1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
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/init.h>
34 #include <linux/sysctl.h>
35 #include <linux/random.h>
36 #include <linux/blkdev.h>
37 #include <linux/socket.h>
38 #include <linux/inet.h>
39 #include <linux/timer.h>
40 #include <linux/kthread.h>
41 #include <linux/delay.h>
44 #include "cluster/heartbeat.h"
45 #include "cluster/nodemanager.h"
46 #include "cluster/tcp.h"
49 #include "dlmcommon.h"
50 #include "dlmdomain.h"
52 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY)
53 #include "cluster/masklog.h"
55 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node);
57 static int dlm_recovery_thread(void *data);
58 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
59 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
60 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm);
61 static int dlm_do_recovery(struct dlm_ctxt *dlm);
63 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm);
64 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node);
65 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
66 static int dlm_request_all_locks(struct dlm_ctxt *dlm,
67 u8 request_from, u8 dead_node);
68 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
70 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res);
71 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
72 const char *lockname, int namelen,
73 int total_locks, u64 cookie,
75 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
76 struct dlm_migratable_lockres *mres,
78 struct dlm_lock_resource *res,
80 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
81 struct dlm_lock_resource *res,
82 struct dlm_migratable_lockres *mres);
83 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm);
84 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm,
85 u8 dead_node, u8 send_to);
86 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node);
87 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
88 struct list_head *list, u8 dead_node);
89 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
90 u8 dead_node, u8 new_master);
91 static void dlm_reco_ast(void *astdata);
92 static void dlm_reco_bast(void *astdata, int blocked_type);
93 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st);
94 static void dlm_request_all_locks_worker(struct dlm_work_item *item,
96 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data);
97 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
98 struct dlm_lock_resource *res,
101 static u64 dlm_get_next_mig_cookie(void);
103 static DEFINE_SPINLOCK(dlm_reco_state_lock);
104 static DEFINE_SPINLOCK(dlm_mig_cookie_lock);
105 static u64 dlm_mig_cookie = 1;
107 static u64 dlm_get_next_mig_cookie(void)
110 spin_lock(&dlm_mig_cookie_lock);
112 if (dlm_mig_cookie == (~0ULL))
116 spin_unlock(&dlm_mig_cookie_lock);
120 static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm,
123 assert_spin_locked(&dlm->spinlock);
124 if (dlm->reco.dead_node != dead_node)
125 mlog(0, "%s: changing dead_node from %u to %u\n",
126 dlm->name, dlm->reco.dead_node, dead_node);
127 dlm->reco.dead_node = dead_node;
130 static inline void dlm_set_reco_master(struct dlm_ctxt *dlm,
133 assert_spin_locked(&dlm->spinlock);
134 mlog(0, "%s: changing new_master from %u to %u\n",
135 dlm->name, dlm->reco.new_master, master);
136 dlm->reco.new_master = master;
139 static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm)
141 assert_spin_locked(&dlm->spinlock);
142 clear_bit(dlm->reco.dead_node, dlm->recovery_map);
143 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
144 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
147 static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
149 spin_lock(&dlm->spinlock);
150 __dlm_reset_recovery(dlm);
151 spin_unlock(&dlm->spinlock);
154 /* Worker function used during recovery. */
155 void dlm_dispatch_work(struct work_struct *work)
157 struct dlm_ctxt *dlm =
158 container_of(work, struct dlm_ctxt, dispatched_work);
160 struct dlm_work_item *item, *next;
161 dlm_workfunc_t *workfunc;
164 spin_lock(&dlm->work_lock);
165 list_splice_init(&dlm->work_list, &tmp_list);
166 spin_unlock(&dlm->work_lock);
168 list_for_each_entry(item, &tmp_list, list) {
171 mlog(0, "%s: work thread has %d work items\n", dlm->name, tot);
173 list_for_each_entry_safe(item, next, &tmp_list, list) {
174 workfunc = item->func;
175 list_del_init(&item->list);
177 /* already have ref on dlm to avoid having
178 * it disappear. just double-check. */
179 BUG_ON(item->dlm != dlm);
181 /* this is allowed to sleep and
182 * call network stuff */
183 workfunc(item, item->data);
194 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
196 /* wake the recovery thread
197 * this will wake the reco thread in one of three places
198 * 1) sleeping with no recovery happening
199 * 2) sleeping with recovery mastered elsewhere
200 * 3) recovery mastered here, waiting on reco data */
202 wake_up(&dlm->dlm_reco_thread_wq);
205 /* Launch the recovery thread */
206 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
208 mlog(0, "starting dlm recovery thread...\n");
210 dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
212 if (IS_ERR(dlm->dlm_reco_thread_task)) {
213 mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
214 dlm->dlm_reco_thread_task = NULL;
221 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
223 if (dlm->dlm_reco_thread_task) {
224 mlog(0, "waiting for dlm recovery thread to exit\n");
225 kthread_stop(dlm->dlm_reco_thread_task);
226 dlm->dlm_reco_thread_task = NULL;
233 * this is lame, but here's how recovery works...
234 * 1) all recovery threads cluster wide will work on recovering
236 * 2) negotiate who will take over all the locks for the dead node.
237 * thats right... ALL the locks.
238 * 3) once a new master is chosen, everyone scans all locks
239 * and moves aside those mastered by the dead guy
240 * 4) each of these locks should be locked until recovery is done
241 * 5) the new master collects up all of secondary lock queue info
242 * one lock at a time, forcing each node to communicate back
244 * 6) each secondary lock queue responds with the full known lock info
245 * 7) once the new master has run all its locks, it sends a ALLDONE!
246 * message to everyone
247 * 8) upon receiving this message, the secondary queue node unlocks
248 * and responds to the ALLDONE
249 * 9) once the new master gets responses from everyone, he unlocks
250 * everything and recovery for this dead node is done
251 *10) go back to 2) while there are still dead nodes
255 static void dlm_print_reco_node_status(struct dlm_ctxt *dlm)
257 struct dlm_reco_node_data *ndata;
258 struct dlm_lock_resource *res;
260 mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n",
261 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
262 dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive",
263 dlm->reco.dead_node, dlm->reco.new_master);
265 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
266 char *st = "unknown";
267 switch (ndata->state) {
268 case DLM_RECO_NODE_DATA_INIT:
271 case DLM_RECO_NODE_DATA_REQUESTING:
274 case DLM_RECO_NODE_DATA_DEAD:
277 case DLM_RECO_NODE_DATA_RECEIVING:
280 case DLM_RECO_NODE_DATA_REQUESTED:
283 case DLM_RECO_NODE_DATA_DONE:
286 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
287 st = "finalize-sent";
293 mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n",
294 dlm->name, ndata->node_num, st);
296 list_for_each_entry(res, &dlm->reco.resources, recovering) {
297 mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n",
298 dlm->name, res->lockname.len, res->lockname.name);
302 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
304 static int dlm_recovery_thread(void *data)
307 struct dlm_ctxt *dlm = data;
308 unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
310 mlog(0, "dlm thread running for %s...\n", dlm->name);
312 while (!kthread_should_stop()) {
313 if (dlm_domain_fully_joined(dlm)) {
314 status = dlm_do_recovery(dlm);
315 if (status == -EAGAIN) {
316 /* do not sleep, recheck immediately. */
323 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
324 kthread_should_stop(),
328 mlog(0, "quitting DLM recovery thread\n");
332 /* returns true when the recovery master has contacted us */
333 static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
336 spin_lock(&dlm->spinlock);
337 ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
338 spin_unlock(&dlm->spinlock);
342 /* returns true if node is no longer in the domain
343 * could be dead or just not joined */
344 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
347 spin_lock(&dlm->spinlock);
348 dead = !test_bit(node, dlm->domain_map);
349 spin_unlock(&dlm->spinlock);
353 /* returns true if node is no longer in the domain
354 * could be dead or just not joined */
355 static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node)
358 spin_lock(&dlm->spinlock);
359 recovered = !test_bit(node, dlm->recovery_map);
360 spin_unlock(&dlm->spinlock);
365 void dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout)
367 if (dlm_is_node_dead(dlm, node))
370 printk(KERN_NOTICE "o2dlm: Waiting on the death of node %u in "
371 "domain %s\n", node, dlm->name);
374 wait_event_timeout(dlm->dlm_reco_thread_wq,
375 dlm_is_node_dead(dlm, node),
376 msecs_to_jiffies(timeout));
378 wait_event(dlm->dlm_reco_thread_wq,
379 dlm_is_node_dead(dlm, node));
382 void dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout)
384 if (dlm_is_node_recovered(dlm, node))
387 printk(KERN_NOTICE "o2dlm: Waiting on the recovery of node %u in "
388 "domain %s\n", node, dlm->name);
391 wait_event_timeout(dlm->dlm_reco_thread_wq,
392 dlm_is_node_recovered(dlm, node),
393 msecs_to_jiffies(timeout));
395 wait_event(dlm->dlm_reco_thread_wq,
396 dlm_is_node_recovered(dlm, node));
399 /* callers of the top-level api calls (dlmlock/dlmunlock) should
400 * block on the dlm->reco.event when recovery is in progress.
401 * the dlm recovery thread will set this state when it begins
402 * recovering a dead node (as the new master or not) and clear
403 * the state and wake as soon as all affected lock resources have
404 * been marked with the RECOVERY flag */
405 static int dlm_in_recovery(struct dlm_ctxt *dlm)
408 spin_lock(&dlm->spinlock);
409 in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
410 spin_unlock(&dlm->spinlock);
415 void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
417 if (dlm_in_recovery(dlm)) {
418 mlog(0, "%s: reco thread %d in recovery: "
419 "state=%d, master=%u, dead=%u\n",
420 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
421 dlm->reco.state, dlm->reco.new_master,
422 dlm->reco.dead_node);
424 wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
427 static void dlm_begin_recovery(struct dlm_ctxt *dlm)
429 spin_lock(&dlm->spinlock);
430 BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
431 printk(KERN_NOTICE "o2dlm: Begin recovery on domain %s for node %u\n",
432 dlm->name, dlm->reco.dead_node);
433 dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
434 spin_unlock(&dlm->spinlock);
437 static void dlm_end_recovery(struct dlm_ctxt *dlm)
439 spin_lock(&dlm->spinlock);
440 BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
441 dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
442 spin_unlock(&dlm->spinlock);
443 printk(KERN_NOTICE "o2dlm: End recovery on domain %s\n", dlm->name);
444 wake_up(&dlm->reco.event);
447 static void dlm_print_recovery_master(struct dlm_ctxt *dlm)
449 printk(KERN_NOTICE "o2dlm: Node %u (%s) is the Recovery Master for the "
450 "dead node %u in domain %s\n", dlm->reco.new_master,
451 (dlm->node_num == dlm->reco.new_master ? "me" : "he"),
452 dlm->reco.dead_node, dlm->name);
455 static int dlm_do_recovery(struct dlm_ctxt *dlm)
460 spin_lock(&dlm->spinlock);
462 /* check to see if the new master has died */
463 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
464 test_bit(dlm->reco.new_master, dlm->recovery_map)) {
465 mlog(0, "new master %u died while recovering %u!\n",
466 dlm->reco.new_master, dlm->reco.dead_node);
467 /* unset the new_master, leave dead_node */
468 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
471 /* select a target to recover */
472 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
475 bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES, 0);
476 if (bit >= O2NM_MAX_NODES || bit < 0)
477 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
479 dlm_set_reco_dead_node(dlm, bit);
480 } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
482 mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
483 dlm->reco.dead_node);
484 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
487 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
488 // mlog(0, "nothing to recover! sleeping now!\n");
489 spin_unlock(&dlm->spinlock);
490 /* return to main thread loop and sleep. */
493 mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n",
494 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
495 dlm->reco.dead_node);
496 spin_unlock(&dlm->spinlock);
498 /* take write barrier */
499 /* (stops the list reshuffling thread, proxy ast handling) */
500 dlm_begin_recovery(dlm);
502 if (dlm->reco.new_master == dlm->node_num)
505 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
506 /* choose a new master, returns 0 if this node
507 * is the master, -EEXIST if it's another node.
508 * this does not return until a new master is chosen
509 * or recovery completes entirely. */
510 ret = dlm_pick_recovery_master(dlm);
512 /* already notified everyone. go. */
515 mlog(0, "another node will master this recovery session.\n");
518 dlm_print_recovery_master(dlm);
520 /* it is safe to start everything back up here
521 * because all of the dead node's lock resources
522 * have been marked as in-recovery */
523 dlm_end_recovery(dlm);
525 /* sleep out in main dlm_recovery_thread loop. */
529 dlm_print_recovery_master(dlm);
531 status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
533 /* we should never hit this anymore */
534 mlog(ML_ERROR, "%s: Error %d remastering locks for node %u, "
535 "retrying.\n", dlm->name, status, dlm->reco.dead_node);
536 /* yield a bit to allow any final network messages
537 * to get handled on remaining nodes */
540 /* success! see if any other nodes need recovery */
541 mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
542 dlm->name, dlm->reco.dead_node, dlm->node_num);
543 spin_lock(&dlm->spinlock);
544 __dlm_reset_recovery(dlm);
545 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
546 spin_unlock(&dlm->spinlock);
548 dlm_end_recovery(dlm);
550 /* continue and look for another dead node */
554 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
557 struct dlm_reco_node_data *ndata;
563 /* we have become recovery master. there is no escaping
564 * this, so just keep trying until we get it. */
565 status = dlm_init_recovery_area(dlm, dead_node);
567 mlog(ML_ERROR, "%s: failed to alloc recovery area, "
568 "retrying\n", dlm->name);
571 } while (status != 0);
573 /* safe to access the node data list without a lock, since this
574 * process is the only one to change the list */
575 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
576 BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
577 ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
579 mlog(0, "%s: Requesting lock info from node %u\n", dlm->name,
582 if (ndata->node_num == dlm->node_num) {
583 ndata->state = DLM_RECO_NODE_DATA_DONE;
588 status = dlm_request_all_locks(dlm, ndata->node_num,
592 if (dlm_is_host_down(status)) {
593 /* node died, ignore it for recovery */
595 ndata->state = DLM_RECO_NODE_DATA_DEAD;
596 /* wait for the domain map to catch up
597 * with the network state. */
598 wait_event_timeout(dlm->dlm_reco_thread_wq,
599 dlm_is_node_dead(dlm,
601 msecs_to_jiffies(1000));
602 mlog(0, "waited 1 sec for %u, "
603 "dead? %s\n", ndata->node_num,
604 dlm_is_node_dead(dlm, ndata->node_num) ?
607 /* -ENOMEM on the other node */
608 mlog(0, "%s: node %u returned "
609 "%d during recovery, retrying "
610 "after a short wait\n",
611 dlm->name, ndata->node_num,
616 } while (status != 0);
618 spin_lock(&dlm_reco_state_lock);
619 switch (ndata->state) {
620 case DLM_RECO_NODE_DATA_INIT:
621 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
622 case DLM_RECO_NODE_DATA_REQUESTED:
625 case DLM_RECO_NODE_DATA_DEAD:
626 mlog(0, "node %u died after requesting "
627 "recovery info for node %u\n",
628 ndata->node_num, dead_node);
629 /* fine. don't need this node's info.
630 * continue without it. */
632 case DLM_RECO_NODE_DATA_REQUESTING:
633 ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
634 mlog(0, "now receiving recovery data from "
635 "node %u for dead node %u\n",
636 ndata->node_num, dead_node);
638 case DLM_RECO_NODE_DATA_RECEIVING:
639 mlog(0, "already receiving recovery data from "
640 "node %u for dead node %u\n",
641 ndata->node_num, dead_node);
643 case DLM_RECO_NODE_DATA_DONE:
644 mlog(0, "already DONE receiving recovery data "
645 "from node %u for dead node %u\n",
646 ndata->node_num, dead_node);
649 spin_unlock(&dlm_reco_state_lock);
652 mlog(0, "%s: Done requesting all lock info\n", dlm->name);
654 /* nodes should be sending reco data now
655 * just need to wait */
658 /* check all the nodes now to see if we are
659 * done, or if anyone died */
661 spin_lock(&dlm_reco_state_lock);
662 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
663 mlog(0, "checking recovery state of node %u\n",
665 switch (ndata->state) {
666 case DLM_RECO_NODE_DATA_INIT:
667 case DLM_RECO_NODE_DATA_REQUESTING:
668 mlog(ML_ERROR, "bad ndata state for "
669 "node %u: state=%d\n",
670 ndata->node_num, ndata->state);
673 case DLM_RECO_NODE_DATA_DEAD:
674 mlog(0, "node %u died after "
675 "requesting recovery info for "
676 "node %u\n", ndata->node_num,
679 case DLM_RECO_NODE_DATA_RECEIVING:
680 case DLM_RECO_NODE_DATA_REQUESTED:
681 mlog(0, "%s: node %u still in state %s\n",
682 dlm->name, ndata->node_num,
683 ndata->state==DLM_RECO_NODE_DATA_RECEIVING ?
684 "receiving" : "requested");
687 case DLM_RECO_NODE_DATA_DONE:
688 mlog(0, "%s: node %u state is done\n",
689 dlm->name, ndata->node_num);
691 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
692 mlog(0, "%s: node %u state is finalize\n",
693 dlm->name, ndata->node_num);
697 spin_unlock(&dlm_reco_state_lock);
699 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
700 all_nodes_done?"yes":"no");
701 if (all_nodes_done) {
704 /* Set this flag on recovery master to avoid
705 * a new recovery for another dead node start
706 * before the recovery is not done. That may
707 * cause recovery hung.*/
708 spin_lock(&dlm->spinlock);
709 dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
710 spin_unlock(&dlm->spinlock);
712 /* all nodes are now in DLM_RECO_NODE_DATA_DONE state
713 * just send a finalize message to everyone and
715 mlog(0, "all nodes are done! send finalize\n");
716 ret = dlm_send_finalize_reco_message(dlm);
720 spin_lock(&dlm->spinlock);
721 dlm_finish_local_lockres_recovery(dlm, dead_node,
723 spin_unlock(&dlm->spinlock);
724 mlog(0, "should be done with recovery!\n");
726 mlog(0, "finishing recovery of %s at %lu, "
727 "dead=%u, this=%u, new=%u\n", dlm->name,
728 jiffies, dlm->reco.dead_node,
729 dlm->node_num, dlm->reco.new_master);
732 /* rescan everything marked dirty along the way */
733 dlm_kick_thread(dlm, NULL);
736 /* wait to be signalled, with periodic timeout
737 * to check for node death */
738 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
739 kthread_should_stop(),
740 msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
745 dlm_destroy_recovery_area(dlm, dead_node);
750 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
753 struct dlm_reco_node_data *ndata;
755 spin_lock(&dlm->spinlock);
756 memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
757 /* nodes can only be removed (by dying) after dropping
758 * this lock, and death will be trapped later, so this should do */
759 spin_unlock(&dlm->spinlock);
762 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
763 if (num >= O2NM_MAX_NODES) {
766 BUG_ON(num == dead_node);
768 ndata = kzalloc(sizeof(*ndata), GFP_NOFS);
770 dlm_destroy_recovery_area(dlm, dead_node);
773 ndata->node_num = num;
774 ndata->state = DLM_RECO_NODE_DATA_INIT;
775 spin_lock(&dlm_reco_state_lock);
776 list_add_tail(&ndata->list, &dlm->reco.node_data);
777 spin_unlock(&dlm_reco_state_lock);
784 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
786 struct dlm_reco_node_data *ndata, *next;
789 spin_lock(&dlm_reco_state_lock);
790 list_splice_init(&dlm->reco.node_data, &tmplist);
791 spin_unlock(&dlm_reco_state_lock);
793 list_for_each_entry_safe(ndata, next, &tmplist, list) {
794 list_del_init(&ndata->list);
799 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
802 struct dlm_lock_request lr;
808 mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
809 "to %u\n", dead_node, request_from);
811 memset(&lr, 0, sizeof(lr));
812 lr.node_idx = dlm->node_num;
813 lr.dead_node = dead_node;
817 ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
818 &lr, sizeof(lr), request_from, NULL);
820 /* negative status is handled by caller */
822 mlog(ML_ERROR, "%s: Error %d send LOCK_REQUEST to node %u "
823 "to recover dead node %u\n", dlm->name, ret,
824 request_from, dead_node);
825 // return from here, then
826 // sleep until all received or error
831 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
834 struct dlm_ctxt *dlm = data;
835 struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
837 struct dlm_work_item *item = NULL;
842 if (lr->dead_node != dlm->reco.dead_node) {
843 mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local "
844 "dead_node is %u\n", dlm->name, lr->node_idx,
845 lr->dead_node, dlm->reco.dead_node);
846 dlm_print_reco_node_status(dlm);
851 BUG_ON(lr->dead_node != dlm->reco.dead_node);
853 item = kzalloc(sizeof(*item), GFP_NOFS);
859 /* this will get freed by dlm_request_all_locks_worker */
860 buf = (char *) __get_free_page(GFP_NOFS);
867 /* queue up work for dlm_request_all_locks_worker */
868 dlm_grab(dlm); /* get an extra ref for the work item */
869 dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
870 item->u.ral.reco_master = lr->node_idx;
871 item->u.ral.dead_node = lr->dead_node;
872 spin_lock(&dlm->work_lock);
873 list_add_tail(&item->list, &dlm->work_list);
874 spin_unlock(&dlm->work_lock);
875 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
881 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
883 struct dlm_migratable_lockres *mres;
884 struct dlm_lock_resource *res;
885 struct dlm_ctxt *dlm;
886 LIST_HEAD(resources);
888 u8 dead_node, reco_master;
889 int skip_all_done = 0;
892 dead_node = item->u.ral.dead_node;
893 reco_master = item->u.ral.reco_master;
894 mres = (struct dlm_migratable_lockres *)data;
896 mlog(0, "%s: recovery worker started, dead=%u, master=%u\n",
897 dlm->name, dead_node, reco_master);
899 if (dead_node != dlm->reco.dead_node ||
900 reco_master != dlm->reco.new_master) {
901 /* worker could have been created before the recovery master
902 * died. if so, do not continue, but do not error. */
903 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
904 mlog(ML_NOTICE, "%s: will not send recovery state, "
905 "recovery master %u died, thread=(dead=%u,mas=%u)"
906 " current=(dead=%u,mas=%u)\n", dlm->name,
907 reco_master, dead_node, reco_master,
908 dlm->reco.dead_node, dlm->reco.new_master);
910 mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, "
911 "master=%u), request(dead=%u, master=%u)\n",
912 dlm->name, dlm->reco.dead_node,
913 dlm->reco.new_master, dead_node, reco_master);
918 /* lock resources should have already been moved to the
919 * dlm->reco.resources list. now move items from that list
920 * to a temp list if the dead owner matches. note that the
921 * whole cluster recovers only one node at a time, so we
922 * can safely move UNKNOWN lock resources for each recovery
924 dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
926 /* now we can begin blasting lockreses without the dlm lock */
928 /* any errors returned will be due to the new_master dying,
929 * the dlm_reco_thread should detect this */
930 list_for_each_entry(res, &resources, recovering) {
931 ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
934 mlog(ML_ERROR, "%s: node %u went down while sending "
935 "recovery state for dead node %u, ret=%d\n", dlm->name,
936 reco_master, dead_node, ret);
942 /* move the resources back to the list */
943 spin_lock(&dlm->spinlock);
944 list_splice_init(&resources, &dlm->reco.resources);
945 spin_unlock(&dlm->spinlock);
947 if (!skip_all_done) {
948 ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
950 mlog(ML_ERROR, "%s: node %u went down while sending "
951 "recovery all-done for dead node %u, ret=%d\n",
952 dlm->name, reco_master, dead_node, ret);
956 free_page((unsigned long)data);
960 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
963 struct dlm_reco_data_done done_msg;
965 memset(&done_msg, 0, sizeof(done_msg));
966 done_msg.node_idx = dlm->node_num;
967 done_msg.dead_node = dead_node;
968 mlog(0, "sending DATA DONE message to %u, "
969 "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
972 ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
973 sizeof(done_msg), send_to, &tmpret);
975 mlog(ML_ERROR, "%s: Error %d send RECO_DATA_DONE to node %u "
976 "to recover dead node %u\n", dlm->name, ret, send_to,
978 if (!dlm_is_host_down(ret)) {
987 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
990 struct dlm_ctxt *dlm = data;
991 struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
992 struct dlm_reco_node_data *ndata = NULL;
998 mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
999 "node_idx=%u, this node=%u\n", done->dead_node,
1000 dlm->reco.dead_node, done->node_idx, dlm->node_num);
1002 mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node),
1003 "Got DATA DONE: dead_node=%u, reco.dead_node=%u, "
1004 "node_idx=%u, this node=%u\n", done->dead_node,
1005 dlm->reco.dead_node, done->node_idx, dlm->node_num);
1007 spin_lock(&dlm_reco_state_lock);
1008 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
1009 if (ndata->node_num != done->node_idx)
1012 switch (ndata->state) {
1013 /* should have moved beyond INIT but not to FINALIZE yet */
1014 case DLM_RECO_NODE_DATA_INIT:
1015 case DLM_RECO_NODE_DATA_DEAD:
1016 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
1017 mlog(ML_ERROR, "bad ndata state for node %u:"
1018 " state=%d\n", ndata->node_num,
1022 /* these states are possible at this point, anywhere along
1023 * the line of recovery */
1024 case DLM_RECO_NODE_DATA_DONE:
1025 case DLM_RECO_NODE_DATA_RECEIVING:
1026 case DLM_RECO_NODE_DATA_REQUESTED:
1027 case DLM_RECO_NODE_DATA_REQUESTING:
1028 mlog(0, "node %u is DONE sending "
1032 ndata->state = DLM_RECO_NODE_DATA_DONE;
1037 spin_unlock(&dlm_reco_state_lock);
1039 /* wake the recovery thread, some node is done */
1041 dlm_kick_recovery_thread(dlm);
1044 mlog(ML_ERROR, "failed to find recovery node data for node "
1045 "%u\n", done->node_idx);
1048 mlog(0, "leaving reco data done handler, ret=%d\n", ret);
1052 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
1053 struct list_head *list,
1056 struct dlm_lock_resource *res, *next;
1057 struct dlm_lock *lock;
1059 spin_lock(&dlm->spinlock);
1060 list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
1061 /* always prune any $RECOVERY entries for dead nodes,
1062 * otherwise hangs can occur during later recovery */
1063 if (dlm_is_recovery_lock(res->lockname.name,
1064 res->lockname.len)) {
1065 spin_lock(&res->spinlock);
1066 list_for_each_entry(lock, &res->granted, list) {
1067 if (lock->ml.node == dead_node) {
1068 mlog(0, "AHA! there was "
1069 "a $RECOVERY lock for dead "
1071 dead_node, dlm->name);
1072 list_del_init(&lock->list);
1077 spin_unlock(&res->spinlock);
1081 if (res->owner == dead_node) {
1082 mlog(0, "found lockres owned by dead node while "
1083 "doing recovery for node %u. sending it.\n",
1085 list_move_tail(&res->recovering, list);
1086 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1087 mlog(0, "found UNKNOWN owner while doing recovery "
1088 "for node %u. sending it.\n", dead_node);
1089 list_move_tail(&res->recovering, list);
1092 spin_unlock(&dlm->spinlock);
1095 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
1097 int total_locks = 0;
1098 struct list_head *iter, *queue = &res->granted;
1101 for (i=0; i<3; i++) {
1102 list_for_each(iter, queue)
1110 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
1111 struct dlm_migratable_lockres *mres,
1113 struct dlm_lock_resource *res,
1116 u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
1117 int mres_total_locks = be32_to_cpu(mres->total_locks);
1118 int sz, ret = 0, status = 0;
1119 u8 orig_flags = mres->flags,
1120 orig_master = mres->master;
1122 BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
1123 if (!mres->num_locks)
1126 sz = sizeof(struct dlm_migratable_lockres) +
1127 (mres->num_locks * sizeof(struct dlm_migratable_lock));
1129 /* add an all-done flag if we reached the last lock */
1130 orig_flags = mres->flags;
1131 BUG_ON(total_locks > mres_total_locks);
1132 if (total_locks == mres_total_locks)
1133 mres->flags |= DLM_MRES_ALL_DONE;
1135 mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n",
1136 dlm->name, res->lockname.len, res->lockname.name,
1137 orig_flags & DLM_MRES_MIGRATION ? "migration" : "recovery",
1141 ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
1142 sz, send_to, &status);
1144 /* XXX: negative status is not handled.
1145 * this will end up killing this node. */
1146 mlog(ML_ERROR, "%s: res %.*s, Error %d send MIG_LOCKRES to "
1147 "node %u (%s)\n", dlm->name, mres->lockname_len,
1148 mres->lockname, ret, send_to,
1149 (orig_flags & DLM_MRES_MIGRATION ?
1150 "migration" : "recovery"));
1152 /* might get an -ENOMEM back here */
1157 if (ret == -EFAULT) {
1158 mlog(ML_ERROR, "node %u told me to kill "
1159 "myself!\n", send_to);
1165 /* zero and reinit the message buffer */
1166 dlm_init_migratable_lockres(mres, res->lockname.name,
1167 res->lockname.len, mres_total_locks,
1168 mig_cookie, orig_flags, orig_master);
1172 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
1173 const char *lockname, int namelen,
1174 int total_locks, u64 cookie,
1175 u8 flags, u8 master)
1177 /* mres here is one full page */
1179 mres->lockname_len = namelen;
1180 memcpy(mres->lockname, lockname, namelen);
1181 mres->num_locks = 0;
1182 mres->total_locks = cpu_to_be32(total_locks);
1183 mres->mig_cookie = cpu_to_be64(cookie);
1184 mres->flags = flags;
1185 mres->master = master;
1188 static void dlm_prepare_lvb_for_migration(struct dlm_lock *lock,
1189 struct dlm_migratable_lockres *mres,
1195 /* Ignore lvb in all locks in the blocked list */
1196 if (queue == DLM_BLOCKED_LIST)
1199 /* Only consider lvbs in locks with granted EX or PR lock levels */
1200 if (lock->ml.type != LKM_EXMODE && lock->ml.type != LKM_PRMODE)
1203 if (dlm_lvb_is_empty(mres->lvb)) {
1204 memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1208 /* Ensure the lvb copied for migration matches in other valid locks */
1209 if (!memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))
1212 mlog(ML_ERROR, "Mismatched lvb in lock cookie=%u:%llu, name=%.*s, "
1214 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
1215 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
1216 lock->lockres->lockname.len, lock->lockres->lockname.name,
1218 dlm_print_one_lock_resource(lock->lockres);
1222 /* returns 1 if this lock fills the network structure,
1224 static int dlm_add_lock_to_array(struct dlm_lock *lock,
1225 struct dlm_migratable_lockres *mres, int queue)
1227 struct dlm_migratable_lock *ml;
1228 int lock_num = mres->num_locks;
1230 ml = &(mres->ml[lock_num]);
1231 ml->cookie = lock->ml.cookie;
1232 ml->type = lock->ml.type;
1233 ml->convert_type = lock->ml.convert_type;
1234 ml->highest_blocked = lock->ml.highest_blocked;
1237 ml->flags = lock->lksb->flags;
1238 dlm_prepare_lvb_for_migration(lock, mres, queue);
1240 ml->node = lock->ml.node;
1242 /* we reached the max, send this network message */
1243 if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1248 static void dlm_add_dummy_lock(struct dlm_ctxt *dlm,
1249 struct dlm_migratable_lockres *mres)
1251 struct dlm_lock dummy;
1252 memset(&dummy, 0, sizeof(dummy));
1253 dummy.ml.cookie = 0;
1254 dummy.ml.type = LKM_IVMODE;
1255 dummy.ml.convert_type = LKM_IVMODE;
1256 dummy.ml.highest_blocked = LKM_IVMODE;
1258 dummy.ml.node = dlm->node_num;
1259 dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST);
1262 static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm,
1263 struct dlm_migratable_lock *ml,
1266 if (unlikely(ml->cookie == 0 &&
1267 ml->type == LKM_IVMODE &&
1268 ml->convert_type == LKM_IVMODE &&
1269 ml->highest_blocked == LKM_IVMODE &&
1270 ml->list == DLM_BLOCKED_LIST)) {
1271 *nodenum = ml->node;
1277 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1278 struct dlm_migratable_lockres *mres,
1279 u8 send_to, u8 flags)
1281 struct list_head *queue;
1284 struct dlm_lock *lock;
1287 BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1289 mlog(0, "sending to %u\n", send_to);
1291 total_locks = dlm_num_locks_in_lockres(res);
1292 if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1293 /* rare, but possible */
1294 mlog(0, "argh. lockres has %d locks. this will "
1295 "require more than one network packet to "
1296 "migrate\n", total_locks);
1297 mig_cookie = dlm_get_next_mig_cookie();
1300 dlm_init_migratable_lockres(mres, res->lockname.name,
1301 res->lockname.len, total_locks,
1302 mig_cookie, flags, res->owner);
1305 for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1306 queue = dlm_list_idx_to_ptr(res, i);
1307 list_for_each_entry(lock, queue, list) {
1308 /* add another lock. */
1310 if (!dlm_add_lock_to_array(lock, mres, i))
1313 /* this filled the lock message,
1314 * we must send it immediately. */
1315 ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1321 if (total_locks == 0) {
1322 /* send a dummy lock to indicate a mastery reference only */
1323 mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n",
1324 dlm->name, res->lockname.len, res->lockname.name,
1325 send_to, flags & DLM_MRES_RECOVERY ? "recovery" :
1327 dlm_add_dummy_lock(dlm, mres);
1329 /* flush any remaining locks */
1330 ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1336 mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n",
1338 if (!dlm_is_host_down(ret))
1340 mlog(0, "%s: node %u went down while sending %s "
1341 "lockres %.*s\n", dlm->name, send_to,
1342 flags & DLM_MRES_RECOVERY ? "recovery" : "migration",
1343 res->lockname.len, res->lockname.name);
1350 * this message will contain no more than one page worth of
1351 * recovery data, and it will work on only one lockres.
1352 * there may be many locks in this page, and we may need to wait
1353 * for additional packets to complete all the locks (rare, but
1357 * NOTE: the allocation error cases here are scary
1358 * we really cannot afford to fail an alloc in recovery
1359 * do we spin? returning an error only delays the problem really
1362 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
1365 struct dlm_ctxt *dlm = data;
1366 struct dlm_migratable_lockres *mres =
1367 (struct dlm_migratable_lockres *)msg->buf;
1372 struct dlm_work_item *item = NULL;
1373 struct dlm_lock_resource *res = NULL;
1378 BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1380 real_master = mres->master;
1381 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1382 /* cannot migrate a lockres with no master */
1383 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1386 mlog(0, "%s message received from node %u\n",
1387 (mres->flags & DLM_MRES_RECOVERY) ?
1388 "recovery" : "migration", mres->master);
1389 if (mres->flags & DLM_MRES_ALL_DONE)
1390 mlog(0, "all done flag. all lockres data received!\n");
1393 buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS);
1394 item = kzalloc(sizeof(*item), GFP_NOFS);
1398 /* lookup the lock to see if we have a secondary queue for this
1399 * already... just add the locks in and this will have its owner
1400 * and RECOVERY flag changed when it completes. */
1401 res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len);
1403 /* this will get a ref on res */
1404 /* mark it as recovering/migrating and hash it */
1405 spin_lock(&res->spinlock);
1406 if (mres->flags & DLM_MRES_RECOVERY) {
1407 res->state |= DLM_LOCK_RES_RECOVERING;
1409 if (res->state & DLM_LOCK_RES_MIGRATING) {
1410 /* this is at least the second
1411 * lockres message */
1412 mlog(0, "lock %.*s is already migrating\n",
1415 } else if (res->state & DLM_LOCK_RES_RECOVERING) {
1416 /* caller should BUG */
1417 mlog(ML_ERROR, "node is attempting to migrate "
1418 "lock %.*s, but marked as recovering!\n",
1419 mres->lockname_len, mres->lockname);
1421 spin_unlock(&res->spinlock);
1422 dlm_lockres_put(res);
1425 res->state |= DLM_LOCK_RES_MIGRATING;
1427 spin_unlock(&res->spinlock);
1429 /* need to allocate, just like if it was
1430 * mastered here normally */
1431 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1435 /* to match the ref that we would have gotten if
1436 * dlm_lookup_lockres had succeeded */
1437 dlm_lockres_get(res);
1439 /* mark it as recovering/migrating and hash it */
1440 if (mres->flags & DLM_MRES_RECOVERY)
1441 res->state |= DLM_LOCK_RES_RECOVERING;
1443 res->state |= DLM_LOCK_RES_MIGRATING;
1445 spin_lock(&dlm->spinlock);
1446 __dlm_insert_lockres(dlm, res);
1447 spin_unlock(&dlm->spinlock);
1449 /* Add an extra ref for this lock-less lockres lest the
1450 * dlm_thread purges it before we get the chance to add
1452 dlm_lockres_get(res);
1454 /* There are three refs that need to be put.
1456 * 2. kref_init in dlm_new_lockres()->dlm_init_lockres().
1457 * 3. dlm_lookup_lockres()
1458 * The first one is handled at the end of this function. The
1459 * other two are handled in the worker thread after locks have
1460 * been attached. Yes, we don't wait for purge time to match
1461 * kref_init. The lockres will still have atleast one ref
1462 * added because it is in the hash __dlm_insert_lockres() */
1465 /* now that the new lockres is inserted,
1466 * make it usable by other processes */
1467 spin_lock(&res->spinlock);
1468 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1469 spin_unlock(&res->spinlock);
1473 /* at this point we have allocated everything we need,
1474 * and we have a hashed lockres with an extra ref and
1475 * the proper res->state flags. */
1477 spin_lock(&res->spinlock);
1478 /* drop this either when master requery finds a different master
1479 * or when a lock is added by the recovery worker */
1480 dlm_lockres_grab_inflight_ref(dlm, res);
1481 if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1482 /* migration cannot have an unknown master */
1483 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1484 mlog(0, "recovery has passed me a lockres with an "
1485 "unknown owner.. will need to requery: "
1486 "%.*s\n", mres->lockname_len, mres->lockname);
1488 /* take a reference now to pin the lockres, drop it
1489 * when locks are added in the worker */
1490 dlm_change_lockres_owner(dlm, res, dlm->node_num);
1492 spin_unlock(&res->spinlock);
1494 /* queue up work for dlm_mig_lockres_worker */
1495 dlm_grab(dlm); /* get an extra ref for the work item */
1496 memcpy(buf, msg->buf, be16_to_cpu(msg->data_len)); /* copy the whole message */
1497 dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1498 item->u.ml.lockres = res; /* already have a ref */
1499 item->u.ml.real_master = real_master;
1500 item->u.ml.extra_ref = extra_refs;
1501 spin_lock(&dlm->work_lock);
1502 list_add_tail(&item->list, &dlm->work_list);
1503 spin_unlock(&dlm->work_lock);
1504 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
1507 /* One extra ref taken needs to be put here */
1509 dlm_lockres_put(res);
1522 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1524 struct dlm_ctxt *dlm;
1525 struct dlm_migratable_lockres *mres;
1527 struct dlm_lock_resource *res;
1532 mres = (struct dlm_migratable_lockres *)data;
1534 res = item->u.ml.lockres;
1535 real_master = item->u.ml.real_master;
1536 extra_ref = item->u.ml.extra_ref;
1538 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1539 /* this case is super-rare. only occurs if
1540 * node death happens during migration. */
1542 ret = dlm_lockres_master_requery(dlm, res, &real_master);
1544 mlog(0, "dlm_lockres_master_requery ret=%d\n",
1548 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1549 mlog(0, "lockres %.*s not claimed. "
1550 "this node will take it.\n",
1551 res->lockname.len, res->lockname.name);
1553 spin_lock(&res->spinlock);
1554 dlm_lockres_drop_inflight_ref(dlm, res);
1555 spin_unlock(&res->spinlock);
1556 mlog(0, "master needs to respond to sender "
1557 "that node %u still owns %.*s\n",
1558 real_master, res->lockname.len,
1559 res->lockname.name);
1560 /* cannot touch this lockres */
1565 ret = dlm_process_recovery_data(dlm, res, mres);
1567 mlog(0, "dlm_process_recovery_data returned %d\n", ret);
1569 mlog(0, "dlm_process_recovery_data succeeded\n");
1571 if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1572 (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1573 ret = dlm_finish_migration(dlm, res, mres->master);
1579 /* See comment in dlm_mig_lockres_handler() */
1582 dlm_lockres_put(res);
1583 dlm_lockres_put(res);
1590 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1591 struct dlm_lock_resource *res,
1594 struct dlm_node_iter iter;
1598 *real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1600 /* we only reach here if one of the two nodes in a
1601 * migration died while the migration was in progress.
1602 * at this point we need to requery the master. we
1603 * know that the new_master got as far as creating
1604 * an mle on at least one node, but we do not know
1605 * if any nodes had actually cleared the mle and set
1606 * the master to the new_master. the old master
1607 * is supposed to set the owner to UNKNOWN in the
1608 * event of a new_master death, so the only possible
1609 * responses that we can get from nodes here are
1610 * that the master is new_master, or that the master
1612 * if all nodes come back with UNKNOWN then we know
1613 * the lock needs remastering here.
1614 * if any node comes back with a valid master, check
1615 * to see if that master is the one that we are
1616 * recovering. if so, then the new_master died and
1617 * we need to remaster this lock. if not, then the
1618 * new_master survived and that node will respond to
1619 * other nodes about the owner.
1620 * if there is an owner, this node needs to dump this
1621 * lockres and alert the sender that this lockres
1623 spin_lock(&dlm->spinlock);
1624 dlm_node_iter_init(dlm->domain_map, &iter);
1625 spin_unlock(&dlm->spinlock);
1627 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1628 /* do not send to self */
1629 if (nodenum == dlm->node_num)
1631 ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1634 if (!dlm_is_host_down(ret))
1636 /* host is down, so answer for that node would be
1637 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
1639 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1640 mlog(0, "lock master is %u\n", *real_master);
1648 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1649 u8 nodenum, u8 *real_master)
1652 struct dlm_master_requery req;
1653 int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1655 memset(&req, 0, sizeof(req));
1656 req.node_idx = dlm->node_num;
1657 req.namelen = res->lockname.len;
1658 memcpy(req.name, res->lockname.name, res->lockname.len);
1660 ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1661 &req, sizeof(req), nodenum, &status);
1662 /* XXX: negative status not handled properly here. */
1664 mlog(ML_ERROR, "Error %d when sending message %u (key "
1665 "0x%x) to node %u\n", ret, DLM_MASTER_REQUERY_MSG,
1669 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1670 *real_master = (u8) (status & 0xff);
1671 mlog(0, "node %u responded to master requery with %u\n",
1672 nodenum, *real_master);
1679 /* this function cannot error, so unless the sending
1680 * or receiving of the message failed, the owner can
1682 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
1685 struct dlm_ctxt *dlm = data;
1686 struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1687 struct dlm_lock_resource *res = NULL;
1689 int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1690 u32 flags = DLM_ASSERT_MASTER_REQUERY;
1692 if (!dlm_grab(dlm)) {
1693 /* since the domain has gone away on this
1694 * node, the proper response is UNKNOWN */
1698 hash = dlm_lockid_hash(req->name, req->namelen);
1700 spin_lock(&dlm->spinlock);
1701 res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash);
1703 spin_lock(&res->spinlock);
1704 master = res->owner;
1705 if (master == dlm->node_num) {
1706 int ret = dlm_dispatch_assert_master(dlm, res,
1709 mlog_errno(-ENOMEM);
1713 } else /* put.. incase we are not the master */
1714 dlm_lockres_put(res);
1715 spin_unlock(&res->spinlock);
1717 spin_unlock(&dlm->spinlock);
1723 static inline struct list_head *
1724 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1726 struct list_head *ret;
1727 BUG_ON(list_num < 0);
1728 BUG_ON(list_num > 2);
1729 ret = &(res->granted);
1733 /* TODO: do ast flush business
1734 * TODO: do MIGRATING and RECOVERING spinning
1738 * NOTE about in-flight requests during migration:
1740 * Before attempting the migrate, the master has marked the lockres as
1741 * MIGRATING and then flushed all of its pending ASTS. So any in-flight
1742 * requests either got queued before the MIGRATING flag got set, in which
1743 * case the lock data will reflect the change and a return message is on
1744 * the way, or the request failed to get in before MIGRATING got set. In
1745 * this case, the caller will be told to spin and wait for the MIGRATING
1746 * flag to be dropped, then recheck the master.
1747 * This holds true for the convert, cancel and unlock cases, and since lvb
1748 * updates are tied to these same messages, it applies to lvb updates as
1749 * well. For the lock case, there is no way a lock can be on the master
1750 * queue and not be on the secondary queue since the lock is always added
1751 * locally first. This means that the new target node will never be sent
1752 * a lock that he doesn't already have on the list.
1753 * In total, this means that the local lock is correct and should not be
1754 * updated to match the one sent by the master. Any messages sent back
1755 * from the master before the MIGRATING flag will bring the lock properly
1756 * up-to-date, and the change will be ordered properly for the waiter.
1757 * We will *not* attempt to modify the lock underneath the waiter.
1760 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1761 struct dlm_lock_resource *res,
1762 struct dlm_migratable_lockres *mres)
1764 struct dlm_migratable_lock *ml;
1765 struct list_head *queue, *iter;
1766 struct list_head *tmpq = NULL;
1767 struct dlm_lock *newlock = NULL;
1768 struct dlm_lockstatus *lksb = NULL;
1771 struct dlm_lock *lock;
1772 u8 from = O2NM_MAX_NODES;
1773 unsigned int added = 0;
1776 mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1777 for (i=0; i<mres->num_locks; i++) {
1778 ml = &(mres->ml[i]);
1780 if (dlm_is_dummy_lock(dlm, ml, &from)) {
1781 /* placeholder, just need to set the refmap bit */
1782 BUG_ON(mres->num_locks != 1);
1783 mlog(0, "%s:%.*s: dummy lock for %u\n",
1784 dlm->name, mres->lockname_len, mres->lockname,
1786 spin_lock(&res->spinlock);
1787 dlm_lockres_set_refmap_bit(dlm, res, from);
1788 spin_unlock(&res->spinlock);
1792 BUG_ON(ml->highest_blocked != LKM_IVMODE);
1796 queue = dlm_list_num_to_pointer(res, ml->list);
1799 /* if the lock is for the local node it needs to
1800 * be moved to the proper location within the queue.
1801 * do not allocate a new lock structure. */
1802 if (ml->node == dlm->node_num) {
1803 /* MIGRATION ONLY! */
1804 BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1807 spin_lock(&res->spinlock);
1808 for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) {
1809 tmpq = dlm_list_idx_to_ptr(res, j);
1810 list_for_each(iter, tmpq) {
1811 lock = list_entry(iter,
1812 struct dlm_lock, list);
1813 if (lock->ml.cookie == ml->cookie)
1821 /* lock is always created locally first, and
1822 * destroyed locally last. it must be on the list */
1825 mlog(ML_ERROR, "Could not find local lock "
1826 "with cookie %u:%llu, node %u, "
1827 "list %u, flags 0x%x, type %d, "
1828 "conv %d, highest blocked %d\n",
1829 dlm_get_lock_cookie_node(be64_to_cpu(c)),
1830 dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1831 ml->node, ml->list, ml->flags, ml->type,
1832 ml->convert_type, ml->highest_blocked);
1833 __dlm_print_one_lock_resource(res);
1837 if (lock->ml.node != ml->node) {
1838 c = lock->ml.cookie;
1839 mlog(ML_ERROR, "Mismatched node# in lock "
1840 "cookie %u:%llu, name %.*s, node %u\n",
1841 dlm_get_lock_cookie_node(be64_to_cpu(c)),
1842 dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1843 res->lockname.len, res->lockname.name,
1846 mlog(ML_ERROR, "Migrate lock cookie %u:%llu, "
1847 "node %u, list %u, flags 0x%x, type %d, "
1848 "conv %d, highest blocked %d\n",
1849 dlm_get_lock_cookie_node(be64_to_cpu(c)),
1850 dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1851 ml->node, ml->list, ml->flags, ml->type,
1852 ml->convert_type, ml->highest_blocked);
1853 __dlm_print_one_lock_resource(res);
1857 if (tmpq != queue) {
1859 mlog(0, "Lock cookie %u:%llu was on list %u "
1860 "instead of list %u for %.*s\n",
1861 dlm_get_lock_cookie_node(be64_to_cpu(c)),
1862 dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1863 j, ml->list, res->lockname.len,
1864 res->lockname.name);
1865 __dlm_print_one_lock_resource(res);
1866 spin_unlock(&res->spinlock);
1870 /* see NOTE above about why we do not update
1871 * to match the master here */
1873 /* move the lock to its proper place */
1874 /* do not alter lock refcount. switching lists. */
1875 list_move_tail(&lock->list, queue);
1876 spin_unlock(&res->spinlock);
1879 mlog(0, "just reordered a local lock!\n");
1883 /* lock is for another node. */
1884 newlock = dlm_new_lock(ml->type, ml->node,
1885 be64_to_cpu(ml->cookie), NULL);
1890 lksb = newlock->lksb;
1891 dlm_lock_attach_lockres(newlock, res);
1893 if (ml->convert_type != LKM_IVMODE) {
1894 BUG_ON(queue != &res->converting);
1895 newlock->ml.convert_type = ml->convert_type;
1897 lksb->flags |= (ml->flags &
1898 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1900 if (ml->type == LKM_NLMODE)
1903 if (!dlm_lvb_is_empty(mres->lvb)) {
1904 if (lksb->flags & DLM_LKSB_PUT_LVB) {
1905 /* other node was trying to update
1906 * lvb when node died. recreate the
1907 * lksb with the updated lvb. */
1908 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1909 /* the lock resource lvb update must happen
1910 * NOW, before the spinlock is dropped.
1911 * we no longer wait for the AST to update
1913 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1915 /* otherwise, the node is sending its
1916 * most recent valid lvb info */
1917 BUG_ON(ml->type != LKM_EXMODE &&
1918 ml->type != LKM_PRMODE);
1919 if (!dlm_lvb_is_empty(res->lvb) &&
1920 (ml->type == LKM_EXMODE ||
1921 memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1923 mlog(ML_ERROR, "%s:%.*s: received bad "
1924 "lvb! type=%d\n", dlm->name,
1926 res->lockname.name, ml->type);
1927 printk("lockres lvb=[");
1928 for (i=0; i<DLM_LVB_LEN; i++)
1929 printk("%02x", res->lvb[i]);
1930 printk("]\nmigrated lvb=[");
1931 for (i=0; i<DLM_LVB_LEN; i++)
1932 printk("%02x", mres->lvb[i]);
1934 dlm_print_one_lock_resource(res);
1937 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1943 * wrt lock queue ordering and recovery:
1944 * 1. order of locks on granted queue is
1946 * 2. order of locks on converting queue is
1947 * LOST with the node death. sorry charlie.
1948 * 3. order of locks on the blocked queue is
1950 * order of locks does not affect integrity, it
1951 * just means that a lock request may get pushed
1952 * back in line as a result of the node death.
1953 * also note that for a given node the lock order
1954 * for its secondary queue locks is preserved
1955 * relative to each other, but clearly *not*
1956 * preserved relative to locks from other nodes.
1959 spin_lock(&res->spinlock);
1960 list_for_each_entry(lock, queue, list) {
1961 if (lock->ml.cookie == ml->cookie) {
1962 c = lock->ml.cookie;
1963 mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already "
1964 "exists on this lockres!\n", dlm->name,
1965 res->lockname.len, res->lockname.name,
1966 dlm_get_lock_cookie_node(be64_to_cpu(c)),
1967 dlm_get_lock_cookie_seq(be64_to_cpu(c)));
1969 mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, "
1970 "node=%u, cookie=%u:%llu, queue=%d\n",
1971 ml->type, ml->convert_type, ml->node,
1972 dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)),
1973 dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)),
1976 __dlm_print_one_lock_resource(res);
1982 dlm_lock_get(newlock);
1983 list_add_tail(&newlock->list, queue);
1984 mlog(0, "%s:%.*s: added lock for node %u, "
1985 "setting refmap bit\n", dlm->name,
1986 res->lockname.len, res->lockname.name, ml->node);
1987 dlm_lockres_set_refmap_bit(dlm, res, ml->node);
1990 spin_unlock(&res->spinlock);
1992 mlog(0, "done running all the locks\n");
1995 /* balance the ref taken when the work was queued */
1996 spin_lock(&res->spinlock);
1997 dlm_lockres_drop_inflight_ref(dlm, res);
1998 spin_unlock(&res->spinlock);
2003 dlm_lock_put(newlock);
2009 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
2010 struct dlm_lock_resource *res)
2013 struct list_head *queue;
2014 struct dlm_lock *lock, *next;
2016 assert_spin_locked(&dlm->spinlock);
2017 assert_spin_locked(&res->spinlock);
2018 res->state |= DLM_LOCK_RES_RECOVERING;
2019 if (!list_empty(&res->recovering)) {
2021 "Recovering res %s:%.*s, is already on recovery list!\n",
2022 dlm->name, res->lockname.len, res->lockname.name);
2023 list_del_init(&res->recovering);
2024 dlm_lockres_put(res);
2026 /* We need to hold a reference while on the recovery list */
2027 dlm_lockres_get(res);
2028 list_add_tail(&res->recovering, &dlm->reco.resources);
2030 /* find any pending locks and put them back on proper list */
2031 for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
2032 queue = dlm_list_idx_to_ptr(res, i);
2033 list_for_each_entry_safe(lock, next, queue, list) {
2035 if (lock->convert_pending) {
2036 /* move converting lock back to granted */
2037 BUG_ON(i != DLM_CONVERTING_LIST);
2038 mlog(0, "node died with convert pending "
2039 "on %.*s. move back to granted list.\n",
2040 res->lockname.len, res->lockname.name);
2041 dlm_revert_pending_convert(res, lock);
2042 lock->convert_pending = 0;
2043 } else if (lock->lock_pending) {
2044 /* remove pending lock requests completely */
2045 BUG_ON(i != DLM_BLOCKED_LIST);
2046 mlog(0, "node died with lock pending "
2047 "on %.*s. remove from blocked list and skip.\n",
2048 res->lockname.len, res->lockname.name);
2049 /* lock will be floating until ref in
2050 * dlmlock_remote is freed after the network
2051 * call returns. ok for it to not be on any
2052 * list since no ast can be called
2053 * (the master is dead). */
2054 dlm_revert_pending_lock(res, lock);
2055 lock->lock_pending = 0;
2056 } else if (lock->unlock_pending) {
2057 /* if an unlock was in progress, treat as
2058 * if this had completed successfully
2059 * before sending this lock state to the
2060 * new master. note that the dlm_unlock
2061 * call is still responsible for calling
2062 * the unlockast. that will happen after
2063 * the network call times out. for now,
2064 * just move lists to prepare the new
2065 * recovery master. */
2066 BUG_ON(i != DLM_GRANTED_LIST);
2067 mlog(0, "node died with unlock pending "
2068 "on %.*s. remove from blocked list and skip.\n",
2069 res->lockname.len, res->lockname.name);
2070 dlm_commit_pending_unlock(res, lock);
2071 lock->unlock_pending = 0;
2072 } else if (lock->cancel_pending) {
2073 /* if a cancel was in progress, treat as
2074 * if this had completed successfully
2075 * before sending this lock state to the
2077 BUG_ON(i != DLM_CONVERTING_LIST);
2078 mlog(0, "node died with cancel pending "
2079 "on %.*s. move back to granted list.\n",
2080 res->lockname.len, res->lockname.name);
2081 dlm_commit_pending_cancel(res, lock);
2082 lock->cancel_pending = 0;
2091 /* removes all recovered locks from the recovery list.
2092 * sets the res->owner to the new master.
2093 * unsets the RECOVERY flag and wakes waiters. */
2094 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
2095 u8 dead_node, u8 new_master)
2098 struct hlist_head *bucket;
2099 struct dlm_lock_resource *res, *next;
2101 assert_spin_locked(&dlm->spinlock);
2103 list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
2104 if (res->owner == dead_node) {
2105 mlog(0, "%s: res %.*s, Changing owner from %u to %u\n",
2106 dlm->name, res->lockname.len, res->lockname.name,
2107 res->owner, new_master);
2108 list_del_init(&res->recovering);
2109 spin_lock(&res->spinlock);
2110 /* new_master has our reference from
2111 * the lock state sent during recovery */
2112 dlm_change_lockres_owner(dlm, res, new_master);
2113 res->state &= ~DLM_LOCK_RES_RECOVERING;
2114 if (__dlm_lockres_has_locks(res))
2115 __dlm_dirty_lockres(dlm, res);
2116 spin_unlock(&res->spinlock);
2118 dlm_lockres_put(res);
2122 /* this will become unnecessary eventually, but
2123 * for now we need to run the whole hash, clear
2124 * the RECOVERING state and set the owner
2126 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2127 bucket = dlm_lockres_hash(dlm, i);
2128 hlist_for_each_entry(res, bucket, hash_node) {
2129 if (!(res->state & DLM_LOCK_RES_RECOVERING))
2132 if (res->owner != dead_node &&
2133 res->owner != dlm->node_num)
2136 if (!list_empty(&res->recovering)) {
2137 list_del_init(&res->recovering);
2138 dlm_lockres_put(res);
2141 /* new_master has our reference from
2142 * the lock state sent during recovery */
2143 mlog(0, "%s: res %.*s, Changing owner from %u to %u\n",
2144 dlm->name, res->lockname.len, res->lockname.name,
2145 res->owner, new_master);
2146 spin_lock(&res->spinlock);
2147 dlm_change_lockres_owner(dlm, res, new_master);
2148 res->state &= ~DLM_LOCK_RES_RECOVERING;
2149 if (__dlm_lockres_has_locks(res))
2150 __dlm_dirty_lockres(dlm, res);
2151 spin_unlock(&res->spinlock);
2157 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
2160 if (lock->ml.type != LKM_EXMODE &&
2161 lock->ml.type != LKM_PRMODE)
2163 } else if (lock->ml.type == LKM_EXMODE)
2168 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
2169 struct dlm_lock_resource *res, u8 dead_node)
2171 struct list_head *queue;
2172 struct dlm_lock *lock;
2173 int blank_lvb = 0, local = 0;
2177 assert_spin_locked(&dlm->spinlock);
2178 assert_spin_locked(&res->spinlock);
2180 if (res->owner == dlm->node_num)
2181 /* if this node owned the lockres, and if the dead node
2182 * had an EX when he died, blank out the lvb */
2183 search_node = dead_node;
2185 /* if this is a secondary lockres, and we had no EX or PR
2186 * locks granted, we can no longer trust the lvb */
2187 search_node = dlm->node_num;
2188 local = 1; /* check local state for valid lvb */
2191 for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
2192 queue = dlm_list_idx_to_ptr(res, i);
2193 list_for_each_entry(lock, queue, list) {
2194 if (lock->ml.node == search_node) {
2195 if (dlm_lvb_needs_invalidation(lock, local)) {
2196 /* zero the lksb lvb and lockres lvb */
2198 memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
2205 mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
2206 res->lockname.len, res->lockname.name, dead_node);
2207 memset(res->lvb, 0, DLM_LVB_LEN);
2211 static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
2212 struct dlm_lock_resource *res, u8 dead_node)
2214 struct dlm_lock *lock, *next;
2215 unsigned int freed = 0;
2217 /* this node is the lockres master:
2218 * 1) remove any stale locks for the dead node
2219 * 2) if the dead node had an EX when he died, blank out the lvb
2221 assert_spin_locked(&dlm->spinlock);
2222 assert_spin_locked(&res->spinlock);
2224 /* We do two dlm_lock_put(). One for removing from list and the other is
2225 * to force the DLM_UNLOCK_FREE_LOCK action so as to free the locks */
2227 /* TODO: check pending_asts, pending_basts here */
2228 list_for_each_entry_safe(lock, next, &res->granted, list) {
2229 if (lock->ml.node == dead_node) {
2230 list_del_init(&lock->list);
2232 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2237 list_for_each_entry_safe(lock, next, &res->converting, list) {
2238 if (lock->ml.node == dead_node) {
2239 list_del_init(&lock->list);
2241 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2246 list_for_each_entry_safe(lock, next, &res->blocked, list) {
2247 if (lock->ml.node == dead_node) {
2248 list_del_init(&lock->list);
2250 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2257 mlog(0, "%s:%.*s: freed %u locks for dead node %u, "
2258 "dropping ref from lockres\n", dlm->name,
2259 res->lockname.len, res->lockname.name, freed, dead_node);
2260 if(!test_bit(dead_node, res->refmap)) {
2261 mlog(ML_ERROR, "%s:%.*s: freed %u locks for dead node %u, "
2262 "but ref was not set\n", dlm->name,
2263 res->lockname.len, res->lockname.name, freed, dead_node);
2264 __dlm_print_one_lock_resource(res);
2266 dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2267 } else if (test_bit(dead_node, res->refmap)) {
2268 mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2269 "no locks and had not purged before dying\n", dlm->name,
2270 res->lockname.len, res->lockname.name, dead_node);
2271 dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2274 /* do not kick thread yet */
2275 __dlm_dirty_lockres(dlm, res);
2278 /* if this node is the recovery master, and there are no
2279 * locks for a given lockres owned by this node that are in
2280 * either PR or EX mode, zero out the lvb before requesting.
2285 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
2287 struct dlm_lock_resource *res;
2289 struct hlist_head *bucket;
2290 struct dlm_lock *lock;
2293 /* purge any stale mles */
2294 dlm_clean_master_list(dlm, dead_node);
2297 * now clean up all lock resources. there are two rules:
2299 * 1) if the dead node was the master, move the lockres
2300 * to the recovering list. set the RECOVERING flag.
2301 * this lockres needs to be cleaned up before it can
2304 * 2) if this node was the master, remove all locks from
2305 * each of the lockres queues that were owned by the
2306 * dead node. once recovery finishes, the dlm thread
2307 * can be kicked again to see if any ASTs or BASTs
2308 * need to be fired as a result.
2310 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2311 bucket = dlm_lockres_hash(dlm, i);
2312 hlist_for_each_entry(res, bucket, hash_node) {
2313 /* always prune any $RECOVERY entries for dead nodes,
2314 * otherwise hangs can occur during later recovery */
2315 if (dlm_is_recovery_lock(res->lockname.name,
2316 res->lockname.len)) {
2317 spin_lock(&res->spinlock);
2318 list_for_each_entry(lock, &res->granted, list) {
2319 if (lock->ml.node == dead_node) {
2320 mlog(0, "AHA! there was "
2321 "a $RECOVERY lock for dead "
2323 dead_node, dlm->name);
2324 list_del_init(&lock->list);
2329 spin_unlock(&res->spinlock);
2332 spin_lock(&res->spinlock);
2333 /* zero the lvb if necessary */
2334 dlm_revalidate_lvb(dlm, res, dead_node);
2335 if (res->owner == dead_node) {
2336 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
2337 mlog(ML_NOTICE, "%s: res %.*s, Skip "
2338 "recovery as it is being freed\n",
2339 dlm->name, res->lockname.len,
2340 res->lockname.name);
2342 dlm_move_lockres_to_recovery_list(dlm,
2345 } else if (res->owner == dlm->node_num) {
2346 dlm_free_dead_locks(dlm, res, dead_node);
2347 __dlm_lockres_calc_usage(dlm, res);
2349 spin_unlock(&res->spinlock);
2355 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
2357 assert_spin_locked(&dlm->spinlock);
2359 if (dlm->reco.new_master == idx) {
2360 mlog(0, "%s: recovery master %d just died\n",
2362 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2363 /* finalize1 was reached, so it is safe to clear
2364 * the new_master and dead_node. that recovery
2366 mlog(0, "%s: dead master %d had reached "
2367 "finalize1 state, clearing\n", dlm->name, idx);
2368 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2369 __dlm_reset_recovery(dlm);
2373 /* Clean up join state on node death. */
2374 if (dlm->joining_node == idx) {
2375 mlog(0, "Clearing join state for node %u\n", idx);
2376 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
2379 /* check to see if the node is already considered dead */
2380 if (!test_bit(idx, dlm->live_nodes_map)) {
2381 mlog(0, "for domain %s, node %d is already dead. "
2382 "another node likely did recovery already.\n",
2387 /* check to see if we do not care about this node */
2388 if (!test_bit(idx, dlm->domain_map)) {
2389 /* This also catches the case that we get a node down
2390 * but haven't joined the domain yet. */
2391 mlog(0, "node %u already removed from domain!\n", idx);
2395 clear_bit(idx, dlm->live_nodes_map);
2397 /* make sure local cleanup occurs before the heartbeat events */
2398 if (!test_bit(idx, dlm->recovery_map))
2399 dlm_do_local_recovery_cleanup(dlm, idx);
2401 /* notify anything attached to the heartbeat events */
2402 dlm_hb_event_notify_attached(dlm, idx, 0);
2404 mlog(0, "node %u being removed from domain map!\n", idx);
2405 clear_bit(idx, dlm->domain_map);
2406 clear_bit(idx, dlm->exit_domain_map);
2407 /* wake up migration waiters if a node goes down.
2408 * perhaps later we can genericize this for other waiters. */
2409 wake_up(&dlm->migration_wq);
2411 if (test_bit(idx, dlm->recovery_map))
2412 mlog(0, "domain %s, node %u already added "
2413 "to recovery map!\n", dlm->name, idx);
2415 set_bit(idx, dlm->recovery_map);
2418 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
2420 struct dlm_ctxt *dlm = data;
2426 * This will notify any dlm users that a node in our domain
2427 * went away without notifying us first.
2429 if (test_bit(idx, dlm->domain_map))
2430 dlm_fire_domain_eviction_callbacks(dlm, idx);
2432 spin_lock(&dlm->spinlock);
2433 __dlm_hb_node_down(dlm, idx);
2434 spin_unlock(&dlm->spinlock);
2439 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
2441 struct dlm_ctxt *dlm = data;
2446 spin_lock(&dlm->spinlock);
2447 set_bit(idx, dlm->live_nodes_map);
2448 /* do NOT notify mle attached to the heartbeat events.
2449 * new nodes are not interesting in mastery until joined. */
2450 spin_unlock(&dlm->spinlock);
2455 static void dlm_reco_ast(void *astdata)
2457 struct dlm_ctxt *dlm = astdata;
2458 mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2459 dlm->node_num, dlm->name);
2461 static void dlm_reco_bast(void *astdata, int blocked_type)
2463 struct dlm_ctxt *dlm = astdata;
2464 mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2465 dlm->node_num, dlm->name);
2467 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
2469 mlog(0, "unlockast for recovery lock fired!\n");
2473 * dlm_pick_recovery_master will continually attempt to use
2474 * dlmlock() on the special "$RECOVERY" lockres with the
2475 * LKM_NOQUEUE flag to get an EX. every thread that enters
2476 * this function on each node racing to become the recovery
2477 * master will not stop attempting this until either:
2478 * a) this node gets the EX (and becomes the recovery master),
2479 * or b) dlm->reco.new_master gets set to some nodenum
2480 * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2481 * so each time a recovery master is needed, the entire cluster
2482 * will sync at this point. if the new master dies, that will
2483 * be detected in dlm_do_recovery */
2484 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2486 enum dlm_status ret;
2487 struct dlm_lockstatus lksb;
2488 int status = -EINVAL;
2490 mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2491 dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2493 memset(&lksb, 0, sizeof(lksb));
2495 ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
2496 DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN,
2497 dlm_reco_ast, dlm, dlm_reco_bast);
2499 mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2500 dlm->name, ret, lksb.status);
2502 if (ret == DLM_NORMAL) {
2503 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2504 dlm->name, dlm->node_num);
2506 /* got the EX lock. check to see if another node
2507 * just became the reco master */
2508 if (dlm_reco_master_ready(dlm)) {
2509 mlog(0, "%s: got reco EX lock, but %u will "
2510 "do the recovery\n", dlm->name,
2511 dlm->reco.new_master);
2516 /* see if recovery was already finished elsewhere */
2517 spin_lock(&dlm->spinlock);
2518 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
2520 mlog(0, "%s: got reco EX lock, but "
2521 "node got recovered already\n", dlm->name);
2522 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2523 mlog(ML_ERROR, "%s: new master is %u "
2524 "but no dead node!\n",
2525 dlm->name, dlm->reco.new_master);
2529 spin_unlock(&dlm->spinlock);
2532 /* if this node has actually become the recovery master,
2533 * set the master and send the messages to begin recovery */
2535 mlog(0, "%s: dead=%u, this=%u, sending "
2536 "begin_reco now\n", dlm->name,
2537 dlm->reco.dead_node, dlm->node_num);
2538 status = dlm_send_begin_reco_message(dlm,
2539 dlm->reco.dead_node);
2540 /* this always succeeds */
2543 /* set the new_master to this node */
2544 spin_lock(&dlm->spinlock);
2545 dlm_set_reco_master(dlm, dlm->node_num);
2546 spin_unlock(&dlm->spinlock);
2549 /* recovery lock is a special case. ast will not get fired,
2550 * so just go ahead and unlock it. */
2551 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
2552 if (ret == DLM_DENIED) {
2553 mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2554 ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2556 if (ret != DLM_NORMAL) {
2557 /* this would really suck. this could only happen
2558 * if there was a network error during the unlock
2559 * because of node death. this means the unlock
2560 * is actually "done" and the lock structure is
2561 * even freed. we can continue, but only
2562 * because this specific lock name is special. */
2563 mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
2565 } else if (ret == DLM_NOTQUEUED) {
2566 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2567 dlm->name, dlm->node_num);
2568 /* another node is master. wait on
2569 * reco.new_master != O2NM_INVALID_NODE_NUM
2570 * for at most one second */
2571 wait_event_timeout(dlm->dlm_reco_thread_wq,
2572 dlm_reco_master_ready(dlm),
2573 msecs_to_jiffies(1000));
2574 if (!dlm_reco_master_ready(dlm)) {
2575 mlog(0, "%s: reco master taking awhile\n",
2579 /* another node has informed this one that it is reco master */
2580 mlog(0, "%s: reco master %u is ready to recover %u\n",
2581 dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
2583 } else if (ret == DLM_RECOVERING) {
2584 mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n",
2585 dlm->name, dlm->node_num);
2588 struct dlm_lock_resource *res;
2590 /* dlmlock returned something other than NOTQUEUED or NORMAL */
2591 mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2592 "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2593 dlm_errname(lksb.status));
2594 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2595 DLM_RECOVERY_LOCK_NAME_LEN);
2597 dlm_print_one_lock_resource(res);
2598 dlm_lockres_put(res);
2600 mlog(ML_ERROR, "recovery lock not found\n");
2608 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2610 struct dlm_begin_reco br;
2612 struct dlm_node_iter iter;
2616 mlog(0, "%s: dead node is %u\n", dlm->name, dead_node);
2618 spin_lock(&dlm->spinlock);
2619 dlm_node_iter_init(dlm->domain_map, &iter);
2620 spin_unlock(&dlm->spinlock);
2622 clear_bit(dead_node, iter.node_map);
2624 memset(&br, 0, sizeof(br));
2625 br.node_idx = dlm->node_num;
2626 br.dead_node = dead_node;
2628 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2630 if (nodenum == dead_node) {
2631 mlog(0, "not sending begin reco to dead node "
2635 if (nodenum == dlm->node_num) {
2636 mlog(0, "not sending begin reco to self\n");
2641 mlog(0, "attempting to send begin reco msg to %d\n",
2643 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2644 &br, sizeof(br), nodenum, &status);
2645 /* negative status is handled ok by caller here */
2648 if (dlm_is_host_down(ret)) {
2649 /* node is down. not involved in recovery
2650 * so just keep going */
2651 mlog(ML_NOTICE, "%s: node %u was down when sending "
2652 "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2657 * Prior to commit aad1b15310b9bcd59fa81ab8f2b1513b59553ea8,
2658 * dlm_begin_reco_handler() returned EAGAIN and not -EAGAIN.
2659 * We are handling both for compatibility reasons.
2661 if (ret == -EAGAIN || ret == EAGAIN) {
2662 mlog(0, "%s: trying to start recovery of node "
2663 "%u, but node %u is waiting for last recovery "
2664 "to complete, backoff for a bit\n", dlm->name,
2665 dead_node, nodenum);
2670 struct dlm_lock_resource *res;
2672 /* this is now a serious problem, possibly ENOMEM
2673 * in the network stack. must retry */
2675 mlog(ML_ERROR, "begin reco of dlm %s to node %u "
2676 "returned %d\n", dlm->name, nodenum, ret);
2677 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2678 DLM_RECOVERY_LOCK_NAME_LEN);
2680 dlm_print_one_lock_resource(res);
2681 dlm_lockres_put(res);
2683 mlog(ML_ERROR, "recovery lock not found\n");
2685 /* sleep for a bit in hopes that we can avoid
2695 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2698 struct dlm_ctxt *dlm = data;
2699 struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2701 /* ok to return 0, domain has gone away */
2705 spin_lock(&dlm->spinlock);
2706 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2707 mlog(0, "%s: node %u wants to recover node %u (%u:%u) "
2708 "but this node is in finalize state, waiting on finalize2\n",
2709 dlm->name, br->node_idx, br->dead_node,
2710 dlm->reco.dead_node, dlm->reco.new_master);
2711 spin_unlock(&dlm->spinlock);
2714 spin_unlock(&dlm->spinlock);
2716 mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n",
2717 dlm->name, br->node_idx, br->dead_node,
2718 dlm->reco.dead_node, dlm->reco.new_master);
2720 dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2722 spin_lock(&dlm->spinlock);
2723 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2724 if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2725 mlog(0, "%s: new_master %u died, changing "
2726 "to %u\n", dlm->name, dlm->reco.new_master,
2729 mlog(0, "%s: new_master %u NOT DEAD, changing "
2730 "to %u\n", dlm->name, dlm->reco.new_master,
2732 /* may not have seen the new master as dead yet */
2735 if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2736 mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2737 "node %u changing it to %u\n", dlm->name,
2738 dlm->reco.dead_node, br->node_idx, br->dead_node);
2740 dlm_set_reco_master(dlm, br->node_idx);
2741 dlm_set_reco_dead_node(dlm, br->dead_node);
2742 if (!test_bit(br->dead_node, dlm->recovery_map)) {
2743 mlog(0, "recovery master %u sees %u as dead, but this "
2744 "node has not yet. marking %u as dead\n",
2745 br->node_idx, br->dead_node, br->dead_node);
2746 if (!test_bit(br->dead_node, dlm->domain_map) ||
2747 !test_bit(br->dead_node, dlm->live_nodes_map))
2748 mlog(0, "%u not in domain/live_nodes map "
2749 "so setting it in reco map manually\n",
2751 /* force the recovery cleanup in __dlm_hb_node_down
2752 * both of these will be cleared in a moment */
2753 set_bit(br->dead_node, dlm->domain_map);
2754 set_bit(br->dead_node, dlm->live_nodes_map);
2755 __dlm_hb_node_down(dlm, br->dead_node);
2757 spin_unlock(&dlm->spinlock);
2759 dlm_kick_recovery_thread(dlm);
2761 mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n",
2762 dlm->name, br->node_idx, br->dead_node,
2763 dlm->reco.dead_node, dlm->reco.new_master);
2769 #define DLM_FINALIZE_STAGE2 0x01
2770 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2773 struct dlm_finalize_reco fr;
2774 struct dlm_node_iter iter;
2779 mlog(0, "finishing recovery for node %s:%u, "
2780 "stage %d\n", dlm->name, dlm->reco.dead_node, stage);
2782 spin_lock(&dlm->spinlock);
2783 dlm_node_iter_init(dlm->domain_map, &iter);
2784 spin_unlock(&dlm->spinlock);
2787 memset(&fr, 0, sizeof(fr));
2788 fr.node_idx = dlm->node_num;
2789 fr.dead_node = dlm->reco.dead_node;
2791 fr.flags |= DLM_FINALIZE_STAGE2;
2793 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2794 if (nodenum == dlm->node_num)
2796 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2797 &fr, sizeof(fr), nodenum, &status);
2801 mlog(ML_ERROR, "Error %d when sending message %u (key "
2802 "0x%x) to node %u\n", ret, DLM_FINALIZE_RECO_MSG,
2804 if (dlm_is_host_down(ret)) {
2805 /* this has no effect on this recovery
2806 * session, so set the status to zero to
2807 * finish out the last recovery */
2808 mlog(ML_ERROR, "node %u went down after this "
2809 "node finished recovery.\n", nodenum);
2817 /* reset the node_iter back to the top and send finalize2 */
2826 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2829 struct dlm_ctxt *dlm = data;
2830 struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2833 /* ok to return 0, domain has gone away */
2837 if (fr->flags & DLM_FINALIZE_STAGE2)
2840 mlog(0, "%s: node %u finalizing recovery stage%d of "
2841 "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage,
2842 fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master);
2844 spin_lock(&dlm->spinlock);
2846 if (dlm->reco.new_master != fr->node_idx) {
2847 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2848 "%u is supposed to be the new master, dead=%u\n",
2849 fr->node_idx, dlm->reco.new_master, fr->dead_node);
2852 if (dlm->reco.dead_node != fr->dead_node) {
2853 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2854 "node %u, but node %u is supposed to be dead\n",
2855 fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2861 dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2862 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2863 mlog(ML_ERROR, "%s: received finalize1 from "
2864 "new master %u for dead node %u, but "
2865 "this node has already received it!\n",
2866 dlm->name, fr->node_idx, fr->dead_node);
2867 dlm_print_reco_node_status(dlm);
2870 dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
2871 spin_unlock(&dlm->spinlock);
2874 if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) {
2875 mlog(ML_ERROR, "%s: received finalize2 from "
2876 "new master %u for dead node %u, but "
2877 "this node did not have finalize1!\n",
2878 dlm->name, fr->node_idx, fr->dead_node);
2879 dlm_print_reco_node_status(dlm);
2882 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2883 __dlm_reset_recovery(dlm);
2884 spin_unlock(&dlm->spinlock);
2885 dlm_kick_recovery_thread(dlm);
2891 mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n",
2892 dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master);