1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * defines domain join / leave apis
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.
27 #include <linux/module.h>
28 #include <linux/types.h>
29 #include <linux/slab.h>
30 #include <linux/highmem.h>
31 #include <linux/utsname.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/delay.h>
35 #include <linux/err.h>
37 #include "cluster/heartbeat.h"
38 #include "cluster/nodemanager.h"
39 #include "cluster/tcp.h"
42 #include "dlmcommon.h"
44 #include "dlmdomain.h"
48 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
49 #include "cluster/masklog.h"
51 static void dlm_free_pagevec(void **vec
, int pages
)
54 free_page((unsigned long)vec
[pages
]);
58 static void **dlm_alloc_pagevec(int pages
)
60 void **vec
= kmalloc(pages
* sizeof(void *), GFP_KERNEL
);
66 for (i
= 0; i
< pages
; i
++)
67 if (!(vec
[i
] = (void *)__get_free_page(GFP_KERNEL
)))
70 mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n",
71 pages
, DLM_HASH_PAGES
, (unsigned long)DLM_BUCKETS_PER_PAGE
);
74 dlm_free_pagevec(vec
, i
);
80 * spinlock lock ordering: if multiple locks are needed, obey this ordering:
82 * struct dlm_ctxt->spinlock
83 * struct dlm_lock_resource->spinlock
84 * struct dlm_ctxt->master_lock
85 * struct dlm_ctxt->ast_lock
86 * dlm_master_list_entry->spinlock
91 spinlock_t dlm_domain_lock
= SPIN_LOCK_UNLOCKED
;
92 LIST_HEAD(dlm_domains
);
93 static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events
);
95 #define DLM_DOMAIN_BACKOFF_MS 200
97 static int dlm_query_join_handler(struct o2net_msg
*msg
, u32 len
, void *data
);
98 static int dlm_assert_joined_handler(struct o2net_msg
*msg
, u32 len
, void *data
);
99 static int dlm_cancel_join_handler(struct o2net_msg
*msg
, u32 len
, void *data
);
100 static int dlm_exit_domain_handler(struct o2net_msg
*msg
, u32 len
, void *data
);
102 static void dlm_unregister_domain_handlers(struct dlm_ctxt
*dlm
);
104 void __dlm_unhash_lockres(struct dlm_lock_resource
*lockres
)
106 hlist_del_init(&lockres
->hash_node
);
107 dlm_lockres_put(lockres
);
110 void __dlm_insert_lockres(struct dlm_ctxt
*dlm
,
111 struct dlm_lock_resource
*res
)
113 struct hlist_head
*bucket
;
116 assert_spin_locked(&dlm
->spinlock
);
119 bucket
= dlm_lockres_hash(dlm
, q
->hash
);
121 /* get a reference for our hashtable */
122 dlm_lockres_get(res
);
124 hlist_add_head(&res
->hash_node
, bucket
);
127 struct dlm_lock_resource
* __dlm_lookup_lockres(struct dlm_ctxt
*dlm
,
132 struct hlist_head
*bucket
;
133 struct hlist_node
*list
;
135 mlog_entry("%.*s\n", len
, name
);
137 assert_spin_locked(&dlm
->spinlock
);
139 bucket
= dlm_lockres_hash(dlm
, hash
);
141 hlist_for_each(list
, bucket
) {
142 struct dlm_lock_resource
*res
= hlist_entry(list
,
143 struct dlm_lock_resource
, hash_node
);
144 if (res
->lockname
.name
[0] != name
[0])
146 if (unlikely(res
->lockname
.len
!= len
))
148 if (memcmp(res
->lockname
.name
+ 1, name
+ 1, len
- 1))
150 dlm_lockres_get(res
);
156 struct dlm_lock_resource
* dlm_lookup_lockres(struct dlm_ctxt
*dlm
,
160 struct dlm_lock_resource
*res
;
161 unsigned int hash
= dlm_lockid_hash(name
, len
);
163 spin_lock(&dlm
->spinlock
);
164 res
= __dlm_lookup_lockres(dlm
, name
, len
, hash
);
165 spin_unlock(&dlm
->spinlock
);
169 static struct dlm_ctxt
* __dlm_lookup_domain_full(const char *domain
, int len
)
171 struct dlm_ctxt
*tmp
= NULL
;
172 struct list_head
*iter
;
174 assert_spin_locked(&dlm_domain_lock
);
176 /* tmp->name here is always NULL terminated,
177 * but domain may not be! */
178 list_for_each(iter
, &dlm_domains
) {
179 tmp
= list_entry (iter
, struct dlm_ctxt
, list
);
180 if (strlen(tmp
->name
) == len
&&
181 memcmp(tmp
->name
, domain
, len
)==0)
189 /* For null terminated domain strings ONLY */
190 static struct dlm_ctxt
* __dlm_lookup_domain(const char *domain
)
192 assert_spin_locked(&dlm_domain_lock
);
194 return __dlm_lookup_domain_full(domain
, strlen(domain
));
198 /* returns true on one of two conditions:
199 * 1) the domain does not exist
200 * 2) the domain exists and it's state is "joined" */
201 static int dlm_wait_on_domain_helper(const char *domain
)
204 struct dlm_ctxt
*tmp
= NULL
;
206 spin_lock(&dlm_domain_lock
);
208 tmp
= __dlm_lookup_domain(domain
);
211 else if (tmp
->dlm_state
== DLM_CTXT_JOINED
)
214 spin_unlock(&dlm_domain_lock
);
218 static void dlm_free_ctxt_mem(struct dlm_ctxt
*dlm
)
220 if (dlm
->lockres_hash
)
221 dlm_free_pagevec((void **)dlm
->lockres_hash
, DLM_HASH_PAGES
);
229 /* A little strange - this function will be called while holding
230 * dlm_domain_lock and is expected to be holding it on the way out. We
231 * will however drop and reacquire it multiple times */
232 static void dlm_ctxt_release(struct kref
*kref
)
234 struct dlm_ctxt
*dlm
;
236 dlm
= container_of(kref
, struct dlm_ctxt
, dlm_refs
);
238 BUG_ON(dlm
->num_joins
);
239 BUG_ON(dlm
->dlm_state
== DLM_CTXT_JOINED
);
241 /* we may still be in the list if we hit an error during join. */
242 list_del_init(&dlm
->list
);
244 spin_unlock(&dlm_domain_lock
);
246 mlog(0, "freeing memory from domain %s\n", dlm
->name
);
248 wake_up(&dlm_domain_events
);
250 dlm_free_ctxt_mem(dlm
);
252 spin_lock(&dlm_domain_lock
);
255 void dlm_put(struct dlm_ctxt
*dlm
)
257 spin_lock(&dlm_domain_lock
);
258 kref_put(&dlm
->dlm_refs
, dlm_ctxt_release
);
259 spin_unlock(&dlm_domain_lock
);
262 static void __dlm_get(struct dlm_ctxt
*dlm
)
264 kref_get(&dlm
->dlm_refs
);
267 /* given a questionable reference to a dlm object, gets a reference if
268 * it can find it in the list, otherwise returns NULL in which case
269 * you shouldn't trust your pointer. */
270 struct dlm_ctxt
*dlm_grab(struct dlm_ctxt
*dlm
)
272 struct list_head
*iter
;
273 struct dlm_ctxt
*target
= NULL
;
275 spin_lock(&dlm_domain_lock
);
277 list_for_each(iter
, &dlm_domains
) {
278 target
= list_entry (iter
, struct dlm_ctxt
, list
);
288 spin_unlock(&dlm_domain_lock
);
293 int dlm_domain_fully_joined(struct dlm_ctxt
*dlm
)
297 spin_lock(&dlm_domain_lock
);
298 ret
= (dlm
->dlm_state
== DLM_CTXT_JOINED
) ||
299 (dlm
->dlm_state
== DLM_CTXT_IN_SHUTDOWN
);
300 spin_unlock(&dlm_domain_lock
);
305 static void dlm_destroy_dlm_worker(struct dlm_ctxt
*dlm
)
307 if (dlm
->dlm_worker
) {
308 flush_workqueue(dlm
->dlm_worker
);
309 destroy_workqueue(dlm
->dlm_worker
);
310 dlm
->dlm_worker
= NULL
;
314 static void dlm_complete_dlm_shutdown(struct dlm_ctxt
*dlm
)
316 dlm_unregister_domain_handlers(dlm
);
317 dlm_complete_thread(dlm
);
318 dlm_complete_recovery_thread(dlm
);
319 dlm_destroy_dlm_worker(dlm
);
321 /* We've left the domain. Now we can take ourselves out of the
322 * list and allow the kref stuff to help us free the
324 spin_lock(&dlm_domain_lock
);
325 list_del_init(&dlm
->list
);
326 spin_unlock(&dlm_domain_lock
);
328 /* Wake up anyone waiting for us to remove this domain */
329 wake_up(&dlm_domain_events
);
332 static void dlm_migrate_all_locks(struct dlm_ctxt
*dlm
)
335 struct dlm_lock_resource
*res
;
337 mlog(0, "Migrating locks from domain %s\n", dlm
->name
);
339 spin_lock(&dlm
->spinlock
);
340 for (i
= 0; i
< DLM_HASH_BUCKETS
; i
++) {
341 while (!hlist_empty(dlm_lockres_hash(dlm
, i
))) {
342 res
= hlist_entry(dlm_lockres_hash(dlm
, i
)->first
,
343 struct dlm_lock_resource
, hash_node
);
344 /* need reference when manually grabbing lockres */
345 dlm_lockres_get(res
);
346 /* this should unhash the lockres
347 * and exit with dlm->spinlock */
348 mlog(0, "purging res=%p\n", res
);
349 if (dlm_lockres_is_dirty(dlm
, res
)) {
350 /* HACK! this should absolutely go.
351 * need to figure out why some empty
352 * lockreses are still marked dirty */
353 mlog(ML_ERROR
, "lockres %.*s dirty!\n",
354 res
->lockname
.len
, res
->lockname
.name
);
356 spin_unlock(&dlm
->spinlock
);
357 dlm_kick_thread(dlm
, res
);
358 wait_event(dlm
->ast_wq
, !dlm_lockres_is_dirty(dlm
, res
));
359 dlm_lockres_put(res
);
362 dlm_purge_lockres(dlm
, res
);
363 dlm_lockres_put(res
);
366 spin_unlock(&dlm
->spinlock
);
368 mlog(0, "DONE Migrating locks from domain %s\n", dlm
->name
);
371 static int dlm_no_joining_node(struct dlm_ctxt
*dlm
)
375 spin_lock(&dlm
->spinlock
);
376 ret
= dlm
->joining_node
== DLM_LOCK_RES_OWNER_UNKNOWN
;
377 spin_unlock(&dlm
->spinlock
);
382 static void dlm_mark_domain_leaving(struct dlm_ctxt
*dlm
)
384 /* Yikes, a double spinlock! I need domain_lock for the dlm
385 * state and the dlm spinlock for join state... Sorry! */
387 spin_lock(&dlm_domain_lock
);
388 spin_lock(&dlm
->spinlock
);
390 if (dlm
->joining_node
!= DLM_LOCK_RES_OWNER_UNKNOWN
) {
391 mlog(0, "Node %d is joining, we wait on it.\n",
393 spin_unlock(&dlm
->spinlock
);
394 spin_unlock(&dlm_domain_lock
);
396 wait_event(dlm
->dlm_join_events
, dlm_no_joining_node(dlm
));
400 dlm
->dlm_state
= DLM_CTXT_LEAVING
;
401 spin_unlock(&dlm
->spinlock
);
402 spin_unlock(&dlm_domain_lock
);
405 static void __dlm_print_nodes(struct dlm_ctxt
*dlm
)
409 assert_spin_locked(&dlm
->spinlock
);
411 mlog(ML_NOTICE
, "Nodes in my domain (\"%s\"):\n", dlm
->name
);
413 while ((node
= find_next_bit(dlm
->domain_map
, O2NM_MAX_NODES
,
414 node
+ 1)) < O2NM_MAX_NODES
) {
415 mlog(ML_NOTICE
, " node %d\n", node
);
419 static int dlm_exit_domain_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
421 struct dlm_ctxt
*dlm
= data
;
423 struct dlm_exit_domain
*exit_msg
= (struct dlm_exit_domain
*) msg
->buf
;
425 mlog_entry("%p %u %p", msg
, len
, data
);
430 node
= exit_msg
->node_idx
;
432 mlog(0, "Node %u leaves domain %s\n", node
, dlm
->name
);
434 spin_lock(&dlm
->spinlock
);
435 clear_bit(node
, dlm
->domain_map
);
436 __dlm_print_nodes(dlm
);
438 /* notify anything attached to the heartbeat events */
439 dlm_hb_event_notify_attached(dlm
, node
, 0);
441 spin_unlock(&dlm
->spinlock
);
448 static int dlm_send_one_domain_exit(struct dlm_ctxt
*dlm
,
452 struct dlm_exit_domain leave_msg
;
454 mlog(0, "Asking node %u if we can leave the domain %s me = %u\n",
455 node
, dlm
->name
, dlm
->node_num
);
457 memset(&leave_msg
, 0, sizeof(leave_msg
));
458 leave_msg
.node_idx
= dlm
->node_num
;
460 status
= o2net_send_message(DLM_EXIT_DOMAIN_MSG
, dlm
->key
,
461 &leave_msg
, sizeof(leave_msg
), node
,
464 mlog(0, "status return %d from o2net_send_message\n", status
);
470 static void dlm_leave_domain(struct dlm_ctxt
*dlm
)
472 int node
, clear_node
, status
;
474 /* At this point we've migrated away all our locks and won't
475 * accept mastership of new ones. The dlm is responsible for
476 * almost nothing now. We make sure not to confuse any joining
477 * nodes and then commence shutdown procedure. */
479 spin_lock(&dlm
->spinlock
);
480 /* Clear ourselves from the domain map */
481 clear_bit(dlm
->node_num
, dlm
->domain_map
);
482 while ((node
= find_next_bit(dlm
->domain_map
, O2NM_MAX_NODES
,
483 0)) < O2NM_MAX_NODES
) {
484 /* Drop the dlm spinlock. This is safe wrt the domain_map.
485 * -nodes cannot be added now as the
486 * query_join_handlers knows to respond with OK_NO_MAP
487 * -we catch the right network errors if a node is
488 * removed from the map while we're sending him the
490 spin_unlock(&dlm
->spinlock
);
494 status
= dlm_send_one_domain_exit(dlm
, node
);
496 status
!= -ENOPROTOOPT
&&
497 status
!= -ENOTCONN
) {
498 mlog(ML_NOTICE
, "Error %d sending domain exit message "
499 "to node %d\n", status
, node
);
501 /* Not sure what to do here but lets sleep for
502 * a bit in case this was a transient
504 msleep(DLM_DOMAIN_BACKOFF_MS
);
508 spin_lock(&dlm
->spinlock
);
509 /* If we're not clearing the node bit then we intend
510 * to loop back around to try again. */
512 clear_bit(node
, dlm
->domain_map
);
514 spin_unlock(&dlm
->spinlock
);
517 int dlm_joined(struct dlm_ctxt
*dlm
)
521 spin_lock(&dlm_domain_lock
);
523 if (dlm
->dlm_state
== DLM_CTXT_JOINED
)
526 spin_unlock(&dlm_domain_lock
);
531 int dlm_shutting_down(struct dlm_ctxt
*dlm
)
535 spin_lock(&dlm_domain_lock
);
537 if (dlm
->dlm_state
== DLM_CTXT_IN_SHUTDOWN
)
540 spin_unlock(&dlm_domain_lock
);
545 void dlm_unregister_domain(struct dlm_ctxt
*dlm
)
549 spin_lock(&dlm_domain_lock
);
550 BUG_ON(dlm
->dlm_state
!= DLM_CTXT_JOINED
);
551 BUG_ON(!dlm
->num_joins
);
554 if (!dlm
->num_joins
) {
555 /* We mark it "in shutdown" now so new register
556 * requests wait until we've completely left the
557 * domain. Don't use DLM_CTXT_LEAVING yet as we still
558 * want new domain joins to communicate with us at
559 * least until we've completed migration of our
561 dlm
->dlm_state
= DLM_CTXT_IN_SHUTDOWN
;
564 spin_unlock(&dlm_domain_lock
);
567 mlog(0, "shutting down domain %s\n", dlm
->name
);
569 /* We changed dlm state, notify the thread */
570 dlm_kick_thread(dlm
, NULL
);
572 dlm_migrate_all_locks(dlm
);
573 dlm_mark_domain_leaving(dlm
);
574 dlm_leave_domain(dlm
);
575 dlm_complete_dlm_shutdown(dlm
);
579 EXPORT_SYMBOL_GPL(dlm_unregister_domain
);
581 static int dlm_query_join_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
583 struct dlm_query_join_request
*query
;
584 enum dlm_query_join_response response
;
585 struct dlm_ctxt
*dlm
= NULL
;
587 query
= (struct dlm_query_join_request
*) msg
->buf
;
589 mlog(0, "node %u wants to join domain %s\n", query
->node_idx
,
593 * If heartbeat doesn't consider the node live, tell it
594 * to back off and try again. This gives heartbeat a chance
597 if (!o2hb_check_node_heartbeating(query
->node_idx
)) {
598 mlog(0, "node %u is not in our live map yet\n",
601 response
= JOIN_DISALLOW
;
605 response
= JOIN_OK_NO_MAP
;
607 spin_lock(&dlm_domain_lock
);
608 dlm
= __dlm_lookup_domain_full(query
->domain
, query
->name_len
);
609 /* Once the dlm ctxt is marked as leaving then we don't want
610 * to be put in someone's domain map.
611 * Also, explicitly disallow joining at certain troublesome
612 * times (ie. during recovery). */
613 if (dlm
&& dlm
->dlm_state
!= DLM_CTXT_LEAVING
) {
614 int bit
= query
->node_idx
;
615 spin_lock(&dlm
->spinlock
);
617 if (dlm
->dlm_state
== DLM_CTXT_NEW
&&
618 dlm
->joining_node
== DLM_LOCK_RES_OWNER_UNKNOWN
) {
619 /*If this is a brand new context and we
620 * haven't started our join process yet, then
621 * the other node won the race. */
622 response
= JOIN_OK_NO_MAP
;
623 } else if (dlm
->joining_node
!= DLM_LOCK_RES_OWNER_UNKNOWN
) {
624 /* Disallow parallel joins. */
625 response
= JOIN_DISALLOW
;
626 } else if (dlm
->reco
.state
& DLM_RECO_STATE_ACTIVE
) {
627 mlog(ML_NOTICE
, "node %u trying to join, but recovery "
628 "is ongoing.\n", bit
);
629 response
= JOIN_DISALLOW
;
630 } else if (test_bit(bit
, dlm
->recovery_map
)) {
631 mlog(ML_NOTICE
, "node %u trying to join, but it "
632 "still needs recovery.\n", bit
);
633 response
= JOIN_DISALLOW
;
634 } else if (test_bit(bit
, dlm
->domain_map
)) {
635 mlog(ML_NOTICE
, "node %u trying to join, but it "
636 "is still in the domain! needs recovery?\n",
638 response
= JOIN_DISALLOW
;
640 /* Alright we're fully a part of this domain
641 * so we keep some state as to who's joining
642 * and indicate to him that needs to be fixed
645 __dlm_set_joining_node(dlm
, query
->node_idx
);
648 spin_unlock(&dlm
->spinlock
);
650 spin_unlock(&dlm_domain_lock
);
653 mlog(0, "We respond with %u\n", response
);
658 static int dlm_assert_joined_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
660 struct dlm_assert_joined
*assert;
661 struct dlm_ctxt
*dlm
= NULL
;
663 assert = (struct dlm_assert_joined
*) msg
->buf
;
665 mlog(0, "node %u asserts join on domain %s\n", assert->node_idx
,
668 spin_lock(&dlm_domain_lock
);
669 dlm
= __dlm_lookup_domain_full(assert->domain
, assert->name_len
);
670 /* XXX should we consider no dlm ctxt an error? */
672 spin_lock(&dlm
->spinlock
);
674 /* Alright, this node has officially joined our
675 * domain. Set him in the map and clean up our
676 * leftover join state. */
677 BUG_ON(dlm
->joining_node
!= assert->node_idx
);
678 set_bit(assert->node_idx
, dlm
->domain_map
);
679 __dlm_set_joining_node(dlm
, DLM_LOCK_RES_OWNER_UNKNOWN
);
681 __dlm_print_nodes(dlm
);
683 /* notify anything attached to the heartbeat events */
684 dlm_hb_event_notify_attached(dlm
, assert->node_idx
, 1);
686 spin_unlock(&dlm
->spinlock
);
688 spin_unlock(&dlm_domain_lock
);
693 static int dlm_cancel_join_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
695 struct dlm_cancel_join
*cancel
;
696 struct dlm_ctxt
*dlm
= NULL
;
698 cancel
= (struct dlm_cancel_join
*) msg
->buf
;
700 mlog(0, "node %u cancels join on domain %s\n", cancel
->node_idx
,
703 spin_lock(&dlm_domain_lock
);
704 dlm
= __dlm_lookup_domain_full(cancel
->domain
, cancel
->name_len
);
707 spin_lock(&dlm
->spinlock
);
709 /* Yikes, this guy wants to cancel his join. No
710 * problem, we simply cleanup our join state. */
711 BUG_ON(dlm
->joining_node
!= cancel
->node_idx
);
712 __dlm_set_joining_node(dlm
, DLM_LOCK_RES_OWNER_UNKNOWN
);
714 spin_unlock(&dlm
->spinlock
);
716 spin_unlock(&dlm_domain_lock
);
721 static int dlm_send_one_join_cancel(struct dlm_ctxt
*dlm
,
725 struct dlm_cancel_join cancel_msg
;
727 memset(&cancel_msg
, 0, sizeof(cancel_msg
));
728 cancel_msg
.node_idx
= dlm
->node_num
;
729 cancel_msg
.name_len
= strlen(dlm
->name
);
730 memcpy(cancel_msg
.domain
, dlm
->name
, cancel_msg
.name_len
);
732 status
= o2net_send_message(DLM_CANCEL_JOIN_MSG
, DLM_MOD_KEY
,
733 &cancel_msg
, sizeof(cancel_msg
), node
,
744 /* map_size should be in bytes. */
745 static int dlm_send_join_cancels(struct dlm_ctxt
*dlm
,
746 unsigned long *node_map
,
747 unsigned int map_size
)
752 if (map_size
!= (BITS_TO_LONGS(O2NM_MAX_NODES
) *
753 sizeof(unsigned long))) {
755 "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
756 map_size
, BITS_TO_LONGS(O2NM_MAX_NODES
));
762 while ((node
= find_next_bit(node_map
, O2NM_MAX_NODES
,
763 node
+ 1)) < O2NM_MAX_NODES
) {
764 if (node
== dlm
->node_num
)
767 tmpstat
= dlm_send_one_join_cancel(dlm
, node
);
769 mlog(ML_ERROR
, "Error return %d cancelling join on "
770 "node %d\n", tmpstat
, node
);
781 static int dlm_request_join(struct dlm_ctxt
*dlm
,
783 enum dlm_query_join_response
*response
)
786 struct dlm_query_join_request join_msg
;
788 mlog(0, "querying node %d\n", node
);
790 memset(&join_msg
, 0, sizeof(join_msg
));
791 join_msg
.node_idx
= dlm
->node_num
;
792 join_msg
.name_len
= strlen(dlm
->name
);
793 memcpy(join_msg
.domain
, dlm
->name
, join_msg
.name_len
);
795 status
= o2net_send_message(DLM_QUERY_JOIN_MSG
, DLM_MOD_KEY
, &join_msg
,
796 sizeof(join_msg
), node
, &retval
);
797 if (status
< 0 && status
!= -ENOPROTOOPT
) {
802 /* -ENOPROTOOPT from the net code means the other side isn't
803 listening for our message type -- that's fine, it means
804 his dlm isn't up, so we can consider him a 'yes' but not
805 joined into the domain. */
806 if (status
== -ENOPROTOOPT
) {
808 *response
= JOIN_OK_NO_MAP
;
809 } else if (retval
== JOIN_DISALLOW
||
811 retval
== JOIN_OK_NO_MAP
) {
815 mlog(ML_ERROR
, "invalid response %d from node %u\n", retval
,
819 mlog(0, "status %d, node %d response is %d\n", status
, node
,
826 static int dlm_send_one_join_assert(struct dlm_ctxt
*dlm
,
830 struct dlm_assert_joined assert_msg
;
832 mlog(0, "Sending join assert to node %u\n", node
);
834 memset(&assert_msg
, 0, sizeof(assert_msg
));
835 assert_msg
.node_idx
= dlm
->node_num
;
836 assert_msg
.name_len
= strlen(dlm
->name
);
837 memcpy(assert_msg
.domain
, dlm
->name
, assert_msg
.name_len
);
839 status
= o2net_send_message(DLM_ASSERT_JOINED_MSG
, DLM_MOD_KEY
,
840 &assert_msg
, sizeof(assert_msg
), node
,
848 static void dlm_send_join_asserts(struct dlm_ctxt
*dlm
,
849 unsigned long *node_map
)
851 int status
, node
, live
;
855 while ((node
= find_next_bit(node_map
, O2NM_MAX_NODES
,
856 node
+ 1)) < O2NM_MAX_NODES
) {
857 if (node
== dlm
->node_num
)
861 /* It is very important that this message be
862 * received so we spin until either the node
863 * has died or it gets the message. */
864 status
= dlm_send_one_join_assert(dlm
, node
);
866 spin_lock(&dlm
->spinlock
);
867 live
= test_bit(node
, dlm
->live_nodes_map
);
868 spin_unlock(&dlm
->spinlock
);
871 mlog(ML_ERROR
, "Error return %d asserting "
872 "join on node %d\n", status
, node
);
874 /* give us some time between errors... */
876 msleep(DLM_DOMAIN_BACKOFF_MS
);
878 } while (status
&& live
);
882 struct domain_join_ctxt
{
883 unsigned long live_map
[BITS_TO_LONGS(O2NM_MAX_NODES
)];
884 unsigned long yes_resp_map
[BITS_TO_LONGS(O2NM_MAX_NODES
)];
887 static int dlm_should_restart_join(struct dlm_ctxt
*dlm
,
888 struct domain_join_ctxt
*ctxt
,
889 enum dlm_query_join_response response
)
893 if (response
== JOIN_DISALLOW
) {
894 mlog(0, "Latest response of disallow -- should restart\n");
898 spin_lock(&dlm
->spinlock
);
899 /* For now, we restart the process if the node maps have
901 ret
= memcmp(ctxt
->live_map
, dlm
->live_nodes_map
,
902 sizeof(dlm
->live_nodes_map
));
903 spin_unlock(&dlm
->spinlock
);
906 mlog(0, "Node maps changed -- should restart\n");
911 static int dlm_try_to_join_domain(struct dlm_ctxt
*dlm
)
913 int status
= 0, tmpstat
, node
;
914 struct domain_join_ctxt
*ctxt
;
915 enum dlm_query_join_response response
;
917 mlog_entry("%p", dlm
);
919 ctxt
= kcalloc(1, sizeof(*ctxt
), GFP_KERNEL
);
926 /* group sem locking should work for us here -- we're already
927 * registered for heartbeat events so filling this should be
928 * atomic wrt getting those handlers called. */
929 o2hb_fill_node_map(dlm
->live_nodes_map
, sizeof(dlm
->live_nodes_map
));
931 spin_lock(&dlm
->spinlock
);
932 memcpy(ctxt
->live_map
, dlm
->live_nodes_map
, sizeof(ctxt
->live_map
));
934 __dlm_set_joining_node(dlm
, dlm
->node_num
);
936 spin_unlock(&dlm
->spinlock
);
939 while ((node
= find_next_bit(ctxt
->live_map
, O2NM_MAX_NODES
,
940 node
+ 1)) < O2NM_MAX_NODES
) {
941 if (node
== dlm
->node_num
)
944 status
= dlm_request_join(dlm
, node
, &response
);
950 /* Ok, either we got a response or the node doesn't have a
952 if (response
== JOIN_OK
)
953 set_bit(node
, ctxt
->yes_resp_map
);
955 if (dlm_should_restart_join(dlm
, ctxt
, response
)) {
961 mlog(0, "Yay, done querying nodes!\n");
963 /* Yay, everyone agree's we can join the domain. My domain is
964 * comprised of all nodes who were put in the
965 * yes_resp_map. Copy that into our domain map and send a join
966 * assert message to clean up everyone elses state. */
967 spin_lock(&dlm
->spinlock
);
968 memcpy(dlm
->domain_map
, ctxt
->yes_resp_map
,
969 sizeof(ctxt
->yes_resp_map
));
970 set_bit(dlm
->node_num
, dlm
->domain_map
);
971 spin_unlock(&dlm
->spinlock
);
973 dlm_send_join_asserts(dlm
, ctxt
->yes_resp_map
);
975 /* Joined state *must* be set before the joining node
976 * information, otherwise the query_join handler may read no
977 * current joiner but a state of NEW and tell joining nodes
978 * we're not in the domain. */
979 spin_lock(&dlm_domain_lock
);
980 dlm
->dlm_state
= DLM_CTXT_JOINED
;
982 spin_unlock(&dlm_domain_lock
);
985 spin_lock(&dlm
->spinlock
);
986 __dlm_set_joining_node(dlm
, DLM_LOCK_RES_OWNER_UNKNOWN
);
988 __dlm_print_nodes(dlm
);
989 spin_unlock(&dlm
->spinlock
);
992 /* Do we need to send a cancel message to any nodes? */
994 tmpstat
= dlm_send_join_cancels(dlm
,
996 sizeof(ctxt
->yes_resp_map
));
1003 mlog(0, "returning %d\n", status
);
1007 static void dlm_unregister_domain_handlers(struct dlm_ctxt
*dlm
)
1009 o2hb_unregister_callback(&dlm
->dlm_hb_up
);
1010 o2hb_unregister_callback(&dlm
->dlm_hb_down
);
1011 o2net_unregister_handler_list(&dlm
->dlm_domain_handlers
);
1014 static int dlm_register_domain_handlers(struct dlm_ctxt
*dlm
)
1018 mlog(0, "registering handlers.\n");
1020 o2hb_setup_callback(&dlm
->dlm_hb_down
, O2HB_NODE_DOWN_CB
,
1021 dlm_hb_node_down_cb
, dlm
, DLM_HB_NODE_DOWN_PRI
);
1022 status
= o2hb_register_callback(&dlm
->dlm_hb_down
);
1026 o2hb_setup_callback(&dlm
->dlm_hb_up
, O2HB_NODE_UP_CB
,
1027 dlm_hb_node_up_cb
, dlm
, DLM_HB_NODE_UP_PRI
);
1028 status
= o2hb_register_callback(&dlm
->dlm_hb_up
);
1032 status
= o2net_register_handler(DLM_MASTER_REQUEST_MSG
, dlm
->key
,
1033 sizeof(struct dlm_master_request
),
1034 dlm_master_request_handler
,
1035 dlm
, &dlm
->dlm_domain_handlers
);
1039 status
= o2net_register_handler(DLM_ASSERT_MASTER_MSG
, dlm
->key
,
1040 sizeof(struct dlm_assert_master
),
1041 dlm_assert_master_handler
,
1042 dlm
, &dlm
->dlm_domain_handlers
);
1046 status
= o2net_register_handler(DLM_CREATE_LOCK_MSG
, dlm
->key
,
1047 sizeof(struct dlm_create_lock
),
1048 dlm_create_lock_handler
,
1049 dlm
, &dlm
->dlm_domain_handlers
);
1053 status
= o2net_register_handler(DLM_CONVERT_LOCK_MSG
, dlm
->key
,
1054 DLM_CONVERT_LOCK_MAX_LEN
,
1055 dlm_convert_lock_handler
,
1056 dlm
, &dlm
->dlm_domain_handlers
);
1060 status
= o2net_register_handler(DLM_UNLOCK_LOCK_MSG
, dlm
->key
,
1061 DLM_UNLOCK_LOCK_MAX_LEN
,
1062 dlm_unlock_lock_handler
,
1063 dlm
, &dlm
->dlm_domain_handlers
);
1067 status
= o2net_register_handler(DLM_PROXY_AST_MSG
, dlm
->key
,
1068 DLM_PROXY_AST_MAX_LEN
,
1069 dlm_proxy_ast_handler
,
1070 dlm
, &dlm
->dlm_domain_handlers
);
1074 status
= o2net_register_handler(DLM_EXIT_DOMAIN_MSG
, dlm
->key
,
1075 sizeof(struct dlm_exit_domain
),
1076 dlm_exit_domain_handler
,
1077 dlm
, &dlm
->dlm_domain_handlers
);
1081 status
= o2net_register_handler(DLM_MIGRATE_REQUEST_MSG
, dlm
->key
,
1082 sizeof(struct dlm_migrate_request
),
1083 dlm_migrate_request_handler
,
1084 dlm
, &dlm
->dlm_domain_handlers
);
1088 status
= o2net_register_handler(DLM_MIG_LOCKRES_MSG
, dlm
->key
,
1089 DLM_MIG_LOCKRES_MAX_LEN
,
1090 dlm_mig_lockres_handler
,
1091 dlm
, &dlm
->dlm_domain_handlers
);
1095 status
= o2net_register_handler(DLM_MASTER_REQUERY_MSG
, dlm
->key
,
1096 sizeof(struct dlm_master_requery
),
1097 dlm_master_requery_handler
,
1098 dlm
, &dlm
->dlm_domain_handlers
);
1102 status
= o2net_register_handler(DLM_LOCK_REQUEST_MSG
, dlm
->key
,
1103 sizeof(struct dlm_lock_request
),
1104 dlm_request_all_locks_handler
,
1105 dlm
, &dlm
->dlm_domain_handlers
);
1109 status
= o2net_register_handler(DLM_RECO_DATA_DONE_MSG
, dlm
->key
,
1110 sizeof(struct dlm_reco_data_done
),
1111 dlm_reco_data_done_handler
,
1112 dlm
, &dlm
->dlm_domain_handlers
);
1116 status
= o2net_register_handler(DLM_BEGIN_RECO_MSG
, dlm
->key
,
1117 sizeof(struct dlm_begin_reco
),
1118 dlm_begin_reco_handler
,
1119 dlm
, &dlm
->dlm_domain_handlers
);
1123 status
= o2net_register_handler(DLM_FINALIZE_RECO_MSG
, dlm
->key
,
1124 sizeof(struct dlm_finalize_reco
),
1125 dlm_finalize_reco_handler
,
1126 dlm
, &dlm
->dlm_domain_handlers
);
1132 dlm_unregister_domain_handlers(dlm
);
1137 static int dlm_join_domain(struct dlm_ctxt
*dlm
)
1143 mlog(0, "Join domain %s\n", dlm
->name
);
1145 status
= dlm_register_domain_handlers(dlm
);
1151 status
= dlm_launch_thread(dlm
);
1157 status
= dlm_launch_recovery_thread(dlm
);
1163 dlm
->dlm_worker
= create_singlethread_workqueue("dlm_wq");
1164 if (!dlm
->dlm_worker
) {
1171 unsigned int backoff
;
1172 status
= dlm_try_to_join_domain(dlm
);
1174 /* If we're racing another node to the join, then we
1175 * need to back off temporarily and let them
1177 if (status
== -EAGAIN
) {
1178 if (signal_pending(current
)) {
1179 status
= -ERESTARTSYS
;
1185 * <dale> No, after you!
1187 * <dale> But you first!
1190 backoff
= (unsigned int)(jiffies
& 0x3);
1191 backoff
*= DLM_DOMAIN_BACKOFF_MS
;
1192 mlog(0, "backoff %d\n", backoff
);
1195 } while (status
== -EAGAIN
);
1204 wake_up(&dlm_domain_events
);
1207 dlm_unregister_domain_handlers(dlm
);
1208 dlm_complete_thread(dlm
);
1209 dlm_complete_recovery_thread(dlm
);
1210 dlm_destroy_dlm_worker(dlm
);
1216 static struct dlm_ctxt
*dlm_alloc_ctxt(const char *domain
,
1220 struct dlm_ctxt
*dlm
= NULL
;
1222 dlm
= kcalloc(1, sizeof(*dlm
), GFP_KERNEL
);
1224 mlog_errno(-ENOMEM
);
1228 dlm
->name
= kmalloc(strlen(domain
) + 1, GFP_KERNEL
);
1229 if (dlm
->name
== NULL
) {
1230 mlog_errno(-ENOMEM
);
1236 dlm
->lockres_hash
= (struct hlist_head
**)dlm_alloc_pagevec(DLM_HASH_PAGES
);
1237 if (!dlm
->lockres_hash
) {
1238 mlog_errno(-ENOMEM
);
1245 for (i
= 0; i
< DLM_HASH_BUCKETS
; i
++)
1246 INIT_HLIST_HEAD(dlm_lockres_hash(dlm
, i
));
1248 strcpy(dlm
->name
, domain
);
1250 dlm
->node_num
= o2nm_this_node();
1252 spin_lock_init(&dlm
->spinlock
);
1253 spin_lock_init(&dlm
->master_lock
);
1254 spin_lock_init(&dlm
->ast_lock
);
1255 INIT_LIST_HEAD(&dlm
->list
);
1256 INIT_LIST_HEAD(&dlm
->dirty_list
);
1257 INIT_LIST_HEAD(&dlm
->reco
.resources
);
1258 INIT_LIST_HEAD(&dlm
->reco
.received
);
1259 INIT_LIST_HEAD(&dlm
->reco
.node_data
);
1260 INIT_LIST_HEAD(&dlm
->purge_list
);
1261 INIT_LIST_HEAD(&dlm
->dlm_domain_handlers
);
1262 dlm
->reco
.state
= 0;
1264 INIT_LIST_HEAD(&dlm
->pending_asts
);
1265 INIT_LIST_HEAD(&dlm
->pending_basts
);
1267 mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
1268 dlm
->recovery_map
, &(dlm
->recovery_map
[0]));
1270 memset(dlm
->recovery_map
, 0, sizeof(dlm
->recovery_map
));
1271 memset(dlm
->live_nodes_map
, 0, sizeof(dlm
->live_nodes_map
));
1272 memset(dlm
->domain_map
, 0, sizeof(dlm
->domain_map
));
1274 dlm
->dlm_thread_task
= NULL
;
1275 dlm
->dlm_reco_thread_task
= NULL
;
1276 dlm
->dlm_worker
= NULL
;
1277 init_waitqueue_head(&dlm
->dlm_thread_wq
);
1278 init_waitqueue_head(&dlm
->dlm_reco_thread_wq
);
1279 init_waitqueue_head(&dlm
->reco
.event
);
1280 init_waitqueue_head(&dlm
->ast_wq
);
1281 init_waitqueue_head(&dlm
->migration_wq
);
1282 INIT_LIST_HEAD(&dlm
->master_list
);
1283 INIT_LIST_HEAD(&dlm
->mle_hb_events
);
1285 dlm
->joining_node
= DLM_LOCK_RES_OWNER_UNKNOWN
;
1286 init_waitqueue_head(&dlm
->dlm_join_events
);
1288 dlm
->reco
.new_master
= O2NM_INVALID_NODE_NUM
;
1289 dlm
->reco
.dead_node
= O2NM_INVALID_NODE_NUM
;
1290 atomic_set(&dlm
->local_resources
, 0);
1291 atomic_set(&dlm
->remote_resources
, 0);
1292 atomic_set(&dlm
->unknown_resources
, 0);
1294 spin_lock_init(&dlm
->work_lock
);
1295 INIT_LIST_HEAD(&dlm
->work_list
);
1296 INIT_WORK(&dlm
->dispatched_work
, dlm_dispatch_work
, dlm
);
1298 kref_init(&dlm
->dlm_refs
);
1299 dlm
->dlm_state
= DLM_CTXT_NEW
;
1301 INIT_LIST_HEAD(&dlm
->dlm_eviction_callbacks
);
1303 mlog(0, "context init: refcount %u\n",
1304 atomic_read(&dlm
->dlm_refs
.refcount
));
1311 * dlm_register_domain: one-time setup per "domain"
1313 struct dlm_ctxt
* dlm_register_domain(const char *domain
,
1317 struct dlm_ctxt
*dlm
= NULL
;
1318 struct dlm_ctxt
*new_ctxt
= NULL
;
1320 if (strlen(domain
) > O2NM_MAX_NAME_LEN
) {
1321 ret
= -ENAMETOOLONG
;
1322 mlog(ML_ERROR
, "domain name length too long\n");
1326 if (!o2hb_check_local_node_heartbeating()) {
1327 mlog(ML_ERROR
, "the local node has not been configured, or is "
1328 "not heartbeating\n");
1333 mlog(0, "register called for domain \"%s\"\n", domain
);
1337 if (signal_pending(current
)) {
1343 spin_lock(&dlm_domain_lock
);
1345 dlm
= __dlm_lookup_domain(domain
);
1347 if (dlm
->dlm_state
!= DLM_CTXT_JOINED
) {
1348 spin_unlock(&dlm_domain_lock
);
1350 mlog(0, "This ctxt is not joined yet!\n");
1351 wait_event_interruptible(dlm_domain_events
,
1352 dlm_wait_on_domain_helper(
1360 spin_unlock(&dlm_domain_lock
);
1368 spin_unlock(&dlm_domain_lock
);
1370 new_ctxt
= dlm_alloc_ctxt(domain
, key
);
1379 /* a little variable switch-a-roo here... */
1383 /* add the new domain */
1384 list_add_tail(&dlm
->list
, &dlm_domains
);
1385 spin_unlock(&dlm_domain_lock
);
1387 ret
= dlm_join_domain(dlm
);
1397 dlm_free_ctxt_mem(new_ctxt
);
1404 EXPORT_SYMBOL_GPL(dlm_register_domain
);
1406 static LIST_HEAD(dlm_join_handlers
);
1408 static void dlm_unregister_net_handlers(void)
1410 o2net_unregister_handler_list(&dlm_join_handlers
);
1413 static int dlm_register_net_handlers(void)
1417 status
= o2net_register_handler(DLM_QUERY_JOIN_MSG
, DLM_MOD_KEY
,
1418 sizeof(struct dlm_query_join_request
),
1419 dlm_query_join_handler
,
1420 NULL
, &dlm_join_handlers
);
1424 status
= o2net_register_handler(DLM_ASSERT_JOINED_MSG
, DLM_MOD_KEY
,
1425 sizeof(struct dlm_assert_joined
),
1426 dlm_assert_joined_handler
,
1427 NULL
, &dlm_join_handlers
);
1431 status
= o2net_register_handler(DLM_CANCEL_JOIN_MSG
, DLM_MOD_KEY
,
1432 sizeof(struct dlm_cancel_join
),
1433 dlm_cancel_join_handler
,
1434 NULL
, &dlm_join_handlers
);
1438 dlm_unregister_net_handlers();
1443 /* Domain eviction callback handling.
1445 * The file system requires notification of node death *before* the
1446 * dlm completes it's recovery work, otherwise it may be able to
1447 * acquire locks on resources requiring recovery. Since the dlm can
1448 * evict a node from it's domain *before* heartbeat fires, a similar
1449 * mechanism is required. */
1451 /* Eviction is not expected to happen often, so a per-domain lock is
1452 * not necessary. Eviction callbacks are allowed to sleep for short
1453 * periods of time. */
1454 static DECLARE_RWSEM(dlm_callback_sem
);
1456 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt
*dlm
,
1459 struct list_head
*iter
;
1460 struct dlm_eviction_cb
*cb
;
1462 down_read(&dlm_callback_sem
);
1463 list_for_each(iter
, &dlm
->dlm_eviction_callbacks
) {
1464 cb
= list_entry(iter
, struct dlm_eviction_cb
, ec_item
);
1466 cb
->ec_func(node_num
, cb
->ec_data
);
1468 up_read(&dlm_callback_sem
);
1471 void dlm_setup_eviction_cb(struct dlm_eviction_cb
*cb
,
1472 dlm_eviction_func
*f
,
1475 INIT_LIST_HEAD(&cb
->ec_item
);
1479 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb
);
1481 void dlm_register_eviction_cb(struct dlm_ctxt
*dlm
,
1482 struct dlm_eviction_cb
*cb
)
1484 down_write(&dlm_callback_sem
);
1485 list_add_tail(&cb
->ec_item
, &dlm
->dlm_eviction_callbacks
);
1486 up_write(&dlm_callback_sem
);
1488 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb
);
1490 void dlm_unregister_eviction_cb(struct dlm_eviction_cb
*cb
)
1492 down_write(&dlm_callback_sem
);
1493 list_del_init(&cb
->ec_item
);
1494 up_write(&dlm_callback_sem
);
1496 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb
);
1498 static int __init
dlm_init(void)
1502 dlm_print_version();
1504 status
= dlm_init_mle_cache();
1508 status
= dlm_register_net_handlers();
1510 dlm_destroy_mle_cache();
1517 static void __exit
dlm_exit (void)
1519 dlm_unregister_net_handlers();
1520 dlm_destroy_mle_cache();
1523 MODULE_AUTHOR("Oracle");
1524 MODULE_LICENSE("GPL");
1526 module_init(dlm_init
);
1527 module_exit(dlm_exit
);