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
, (unsigned long)DLM_HASH_PAGES
,
72 (unsigned long)DLM_BUCKETS_PER_PAGE
);
75 dlm_free_pagevec(vec
, i
);
81 * spinlock lock ordering: if multiple locks are needed, obey this ordering:
83 * struct dlm_ctxt->spinlock
84 * struct dlm_lock_resource->spinlock
85 * struct dlm_ctxt->master_lock
86 * struct dlm_ctxt->ast_lock
87 * dlm_master_list_entry->spinlock
92 DEFINE_SPINLOCK(dlm_domain_lock
);
93 LIST_HEAD(dlm_domains
);
94 static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events
);
96 #define DLM_DOMAIN_BACKOFF_MS 200
98 static int dlm_query_join_handler(struct o2net_msg
*msg
, u32 len
, void *data
);
99 static int dlm_assert_joined_handler(struct o2net_msg
*msg
, u32 len
, void *data
);
100 static int dlm_cancel_join_handler(struct o2net_msg
*msg
, u32 len
, void *data
);
101 static int dlm_exit_domain_handler(struct o2net_msg
*msg
, u32 len
, void *data
);
103 static void dlm_unregister_domain_handlers(struct dlm_ctxt
*dlm
);
105 void __dlm_unhash_lockres(struct dlm_lock_resource
*lockres
)
107 hlist_del_init(&lockres
->hash_node
);
108 dlm_lockres_put(lockres
);
111 void __dlm_insert_lockres(struct dlm_ctxt
*dlm
,
112 struct dlm_lock_resource
*res
)
114 struct hlist_head
*bucket
;
117 assert_spin_locked(&dlm
->spinlock
);
120 bucket
= dlm_lockres_hash(dlm
, q
->hash
);
122 /* get a reference for our hashtable */
123 dlm_lockres_get(res
);
125 hlist_add_head(&res
->hash_node
, bucket
);
128 struct dlm_lock_resource
* __dlm_lookup_lockres(struct dlm_ctxt
*dlm
,
133 struct hlist_head
*bucket
;
134 struct hlist_node
*list
;
136 mlog_entry("%.*s\n", len
, name
);
138 assert_spin_locked(&dlm
->spinlock
);
140 bucket
= dlm_lockres_hash(dlm
, hash
);
142 hlist_for_each(list
, bucket
) {
143 struct dlm_lock_resource
*res
= hlist_entry(list
,
144 struct dlm_lock_resource
, hash_node
);
145 if (res
->lockname
.name
[0] != name
[0])
147 if (unlikely(res
->lockname
.len
!= len
))
149 if (memcmp(res
->lockname
.name
+ 1, name
+ 1, len
- 1))
151 dlm_lockres_get(res
);
157 struct dlm_lock_resource
* dlm_lookup_lockres(struct dlm_ctxt
*dlm
,
161 struct dlm_lock_resource
*res
;
162 unsigned int hash
= dlm_lockid_hash(name
, len
);
164 spin_lock(&dlm
->spinlock
);
165 res
= __dlm_lookup_lockres(dlm
, name
, len
, hash
);
166 spin_unlock(&dlm
->spinlock
);
170 static struct dlm_ctxt
* __dlm_lookup_domain_full(const char *domain
, int len
)
172 struct dlm_ctxt
*tmp
= NULL
;
173 struct list_head
*iter
;
175 assert_spin_locked(&dlm_domain_lock
);
177 /* tmp->name here is always NULL terminated,
178 * but domain may not be! */
179 list_for_each(iter
, &dlm_domains
) {
180 tmp
= list_entry (iter
, struct dlm_ctxt
, list
);
181 if (strlen(tmp
->name
) == len
&&
182 memcmp(tmp
->name
, domain
, len
)==0)
190 /* For null terminated domain strings ONLY */
191 static struct dlm_ctxt
* __dlm_lookup_domain(const char *domain
)
193 assert_spin_locked(&dlm_domain_lock
);
195 return __dlm_lookup_domain_full(domain
, strlen(domain
));
199 /* returns true on one of two conditions:
200 * 1) the domain does not exist
201 * 2) the domain exists and it's state is "joined" */
202 static int dlm_wait_on_domain_helper(const char *domain
)
205 struct dlm_ctxt
*tmp
= NULL
;
207 spin_lock(&dlm_domain_lock
);
209 tmp
= __dlm_lookup_domain(domain
);
212 else if (tmp
->dlm_state
== DLM_CTXT_JOINED
)
215 spin_unlock(&dlm_domain_lock
);
219 static void dlm_free_ctxt_mem(struct dlm_ctxt
*dlm
)
221 if (dlm
->lockres_hash
)
222 dlm_free_pagevec((void **)dlm
->lockres_hash
, DLM_HASH_PAGES
);
230 /* A little strange - this function will be called while holding
231 * dlm_domain_lock and is expected to be holding it on the way out. We
232 * will however drop and reacquire it multiple times */
233 static void dlm_ctxt_release(struct kref
*kref
)
235 struct dlm_ctxt
*dlm
;
237 dlm
= container_of(kref
, struct dlm_ctxt
, dlm_refs
);
239 BUG_ON(dlm
->num_joins
);
240 BUG_ON(dlm
->dlm_state
== DLM_CTXT_JOINED
);
242 /* we may still be in the list if we hit an error during join. */
243 list_del_init(&dlm
->list
);
245 spin_unlock(&dlm_domain_lock
);
247 mlog(0, "freeing memory from domain %s\n", dlm
->name
);
249 wake_up(&dlm_domain_events
);
251 dlm_free_ctxt_mem(dlm
);
253 spin_lock(&dlm_domain_lock
);
256 void dlm_put(struct dlm_ctxt
*dlm
)
258 spin_lock(&dlm_domain_lock
);
259 kref_put(&dlm
->dlm_refs
, dlm_ctxt_release
);
260 spin_unlock(&dlm_domain_lock
);
263 static void __dlm_get(struct dlm_ctxt
*dlm
)
265 kref_get(&dlm
->dlm_refs
);
268 /* given a questionable reference to a dlm object, gets a reference if
269 * it can find it in the list, otherwise returns NULL in which case
270 * you shouldn't trust your pointer. */
271 struct dlm_ctxt
*dlm_grab(struct dlm_ctxt
*dlm
)
273 struct list_head
*iter
;
274 struct dlm_ctxt
*target
= NULL
;
276 spin_lock(&dlm_domain_lock
);
278 list_for_each(iter
, &dlm_domains
) {
279 target
= list_entry (iter
, struct dlm_ctxt
, list
);
289 spin_unlock(&dlm_domain_lock
);
294 int dlm_domain_fully_joined(struct dlm_ctxt
*dlm
)
298 spin_lock(&dlm_domain_lock
);
299 ret
= (dlm
->dlm_state
== DLM_CTXT_JOINED
) ||
300 (dlm
->dlm_state
== DLM_CTXT_IN_SHUTDOWN
);
301 spin_unlock(&dlm_domain_lock
);
306 static void dlm_destroy_dlm_worker(struct dlm_ctxt
*dlm
)
308 if (dlm
->dlm_worker
) {
309 flush_workqueue(dlm
->dlm_worker
);
310 destroy_workqueue(dlm
->dlm_worker
);
311 dlm
->dlm_worker
= NULL
;
315 static void dlm_complete_dlm_shutdown(struct dlm_ctxt
*dlm
)
317 dlm_unregister_domain_handlers(dlm
);
318 dlm_complete_thread(dlm
);
319 dlm_complete_recovery_thread(dlm
);
320 dlm_destroy_dlm_worker(dlm
);
322 /* We've left the domain. Now we can take ourselves out of the
323 * list and allow the kref stuff to help us free the
325 spin_lock(&dlm_domain_lock
);
326 list_del_init(&dlm
->list
);
327 spin_unlock(&dlm_domain_lock
);
329 /* Wake up anyone waiting for us to remove this domain */
330 wake_up(&dlm_domain_events
);
333 static void dlm_migrate_all_locks(struct dlm_ctxt
*dlm
)
336 struct dlm_lock_resource
*res
;
338 mlog(0, "Migrating locks from domain %s\n", dlm
->name
);
340 spin_lock(&dlm
->spinlock
);
341 for (i
= 0; i
< DLM_HASH_BUCKETS
; i
++) {
342 while (!hlist_empty(dlm_lockres_hash(dlm
, i
))) {
343 res
= hlist_entry(dlm_lockres_hash(dlm
, i
)->first
,
344 struct dlm_lock_resource
, hash_node
);
345 /* need reference when manually grabbing lockres */
346 dlm_lockres_get(res
);
347 /* this should unhash the lockres
348 * and exit with dlm->spinlock */
349 mlog(0, "purging res=%p\n", res
);
350 if (dlm_lockres_is_dirty(dlm
, res
)) {
351 /* HACK! this should absolutely go.
352 * need to figure out why some empty
353 * lockreses are still marked dirty */
354 mlog(ML_ERROR
, "lockres %.*s dirty!\n",
355 res
->lockname
.len
, res
->lockname
.name
);
357 spin_unlock(&dlm
->spinlock
);
358 dlm_kick_thread(dlm
, res
);
359 wait_event(dlm
->ast_wq
, !dlm_lockres_is_dirty(dlm
, res
));
360 dlm_lockres_put(res
);
363 dlm_purge_lockres(dlm
, res
);
364 dlm_lockres_put(res
);
367 spin_unlock(&dlm
->spinlock
);
369 mlog(0, "DONE Migrating locks from domain %s\n", dlm
->name
);
372 static int dlm_no_joining_node(struct dlm_ctxt
*dlm
)
376 spin_lock(&dlm
->spinlock
);
377 ret
= dlm
->joining_node
== DLM_LOCK_RES_OWNER_UNKNOWN
;
378 spin_unlock(&dlm
->spinlock
);
383 static void dlm_mark_domain_leaving(struct dlm_ctxt
*dlm
)
385 /* Yikes, a double spinlock! I need domain_lock for the dlm
386 * state and the dlm spinlock for join state... Sorry! */
388 spin_lock(&dlm_domain_lock
);
389 spin_lock(&dlm
->spinlock
);
391 if (dlm
->joining_node
!= DLM_LOCK_RES_OWNER_UNKNOWN
) {
392 mlog(0, "Node %d is joining, we wait on it.\n",
394 spin_unlock(&dlm
->spinlock
);
395 spin_unlock(&dlm_domain_lock
);
397 wait_event(dlm
->dlm_join_events
, dlm_no_joining_node(dlm
));
401 dlm
->dlm_state
= DLM_CTXT_LEAVING
;
402 spin_unlock(&dlm
->spinlock
);
403 spin_unlock(&dlm_domain_lock
);
406 static void __dlm_print_nodes(struct dlm_ctxt
*dlm
)
410 assert_spin_locked(&dlm
->spinlock
);
412 printk(KERN_INFO
"ocfs2_dlm: Nodes in domain (\"%s\"): ", dlm
->name
);
414 while ((node
= find_next_bit(dlm
->domain_map
, O2NM_MAX_NODES
,
415 node
+ 1)) < O2NM_MAX_NODES
) {
421 static int dlm_exit_domain_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
423 struct dlm_ctxt
*dlm
= data
;
425 struct dlm_exit_domain
*exit_msg
= (struct dlm_exit_domain
*) msg
->buf
;
427 mlog_entry("%p %u %p", msg
, len
, data
);
432 node
= exit_msg
->node_idx
;
434 printk(KERN_INFO
"ocfs2_dlm: Node %u leaves domain %s\n", node
, dlm
->name
);
436 spin_lock(&dlm
->spinlock
);
437 clear_bit(node
, dlm
->domain_map
);
438 __dlm_print_nodes(dlm
);
440 /* notify anything attached to the heartbeat events */
441 dlm_hb_event_notify_attached(dlm
, node
, 0);
443 spin_unlock(&dlm
->spinlock
);
450 static int dlm_send_one_domain_exit(struct dlm_ctxt
*dlm
,
454 struct dlm_exit_domain leave_msg
;
456 mlog(0, "Asking node %u if we can leave the domain %s me = %u\n",
457 node
, dlm
->name
, dlm
->node_num
);
459 memset(&leave_msg
, 0, sizeof(leave_msg
));
460 leave_msg
.node_idx
= dlm
->node_num
;
462 status
= o2net_send_message(DLM_EXIT_DOMAIN_MSG
, dlm
->key
,
463 &leave_msg
, sizeof(leave_msg
), node
,
466 mlog(0, "status return %d from o2net_send_message\n", status
);
472 static void dlm_leave_domain(struct dlm_ctxt
*dlm
)
474 int node
, clear_node
, status
;
476 /* At this point we've migrated away all our locks and won't
477 * accept mastership of new ones. The dlm is responsible for
478 * almost nothing now. We make sure not to confuse any joining
479 * nodes and then commence shutdown procedure. */
481 spin_lock(&dlm
->spinlock
);
482 /* Clear ourselves from the domain map */
483 clear_bit(dlm
->node_num
, dlm
->domain_map
);
484 while ((node
= find_next_bit(dlm
->domain_map
, O2NM_MAX_NODES
,
485 0)) < O2NM_MAX_NODES
) {
486 /* Drop the dlm spinlock. This is safe wrt the domain_map.
487 * -nodes cannot be added now as the
488 * query_join_handlers knows to respond with OK_NO_MAP
489 * -we catch the right network errors if a node is
490 * removed from the map while we're sending him the
492 spin_unlock(&dlm
->spinlock
);
496 status
= dlm_send_one_domain_exit(dlm
, node
);
498 status
!= -ENOPROTOOPT
&&
499 status
!= -ENOTCONN
) {
500 mlog(ML_NOTICE
, "Error %d sending domain exit message "
501 "to node %d\n", status
, node
);
503 /* Not sure what to do here but lets sleep for
504 * a bit in case this was a transient
506 msleep(DLM_DOMAIN_BACKOFF_MS
);
510 spin_lock(&dlm
->spinlock
);
511 /* If we're not clearing the node bit then we intend
512 * to loop back around to try again. */
514 clear_bit(node
, dlm
->domain_map
);
516 spin_unlock(&dlm
->spinlock
);
519 int dlm_joined(struct dlm_ctxt
*dlm
)
523 spin_lock(&dlm_domain_lock
);
525 if (dlm
->dlm_state
== DLM_CTXT_JOINED
)
528 spin_unlock(&dlm_domain_lock
);
533 int dlm_shutting_down(struct dlm_ctxt
*dlm
)
537 spin_lock(&dlm_domain_lock
);
539 if (dlm
->dlm_state
== DLM_CTXT_IN_SHUTDOWN
)
542 spin_unlock(&dlm_domain_lock
);
547 void dlm_unregister_domain(struct dlm_ctxt
*dlm
)
551 spin_lock(&dlm_domain_lock
);
552 BUG_ON(dlm
->dlm_state
!= DLM_CTXT_JOINED
);
553 BUG_ON(!dlm
->num_joins
);
556 if (!dlm
->num_joins
) {
557 /* We mark it "in shutdown" now so new register
558 * requests wait until we've completely left the
559 * domain. Don't use DLM_CTXT_LEAVING yet as we still
560 * want new domain joins to communicate with us at
561 * least until we've completed migration of our
563 dlm
->dlm_state
= DLM_CTXT_IN_SHUTDOWN
;
566 spin_unlock(&dlm_domain_lock
);
569 mlog(0, "shutting down domain %s\n", dlm
->name
);
571 /* We changed dlm state, notify the thread */
572 dlm_kick_thread(dlm
, NULL
);
574 dlm_migrate_all_locks(dlm
);
575 dlm_mark_domain_leaving(dlm
);
576 dlm_leave_domain(dlm
);
577 dlm_complete_dlm_shutdown(dlm
);
581 EXPORT_SYMBOL_GPL(dlm_unregister_domain
);
583 static int dlm_query_join_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
585 struct dlm_query_join_request
*query
;
586 enum dlm_query_join_response response
;
587 struct dlm_ctxt
*dlm
= NULL
;
589 query
= (struct dlm_query_join_request
*) msg
->buf
;
591 mlog(0, "node %u wants to join domain %s\n", query
->node_idx
,
595 * If heartbeat doesn't consider the node live, tell it
596 * to back off and try again. This gives heartbeat a chance
599 if (!o2hb_check_node_heartbeating(query
->node_idx
)) {
600 mlog(0, "node %u is not in our live map yet\n",
603 response
= JOIN_DISALLOW
;
607 response
= JOIN_OK_NO_MAP
;
609 spin_lock(&dlm_domain_lock
);
610 dlm
= __dlm_lookup_domain_full(query
->domain
, query
->name_len
);
611 /* Once the dlm ctxt is marked as leaving then we don't want
612 * to be put in someone's domain map.
613 * Also, explicitly disallow joining at certain troublesome
614 * times (ie. during recovery). */
615 if (dlm
&& dlm
->dlm_state
!= DLM_CTXT_LEAVING
) {
616 int bit
= query
->node_idx
;
617 spin_lock(&dlm
->spinlock
);
619 if (dlm
->dlm_state
== DLM_CTXT_NEW
&&
620 dlm
->joining_node
== DLM_LOCK_RES_OWNER_UNKNOWN
) {
621 /*If this is a brand new context and we
622 * haven't started our join process yet, then
623 * the other node won the race. */
624 response
= JOIN_OK_NO_MAP
;
625 } else if (dlm
->joining_node
!= DLM_LOCK_RES_OWNER_UNKNOWN
) {
626 /* Disallow parallel joins. */
627 response
= JOIN_DISALLOW
;
628 } else if (dlm
->reco
.state
& DLM_RECO_STATE_ACTIVE
) {
629 mlog(ML_NOTICE
, "node %u trying to join, but recovery "
630 "is ongoing.\n", bit
);
631 response
= JOIN_DISALLOW
;
632 } else if (test_bit(bit
, dlm
->recovery_map
)) {
633 mlog(ML_NOTICE
, "node %u trying to join, but it "
634 "still needs recovery.\n", bit
);
635 response
= JOIN_DISALLOW
;
636 } else if (test_bit(bit
, dlm
->domain_map
)) {
637 mlog(ML_NOTICE
, "node %u trying to join, but it "
638 "is still in the domain! needs recovery?\n",
640 response
= JOIN_DISALLOW
;
642 /* Alright we're fully a part of this domain
643 * so we keep some state as to who's joining
644 * and indicate to him that needs to be fixed
647 __dlm_set_joining_node(dlm
, query
->node_idx
);
650 spin_unlock(&dlm
->spinlock
);
652 spin_unlock(&dlm_domain_lock
);
655 mlog(0, "We respond with %u\n", response
);
660 static int dlm_assert_joined_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
662 struct dlm_assert_joined
*assert;
663 struct dlm_ctxt
*dlm
= NULL
;
665 assert = (struct dlm_assert_joined
*) msg
->buf
;
667 mlog(0, "node %u asserts join on domain %s\n", assert->node_idx
,
670 spin_lock(&dlm_domain_lock
);
671 dlm
= __dlm_lookup_domain_full(assert->domain
, assert->name_len
);
672 /* XXX should we consider no dlm ctxt an error? */
674 spin_lock(&dlm
->spinlock
);
676 /* Alright, this node has officially joined our
677 * domain. Set him in the map and clean up our
678 * leftover join state. */
679 BUG_ON(dlm
->joining_node
!= assert->node_idx
);
680 set_bit(assert->node_idx
, dlm
->domain_map
);
681 __dlm_set_joining_node(dlm
, DLM_LOCK_RES_OWNER_UNKNOWN
);
683 printk(KERN_INFO
"ocfs2_dlm: Node %u joins domain %s\n",
684 assert->node_idx
, dlm
->name
);
685 __dlm_print_nodes(dlm
);
687 /* notify anything attached to the heartbeat events */
688 dlm_hb_event_notify_attached(dlm
, assert->node_idx
, 1);
690 spin_unlock(&dlm
->spinlock
);
692 spin_unlock(&dlm_domain_lock
);
697 static int dlm_cancel_join_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
699 struct dlm_cancel_join
*cancel
;
700 struct dlm_ctxt
*dlm
= NULL
;
702 cancel
= (struct dlm_cancel_join
*) msg
->buf
;
704 mlog(0, "node %u cancels join on domain %s\n", cancel
->node_idx
,
707 spin_lock(&dlm_domain_lock
);
708 dlm
= __dlm_lookup_domain_full(cancel
->domain
, cancel
->name_len
);
711 spin_lock(&dlm
->spinlock
);
713 /* Yikes, this guy wants to cancel his join. No
714 * problem, we simply cleanup our join state. */
715 BUG_ON(dlm
->joining_node
!= cancel
->node_idx
);
716 __dlm_set_joining_node(dlm
, DLM_LOCK_RES_OWNER_UNKNOWN
);
718 spin_unlock(&dlm
->spinlock
);
720 spin_unlock(&dlm_domain_lock
);
725 static int dlm_send_one_join_cancel(struct dlm_ctxt
*dlm
,
729 struct dlm_cancel_join cancel_msg
;
731 memset(&cancel_msg
, 0, sizeof(cancel_msg
));
732 cancel_msg
.node_idx
= dlm
->node_num
;
733 cancel_msg
.name_len
= strlen(dlm
->name
);
734 memcpy(cancel_msg
.domain
, dlm
->name
, cancel_msg
.name_len
);
736 status
= o2net_send_message(DLM_CANCEL_JOIN_MSG
, DLM_MOD_KEY
,
737 &cancel_msg
, sizeof(cancel_msg
), node
,
748 /* map_size should be in bytes. */
749 static int dlm_send_join_cancels(struct dlm_ctxt
*dlm
,
750 unsigned long *node_map
,
751 unsigned int map_size
)
756 if (map_size
!= (BITS_TO_LONGS(O2NM_MAX_NODES
) *
757 sizeof(unsigned long))) {
759 "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
760 map_size
, BITS_TO_LONGS(O2NM_MAX_NODES
));
766 while ((node
= find_next_bit(node_map
, O2NM_MAX_NODES
,
767 node
+ 1)) < O2NM_MAX_NODES
) {
768 if (node
== dlm
->node_num
)
771 tmpstat
= dlm_send_one_join_cancel(dlm
, node
);
773 mlog(ML_ERROR
, "Error return %d cancelling join on "
774 "node %d\n", tmpstat
, node
);
785 static int dlm_request_join(struct dlm_ctxt
*dlm
,
787 enum dlm_query_join_response
*response
)
790 struct dlm_query_join_request join_msg
;
792 mlog(0, "querying node %d\n", node
);
794 memset(&join_msg
, 0, sizeof(join_msg
));
795 join_msg
.node_idx
= dlm
->node_num
;
796 join_msg
.name_len
= strlen(dlm
->name
);
797 memcpy(join_msg
.domain
, dlm
->name
, join_msg
.name_len
);
799 status
= o2net_send_message(DLM_QUERY_JOIN_MSG
, DLM_MOD_KEY
, &join_msg
,
800 sizeof(join_msg
), node
, &retval
);
801 if (status
< 0 && status
!= -ENOPROTOOPT
) {
806 /* -ENOPROTOOPT from the net code means the other side isn't
807 listening for our message type -- that's fine, it means
808 his dlm isn't up, so we can consider him a 'yes' but not
809 joined into the domain. */
810 if (status
== -ENOPROTOOPT
) {
812 *response
= JOIN_OK_NO_MAP
;
813 } else if (retval
== JOIN_DISALLOW
||
815 retval
== JOIN_OK_NO_MAP
) {
819 mlog(ML_ERROR
, "invalid response %d from node %u\n", retval
,
823 mlog(0, "status %d, node %d response is %d\n", status
, node
,
830 static int dlm_send_one_join_assert(struct dlm_ctxt
*dlm
,
834 struct dlm_assert_joined assert_msg
;
836 mlog(0, "Sending join assert to node %u\n", node
);
838 memset(&assert_msg
, 0, sizeof(assert_msg
));
839 assert_msg
.node_idx
= dlm
->node_num
;
840 assert_msg
.name_len
= strlen(dlm
->name
);
841 memcpy(assert_msg
.domain
, dlm
->name
, assert_msg
.name_len
);
843 status
= o2net_send_message(DLM_ASSERT_JOINED_MSG
, DLM_MOD_KEY
,
844 &assert_msg
, sizeof(assert_msg
), node
,
852 static void dlm_send_join_asserts(struct dlm_ctxt
*dlm
,
853 unsigned long *node_map
)
855 int status
, node
, live
;
859 while ((node
= find_next_bit(node_map
, O2NM_MAX_NODES
,
860 node
+ 1)) < O2NM_MAX_NODES
) {
861 if (node
== dlm
->node_num
)
865 /* It is very important that this message be
866 * received so we spin until either the node
867 * has died or it gets the message. */
868 status
= dlm_send_one_join_assert(dlm
, node
);
870 spin_lock(&dlm
->spinlock
);
871 live
= test_bit(node
, dlm
->live_nodes_map
);
872 spin_unlock(&dlm
->spinlock
);
875 mlog(ML_ERROR
, "Error return %d asserting "
876 "join on node %d\n", status
, node
);
878 /* give us some time between errors... */
880 msleep(DLM_DOMAIN_BACKOFF_MS
);
882 } while (status
&& live
);
886 struct domain_join_ctxt
{
887 unsigned long live_map
[BITS_TO_LONGS(O2NM_MAX_NODES
)];
888 unsigned long yes_resp_map
[BITS_TO_LONGS(O2NM_MAX_NODES
)];
891 static int dlm_should_restart_join(struct dlm_ctxt
*dlm
,
892 struct domain_join_ctxt
*ctxt
,
893 enum dlm_query_join_response response
)
897 if (response
== JOIN_DISALLOW
) {
898 mlog(0, "Latest response of disallow -- should restart\n");
902 spin_lock(&dlm
->spinlock
);
903 /* For now, we restart the process if the node maps have
905 ret
= memcmp(ctxt
->live_map
, dlm
->live_nodes_map
,
906 sizeof(dlm
->live_nodes_map
));
907 spin_unlock(&dlm
->spinlock
);
910 mlog(0, "Node maps changed -- should restart\n");
915 static int dlm_try_to_join_domain(struct dlm_ctxt
*dlm
)
917 int status
= 0, tmpstat
, node
;
918 struct domain_join_ctxt
*ctxt
;
919 enum dlm_query_join_response response
;
921 mlog_entry("%p", dlm
);
923 ctxt
= kcalloc(1, sizeof(*ctxt
), GFP_KERNEL
);
930 /* group sem locking should work for us here -- we're already
931 * registered for heartbeat events so filling this should be
932 * atomic wrt getting those handlers called. */
933 o2hb_fill_node_map(dlm
->live_nodes_map
, sizeof(dlm
->live_nodes_map
));
935 spin_lock(&dlm
->spinlock
);
936 memcpy(ctxt
->live_map
, dlm
->live_nodes_map
, sizeof(ctxt
->live_map
));
938 __dlm_set_joining_node(dlm
, dlm
->node_num
);
940 spin_unlock(&dlm
->spinlock
);
943 while ((node
= find_next_bit(ctxt
->live_map
, O2NM_MAX_NODES
,
944 node
+ 1)) < O2NM_MAX_NODES
) {
945 if (node
== dlm
->node_num
)
948 status
= dlm_request_join(dlm
, node
, &response
);
954 /* Ok, either we got a response or the node doesn't have a
956 if (response
== JOIN_OK
)
957 set_bit(node
, ctxt
->yes_resp_map
);
959 if (dlm_should_restart_join(dlm
, ctxt
, response
)) {
965 mlog(0, "Yay, done querying nodes!\n");
967 /* Yay, everyone agree's we can join the domain. My domain is
968 * comprised of all nodes who were put in the
969 * yes_resp_map. Copy that into our domain map and send a join
970 * assert message to clean up everyone elses state. */
971 spin_lock(&dlm
->spinlock
);
972 memcpy(dlm
->domain_map
, ctxt
->yes_resp_map
,
973 sizeof(ctxt
->yes_resp_map
));
974 set_bit(dlm
->node_num
, dlm
->domain_map
);
975 spin_unlock(&dlm
->spinlock
);
977 dlm_send_join_asserts(dlm
, ctxt
->yes_resp_map
);
979 /* Joined state *must* be set before the joining node
980 * information, otherwise the query_join handler may read no
981 * current joiner but a state of NEW and tell joining nodes
982 * we're not in the domain. */
983 spin_lock(&dlm_domain_lock
);
984 dlm
->dlm_state
= DLM_CTXT_JOINED
;
986 spin_unlock(&dlm_domain_lock
);
989 spin_lock(&dlm
->spinlock
);
990 __dlm_set_joining_node(dlm
, DLM_LOCK_RES_OWNER_UNKNOWN
);
992 __dlm_print_nodes(dlm
);
993 spin_unlock(&dlm
->spinlock
);
996 /* Do we need to send a cancel message to any nodes? */
998 tmpstat
= dlm_send_join_cancels(dlm
,
1000 sizeof(ctxt
->yes_resp_map
));
1002 mlog_errno(tmpstat
);
1007 mlog(0, "returning %d\n", status
);
1011 static void dlm_unregister_domain_handlers(struct dlm_ctxt
*dlm
)
1013 o2hb_unregister_callback(&dlm
->dlm_hb_up
);
1014 o2hb_unregister_callback(&dlm
->dlm_hb_down
);
1015 o2net_unregister_handler_list(&dlm
->dlm_domain_handlers
);
1018 static int dlm_register_domain_handlers(struct dlm_ctxt
*dlm
)
1022 mlog(0, "registering handlers.\n");
1024 o2hb_setup_callback(&dlm
->dlm_hb_down
, O2HB_NODE_DOWN_CB
,
1025 dlm_hb_node_down_cb
, dlm
, DLM_HB_NODE_DOWN_PRI
);
1026 status
= o2hb_register_callback(&dlm
->dlm_hb_down
);
1030 o2hb_setup_callback(&dlm
->dlm_hb_up
, O2HB_NODE_UP_CB
,
1031 dlm_hb_node_up_cb
, dlm
, DLM_HB_NODE_UP_PRI
);
1032 status
= o2hb_register_callback(&dlm
->dlm_hb_up
);
1036 status
= o2net_register_handler(DLM_MASTER_REQUEST_MSG
, dlm
->key
,
1037 sizeof(struct dlm_master_request
),
1038 dlm_master_request_handler
,
1039 dlm
, &dlm
->dlm_domain_handlers
);
1043 status
= o2net_register_handler(DLM_ASSERT_MASTER_MSG
, dlm
->key
,
1044 sizeof(struct dlm_assert_master
),
1045 dlm_assert_master_handler
,
1046 dlm
, &dlm
->dlm_domain_handlers
);
1050 status
= o2net_register_handler(DLM_CREATE_LOCK_MSG
, dlm
->key
,
1051 sizeof(struct dlm_create_lock
),
1052 dlm_create_lock_handler
,
1053 dlm
, &dlm
->dlm_domain_handlers
);
1057 status
= o2net_register_handler(DLM_CONVERT_LOCK_MSG
, dlm
->key
,
1058 DLM_CONVERT_LOCK_MAX_LEN
,
1059 dlm_convert_lock_handler
,
1060 dlm
, &dlm
->dlm_domain_handlers
);
1064 status
= o2net_register_handler(DLM_UNLOCK_LOCK_MSG
, dlm
->key
,
1065 DLM_UNLOCK_LOCK_MAX_LEN
,
1066 dlm_unlock_lock_handler
,
1067 dlm
, &dlm
->dlm_domain_handlers
);
1071 status
= o2net_register_handler(DLM_PROXY_AST_MSG
, dlm
->key
,
1072 DLM_PROXY_AST_MAX_LEN
,
1073 dlm_proxy_ast_handler
,
1074 dlm
, &dlm
->dlm_domain_handlers
);
1078 status
= o2net_register_handler(DLM_EXIT_DOMAIN_MSG
, dlm
->key
,
1079 sizeof(struct dlm_exit_domain
),
1080 dlm_exit_domain_handler
,
1081 dlm
, &dlm
->dlm_domain_handlers
);
1085 status
= o2net_register_handler(DLM_MIGRATE_REQUEST_MSG
, dlm
->key
,
1086 sizeof(struct dlm_migrate_request
),
1087 dlm_migrate_request_handler
,
1088 dlm
, &dlm
->dlm_domain_handlers
);
1092 status
= o2net_register_handler(DLM_MIG_LOCKRES_MSG
, dlm
->key
,
1093 DLM_MIG_LOCKRES_MAX_LEN
,
1094 dlm_mig_lockres_handler
,
1095 dlm
, &dlm
->dlm_domain_handlers
);
1099 status
= o2net_register_handler(DLM_MASTER_REQUERY_MSG
, dlm
->key
,
1100 sizeof(struct dlm_master_requery
),
1101 dlm_master_requery_handler
,
1102 dlm
, &dlm
->dlm_domain_handlers
);
1106 status
= o2net_register_handler(DLM_LOCK_REQUEST_MSG
, dlm
->key
,
1107 sizeof(struct dlm_lock_request
),
1108 dlm_request_all_locks_handler
,
1109 dlm
, &dlm
->dlm_domain_handlers
);
1113 status
= o2net_register_handler(DLM_RECO_DATA_DONE_MSG
, dlm
->key
,
1114 sizeof(struct dlm_reco_data_done
),
1115 dlm_reco_data_done_handler
,
1116 dlm
, &dlm
->dlm_domain_handlers
);
1120 status
= o2net_register_handler(DLM_BEGIN_RECO_MSG
, dlm
->key
,
1121 sizeof(struct dlm_begin_reco
),
1122 dlm_begin_reco_handler
,
1123 dlm
, &dlm
->dlm_domain_handlers
);
1127 status
= o2net_register_handler(DLM_FINALIZE_RECO_MSG
, dlm
->key
,
1128 sizeof(struct dlm_finalize_reco
),
1129 dlm_finalize_reco_handler
,
1130 dlm
, &dlm
->dlm_domain_handlers
);
1136 dlm_unregister_domain_handlers(dlm
);
1141 static int dlm_join_domain(struct dlm_ctxt
*dlm
)
1147 mlog(0, "Join domain %s\n", dlm
->name
);
1149 status
= dlm_register_domain_handlers(dlm
);
1155 status
= dlm_launch_thread(dlm
);
1161 status
= dlm_launch_recovery_thread(dlm
);
1167 dlm
->dlm_worker
= create_singlethread_workqueue("dlm_wq");
1168 if (!dlm
->dlm_worker
) {
1175 unsigned int backoff
;
1176 status
= dlm_try_to_join_domain(dlm
);
1178 /* If we're racing another node to the join, then we
1179 * need to back off temporarily and let them
1181 if (status
== -EAGAIN
) {
1182 if (signal_pending(current
)) {
1183 status
= -ERESTARTSYS
;
1189 * <dale> No, after you!
1191 * <dale> But you first!
1194 backoff
= (unsigned int)(jiffies
& 0x3);
1195 backoff
*= DLM_DOMAIN_BACKOFF_MS
;
1196 mlog(0, "backoff %d\n", backoff
);
1199 } while (status
== -EAGAIN
);
1208 wake_up(&dlm_domain_events
);
1211 dlm_unregister_domain_handlers(dlm
);
1212 dlm_complete_thread(dlm
);
1213 dlm_complete_recovery_thread(dlm
);
1214 dlm_destroy_dlm_worker(dlm
);
1220 static struct dlm_ctxt
*dlm_alloc_ctxt(const char *domain
,
1224 struct dlm_ctxt
*dlm
= NULL
;
1226 dlm
= kcalloc(1, sizeof(*dlm
), GFP_KERNEL
);
1228 mlog_errno(-ENOMEM
);
1232 dlm
->name
= kmalloc(strlen(domain
) + 1, GFP_KERNEL
);
1233 if (dlm
->name
== NULL
) {
1234 mlog_errno(-ENOMEM
);
1240 dlm
->lockres_hash
= (struct hlist_head
**)dlm_alloc_pagevec(DLM_HASH_PAGES
);
1241 if (!dlm
->lockres_hash
) {
1242 mlog_errno(-ENOMEM
);
1249 for (i
= 0; i
< DLM_HASH_BUCKETS
; i
++)
1250 INIT_HLIST_HEAD(dlm_lockres_hash(dlm
, i
));
1252 strcpy(dlm
->name
, domain
);
1254 dlm
->node_num
= o2nm_this_node();
1256 spin_lock_init(&dlm
->spinlock
);
1257 spin_lock_init(&dlm
->master_lock
);
1258 spin_lock_init(&dlm
->ast_lock
);
1259 INIT_LIST_HEAD(&dlm
->list
);
1260 INIT_LIST_HEAD(&dlm
->dirty_list
);
1261 INIT_LIST_HEAD(&dlm
->reco
.resources
);
1262 INIT_LIST_HEAD(&dlm
->reco
.received
);
1263 INIT_LIST_HEAD(&dlm
->reco
.node_data
);
1264 INIT_LIST_HEAD(&dlm
->purge_list
);
1265 INIT_LIST_HEAD(&dlm
->dlm_domain_handlers
);
1266 dlm
->reco
.state
= 0;
1268 INIT_LIST_HEAD(&dlm
->pending_asts
);
1269 INIT_LIST_HEAD(&dlm
->pending_basts
);
1271 mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
1272 dlm
->recovery_map
, &(dlm
->recovery_map
[0]));
1274 memset(dlm
->recovery_map
, 0, sizeof(dlm
->recovery_map
));
1275 memset(dlm
->live_nodes_map
, 0, sizeof(dlm
->live_nodes_map
));
1276 memset(dlm
->domain_map
, 0, sizeof(dlm
->domain_map
));
1278 dlm
->dlm_thread_task
= NULL
;
1279 dlm
->dlm_reco_thread_task
= NULL
;
1280 dlm
->dlm_worker
= NULL
;
1281 init_waitqueue_head(&dlm
->dlm_thread_wq
);
1282 init_waitqueue_head(&dlm
->dlm_reco_thread_wq
);
1283 init_waitqueue_head(&dlm
->reco
.event
);
1284 init_waitqueue_head(&dlm
->ast_wq
);
1285 init_waitqueue_head(&dlm
->migration_wq
);
1286 INIT_LIST_HEAD(&dlm
->master_list
);
1287 INIT_LIST_HEAD(&dlm
->mle_hb_events
);
1289 dlm
->joining_node
= DLM_LOCK_RES_OWNER_UNKNOWN
;
1290 init_waitqueue_head(&dlm
->dlm_join_events
);
1292 dlm
->reco
.new_master
= O2NM_INVALID_NODE_NUM
;
1293 dlm
->reco
.dead_node
= O2NM_INVALID_NODE_NUM
;
1294 atomic_set(&dlm
->local_resources
, 0);
1295 atomic_set(&dlm
->remote_resources
, 0);
1296 atomic_set(&dlm
->unknown_resources
, 0);
1298 spin_lock_init(&dlm
->work_lock
);
1299 INIT_LIST_HEAD(&dlm
->work_list
);
1300 INIT_WORK(&dlm
->dispatched_work
, dlm_dispatch_work
);
1302 kref_init(&dlm
->dlm_refs
);
1303 dlm
->dlm_state
= DLM_CTXT_NEW
;
1305 INIT_LIST_HEAD(&dlm
->dlm_eviction_callbacks
);
1307 mlog(0, "context init: refcount %u\n",
1308 atomic_read(&dlm
->dlm_refs
.refcount
));
1315 * dlm_register_domain: one-time setup per "domain"
1317 struct dlm_ctxt
* dlm_register_domain(const char *domain
,
1321 struct dlm_ctxt
*dlm
= NULL
;
1322 struct dlm_ctxt
*new_ctxt
= NULL
;
1324 if (strlen(domain
) > O2NM_MAX_NAME_LEN
) {
1325 ret
= -ENAMETOOLONG
;
1326 mlog(ML_ERROR
, "domain name length too long\n");
1330 if (!o2hb_check_local_node_heartbeating()) {
1331 mlog(ML_ERROR
, "the local node has not been configured, or is "
1332 "not heartbeating\n");
1337 mlog(0, "register called for domain \"%s\"\n", domain
);
1341 if (signal_pending(current
)) {
1347 spin_lock(&dlm_domain_lock
);
1349 dlm
= __dlm_lookup_domain(domain
);
1351 if (dlm
->dlm_state
!= DLM_CTXT_JOINED
) {
1352 spin_unlock(&dlm_domain_lock
);
1354 mlog(0, "This ctxt is not joined yet!\n");
1355 wait_event_interruptible(dlm_domain_events
,
1356 dlm_wait_on_domain_helper(
1364 spin_unlock(&dlm_domain_lock
);
1372 spin_unlock(&dlm_domain_lock
);
1374 new_ctxt
= dlm_alloc_ctxt(domain
, key
);
1383 /* a little variable switch-a-roo here... */
1387 /* add the new domain */
1388 list_add_tail(&dlm
->list
, &dlm_domains
);
1389 spin_unlock(&dlm_domain_lock
);
1391 ret
= dlm_join_domain(dlm
);
1401 dlm_free_ctxt_mem(new_ctxt
);
1408 EXPORT_SYMBOL_GPL(dlm_register_domain
);
1410 static LIST_HEAD(dlm_join_handlers
);
1412 static void dlm_unregister_net_handlers(void)
1414 o2net_unregister_handler_list(&dlm_join_handlers
);
1417 static int dlm_register_net_handlers(void)
1421 status
= o2net_register_handler(DLM_QUERY_JOIN_MSG
, DLM_MOD_KEY
,
1422 sizeof(struct dlm_query_join_request
),
1423 dlm_query_join_handler
,
1424 NULL
, &dlm_join_handlers
);
1428 status
= o2net_register_handler(DLM_ASSERT_JOINED_MSG
, DLM_MOD_KEY
,
1429 sizeof(struct dlm_assert_joined
),
1430 dlm_assert_joined_handler
,
1431 NULL
, &dlm_join_handlers
);
1435 status
= o2net_register_handler(DLM_CANCEL_JOIN_MSG
, DLM_MOD_KEY
,
1436 sizeof(struct dlm_cancel_join
),
1437 dlm_cancel_join_handler
,
1438 NULL
, &dlm_join_handlers
);
1442 dlm_unregister_net_handlers();
1447 /* Domain eviction callback handling.
1449 * The file system requires notification of node death *before* the
1450 * dlm completes it's recovery work, otherwise it may be able to
1451 * acquire locks on resources requiring recovery. Since the dlm can
1452 * evict a node from it's domain *before* heartbeat fires, a similar
1453 * mechanism is required. */
1455 /* Eviction is not expected to happen often, so a per-domain lock is
1456 * not necessary. Eviction callbacks are allowed to sleep for short
1457 * periods of time. */
1458 static DECLARE_RWSEM(dlm_callback_sem
);
1460 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt
*dlm
,
1463 struct list_head
*iter
;
1464 struct dlm_eviction_cb
*cb
;
1466 down_read(&dlm_callback_sem
);
1467 list_for_each(iter
, &dlm
->dlm_eviction_callbacks
) {
1468 cb
= list_entry(iter
, struct dlm_eviction_cb
, ec_item
);
1470 cb
->ec_func(node_num
, cb
->ec_data
);
1472 up_read(&dlm_callback_sem
);
1475 void dlm_setup_eviction_cb(struct dlm_eviction_cb
*cb
,
1476 dlm_eviction_func
*f
,
1479 INIT_LIST_HEAD(&cb
->ec_item
);
1483 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb
);
1485 void dlm_register_eviction_cb(struct dlm_ctxt
*dlm
,
1486 struct dlm_eviction_cb
*cb
)
1488 down_write(&dlm_callback_sem
);
1489 list_add_tail(&cb
->ec_item
, &dlm
->dlm_eviction_callbacks
);
1490 up_write(&dlm_callback_sem
);
1492 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb
);
1494 void dlm_unregister_eviction_cb(struct dlm_eviction_cb
*cb
)
1496 down_write(&dlm_callback_sem
);
1497 list_del_init(&cb
->ec_item
);
1498 up_write(&dlm_callback_sem
);
1500 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb
);
1502 static int __init
dlm_init(void)
1506 dlm_print_version();
1508 status
= dlm_init_mle_cache();
1512 status
= dlm_register_net_handlers();
1514 dlm_destroy_mle_cache();
1521 static void __exit
dlm_exit (void)
1523 dlm_unregister_net_handlers();
1524 dlm_destroy_mle_cache();
1527 MODULE_AUTHOR("Oracle");
1528 MODULE_LICENSE("GPL");
1530 module_init(dlm_init
);
1531 module_exit(dlm_exit
);