */
319 cfs_wi_sched_lock(sched);
323 cfs_wi_sched_unlock(sched);
324 cfs_wait_event_interruptible_exclusive(sched->ws_waitq,
325 !cfs_wi_sched_cansleep(sched), rc);
326 cfs_wi_sched_lock(sched);
329 cfs_wi_sched_unlock(sched);
331 spin_lock(&cfs_wi_data.wi_glock);
332 sched->ws_nthreads--;
333 spin_unlock(&cfs_wi_data.wi_glock);
338 #else /* __KERNEL__ */
341 cfs_wi_check_events (void)
346 spin_lock(&cfs_wi_data.wi_glock);
349 struct cfs_wi_sched *sched = NULL;
350 struct cfs_wi_sched *tmp;
352 /** rerunq is always empty for userspace */
353 cfs_list_for_each_entry(tmp,
354 &cfs_wi_data.wi_scheds, ws_list) {
355 if (!cfs_list_empty(&tmp->ws_runq)) {
364 wi = cfs_list_entry(sched->ws_runq.next,
365 cfs_workitem_t, wi_list);
366 cfs_list_del_init(&wi->wi_list);
368 LASSERT(sched->ws_nscheduled > 0);
369 sched->ws_nscheduled--;
371 LASSERT(wi->wi_scheduled);
372 wi->wi_scheduled = 0;
373 spin_unlock(&cfs_wi_data.wi_glock);
376 (*wi->wi_action) (wi);
378 spin_lock(&cfs_wi_data.wi_glock);
381 spin_unlock(&cfs_wi_data.wi_glock);
388 cfs_wi_sched_destroy(struct cfs_wi_sched *sched)
392 LASSERT(cfs_wi_data.wi_init);
393 LASSERT(!cfs_wi_data.wi_stopping);
395 spin_lock(&cfs_wi_data.wi_glock);
396 if (sched->ws_stopping) {
397 CDEBUG(D_INFO, "%s is in progress of stopping\n",
399 spin_unlock(&cfs_wi_data.wi_glock);
403 LASSERT(!cfs_list_empty(&sched->ws_list));
404 sched->ws_stopping = 1;
406 spin_unlock(&cfs_wi_data.wi_glock);
410 cfs_waitq_broadcast(&sched->ws_waitq);
412 spin_lock(&cfs_wi_data.wi_glock);
413 while (sched->ws_nthreads > 0) {
414 CDEBUG(IS_PO2(++i) ? D_WARNING : D_NET,
415 "waiting for %d threads of WI sched[%s] to terminate\n",
416 sched->ws_nthreads, sched->ws_name);
418 spin_unlock(&cfs_wi_data.wi_glock);
419 cfs_pause(cfs_time_seconds(1) / 20);
420 spin_lock(&cfs_wi_data.wi_glock);
423 cfs_list_del(&sched->ws_list);
425 spin_unlock(&cfs_wi_data.wi_glock);
429 LASSERT(sched->ws_nscheduled == 0);
431 LIBCFS_FREE(sched, sizeof(*sched));
433 EXPORT_SYMBOL(cfs_wi_sched_destroy);
436 cfs_wi_sched_create(char *name, struct cfs_cpt_table *cptab,
437 int cpt, int nthrs, struct cfs_wi_sched **sched_pp)
439 struct cfs_wi_sched *sched;
442 LASSERT(cfs_wi_data.wi_init);
443 LASSERT(!cfs_wi_data.wi_stopping);
444 LASSERT(cptab == NULL || cpt == CFS_CPT_ANY ||
445 (cpt >= 0 && cpt < cfs_cpt_number(cptab)));
447 LIBCFS_ALLOC(sched, sizeof(*sched));
451 strncpy(sched->ws_name, name, CFS_WS_NAME_LEN);
452 sched->ws_cptab = cptab;
456 spin_lock_init(&sched->ws_lock);
457 cfs_waitq_init(&sched->ws_waitq);
459 CFS_INIT_LIST_HEAD(&sched->ws_runq);
460 CFS_INIT_LIST_HEAD(&sched->ws_rerunq);
461 CFS_INIT_LIST_HEAD(&sched->ws_list);
466 spin_lock(&cfs_wi_data.wi_glock);
467 while (sched->ws_starting > 0) {
468 spin_unlock(&cfs_wi_data.wi_glock);
470 spin_lock(&cfs_wi_data.wi_glock);
473 sched->ws_starting++;
474 spin_unlock(&cfs_wi_data.wi_glock);
476 rc = cfs_create_thread(cfs_wi_scheduler, sched, 0);
482 CERROR("Failed to create thread for WI scheduler %s: %d\n",
485 spin_lock(&cfs_wi_data.wi_glock);
487 /* make up for cfs_wi_sched_destroy */
488 cfs_list_add(&sched->ws_list, &cfs_wi_data.wi_scheds);
489 sched->ws_starting--;
491 spin_unlock(&cfs_wi_data.wi_glock);
493 cfs_wi_sched_destroy(sched);
499 spin_lock(&cfs_wi_data.wi_glock);
500 cfs_list_add(&sched->ws_list, &cfs_wi_data.wi_scheds);
501 spin_unlock(&cfs_wi_data.wi_glock);
506 EXPORT_SYMBOL(cfs_wi_sched_create);
511 memset(&cfs_wi_data, 0, sizeof(cfs_wi_data));
513 spin_lock_init(&cfs_wi_data.wi_glock);
514 CFS_INIT_LIST_HEAD(&cfs_wi_data.wi_scheds);
515 cfs_wi_data.wi_init = 1;
521 cfs_wi_shutdown (void)
523 struct cfs_wi_sched *sched;
525 spin_lock(&cfs_wi_data.wi_glock);
526 cfs_wi_data.wi_stopping = 1;
527 spin_unlock(&cfs_wi_data.wi_glock);
530 /* nobody should contend on this list */
531 cfs_list_for_each_entry(sched, &cfs_wi_data.wi_scheds, ws_list) {
532 sched->ws_stopping = 1;
533 cfs_waitq_broadcast(&sched->ws_waitq);
536 cfs_list_for_each_entry(sched, &cfs_wi_data.wi_scheds, ws_list) {
537 spin_lock(&cfs_wi_data.wi_glock);
539 while (sched->ws_nthreads != 0) {
540 spin_unlock(&cfs_wi_data.wi_glock);
541 cfs_pause(cfs_time_seconds(1) / 20);
542 spin_lock(&cfs_wi_data.wi_glock);
544 spin_unlock(&cfs_wi_data.wi_glock);
547 while (!cfs_list_empty(&cfs_wi_data.wi_scheds)) {
548 sched = cfs_list_entry(cfs_wi_data.wi_scheds.next,
549 struct cfs_wi_sched, ws_list);
550 cfs_list_del(&sched->ws_list);
551 LIBCFS_FREE(sched, sizeof(*sched));
554 cfs_wi_data.wi_stopping = 0;
555 cfs_wi_data.wi_init = 0;