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;
251 /* CPT affinity scheduler? */
252 if (sched->ws_cptab != NULL)
253 cfs_cpt_bind(sched->ws_cptab, sched->ws_cpt);
255 spin_lock(&cfs_wi_data.wi_glock);
257 LASSERT(sched->ws_starting == 1);
258 sched->ws_starting--;
259 sched->ws_nthreads++;
261 spin_unlock(&cfs_wi_data.wi_glock);
263 cfs_wi_sched_lock(sched);
265 while (!sched->ws_stopping) {
270 while (!cfs_list_empty(&sched->ws_runq) &&
271 nloops < CFS_WI_RESCHED) {
272 wi = cfs_list_entry(sched->ws_runq.next,
273 cfs_workitem_t, wi_list);
274 LASSERT(wi->wi_scheduled && !wi->wi_running);
276 cfs_list_del_init(&wi->wi_list);
278 LASSERT(sched->ws_nscheduled > 0);
279 sched->ws_nscheduled--;
282 wi->wi_scheduled = 0;
285 cfs_wi_sched_unlock(sched);
288 rc = (*wi->wi_action) (wi);
290 cfs_wi_sched_lock(sched);
291 if (rc != 0) /* WI should be dead, even be freed! */
295 if (cfs_list_empty(&wi->wi_list))
298 LASSERT(wi->wi_scheduled);
299 /* wi is rescheduled, should be on rerunq now, we
300 * move it to runq so it can run action now */
301 cfs_list_move_tail(&wi->wi_list, &sched->ws_runq);
304 if (!cfs_list_empty(&sched->ws_runq)) {
305 cfs_wi_sched_unlock(sched);
306 /* don't sleep because some workitems still
307 * expect me to come back soon */
309 cfs_wi_sched_lock(sched);
313 cfs_wi_sched_unlock(sched);
314 cfs_wait_event_interruptible_exclusive(sched->ws_waitq,
315 !cfs_wi_sched_cansleep(sched), rc);
316 cfs_wi_sched_lock(sched);
319 cfs_wi_sched_unlock(sched);
321 spin_lock(&cfs_wi_data.wi_glock);
322 sched->ws_nthreads--;
323 spin_unlock(&cfs_wi_data.wi_glock);
328 #else /* __KERNEL__ */
331 cfs_wi_check_events (void)
336 spin_lock(&cfs_wi_data.wi_glock);
339 struct cfs_wi_sched *sched = NULL;
340 struct cfs_wi_sched *tmp;
342 /** rerunq is always empty for userspace */
343 cfs_list_for_each_entry(tmp,
344 &cfs_wi_data.wi_scheds, ws_list) {
345 if (!cfs_list_empty(&tmp->ws_runq)) {
354 wi = cfs_list_entry(sched->ws_runq.next,
355 cfs_workitem_t, wi_list);
356 cfs_list_del_init(&wi->wi_list);
358 LASSERT(sched->ws_nscheduled > 0);
359 sched->ws_nscheduled--;
361 LASSERT(wi->wi_scheduled);
362 wi->wi_scheduled = 0;
363 spin_unlock(&cfs_wi_data.wi_glock);
366 (*wi->wi_action) (wi);
368 spin_lock(&cfs_wi_data.wi_glock);
371 spin_unlock(&cfs_wi_data.wi_glock);
378 cfs_wi_sched_destroy(struct cfs_wi_sched *sched)
382 LASSERT(cfs_wi_data.wi_init);
383 LASSERT(!cfs_wi_data.wi_stopping);
385 spin_lock(&cfs_wi_data.wi_glock);
386 if (sched->ws_stopping) {
387 CDEBUG(D_INFO, "%s is in progress of stopping\n",
389 spin_unlock(&cfs_wi_data.wi_glock);
393 LASSERT(!cfs_list_empty(&sched->ws_list));
394 sched->ws_stopping = 1;
396 spin_unlock(&cfs_wi_data.wi_glock);
400 cfs_waitq_broadcast(&sched->ws_waitq);
402 spin_lock(&cfs_wi_data.wi_glock);
403 while (sched->ws_nthreads > 0) {
404 CDEBUG(IS_PO2(++i) ? D_WARNING : D_NET,
405 "waiting for %d threads of WI sched[%s] to terminate\n",
406 sched->ws_nthreads, sched->ws_name);
408 spin_unlock(&cfs_wi_data.wi_glock);
409 cfs_pause(cfs_time_seconds(1) / 20);
410 spin_lock(&cfs_wi_data.wi_glock);
413 cfs_list_del(&sched->ws_list);
415 spin_unlock(&cfs_wi_data.wi_glock);
419 LASSERT(sched->ws_nscheduled == 0);
421 LIBCFS_FREE(sched, sizeof(*sched));
423 EXPORT_SYMBOL(cfs_wi_sched_destroy);
426 cfs_wi_sched_create(char *name, struct cfs_cpt_table *cptab,
427 int cpt, int nthrs, struct cfs_wi_sched **sched_pp)
429 struct cfs_wi_sched *sched;
432 LASSERT(cfs_wi_data.wi_init);
433 LASSERT(!cfs_wi_data.wi_stopping);
434 LASSERT(cptab == NULL || cpt == CFS_CPT_ANY ||
435 (cpt >= 0 && cpt < cfs_cpt_number(cptab)));
437 LIBCFS_ALLOC(sched, sizeof(*sched));
441 strncpy(sched->ws_name, name, CFS_WS_NAME_LEN);
442 sched->ws_cptab = cptab;
446 spin_lock_init(&sched->ws_lock);
447 cfs_waitq_init(&sched->ws_waitq);
449 CFS_INIT_LIST_HEAD(&sched->ws_runq);
450 CFS_INIT_LIST_HEAD(&sched->ws_rerunq);
451 CFS_INIT_LIST_HEAD(&sched->ws_list);
458 spin_lock(&cfs_wi_data.wi_glock);
459 while (sched->ws_starting > 0) {
460 spin_unlock(&cfs_wi_data.wi_glock);
462 spin_lock(&cfs_wi_data.wi_glock);
465 sched->ws_starting++;
466 spin_unlock(&cfs_wi_data.wi_glock);
468 if (sched->ws_cptab != NULL && sched->ws_cpt >= 0) {
469 snprintf(name, sizeof(name), "%s_%02d_%02d",
470 sched->ws_name, sched->ws_cpt,
473 snprintf(name, sizeof(name), "%s_%02d",
474 sched->ws_name, sched->ws_nthreads);
477 task = kthread_run(cfs_wi_scheduler, sched, name);
484 CERROR("Failed to create thread for WI scheduler %s: %d\n",
487 spin_lock(&cfs_wi_data.wi_glock);
489 /* make up for cfs_wi_sched_destroy */
490 cfs_list_add(&sched->ws_list, &cfs_wi_data.wi_scheds);
491 sched->ws_starting--;
493 spin_unlock(&cfs_wi_data.wi_glock);
495 cfs_wi_sched_destroy(sched);
501 spin_lock(&cfs_wi_data.wi_glock);
502 cfs_list_add(&sched->ws_list, &cfs_wi_data.wi_scheds);
503 spin_unlock(&cfs_wi_data.wi_glock);
508 EXPORT_SYMBOL(cfs_wi_sched_create);
513 memset(&cfs_wi_data, 0, sizeof(cfs_wi_data));
515 spin_lock_init(&cfs_wi_data.wi_glock);
516 CFS_INIT_LIST_HEAD(&cfs_wi_data.wi_scheds);
517 cfs_wi_data.wi_init = 1;
523 cfs_wi_shutdown (void)
525 struct cfs_wi_sched *sched;
527 spin_lock(&cfs_wi_data.wi_glock);
528 cfs_wi_data.wi_stopping = 1;
529 spin_unlock(&cfs_wi_data.wi_glock);
532 /* nobody should contend on this list */
533 cfs_list_for_each_entry(sched, &cfs_wi_data.wi_scheds, ws_list) {
534 sched->ws_stopping = 1;
535 cfs_waitq_broadcast(&sched->ws_waitq);
538 cfs_list_for_each_entry(sched, &cfs_wi_data.wi_scheds, ws_list) {
539 spin_lock(&cfs_wi_data.wi_glock);
541 while (sched->ws_nthreads != 0) {
542 spin_unlock(&cfs_wi_data.wi_glock);
543 cfs_pause(cfs_time_seconds(1) / 20);
544 spin_lock(&cfs_wi_data.wi_glock);
546 spin_unlock(&cfs_wi_data.wi_glock);
549 while (!cfs_list_empty(&cfs_wi_data.wi_scheds)) {
550 sched = cfs_list_entry(cfs_wi_data.wi_scheds.next,
551 struct cfs_wi_sched, ws_list);
552 cfs_list_del(&sched->ws_list);
553 LIBCFS_FREE(sched, sizeof(*sched));
556 cfs_wi_data.wi_stopping = 0;
557 cfs_wi_data.wi_init = 0;