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 int dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout)
368 mlog(ML_NOTICE, "%s: waiting %dms for notification of "
369 "death of node %u\n", dlm->name, timeout, node);
370 wait_event_timeout(dlm->dlm_reco_thread_wq,
371 dlm_is_node_dead(dlm, node),
372 msecs_to_jiffies(timeout));
374 mlog(ML_NOTICE, "%s: waiting indefinitely for notification "
375 "of death of node %u\n", dlm->name, node);
376 wait_event(dlm->dlm_reco_thread_wq,
377 dlm_is_node_dead(dlm, node));
379 /* for now, return 0 */
383 int dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout)
386 mlog(0, "%s: waiting %dms for notification of "
387 "recovery of node %u\n", dlm->name, timeout, node);
388 wait_event_timeout(dlm->dlm_reco_thread_wq,
389 dlm_is_node_recovered(dlm, node),
390 msecs_to_jiffies(timeout));
392 mlog(0, "%s: waiting indefinitely for notification "
393 "of recovery of node %u\n", dlm->name, node);
394 wait_event(dlm->dlm_reco_thread_wq,
395 dlm_is_node_recovered(dlm, node));
397 /* for now, return 0 */
401 /* callers of the top-level api calls (dlmlock/dlmunlock) should
402 * block on the dlm->reco.event when recovery is in progress.
403 * the dlm recovery thread will set this state when it begins
404 * recovering a dead node (as the new master or not) and clear
405 * the state and wake as soon as all affected lock resources have
406 * been marked with the RECOVERY flag */
407 static int dlm_in_recovery(struct dlm_ctxt *dlm)
410 spin_lock(&dlm->spinlock);
411 in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
412 spin_unlock(&dlm->spinlock);
417 void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
419 if (dlm_in_recovery(dlm)) {
420 mlog(0, "%s: reco thread %d in recovery: "
421 "state=%d, master=%u, dead=%u\n",
422 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
423 dlm->reco.state, dlm->reco.new_master,
424 dlm->reco.dead_node);
426 wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
429 static void dlm_begin_recovery(struct dlm_ctxt *dlm)
431 spin_lock(&dlm->spinlock);
432 BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
433 printk(KERN_NOTICE "o2dlm: Begin recovery on domain %s for node %u\n",
434 dlm->name, dlm->reco.dead_node);
435 dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
436 spin_unlock(&dlm->spinlock);
439 static void dlm_end_recovery(struct dlm_ctxt *dlm)
441 spin_lock(&dlm->spinlock);
442 BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
443 dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
444 spin_unlock(&dlm->spinlock);
445 printk(KERN_NOTICE "o2dlm: End recovery on domain %s\n", dlm->name);
446 wake_up(&dlm->reco.event);
449 static void dlm_print_recovery_master(struct dlm_ctxt *dlm)
451 printk(KERN_NOTICE "o2dlm: Node %u (%s) is the Recovery Master for the "
452 "dead node %u in domain %s\n", dlm->reco.new_master,
453 (dlm->node_num == dlm->reco.new_master ? "me" : "he"),
454 dlm->reco.dead_node, dlm->name);
457 static int dlm_do_recovery(struct dlm_ctxt *dlm)
462 spin_lock(&dlm->spinlock);
464 /* check to see if the new master has died */
465 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
466 test_bit(dlm->reco.new_master, dlm->recovery_map)) {
467 mlog(0, "new master %u died while recovering %u!\n",
468 dlm->reco.new_master, dlm->reco.dead_node);
469 /* unset the new_master, leave dead_node */
470 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
473 /* select a target to recover */
474 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
477 bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES, 0);
478 if (bit >= O2NM_MAX_NODES || bit < 0)
479 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
481 dlm_set_reco_dead_node(dlm, bit);
482 } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
484 mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
485 dlm->reco.dead_node);
486 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
489 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
490 // mlog(0, "nothing to recover! sleeping now!\n");
491 spin_unlock(&dlm->spinlock);
492 /* return to main thread loop and sleep. */
495 mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n",
496 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
497 dlm->reco.dead_node);
498 spin_unlock(&dlm->spinlock);
500 /* take write barrier */
501 /* (stops the list reshuffling thread, proxy ast handling) */
502 dlm_begin_recovery(dlm);
504 if (dlm->reco.new_master == dlm->node_num)
507 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
508 /* choose a new master, returns 0 if this node
509 * is the master, -EEXIST if it's another node.
510 * this does not return until a new master is chosen
511 * or recovery completes entirely. */
512 ret = dlm_pick_recovery_master(dlm);
514 /* already notified everyone. go. */
517 mlog(0, "another node will master this recovery session.\n");
520 dlm_print_recovery_master(dlm);
522 /* it is safe to start everything back up here
523 * because all of the dead node's lock resources
524 * have been marked as in-recovery */
525 dlm_end_recovery(dlm);
527 /* sleep out in main dlm_recovery_thread loop. */
531 dlm_print_recovery_master(dlm);
533 status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
535 /* we should never hit this anymore */
536 mlog(ML_ERROR, "%s: Error %d remastering locks for node %u, "
537 "retrying.\n", dlm->name, status, dlm->reco.dead_node);
538 /* yield a bit to allow any final network messages
539 * to get handled on remaining nodes */
542 /* success! see if any other nodes need recovery */
543 mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
544 dlm->name, dlm->reco.dead_node, dlm->node_num);
545 dlm_reset_recovery(dlm);
547 dlm_end_recovery(dlm);
549 /* continue and look for another dead node */
553 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
556 struct dlm_reco_node_data *ndata;
562 /* we have become recovery master. there is no escaping
563 * this, so just keep trying until we get it. */
564 status = dlm_init_recovery_area(dlm, dead_node);
566 mlog(ML_ERROR, "%s: failed to alloc recovery area, "
567 "retrying\n", dlm->name);
570 } while (status != 0);
572 /* safe to access the node data list without a lock, since this
573 * process is the only one to change the list */
574 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
575 BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
576 ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
578 mlog(0, "%s: Requesting lock info from node %u\n", dlm->name,
581 if (ndata->node_num == dlm->node_num) {
582 ndata->state = DLM_RECO_NODE_DATA_DONE;
587 status = dlm_request_all_locks(dlm, ndata->node_num,
591 if (dlm_is_host_down(status)) {
592 /* node died, ignore it for recovery */
594 ndata->state = DLM_RECO_NODE_DATA_DEAD;
595 /* wait for the domain map to catch up
596 * with the network state. */
597 wait_event_timeout(dlm->dlm_reco_thread_wq,
598 dlm_is_node_dead(dlm,
600 msecs_to_jiffies(1000));
601 mlog(0, "waited 1 sec for %u, "
602 "dead? %s\n", ndata->node_num,
603 dlm_is_node_dead(dlm, ndata->node_num) ?
606 /* -ENOMEM on the other node */
607 mlog(0, "%s: node %u returned "
608 "%d during recovery, retrying "
609 "after a short wait\n",
610 dlm->name, ndata->node_num,
615 } while (status != 0);
617 spin_lock(&dlm_reco_state_lock);
618 switch (ndata->state) {
619 case DLM_RECO_NODE_DATA_INIT:
620 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
621 case DLM_RECO_NODE_DATA_REQUESTED:
624 case DLM_RECO_NODE_DATA_DEAD:
625 mlog(0, "node %u died after requesting "
626 "recovery info for node %u\n",
627 ndata->node_num, dead_node);
628 /* fine. don't need this node's info.
629 * continue without it. */
631 case DLM_RECO_NODE_DATA_REQUESTING:
632 ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
633 mlog(0, "now receiving recovery data from "
634 "node %u for dead node %u\n",
635 ndata->node_num, dead_node);
637 case DLM_RECO_NODE_DATA_RECEIVING:
638 mlog(0, "already receiving recovery data from "
639 "node %u for dead node %u\n",
640 ndata->node_num, dead_node);
642 case DLM_RECO_NODE_DATA_DONE:
643 mlog(0, "already DONE receiving recovery data "
644 "from node %u for dead node %u\n",
645 ndata->node_num, dead_node);
648 spin_unlock(&dlm_reco_state_lock);
651 mlog(0, "%s: Done requesting all lock info\n", dlm->name);
653 /* nodes should be sending reco data now
654 * just need to wait */
657 /* check all the nodes now to see if we are
658 * done, or if anyone died */
660 spin_lock(&dlm_reco_state_lock);
661 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
662 mlog(0, "checking recovery state of node %u\n",
664 switch (ndata->state) {
665 case DLM_RECO_NODE_DATA_INIT:
666 case DLM_RECO_NODE_DATA_REQUESTING:
667 mlog(ML_ERROR, "bad ndata state for "
668 "node %u: state=%d\n",
669 ndata->node_num, ndata->state);
672 case DLM_RECO_NODE_DATA_DEAD:
673 mlog(0, "node %u died after "
674 "requesting recovery info for "
675 "node %u\n", ndata->node_num,
678 case DLM_RECO_NODE_DATA_RECEIVING:
679 case DLM_RECO_NODE_DATA_REQUESTED:
680 mlog(0, "%s: node %u still in state %s\n",
681 dlm->name, ndata->node_num,
682 ndata->state==DLM_RECO_NODE_DATA_RECEIVING ?
683 "receiving" : "requested");
686 case DLM_RECO_NODE_DATA_DONE:
687 mlog(0, "%s: node %u state is done\n",
688 dlm->name, ndata->node_num);
690 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
691 mlog(0, "%s: node %u state is finalize\n",
692 dlm->name, ndata->node_num);
696 spin_unlock(&dlm_reco_state_lock);
698 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
699 all_nodes_done?"yes":"no");
700 if (all_nodes_done) {
703 /* all nodes are now in DLM_RECO_NODE_DATA_DONE state
704 * just send a finalize message to everyone and
706 mlog(0, "all nodes are done! send finalize\n");
707 ret = dlm_send_finalize_reco_message(dlm);
711 spin_lock(&dlm->spinlock);
712 dlm_finish_local_lockres_recovery(dlm, dead_node,
714 spin_unlock(&dlm->spinlock);
715 mlog(0, "should be done with recovery!\n");
717 mlog(0, "finishing recovery of %s at %lu, "
718 "dead=%u, this=%u, new=%u\n", dlm->name,
719 jiffies, dlm->reco.dead_node,
720 dlm->node_num, dlm->reco.new_master);
723 /* rescan everything marked dirty along the way */
724 dlm_kick_thread(dlm, NULL);
727 /* wait to be signalled, with periodic timeout
728 * to check for node death */
729 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
730 kthread_should_stop(),
731 msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
736 dlm_destroy_recovery_area(dlm, dead_node);
741 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
744 struct dlm_reco_node_data *ndata;
746 spin_lock(&dlm->spinlock);
747 memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
748 /* nodes can only be removed (by dying) after dropping
749 * this lock, and death will be trapped later, so this should do */
750 spin_unlock(&dlm->spinlock);
753 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
754 if (num >= O2NM_MAX_NODES) {
757 BUG_ON(num == dead_node);
759 ndata = kzalloc(sizeof(*ndata), GFP_NOFS);
761 dlm_destroy_recovery_area(dlm, dead_node);
764 ndata->node_num = num;
765 ndata->state = DLM_RECO_NODE_DATA_INIT;
766 spin_lock(&dlm_reco_state_lock);
767 list_add_tail(&ndata->list, &dlm->reco.node_data);
768 spin_unlock(&dlm_reco_state_lock);
775 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
777 struct dlm_reco_node_data *ndata, *next;
780 spin_lock(&dlm_reco_state_lock);
781 list_splice_init(&dlm->reco.node_data, &tmplist);
782 spin_unlock(&dlm_reco_state_lock);
784 list_for_each_entry_safe(ndata, next, &tmplist, list) {
785 list_del_init(&ndata->list);
790 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
793 struct dlm_lock_request lr;
799 mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
800 "to %u\n", dead_node, request_from);
802 memset(&lr, 0, sizeof(lr));
803 lr.node_idx = dlm->node_num;
804 lr.dead_node = dead_node;
808 ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
809 &lr, sizeof(lr), request_from, NULL);
811 /* negative status is handled by caller */
813 mlog(ML_ERROR, "%s: Error %d send LOCK_REQUEST to node %u "
814 "to recover dead node %u\n", dlm->name, ret,
815 request_from, dead_node);
816 // return from here, then
817 // sleep until all received or error
822 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
825 struct dlm_ctxt *dlm = data;
826 struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
828 struct dlm_work_item *item = NULL;
833 if (lr->dead_node != dlm->reco.dead_node) {
834 mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local "
835 "dead_node is %u\n", dlm->name, lr->node_idx,
836 lr->dead_node, dlm->reco.dead_node);
837 dlm_print_reco_node_status(dlm);
842 BUG_ON(lr->dead_node != dlm->reco.dead_node);
844 item = kzalloc(sizeof(*item), GFP_NOFS);
850 /* this will get freed by dlm_request_all_locks_worker */
851 buf = (char *) __get_free_page(GFP_NOFS);
858 /* queue up work for dlm_request_all_locks_worker */
859 dlm_grab(dlm); /* get an extra ref for the work item */
860 dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
861 item->u.ral.reco_master = lr->node_idx;
862 item->u.ral.dead_node = lr->dead_node;
863 spin_lock(&dlm->work_lock);
864 list_add_tail(&item->list, &dlm->work_list);
865 spin_unlock(&dlm->work_lock);
866 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
872 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
874 struct dlm_migratable_lockres *mres;
875 struct dlm_lock_resource *res;
876 struct dlm_ctxt *dlm;
877 LIST_HEAD(resources);
879 u8 dead_node, reco_master;
880 int skip_all_done = 0;
883 dead_node = item->u.ral.dead_node;
884 reco_master = item->u.ral.reco_master;
885 mres = (struct dlm_migratable_lockres *)data;
887 mlog(0, "%s: recovery worker started, dead=%u, master=%u\n",
888 dlm->name, dead_node, reco_master);
890 if (dead_node != dlm->reco.dead_node ||
891 reco_master != dlm->reco.new_master) {
892 /* worker could have been created before the recovery master
893 * died. if so, do not continue, but do not error. */
894 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
895 mlog(ML_NOTICE, "%s: will not send recovery state, "
896 "recovery master %u died, thread=(dead=%u,mas=%u)"
897 " current=(dead=%u,mas=%u)\n", dlm->name,
898 reco_master, dead_node, reco_master,
899 dlm->reco.dead_node, dlm->reco.new_master);
901 mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, "
902 "master=%u), request(dead=%u, master=%u)\n",
903 dlm->name, dlm->reco.dead_node,
904 dlm->reco.new_master, dead_node, reco_master);
909 /* lock resources should have already been moved to the
910 * dlm->reco.resources list. now move items from that list
911 * to a temp list if the dead owner matches. note that the
912 * whole cluster recovers only one node at a time, so we
913 * can safely move UNKNOWN lock resources for each recovery
915 dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
917 /* now we can begin blasting lockreses without the dlm lock */
919 /* any errors returned will be due to the new_master dying,
920 * the dlm_reco_thread should detect this */
921 list_for_each_entry(res, &resources, recovering) {
922 ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
925 mlog(ML_ERROR, "%s: node %u went down while sending "
926 "recovery state for dead node %u, ret=%d\n", dlm->name,
927 reco_master, dead_node, ret);
933 /* move the resources back to the list */
934 spin_lock(&dlm->spinlock);
935 list_splice_init(&resources, &dlm->reco.resources);
936 spin_unlock(&dlm->spinlock);
938 if (!skip_all_done) {
939 ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
941 mlog(ML_ERROR, "%s: node %u went down while sending "
942 "recovery all-done for dead node %u, ret=%d\n",
943 dlm->name, reco_master, dead_node, ret);
947 free_page((unsigned long)data);
951 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
954 struct dlm_reco_data_done done_msg;
956 memset(&done_msg, 0, sizeof(done_msg));
957 done_msg.node_idx = dlm->node_num;
958 done_msg.dead_node = dead_node;
959 mlog(0, "sending DATA DONE message to %u, "
960 "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
963 ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
964 sizeof(done_msg), send_to, &tmpret);
966 mlog(ML_ERROR, "%s: Error %d send RECO_DATA_DONE to node %u "
967 "to recover dead node %u\n", dlm->name, ret, send_to,
969 if (!dlm_is_host_down(ret)) {
978 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
981 struct dlm_ctxt *dlm = data;
982 struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
983 struct dlm_reco_node_data *ndata = NULL;
989 mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
990 "node_idx=%u, this node=%u\n", done->dead_node,
991 dlm->reco.dead_node, done->node_idx, dlm->node_num);
993 mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node),
994 "Got DATA DONE: dead_node=%u, reco.dead_node=%u, "
995 "node_idx=%u, this node=%u\n", done->dead_node,
996 dlm->reco.dead_node, done->node_idx, dlm->node_num);
998 spin_lock(&dlm_reco_state_lock);
999 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
1000 if (ndata->node_num != done->node_idx)
1003 switch (ndata->state) {
1004 /* should have moved beyond INIT but not to FINALIZE yet */
1005 case DLM_RECO_NODE_DATA_INIT:
1006 case DLM_RECO_NODE_DATA_DEAD:
1007 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
1008 mlog(ML_ERROR, "bad ndata state for node %u:"
1009 " state=%d\n", ndata->node_num,
1013 /* these states are possible at this point, anywhere along
1014 * the line of recovery */
1015 case DLM_RECO_NODE_DATA_DONE:
1016 case DLM_RECO_NODE_DATA_RECEIVING:
1017 case DLM_RECO_NODE_DATA_REQUESTED:
1018 case DLM_RECO_NODE_DATA_REQUESTING:
1019 mlog(0, "node %u is DONE sending "
1023 ndata->state = DLM_RECO_NODE_DATA_DONE;
1028 spin_unlock(&dlm_reco_state_lock);
1030 /* wake the recovery thread, some node is done */
1032 dlm_kick_recovery_thread(dlm);
1035 mlog(ML_ERROR, "failed to find recovery node data for node "
1036 "%u\n", done->node_idx);
1039 mlog(0, "leaving reco data done handler, ret=%d\n", ret);
1043 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
1044 struct list_head *list,
1047 struct dlm_lock_resource *res, *next;
1048 struct dlm_lock *lock;
1050 spin_lock(&dlm->spinlock);
1051 list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
1052 /* always prune any $RECOVERY entries for dead nodes,
1053 * otherwise hangs can occur during later recovery */
1054 if (dlm_is_recovery_lock(res->lockname.name,
1055 res->lockname.len)) {
1056 spin_lock(&res->spinlock);
1057 list_for_each_entry(lock, &res->granted, list) {
1058 if (lock->ml.node == dead_node) {
1059 mlog(0, "AHA! there was "
1060 "a $RECOVERY lock for dead "
1062 dead_node, dlm->name);
1063 list_del_init(&lock->list);
1068 spin_unlock(&res->spinlock);
1072 if (res->owner == dead_node) {
1073 mlog(0, "found lockres owned by dead node while "
1074 "doing recovery for node %u. sending it.\n",
1076 list_move_tail(&res->recovering, list);
1077 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1078 mlog(0, "found UNKNOWN owner while doing recovery "
1079 "for node %u. sending it.\n", dead_node);
1080 list_move_tail(&res->recovering, list);
1083 spin_unlock(&dlm->spinlock);
1086 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
1088 int total_locks = 0;
1089 struct list_head *iter, *queue = &res->granted;
1092 for (i=0; i<3; i++) {
1093 list_for_each(iter, queue)
1101 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
1102 struct dlm_migratable_lockres *mres,
1104 struct dlm_lock_resource *res,
1107 u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
1108 int mres_total_locks = be32_to_cpu(mres->total_locks);
1109 int sz, ret = 0, status = 0;
1110 u8 orig_flags = mres->flags,
1111 orig_master = mres->master;
1113 BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
1114 if (!mres->num_locks)
1117 sz = sizeof(struct dlm_migratable_lockres) +
1118 (mres->num_locks * sizeof(struct dlm_migratable_lock));
1120 /* add an all-done flag if we reached the last lock */
1121 orig_flags = mres->flags;
1122 BUG_ON(total_locks > mres_total_locks);
1123 if (total_locks == mres_total_locks)
1124 mres->flags |= DLM_MRES_ALL_DONE;
1126 mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n",
1127 dlm->name, res->lockname.len, res->lockname.name,
1128 orig_flags & DLM_MRES_MIGRATION ? "migration" : "recovery",
1132 ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
1133 sz, send_to, &status);
1135 /* XXX: negative status is not handled.
1136 * this will end up killing this node. */
1137 mlog(ML_ERROR, "%s: res %.*s, Error %d send MIG_LOCKRES to "
1138 "node %u (%s)\n", dlm->name, mres->lockname_len,
1139 mres->lockname, ret, send_to,
1140 (orig_flags & DLM_MRES_MIGRATION ?
1141 "migration" : "recovery"));
1143 /* might get an -ENOMEM back here */
1148 if (ret == -EFAULT) {
1149 mlog(ML_ERROR, "node %u told me to kill "
1150 "myself!\n", send_to);
1156 /* zero and reinit the message buffer */
1157 dlm_init_migratable_lockres(mres, res->lockname.name,
1158 res->lockname.len, mres_total_locks,
1159 mig_cookie, orig_flags, orig_master);
1163 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
1164 const char *lockname, int namelen,
1165 int total_locks, u64 cookie,
1166 u8 flags, u8 master)
1168 /* mres here is one full page */
1170 mres->lockname_len = namelen;
1171 memcpy(mres->lockname, lockname, namelen);
1172 mres->num_locks = 0;
1173 mres->total_locks = cpu_to_be32(total_locks);
1174 mres->mig_cookie = cpu_to_be64(cookie);
1175 mres->flags = flags;
1176 mres->master = master;
1179 static void dlm_prepare_lvb_for_migration(struct dlm_lock *lock,
1180 struct dlm_migratable_lockres *mres,
1186 /* Ignore lvb in all locks in the blocked list */
1187 if (queue == DLM_BLOCKED_LIST)
1190 /* Only consider lvbs in locks with granted EX or PR lock levels */
1191 if (lock->ml.type != LKM_EXMODE && lock->ml.type != LKM_PRMODE)
1194 if (dlm_lvb_is_empty(mres->lvb)) {
1195 memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1199 /* Ensure the lvb copied for migration matches in other valid locks */
1200 if (!memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))
1203 mlog(ML_ERROR, "Mismatched lvb in lock cookie=%u:%llu, name=%.*s, "
1205 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
1206 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
1207 lock->lockres->lockname.len, lock->lockres->lockname.name,
1209 dlm_print_one_lock_resource(lock->lockres);
1213 /* returns 1 if this lock fills the network structure,
1215 static int dlm_add_lock_to_array(struct dlm_lock *lock,
1216 struct dlm_migratable_lockres *mres, int queue)
1218 struct dlm_migratable_lock *ml;
1219 int lock_num = mres->num_locks;
1221 ml = &(mres->ml[lock_num]);
1222 ml->cookie = lock->ml.cookie;
1223 ml->type = lock->ml.type;
1224 ml->convert_type = lock->ml.convert_type;
1225 ml->highest_blocked = lock->ml.highest_blocked;
1228 ml->flags = lock->lksb->flags;
1229 dlm_prepare_lvb_for_migration(lock, mres, queue);
1231 ml->node = lock->ml.node;
1233 /* we reached the max, send this network message */
1234 if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1239 static void dlm_add_dummy_lock(struct dlm_ctxt *dlm,
1240 struct dlm_migratable_lockres *mres)
1242 struct dlm_lock dummy;
1243 memset(&dummy, 0, sizeof(dummy));
1244 dummy.ml.cookie = 0;
1245 dummy.ml.type = LKM_IVMODE;
1246 dummy.ml.convert_type = LKM_IVMODE;
1247 dummy.ml.highest_blocked = LKM_IVMODE;
1249 dummy.ml.node = dlm->node_num;
1250 dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST);
1253 static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm,
1254 struct dlm_migratable_lock *ml,
1257 if (unlikely(ml->cookie == 0 &&
1258 ml->type == LKM_IVMODE &&
1259 ml->convert_type == LKM_IVMODE &&
1260 ml->highest_blocked == LKM_IVMODE &&
1261 ml->list == DLM_BLOCKED_LIST)) {
1262 *nodenum = ml->node;
1268 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1269 struct dlm_migratable_lockres *mres,
1270 u8 send_to, u8 flags)
1272 struct list_head *queue;
1275 struct dlm_lock *lock;
1278 BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1280 mlog(0, "sending to %u\n", send_to);
1282 total_locks = dlm_num_locks_in_lockres(res);
1283 if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1284 /* rare, but possible */
1285 mlog(0, "argh. lockres has %d locks. this will "
1286 "require more than one network packet to "
1287 "migrate\n", total_locks);
1288 mig_cookie = dlm_get_next_mig_cookie();
1291 dlm_init_migratable_lockres(mres, res->lockname.name,
1292 res->lockname.len, total_locks,
1293 mig_cookie, flags, res->owner);
1296 for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1297 queue = dlm_list_idx_to_ptr(res, i);
1298 list_for_each_entry(lock, queue, list) {
1299 /* add another lock. */
1301 if (!dlm_add_lock_to_array(lock, mres, i))
1304 /* this filled the lock message,
1305 * we must send it immediately. */
1306 ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1312 if (total_locks == 0) {
1313 /* send a dummy lock to indicate a mastery reference only */
1314 mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n",
1315 dlm->name, res->lockname.len, res->lockname.name,
1316 send_to, flags & DLM_MRES_RECOVERY ? "recovery" :
1318 dlm_add_dummy_lock(dlm, mres);
1320 /* flush any remaining locks */
1321 ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1327 mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n",
1329 if (!dlm_is_host_down(ret))
1331 mlog(0, "%s: node %u went down while sending %s "
1332 "lockres %.*s\n", dlm->name, send_to,
1333 flags & DLM_MRES_RECOVERY ? "recovery" : "migration",
1334 res->lockname.len, res->lockname.name);
1341 * this message will contain no more than one page worth of
1342 * recovery data, and it will work on only one lockres.
1343 * there may be many locks in this page, and we may need to wait
1344 * for additional packets to complete all the locks (rare, but
1348 * NOTE: the allocation error cases here are scary
1349 * we really cannot afford to fail an alloc in recovery
1350 * do we spin? returning an error only delays the problem really
1353 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
1356 struct dlm_ctxt *dlm = data;
1357 struct dlm_migratable_lockres *mres =
1358 (struct dlm_migratable_lockres *)msg->buf;
1363 struct dlm_work_item *item = NULL;
1364 struct dlm_lock_resource *res = NULL;
1369 BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1371 real_master = mres->master;
1372 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1373 /* cannot migrate a lockres with no master */
1374 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1377 mlog(0, "%s message received from node %u\n",
1378 (mres->flags & DLM_MRES_RECOVERY) ?
1379 "recovery" : "migration", mres->master);
1380 if (mres->flags & DLM_MRES_ALL_DONE)
1381 mlog(0, "all done flag. all lockres data received!\n");
1384 buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS);
1385 item = kzalloc(sizeof(*item), GFP_NOFS);
1389 /* lookup the lock to see if we have a secondary queue for this
1390 * already... just add the locks in and this will have its owner
1391 * and RECOVERY flag changed when it completes. */
1392 res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len);
1394 /* this will get a ref on res */
1395 /* mark it as recovering/migrating and hash it */
1396 spin_lock(&res->spinlock);
1397 if (mres->flags & DLM_MRES_RECOVERY) {
1398 res->state |= DLM_LOCK_RES_RECOVERING;
1400 if (res->state & DLM_LOCK_RES_MIGRATING) {
1401 /* this is at least the second
1402 * lockres message */
1403 mlog(0, "lock %.*s is already migrating\n",
1406 } else if (res->state & DLM_LOCK_RES_RECOVERING) {
1407 /* caller should BUG */
1408 mlog(ML_ERROR, "node is attempting to migrate "
1409 "lock %.*s, but marked as recovering!\n",
1410 mres->lockname_len, mres->lockname);
1412 spin_unlock(&res->spinlock);
1415 res->state |= DLM_LOCK_RES_MIGRATING;
1417 spin_unlock(&res->spinlock);
1419 /* need to allocate, just like if it was
1420 * mastered here normally */
1421 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1425 /* to match the ref that we would have gotten if
1426 * dlm_lookup_lockres had succeeded */
1427 dlm_lockres_get(res);
1429 /* mark it as recovering/migrating and hash it */
1430 if (mres->flags & DLM_MRES_RECOVERY)
1431 res->state |= DLM_LOCK_RES_RECOVERING;
1433 res->state |= DLM_LOCK_RES_MIGRATING;
1435 spin_lock(&dlm->spinlock);
1436 __dlm_insert_lockres(dlm, res);
1437 spin_unlock(&dlm->spinlock);
1439 /* Add an extra ref for this lock-less lockres lest the
1440 * dlm_thread purges it before we get the chance to add
1442 dlm_lockres_get(res);
1444 /* There are three refs that need to be put.
1446 * 2. kref_init in dlm_new_lockres()->dlm_init_lockres().
1447 * 3. dlm_lookup_lockres()
1448 * The first one is handled at the end of this function. The
1449 * other two are handled in the worker thread after locks have
1450 * been attached. Yes, we don't wait for purge time to match
1451 * kref_init. The lockres will still have atleast one ref
1452 * added because it is in the hash __dlm_insert_lockres() */
1455 /* now that the new lockres is inserted,
1456 * make it usable by other processes */
1457 spin_lock(&res->spinlock);
1458 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1459 spin_unlock(&res->spinlock);
1463 /* at this point we have allocated everything we need,
1464 * and we have a hashed lockres with an extra ref and
1465 * the proper res->state flags. */
1467 spin_lock(&res->spinlock);
1468 /* drop this either when master requery finds a different master
1469 * or when a lock is added by the recovery worker */
1470 dlm_lockres_grab_inflight_ref(dlm, res);
1471 if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1472 /* migration cannot have an unknown master */
1473 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1474 mlog(0, "recovery has passed me a lockres with an "
1475 "unknown owner.. will need to requery: "
1476 "%.*s\n", mres->lockname_len, mres->lockname);
1478 /* take a reference now to pin the lockres, drop it
1479 * when locks are added in the worker */
1480 dlm_change_lockres_owner(dlm, res, dlm->node_num);
1482 spin_unlock(&res->spinlock);
1484 /* queue up work for dlm_mig_lockres_worker */
1485 dlm_grab(dlm); /* get an extra ref for the work item */
1486 memcpy(buf, msg->buf, be16_to_cpu(msg->data_len)); /* copy the whole message */
1487 dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1488 item->u.ml.lockres = res; /* already have a ref */
1489 item->u.ml.real_master = real_master;
1490 item->u.ml.extra_ref = extra_refs;
1491 spin_lock(&dlm->work_lock);
1492 list_add_tail(&item->list, &dlm->work_list);
1493 spin_unlock(&dlm->work_lock);
1494 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
1497 /* One extra ref taken needs to be put here */
1499 dlm_lockres_put(res);
1514 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1516 struct dlm_ctxt *dlm;
1517 struct dlm_migratable_lockres *mres;
1519 struct dlm_lock_resource *res;
1524 mres = (struct dlm_migratable_lockres *)data;
1526 res = item->u.ml.lockres;
1527 real_master = item->u.ml.real_master;
1528 extra_ref = item->u.ml.extra_ref;
1530 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1531 /* this case is super-rare. only occurs if
1532 * node death happens during migration. */
1534 ret = dlm_lockres_master_requery(dlm, res, &real_master);
1536 mlog(0, "dlm_lockres_master_requery ret=%d\n",
1540 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1541 mlog(0, "lockres %.*s not claimed. "
1542 "this node will take it.\n",
1543 res->lockname.len, res->lockname.name);
1545 spin_lock(&res->spinlock);
1546 dlm_lockres_drop_inflight_ref(dlm, res);
1547 spin_unlock(&res->spinlock);
1548 mlog(0, "master needs to respond to sender "
1549 "that node %u still owns %.*s\n",
1550 real_master, res->lockname.len,
1551 res->lockname.name);
1552 /* cannot touch this lockres */
1557 ret = dlm_process_recovery_data(dlm, res, mres);
1559 mlog(0, "dlm_process_recovery_data returned %d\n", ret);
1561 mlog(0, "dlm_process_recovery_data succeeded\n");
1563 if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1564 (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1565 ret = dlm_finish_migration(dlm, res, mres->master);
1571 /* See comment in dlm_mig_lockres_handler() */
1574 dlm_lockres_put(res);
1575 dlm_lockres_put(res);
1582 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1583 struct dlm_lock_resource *res,
1586 struct dlm_node_iter iter;
1590 *real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1592 /* we only reach here if one of the two nodes in a
1593 * migration died while the migration was in progress.
1594 * at this point we need to requery the master. we
1595 * know that the new_master got as far as creating
1596 * an mle on at least one node, but we do not know
1597 * if any nodes had actually cleared the mle and set
1598 * the master to the new_master. the old master
1599 * is supposed to set the owner to UNKNOWN in the
1600 * event of a new_master death, so the only possible
1601 * responses that we can get from nodes here are
1602 * that the master is new_master, or that the master
1604 * if all nodes come back with UNKNOWN then we know
1605 * the lock needs remastering here.
1606 * if any node comes back with a valid master, check
1607 * to see if that master is the one that we are
1608 * recovering. if so, then the new_master died and
1609 * we need to remaster this lock. if not, then the
1610 * new_master survived and that node will respond to
1611 * other nodes about the owner.
1612 * if there is an owner, this node needs to dump this
1613 * lockres and alert the sender that this lockres
1615 spin_lock(&dlm->spinlock);
1616 dlm_node_iter_init(dlm->domain_map, &iter);
1617 spin_unlock(&dlm->spinlock);
1619 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1620 /* do not send to self */
1621 if (nodenum == dlm->node_num)
1623 ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1626 if (!dlm_is_host_down(ret))
1628 /* host is down, so answer for that node would be
1629 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
1631 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1632 mlog(0, "lock master is %u\n", *real_master);
1640 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1641 u8 nodenum, u8 *real_master)
1644 struct dlm_master_requery req;
1645 int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1647 memset(&req, 0, sizeof(req));
1648 req.node_idx = dlm->node_num;
1649 req.namelen = res->lockname.len;
1650 memcpy(req.name, res->lockname.name, res->lockname.len);
1652 ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1653 &req, sizeof(req), nodenum, &status);
1654 /* XXX: negative status not handled properly here. */
1656 mlog(ML_ERROR, "Error %d when sending message %u (key "
1657 "0x%x) to node %u\n", ret, DLM_MASTER_REQUERY_MSG,
1661 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1662 *real_master = (u8) (status & 0xff);
1663 mlog(0, "node %u responded to master requery with %u\n",
1664 nodenum, *real_master);
1671 /* this function cannot error, so unless the sending
1672 * or receiving of the message failed, the owner can
1674 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
1677 struct dlm_ctxt *dlm = data;
1678 struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1679 struct dlm_lock_resource *res = NULL;
1681 int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1682 u32 flags = DLM_ASSERT_MASTER_REQUERY;
1684 if (!dlm_grab(dlm)) {
1685 /* since the domain has gone away on this
1686 * node, the proper response is UNKNOWN */
1690 hash = dlm_lockid_hash(req->name, req->namelen);
1692 spin_lock(&dlm->spinlock);
1693 res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash);
1695 spin_lock(&res->spinlock);
1696 master = res->owner;
1697 if (master == dlm->node_num) {
1698 int ret = dlm_dispatch_assert_master(dlm, res,
1701 mlog_errno(-ENOMEM);
1705 } else /* put.. incase we are not the master */
1706 dlm_lockres_put(res);
1707 spin_unlock(&res->spinlock);
1709 spin_unlock(&dlm->spinlock);
1715 static inline struct list_head *
1716 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1718 struct list_head *ret;
1719 BUG_ON(list_num < 0);
1720 BUG_ON(list_num > 2);
1721 ret = &(res->granted);
1725 /* TODO: do ast flush business
1726 * TODO: do MIGRATING and RECOVERING spinning
1730 * NOTE about in-flight requests during migration:
1732 * Before attempting the migrate, the master has marked the lockres as
1733 * MIGRATING and then flushed all of its pending ASTS. So any in-flight
1734 * requests either got queued before the MIGRATING flag got set, in which
1735 * case the lock data will reflect the change and a return message is on
1736 * the way, or the request failed to get in before MIGRATING got set. In
1737 * this case, the caller will be told to spin and wait for the MIGRATING
1738 * flag to be dropped, then recheck the master.
1739 * This holds true for the convert, cancel and unlock cases, and since lvb
1740 * updates are tied to these same messages, it applies to lvb updates as
1741 * well. For the lock case, there is no way a lock can be on the master
1742 * queue and not be on the secondary queue since the lock is always added
1743 * locally first. This means that the new target node will never be sent
1744 * a lock that he doesn't already have on the list.
1745 * In total, this means that the local lock is correct and should not be
1746 * updated to match the one sent by the master. Any messages sent back
1747 * from the master before the MIGRATING flag will bring the lock properly
1748 * up-to-date, and the change will be ordered properly for the waiter.
1749 * We will *not* attempt to modify the lock underneath the waiter.
1752 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1753 struct dlm_lock_resource *res,
1754 struct dlm_migratable_lockres *mres)
1756 struct dlm_migratable_lock *ml;
1757 struct list_head *queue;
1758 struct list_head *tmpq = NULL;
1759 struct dlm_lock *newlock = NULL;
1760 struct dlm_lockstatus *lksb = NULL;
1763 struct dlm_lock *lock = NULL;
1764 u8 from = O2NM_MAX_NODES;
1765 unsigned int added = 0;
1768 mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1769 for (i=0; i<mres->num_locks; i++) {
1770 ml = &(mres->ml[i]);
1772 if (dlm_is_dummy_lock(dlm, ml, &from)) {
1773 /* placeholder, just need to set the refmap bit */
1774 BUG_ON(mres->num_locks != 1);
1775 mlog(0, "%s:%.*s: dummy lock for %u\n",
1776 dlm->name, mres->lockname_len, mres->lockname,
1778 spin_lock(&res->spinlock);
1779 dlm_lockres_set_refmap_bit(from, res);
1780 spin_unlock(&res->spinlock);
1784 BUG_ON(ml->highest_blocked != LKM_IVMODE);
1788 queue = dlm_list_num_to_pointer(res, ml->list);
1791 /* if the lock is for the local node it needs to
1792 * be moved to the proper location within the queue.
1793 * do not allocate a new lock structure. */
1794 if (ml->node == dlm->node_num) {
1795 /* MIGRATION ONLY! */
1796 BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1798 spin_lock(&res->spinlock);
1799 for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) {
1800 tmpq = dlm_list_idx_to_ptr(res, j);
1801 list_for_each_entry(lock, tmpq, list) {
1802 if (lock->ml.cookie != ml->cookie)
1811 /* lock is always created locally first, and
1812 * destroyed locally last. it must be on the list */
1815 mlog(ML_ERROR, "Could not find local lock "
1816 "with cookie %u:%llu, node %u, "
1817 "list %u, flags 0x%x, type %d, "
1818 "conv %d, highest blocked %d\n",
1819 dlm_get_lock_cookie_node(be64_to_cpu(c)),
1820 dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1821 ml->node, ml->list, ml->flags, ml->type,
1822 ml->convert_type, ml->highest_blocked);
1823 __dlm_print_one_lock_resource(res);
1827 if (lock->ml.node != ml->node) {
1828 c = lock->ml.cookie;
1829 mlog(ML_ERROR, "Mismatched node# in lock "
1830 "cookie %u:%llu, name %.*s, node %u\n",
1831 dlm_get_lock_cookie_node(be64_to_cpu(c)),
1832 dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1833 res->lockname.len, res->lockname.name,
1836 mlog(ML_ERROR, "Migrate lock cookie %u:%llu, "
1837 "node %u, list %u, flags 0x%x, type %d, "
1838 "conv %d, highest blocked %d\n",
1839 dlm_get_lock_cookie_node(be64_to_cpu(c)),
1840 dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1841 ml->node, ml->list, ml->flags, ml->type,
1842 ml->convert_type, ml->highest_blocked);
1843 __dlm_print_one_lock_resource(res);
1847 if (tmpq != queue) {
1849 mlog(0, "Lock cookie %u:%llu was on list %u "
1850 "instead of list %u for %.*s\n",
1851 dlm_get_lock_cookie_node(be64_to_cpu(c)),
1852 dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1853 j, ml->list, res->lockname.len,
1854 res->lockname.name);
1855 __dlm_print_one_lock_resource(res);
1856 spin_unlock(&res->spinlock);
1860 /* see NOTE above about why we do not update
1861 * to match the master here */
1863 /* move the lock to its proper place */
1864 /* do not alter lock refcount. switching lists. */
1865 list_move_tail(&lock->list, queue);
1866 spin_unlock(&res->spinlock);
1869 mlog(0, "just reordered a local lock!\n");
1873 /* lock is for another node. */
1874 newlock = dlm_new_lock(ml->type, ml->node,
1875 be64_to_cpu(ml->cookie), NULL);
1880 lksb = newlock->lksb;
1881 dlm_lock_attach_lockres(newlock, res);
1883 if (ml->convert_type != LKM_IVMODE) {
1884 BUG_ON(queue != &res->converting);
1885 newlock->ml.convert_type = ml->convert_type;
1887 lksb->flags |= (ml->flags &
1888 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1890 if (ml->type == LKM_NLMODE)
1893 if (!dlm_lvb_is_empty(mres->lvb)) {
1894 if (lksb->flags & DLM_LKSB_PUT_LVB) {
1895 /* other node was trying to update
1896 * lvb when node died. recreate the
1897 * lksb with the updated lvb. */
1898 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1899 /* the lock resource lvb update must happen
1900 * NOW, before the spinlock is dropped.
1901 * we no longer wait for the AST to update
1903 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1905 /* otherwise, the node is sending its
1906 * most recent valid lvb info */
1907 BUG_ON(ml->type != LKM_EXMODE &&
1908 ml->type != LKM_PRMODE);
1909 if (!dlm_lvb_is_empty(res->lvb) &&
1910 (ml->type == LKM_EXMODE ||
1911 memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1913 mlog(ML_ERROR, "%s:%.*s: received bad "
1914 "lvb! type=%d\n", dlm->name,
1916 res->lockname.name, ml->type);
1917 printk("lockres lvb=[");
1918 for (i=0; i<DLM_LVB_LEN; i++)
1919 printk("%02x", res->lvb[i]);
1920 printk("]\nmigrated lvb=[");
1921 for (i=0; i<DLM_LVB_LEN; i++)
1922 printk("%02x", mres->lvb[i]);
1924 dlm_print_one_lock_resource(res);
1927 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1933 * wrt lock queue ordering and recovery:
1934 * 1. order of locks on granted queue is
1936 * 2. order of locks on converting queue is
1937 * LOST with the node death. sorry charlie.
1938 * 3. order of locks on the blocked queue is
1940 * order of locks does not affect integrity, it
1941 * just means that a lock request may get pushed
1942 * back in line as a result of the node death.
1943 * also note that for a given node the lock order
1944 * for its secondary queue locks is preserved
1945 * relative to each other, but clearly *not*
1946 * preserved relative to locks from other nodes.
1949 spin_lock(&res->spinlock);
1950 list_for_each_entry(lock, queue, list) {
1951 if (lock->ml.cookie == ml->cookie) {
1952 c = lock->ml.cookie;
1953 mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already "
1954 "exists on this lockres!\n", dlm->name,
1955 res->lockname.len, res->lockname.name,
1956 dlm_get_lock_cookie_node(be64_to_cpu(c)),
1957 dlm_get_lock_cookie_seq(be64_to_cpu(c)));
1959 mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, "
1960 "node=%u, cookie=%u:%llu, queue=%d\n",
1961 ml->type, ml->convert_type, ml->node,
1962 dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)),
1963 dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)),
1966 __dlm_print_one_lock_resource(res);
1972 dlm_lock_get(newlock);
1973 list_add_tail(&newlock->list, queue);
1974 mlog(0, "%s:%.*s: added lock for node %u, "
1975 "setting refmap bit\n", dlm->name,
1976 res->lockname.len, res->lockname.name, ml->node);
1977 dlm_lockres_set_refmap_bit(ml->node, res);
1980 spin_unlock(&res->spinlock);
1982 mlog(0, "done running all the locks\n");
1985 /* balance the ref taken when the work was queued */
1986 spin_lock(&res->spinlock);
1987 dlm_lockres_drop_inflight_ref(dlm, res);
1988 spin_unlock(&res->spinlock);
1993 dlm_lock_put(newlock);
1999 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
2000 struct dlm_lock_resource *res)
2003 struct list_head *queue;
2004 struct dlm_lock *lock, *next;
2006 assert_spin_locked(&dlm->spinlock);
2007 assert_spin_locked(&res->spinlock);
2008 res->state |= DLM_LOCK_RES_RECOVERING;
2009 if (!list_empty(&res->recovering)) {
2011 "Recovering res %s:%.*s, is already on recovery list!\n",
2012 dlm->name, res->lockname.len, res->lockname.name);
2013 list_del_init(&res->recovering);
2014 dlm_lockres_put(res);
2016 /* We need to hold a reference while on the recovery list */
2017 dlm_lockres_get(res);
2018 list_add_tail(&res->recovering, &dlm->reco.resources);
2020 /* find any pending locks and put them back on proper list */
2021 for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
2022 queue = dlm_list_idx_to_ptr(res, i);
2023 list_for_each_entry_safe(lock, next, queue, list) {
2025 if (lock->convert_pending) {
2026 /* move converting lock back to granted */
2027 BUG_ON(i != DLM_CONVERTING_LIST);
2028 mlog(0, "node died with convert pending "
2029 "on %.*s. move back to granted list.\n",
2030 res->lockname.len, res->lockname.name);
2031 dlm_revert_pending_convert(res, lock);
2032 lock->convert_pending = 0;
2033 } else if (lock->lock_pending) {
2034 /* remove pending lock requests completely */
2035 BUG_ON(i != DLM_BLOCKED_LIST);
2036 mlog(0, "node died with lock pending "
2037 "on %.*s. remove from blocked list and skip.\n",
2038 res->lockname.len, res->lockname.name);
2039 /* lock will be floating until ref in
2040 * dlmlock_remote is freed after the network
2041 * call returns. ok for it to not be on any
2042 * list since no ast can be called
2043 * (the master is dead). */
2044 dlm_revert_pending_lock(res, lock);
2045 lock->lock_pending = 0;
2046 } else if (lock->unlock_pending) {
2047 /* if an unlock was in progress, treat as
2048 * if this had completed successfully
2049 * before sending this lock state to the
2050 * new master. note that the dlm_unlock
2051 * call is still responsible for calling
2052 * the unlockast. that will happen after
2053 * the network call times out. for now,
2054 * just move lists to prepare the new
2055 * recovery master. */
2056 BUG_ON(i != DLM_GRANTED_LIST);
2057 mlog(0, "node died with unlock pending "
2058 "on %.*s. remove from blocked list and skip.\n",
2059 res->lockname.len, res->lockname.name);
2060 dlm_commit_pending_unlock(res, lock);
2061 lock->unlock_pending = 0;
2062 } else if (lock->cancel_pending) {
2063 /* if a cancel was in progress, treat as
2064 * if this had completed successfully
2065 * before sending this lock state to the
2067 BUG_ON(i != DLM_CONVERTING_LIST);
2068 mlog(0, "node died with cancel pending "
2069 "on %.*s. move back to granted list.\n",
2070 res->lockname.len, res->lockname.name);
2071 dlm_commit_pending_cancel(res, lock);
2072 lock->cancel_pending = 0;
2081 /* removes all recovered locks from the recovery list.
2082 * sets the res->owner to the new master.
2083 * unsets the RECOVERY flag and wakes waiters. */
2084 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
2085 u8 dead_node, u8 new_master)
2088 struct hlist_node *hash_iter;
2089 struct hlist_head *bucket;
2090 struct dlm_lock_resource *res, *next;
2092 assert_spin_locked(&dlm->spinlock);
2094 list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
2095 if (res->owner == dead_node) {
2096 mlog(0, "%s: res %.*s, Changing owner from %u to %u\n",
2097 dlm->name, res->lockname.len, res->lockname.name,
2098 res->owner, new_master);
2099 list_del_init(&res->recovering);
2100 spin_lock(&res->spinlock);
2101 /* new_master has our reference from
2102 * the lock state sent during recovery */
2103 dlm_change_lockres_owner(dlm, res, new_master);
2104 res->state &= ~DLM_LOCK_RES_RECOVERING;
2105 if (__dlm_lockres_has_locks(res))
2106 __dlm_dirty_lockres(dlm, res);
2107 spin_unlock(&res->spinlock);
2109 dlm_lockres_put(res);
2113 /* this will become unnecessary eventually, but
2114 * for now we need to run the whole hash, clear
2115 * the RECOVERING state and set the owner
2117 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2118 bucket = dlm_lockres_hash(dlm, i);
2119 hlist_for_each_entry(res, hash_iter, bucket, hash_node) {
2120 if (!(res->state & DLM_LOCK_RES_RECOVERING))
2123 if (res->owner != dead_node &&
2124 res->owner != dlm->node_num)
2127 if (!list_empty(&res->recovering)) {
2128 list_del_init(&res->recovering);
2129 dlm_lockres_put(res);
2132 /* new_master has our reference from
2133 * the lock state sent during recovery */
2134 mlog(0, "%s: res %.*s, Changing owner from %u to %u\n",
2135 dlm->name, res->lockname.len, res->lockname.name,
2136 res->owner, new_master);
2137 spin_lock(&res->spinlock);
2138 dlm_change_lockres_owner(dlm, res, new_master);
2139 res->state &= ~DLM_LOCK_RES_RECOVERING;
2140 if (__dlm_lockres_has_locks(res))
2141 __dlm_dirty_lockres(dlm, res);
2142 spin_unlock(&res->spinlock);
2148 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
2151 if (lock->ml.type != LKM_EXMODE &&
2152 lock->ml.type != LKM_PRMODE)
2154 } else if (lock->ml.type == LKM_EXMODE)
2159 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
2160 struct dlm_lock_resource *res, u8 dead_node)
2162 struct list_head *queue;
2163 struct dlm_lock *lock;
2164 int blank_lvb = 0, local = 0;
2168 assert_spin_locked(&dlm->spinlock);
2169 assert_spin_locked(&res->spinlock);
2171 if (res->owner == dlm->node_num)
2172 /* if this node owned the lockres, and if the dead node
2173 * had an EX when he died, blank out the lvb */
2174 search_node = dead_node;
2176 /* if this is a secondary lockres, and we had no EX or PR
2177 * locks granted, we can no longer trust the lvb */
2178 search_node = dlm->node_num;
2179 local = 1; /* check local state for valid lvb */
2182 for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
2183 queue = dlm_list_idx_to_ptr(res, i);
2184 list_for_each_entry(lock, queue, list) {
2185 if (lock->ml.node == search_node) {
2186 if (dlm_lvb_needs_invalidation(lock, local)) {
2187 /* zero the lksb lvb and lockres lvb */
2189 memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
2196 mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
2197 res->lockname.len, res->lockname.name, dead_node);
2198 memset(res->lvb, 0, DLM_LVB_LEN);
2202 static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
2203 struct dlm_lock_resource *res, u8 dead_node)
2205 struct dlm_lock *lock, *next;
2206 unsigned int freed = 0;
2208 /* this node is the lockres master:
2209 * 1) remove any stale locks for the dead node
2210 * 2) if the dead node had an EX when he died, blank out the lvb
2212 assert_spin_locked(&dlm->spinlock);
2213 assert_spin_locked(&res->spinlock);
2215 /* We do two dlm_lock_put(). One for removing from list and the other is
2216 * to force the DLM_UNLOCK_FREE_LOCK action so as to free the locks */
2218 /* TODO: check pending_asts, pending_basts here */
2219 list_for_each_entry_safe(lock, next, &res->granted, list) {
2220 if (lock->ml.node == dead_node) {
2221 list_del_init(&lock->list);
2223 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2228 list_for_each_entry_safe(lock, next, &res->converting, 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->blocked, list) {
2238 if (lock->ml.node == dead_node) {
2239 list_del_init(&lock->list);
2241 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2248 mlog(0, "%s:%.*s: freed %u locks for dead node %u, "
2249 "dropping ref from lockres\n", dlm->name,
2250 res->lockname.len, res->lockname.name, freed, dead_node);
2251 if(!test_bit(dead_node, res->refmap)) {
2252 mlog(ML_ERROR, "%s:%.*s: freed %u locks for dead node %u, "
2253 "but ref was not set\n", dlm->name,
2254 res->lockname.len, res->lockname.name, freed, dead_node);
2255 __dlm_print_one_lock_resource(res);
2257 dlm_lockres_clear_refmap_bit(dead_node, res);
2258 } else if (test_bit(dead_node, res->refmap)) {
2259 mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2260 "no locks and had not purged before dying\n", dlm->name,
2261 res->lockname.len, res->lockname.name, dead_node);
2262 dlm_lockres_clear_refmap_bit(dead_node, res);
2265 /* do not kick thread yet */
2266 __dlm_dirty_lockres(dlm, res);
2269 /* if this node is the recovery master, and there are no
2270 * locks for a given lockres owned by this node that are in
2271 * either PR or EX mode, zero out the lvb before requesting.
2276 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
2278 struct hlist_node *iter;
2279 struct dlm_lock_resource *res;
2281 struct hlist_head *bucket;
2282 struct dlm_lock *lock;
2285 /* purge any stale mles */
2286 dlm_clean_master_list(dlm, dead_node);
2289 * now clean up all lock resources. there are two rules:
2291 * 1) if the dead node was the master, move the lockres
2292 * to the recovering list. set the RECOVERING flag.
2293 * this lockres needs to be cleaned up before it can
2296 * 2) if this node was the master, remove all locks from
2297 * each of the lockres queues that were owned by the
2298 * dead node. once recovery finishes, the dlm thread
2299 * can be kicked again to see if any ASTs or BASTs
2300 * need to be fired as a result.
2302 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2303 bucket = dlm_lockres_hash(dlm, i);
2304 hlist_for_each_entry(res, iter, bucket, hash_node) {
2305 /* always prune any $RECOVERY entries for dead nodes,
2306 * otherwise hangs can occur during later recovery */
2307 if (dlm_is_recovery_lock(res->lockname.name,
2308 res->lockname.len)) {
2309 spin_lock(&res->spinlock);
2310 list_for_each_entry(lock, &res->granted, list) {
2311 if (lock->ml.node == dead_node) {
2312 mlog(0, "AHA! there was "
2313 "a $RECOVERY lock for dead "
2315 dead_node, dlm->name);
2316 list_del_init(&lock->list);
2321 spin_unlock(&res->spinlock);
2324 spin_lock(&res->spinlock);
2325 /* zero the lvb if necessary */
2326 dlm_revalidate_lvb(dlm, res, dead_node);
2327 if (res->owner == dead_node) {
2328 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
2329 mlog(ML_NOTICE, "%s: res %.*s, Skip "
2330 "recovery as it is being freed\n",
2331 dlm->name, res->lockname.len,
2332 res->lockname.name);
2334 dlm_move_lockres_to_recovery_list(dlm,
2337 } else if (res->owner == dlm->node_num) {
2338 dlm_free_dead_locks(dlm, res, dead_node);
2339 __dlm_lockres_calc_usage(dlm, res);
2341 spin_unlock(&res->spinlock);
2347 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
2349 assert_spin_locked(&dlm->spinlock);
2351 if (dlm->reco.new_master == idx) {
2352 mlog(0, "%s: recovery master %d just died\n",
2354 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2355 /* finalize1 was reached, so it is safe to clear
2356 * the new_master and dead_node. that recovery
2358 mlog(0, "%s: dead master %d had reached "
2359 "finalize1 state, clearing\n", dlm->name, idx);
2360 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2361 __dlm_reset_recovery(dlm);
2365 /* Clean up join state on node death. */
2366 if (dlm->joining_node == idx) {
2367 mlog(0, "Clearing join state for node %u\n", idx);
2368 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
2371 /* check to see if the node is already considered dead */
2372 if (!test_bit(idx, dlm->live_nodes_map)) {
2373 mlog(0, "for domain %s, node %d is already dead. "
2374 "another node likely did recovery already.\n",
2379 /* check to see if we do not care about this node */
2380 if (!test_bit(idx, dlm->domain_map)) {
2381 /* This also catches the case that we get a node down
2382 * but haven't joined the domain yet. */
2383 mlog(0, "node %u already removed from domain!\n", idx);
2387 clear_bit(idx, dlm->live_nodes_map);
2389 /* make sure local cleanup occurs before the heartbeat events */
2390 if (!test_bit(idx, dlm->recovery_map))
2391 dlm_do_local_recovery_cleanup(dlm, idx);
2393 /* notify anything attached to the heartbeat events */
2394 dlm_hb_event_notify_attached(dlm, idx, 0);
2396 mlog(0, "node %u being removed from domain map!\n", idx);
2397 clear_bit(idx, dlm->domain_map);
2398 clear_bit(idx, dlm->exit_domain_map);
2399 /* wake up migration waiters if a node goes down.
2400 * perhaps later we can genericize this for other waiters. */
2401 wake_up(&dlm->migration_wq);
2403 if (test_bit(idx, dlm->recovery_map))
2404 mlog(0, "domain %s, node %u already added "
2405 "to recovery map!\n", dlm->name, idx);
2407 set_bit(idx, dlm->recovery_map);
2410 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
2412 struct dlm_ctxt *dlm = data;
2418 * This will notify any dlm users that a node in our domain
2419 * went away without notifying us first.
2421 if (test_bit(idx, dlm->domain_map))
2422 dlm_fire_domain_eviction_callbacks(dlm, idx);
2424 spin_lock(&dlm->spinlock);
2425 __dlm_hb_node_down(dlm, idx);
2426 spin_unlock(&dlm->spinlock);
2431 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
2433 struct dlm_ctxt *dlm = data;
2438 spin_lock(&dlm->spinlock);
2439 set_bit(idx, dlm->live_nodes_map);
2440 /* do NOT notify mle attached to the heartbeat events.
2441 * new nodes are not interesting in mastery until joined. */
2442 spin_unlock(&dlm->spinlock);
2447 static void dlm_reco_ast(void *astdata)
2449 struct dlm_ctxt *dlm = astdata;
2450 mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2451 dlm->node_num, dlm->name);
2453 static void dlm_reco_bast(void *astdata, int blocked_type)
2455 struct dlm_ctxt *dlm = astdata;
2456 mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2457 dlm->node_num, dlm->name);
2459 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
2461 mlog(0, "unlockast for recovery lock fired!\n");
2465 * dlm_pick_recovery_master will continually attempt to use
2466 * dlmlock() on the special "$RECOVERY" lockres with the
2467 * LKM_NOQUEUE flag to get an EX. every thread that enters
2468 * this function on each node racing to become the recovery
2469 * master will not stop attempting this until either:
2470 * a) this node gets the EX (and becomes the recovery master),
2471 * or b) dlm->reco.new_master gets set to some nodenum
2472 * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2473 * so each time a recovery master is needed, the entire cluster
2474 * will sync at this point. if the new master dies, that will
2475 * be detected in dlm_do_recovery */
2476 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2478 enum dlm_status ret;
2479 struct dlm_lockstatus lksb;
2480 int status = -EINVAL;
2482 mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2483 dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2485 memset(&lksb, 0, sizeof(lksb));
2487 ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
2488 DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN,
2489 dlm_reco_ast, dlm, dlm_reco_bast);
2491 mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2492 dlm->name, ret, lksb.status);
2494 if (ret == DLM_NORMAL) {
2495 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2496 dlm->name, dlm->node_num);
2498 /* got the EX lock. check to see if another node
2499 * just became the reco master */
2500 if (dlm_reco_master_ready(dlm)) {
2501 mlog(0, "%s: got reco EX lock, but %u will "
2502 "do the recovery\n", dlm->name,
2503 dlm->reco.new_master);
2508 /* see if recovery was already finished elsewhere */
2509 spin_lock(&dlm->spinlock);
2510 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
2512 mlog(0, "%s: got reco EX lock, but "
2513 "node got recovered already\n", dlm->name);
2514 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2515 mlog(ML_ERROR, "%s: new master is %u "
2516 "but no dead node!\n",
2517 dlm->name, dlm->reco.new_master);
2521 spin_unlock(&dlm->spinlock);
2524 /* if this node has actually become the recovery master,
2525 * set the master and send the messages to begin recovery */
2527 mlog(0, "%s: dead=%u, this=%u, sending "
2528 "begin_reco now\n", dlm->name,
2529 dlm->reco.dead_node, dlm->node_num);
2530 status = dlm_send_begin_reco_message(dlm,
2531 dlm->reco.dead_node);
2532 /* this always succeeds */
2535 /* set the new_master to this node */
2536 spin_lock(&dlm->spinlock);
2537 dlm_set_reco_master(dlm, dlm->node_num);
2538 spin_unlock(&dlm->spinlock);
2541 /* recovery lock is a special case. ast will not get fired,
2542 * so just go ahead and unlock it. */
2543 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
2544 if (ret == DLM_DENIED) {
2545 mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2546 ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2548 if (ret != DLM_NORMAL) {
2549 /* this would really suck. this could only happen
2550 * if there was a network error during the unlock
2551 * because of node death. this means the unlock
2552 * is actually "done" and the lock structure is
2553 * even freed. we can continue, but only
2554 * because this specific lock name is special. */
2555 mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
2557 } else if (ret == DLM_NOTQUEUED) {
2558 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2559 dlm->name, dlm->node_num);
2560 /* another node is master. wait on
2561 * reco.new_master != O2NM_INVALID_NODE_NUM
2562 * for at most one second */
2563 wait_event_timeout(dlm->dlm_reco_thread_wq,
2564 dlm_reco_master_ready(dlm),
2565 msecs_to_jiffies(1000));
2566 if (!dlm_reco_master_ready(dlm)) {
2567 mlog(0, "%s: reco master taking awhile\n",
2571 /* another node has informed this one that it is reco master */
2572 mlog(0, "%s: reco master %u is ready to recover %u\n",
2573 dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
2575 } else if (ret == DLM_RECOVERING) {
2576 mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n",
2577 dlm->name, dlm->node_num);
2580 struct dlm_lock_resource *res;
2582 /* dlmlock returned something other than NOTQUEUED or NORMAL */
2583 mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2584 "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2585 dlm_errname(lksb.status));
2586 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2587 DLM_RECOVERY_LOCK_NAME_LEN);
2589 dlm_print_one_lock_resource(res);
2590 dlm_lockres_put(res);
2592 mlog(ML_ERROR, "recovery lock not found\n");
2600 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2602 struct dlm_begin_reco br;
2604 struct dlm_node_iter iter;
2608 mlog(0, "%s: dead node is %u\n", dlm->name, dead_node);
2610 spin_lock(&dlm->spinlock);
2611 dlm_node_iter_init(dlm->domain_map, &iter);
2612 spin_unlock(&dlm->spinlock);
2614 clear_bit(dead_node, iter.node_map);
2616 memset(&br, 0, sizeof(br));
2617 br.node_idx = dlm->node_num;
2618 br.dead_node = dead_node;
2620 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2622 if (nodenum == dead_node) {
2623 mlog(0, "not sending begin reco to dead node "
2627 if (nodenum == dlm->node_num) {
2628 mlog(0, "not sending begin reco to self\n");
2633 mlog(0, "attempting to send begin reco msg to %d\n",
2635 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2636 &br, sizeof(br), nodenum, &status);
2637 /* negative status is handled ok by caller here */
2640 if (dlm_is_host_down(ret)) {
2641 /* node is down. not involved in recovery
2642 * so just keep going */
2643 mlog(ML_NOTICE, "%s: node %u was down when sending "
2644 "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2649 * Prior to commit aad1b15310b9bcd59fa81ab8f2b1513b59553ea8,
2650 * dlm_begin_reco_handler() returned EAGAIN and not -EAGAIN.
2651 * We are handling both for compatibility reasons.
2653 if (ret == -EAGAIN || ret == EAGAIN) {
2654 mlog(0, "%s: trying to start recovery of node "
2655 "%u, but node %u is waiting for last recovery "
2656 "to complete, backoff for a bit\n", dlm->name,
2657 dead_node, nodenum);
2662 struct dlm_lock_resource *res;
2664 /* this is now a serious problem, possibly ENOMEM
2665 * in the network stack. must retry */
2667 mlog(ML_ERROR, "begin reco of dlm %s to node %u "
2668 "returned %d\n", dlm->name, nodenum, ret);
2669 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2670 DLM_RECOVERY_LOCK_NAME_LEN);
2672 dlm_print_one_lock_resource(res);
2673 dlm_lockres_put(res);
2675 mlog(ML_ERROR, "recovery lock not found\n");
2677 /* sleep for a bit in hopes that we can avoid
2687 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2690 struct dlm_ctxt *dlm = data;
2691 struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2693 /* ok to return 0, domain has gone away */
2697 spin_lock(&dlm->spinlock);
2698 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2699 mlog(0, "%s: node %u wants to recover node %u (%u:%u) "
2700 "but this node is in finalize state, waiting on finalize2\n",
2701 dlm->name, br->node_idx, br->dead_node,
2702 dlm->reco.dead_node, dlm->reco.new_master);
2703 spin_unlock(&dlm->spinlock);
2706 spin_unlock(&dlm->spinlock);
2708 mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n",
2709 dlm->name, br->node_idx, br->dead_node,
2710 dlm->reco.dead_node, dlm->reco.new_master);
2712 dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2714 spin_lock(&dlm->spinlock);
2715 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2716 if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2717 mlog(0, "%s: new_master %u died, changing "
2718 "to %u\n", dlm->name, dlm->reco.new_master,
2721 mlog(0, "%s: new_master %u NOT DEAD, changing "
2722 "to %u\n", dlm->name, dlm->reco.new_master,
2724 /* may not have seen the new master as dead yet */
2727 if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2728 mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2729 "node %u changing it to %u\n", dlm->name,
2730 dlm->reco.dead_node, br->node_idx, br->dead_node);
2732 dlm_set_reco_master(dlm, br->node_idx);
2733 dlm_set_reco_dead_node(dlm, br->dead_node);
2734 if (!test_bit(br->dead_node, dlm->recovery_map)) {
2735 mlog(0, "recovery master %u sees %u as dead, but this "
2736 "node has not yet. marking %u as dead\n",
2737 br->node_idx, br->dead_node, br->dead_node);
2738 if (!test_bit(br->dead_node, dlm->domain_map) ||
2739 !test_bit(br->dead_node, dlm->live_nodes_map))
2740 mlog(0, "%u not in domain/live_nodes map "
2741 "so setting it in reco map manually\n",
2743 /* force the recovery cleanup in __dlm_hb_node_down
2744 * both of these will be cleared in a moment */
2745 set_bit(br->dead_node, dlm->domain_map);
2746 set_bit(br->dead_node, dlm->live_nodes_map);
2747 __dlm_hb_node_down(dlm, br->dead_node);
2749 spin_unlock(&dlm->spinlock);
2751 dlm_kick_recovery_thread(dlm);
2753 mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n",
2754 dlm->name, br->node_idx, br->dead_node,
2755 dlm->reco.dead_node, dlm->reco.new_master);
2761 #define DLM_FINALIZE_STAGE2 0x01
2762 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2765 struct dlm_finalize_reco fr;
2766 struct dlm_node_iter iter;
2771 mlog(0, "finishing recovery for node %s:%u, "
2772 "stage %d\n", dlm->name, dlm->reco.dead_node, stage);
2774 spin_lock(&dlm->spinlock);
2775 dlm_node_iter_init(dlm->domain_map, &iter);
2776 spin_unlock(&dlm->spinlock);
2779 memset(&fr, 0, sizeof(fr));
2780 fr.node_idx = dlm->node_num;
2781 fr.dead_node = dlm->reco.dead_node;
2783 fr.flags |= DLM_FINALIZE_STAGE2;
2785 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2786 if (nodenum == dlm->node_num)
2788 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2789 &fr, sizeof(fr), nodenum, &status);
2793 mlog(ML_ERROR, "Error %d when sending message %u (key "
2794 "0x%x) to node %u\n", ret, DLM_FINALIZE_RECO_MSG,
2796 if (dlm_is_host_down(ret)) {
2797 /* this has no effect on this recovery
2798 * session, so set the status to zero to
2799 * finish out the last recovery */
2800 mlog(ML_ERROR, "node %u went down after this "
2801 "node finished recovery.\n", nodenum);
2809 /* reset the node_iter back to the top and send finalize2 */
2818 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2821 struct dlm_ctxt *dlm = data;
2822 struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2825 /* ok to return 0, domain has gone away */
2829 if (fr->flags & DLM_FINALIZE_STAGE2)
2832 mlog(0, "%s: node %u finalizing recovery stage%d of "
2833 "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage,
2834 fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master);
2836 spin_lock(&dlm->spinlock);
2838 if (dlm->reco.new_master != fr->node_idx) {
2839 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2840 "%u is supposed to be the new master, dead=%u\n",
2841 fr->node_idx, dlm->reco.new_master, fr->dead_node);
2844 if (dlm->reco.dead_node != fr->dead_node) {
2845 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2846 "node %u, but node %u is supposed to be dead\n",
2847 fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2853 dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2854 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2855 mlog(ML_ERROR, "%s: received finalize1 from "
2856 "new master %u for dead node %u, but "
2857 "this node has already received it!\n",
2858 dlm->name, fr->node_idx, fr->dead_node);
2859 dlm_print_reco_node_status(dlm);
2862 dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
2863 spin_unlock(&dlm->spinlock);
2866 if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) {
2867 mlog(ML_ERROR, "%s: received finalize2 from "
2868 "new master %u for dead node %u, but "
2869 "this node did not have finalize1!\n",
2870 dlm->name, fr->node_idx, fr->dead_node);
2871 dlm_print_reco_node_status(dlm);
2874 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2875 spin_unlock(&dlm->spinlock);
2876 dlm_reset_recovery(dlm);
2877 dlm_kick_recovery_thread(dlm);
2883 mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n",
2884 dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master);