1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * standalone DLM module
8 * Copyright (C) 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
28 #include <linux/module.h>
30 #include <linux/types.h>
31 #include <linux/slab.h>
32 #include <linux/highmem.h>
33 #include <linux/utsname.h>
34 #include <linux/init.h>
35 #include <linux/sysctl.h>
36 #include <linux/random.h>
37 #include <linux/blkdev.h>
38 #include <linux/socket.h>
39 #include <linux/inet.h>
40 #include <linux/timer.h>
41 #include <linux/kthread.h>
42 #include <linux/delay.h>
45 #include "cluster/heartbeat.h"
46 #include "cluster/nodemanager.h"
47 #include "cluster/tcp.h"
50 #include "dlmcommon.h"
51 #include "dlmdomain.h"
53 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
54 #include "cluster/masklog.h"
56 static int dlm_thread(void *data);
58 static void dlm_flush_asts(struct dlm_ctxt *dlm);
60 #define dlm_lock_is_remote(dlm, lock) ((lock)->ml.node != (dlm)->node_num)
62 /* will exit holding res->spinlock, but may drop in function */
63 /* waits until flags are cleared on res->state */
64 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
66 DECLARE_WAITQUEUE(wait, current);
68 assert_spin_locked(&res->spinlock);
70 add_wait_queue(&res->wq, &wait);
72 set_current_state(TASK_UNINTERRUPTIBLE);
73 if (res->state & flags) {
74 spin_unlock(&res->spinlock);
76 spin_lock(&res->spinlock);
79 remove_wait_queue(&res->wq, &wait);
80 current->state = TASK_RUNNING;
84 static int __dlm_lockres_unused(struct dlm_lock_resource *res)
86 if (list_empty(&res->granted) &&
87 list_empty(&res->converting) &&
88 list_empty(&res->blocked) &&
89 list_empty(&res->dirty))
95 /* Call whenever you may have added or deleted something from one of
96 * the lockres queue's. This will figure out whether it belongs on the
97 * unused list or not and does the appropriate thing. */
98 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
99 struct dlm_lock_resource *res)
101 mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
103 assert_spin_locked(&dlm->spinlock);
104 assert_spin_locked(&res->spinlock);
106 if (__dlm_lockres_unused(res)){
107 if (list_empty(&res->purge)) {
108 mlog(0, "putting lockres %.*s from purge list\n",
109 res->lockname.len, res->lockname.name);
111 res->last_used = jiffies;
112 list_add_tail(&res->purge, &dlm->purge_list);
115 } else if (!list_empty(&res->purge)) {
116 mlog(0, "removing lockres %.*s from purge list\n",
117 res->lockname.len, res->lockname.name);
119 list_del_init(&res->purge);
124 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
125 struct dlm_lock_resource *res)
127 mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
128 spin_lock(&dlm->spinlock);
129 spin_lock(&res->spinlock);
131 __dlm_lockres_calc_usage(dlm, res);
133 spin_unlock(&res->spinlock);
134 spin_unlock(&dlm->spinlock);
137 /* TODO: Eventual API: Called with the dlm spinlock held, may drop it
138 * to do migration, but will re-acquire before exit. */
139 void dlm_purge_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *lockres)
144 spin_lock(&lockres->spinlock);
145 master = lockres->owner == dlm->node_num;
146 spin_unlock(&lockres->spinlock);
148 mlog(0, "purging lockres %.*s, master = %d\n", lockres->lockname.len,
149 lockres->lockname.name, master);
151 /* Non master is the easy case -- no migration required, just
156 /* Wheee! Migrate lockres here! */
157 spin_unlock(&dlm->spinlock);
160 ret = dlm_migrate_lockres(dlm, lockres, O2NM_MAX_NODES);
161 if (ret == -ENOTEMPTY) {
162 mlog(ML_ERROR, "lockres %.*s still has local locks!\n",
163 lockres->lockname.len, lockres->lockname.name);
166 } else if (ret < 0) {
167 mlog(ML_NOTICE, "lockres %.*s: migrate failed, retrying\n",
168 lockres->lockname.len, lockres->lockname.name);
173 spin_lock(&dlm->spinlock);
176 if (!list_empty(&lockres->purge)) {
177 list_del_init(&lockres->purge);
180 __dlm_unhash_lockres(lockres);
183 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
186 unsigned int run_max, unused;
187 unsigned long purge_jiffies;
188 struct dlm_lock_resource *lockres;
190 spin_lock(&dlm->spinlock);
191 run_max = dlm->purge_count;
193 while(run_max && !list_empty(&dlm->purge_list)) {
196 lockres = list_entry(dlm->purge_list.next,
197 struct dlm_lock_resource, purge);
199 /* Status of the lockres *might* change so double
200 * check. If the lockres is unused, holding the dlm
201 * spinlock will prevent people from getting and more
202 * refs on it -- there's no need to keep the lockres
204 spin_lock(&lockres->spinlock);
205 unused = __dlm_lockres_unused(lockres);
206 spin_unlock(&lockres->spinlock);
211 purge_jiffies = lockres->last_used +
212 msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
214 /* Make sure that we want to be processing this guy at
216 if (!purge_now && time_after(purge_jiffies, jiffies)) {
217 /* Since resources are added to the purge list
218 * in tail order, we can stop at the first
219 * unpurgable resource -- anyone added after
220 * him will have a greater last_used value */
224 list_del_init(&lockres->purge);
227 /* This may drop and reacquire the dlm spinlock if it
228 * has to do migration. */
229 mlog(0, "calling dlm_purge_lockres!\n");
230 dlm_purge_lockres(dlm, lockres);
231 mlog(0, "DONE calling dlm_purge_lockres!\n");
233 /* Avoid adding any scheduling latencies */
234 cond_resched_lock(&dlm->spinlock);
237 spin_unlock(&dlm->spinlock);
240 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
241 struct dlm_lock_resource *res)
243 struct dlm_lock *lock, *target;
244 struct list_head *iter;
245 struct list_head *head;
248 //mlog(0, "res->lockname.len=%d\n", res->lockname.len);
249 //mlog(0, "res->lockname.name=%p\n", res->lockname.name);
250 //mlog(0, "shuffle res %.*s\n", res->lockname.len,
251 // res->lockname.name);
253 /* because this function is called with the lockres
254 * spinlock, and because we know that it is not migrating/
255 * recovering/in-progress, it is fine to reserve asts and
256 * basts right before queueing them all throughout */
257 assert_spin_locked(&res->spinlock);
258 BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
259 DLM_LOCK_RES_RECOVERING|
260 DLM_LOCK_RES_IN_PROGRESS)));
263 if (list_empty(&res->converting))
265 mlog(0, "res %.*s has locks on a convert queue\n", res->lockname.len,
268 target = list_entry(res->converting.next, struct dlm_lock, list);
269 if (target->ml.convert_type == LKM_IVMODE) {
270 mlog(ML_ERROR, "%.*s: converting a lock with no "
271 "convert_type!\n", res->lockname.len, res->lockname.name);
274 head = &res->granted;
275 list_for_each(iter, head) {
276 lock = list_entry(iter, struct dlm_lock, list);
279 if (!dlm_lock_compatible(lock->ml.type,
280 target->ml.convert_type)) {
282 /* queue the BAST if not already */
283 if (lock->ml.highest_blocked == LKM_IVMODE) {
284 __dlm_lockres_reserve_ast(res);
285 dlm_queue_bast(dlm, lock);
287 /* update the highest_blocked if needed */
288 if (lock->ml.highest_blocked < target->ml.convert_type)
289 lock->ml.highest_blocked =
290 target->ml.convert_type;
293 head = &res->converting;
294 list_for_each(iter, head) {
295 lock = list_entry(iter, struct dlm_lock, list);
298 if (!dlm_lock_compatible(lock->ml.type,
299 target->ml.convert_type)) {
301 if (lock->ml.highest_blocked == LKM_IVMODE) {
302 __dlm_lockres_reserve_ast(res);
303 dlm_queue_bast(dlm, lock);
305 if (lock->ml.highest_blocked < target->ml.convert_type)
306 lock->ml.highest_blocked =
307 target->ml.convert_type;
311 /* we can convert the lock */
313 spin_lock(&target->spinlock);
314 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
316 mlog(0, "calling ast for converting lock: %.*s, have: %d, "
317 "granting: %d, node: %u\n", res->lockname.len,
318 res->lockname.name, target->ml.type,
319 target->ml.convert_type, target->ml.node);
321 target->ml.type = target->ml.convert_type;
322 target->ml.convert_type = LKM_IVMODE;
323 list_move_tail(&target->list, &res->granted);
325 BUG_ON(!target->lksb);
326 target->lksb->status = DLM_NORMAL;
328 spin_unlock(&target->spinlock);
330 __dlm_lockres_reserve_ast(res);
331 dlm_queue_ast(dlm, target);
332 /* go back and check for more */
337 if (list_empty(&res->blocked))
339 target = list_entry(res->blocked.next, struct dlm_lock, list);
341 head = &res->granted;
342 list_for_each(iter, head) {
343 lock = list_entry(iter, struct dlm_lock, list);
346 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
348 if (lock->ml.highest_blocked == LKM_IVMODE) {
349 __dlm_lockres_reserve_ast(res);
350 dlm_queue_bast(dlm, lock);
352 if (lock->ml.highest_blocked < target->ml.type)
353 lock->ml.highest_blocked = target->ml.type;
357 head = &res->converting;
358 list_for_each(iter, head) {
359 lock = list_entry(iter, struct dlm_lock, list);
362 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
364 if (lock->ml.highest_blocked == LKM_IVMODE) {
365 __dlm_lockres_reserve_ast(res);
366 dlm_queue_bast(dlm, lock);
368 if (lock->ml.highest_blocked < target->ml.type)
369 lock->ml.highest_blocked = target->ml.type;
373 /* we can grant the blocked lock (only
374 * possible if converting list empty) */
376 spin_lock(&target->spinlock);
377 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
379 mlog(0, "calling ast for blocked lock: %.*s, granting: %d, "
380 "node: %u\n", res->lockname.len, res->lockname.name,
381 target->ml.type, target->ml.node);
383 // target->ml.type is already correct
384 list_move_tail(&target->list, &res->granted);
386 BUG_ON(!target->lksb);
387 target->lksb->status = DLM_NORMAL;
389 spin_unlock(&target->spinlock);
391 __dlm_lockres_reserve_ast(res);
392 dlm_queue_ast(dlm, target);
393 /* go back and check for more */
401 /* must have NO locks when calling this with res !=NULL * */
402 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
404 mlog_entry("dlm=%p, res=%p\n", dlm, res);
406 spin_lock(&dlm->spinlock);
407 spin_lock(&res->spinlock);
408 __dlm_dirty_lockres(dlm, res);
409 spin_unlock(&res->spinlock);
410 spin_unlock(&dlm->spinlock);
412 wake_up(&dlm->dlm_thread_wq);
415 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
417 mlog_entry("dlm=%p, res=%p\n", dlm, res);
419 assert_spin_locked(&dlm->spinlock);
420 assert_spin_locked(&res->spinlock);
422 /* don't shuffle secondary queues */
423 if ((res->owner == dlm->node_num) &&
424 !(res->state & DLM_LOCK_RES_DIRTY)) {
425 list_add_tail(&res->dirty, &dlm->dirty_list);
426 res->state |= DLM_LOCK_RES_DIRTY;
431 /* Launch the NM thread for the mounted volume */
432 int dlm_launch_thread(struct dlm_ctxt *dlm)
434 mlog(0, "starting dlm thread...\n");
436 dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
437 if (IS_ERR(dlm->dlm_thread_task)) {
438 mlog_errno(PTR_ERR(dlm->dlm_thread_task));
439 dlm->dlm_thread_task = NULL;
446 void dlm_complete_thread(struct dlm_ctxt *dlm)
448 if (dlm->dlm_thread_task) {
449 mlog(ML_KTHREAD, "waiting for dlm thread to exit\n");
450 kthread_stop(dlm->dlm_thread_task);
451 dlm->dlm_thread_task = NULL;
455 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
459 spin_lock(&dlm->spinlock);
460 empty = list_empty(&dlm->dirty_list);
461 spin_unlock(&dlm->spinlock);
466 static void dlm_flush_asts(struct dlm_ctxt *dlm)
469 struct dlm_lock *lock;
470 struct dlm_lock_resource *res;
473 spin_lock(&dlm->ast_lock);
474 while (!list_empty(&dlm->pending_asts)) {
475 lock = list_entry(dlm->pending_asts.next,
476 struct dlm_lock, ast_list);
477 /* get an extra ref on lock */
480 mlog(0, "delivering an ast for this lockres\n");
482 BUG_ON(!lock->ast_pending);
484 /* remove from list (including ref) */
485 list_del_init(&lock->ast_list);
487 spin_unlock(&dlm->ast_lock);
489 if (lock->ml.node != dlm->node_num) {
490 ret = dlm_do_remote_ast(dlm, res, lock);
494 dlm_do_local_ast(dlm, res, lock);
496 spin_lock(&dlm->ast_lock);
498 /* possible that another ast was queued while
499 * we were delivering the last one */
500 if (!list_empty(&lock->ast_list)) {
501 mlog(0, "aha another ast got queued while "
502 "we were finishing the last one. will "
503 "keep the ast_pending flag set.\n");
505 lock->ast_pending = 0;
507 /* drop the extra ref.
508 * this may drop it completely. */
510 dlm_lockres_release_ast(dlm, res);
513 while (!list_empty(&dlm->pending_basts)) {
514 lock = list_entry(dlm->pending_basts.next,
515 struct dlm_lock, bast_list);
516 /* get an extra ref on lock */
520 BUG_ON(!lock->bast_pending);
522 /* get the highest blocked lock, and reset */
523 spin_lock(&lock->spinlock);
524 BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
525 hi = lock->ml.highest_blocked;
526 lock->ml.highest_blocked = LKM_IVMODE;
527 spin_unlock(&lock->spinlock);
529 /* remove from list (including ref) */
530 list_del_init(&lock->bast_list);
532 spin_unlock(&dlm->ast_lock);
534 mlog(0, "delivering a bast for this lockres "
535 "(blocked = %d\n", hi);
537 if (lock->ml.node != dlm->node_num) {
538 ret = dlm_send_proxy_bast(dlm, res, lock, hi);
542 dlm_do_local_bast(dlm, res, lock, hi);
544 spin_lock(&dlm->ast_lock);
546 /* possible that another bast was queued while
547 * we were delivering the last one */
548 if (!list_empty(&lock->bast_list)) {
549 mlog(0, "aha another bast got queued while "
550 "we were finishing the last one. will "
551 "keep the bast_pending flag set.\n");
553 lock->bast_pending = 0;
555 /* drop the extra ref.
556 * this may drop it completely. */
558 dlm_lockres_release_ast(dlm, res);
560 wake_up(&dlm->ast_wq);
561 spin_unlock(&dlm->ast_lock);
565 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
566 #define DLM_THREAD_MAX_DIRTY 100
567 #define DLM_THREAD_MAX_ASTS 10
569 static int dlm_thread(void *data)
571 struct dlm_lock_resource *res;
572 struct dlm_ctxt *dlm = data;
573 unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
575 mlog(0, "dlm thread running for %s...\n", dlm->name);
577 while (!kthread_should_stop()) {
578 int n = DLM_THREAD_MAX_DIRTY;
580 /* dlm_shutting_down is very point-in-time, but that
581 * doesn't matter as we'll just loop back around if we
582 * get false on the leading edge of a state
584 dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
586 /* We really don't want to hold dlm->spinlock while
587 * calling dlm_shuffle_lists on each lockres that
588 * needs to have its queues adjusted and AST/BASTs
589 * run. So let's pull each entry off the dirty_list
590 * and drop dlm->spinlock ASAP. Once off the list,
591 * res->spinlock needs to be taken again to protect
592 * the queues while calling dlm_shuffle_lists. */
593 spin_lock(&dlm->spinlock);
594 while (!list_empty(&dlm->dirty_list)) {
596 res = list_entry(dlm->dirty_list.next,
597 struct dlm_lock_resource, dirty);
599 /* peel a lockres off, remove it from the list,
600 * unset the dirty flag and drop the dlm lock */
602 dlm_lockres_get(res);
604 spin_lock(&res->spinlock);
605 res->state &= ~DLM_LOCK_RES_DIRTY;
606 list_del_init(&res->dirty);
607 spin_unlock(&res->spinlock);
608 spin_unlock(&dlm->spinlock);
610 /* lockres can be re-dirtied/re-added to the
611 * dirty_list in this gap, but that is ok */
613 spin_lock(&res->spinlock);
614 if (res->owner != dlm->node_num) {
615 __dlm_print_one_lock_resource(res);
616 mlog(ML_ERROR, "inprog:%s, mig:%s, reco:%s, dirty:%s\n",
617 res->state & DLM_LOCK_RES_IN_PROGRESS ? "yes" : "no",
618 res->state & DLM_LOCK_RES_MIGRATING ? "yes" : "no",
619 res->state & DLM_LOCK_RES_RECOVERING ? "yes" : "no",
620 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
622 BUG_ON(res->owner != dlm->node_num);
624 /* it is now ok to move lockreses in these states
625 * to the dirty list, assuming that they will only be
626 * dirty for a short while. */
627 if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
628 DLM_LOCK_RES_MIGRATING |
629 DLM_LOCK_RES_RECOVERING)) {
630 /* move it to the tail and keep going */
631 spin_unlock(&res->spinlock);
632 mlog(0, "delaying list shuffling for in-"
633 "progress lockres %.*s, state=%d\n",
634 res->lockname.len, res->lockname.name,
640 /* at this point the lockres is not migrating/
641 * recovering/in-progress. we have the lockres
642 * spinlock and do NOT have the dlm lock.
643 * safe to reserve/queue asts and run the lists. */
645 mlog(0, "calling dlm_shuffle_lists with dlm=%s, "
646 "res=%.*s\n", dlm->name,
647 res->lockname.len, res->lockname.name);
649 /* called while holding lockres lock */
650 dlm_shuffle_lists(dlm, res);
651 spin_unlock(&res->spinlock);
653 dlm_lockres_calc_usage(dlm, res);
657 spin_lock(&dlm->spinlock);
658 /* if the lock was in-progress, stick
659 * it on the back of the list */
661 spin_lock(&res->spinlock);
662 list_add_tail(&res->dirty, &dlm->dirty_list);
663 res->state |= DLM_LOCK_RES_DIRTY;
664 spin_unlock(&res->spinlock);
666 dlm_lockres_put(res);
668 /* unlikely, but we may need to give time to
671 mlog(0, "throttling dlm_thread\n");
676 spin_unlock(&dlm->spinlock);
679 /* yield and continue right away if there is more work to do */
685 wait_event_interruptible_timeout(dlm->dlm_thread_wq,
686 !dlm_dirty_list_empty(dlm) ||
687 kthread_should_stop(),
691 mlog(0, "quitting DLM thread\n");