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/highmem.h>
32 #include <linux/init.h>
33 #include <linux/sysctl.h>
34 #include <linux/random.h>
35 #include <linux/blkdev.h>
36 #include <linux/socket.h>
37 #include <linux/inet.h>
38 #include <linux/timer.h>
39 #include <linux/kthread.h>
40 #include <linux/delay.h>
43 #include "cluster/heartbeat.h"
44 #include "cluster/nodemanager.h"
45 #include "cluster/tcp.h"
48 #include "dlmcommon.h"
49 #include "dlmdomain.h"
51 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
52 #include "cluster/masklog.h"
54 static int dlm_thread(void *data);
55 static void dlm_flush_asts(struct dlm_ctxt *dlm);
57 #define dlm_lock_is_remote(dlm, lock) ((lock)->ml.node != (dlm)->node_num)
59 /* will exit holding res->spinlock, but may drop in function */
60 /* waits until flags are cleared on res->state */
61 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
63 DECLARE_WAITQUEUE(wait, current);
65 assert_spin_locked(&res->spinlock);
67 add_wait_queue(&res->wq, &wait);
69 set_current_state(TASK_UNINTERRUPTIBLE);
70 if (res->state & flags) {
71 spin_unlock(&res->spinlock);
73 spin_lock(&res->spinlock);
76 remove_wait_queue(&res->wq, &wait);
77 __set_current_state(TASK_RUNNING);
80 int __dlm_lockres_has_locks(struct dlm_lock_resource *res)
82 if (list_empty(&res->granted) &&
83 list_empty(&res->converting) &&
84 list_empty(&res->blocked))
89 /* "unused": the lockres has no locks, is not on the dirty list,
90 * has no inflight locks (in the gap between mastery and acquiring
91 * the first lock), and has no bits in its refmap.
92 * truly ready to be freed. */
93 int __dlm_lockres_unused(struct dlm_lock_resource *res)
97 if (__dlm_lockres_has_locks(res))
100 if (!list_empty(&res->dirty) || res->state & DLM_LOCK_RES_DIRTY)
103 if (res->state & DLM_LOCK_RES_RECOVERING)
106 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
107 if (bit < O2NM_MAX_NODES)
111 * since the bit for dlm->node_num is not set, inflight_locks better
114 BUG_ON(res->inflight_locks != 0);
119 /* Call whenever you may have added or deleted something from one of
120 * the lockres queue's. This will figure out whether it belongs on the
121 * unused list or not and does the appropriate thing. */
122 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
123 struct dlm_lock_resource *res)
125 assert_spin_locked(&dlm->spinlock);
126 assert_spin_locked(&res->spinlock);
128 if (__dlm_lockres_unused(res)){
129 if (list_empty(&res->purge)) {
130 mlog(0, "%s: Adding res %.*s to purge list\n",
131 dlm->name, res->lockname.len, res->lockname.name);
133 res->last_used = jiffies;
134 dlm_lockres_get(res);
135 list_add_tail(&res->purge, &dlm->purge_list);
138 } else if (!list_empty(&res->purge)) {
139 mlog(0, "%s: Removing res %.*s from purge list\n",
140 dlm->name, res->lockname.len, res->lockname.name);
142 list_del_init(&res->purge);
143 dlm_lockres_put(res);
148 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
149 struct dlm_lock_resource *res)
151 spin_lock(&dlm->spinlock);
152 spin_lock(&res->spinlock);
154 __dlm_lockres_calc_usage(dlm, res);
156 spin_unlock(&res->spinlock);
157 spin_unlock(&dlm->spinlock);
160 static void dlm_purge_lockres(struct dlm_ctxt *dlm,
161 struct dlm_lock_resource *res)
166 assert_spin_locked(&dlm->spinlock);
167 assert_spin_locked(&res->spinlock);
169 master = (res->owner == dlm->node_num);
171 mlog(0, "%s: Purging res %.*s, master %d\n", dlm->name,
172 res->lockname.len, res->lockname.name, master);
175 res->state |= DLM_LOCK_RES_DROPPING_REF;
176 /* drop spinlock... retake below */
177 spin_unlock(&res->spinlock);
178 spin_unlock(&dlm->spinlock);
180 spin_lock(&res->spinlock);
181 /* This ensures that clear refmap is sent after the set */
182 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
183 spin_unlock(&res->spinlock);
185 /* clear our bit from the master's refmap, ignore errors */
186 ret = dlm_drop_lockres_ref(dlm, res);
188 if (!dlm_is_host_down(ret))
191 spin_lock(&dlm->spinlock);
192 spin_lock(&res->spinlock);
195 if (!list_empty(&res->purge)) {
196 mlog(0, "%s: Removing res %.*s from purgelist, master %d\n",
197 dlm->name, res->lockname.len, res->lockname.name, master);
198 list_del_init(&res->purge);
199 dlm_lockres_put(res);
203 if (!__dlm_lockres_unused(res)) {
204 mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
205 dlm->name, res->lockname.len, res->lockname.name);
206 __dlm_print_one_lock_resource(res);
210 __dlm_unhash_lockres(dlm, res);
212 /* lockres is not in the hash now. drop the flag and wake up
213 * any processes waiting in dlm_get_lock_resource. */
215 res->state &= ~DLM_LOCK_RES_DROPPING_REF;
216 spin_unlock(&res->spinlock);
219 spin_unlock(&res->spinlock);
222 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
225 unsigned int run_max, unused;
226 unsigned long purge_jiffies;
227 struct dlm_lock_resource *lockres;
229 spin_lock(&dlm->spinlock);
230 run_max = dlm->purge_count;
232 while(run_max && !list_empty(&dlm->purge_list)) {
235 lockres = list_entry(dlm->purge_list.next,
236 struct dlm_lock_resource, purge);
238 spin_lock(&lockres->spinlock);
240 purge_jiffies = lockres->last_used +
241 msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
243 /* Make sure that we want to be processing this guy at
245 if (!purge_now && time_after(purge_jiffies, jiffies)) {
246 /* Since resources are added to the purge list
247 * in tail order, we can stop at the first
248 * unpurgable resource -- anyone added after
249 * him will have a greater last_used value */
250 spin_unlock(&lockres->spinlock);
254 /* Status of the lockres *might* change so double
255 * check. If the lockres is unused, holding the dlm
256 * spinlock will prevent people from getting and more
258 unused = __dlm_lockres_unused(lockres);
260 (lockres->state & DLM_LOCK_RES_MIGRATING)) {
261 mlog(0, "%s: res %.*s is in use or being remastered, "
262 "used %d, state %d\n", dlm->name,
263 lockres->lockname.len, lockres->lockname.name,
264 !unused, lockres->state);
265 list_move_tail(&dlm->purge_list, &lockres->purge);
266 spin_unlock(&lockres->spinlock);
270 dlm_lockres_get(lockres);
272 dlm_purge_lockres(dlm, lockres);
274 dlm_lockres_put(lockres);
276 /* Avoid adding any scheduling latencies */
277 cond_resched_lock(&dlm->spinlock);
280 spin_unlock(&dlm->spinlock);
283 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
284 struct dlm_lock_resource *res)
286 struct dlm_lock *lock, *target;
287 struct list_head *iter;
288 struct list_head *head;
292 * Because this function is called with the lockres
293 * spinlock, and because we know that it is not migrating/
294 * recovering/in-progress, it is fine to reserve asts and
295 * basts right before queueing them all throughout
297 assert_spin_locked(&dlm->ast_lock);
298 assert_spin_locked(&res->spinlock);
299 BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
300 DLM_LOCK_RES_RECOVERING|
301 DLM_LOCK_RES_IN_PROGRESS)));
304 if (list_empty(&res->converting))
306 mlog(0, "%s: res %.*s has locks on the convert queue\n", dlm->name,
307 res->lockname.len, res->lockname.name);
309 target = list_entry(res->converting.next, struct dlm_lock, list);
310 if (target->ml.convert_type == LKM_IVMODE) {
311 mlog(ML_ERROR, "%s: res %.*s converting lock to invalid mode\n",
312 dlm->name, res->lockname.len, res->lockname.name);
315 head = &res->granted;
316 list_for_each(iter, head) {
317 lock = list_entry(iter, struct dlm_lock, list);
320 if (!dlm_lock_compatible(lock->ml.type,
321 target->ml.convert_type)) {
323 /* queue the BAST if not already */
324 if (lock->ml.highest_blocked == LKM_IVMODE) {
325 __dlm_lockres_reserve_ast(res);
326 __dlm_queue_bast(dlm, lock);
328 /* update the highest_blocked if needed */
329 if (lock->ml.highest_blocked < target->ml.convert_type)
330 lock->ml.highest_blocked =
331 target->ml.convert_type;
334 head = &res->converting;
335 list_for_each(iter, head) {
336 lock = list_entry(iter, struct dlm_lock, list);
339 if (!dlm_lock_compatible(lock->ml.type,
340 target->ml.convert_type)) {
342 if (lock->ml.highest_blocked == LKM_IVMODE) {
343 __dlm_lockres_reserve_ast(res);
344 __dlm_queue_bast(dlm, lock);
346 if (lock->ml.highest_blocked < target->ml.convert_type)
347 lock->ml.highest_blocked =
348 target->ml.convert_type;
352 /* we can convert the lock */
354 spin_lock(&target->spinlock);
355 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
357 mlog(0, "%s: res %.*s, AST for Converting lock %u:%llu, type "
358 "%d => %d, node %u\n", dlm->name, res->lockname.len,
360 dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
361 dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
363 target->ml.convert_type, target->ml.node);
365 target->ml.type = target->ml.convert_type;
366 target->ml.convert_type = LKM_IVMODE;
367 list_move_tail(&target->list, &res->granted);
369 BUG_ON(!target->lksb);
370 target->lksb->status = DLM_NORMAL;
372 spin_unlock(&target->spinlock);
374 __dlm_lockres_reserve_ast(res);
375 __dlm_queue_ast(dlm, target);
376 /* go back and check for more */
381 if (list_empty(&res->blocked))
383 target = list_entry(res->blocked.next, struct dlm_lock, list);
385 head = &res->granted;
386 list_for_each(iter, head) {
387 lock = list_entry(iter, struct dlm_lock, list);
390 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
392 if (lock->ml.highest_blocked == LKM_IVMODE) {
393 __dlm_lockres_reserve_ast(res);
394 __dlm_queue_bast(dlm, lock);
396 if (lock->ml.highest_blocked < target->ml.type)
397 lock->ml.highest_blocked = target->ml.type;
401 head = &res->converting;
402 list_for_each(iter, head) {
403 lock = list_entry(iter, struct dlm_lock, list);
406 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
408 if (lock->ml.highest_blocked == LKM_IVMODE) {
409 __dlm_lockres_reserve_ast(res);
410 __dlm_queue_bast(dlm, lock);
412 if (lock->ml.highest_blocked < target->ml.type)
413 lock->ml.highest_blocked = target->ml.type;
417 /* we can grant the blocked lock (only
418 * possible if converting list empty) */
420 spin_lock(&target->spinlock);
421 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
423 mlog(0, "%s: res %.*s, AST for Blocked lock %u:%llu, type %d, "
424 "node %u\n", dlm->name, res->lockname.len,
426 dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
427 dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
428 target->ml.type, target->ml.node);
430 /* target->ml.type is already correct */
431 list_move_tail(&target->list, &res->granted);
433 BUG_ON(!target->lksb);
434 target->lksb->status = DLM_NORMAL;
436 spin_unlock(&target->spinlock);
438 __dlm_lockres_reserve_ast(res);
439 __dlm_queue_ast(dlm, target);
440 /* go back and check for more */
448 /* must have NO locks when calling this with res !=NULL * */
449 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
452 spin_lock(&dlm->spinlock);
453 spin_lock(&res->spinlock);
454 __dlm_dirty_lockres(dlm, res);
455 spin_unlock(&res->spinlock);
456 spin_unlock(&dlm->spinlock);
458 wake_up(&dlm->dlm_thread_wq);
461 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
463 assert_spin_locked(&dlm->spinlock);
464 assert_spin_locked(&res->spinlock);
466 /* don't shuffle secondary queues */
467 if ((res->owner == dlm->node_num)) {
468 if (res->state & (DLM_LOCK_RES_MIGRATING |
469 DLM_LOCK_RES_BLOCK_DIRTY))
472 if (list_empty(&res->dirty)) {
473 /* ref for dirty_list */
474 dlm_lockres_get(res);
475 list_add_tail(&res->dirty, &dlm->dirty_list);
476 res->state |= DLM_LOCK_RES_DIRTY;
480 mlog(0, "%s: res %.*s\n", dlm->name, res->lockname.len,
485 /* Launch the NM thread for the mounted volume */
486 int dlm_launch_thread(struct dlm_ctxt *dlm)
488 mlog(0, "Starting dlm_thread...\n");
490 dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
491 if (IS_ERR(dlm->dlm_thread_task)) {
492 mlog_errno(PTR_ERR(dlm->dlm_thread_task));
493 dlm->dlm_thread_task = NULL;
500 void dlm_complete_thread(struct dlm_ctxt *dlm)
502 if (dlm->dlm_thread_task) {
503 mlog(ML_KTHREAD, "Waiting for dlm thread to exit\n");
504 kthread_stop(dlm->dlm_thread_task);
505 dlm->dlm_thread_task = NULL;
509 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
513 spin_lock(&dlm->spinlock);
514 empty = list_empty(&dlm->dirty_list);
515 spin_unlock(&dlm->spinlock);
520 static void dlm_flush_asts(struct dlm_ctxt *dlm)
523 struct dlm_lock *lock;
524 struct dlm_lock_resource *res;
527 spin_lock(&dlm->ast_lock);
528 while (!list_empty(&dlm->pending_asts)) {
529 lock = list_entry(dlm->pending_asts.next,
530 struct dlm_lock, ast_list);
531 /* get an extra ref on lock */
534 mlog(0, "%s: res %.*s, Flush AST for lock %u:%llu, type %d, "
535 "node %u\n", dlm->name, res->lockname.len,
537 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
538 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
539 lock->ml.type, lock->ml.node);
541 BUG_ON(!lock->ast_pending);
543 /* remove from list (including ref) */
544 list_del_init(&lock->ast_list);
546 spin_unlock(&dlm->ast_lock);
548 if (lock->ml.node != dlm->node_num) {
549 ret = dlm_do_remote_ast(dlm, res, lock);
553 dlm_do_local_ast(dlm, res, lock);
555 spin_lock(&dlm->ast_lock);
557 /* possible that another ast was queued while
558 * we were delivering the last one */
559 if (!list_empty(&lock->ast_list)) {
560 mlog(0, "%s: res %.*s, AST queued while flushing last "
561 "one\n", dlm->name, res->lockname.len,
564 lock->ast_pending = 0;
566 /* drop the extra ref.
567 * this may drop it completely. */
569 dlm_lockres_release_ast(dlm, res);
572 while (!list_empty(&dlm->pending_basts)) {
573 lock = list_entry(dlm->pending_basts.next,
574 struct dlm_lock, bast_list);
575 /* get an extra ref on lock */
579 BUG_ON(!lock->bast_pending);
581 /* get the highest blocked lock, and reset */
582 spin_lock(&lock->spinlock);
583 BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
584 hi = lock->ml.highest_blocked;
585 lock->ml.highest_blocked = LKM_IVMODE;
586 spin_unlock(&lock->spinlock);
588 /* remove from list (including ref) */
589 list_del_init(&lock->bast_list);
591 spin_unlock(&dlm->ast_lock);
593 mlog(0, "%s: res %.*s, Flush BAST for lock %u:%llu, "
594 "blocked %d, node %u\n",
595 dlm->name, res->lockname.len, res->lockname.name,
596 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
597 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
600 if (lock->ml.node != dlm->node_num) {
601 ret = dlm_send_proxy_bast(dlm, res, lock, hi);
605 dlm_do_local_bast(dlm, res, lock, hi);
607 spin_lock(&dlm->ast_lock);
609 /* possible that another bast was queued while
610 * we were delivering the last one */
611 if (!list_empty(&lock->bast_list)) {
612 mlog(0, "%s: res %.*s, BAST queued while flushing last "
613 "one\n", dlm->name, res->lockname.len,
616 lock->bast_pending = 0;
618 /* drop the extra ref.
619 * this may drop it completely. */
621 dlm_lockres_release_ast(dlm, res);
623 wake_up(&dlm->ast_wq);
624 spin_unlock(&dlm->ast_lock);
628 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
629 #define DLM_THREAD_MAX_DIRTY 100
630 #define DLM_THREAD_MAX_ASTS 10
632 static int dlm_thread(void *data)
634 struct dlm_lock_resource *res;
635 struct dlm_ctxt *dlm = data;
636 unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
638 mlog(0, "dlm thread running for %s...\n", dlm->name);
640 while (!kthread_should_stop()) {
641 int n = DLM_THREAD_MAX_DIRTY;
643 /* dlm_shutting_down is very point-in-time, but that
644 * doesn't matter as we'll just loop back around if we
645 * get false on the leading edge of a state
647 dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
649 /* We really don't want to hold dlm->spinlock while
650 * calling dlm_shuffle_lists on each lockres that
651 * needs to have its queues adjusted and AST/BASTs
652 * run. So let's pull each entry off the dirty_list
653 * and drop dlm->spinlock ASAP. Once off the list,
654 * res->spinlock needs to be taken again to protect
655 * the queues while calling dlm_shuffle_lists. */
656 spin_lock(&dlm->spinlock);
657 while (!list_empty(&dlm->dirty_list)) {
659 res = list_entry(dlm->dirty_list.next,
660 struct dlm_lock_resource, dirty);
662 /* peel a lockres off, remove it from the list,
663 * unset the dirty flag and drop the dlm lock */
665 dlm_lockres_get(res);
667 spin_lock(&res->spinlock);
668 /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
669 list_del_init(&res->dirty);
670 spin_unlock(&res->spinlock);
671 spin_unlock(&dlm->spinlock);
672 /* Drop dirty_list ref */
673 dlm_lockres_put(res);
675 /* lockres can be re-dirtied/re-added to the
676 * dirty_list in this gap, but that is ok */
678 spin_lock(&dlm->ast_lock);
679 spin_lock(&res->spinlock);
680 if (res->owner != dlm->node_num) {
681 __dlm_print_one_lock_resource(res);
682 mlog(ML_ERROR, "%s: inprog %d, mig %d, reco %d,"
683 " dirty %d\n", dlm->name,
684 !!(res->state & DLM_LOCK_RES_IN_PROGRESS),
685 !!(res->state & DLM_LOCK_RES_MIGRATING),
686 !!(res->state & DLM_LOCK_RES_RECOVERING),
687 !!(res->state & DLM_LOCK_RES_DIRTY));
689 BUG_ON(res->owner != dlm->node_num);
691 /* it is now ok to move lockreses in these states
692 * to the dirty list, assuming that they will only be
693 * dirty for a short while. */
694 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
695 if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
696 DLM_LOCK_RES_RECOVERING)) {
697 /* move it to the tail and keep going */
698 res->state &= ~DLM_LOCK_RES_DIRTY;
699 spin_unlock(&res->spinlock);
700 spin_unlock(&dlm->ast_lock);
701 mlog(0, "%s: res %.*s, inprogress, delay list "
702 "shuffle, state %d\n", dlm->name,
703 res->lockname.len, res->lockname.name,
709 /* at this point the lockres is not migrating/
710 * recovering/in-progress. we have the lockres
711 * spinlock and do NOT have the dlm lock.
712 * safe to reserve/queue asts and run the lists. */
714 /* called while holding lockres lock */
715 dlm_shuffle_lists(dlm, res);
716 res->state &= ~DLM_LOCK_RES_DIRTY;
717 spin_unlock(&res->spinlock);
718 spin_unlock(&dlm->ast_lock);
720 dlm_lockres_calc_usage(dlm, res);
724 spin_lock(&dlm->spinlock);
725 /* if the lock was in-progress, stick
726 * it on the back of the list */
728 spin_lock(&res->spinlock);
729 __dlm_dirty_lockres(dlm, res);
730 spin_unlock(&res->spinlock);
732 dlm_lockres_put(res);
734 /* unlikely, but we may need to give time to
737 mlog(0, "%s: Throttling dlm thread\n",
743 spin_unlock(&dlm->spinlock);
746 /* yield and continue right away if there is more work to do */
752 wait_event_interruptible_timeout(dlm->dlm_thread_wq,
753 !dlm_dirty_list_empty(dlm) ||
754 kthread_should_stop(),
758 mlog(0, "quitting DLM thread\n");