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/utsname.h>
34 #include <linux/init.h>
35 #include <linux/sysctl.h>
36 #include <linux/random.h>
37 #include <linux/blkdev.h>
38 #include <linux/socket.h>
39 #include <linux/inet.h>
40 #include <linux/timer.h>
41 #include <linux/kthread.h>
42 #include <linux/delay.h>
45 #include "cluster/heartbeat.h"
46 #include "cluster/nodemanager.h"
47 #include "cluster/tcp.h"
50 #include "dlmcommon.h"
51 #include "dlmdomain.h"
53 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY)
54 #include "cluster/masklog.h"
56 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt
*dlm
, u8 dead_node
);
58 static int dlm_recovery_thread(void *data
);
59 void dlm_complete_recovery_thread(struct dlm_ctxt
*dlm
);
60 int dlm_launch_recovery_thread(struct dlm_ctxt
*dlm
);
61 void dlm_kick_recovery_thread(struct dlm_ctxt
*dlm
);
62 static int dlm_do_recovery(struct dlm_ctxt
*dlm
);
64 static int dlm_pick_recovery_master(struct dlm_ctxt
*dlm
);
65 static int dlm_remaster_locks(struct dlm_ctxt
*dlm
, u8 dead_node
);
66 static int dlm_init_recovery_area(struct dlm_ctxt
*dlm
, u8 dead_node
);
67 static int dlm_request_all_locks(struct dlm_ctxt
*dlm
,
68 u8 request_from
, u8 dead_node
);
69 static void dlm_destroy_recovery_area(struct dlm_ctxt
*dlm
, u8 dead_node
);
71 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource
*res
);
72 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres
*mres
,
73 const char *lockname
, int namelen
,
74 int total_locks
, u64 cookie
,
76 static int dlm_send_mig_lockres_msg(struct dlm_ctxt
*dlm
,
77 struct dlm_migratable_lockres
*mres
,
79 struct dlm_lock_resource
*res
,
81 static int dlm_process_recovery_data(struct dlm_ctxt
*dlm
,
82 struct dlm_lock_resource
*res
,
83 struct dlm_migratable_lockres
*mres
);
84 static int dlm_send_finalize_reco_message(struct dlm_ctxt
*dlm
);
85 static int dlm_send_all_done_msg(struct dlm_ctxt
*dlm
,
86 u8 dead_node
, u8 send_to
);
87 static int dlm_send_begin_reco_message(struct dlm_ctxt
*dlm
, u8 dead_node
);
88 static void dlm_move_reco_locks_to_list(struct dlm_ctxt
*dlm
,
89 struct list_head
*list
, u8 dead_node
);
90 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt
*dlm
,
91 u8 dead_node
, u8 new_master
);
92 static void dlm_reco_ast(void *astdata
);
93 static void dlm_reco_bast(void *astdata
, int blocked_type
);
94 static void dlm_reco_unlock_ast(void *astdata
, enum dlm_status st
);
95 static void dlm_request_all_locks_worker(struct dlm_work_item
*item
,
97 static void dlm_mig_lockres_worker(struct dlm_work_item
*item
, void *data
);
99 static u64
dlm_get_next_mig_cookie(void);
101 static spinlock_t dlm_reco_state_lock
= SPIN_LOCK_UNLOCKED
;
102 static spinlock_t dlm_mig_cookie_lock
= SPIN_LOCK_UNLOCKED
;
103 static u64 dlm_mig_cookie
= 1;
105 static u64
dlm_get_next_mig_cookie(void)
108 spin_lock(&dlm_mig_cookie_lock
);
110 if (dlm_mig_cookie
== (~0ULL))
114 spin_unlock(&dlm_mig_cookie_lock
);
118 static inline void dlm_reset_recovery(struct dlm_ctxt
*dlm
)
120 spin_lock(&dlm
->spinlock
);
121 clear_bit(dlm
->reco
.dead_node
, dlm
->recovery_map
);
122 dlm
->reco
.dead_node
= O2NM_INVALID_NODE_NUM
;
123 dlm
->reco
.new_master
= O2NM_INVALID_NODE_NUM
;
124 spin_unlock(&dlm
->spinlock
);
127 /* Worker function used during recovery. */
128 void dlm_dispatch_work(void *data
)
130 struct dlm_ctxt
*dlm
= (struct dlm_ctxt
*)data
;
132 struct list_head
*iter
, *iter2
;
133 struct dlm_work_item
*item
;
134 dlm_workfunc_t
*workfunc
;
136 spin_lock(&dlm
->work_lock
);
137 list_splice_init(&dlm
->work_list
, &tmp_list
);
138 spin_unlock(&dlm
->work_lock
);
140 list_for_each_safe(iter
, iter2
, &tmp_list
) {
141 item
= list_entry(iter
, struct dlm_work_item
, list
);
142 workfunc
= item
->func
;
143 list_del_init(&item
->list
);
145 /* already have ref on dlm to avoid having
146 * it disappear. just double-check. */
147 BUG_ON(item
->dlm
!= dlm
);
149 /* this is allowed to sleep and
150 * call network stuff */
151 workfunc(item
, item
->data
);
162 void dlm_kick_recovery_thread(struct dlm_ctxt
*dlm
)
164 /* wake the recovery thread
165 * this will wake the reco thread in one of three places
166 * 1) sleeping with no recovery happening
167 * 2) sleeping with recovery mastered elsewhere
168 * 3) recovery mastered here, waiting on reco data */
170 wake_up(&dlm
->dlm_reco_thread_wq
);
173 /* Launch the recovery thread */
174 int dlm_launch_recovery_thread(struct dlm_ctxt
*dlm
)
176 mlog(0, "starting dlm recovery thread...\n");
178 dlm
->dlm_reco_thread_task
= kthread_run(dlm_recovery_thread
, dlm
,
180 if (IS_ERR(dlm
->dlm_reco_thread_task
)) {
181 mlog_errno(PTR_ERR(dlm
->dlm_reco_thread_task
));
182 dlm
->dlm_reco_thread_task
= NULL
;
189 void dlm_complete_recovery_thread(struct dlm_ctxt
*dlm
)
191 if (dlm
->dlm_reco_thread_task
) {
192 mlog(0, "waiting for dlm recovery thread to exit\n");
193 kthread_stop(dlm
->dlm_reco_thread_task
);
194 dlm
->dlm_reco_thread_task
= NULL
;
201 * this is lame, but here's how recovery works...
202 * 1) all recovery threads cluster wide will work on recovering
204 * 2) negotiate who will take over all the locks for the dead node.
205 * thats right... ALL the locks.
206 * 3) once a new master is chosen, everyone scans all locks
207 * and moves aside those mastered by the dead guy
208 * 4) each of these locks should be locked until recovery is done
209 * 5) the new master collects up all of secondary lock queue info
210 * one lock at a time, forcing each node to communicate back
212 * 6) each secondary lock queue responds with the full known lock info
213 * 7) once the new master has run all its locks, it sends a ALLDONE!
214 * message to everyone
215 * 8) upon receiving this message, the secondary queue node unlocks
216 * and responds to the ALLDONE
217 * 9) once the new master gets responses from everyone, he unlocks
218 * everything and recovery for this dead node is done
219 *10) go back to 2) while there are still dead nodes
224 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
226 static int dlm_recovery_thread(void *data
)
229 struct dlm_ctxt
*dlm
= data
;
230 unsigned long timeout
= msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS
);
232 mlog(0, "dlm thread running for %s...\n", dlm
->name
);
234 while (!kthread_should_stop()) {
235 if (dlm_joined(dlm
)) {
236 status
= dlm_do_recovery(dlm
);
237 if (status
== -EAGAIN
) {
238 /* do not sleep, recheck immediately. */
245 wait_event_interruptible_timeout(dlm
->dlm_reco_thread_wq
,
246 kthread_should_stop(),
250 mlog(0, "quitting DLM recovery thread\n");
254 /* returns true when the recovery master has contacted us */
255 static int dlm_reco_master_ready(struct dlm_ctxt
*dlm
)
258 spin_lock(&dlm
->spinlock
);
259 ready
= (dlm
->reco
.new_master
!= O2NM_INVALID_NODE_NUM
);
260 spin_unlock(&dlm
->spinlock
);
264 /* returns true if node is no longer in the domain
265 * could be dead or just not joined */
266 int dlm_is_node_dead(struct dlm_ctxt
*dlm
, u8 node
)
269 spin_lock(&dlm
->spinlock
);
270 dead
= test_bit(node
, dlm
->domain_map
);
271 spin_unlock(&dlm
->spinlock
);
275 int dlm_wait_for_node_death(struct dlm_ctxt
*dlm
, u8 node
, int timeout
)
278 mlog(ML_NOTICE
, "%s: waiting %dms for notification of "
279 "death of node %u\n", dlm
->name
, timeout
, node
);
280 wait_event_timeout(dlm
->dlm_reco_thread_wq
,
281 dlm_is_node_dead(dlm
, node
),
282 msecs_to_jiffies(timeout
));
284 mlog(ML_NOTICE
, "%s: waiting indefinitely for notification "
285 "of death of node %u\n", dlm
->name
, node
);
286 wait_event(dlm
->dlm_reco_thread_wq
,
287 dlm_is_node_dead(dlm
, node
));
289 /* for now, return 0 */
293 /* callers of the top-level api calls (dlmlock/dlmunlock) should
294 * block on the dlm->reco.event when recovery is in progress.
295 * the dlm recovery thread will set this state when it begins
296 * recovering a dead node (as the new master or not) and clear
297 * the state and wake as soon as all affected lock resources have
298 * been marked with the RECOVERY flag */
299 static int dlm_in_recovery(struct dlm_ctxt
*dlm
)
302 spin_lock(&dlm
->spinlock
);
303 in_recovery
= !!(dlm
->reco
.state
& DLM_RECO_STATE_ACTIVE
);
304 spin_unlock(&dlm
->spinlock
);
309 void dlm_wait_for_recovery(struct dlm_ctxt
*dlm
)
311 wait_event(dlm
->reco
.event
, !dlm_in_recovery(dlm
));
314 static void dlm_begin_recovery(struct dlm_ctxt
*dlm
)
316 spin_lock(&dlm
->spinlock
);
317 BUG_ON(dlm
->reco
.state
& DLM_RECO_STATE_ACTIVE
);
318 dlm
->reco
.state
|= DLM_RECO_STATE_ACTIVE
;
319 spin_unlock(&dlm
->spinlock
);
322 static void dlm_end_recovery(struct dlm_ctxt
*dlm
)
324 spin_lock(&dlm
->spinlock
);
325 BUG_ON(!(dlm
->reco
.state
& DLM_RECO_STATE_ACTIVE
));
326 dlm
->reco
.state
&= ~DLM_RECO_STATE_ACTIVE
;
327 spin_unlock(&dlm
->spinlock
);
328 wake_up(&dlm
->reco
.event
);
331 static int dlm_do_recovery(struct dlm_ctxt
*dlm
)
336 spin_lock(&dlm
->spinlock
);
338 /* check to see if the new master has died */
339 if (dlm
->reco
.new_master
!= O2NM_INVALID_NODE_NUM
&&
340 test_bit(dlm
->reco
.new_master
, dlm
->recovery_map
)) {
341 mlog(0, "new master %u died while recovering %u!\n",
342 dlm
->reco
.new_master
, dlm
->reco
.dead_node
);
343 /* unset the new_master, leave dead_node */
344 dlm
->reco
.new_master
= O2NM_INVALID_NODE_NUM
;
347 /* select a target to recover */
348 if (dlm
->reco
.dead_node
== O2NM_INVALID_NODE_NUM
) {
351 bit
= find_next_bit (dlm
->recovery_map
, O2NM_MAX_NODES
+1, 0);
352 if (bit
>= O2NM_MAX_NODES
|| bit
< 0)
353 dlm
->reco
.dead_node
= O2NM_INVALID_NODE_NUM
;
355 dlm
->reco
.dead_node
= bit
;
356 } else if (!test_bit(dlm
->reco
.dead_node
, dlm
->recovery_map
)) {
358 mlog(ML_ERROR
, "dead_node %u no longer in recovery map!\n",
359 dlm
->reco
.dead_node
);
360 dlm
->reco
.dead_node
= O2NM_INVALID_NODE_NUM
;
363 if (dlm
->reco
.dead_node
== O2NM_INVALID_NODE_NUM
) {
364 // mlog(0, "nothing to recover! sleeping now!\n");
365 spin_unlock(&dlm
->spinlock
);
366 /* return to main thread loop and sleep. */
369 mlog(0, "recovery thread found node %u in the recovery map!\n",
370 dlm
->reco
.dead_node
);
371 spin_unlock(&dlm
->spinlock
);
373 /* take write barrier */
374 /* (stops the list reshuffling thread, proxy ast handling) */
375 dlm_begin_recovery(dlm
);
377 if (dlm
->reco
.new_master
== dlm
->node_num
)
380 if (dlm
->reco
.new_master
== O2NM_INVALID_NODE_NUM
) {
381 /* choose a new master, returns 0 if this node
382 * is the master, -EEXIST if it's another node.
383 * this does not return until a new master is chosen
384 * or recovery completes entirely. */
385 ret
= dlm_pick_recovery_master(dlm
);
387 /* already notified everyone. go. */
390 mlog(0, "another node will master this recovery session.\n");
392 mlog(0, "dlm=%s, new_master=%u, this node=%u, dead_node=%u\n",
393 dlm
->name
, dlm
->reco
.new_master
,
394 dlm
->node_num
, dlm
->reco
.dead_node
);
396 /* it is safe to start everything back up here
397 * because all of the dead node's lock resources
398 * have been marked as in-recovery */
399 dlm_end_recovery(dlm
);
401 /* sleep out in main dlm_recovery_thread loop. */
405 mlog(0, "mastering recovery of %s:%u here(this=%u)!\n",
406 dlm
->name
, dlm
->reco
.dead_node
, dlm
->node_num
);
408 status
= dlm_remaster_locks(dlm
, dlm
->reco
.dead_node
);
410 mlog(ML_ERROR
, "error %d remastering locks for node %u, "
411 "retrying.\n", status
, dlm
->reco
.dead_node
);
412 /* yield a bit to allow any final network messages
413 * to get handled on remaining nodes */
416 /* success! see if any other nodes need recovery */
417 mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
418 dlm
->name
, dlm
->reco
.dead_node
, dlm
->node_num
);
419 dlm_reset_recovery(dlm
);
421 dlm_end_recovery(dlm
);
423 /* continue and look for another dead node */
427 static int dlm_remaster_locks(struct dlm_ctxt
*dlm
, u8 dead_node
)
430 struct dlm_reco_node_data
*ndata
;
431 struct list_head
*iter
;
436 status
= dlm_init_recovery_area(dlm
, dead_node
);
440 /* safe to access the node data list without a lock, since this
441 * process is the only one to change the list */
442 list_for_each(iter
, &dlm
->reco
.node_data
) {
443 ndata
= list_entry (iter
, struct dlm_reco_node_data
, list
);
444 BUG_ON(ndata
->state
!= DLM_RECO_NODE_DATA_INIT
);
445 ndata
->state
= DLM_RECO_NODE_DATA_REQUESTING
;
447 mlog(0, "requesting lock info from node %u\n",
450 if (ndata
->node_num
== dlm
->node_num
) {
451 ndata
->state
= DLM_RECO_NODE_DATA_DONE
;
455 status
= dlm_request_all_locks(dlm
, ndata
->node_num
, dead_node
);
458 if (dlm_is_host_down(status
))
459 ndata
->state
= DLM_RECO_NODE_DATA_DEAD
;
466 switch (ndata
->state
) {
467 case DLM_RECO_NODE_DATA_INIT
:
468 case DLM_RECO_NODE_DATA_FINALIZE_SENT
:
469 case DLM_RECO_NODE_DATA_REQUESTED
:
472 case DLM_RECO_NODE_DATA_DEAD
:
473 mlog(0, "node %u died after requesting "
474 "recovery info for node %u\n",
475 ndata
->node_num
, dead_node
);
480 case DLM_RECO_NODE_DATA_REQUESTING
:
481 ndata
->state
= DLM_RECO_NODE_DATA_REQUESTED
;
482 mlog(0, "now receiving recovery data from "
483 "node %u for dead node %u\n",
484 ndata
->node_num
, dead_node
);
486 case DLM_RECO_NODE_DATA_RECEIVING
:
487 mlog(0, "already receiving recovery data from "
488 "node %u for dead node %u\n",
489 ndata
->node_num
, dead_node
);
491 case DLM_RECO_NODE_DATA_DONE
:
492 mlog(0, "already DONE receiving recovery data "
493 "from node %u for dead node %u\n",
494 ndata
->node_num
, dead_node
);
499 mlog(0, "done requesting all lock info\n");
501 /* nodes should be sending reco data now
502 * just need to wait */
505 /* check all the nodes now to see if we are
506 * done, or if anyone died */
508 spin_lock(&dlm_reco_state_lock
);
509 list_for_each(iter
, &dlm
->reco
.node_data
) {
510 ndata
= list_entry (iter
, struct dlm_reco_node_data
, list
);
512 mlog(0, "checking recovery state of node %u\n",
514 switch (ndata
->state
) {
515 case DLM_RECO_NODE_DATA_INIT
:
516 case DLM_RECO_NODE_DATA_REQUESTING
:
517 mlog(ML_ERROR
, "bad ndata state for "
518 "node %u: state=%d\n",
519 ndata
->node_num
, ndata
->state
);
522 case DLM_RECO_NODE_DATA_DEAD
:
523 mlog(ML_NOTICE
, "node %u died after "
524 "requesting recovery info for "
525 "node %u\n", ndata
->node_num
,
527 spin_unlock(&dlm_reco_state_lock
);
531 /* instead of spinning like crazy here,
532 * wait for the domain map to catch up
533 * with the network state. otherwise this
534 * can be hit hundreds of times before
535 * the node is really seen as dead. */
536 wait_event_timeout(dlm
->dlm_reco_thread_wq
,
537 dlm_is_node_dead(dlm
,
539 msecs_to_jiffies(1000));
540 mlog(0, "waited 1 sec for %u, "
541 "dead? %s\n", ndata
->node_num
,
542 dlm_is_node_dead(dlm
, ndata
->node_num
) ?
545 case DLM_RECO_NODE_DATA_RECEIVING
:
546 case DLM_RECO_NODE_DATA_REQUESTED
:
549 case DLM_RECO_NODE_DATA_DONE
:
551 case DLM_RECO_NODE_DATA_FINALIZE_SENT
:
555 spin_unlock(&dlm_reco_state_lock
);
557 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass
,
558 all_nodes_done
?"yes":"no");
559 if (all_nodes_done
) {
562 /* all nodes are now in DLM_RECO_NODE_DATA_DONE state
563 * just send a finalize message to everyone and
565 mlog(0, "all nodes are done! send finalize\n");
566 ret
= dlm_send_finalize_reco_message(dlm
);
570 spin_lock(&dlm
->spinlock
);
571 dlm_finish_local_lockres_recovery(dlm
, dead_node
,
573 spin_unlock(&dlm
->spinlock
);
574 mlog(0, "should be done with recovery!\n");
576 mlog(0, "finishing recovery of %s at %lu, "
577 "dead=%u, this=%u, new=%u\n", dlm
->name
,
578 jiffies
, dlm
->reco
.dead_node
,
579 dlm
->node_num
, dlm
->reco
.new_master
);
582 /* rescan everything marked dirty along the way */
583 dlm_kick_thread(dlm
, NULL
);
586 /* wait to be signalled, with periodic timeout
587 * to check for node death */
588 wait_event_interruptible_timeout(dlm
->dlm_reco_thread_wq
,
589 kthread_should_stop(),
590 msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS
));
596 dlm_destroy_recovery_area(dlm
, dead_node
);
602 static int dlm_init_recovery_area(struct dlm_ctxt
*dlm
, u8 dead_node
)
605 struct dlm_reco_node_data
*ndata
;
607 spin_lock(&dlm
->spinlock
);
608 memcpy(dlm
->reco
.node_map
, dlm
->domain_map
, sizeof(dlm
->domain_map
));
609 /* nodes can only be removed (by dying) after dropping
610 * this lock, and death will be trapped later, so this should do */
611 spin_unlock(&dlm
->spinlock
);
614 num
= find_next_bit (dlm
->reco
.node_map
, O2NM_MAX_NODES
, num
);
615 if (num
>= O2NM_MAX_NODES
) {
618 BUG_ON(num
== dead_node
);
620 ndata
= kcalloc(1, sizeof(*ndata
), GFP_KERNEL
);
622 dlm_destroy_recovery_area(dlm
, dead_node
);
625 ndata
->node_num
= num
;
626 ndata
->state
= DLM_RECO_NODE_DATA_INIT
;
627 spin_lock(&dlm_reco_state_lock
);
628 list_add_tail(&ndata
->list
, &dlm
->reco
.node_data
);
629 spin_unlock(&dlm_reco_state_lock
);
636 static void dlm_destroy_recovery_area(struct dlm_ctxt
*dlm
, u8 dead_node
)
638 struct list_head
*iter
, *iter2
;
639 struct dlm_reco_node_data
*ndata
;
642 spin_lock(&dlm_reco_state_lock
);
643 list_splice_init(&dlm
->reco
.node_data
, &tmplist
);
644 spin_unlock(&dlm_reco_state_lock
);
646 list_for_each_safe(iter
, iter2
, &tmplist
) {
647 ndata
= list_entry (iter
, struct dlm_reco_node_data
, list
);
648 list_del_init(&ndata
->list
);
653 static int dlm_request_all_locks(struct dlm_ctxt
*dlm
, u8 request_from
,
656 struct dlm_lock_request lr
;
662 mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
663 "to %u\n", dead_node
, request_from
);
665 memset(&lr
, 0, sizeof(lr
));
666 lr
.node_idx
= dlm
->node_num
;
667 lr
.dead_node
= dead_node
;
671 ret
= o2net_send_message(DLM_LOCK_REQUEST_MSG
, dlm
->key
,
672 &lr
, sizeof(lr
), request_from
, NULL
);
674 /* negative status is handled by caller */
678 // return from here, then
679 // sleep until all received or error
684 int dlm_request_all_locks_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
686 struct dlm_ctxt
*dlm
= data
;
687 struct dlm_lock_request
*lr
= (struct dlm_lock_request
*)msg
->buf
;
689 struct dlm_work_item
*item
= NULL
;
694 BUG_ON(lr
->dead_node
!= dlm
->reco
.dead_node
);
696 item
= kcalloc(1, sizeof(*item
), GFP_KERNEL
);
702 /* this will get freed by dlm_request_all_locks_worker */
703 buf
= (char *) __get_free_page(GFP_KERNEL
);
710 /* queue up work for dlm_request_all_locks_worker */
711 dlm_grab(dlm
); /* get an extra ref for the work item */
712 dlm_init_work_item(dlm
, item
, dlm_request_all_locks_worker
, buf
);
713 item
->u
.ral
.reco_master
= lr
->node_idx
;
714 item
->u
.ral
.dead_node
= lr
->dead_node
;
715 spin_lock(&dlm
->work_lock
);
716 list_add_tail(&item
->list
, &dlm
->work_list
);
717 spin_unlock(&dlm
->work_lock
);
718 schedule_work(&dlm
->dispatched_work
);
724 static void dlm_request_all_locks_worker(struct dlm_work_item
*item
, void *data
)
726 struct dlm_migratable_lockres
*mres
;
727 struct dlm_lock_resource
*res
;
728 struct dlm_ctxt
*dlm
;
729 LIST_HEAD(resources
);
730 struct list_head
*iter
;
732 u8 dead_node
, reco_master
;
735 dead_node
= item
->u
.ral
.dead_node
;
736 reco_master
= item
->u
.ral
.reco_master
;
737 mres
= (struct dlm_migratable_lockres
*)data
;
739 if (dead_node
!= dlm
->reco
.dead_node
||
740 reco_master
!= dlm
->reco
.new_master
) {
741 /* show extra debug info if the recovery state is messed */
742 mlog(ML_ERROR
, "%s: bad reco state: reco(dead=%u, master=%u), "
743 "request(dead=%u, master=%u)\n",
744 dlm
->name
, dlm
->reco
.dead_node
, dlm
->reco
.new_master
,
745 dead_node
, reco_master
);
746 mlog(ML_ERROR
, "%s: name=%.*s master=%u locks=%u/%u flags=%u "
747 "entry[0]={c=%u:%llu,l=%u,f=%u,t=%d,ct=%d,hb=%d,n=%u}\n",
748 dlm
->name
, mres
->lockname_len
, mres
->lockname
, mres
->master
,
749 mres
->num_locks
, mres
->total_locks
, mres
->flags
,
750 dlm_get_lock_cookie_node(mres
->ml
[0].cookie
),
751 dlm_get_lock_cookie_seq(mres
->ml
[0].cookie
),
752 mres
->ml
[0].list
, mres
->ml
[0].flags
,
753 mres
->ml
[0].type
, mres
->ml
[0].convert_type
,
754 mres
->ml
[0].highest_blocked
, mres
->ml
[0].node
);
757 BUG_ON(dead_node
!= dlm
->reco
.dead_node
);
758 BUG_ON(reco_master
!= dlm
->reco
.new_master
);
760 /* lock resources should have already been moved to the
761 * dlm->reco.resources list. now move items from that list
762 * to a temp list if the dead owner matches. note that the
763 * whole cluster recovers only one node at a time, so we
764 * can safely move UNKNOWN lock resources for each recovery
766 dlm_move_reco_locks_to_list(dlm
, &resources
, dead_node
);
768 /* now we can begin blasting lockreses without the dlm lock */
769 list_for_each(iter
, &resources
) {
770 res
= list_entry (iter
, struct dlm_lock_resource
, recovering
);
771 ret
= dlm_send_one_lockres(dlm
, res
, mres
, reco_master
,
777 /* move the resources back to the list */
778 spin_lock(&dlm
->spinlock
);
779 list_splice_init(&resources
, &dlm
->reco
.resources
);
780 spin_unlock(&dlm
->spinlock
);
782 ret
= dlm_send_all_done_msg(dlm
, dead_node
, reco_master
);
786 free_page((unsigned long)data
);
790 static int dlm_send_all_done_msg(struct dlm_ctxt
*dlm
, u8 dead_node
, u8 send_to
)
793 struct dlm_reco_data_done done_msg
;
795 memset(&done_msg
, 0, sizeof(done_msg
));
796 done_msg
.node_idx
= dlm
->node_num
;
797 done_msg
.dead_node
= dead_node
;
798 mlog(0, "sending DATA DONE message to %u, "
799 "my node=%u, dead node=%u\n", send_to
, done_msg
.node_idx
,
802 ret
= o2net_send_message(DLM_RECO_DATA_DONE_MSG
, dlm
->key
, &done_msg
,
803 sizeof(done_msg
), send_to
, &tmpret
);
804 /* negative status is ignored by the caller */
811 int dlm_reco_data_done_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
813 struct dlm_ctxt
*dlm
= data
;
814 struct dlm_reco_data_done
*done
= (struct dlm_reco_data_done
*)msg
->buf
;
815 struct list_head
*iter
;
816 struct dlm_reco_node_data
*ndata
= NULL
;
822 mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
823 "node_idx=%u, this node=%u\n", done
->dead_node
,
824 dlm
->reco
.dead_node
, done
->node_idx
, dlm
->node_num
);
825 BUG_ON(done
->dead_node
!= dlm
->reco
.dead_node
);
827 spin_lock(&dlm_reco_state_lock
);
828 list_for_each(iter
, &dlm
->reco
.node_data
) {
829 ndata
= list_entry (iter
, struct dlm_reco_node_data
, list
);
830 if (ndata
->node_num
!= done
->node_idx
)
833 switch (ndata
->state
) {
834 /* should have moved beyond INIT but not to FINALIZE yet */
835 case DLM_RECO_NODE_DATA_INIT
:
836 case DLM_RECO_NODE_DATA_DEAD
:
837 case DLM_RECO_NODE_DATA_FINALIZE_SENT
:
838 mlog(ML_ERROR
, "bad ndata state for node %u:"
839 " state=%d\n", ndata
->node_num
,
843 /* these states are possible at this point, anywhere along
844 * the line of recovery */
845 case DLM_RECO_NODE_DATA_DONE
:
846 case DLM_RECO_NODE_DATA_RECEIVING
:
847 case DLM_RECO_NODE_DATA_REQUESTED
:
848 case DLM_RECO_NODE_DATA_REQUESTING
:
849 mlog(0, "node %u is DONE sending "
853 ndata
->state
= DLM_RECO_NODE_DATA_DONE
;
858 spin_unlock(&dlm_reco_state_lock
);
860 /* wake the recovery thread, some node is done */
862 dlm_kick_recovery_thread(dlm
);
865 mlog(ML_ERROR
, "failed to find recovery node data for node "
866 "%u\n", done
->node_idx
);
869 mlog(0, "leaving reco data done handler, ret=%d\n", ret
);
873 static void dlm_move_reco_locks_to_list(struct dlm_ctxt
*dlm
,
874 struct list_head
*list
,
877 struct dlm_lock_resource
*res
;
878 struct list_head
*iter
, *iter2
;
879 struct dlm_lock
*lock
;
881 spin_lock(&dlm
->spinlock
);
882 list_for_each_safe(iter
, iter2
, &dlm
->reco
.resources
) {
883 res
= list_entry (iter
, struct dlm_lock_resource
, recovering
);
884 /* always prune any $RECOVERY entries for dead nodes,
885 * otherwise hangs can occur during later recovery */
886 if (dlm_is_recovery_lock(res
->lockname
.name
,
887 res
->lockname
.len
)) {
888 spin_lock(&res
->spinlock
);
889 list_for_each_entry(lock
, &res
->granted
, list
) {
890 if (lock
->ml
.node
== dead_node
) {
891 mlog(0, "AHA! there was "
892 "a $RECOVERY lock for dead "
894 dead_node
, dlm
->name
);
895 list_del_init(&lock
->list
);
900 spin_unlock(&res
->spinlock
);
904 if (res
->owner
== dead_node
) {
905 mlog(0, "found lockres owned by dead node while "
906 "doing recovery for node %u. sending it.\n",
908 list_move_tail(&res
->recovering
, list
);
909 } else if (res
->owner
== DLM_LOCK_RES_OWNER_UNKNOWN
) {
910 mlog(0, "found UNKNOWN owner while doing recovery "
911 "for node %u. sending it.\n", dead_node
);
912 list_move_tail(&res
->recovering
, list
);
915 spin_unlock(&dlm
->spinlock
);
918 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource
*res
)
921 struct list_head
*iter
, *queue
= &res
->granted
;
924 for (i
=0; i
<3; i
++) {
925 list_for_each(iter
, queue
)
933 static int dlm_send_mig_lockres_msg(struct dlm_ctxt
*dlm
,
934 struct dlm_migratable_lockres
*mres
,
936 struct dlm_lock_resource
*res
,
939 u64 mig_cookie
= be64_to_cpu(mres
->mig_cookie
);
940 int mres_total_locks
= be32_to_cpu(mres
->total_locks
);
941 int sz
, ret
= 0, status
= 0;
942 u8 orig_flags
= mres
->flags
,
943 orig_master
= mres
->master
;
945 BUG_ON(mres
->num_locks
> DLM_MAX_MIGRATABLE_LOCKS
);
946 if (!mres
->num_locks
)
949 sz
= sizeof(struct dlm_migratable_lockres
) +
950 (mres
->num_locks
* sizeof(struct dlm_migratable_lock
));
952 /* add an all-done flag if we reached the last lock */
953 orig_flags
= mres
->flags
;
954 BUG_ON(total_locks
> mres_total_locks
);
955 if (total_locks
== mres_total_locks
)
956 mres
->flags
|= DLM_MRES_ALL_DONE
;
959 ret
= o2net_send_message(DLM_MIG_LOCKRES_MSG
, dlm
->key
, mres
,
960 sz
, send_to
, &status
);
962 /* XXX: negative status is not handled.
963 * this will end up killing this node. */
966 /* might get an -ENOMEM back here */
971 if (ret
== -EFAULT
) {
972 mlog(ML_ERROR
, "node %u told me to kill "
973 "myself!\n", send_to
);
979 /* zero and reinit the message buffer */
980 dlm_init_migratable_lockres(mres
, res
->lockname
.name
,
981 res
->lockname
.len
, mres_total_locks
,
982 mig_cookie
, orig_flags
, orig_master
);
986 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres
*mres
,
987 const char *lockname
, int namelen
,
988 int total_locks
, u64 cookie
,
991 /* mres here is one full page */
992 memset(mres
, 0, PAGE_SIZE
);
993 mres
->lockname_len
= namelen
;
994 memcpy(mres
->lockname
, lockname
, namelen
);
996 mres
->total_locks
= cpu_to_be32(total_locks
);
997 mres
->mig_cookie
= cpu_to_be64(cookie
);
999 mres
->master
= master
;
1003 /* returns 1 if this lock fills the network structure,
1005 static int dlm_add_lock_to_array(struct dlm_lock
*lock
,
1006 struct dlm_migratable_lockres
*mres
, int queue
)
1008 struct dlm_migratable_lock
*ml
;
1009 int lock_num
= mres
->num_locks
;
1011 ml
= &(mres
->ml
[lock_num
]);
1012 ml
->cookie
= lock
->ml
.cookie
;
1013 ml
->type
= lock
->ml
.type
;
1014 ml
->convert_type
= lock
->ml
.convert_type
;
1015 ml
->highest_blocked
= lock
->ml
.highest_blocked
;
1018 ml
->flags
= lock
->lksb
->flags
;
1019 /* send our current lvb */
1020 if (ml
->type
== LKM_EXMODE
||
1021 ml
->type
== LKM_PRMODE
) {
1022 /* if it is already set, this had better be a PR
1023 * and it has to match */
1024 if (mres
->lvb
[0] && (ml
->type
== LKM_EXMODE
||
1025 memcmp(mres
->lvb
, lock
->lksb
->lvb
, DLM_LVB_LEN
))) {
1026 mlog(ML_ERROR
, "mismatched lvbs!\n");
1027 __dlm_print_one_lock_resource(lock
->lockres
);
1030 memcpy(mres
->lvb
, lock
->lksb
->lvb
, DLM_LVB_LEN
);
1033 ml
->node
= lock
->ml
.node
;
1035 /* we reached the max, send this network message */
1036 if (mres
->num_locks
== DLM_MAX_MIGRATABLE_LOCKS
)
1042 int dlm_send_one_lockres(struct dlm_ctxt
*dlm
, struct dlm_lock_resource
*res
,
1043 struct dlm_migratable_lockres
*mres
,
1044 u8 send_to
, u8 flags
)
1046 struct list_head
*queue
, *iter
;
1049 struct dlm_lock
*lock
;
1052 BUG_ON(!(flags
& (DLM_MRES_RECOVERY
|DLM_MRES_MIGRATION
)));
1054 mlog(0, "sending to %u\n", send_to
);
1056 total_locks
= dlm_num_locks_in_lockres(res
);
1057 if (total_locks
> DLM_MAX_MIGRATABLE_LOCKS
) {
1058 /* rare, but possible */
1059 mlog(0, "argh. lockres has %d locks. this will "
1060 "require more than one network packet to "
1061 "migrate\n", total_locks
);
1062 mig_cookie
= dlm_get_next_mig_cookie();
1065 dlm_init_migratable_lockres(mres
, res
->lockname
.name
,
1066 res
->lockname
.len
, total_locks
,
1067 mig_cookie
, flags
, res
->owner
);
1070 for (i
=DLM_GRANTED_LIST
; i
<=DLM_BLOCKED_LIST
; i
++) {
1071 queue
= dlm_list_idx_to_ptr(res
, i
);
1072 list_for_each(iter
, queue
) {
1073 lock
= list_entry (iter
, struct dlm_lock
, list
);
1075 /* add another lock. */
1077 if (!dlm_add_lock_to_array(lock
, mres
, i
))
1080 /* this filled the lock message,
1081 * we must send it immediately. */
1082 ret
= dlm_send_mig_lockres_msg(dlm
, mres
, send_to
,
1086 mlog(ML_ERROR
, "dlm_send_mig_lockres_msg "
1087 "returned %d, TODO\n", ret
);
1092 /* flush any remaining locks */
1093 ret
= dlm_send_mig_lockres_msg(dlm
, mres
, send_to
, res
, total_locks
);
1096 mlog(ML_ERROR
, "dlm_send_mig_lockres_msg returned %d, "
1106 * this message will contain no more than one page worth of
1107 * recovery data, and it will work on only one lockres.
1108 * there may be many locks in this page, and we may need to wait
1109 * for additional packets to complete all the locks (rare, but
1113 * NOTE: the allocation error cases here are scary
1114 * we really cannot afford to fail an alloc in recovery
1115 * do we spin? returning an error only delays the problem really
1118 int dlm_mig_lockres_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
1120 struct dlm_ctxt
*dlm
= data
;
1121 struct dlm_migratable_lockres
*mres
=
1122 (struct dlm_migratable_lockres
*)msg
->buf
;
1126 struct dlm_work_item
*item
= NULL
;
1127 struct dlm_lock_resource
*res
= NULL
;
1132 BUG_ON(!(mres
->flags
& (DLM_MRES_RECOVERY
|DLM_MRES_MIGRATION
)));
1134 real_master
= mres
->master
;
1135 if (real_master
== DLM_LOCK_RES_OWNER_UNKNOWN
) {
1136 /* cannot migrate a lockres with no master */
1137 BUG_ON(!(mres
->flags
& DLM_MRES_RECOVERY
));
1140 mlog(0, "%s message received from node %u\n",
1141 (mres
->flags
& DLM_MRES_RECOVERY
) ?
1142 "recovery" : "migration", mres
->master
);
1143 if (mres
->flags
& DLM_MRES_ALL_DONE
)
1144 mlog(0, "all done flag. all lockres data received!\n");
1147 buf
= kmalloc(be16_to_cpu(msg
->data_len
), GFP_KERNEL
);
1148 item
= kcalloc(1, sizeof(*item
), GFP_KERNEL
);
1152 /* lookup the lock to see if we have a secondary queue for this
1153 * already... just add the locks in and this will have its owner
1154 * and RECOVERY flag changed when it completes. */
1155 res
= dlm_lookup_lockres(dlm
, mres
->lockname
, mres
->lockname_len
);
1157 /* this will get a ref on res */
1158 /* mark it as recovering/migrating and hash it */
1159 spin_lock(&res
->spinlock
);
1160 if (mres
->flags
& DLM_MRES_RECOVERY
) {
1161 res
->state
|= DLM_LOCK_RES_RECOVERING
;
1163 if (res
->state
& DLM_LOCK_RES_MIGRATING
) {
1164 /* this is at least the second
1165 * lockres message */
1166 mlog(0, "lock %.*s is already migrating\n",
1169 } else if (res
->state
& DLM_LOCK_RES_RECOVERING
) {
1170 /* caller should BUG */
1171 mlog(ML_ERROR
, "node is attempting to migrate "
1172 "lock %.*s, but marked as recovering!\n",
1173 mres
->lockname_len
, mres
->lockname
);
1175 spin_unlock(&res
->spinlock
);
1178 res
->state
|= DLM_LOCK_RES_MIGRATING
;
1180 spin_unlock(&res
->spinlock
);
1182 /* need to allocate, just like if it was
1183 * mastered here normally */
1184 res
= dlm_new_lockres(dlm
, mres
->lockname
, mres
->lockname_len
);
1188 /* to match the ref that we would have gotten if
1189 * dlm_lookup_lockres had succeeded */
1190 dlm_lockres_get(res
);
1192 /* mark it as recovering/migrating and hash it */
1193 if (mres
->flags
& DLM_MRES_RECOVERY
)
1194 res
->state
|= DLM_LOCK_RES_RECOVERING
;
1196 res
->state
|= DLM_LOCK_RES_MIGRATING
;
1198 spin_lock(&dlm
->spinlock
);
1199 __dlm_insert_lockres(dlm
, res
);
1200 spin_unlock(&dlm
->spinlock
);
1202 /* now that the new lockres is inserted,
1203 * make it usable by other processes */
1204 spin_lock(&res
->spinlock
);
1205 res
->state
&= ~DLM_LOCK_RES_IN_PROGRESS
;
1206 spin_unlock(&res
->spinlock
);
1208 /* add an extra ref for just-allocated lockres
1209 * otherwise the lockres will be purged immediately */
1210 dlm_lockres_get(res
);
1214 /* at this point we have allocated everything we need,
1215 * and we have a hashed lockres with an extra ref and
1216 * the proper res->state flags. */
1218 if (mres
->master
== DLM_LOCK_RES_OWNER_UNKNOWN
) {
1219 /* migration cannot have an unknown master */
1220 BUG_ON(!(mres
->flags
& DLM_MRES_RECOVERY
));
1221 mlog(0, "recovery has passed me a lockres with an "
1222 "unknown owner.. will need to requery: "
1223 "%.*s\n", mres
->lockname_len
, mres
->lockname
);
1225 spin_lock(&res
->spinlock
);
1226 dlm_change_lockres_owner(dlm
, res
, dlm
->node_num
);
1227 spin_unlock(&res
->spinlock
);
1230 /* queue up work for dlm_mig_lockres_worker */
1231 dlm_grab(dlm
); /* get an extra ref for the work item */
1232 memcpy(buf
, msg
->buf
, be16_to_cpu(msg
->data_len
)); /* copy the whole message */
1233 dlm_init_work_item(dlm
, item
, dlm_mig_lockres_worker
, buf
);
1234 item
->u
.ml
.lockres
= res
; /* already have a ref */
1235 item
->u
.ml
.real_master
= real_master
;
1236 spin_lock(&dlm
->work_lock
);
1237 list_add_tail(&item
->list
, &dlm
->work_list
);
1238 spin_unlock(&dlm
->work_lock
);
1239 schedule_work(&dlm
->dispatched_work
);
1255 static void dlm_mig_lockres_worker(struct dlm_work_item
*item
, void *data
)
1257 struct dlm_ctxt
*dlm
= data
;
1258 struct dlm_migratable_lockres
*mres
;
1260 struct dlm_lock_resource
*res
;
1264 mres
= (struct dlm_migratable_lockres
*)data
;
1266 res
= item
->u
.ml
.lockres
;
1267 real_master
= item
->u
.ml
.real_master
;
1269 if (real_master
== DLM_LOCK_RES_OWNER_UNKNOWN
) {
1270 /* this case is super-rare. only occurs if
1271 * node death happens during migration. */
1273 ret
= dlm_lockres_master_requery(dlm
, res
, &real_master
);
1275 mlog(0, "dlm_lockres_master_requery ret=%d\n",
1279 if (real_master
== DLM_LOCK_RES_OWNER_UNKNOWN
) {
1280 mlog(0, "lockres %.*s not claimed. "
1281 "this node will take it.\n",
1282 res
->lockname
.len
, res
->lockname
.name
);
1284 mlog(0, "master needs to respond to sender "
1285 "that node %u still owns %.*s\n",
1286 real_master
, res
->lockname
.len
,
1287 res
->lockname
.name
);
1288 /* cannot touch this lockres */
1293 ret
= dlm_process_recovery_data(dlm
, res
, mres
);
1295 mlog(0, "dlm_process_recovery_data returned %d\n", ret
);
1297 mlog(0, "dlm_process_recovery_data succeeded\n");
1299 if ((mres
->flags
& (DLM_MRES_MIGRATION
|DLM_MRES_ALL_DONE
)) ==
1300 (DLM_MRES_MIGRATION
|DLM_MRES_ALL_DONE
)) {
1301 ret
= dlm_finish_migration(dlm
, res
, mres
->master
);
1313 int dlm_lockres_master_requery(struct dlm_ctxt
*dlm
,
1314 struct dlm_lock_resource
*res
, u8
*real_master
)
1316 struct dlm_node_iter iter
;
1320 *real_master
= DLM_LOCK_RES_OWNER_UNKNOWN
;
1322 /* we only reach here if one of the two nodes in a
1323 * migration died while the migration was in progress.
1324 * at this point we need to requery the master. we
1325 * know that the new_master got as far as creating
1326 * an mle on at least one node, but we do not know
1327 * if any nodes had actually cleared the mle and set
1328 * the master to the new_master. the old master
1329 * is supposed to set the owner to UNKNOWN in the
1330 * event of a new_master death, so the only possible
1331 * responses that we can get from nodes here are
1332 * that the master is new_master, or that the master
1334 * if all nodes come back with UNKNOWN then we know
1335 * the lock needs remastering here.
1336 * if any node comes back with a valid master, check
1337 * to see if that master is the one that we are
1338 * recovering. if so, then the new_master died and
1339 * we need to remaster this lock. if not, then the
1340 * new_master survived and that node will respond to
1341 * other nodes about the owner.
1342 * if there is an owner, this node needs to dump this
1343 * lockres and alert the sender that this lockres
1345 spin_lock(&dlm
->spinlock
);
1346 dlm_node_iter_init(dlm
->domain_map
, &iter
);
1347 spin_unlock(&dlm
->spinlock
);
1349 while ((nodenum
= dlm_node_iter_next(&iter
)) >= 0) {
1350 /* do not send to self */
1351 if (nodenum
== dlm
->node_num
)
1353 ret
= dlm_do_master_requery(dlm
, res
, nodenum
, real_master
);
1356 if (!dlm_is_host_down(ret
))
1358 /* host is down, so answer for that node would be
1359 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
1361 if (*real_master
!= DLM_LOCK_RES_OWNER_UNKNOWN
) {
1362 mlog(0, "lock master is %u\n", *real_master
);
1370 int dlm_do_master_requery(struct dlm_ctxt
*dlm
, struct dlm_lock_resource
*res
,
1371 u8 nodenum
, u8
*real_master
)
1374 struct dlm_master_requery req
;
1375 int status
= DLM_LOCK_RES_OWNER_UNKNOWN
;
1377 memset(&req
, 0, sizeof(req
));
1378 req
.node_idx
= dlm
->node_num
;
1379 req
.namelen
= res
->lockname
.len
;
1380 memcpy(req
.name
, res
->lockname
.name
, res
->lockname
.len
);
1382 ret
= o2net_send_message(DLM_MASTER_REQUERY_MSG
, dlm
->key
,
1383 &req
, sizeof(req
), nodenum
, &status
);
1384 /* XXX: negative status not handled properly here. */
1389 BUG_ON(status
> DLM_LOCK_RES_OWNER_UNKNOWN
);
1390 *real_master
= (u8
) (status
& 0xff);
1391 mlog(0, "node %u responded to master requery with %u\n",
1392 nodenum
, *real_master
);
1399 /* this function cannot error, so unless the sending
1400 * or receiving of the message failed, the owner can
1402 int dlm_master_requery_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
1404 struct dlm_ctxt
*dlm
= data
;
1405 struct dlm_master_requery
*req
= (struct dlm_master_requery
*)msg
->buf
;
1406 struct dlm_lock_resource
*res
= NULL
;
1407 int master
= DLM_LOCK_RES_OWNER_UNKNOWN
;
1408 u32 flags
= DLM_ASSERT_MASTER_REQUERY
;
1410 if (!dlm_grab(dlm
)) {
1411 /* since the domain has gone away on this
1412 * node, the proper response is UNKNOWN */
1416 spin_lock(&dlm
->spinlock
);
1417 res
= __dlm_lookup_lockres(dlm
, req
->name
, req
->namelen
);
1419 spin_lock(&res
->spinlock
);
1420 master
= res
->owner
;
1421 if (master
== dlm
->node_num
) {
1422 int ret
= dlm_dispatch_assert_master(dlm
, res
,
1425 mlog_errno(-ENOMEM
);
1430 spin_unlock(&res
->spinlock
);
1432 spin_unlock(&dlm
->spinlock
);
1438 static inline struct list_head
*
1439 dlm_list_num_to_pointer(struct dlm_lock_resource
*res
, int list_num
)
1441 struct list_head
*ret
;
1442 BUG_ON(list_num
< 0);
1443 BUG_ON(list_num
> 2);
1444 ret
= &(res
->granted
);
1448 /* TODO: do ast flush business
1449 * TODO: do MIGRATING and RECOVERING spinning
1453 * NOTE about in-flight requests during migration:
1455 * Before attempting the migrate, the master has marked the lockres as
1456 * MIGRATING and then flushed all of its pending ASTS. So any in-flight
1457 * requests either got queued before the MIGRATING flag got set, in which
1458 * case the lock data will reflect the change and a return message is on
1459 * the way, or the request failed to get in before MIGRATING got set. In
1460 * this case, the caller will be told to spin and wait for the MIGRATING
1461 * flag to be dropped, then recheck the master.
1462 * This holds true for the convert, cancel and unlock cases, and since lvb
1463 * updates are tied to these same messages, it applies to lvb updates as
1464 * well. For the lock case, there is no way a lock can be on the master
1465 * queue and not be on the secondary queue since the lock is always added
1466 * locally first. This means that the new target node will never be sent
1467 * a lock that he doesn't already have on the list.
1468 * In total, this means that the local lock is correct and should not be
1469 * updated to match the one sent by the master. Any messages sent back
1470 * from the master before the MIGRATING flag will bring the lock properly
1471 * up-to-date, and the change will be ordered properly for the waiter.
1472 * We will *not* attempt to modify the lock underneath the waiter.
1475 static int dlm_process_recovery_data(struct dlm_ctxt
*dlm
,
1476 struct dlm_lock_resource
*res
,
1477 struct dlm_migratable_lockres
*mres
)
1479 struct dlm_migratable_lock
*ml
;
1480 struct list_head
*queue
;
1481 struct dlm_lock
*newlock
= NULL
;
1482 struct dlm_lockstatus
*lksb
= NULL
;
1485 struct list_head
*iter
;
1486 struct dlm_lock
*lock
= NULL
;
1488 mlog(0, "running %d locks for this lockres\n", mres
->num_locks
);
1489 for (i
=0; i
<mres
->num_locks
; i
++) {
1490 ml
= &(mres
->ml
[i
]);
1491 BUG_ON(ml
->highest_blocked
!= LKM_IVMODE
);
1495 queue
= dlm_list_num_to_pointer(res
, ml
->list
);
1497 /* if the lock is for the local node it needs to
1498 * be moved to the proper location within the queue.
1499 * do not allocate a new lock structure. */
1500 if (ml
->node
== dlm
->node_num
) {
1501 /* MIGRATION ONLY! */
1502 BUG_ON(!(mres
->flags
& DLM_MRES_MIGRATION
));
1504 spin_lock(&res
->spinlock
);
1505 list_for_each(iter
, queue
) {
1506 lock
= list_entry (iter
, struct dlm_lock
, list
);
1507 if (lock
->ml
.cookie
!= ml
->cookie
)
1513 /* lock is always created locally first, and
1514 * destroyed locally last. it must be on the list */
1517 mlog(ML_ERROR
, "could not find local lock "
1518 "with cookie %u:%llu!\n",
1519 dlm_get_lock_cookie_node(c
),
1520 dlm_get_lock_cookie_seq(c
));
1523 BUG_ON(lock
->ml
.node
!= ml
->node
);
1525 /* see NOTE above about why we do not update
1526 * to match the master here */
1528 /* move the lock to its proper place */
1529 /* do not alter lock refcount. switching lists. */
1530 list_move_tail(&lock
->list
, queue
);
1531 spin_unlock(&res
->spinlock
);
1533 mlog(0, "just reordered a local lock!\n");
1537 /* lock is for another node. */
1538 newlock
= dlm_new_lock(ml
->type
, ml
->node
,
1539 be64_to_cpu(ml
->cookie
), NULL
);
1544 lksb
= newlock
->lksb
;
1545 dlm_lock_attach_lockres(newlock
, res
);
1547 if (ml
->convert_type
!= LKM_IVMODE
) {
1548 BUG_ON(queue
!= &res
->converting
);
1549 newlock
->ml
.convert_type
= ml
->convert_type
;
1551 lksb
->flags
|= (ml
->flags
&
1552 (DLM_LKSB_PUT_LVB
|DLM_LKSB_GET_LVB
));
1555 if (lksb
->flags
& DLM_LKSB_PUT_LVB
) {
1556 /* other node was trying to update
1557 * lvb when node died. recreate the
1558 * lksb with the updated lvb. */
1559 memcpy(lksb
->lvb
, mres
->lvb
, DLM_LVB_LEN
);
1561 /* otherwise, the node is sending its
1562 * most recent valid lvb info */
1563 BUG_ON(ml
->type
!= LKM_EXMODE
&&
1564 ml
->type
!= LKM_PRMODE
);
1565 if (res
->lvb
[0] && (ml
->type
== LKM_EXMODE
||
1566 memcmp(res
->lvb
, mres
->lvb
, DLM_LVB_LEN
))) {
1567 mlog(ML_ERROR
, "received bad lvb!\n");
1568 __dlm_print_one_lock_resource(res
);
1571 memcpy(res
->lvb
, mres
->lvb
, DLM_LVB_LEN
);
1577 * wrt lock queue ordering and recovery:
1578 * 1. order of locks on granted queue is
1580 * 2. order of locks on converting queue is
1581 * LOST with the node death. sorry charlie.
1582 * 3. order of locks on the blocked queue is
1584 * order of locks does not affect integrity, it
1585 * just means that a lock request may get pushed
1586 * back in line as a result of the node death.
1587 * also note that for a given node the lock order
1588 * for its secondary queue locks is preserved
1589 * relative to each other, but clearly *not*
1590 * preserved relative to locks from other nodes.
1592 spin_lock(&res
->spinlock
);
1593 dlm_lock_get(newlock
);
1594 list_add_tail(&newlock
->list
, queue
);
1595 spin_unlock(&res
->spinlock
);
1597 mlog(0, "done running all the locks\n");
1603 dlm_lock_put(newlock
);
1610 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt
*dlm
,
1611 struct dlm_lock_resource
*res
)
1614 struct list_head
*queue
, *iter
, *iter2
;
1615 struct dlm_lock
*lock
;
1617 res
->state
|= DLM_LOCK_RES_RECOVERING
;
1618 if (!list_empty(&res
->recovering
))
1619 list_del_init(&res
->recovering
);
1620 list_add_tail(&res
->recovering
, &dlm
->reco
.resources
);
1622 /* find any pending locks and put them back on proper list */
1623 for (i
=DLM_BLOCKED_LIST
; i
>=DLM_GRANTED_LIST
; i
--) {
1624 queue
= dlm_list_idx_to_ptr(res
, i
);
1625 list_for_each_safe(iter
, iter2
, queue
) {
1626 lock
= list_entry (iter
, struct dlm_lock
, list
);
1628 if (lock
->convert_pending
) {
1629 /* move converting lock back to granted */
1630 BUG_ON(i
!= DLM_CONVERTING_LIST
);
1631 mlog(0, "node died with convert pending "
1632 "on %.*s. move back to granted list.\n",
1633 res
->lockname
.len
, res
->lockname
.name
);
1634 dlm_revert_pending_convert(res
, lock
);
1635 lock
->convert_pending
= 0;
1636 } else if (lock
->lock_pending
) {
1637 /* remove pending lock requests completely */
1638 BUG_ON(i
!= DLM_BLOCKED_LIST
);
1639 mlog(0, "node died with lock pending "
1640 "on %.*s. remove from blocked list and skip.\n",
1641 res
->lockname
.len
, res
->lockname
.name
);
1642 /* lock will be floating until ref in
1643 * dlmlock_remote is freed after the network
1644 * call returns. ok for it to not be on any
1645 * list since no ast can be called
1646 * (the master is dead). */
1647 dlm_revert_pending_lock(res
, lock
);
1648 lock
->lock_pending
= 0;
1649 } else if (lock
->unlock_pending
) {
1650 /* if an unlock was in progress, treat as
1651 * if this had completed successfully
1652 * before sending this lock state to the
1653 * new master. note that the dlm_unlock
1654 * call is still responsible for calling
1655 * the unlockast. that will happen after
1656 * the network call times out. for now,
1657 * just move lists to prepare the new
1658 * recovery master. */
1659 BUG_ON(i
!= DLM_GRANTED_LIST
);
1660 mlog(0, "node died with unlock pending "
1661 "on %.*s. remove from blocked list and skip.\n",
1662 res
->lockname
.len
, res
->lockname
.name
);
1663 dlm_commit_pending_unlock(res
, lock
);
1664 lock
->unlock_pending
= 0;
1665 } else if (lock
->cancel_pending
) {
1666 /* if a cancel was in progress, treat as
1667 * if this had completed successfully
1668 * before sending this lock state to the
1670 BUG_ON(i
!= DLM_CONVERTING_LIST
);
1671 mlog(0, "node died with cancel pending "
1672 "on %.*s. move back to granted list.\n",
1673 res
->lockname
.len
, res
->lockname
.name
);
1674 dlm_commit_pending_cancel(res
, lock
);
1675 lock
->cancel_pending
= 0;
1684 /* removes all recovered locks from the recovery list.
1685 * sets the res->owner to the new master.
1686 * unsets the RECOVERY flag and wakes waiters. */
1687 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt
*dlm
,
1688 u8 dead_node
, u8 new_master
)
1691 struct list_head
*iter
, *iter2
;
1692 struct hlist_node
*hash_iter
;
1693 struct hlist_head
*bucket
;
1695 struct dlm_lock_resource
*res
;
1699 assert_spin_locked(&dlm
->spinlock
);
1701 list_for_each_safe(iter
, iter2
, &dlm
->reco
.resources
) {
1702 res
= list_entry (iter
, struct dlm_lock_resource
, recovering
);
1703 if (res
->owner
== dead_node
) {
1704 list_del_init(&res
->recovering
);
1705 spin_lock(&res
->spinlock
);
1706 dlm_change_lockres_owner(dlm
, res
, new_master
);
1707 res
->state
&= ~DLM_LOCK_RES_RECOVERING
;
1708 __dlm_dirty_lockres(dlm
, res
);
1709 spin_unlock(&res
->spinlock
);
1714 /* this will become unnecessary eventually, but
1715 * for now we need to run the whole hash, clear
1716 * the RECOVERING state and set the owner
1718 for (i
= 0; i
< DLM_HASH_BUCKETS
; i
++) {
1719 bucket
= &(dlm
->lockres_hash
[i
]);
1720 hlist_for_each_entry(res
, hash_iter
, bucket
, hash_node
) {
1721 if (res
->state
& DLM_LOCK_RES_RECOVERING
) {
1722 if (res
->owner
== dead_node
) {
1723 mlog(0, "(this=%u) res %.*s owner=%u "
1724 "was not on recovering list, but "
1725 "clearing state anyway\n",
1726 dlm
->node_num
, res
->lockname
.len
,
1727 res
->lockname
.name
, new_master
);
1728 } else if (res
->owner
== dlm
->node_num
) {
1729 mlog(0, "(this=%u) res %.*s owner=%u "
1730 "was not on recovering list, "
1731 "owner is THIS node, clearing\n",
1732 dlm
->node_num
, res
->lockname
.len
,
1733 res
->lockname
.name
, new_master
);
1737 if (!list_empty(&res
->recovering
)) {
1738 mlog(0, "%s:%.*s: lockres was "
1739 "marked RECOVERING, owner=%u\n",
1740 dlm
->name
, res
->lockname
.len
,
1741 res
->lockname
.name
, res
->owner
);
1742 list_del_init(&res
->recovering
);
1744 spin_lock(&res
->spinlock
);
1745 dlm_change_lockres_owner(dlm
, res
, new_master
);
1746 res
->state
&= ~DLM_LOCK_RES_RECOVERING
;
1747 __dlm_dirty_lockres(dlm
, res
);
1748 spin_unlock(&res
->spinlock
);
1755 static inline int dlm_lvb_needs_invalidation(struct dlm_lock
*lock
, int local
)
1758 if (lock
->ml
.type
!= LKM_EXMODE
&&
1759 lock
->ml
.type
!= LKM_PRMODE
)
1761 } else if (lock
->ml
.type
== LKM_EXMODE
)
1766 static void dlm_revalidate_lvb(struct dlm_ctxt
*dlm
,
1767 struct dlm_lock_resource
*res
, u8 dead_node
)
1769 struct list_head
*iter
, *queue
;
1770 struct dlm_lock
*lock
;
1771 int blank_lvb
= 0, local
= 0;
1775 assert_spin_locked(&dlm
->spinlock
);
1776 assert_spin_locked(&res
->spinlock
);
1778 if (res
->owner
== dlm
->node_num
)
1779 /* if this node owned the lockres, and if the dead node
1780 * had an EX when he died, blank out the lvb */
1781 search_node
= dead_node
;
1783 /* if this is a secondary lockres, and we had no EX or PR
1784 * locks granted, we can no longer trust the lvb */
1785 search_node
= dlm
->node_num
;
1786 local
= 1; /* check local state for valid lvb */
1789 for (i
=DLM_GRANTED_LIST
; i
<=DLM_CONVERTING_LIST
; i
++) {
1790 queue
= dlm_list_idx_to_ptr(res
, i
);
1791 list_for_each(iter
, queue
) {
1792 lock
= list_entry (iter
, struct dlm_lock
, list
);
1793 if (lock
->ml
.node
== search_node
) {
1794 if (dlm_lvb_needs_invalidation(lock
, local
)) {
1795 /* zero the lksb lvb and lockres lvb */
1797 memset(lock
->lksb
->lvb
, 0, DLM_LVB_LEN
);
1804 mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
1805 res
->lockname
.len
, res
->lockname
.name
, dead_node
);
1806 memset(res
->lvb
, 0, DLM_LVB_LEN
);
1810 static void dlm_free_dead_locks(struct dlm_ctxt
*dlm
,
1811 struct dlm_lock_resource
*res
, u8 dead_node
)
1813 struct list_head
*iter
, *tmpiter
;
1814 struct dlm_lock
*lock
;
1816 /* this node is the lockres master:
1817 * 1) remove any stale locks for the dead node
1818 * 2) if the dead node had an EX when he died, blank out the lvb
1820 assert_spin_locked(&dlm
->spinlock
);
1821 assert_spin_locked(&res
->spinlock
);
1823 /* TODO: check pending_asts, pending_basts here */
1824 list_for_each_safe(iter
, tmpiter
, &res
->granted
) {
1825 lock
= list_entry (iter
, struct dlm_lock
, list
);
1826 if (lock
->ml
.node
== dead_node
) {
1827 list_del_init(&lock
->list
);
1831 list_for_each_safe(iter
, tmpiter
, &res
->converting
) {
1832 lock
= list_entry (iter
, struct dlm_lock
, list
);
1833 if (lock
->ml
.node
== dead_node
) {
1834 list_del_init(&lock
->list
);
1838 list_for_each_safe(iter
, tmpiter
, &res
->blocked
) {
1839 lock
= list_entry (iter
, struct dlm_lock
, list
);
1840 if (lock
->ml
.node
== dead_node
) {
1841 list_del_init(&lock
->list
);
1846 /* do not kick thread yet */
1847 __dlm_dirty_lockres(dlm
, res
);
1850 /* if this node is the recovery master, and there are no
1851 * locks for a given lockres owned by this node that are in
1852 * either PR or EX mode, zero out the lvb before requesting.
1857 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt
*dlm
, u8 dead_node
)
1859 struct hlist_node
*iter
;
1860 struct dlm_lock_resource
*res
;
1862 struct hlist_head
*bucket
;
1863 struct dlm_lock
*lock
;
1866 /* purge any stale mles */
1867 dlm_clean_master_list(dlm
, dead_node
);
1870 * now clean up all lock resources. there are two rules:
1872 * 1) if the dead node was the master, move the lockres
1873 * to the recovering list. set the RECOVERING flag.
1874 * this lockres needs to be cleaned up before it can
1877 * 2) if this node was the master, remove all locks from
1878 * each of the lockres queues that were owned by the
1879 * dead node. once recovery finishes, the dlm thread
1880 * can be kicked again to see if any ASTs or BASTs
1881 * need to be fired as a result.
1883 for (i
= 0; i
< DLM_HASH_BUCKETS
; i
++) {
1884 bucket
= &(dlm
->lockres_hash
[i
]);
1885 hlist_for_each_entry(res
, iter
, bucket
, hash_node
) {
1886 /* always prune any $RECOVERY entries for dead nodes,
1887 * otherwise hangs can occur during later recovery */
1888 if (dlm_is_recovery_lock(res
->lockname
.name
,
1889 res
->lockname
.len
)) {
1890 spin_lock(&res
->spinlock
);
1891 list_for_each_entry(lock
, &res
->granted
, list
) {
1892 if (lock
->ml
.node
== dead_node
) {
1893 mlog(0, "AHA! there was "
1894 "a $RECOVERY lock for dead "
1896 dead_node
, dlm
->name
);
1897 list_del_init(&lock
->list
);
1902 spin_unlock(&res
->spinlock
);
1905 spin_lock(&res
->spinlock
);
1906 /* zero the lvb if necessary */
1907 dlm_revalidate_lvb(dlm
, res
, dead_node
);
1908 if (res
->owner
== dead_node
)
1909 dlm_move_lockres_to_recovery_list(dlm
, res
);
1910 else if (res
->owner
== dlm
->node_num
) {
1911 dlm_free_dead_locks(dlm
, res
, dead_node
);
1912 __dlm_lockres_calc_usage(dlm
, res
);
1914 spin_unlock(&res
->spinlock
);
1920 static void __dlm_hb_node_down(struct dlm_ctxt
*dlm
, int idx
)
1922 assert_spin_locked(&dlm
->spinlock
);
1924 /* check to see if the node is already considered dead */
1925 if (!test_bit(idx
, dlm
->live_nodes_map
)) {
1926 mlog(0, "for domain %s, node %d is already dead. "
1927 "another node likely did recovery already.\n",
1932 /* check to see if we do not care about this node */
1933 if (!test_bit(idx
, dlm
->domain_map
)) {
1934 /* This also catches the case that we get a node down
1935 * but haven't joined the domain yet. */
1936 mlog(0, "node %u already removed from domain!\n", idx
);
1940 clear_bit(idx
, dlm
->live_nodes_map
);
1942 /* Clean up join state on node death. */
1943 if (dlm
->joining_node
== idx
) {
1944 mlog(0, "Clearing join state for node %u\n", idx
);
1945 __dlm_set_joining_node(dlm
, DLM_LOCK_RES_OWNER_UNKNOWN
);
1948 /* make sure local cleanup occurs before the heartbeat events */
1949 if (!test_bit(idx
, dlm
->recovery_map
))
1950 dlm_do_local_recovery_cleanup(dlm
, idx
);
1952 /* notify anything attached to the heartbeat events */
1953 dlm_hb_event_notify_attached(dlm
, idx
, 0);
1955 mlog(0, "node %u being removed from domain map!\n", idx
);
1956 clear_bit(idx
, dlm
->domain_map
);
1957 /* wake up migration waiters if a node goes down.
1958 * perhaps later we can genericize this for other waiters. */
1959 wake_up(&dlm
->migration_wq
);
1961 if (test_bit(idx
, dlm
->recovery_map
))
1962 mlog(0, "domain %s, node %u already added "
1963 "to recovery map!\n", dlm
->name
, idx
);
1965 set_bit(idx
, dlm
->recovery_map
);
1968 void dlm_hb_node_down_cb(struct o2nm_node
*node
, int idx
, void *data
)
1970 struct dlm_ctxt
*dlm
= data
;
1975 spin_lock(&dlm
->spinlock
);
1976 __dlm_hb_node_down(dlm
, idx
);
1977 spin_unlock(&dlm
->spinlock
);
1982 void dlm_hb_node_up_cb(struct o2nm_node
*node
, int idx
, void *data
)
1984 struct dlm_ctxt
*dlm
= data
;
1989 spin_lock(&dlm
->spinlock
);
1990 set_bit(idx
, dlm
->live_nodes_map
);
1991 /* do NOT notify mle attached to the heartbeat events.
1992 * new nodes are not interesting in mastery until joined. */
1993 spin_unlock(&dlm
->spinlock
);
1998 static void dlm_reco_ast(void *astdata
)
2000 struct dlm_ctxt
*dlm
= astdata
;
2001 mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2002 dlm
->node_num
, dlm
->name
);
2004 static void dlm_reco_bast(void *astdata
, int blocked_type
)
2006 struct dlm_ctxt
*dlm
= astdata
;
2007 mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2008 dlm
->node_num
, dlm
->name
);
2010 static void dlm_reco_unlock_ast(void *astdata
, enum dlm_status st
)
2012 mlog(0, "unlockast for recovery lock fired!\n");
2016 * dlm_pick_recovery_master will continually attempt to use
2017 * dlmlock() on the special "$RECOVERY" lockres with the
2018 * LKM_NOQUEUE flag to get an EX. every thread that enters
2019 * this function on each node racing to become the recovery
2020 * master will not stop attempting this until either:
2021 * a) this node gets the EX (and becomes the recovery master),
2022 * or b) dlm->reco.new_master gets set to some nodenum
2023 * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2024 * so each time a recovery master is needed, the entire cluster
2025 * will sync at this point. if the new master dies, that will
2026 * be detected in dlm_do_recovery */
2027 static int dlm_pick_recovery_master(struct dlm_ctxt
*dlm
)
2029 enum dlm_status ret
;
2030 struct dlm_lockstatus lksb
;
2031 int status
= -EINVAL
;
2033 mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2034 dlm
->name
, jiffies
, dlm
->reco
.dead_node
, dlm
->node_num
);
2036 memset(&lksb
, 0, sizeof(lksb
));
2038 ret
= dlmlock(dlm
, LKM_EXMODE
, &lksb
, LKM_NOQUEUE
|LKM_RECOVERY
,
2039 DLM_RECOVERY_LOCK_NAME
, dlm_reco_ast
, dlm
, dlm_reco_bast
);
2041 mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2042 dlm
->name
, ret
, lksb
.status
);
2044 if (ret
== DLM_NORMAL
) {
2045 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2046 dlm
->name
, dlm
->node_num
);
2048 /* got the EX lock. check to see if another node
2049 * just became the reco master */
2050 if (dlm_reco_master_ready(dlm
)) {
2051 mlog(0, "%s: got reco EX lock, but %u will "
2052 "do the recovery\n", dlm
->name
,
2053 dlm
->reco
.new_master
);
2058 /* see if recovery was already finished elsewhere */
2059 spin_lock(&dlm
->spinlock
);
2060 if (dlm
->reco
.dead_node
== O2NM_INVALID_NODE_NUM
) {
2062 mlog(0, "%s: got reco EX lock, but "
2063 "node got recovered already\n", dlm
->name
);
2064 if (dlm
->reco
.new_master
!= O2NM_INVALID_NODE_NUM
) {
2065 mlog(ML_ERROR
, "%s: new master is %u "
2066 "but no dead node!\n",
2067 dlm
->name
, dlm
->reco
.new_master
);
2071 spin_unlock(&dlm
->spinlock
);
2074 /* if this node has actually become the recovery master,
2075 * set the master and send the messages to begin recovery */
2077 mlog(0, "%s: dead=%u, this=%u, sending "
2078 "begin_reco now\n", dlm
->name
,
2079 dlm
->reco
.dead_node
, dlm
->node_num
);
2080 status
= dlm_send_begin_reco_message(dlm
,
2081 dlm
->reco
.dead_node
);
2082 /* this always succeeds */
2085 /* set the new_master to this node */
2086 spin_lock(&dlm
->spinlock
);
2087 dlm
->reco
.new_master
= dlm
->node_num
;
2088 spin_unlock(&dlm
->spinlock
);
2091 /* recovery lock is a special case. ast will not get fired,
2092 * so just go ahead and unlock it. */
2093 ret
= dlmunlock(dlm
, &lksb
, 0, dlm_reco_unlock_ast
, dlm
);
2094 if (ret
== DLM_DENIED
) {
2095 mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2096 ret
= dlmunlock(dlm
, &lksb
, LKM_CANCEL
, dlm_reco_unlock_ast
, dlm
);
2098 if (ret
!= DLM_NORMAL
) {
2099 /* this would really suck. this could only happen
2100 * if there was a network error during the unlock
2101 * because of node death. this means the unlock
2102 * is actually "done" and the lock structure is
2103 * even freed. we can continue, but only
2104 * because this specific lock name is special. */
2105 mlog(ML_ERROR
, "dlmunlock returned %d\n", ret
);
2107 } else if (ret
== DLM_NOTQUEUED
) {
2108 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2109 dlm
->name
, dlm
->node_num
);
2110 /* another node is master. wait on
2111 * reco.new_master != O2NM_INVALID_NODE_NUM
2112 * for at most one second */
2113 wait_event_timeout(dlm
->dlm_reco_thread_wq
,
2114 dlm_reco_master_ready(dlm
),
2115 msecs_to_jiffies(1000));
2116 if (!dlm_reco_master_ready(dlm
)) {
2117 mlog(0, "%s: reco master taking awhile\n",
2121 /* another node has informed this one that it is reco master */
2122 mlog(0, "%s: reco master %u is ready to recover %u\n",
2123 dlm
->name
, dlm
->reco
.new_master
, dlm
->reco
.dead_node
);
2126 struct dlm_lock_resource
*res
;
2128 /* dlmlock returned something other than NOTQUEUED or NORMAL */
2129 mlog(ML_ERROR
, "%s: got %s from dlmlock($RECOVERY), "
2130 "lksb.status=%s\n", dlm
->name
, dlm_errname(ret
),
2131 dlm_errname(lksb
.status
));
2132 res
= dlm_lookup_lockres(dlm
, DLM_RECOVERY_LOCK_NAME
,
2133 DLM_RECOVERY_LOCK_NAME_LEN
);
2135 dlm_print_one_lock_resource(res
);
2136 dlm_lockres_put(res
);
2138 mlog(ML_ERROR
, "recovery lock not found\n");
2146 static int dlm_send_begin_reco_message(struct dlm_ctxt
*dlm
, u8 dead_node
)
2148 struct dlm_begin_reco br
;
2150 struct dlm_node_iter iter
;
2154 mlog_entry("%u\n", dead_node
);
2156 mlog(0, "dead node is %u\n", dead_node
);
2158 spin_lock(&dlm
->spinlock
);
2159 dlm_node_iter_init(dlm
->domain_map
, &iter
);
2160 spin_unlock(&dlm
->spinlock
);
2162 clear_bit(dead_node
, iter
.node_map
);
2164 memset(&br
, 0, sizeof(br
));
2165 br
.node_idx
= dlm
->node_num
;
2166 br
.dead_node
= dead_node
;
2168 while ((nodenum
= dlm_node_iter_next(&iter
)) >= 0) {
2170 if (nodenum
== dead_node
) {
2171 mlog(0, "not sending begin reco to dead node "
2175 if (nodenum
== dlm
->node_num
) {
2176 mlog(0, "not sending begin reco to self\n");
2181 mlog(0, "attempting to send begin reco msg to %d\n",
2183 ret
= o2net_send_message(DLM_BEGIN_RECO_MSG
, dlm
->key
,
2184 &br
, sizeof(br
), nodenum
, &status
);
2185 /* negative status is handled ok by caller here */
2188 if (dlm_is_host_down(ret
)) {
2189 /* node is down. not involved in recovery
2190 * so just keep going */
2191 mlog(0, "%s: node %u was down when sending "
2192 "begin reco msg (%d)\n", dlm
->name
, nodenum
, ret
);
2196 struct dlm_lock_resource
*res
;
2197 /* this is now a serious problem, possibly ENOMEM
2198 * in the network stack. must retry */
2200 mlog(ML_ERROR
, "begin reco of dlm %s to node %u "
2201 " returned %d\n", dlm
->name
, nodenum
, ret
);
2202 res
= dlm_lookup_lockres(dlm
, DLM_RECOVERY_LOCK_NAME
,
2203 DLM_RECOVERY_LOCK_NAME_LEN
);
2205 dlm_print_one_lock_resource(res
);
2206 dlm_lockres_put(res
);
2208 mlog(ML_ERROR
, "recovery lock not found\n");
2210 /* sleep for a bit in hopes that we can avoid
2220 int dlm_begin_reco_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
2222 struct dlm_ctxt
*dlm
= data
;
2223 struct dlm_begin_reco
*br
= (struct dlm_begin_reco
*)msg
->buf
;
2225 /* ok to return 0, domain has gone away */
2229 mlog(0, "node %u wants to recover node %u\n",
2230 br
->node_idx
, br
->dead_node
);
2232 dlm_fire_domain_eviction_callbacks(dlm
, br
->dead_node
);
2234 spin_lock(&dlm
->spinlock
);
2235 if (dlm
->reco
.new_master
!= O2NM_INVALID_NODE_NUM
) {
2236 if (test_bit(dlm
->reco
.new_master
, dlm
->recovery_map
)) {
2237 mlog(0, "%s: new_master %u died, changing "
2238 "to %u\n", dlm
->name
, dlm
->reco
.new_master
,
2241 mlog(0, "%s: new_master %u NOT DEAD, changing "
2242 "to %u\n", dlm
->name
, dlm
->reco
.new_master
,
2244 /* may not have seen the new master as dead yet */
2247 if (dlm
->reco
.dead_node
!= O2NM_INVALID_NODE_NUM
) {
2248 mlog(ML_NOTICE
, "%s: dead_node previously set to %u, "
2249 "node %u changing it to %u\n", dlm
->name
,
2250 dlm
->reco
.dead_node
, br
->node_idx
, br
->dead_node
);
2252 dlm
->reco
.new_master
= br
->node_idx
;
2253 dlm
->reco
.dead_node
= br
->dead_node
;
2254 if (!test_bit(br
->dead_node
, dlm
->recovery_map
)) {
2255 mlog(0, "recovery master %u sees %u as dead, but this "
2256 "node has not yet. marking %u as dead\n",
2257 br
->node_idx
, br
->dead_node
, br
->dead_node
);
2258 if (!test_bit(br
->dead_node
, dlm
->domain_map
) ||
2259 !test_bit(br
->dead_node
, dlm
->live_nodes_map
))
2260 mlog(0, "%u not in domain/live_nodes map "
2261 "so setting it in reco map manually\n",
2263 /* force the recovery cleanup in __dlm_hb_node_down
2264 * both of these will be cleared in a moment */
2265 set_bit(br
->dead_node
, dlm
->domain_map
);
2266 set_bit(br
->dead_node
, dlm
->live_nodes_map
);
2267 __dlm_hb_node_down(dlm
, br
->dead_node
);
2269 spin_unlock(&dlm
->spinlock
);
2271 dlm_kick_recovery_thread(dlm
);
2276 static int dlm_send_finalize_reco_message(struct dlm_ctxt
*dlm
)
2279 struct dlm_finalize_reco fr
;
2280 struct dlm_node_iter iter
;
2284 mlog(0, "finishing recovery for node %s:%u\n",
2285 dlm
->name
, dlm
->reco
.dead_node
);
2287 spin_lock(&dlm
->spinlock
);
2288 dlm_node_iter_init(dlm
->domain_map
, &iter
);
2289 spin_unlock(&dlm
->spinlock
);
2291 memset(&fr
, 0, sizeof(fr
));
2292 fr
.node_idx
= dlm
->node_num
;
2293 fr
.dead_node
= dlm
->reco
.dead_node
;
2295 while ((nodenum
= dlm_node_iter_next(&iter
)) >= 0) {
2296 if (nodenum
== dlm
->node_num
)
2298 ret
= o2net_send_message(DLM_FINALIZE_RECO_MSG
, dlm
->key
,
2299 &fr
, sizeof(fr
), nodenum
, &status
);
2302 if (dlm_is_host_down(ret
)) {
2303 /* this has no effect on this recovery
2304 * session, so set the status to zero to
2305 * finish out the last recovery */
2306 mlog(ML_ERROR
, "node %u went down after this "
2307 "node finished recovery.\n", nodenum
);
2320 int dlm_finalize_reco_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
2322 struct dlm_ctxt
*dlm
= data
;
2323 struct dlm_finalize_reco
*fr
= (struct dlm_finalize_reco
*)msg
->buf
;
2325 /* ok to return 0, domain has gone away */
2329 mlog(0, "node %u finalizing recovery of node %u\n",
2330 fr
->node_idx
, fr
->dead_node
);
2332 spin_lock(&dlm
->spinlock
);
2334 if (dlm
->reco
.new_master
!= fr
->node_idx
) {
2335 mlog(ML_ERROR
, "node %u sent recovery finalize msg, but node "
2336 "%u is supposed to be the new master, dead=%u\n",
2337 fr
->node_idx
, dlm
->reco
.new_master
, fr
->dead_node
);
2340 if (dlm
->reco
.dead_node
!= fr
->dead_node
) {
2341 mlog(ML_ERROR
, "node %u sent recovery finalize msg for dead "
2342 "node %u, but node %u is supposed to be dead\n",
2343 fr
->node_idx
, fr
->dead_node
, dlm
->reco
.dead_node
);
2347 dlm_finish_local_lockres_recovery(dlm
, fr
->dead_node
, fr
->node_idx
);
2349 spin_unlock(&dlm
->spinlock
);
2351 dlm_reset_recovery(dlm
);
2353 dlm_kick_recovery_thread(dlm
);