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"
45 #include "dlmdomain.h"
49 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
50 #include "cluster/masklog.h"
52 static void dlm_free_pagevec(void **vec, int pages)
55 free_page((unsigned long)vec[pages]);
59 static void **dlm_alloc_pagevec(int pages)
61 void **vec = kmalloc(pages * sizeof(void *), GFP_KERNEL);
67 for (i = 0; i < pages; i++)
68 if (!(vec[i] = (void *)__get_free_page(GFP_KERNEL)))
71 mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n",
72 pages, DLM_HASH_PAGES, (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 spinlock_t dlm_domain_lock = SPIN_LOCK_UNLOCKED;
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_complete_dlm_shutdown(struct dlm_ctxt *dlm)
308 dlm_unregister_domain_handlers(dlm);
309 dlm_complete_thread(dlm);
310 dlm_complete_recovery_thread(dlm);
312 /* We've left the domain. Now we can take ourselves out of the
313 * list and allow the kref stuff to help us free the
315 spin_lock(&dlm_domain_lock);
316 list_del_init(&dlm->list);
317 spin_unlock(&dlm_domain_lock);
319 /* Wake up anyone waiting for us to remove this domain */
320 wake_up(&dlm_domain_events);
323 static void dlm_migrate_all_locks(struct dlm_ctxt *dlm)
326 struct dlm_lock_resource *res;
328 mlog(0, "Migrating locks from domain %s\n", dlm->name);
330 spin_lock(&dlm->spinlock);
331 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
332 while (!hlist_empty(dlm_lockres_hash(dlm, i))) {
333 res = hlist_entry(dlm_lockres_hash(dlm, i)->first,
334 struct dlm_lock_resource, hash_node);
335 /* need reference when manually grabbing lockres */
336 dlm_lockres_get(res);
337 /* this should unhash the lockres
338 * and exit with dlm->spinlock */
339 mlog(0, "purging res=%p\n", res);
340 if (dlm_lockres_is_dirty(dlm, res)) {
341 /* HACK! this should absolutely go.
342 * need to figure out why some empty
343 * lockreses are still marked dirty */
344 mlog(ML_ERROR, "lockres %.*s dirty!\n",
345 res->lockname.len, res->lockname.name);
347 spin_unlock(&dlm->spinlock);
348 dlm_kick_thread(dlm, res);
349 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
350 dlm_lockres_put(res);
353 dlm_purge_lockres(dlm, res);
354 dlm_lockres_put(res);
357 spin_unlock(&dlm->spinlock);
359 mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
362 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
366 spin_lock(&dlm->spinlock);
367 ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
368 spin_unlock(&dlm->spinlock);
373 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
375 /* Yikes, a double spinlock! I need domain_lock for the dlm
376 * state and the dlm spinlock for join state... Sorry! */
378 spin_lock(&dlm_domain_lock);
379 spin_lock(&dlm->spinlock);
381 if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
382 mlog(0, "Node %d is joining, we wait on it.\n",
384 spin_unlock(&dlm->spinlock);
385 spin_unlock(&dlm_domain_lock);
387 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
391 dlm->dlm_state = DLM_CTXT_LEAVING;
392 spin_unlock(&dlm->spinlock);
393 spin_unlock(&dlm_domain_lock);
396 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
400 assert_spin_locked(&dlm->spinlock);
402 mlog(ML_NOTICE, "Nodes in my domain (\"%s\"):\n", dlm->name);
404 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
405 node + 1)) < O2NM_MAX_NODES) {
406 mlog(ML_NOTICE, " node %d\n", node);
410 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data)
412 struct dlm_ctxt *dlm = data;
414 struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
416 mlog_entry("%p %u %p", msg, len, data);
421 node = exit_msg->node_idx;
423 mlog(0, "Node %u leaves domain %s\n", node, dlm->name);
425 spin_lock(&dlm->spinlock);
426 clear_bit(node, dlm->domain_map);
427 __dlm_print_nodes(dlm);
429 /* notify anything attached to the heartbeat events */
430 dlm_hb_event_notify_attached(dlm, node, 0);
432 spin_unlock(&dlm->spinlock);
439 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm,
443 struct dlm_exit_domain leave_msg;
445 mlog(0, "Asking node %u if we can leave the domain %s me = %u\n",
446 node, dlm->name, dlm->node_num);
448 memset(&leave_msg, 0, sizeof(leave_msg));
449 leave_msg.node_idx = dlm->node_num;
451 status = o2net_send_message(DLM_EXIT_DOMAIN_MSG, dlm->key,
452 &leave_msg, sizeof(leave_msg), node,
455 mlog(0, "status return %d from o2net_send_message\n", status);
461 static void dlm_leave_domain(struct dlm_ctxt *dlm)
463 int node, clear_node, status;
465 /* At this point we've migrated away all our locks and won't
466 * accept mastership of new ones. The dlm is responsible for
467 * almost nothing now. We make sure not to confuse any joining
468 * nodes and then commence shutdown procedure. */
470 spin_lock(&dlm->spinlock);
471 /* Clear ourselves from the domain map */
472 clear_bit(dlm->node_num, dlm->domain_map);
473 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
474 0)) < O2NM_MAX_NODES) {
475 /* Drop the dlm spinlock. This is safe wrt the domain_map.
476 * -nodes cannot be added now as the
477 * query_join_handlers knows to respond with OK_NO_MAP
478 * -we catch the right network errors if a node is
479 * removed from the map while we're sending him the
481 spin_unlock(&dlm->spinlock);
485 status = dlm_send_one_domain_exit(dlm, node);
487 status != -ENOPROTOOPT &&
488 status != -ENOTCONN) {
489 mlog(ML_NOTICE, "Error %d sending domain exit message "
490 "to node %d\n", status, node);
492 /* Not sure what to do here but lets sleep for
493 * a bit in case this was a transient
495 msleep(DLM_DOMAIN_BACKOFF_MS);
499 spin_lock(&dlm->spinlock);
500 /* If we're not clearing the node bit then we intend
501 * to loop back around to try again. */
503 clear_bit(node, dlm->domain_map);
505 spin_unlock(&dlm->spinlock);
508 int dlm_joined(struct dlm_ctxt *dlm)
512 spin_lock(&dlm_domain_lock);
514 if (dlm->dlm_state == DLM_CTXT_JOINED)
517 spin_unlock(&dlm_domain_lock);
522 int dlm_shutting_down(struct dlm_ctxt *dlm)
526 spin_lock(&dlm_domain_lock);
528 if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
531 spin_unlock(&dlm_domain_lock);
536 void dlm_unregister_domain(struct dlm_ctxt *dlm)
540 spin_lock(&dlm_domain_lock);
541 BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
542 BUG_ON(!dlm->num_joins);
545 if (!dlm->num_joins) {
546 /* We mark it "in shutdown" now so new register
547 * requests wait until we've completely left the
548 * domain. Don't use DLM_CTXT_LEAVING yet as we still
549 * want new domain joins to communicate with us at
550 * least until we've completed migration of our
552 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
555 spin_unlock(&dlm_domain_lock);
558 mlog(0, "shutting down domain %s\n", dlm->name);
560 /* We changed dlm state, notify the thread */
561 dlm_kick_thread(dlm, NULL);
563 dlm_migrate_all_locks(dlm);
564 dlm_mark_domain_leaving(dlm);
565 dlm_leave_domain(dlm);
566 dlm_complete_dlm_shutdown(dlm);
570 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
572 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data)
574 struct dlm_query_join_request *query;
575 enum dlm_query_join_response response;
576 struct dlm_ctxt *dlm = NULL;
578 query = (struct dlm_query_join_request *) msg->buf;
580 mlog(0, "node %u wants to join domain %s\n", query->node_idx,
584 * If heartbeat doesn't consider the node live, tell it
585 * to back off and try again. This gives heartbeat a chance
588 if (!o2hb_check_node_heartbeating(query->node_idx)) {
589 mlog(0, "node %u is not in our live map yet\n",
592 response = JOIN_DISALLOW;
596 response = JOIN_OK_NO_MAP;
598 spin_lock(&dlm_domain_lock);
599 dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
600 /* Once the dlm ctxt is marked as leaving then we don't want
601 * to be put in someone's domain map.
602 * Also, explicitly disallow joining at certain troublesome
603 * times (ie. during recovery). */
604 if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
605 int bit = query->node_idx;
606 spin_lock(&dlm->spinlock);
608 if (dlm->dlm_state == DLM_CTXT_NEW &&
609 dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
610 /*If this is a brand new context and we
611 * haven't started our join process yet, then
612 * the other node won the race. */
613 response = JOIN_OK_NO_MAP;
614 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
615 /* Disallow parallel joins. */
616 response = JOIN_DISALLOW;
617 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
618 mlog(ML_NOTICE, "node %u trying to join, but recovery "
619 "is ongoing.\n", bit);
620 response = JOIN_DISALLOW;
621 } else if (test_bit(bit, dlm->recovery_map)) {
622 mlog(ML_NOTICE, "node %u trying to join, but it "
623 "still needs recovery.\n", bit);
624 response = JOIN_DISALLOW;
625 } else if (test_bit(bit, dlm->domain_map)) {
626 mlog(ML_NOTICE, "node %u trying to join, but it "
627 "is still in the domain! needs recovery?\n",
629 response = JOIN_DISALLOW;
631 /* Alright we're fully a part of this domain
632 * so we keep some state as to who's joining
633 * and indicate to him that needs to be fixed
636 __dlm_set_joining_node(dlm, query->node_idx);
639 spin_unlock(&dlm->spinlock);
641 spin_unlock(&dlm_domain_lock);
644 mlog(0, "We respond with %u\n", response);
649 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data)
651 struct dlm_assert_joined *assert;
652 struct dlm_ctxt *dlm = NULL;
654 assert = (struct dlm_assert_joined *) msg->buf;
656 mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
659 spin_lock(&dlm_domain_lock);
660 dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
661 /* XXX should we consider no dlm ctxt an error? */
663 spin_lock(&dlm->spinlock);
665 /* Alright, this node has officially joined our
666 * domain. Set him in the map and clean up our
667 * leftover join state. */
668 BUG_ON(dlm->joining_node != assert->node_idx);
669 set_bit(assert->node_idx, dlm->domain_map);
670 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
672 __dlm_print_nodes(dlm);
674 /* notify anything attached to the heartbeat events */
675 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
677 spin_unlock(&dlm->spinlock);
679 spin_unlock(&dlm_domain_lock);
684 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data)
686 struct dlm_cancel_join *cancel;
687 struct dlm_ctxt *dlm = NULL;
689 cancel = (struct dlm_cancel_join *) msg->buf;
691 mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
694 spin_lock(&dlm_domain_lock);
695 dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
698 spin_lock(&dlm->spinlock);
700 /* Yikes, this guy wants to cancel his join. No
701 * problem, we simply cleanup our join state. */
702 BUG_ON(dlm->joining_node != cancel->node_idx);
703 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
705 spin_unlock(&dlm->spinlock);
707 spin_unlock(&dlm_domain_lock);
712 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
716 struct dlm_cancel_join cancel_msg;
718 memset(&cancel_msg, 0, sizeof(cancel_msg));
719 cancel_msg.node_idx = dlm->node_num;
720 cancel_msg.name_len = strlen(dlm->name);
721 memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
723 status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
724 &cancel_msg, sizeof(cancel_msg), node,
735 /* map_size should be in bytes. */
736 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
737 unsigned long *node_map,
738 unsigned int map_size)
743 if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
744 sizeof(unsigned long))) {
746 "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
747 map_size, BITS_TO_LONGS(O2NM_MAX_NODES));
753 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
754 node + 1)) < O2NM_MAX_NODES) {
755 if (node == dlm->node_num)
758 tmpstat = dlm_send_one_join_cancel(dlm, node);
760 mlog(ML_ERROR, "Error return %d cancelling join on "
761 "node %d\n", tmpstat, node);
772 static int dlm_request_join(struct dlm_ctxt *dlm,
774 enum dlm_query_join_response *response)
777 struct dlm_query_join_request join_msg;
779 mlog(0, "querying node %d\n", node);
781 memset(&join_msg, 0, sizeof(join_msg));
782 join_msg.node_idx = dlm->node_num;
783 join_msg.name_len = strlen(dlm->name);
784 memcpy(join_msg.domain, dlm->name, join_msg.name_len);
786 status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
787 sizeof(join_msg), node, &retval);
788 if (status < 0 && status != -ENOPROTOOPT) {
793 /* -ENOPROTOOPT from the net code means the other side isn't
794 listening for our message type -- that's fine, it means
795 his dlm isn't up, so we can consider him a 'yes' but not
796 joined into the domain. */
797 if (status == -ENOPROTOOPT) {
799 *response = JOIN_OK_NO_MAP;
800 } else if (retval == JOIN_DISALLOW ||
802 retval == JOIN_OK_NO_MAP) {
806 mlog(ML_ERROR, "invalid response %d from node %u\n", retval,
810 mlog(0, "status %d, node %d response is %d\n", status, node,
817 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
821 struct dlm_assert_joined assert_msg;
823 mlog(0, "Sending join assert to node %u\n", node);
825 memset(&assert_msg, 0, sizeof(assert_msg));
826 assert_msg.node_idx = dlm->node_num;
827 assert_msg.name_len = strlen(dlm->name);
828 memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
830 status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
831 &assert_msg, sizeof(assert_msg), node,
839 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
840 unsigned long *node_map)
842 int status, node, live;
846 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
847 node + 1)) < O2NM_MAX_NODES) {
848 if (node == dlm->node_num)
852 /* It is very important that this message be
853 * received so we spin until either the node
854 * has died or it gets the message. */
855 status = dlm_send_one_join_assert(dlm, node);
857 spin_lock(&dlm->spinlock);
858 live = test_bit(node, dlm->live_nodes_map);
859 spin_unlock(&dlm->spinlock);
862 mlog(ML_ERROR, "Error return %d asserting "
863 "join on node %d\n", status, node);
865 /* give us some time between errors... */
867 msleep(DLM_DOMAIN_BACKOFF_MS);
869 } while (status && live);
873 struct domain_join_ctxt {
874 unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
875 unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
878 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
879 struct domain_join_ctxt *ctxt,
880 enum dlm_query_join_response response)
884 if (response == JOIN_DISALLOW) {
885 mlog(0, "Latest response of disallow -- should restart\n");
889 spin_lock(&dlm->spinlock);
890 /* For now, we restart the process if the node maps have
892 ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
893 sizeof(dlm->live_nodes_map));
894 spin_unlock(&dlm->spinlock);
897 mlog(0, "Node maps changed -- should restart\n");
902 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
904 int status = 0, tmpstat, node;
905 struct domain_join_ctxt *ctxt;
906 enum dlm_query_join_response response;
908 mlog_entry("%p", dlm);
910 ctxt = kcalloc(1, sizeof(*ctxt), GFP_KERNEL);
917 /* group sem locking should work for us here -- we're already
918 * registered for heartbeat events so filling this should be
919 * atomic wrt getting those handlers called. */
920 o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
922 spin_lock(&dlm->spinlock);
923 memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
925 __dlm_set_joining_node(dlm, dlm->node_num);
927 spin_unlock(&dlm->spinlock);
930 while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
931 node + 1)) < O2NM_MAX_NODES) {
932 if (node == dlm->node_num)
935 status = dlm_request_join(dlm, node, &response);
941 /* Ok, either we got a response or the node doesn't have a
943 if (response == JOIN_OK)
944 set_bit(node, ctxt->yes_resp_map);
946 if (dlm_should_restart_join(dlm, ctxt, response)) {
952 mlog(0, "Yay, done querying nodes!\n");
954 /* Yay, everyone agree's we can join the domain. My domain is
955 * comprised of all nodes who were put in the
956 * yes_resp_map. Copy that into our domain map and send a join
957 * assert message to clean up everyone elses state. */
958 spin_lock(&dlm->spinlock);
959 memcpy(dlm->domain_map, ctxt->yes_resp_map,
960 sizeof(ctxt->yes_resp_map));
961 set_bit(dlm->node_num, dlm->domain_map);
962 spin_unlock(&dlm->spinlock);
964 dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
966 /* Joined state *must* be set before the joining node
967 * information, otherwise the query_join handler may read no
968 * current joiner but a state of NEW and tell joining nodes
969 * we're not in the domain. */
970 spin_lock(&dlm_domain_lock);
971 dlm->dlm_state = DLM_CTXT_JOINED;
973 spin_unlock(&dlm_domain_lock);
976 spin_lock(&dlm->spinlock);
977 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
979 __dlm_print_nodes(dlm);
980 spin_unlock(&dlm->spinlock);
983 /* Do we need to send a cancel message to any nodes? */
985 tmpstat = dlm_send_join_cancels(dlm,
987 sizeof(ctxt->yes_resp_map));
994 mlog(0, "returning %d\n", status);
998 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1000 o2hb_unregister_callback(&dlm->dlm_hb_up);
1001 o2hb_unregister_callback(&dlm->dlm_hb_down);
1002 o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1005 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1009 mlog(0, "registering handlers.\n");
1011 o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1012 dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1013 status = o2hb_register_callback(&dlm->dlm_hb_down);
1017 o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1018 dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1019 status = o2hb_register_callback(&dlm->dlm_hb_up);
1023 status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1024 sizeof(struct dlm_master_request),
1025 dlm_master_request_handler,
1026 dlm, &dlm->dlm_domain_handlers);
1030 status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1031 sizeof(struct dlm_assert_master),
1032 dlm_assert_master_handler,
1033 dlm, &dlm->dlm_domain_handlers);
1037 status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1038 sizeof(struct dlm_create_lock),
1039 dlm_create_lock_handler,
1040 dlm, &dlm->dlm_domain_handlers);
1044 status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1045 DLM_CONVERT_LOCK_MAX_LEN,
1046 dlm_convert_lock_handler,
1047 dlm, &dlm->dlm_domain_handlers);
1051 status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1052 DLM_UNLOCK_LOCK_MAX_LEN,
1053 dlm_unlock_lock_handler,
1054 dlm, &dlm->dlm_domain_handlers);
1058 status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1059 DLM_PROXY_AST_MAX_LEN,
1060 dlm_proxy_ast_handler,
1061 dlm, &dlm->dlm_domain_handlers);
1065 status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1066 sizeof(struct dlm_exit_domain),
1067 dlm_exit_domain_handler,
1068 dlm, &dlm->dlm_domain_handlers);
1072 status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1073 sizeof(struct dlm_migrate_request),
1074 dlm_migrate_request_handler,
1075 dlm, &dlm->dlm_domain_handlers);
1079 status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1080 DLM_MIG_LOCKRES_MAX_LEN,
1081 dlm_mig_lockres_handler,
1082 dlm, &dlm->dlm_domain_handlers);
1086 status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1087 sizeof(struct dlm_master_requery),
1088 dlm_master_requery_handler,
1089 dlm, &dlm->dlm_domain_handlers);
1093 status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1094 sizeof(struct dlm_lock_request),
1095 dlm_request_all_locks_handler,
1096 dlm, &dlm->dlm_domain_handlers);
1100 status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1101 sizeof(struct dlm_reco_data_done),
1102 dlm_reco_data_done_handler,
1103 dlm, &dlm->dlm_domain_handlers);
1107 status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1108 sizeof(struct dlm_begin_reco),
1109 dlm_begin_reco_handler,
1110 dlm, &dlm->dlm_domain_handlers);
1114 status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1115 sizeof(struct dlm_finalize_reco),
1116 dlm_finalize_reco_handler,
1117 dlm, &dlm->dlm_domain_handlers);
1123 dlm_unregister_domain_handlers(dlm);
1128 static int dlm_join_domain(struct dlm_ctxt *dlm)
1134 mlog(0, "Join domain %s\n", dlm->name);
1136 status = dlm_register_domain_handlers(dlm);
1142 status = dlm_launch_thread(dlm);
1148 status = dlm_launch_recovery_thread(dlm);
1155 unsigned int backoff;
1156 status = dlm_try_to_join_domain(dlm);
1158 /* If we're racing another node to the join, then we
1159 * need to back off temporarily and let them
1161 if (status == -EAGAIN) {
1162 if (signal_pending(current)) {
1163 status = -ERESTARTSYS;
1169 * <dale> No, after you!
1171 * <dale> But you first!
1174 backoff = (unsigned int)(jiffies & 0x3);
1175 backoff *= DLM_DOMAIN_BACKOFF_MS;
1176 mlog(0, "backoff %d\n", backoff);
1179 } while (status == -EAGAIN);
1188 wake_up(&dlm_domain_events);
1191 dlm_unregister_domain_handlers(dlm);
1192 dlm_complete_thread(dlm);
1193 dlm_complete_recovery_thread(dlm);
1199 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1203 struct dlm_ctxt *dlm = NULL;
1205 dlm = kcalloc(1, sizeof(*dlm), GFP_KERNEL);
1207 mlog_errno(-ENOMEM);
1211 dlm->name = kmalloc(strlen(domain) + 1, GFP_KERNEL);
1212 if (dlm->name == NULL) {
1213 mlog_errno(-ENOMEM);
1219 dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1220 if (!dlm->lockres_hash) {
1221 mlog_errno(-ENOMEM);
1228 for (i = 0; i < DLM_HASH_BUCKETS; i++)
1229 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1231 strcpy(dlm->name, domain);
1233 dlm->node_num = o2nm_this_node();
1235 spin_lock_init(&dlm->spinlock);
1236 spin_lock_init(&dlm->master_lock);
1237 spin_lock_init(&dlm->ast_lock);
1238 INIT_LIST_HEAD(&dlm->list);
1239 INIT_LIST_HEAD(&dlm->dirty_list);
1240 INIT_LIST_HEAD(&dlm->reco.resources);
1241 INIT_LIST_HEAD(&dlm->reco.received);
1242 INIT_LIST_HEAD(&dlm->reco.node_data);
1243 INIT_LIST_HEAD(&dlm->purge_list);
1244 INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
1245 dlm->reco.state = 0;
1247 INIT_LIST_HEAD(&dlm->pending_asts);
1248 INIT_LIST_HEAD(&dlm->pending_basts);
1250 mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
1251 dlm->recovery_map, &(dlm->recovery_map[0]));
1253 memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
1254 memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
1255 memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
1257 dlm->dlm_thread_task = NULL;
1258 dlm->dlm_reco_thread_task = NULL;
1259 init_waitqueue_head(&dlm->dlm_thread_wq);
1260 init_waitqueue_head(&dlm->dlm_reco_thread_wq);
1261 init_waitqueue_head(&dlm->reco.event);
1262 init_waitqueue_head(&dlm->ast_wq);
1263 init_waitqueue_head(&dlm->migration_wq);
1264 INIT_LIST_HEAD(&dlm->master_list);
1265 INIT_LIST_HEAD(&dlm->mle_hb_events);
1267 dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
1268 init_waitqueue_head(&dlm->dlm_join_events);
1270 dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
1271 dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
1272 atomic_set(&dlm->local_resources, 0);
1273 atomic_set(&dlm->remote_resources, 0);
1274 atomic_set(&dlm->unknown_resources, 0);
1276 spin_lock_init(&dlm->work_lock);
1277 INIT_LIST_HEAD(&dlm->work_list);
1278 INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work, dlm);
1280 kref_init(&dlm->dlm_refs);
1281 dlm->dlm_state = DLM_CTXT_NEW;
1283 INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
1285 mlog(0, "context init: refcount %u\n",
1286 atomic_read(&dlm->dlm_refs.refcount));
1293 * dlm_register_domain: one-time setup per "domain"
1295 struct dlm_ctxt * dlm_register_domain(const char *domain,
1299 struct dlm_ctxt *dlm = NULL;
1300 struct dlm_ctxt *new_ctxt = NULL;
1302 if (strlen(domain) > O2NM_MAX_NAME_LEN) {
1303 ret = -ENAMETOOLONG;
1304 mlog(ML_ERROR, "domain name length too long\n");
1308 if (!o2hb_check_local_node_heartbeating()) {
1309 mlog(ML_ERROR, "the local node has not been configured, or is "
1310 "not heartbeating\n");
1315 mlog(0, "register called for domain \"%s\"\n", domain);
1319 if (signal_pending(current)) {
1325 spin_lock(&dlm_domain_lock);
1327 dlm = __dlm_lookup_domain(domain);
1329 if (dlm->dlm_state != DLM_CTXT_JOINED) {
1330 spin_unlock(&dlm_domain_lock);
1332 mlog(0, "This ctxt is not joined yet!\n");
1333 wait_event_interruptible(dlm_domain_events,
1334 dlm_wait_on_domain_helper(
1342 spin_unlock(&dlm_domain_lock);
1350 spin_unlock(&dlm_domain_lock);
1352 new_ctxt = dlm_alloc_ctxt(domain, key);
1361 /* a little variable switch-a-roo here... */
1365 /* add the new domain */
1366 list_add_tail(&dlm->list, &dlm_domains);
1367 spin_unlock(&dlm_domain_lock);
1369 ret = dlm_join_domain(dlm);
1379 dlm_free_ctxt_mem(new_ctxt);
1386 EXPORT_SYMBOL_GPL(dlm_register_domain);
1388 static LIST_HEAD(dlm_join_handlers);
1390 static void dlm_unregister_net_handlers(void)
1392 o2net_unregister_handler_list(&dlm_join_handlers);
1395 static int dlm_register_net_handlers(void)
1399 status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1400 sizeof(struct dlm_query_join_request),
1401 dlm_query_join_handler,
1402 NULL, &dlm_join_handlers);
1406 status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1407 sizeof(struct dlm_assert_joined),
1408 dlm_assert_joined_handler,
1409 NULL, &dlm_join_handlers);
1413 status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1414 sizeof(struct dlm_cancel_join),
1415 dlm_cancel_join_handler,
1416 NULL, &dlm_join_handlers);
1420 dlm_unregister_net_handlers();
1425 /* Domain eviction callback handling.
1427 * The file system requires notification of node death *before* the
1428 * dlm completes it's recovery work, otherwise it may be able to
1429 * acquire locks on resources requiring recovery. Since the dlm can
1430 * evict a node from it's domain *before* heartbeat fires, a similar
1431 * mechanism is required. */
1433 /* Eviction is not expected to happen often, so a per-domain lock is
1434 * not necessary. Eviction callbacks are allowed to sleep for short
1435 * periods of time. */
1436 static DECLARE_RWSEM(dlm_callback_sem);
1438 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
1441 struct list_head *iter;
1442 struct dlm_eviction_cb *cb;
1444 down_read(&dlm_callback_sem);
1445 list_for_each(iter, &dlm->dlm_eviction_callbacks) {
1446 cb = list_entry(iter, struct dlm_eviction_cb, ec_item);
1448 cb->ec_func(node_num, cb->ec_data);
1450 up_read(&dlm_callback_sem);
1453 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
1454 dlm_eviction_func *f,
1457 INIT_LIST_HEAD(&cb->ec_item);
1461 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
1463 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
1464 struct dlm_eviction_cb *cb)
1466 down_write(&dlm_callback_sem);
1467 list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
1468 up_write(&dlm_callback_sem);
1470 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
1472 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
1474 down_write(&dlm_callback_sem);
1475 list_del_init(&cb->ec_item);
1476 up_write(&dlm_callback_sem);
1478 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
1480 static int __init dlm_init(void)
1484 dlm_print_version();
1486 status = dlm_init_mle_cache();
1490 status = dlm_register_net_handlers();
1492 dlm_destroy_mle_cache();
1499 static void __exit dlm_exit (void)
1501 dlm_unregister_net_handlers();
1502 dlm_destroy_mle_cache();
1505 MODULE_AUTHOR("Oracle");
1506 MODULE_LICENSE("GPL");
1508 module_init(dlm_init);
1509 module_exit(dlm_exit);