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_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 mlog(ML_NOTICE, "Nodes in my domain (\"%s\"):\n", dlm->name);
414 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
415 node + 1)) < O2NM_MAX_NODES) {
416 mlog(ML_NOTICE, " node %d\n", node);
420 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data)
422 struct dlm_ctxt *dlm = data;
424 struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
426 mlog_entry("%p %u %p", msg, len, data);
431 node = exit_msg->node_idx;
433 mlog(0, "Node %u leaves domain %s\n", node, dlm->name);
435 spin_lock(&dlm->spinlock);
436 clear_bit(node, dlm->domain_map);
437 __dlm_print_nodes(dlm);
439 /* notify anything attached to the heartbeat events */
440 dlm_hb_event_notify_attached(dlm, node, 0);
442 spin_unlock(&dlm->spinlock);
449 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm,
453 struct dlm_exit_domain leave_msg;
455 mlog(0, "Asking node %u if we can leave the domain %s me = %u\n",
456 node, dlm->name, dlm->node_num);
458 memset(&leave_msg, 0, sizeof(leave_msg));
459 leave_msg.node_idx = dlm->node_num;
461 status = o2net_send_message(DLM_EXIT_DOMAIN_MSG, dlm->key,
462 &leave_msg, sizeof(leave_msg), node,
465 mlog(0, "status return %d from o2net_send_message\n", status);
471 static void dlm_leave_domain(struct dlm_ctxt *dlm)
473 int node, clear_node, status;
475 /* At this point we've migrated away all our locks and won't
476 * accept mastership of new ones. The dlm is responsible for
477 * almost nothing now. We make sure not to confuse any joining
478 * nodes and then commence shutdown procedure. */
480 spin_lock(&dlm->spinlock);
481 /* Clear ourselves from the domain map */
482 clear_bit(dlm->node_num, dlm->domain_map);
483 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
484 0)) < O2NM_MAX_NODES) {
485 /* Drop the dlm spinlock. This is safe wrt the domain_map.
486 * -nodes cannot be added now as the
487 * query_join_handlers knows to respond with OK_NO_MAP
488 * -we catch the right network errors if a node is
489 * removed from the map while we're sending him the
491 spin_unlock(&dlm->spinlock);
495 status = dlm_send_one_domain_exit(dlm, node);
497 status != -ENOPROTOOPT &&
498 status != -ENOTCONN) {
499 mlog(ML_NOTICE, "Error %d sending domain exit message "
500 "to node %d\n", status, node);
502 /* Not sure what to do here but lets sleep for
503 * a bit in case this was a transient
505 msleep(DLM_DOMAIN_BACKOFF_MS);
509 spin_lock(&dlm->spinlock);
510 /* If we're not clearing the node bit then we intend
511 * to loop back around to try again. */
513 clear_bit(node, dlm->domain_map);
515 spin_unlock(&dlm->spinlock);
518 int dlm_joined(struct dlm_ctxt *dlm)
522 spin_lock(&dlm_domain_lock);
524 if (dlm->dlm_state == DLM_CTXT_JOINED)
527 spin_unlock(&dlm_domain_lock);
532 int dlm_shutting_down(struct dlm_ctxt *dlm)
536 spin_lock(&dlm_domain_lock);
538 if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
541 spin_unlock(&dlm_domain_lock);
546 void dlm_unregister_domain(struct dlm_ctxt *dlm)
550 spin_lock(&dlm_domain_lock);
551 BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
552 BUG_ON(!dlm->num_joins);
555 if (!dlm->num_joins) {
556 /* We mark it "in shutdown" now so new register
557 * requests wait until we've completely left the
558 * domain. Don't use DLM_CTXT_LEAVING yet as we still
559 * want new domain joins to communicate with us at
560 * least until we've completed migration of our
562 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
565 spin_unlock(&dlm_domain_lock);
568 mlog(0, "shutting down domain %s\n", dlm->name);
570 /* We changed dlm state, notify the thread */
571 dlm_kick_thread(dlm, NULL);
573 dlm_migrate_all_locks(dlm);
574 dlm_mark_domain_leaving(dlm);
575 dlm_leave_domain(dlm);
576 dlm_complete_dlm_shutdown(dlm);
580 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
582 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data)
584 struct dlm_query_join_request *query;
585 enum dlm_query_join_response response;
586 struct dlm_ctxt *dlm = NULL;
588 query = (struct dlm_query_join_request *) msg->buf;
590 mlog(0, "node %u wants to join domain %s\n", query->node_idx,
594 * If heartbeat doesn't consider the node live, tell it
595 * to back off and try again. This gives heartbeat a chance
598 if (!o2hb_check_node_heartbeating(query->node_idx)) {
599 mlog(0, "node %u is not in our live map yet\n",
602 response = JOIN_DISALLOW;
606 response = JOIN_OK_NO_MAP;
608 spin_lock(&dlm_domain_lock);
609 dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
610 /* Once the dlm ctxt is marked as leaving then we don't want
611 * to be put in someone's domain map.
612 * Also, explicitly disallow joining at certain troublesome
613 * times (ie. during recovery). */
614 if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
615 int bit = query->node_idx;
616 spin_lock(&dlm->spinlock);
618 if (dlm->dlm_state == DLM_CTXT_NEW &&
619 dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
620 /*If this is a brand new context and we
621 * haven't started our join process yet, then
622 * the other node won the race. */
623 response = JOIN_OK_NO_MAP;
624 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
625 /* Disallow parallel joins. */
626 response = JOIN_DISALLOW;
627 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
628 mlog(ML_NOTICE, "node %u trying to join, but recovery "
629 "is ongoing.\n", bit);
630 response = JOIN_DISALLOW;
631 } else if (test_bit(bit, dlm->recovery_map)) {
632 mlog(ML_NOTICE, "node %u trying to join, but it "
633 "still needs recovery.\n", bit);
634 response = JOIN_DISALLOW;
635 } else if (test_bit(bit, dlm->domain_map)) {
636 mlog(ML_NOTICE, "node %u trying to join, but it "
637 "is still in the domain! needs recovery?\n",
639 response = JOIN_DISALLOW;
641 /* Alright we're fully a part of this domain
642 * so we keep some state as to who's joining
643 * and indicate to him that needs to be fixed
646 __dlm_set_joining_node(dlm, query->node_idx);
649 spin_unlock(&dlm->spinlock);
651 spin_unlock(&dlm_domain_lock);
654 mlog(0, "We respond with %u\n", response);
659 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data)
661 struct dlm_assert_joined *assert;
662 struct dlm_ctxt *dlm = NULL;
664 assert = (struct dlm_assert_joined *) msg->buf;
666 mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
669 spin_lock(&dlm_domain_lock);
670 dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
671 /* XXX should we consider no dlm ctxt an error? */
673 spin_lock(&dlm->spinlock);
675 /* Alright, this node has officially joined our
676 * domain. Set him in the map and clean up our
677 * leftover join state. */
678 BUG_ON(dlm->joining_node != assert->node_idx);
679 set_bit(assert->node_idx, dlm->domain_map);
680 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
682 __dlm_print_nodes(dlm);
684 /* notify anything attached to the heartbeat events */
685 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
687 spin_unlock(&dlm->spinlock);
689 spin_unlock(&dlm_domain_lock);
694 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data)
696 struct dlm_cancel_join *cancel;
697 struct dlm_ctxt *dlm = NULL;
699 cancel = (struct dlm_cancel_join *) msg->buf;
701 mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
704 spin_lock(&dlm_domain_lock);
705 dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
708 spin_lock(&dlm->spinlock);
710 /* Yikes, this guy wants to cancel his join. No
711 * problem, we simply cleanup our join state. */
712 BUG_ON(dlm->joining_node != cancel->node_idx);
713 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
715 spin_unlock(&dlm->spinlock);
717 spin_unlock(&dlm_domain_lock);
722 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
726 struct dlm_cancel_join cancel_msg;
728 memset(&cancel_msg, 0, sizeof(cancel_msg));
729 cancel_msg.node_idx = dlm->node_num;
730 cancel_msg.name_len = strlen(dlm->name);
731 memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
733 status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
734 &cancel_msg, sizeof(cancel_msg), node,
745 /* map_size should be in bytes. */
746 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
747 unsigned long *node_map,
748 unsigned int map_size)
753 if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
754 sizeof(unsigned long))) {
756 "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
757 map_size, BITS_TO_LONGS(O2NM_MAX_NODES));
763 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
764 node + 1)) < O2NM_MAX_NODES) {
765 if (node == dlm->node_num)
768 tmpstat = dlm_send_one_join_cancel(dlm, node);
770 mlog(ML_ERROR, "Error return %d cancelling join on "
771 "node %d\n", tmpstat, node);
782 static int dlm_request_join(struct dlm_ctxt *dlm,
784 enum dlm_query_join_response *response)
787 struct dlm_query_join_request join_msg;
789 mlog(0, "querying node %d\n", node);
791 memset(&join_msg, 0, sizeof(join_msg));
792 join_msg.node_idx = dlm->node_num;
793 join_msg.name_len = strlen(dlm->name);
794 memcpy(join_msg.domain, dlm->name, join_msg.name_len);
796 status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
797 sizeof(join_msg), node, &retval);
798 if (status < 0 && status != -ENOPROTOOPT) {
803 /* -ENOPROTOOPT from the net code means the other side isn't
804 listening for our message type -- that's fine, it means
805 his dlm isn't up, so we can consider him a 'yes' but not
806 joined into the domain. */
807 if (status == -ENOPROTOOPT) {
809 *response = JOIN_OK_NO_MAP;
810 } else if (retval == JOIN_DISALLOW ||
812 retval == JOIN_OK_NO_MAP) {
816 mlog(ML_ERROR, "invalid response %d from node %u\n", retval,
820 mlog(0, "status %d, node %d response is %d\n", status, node,
827 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
831 struct dlm_assert_joined assert_msg;
833 mlog(0, "Sending join assert to node %u\n", node);
835 memset(&assert_msg, 0, sizeof(assert_msg));
836 assert_msg.node_idx = dlm->node_num;
837 assert_msg.name_len = strlen(dlm->name);
838 memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
840 status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
841 &assert_msg, sizeof(assert_msg), node,
849 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
850 unsigned long *node_map)
852 int status, node, live;
856 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
857 node + 1)) < O2NM_MAX_NODES) {
858 if (node == dlm->node_num)
862 /* It is very important that this message be
863 * received so we spin until either the node
864 * has died or it gets the message. */
865 status = dlm_send_one_join_assert(dlm, node);
867 spin_lock(&dlm->spinlock);
868 live = test_bit(node, dlm->live_nodes_map);
869 spin_unlock(&dlm->spinlock);
872 mlog(ML_ERROR, "Error return %d asserting "
873 "join on node %d\n", status, node);
875 /* give us some time between errors... */
877 msleep(DLM_DOMAIN_BACKOFF_MS);
879 } while (status && live);
883 struct domain_join_ctxt {
884 unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
885 unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
888 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
889 struct domain_join_ctxt *ctxt,
890 enum dlm_query_join_response response)
894 if (response == JOIN_DISALLOW) {
895 mlog(0, "Latest response of disallow -- should restart\n");
899 spin_lock(&dlm->spinlock);
900 /* For now, we restart the process if the node maps have
902 ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
903 sizeof(dlm->live_nodes_map));
904 spin_unlock(&dlm->spinlock);
907 mlog(0, "Node maps changed -- should restart\n");
912 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
914 int status = 0, tmpstat, node;
915 struct domain_join_ctxt *ctxt;
916 enum dlm_query_join_response response;
918 mlog_entry("%p", dlm);
920 ctxt = kcalloc(1, sizeof(*ctxt), GFP_KERNEL);
927 /* group sem locking should work for us here -- we're already
928 * registered for heartbeat events so filling this should be
929 * atomic wrt getting those handlers called. */
930 o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
932 spin_lock(&dlm->spinlock);
933 memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
935 __dlm_set_joining_node(dlm, dlm->node_num);
937 spin_unlock(&dlm->spinlock);
940 while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
941 node + 1)) < O2NM_MAX_NODES) {
942 if (node == dlm->node_num)
945 status = dlm_request_join(dlm, node, &response);
951 /* Ok, either we got a response or the node doesn't have a
953 if (response == JOIN_OK)
954 set_bit(node, ctxt->yes_resp_map);
956 if (dlm_should_restart_join(dlm, ctxt, response)) {
962 mlog(0, "Yay, done querying nodes!\n");
964 /* Yay, everyone agree's we can join the domain. My domain is
965 * comprised of all nodes who were put in the
966 * yes_resp_map. Copy that into our domain map and send a join
967 * assert message to clean up everyone elses state. */
968 spin_lock(&dlm->spinlock);
969 memcpy(dlm->domain_map, ctxt->yes_resp_map,
970 sizeof(ctxt->yes_resp_map));
971 set_bit(dlm->node_num, dlm->domain_map);
972 spin_unlock(&dlm->spinlock);
974 dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
976 /* Joined state *must* be set before the joining node
977 * information, otherwise the query_join handler may read no
978 * current joiner but a state of NEW and tell joining nodes
979 * we're not in the domain. */
980 spin_lock(&dlm_domain_lock);
981 dlm->dlm_state = DLM_CTXT_JOINED;
983 spin_unlock(&dlm_domain_lock);
986 spin_lock(&dlm->spinlock);
987 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
989 __dlm_print_nodes(dlm);
990 spin_unlock(&dlm->spinlock);
993 /* Do we need to send a cancel message to any nodes? */
995 tmpstat = dlm_send_join_cancels(dlm,
997 sizeof(ctxt->yes_resp_map));
1004 mlog(0, "returning %d\n", status);
1008 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1010 o2hb_unregister_callback(&dlm->dlm_hb_up);
1011 o2hb_unregister_callback(&dlm->dlm_hb_down);
1012 o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1015 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1019 mlog(0, "registering handlers.\n");
1021 o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1022 dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1023 status = o2hb_register_callback(&dlm->dlm_hb_down);
1027 o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1028 dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1029 status = o2hb_register_callback(&dlm->dlm_hb_up);
1033 status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1034 sizeof(struct dlm_master_request),
1035 dlm_master_request_handler,
1036 dlm, &dlm->dlm_domain_handlers);
1040 status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1041 sizeof(struct dlm_assert_master),
1042 dlm_assert_master_handler,
1043 dlm, &dlm->dlm_domain_handlers);
1047 status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1048 sizeof(struct dlm_create_lock),
1049 dlm_create_lock_handler,
1050 dlm, &dlm->dlm_domain_handlers);
1054 status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1055 DLM_CONVERT_LOCK_MAX_LEN,
1056 dlm_convert_lock_handler,
1057 dlm, &dlm->dlm_domain_handlers);
1061 status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1062 DLM_UNLOCK_LOCK_MAX_LEN,
1063 dlm_unlock_lock_handler,
1064 dlm, &dlm->dlm_domain_handlers);
1068 status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1069 DLM_PROXY_AST_MAX_LEN,
1070 dlm_proxy_ast_handler,
1071 dlm, &dlm->dlm_domain_handlers);
1075 status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1076 sizeof(struct dlm_exit_domain),
1077 dlm_exit_domain_handler,
1078 dlm, &dlm->dlm_domain_handlers);
1082 status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1083 sizeof(struct dlm_migrate_request),
1084 dlm_migrate_request_handler,
1085 dlm, &dlm->dlm_domain_handlers);
1089 status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1090 DLM_MIG_LOCKRES_MAX_LEN,
1091 dlm_mig_lockres_handler,
1092 dlm, &dlm->dlm_domain_handlers);
1096 status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1097 sizeof(struct dlm_master_requery),
1098 dlm_master_requery_handler,
1099 dlm, &dlm->dlm_domain_handlers);
1103 status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1104 sizeof(struct dlm_lock_request),
1105 dlm_request_all_locks_handler,
1106 dlm, &dlm->dlm_domain_handlers);
1110 status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1111 sizeof(struct dlm_reco_data_done),
1112 dlm_reco_data_done_handler,
1113 dlm, &dlm->dlm_domain_handlers);
1117 status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1118 sizeof(struct dlm_begin_reco),
1119 dlm_begin_reco_handler,
1120 dlm, &dlm->dlm_domain_handlers);
1124 status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1125 sizeof(struct dlm_finalize_reco),
1126 dlm_finalize_reco_handler,
1127 dlm, &dlm->dlm_domain_handlers);
1133 dlm_unregister_domain_handlers(dlm);
1138 static int dlm_join_domain(struct dlm_ctxt *dlm)
1144 mlog(0, "Join domain %s\n", dlm->name);
1146 status = dlm_register_domain_handlers(dlm);
1152 status = dlm_launch_thread(dlm);
1158 status = dlm_launch_recovery_thread(dlm);
1164 dlm->dlm_worker = create_singlethread_workqueue("dlm_wq");
1165 if (!dlm->dlm_worker) {
1172 unsigned int backoff;
1173 status = dlm_try_to_join_domain(dlm);
1175 /* If we're racing another node to the join, then we
1176 * need to back off temporarily and let them
1178 if (status == -EAGAIN) {
1179 if (signal_pending(current)) {
1180 status = -ERESTARTSYS;
1186 * <dale> No, after you!
1188 * <dale> But you first!
1191 backoff = (unsigned int)(jiffies & 0x3);
1192 backoff *= DLM_DOMAIN_BACKOFF_MS;
1193 mlog(0, "backoff %d\n", backoff);
1196 } while (status == -EAGAIN);
1205 wake_up(&dlm_domain_events);
1208 dlm_unregister_domain_handlers(dlm);
1209 dlm_complete_thread(dlm);
1210 dlm_complete_recovery_thread(dlm);
1211 dlm_destroy_dlm_worker(dlm);
1217 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1221 struct dlm_ctxt *dlm = NULL;
1223 dlm = kcalloc(1, sizeof(*dlm), GFP_KERNEL);
1225 mlog_errno(-ENOMEM);
1229 dlm->name = kmalloc(strlen(domain) + 1, GFP_KERNEL);
1230 if (dlm->name == NULL) {
1231 mlog_errno(-ENOMEM);
1237 dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1238 if (!dlm->lockres_hash) {
1239 mlog_errno(-ENOMEM);
1246 for (i = 0; i < DLM_HASH_BUCKETS; i++)
1247 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1249 strcpy(dlm->name, domain);
1251 dlm->node_num = o2nm_this_node();
1253 spin_lock_init(&dlm->spinlock);
1254 spin_lock_init(&dlm->master_lock);
1255 spin_lock_init(&dlm->ast_lock);
1256 INIT_LIST_HEAD(&dlm->list);
1257 INIT_LIST_HEAD(&dlm->dirty_list);
1258 INIT_LIST_HEAD(&dlm->reco.resources);
1259 INIT_LIST_HEAD(&dlm->reco.received);
1260 INIT_LIST_HEAD(&dlm->reco.node_data);
1261 INIT_LIST_HEAD(&dlm->purge_list);
1262 INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
1263 dlm->reco.state = 0;
1265 INIT_LIST_HEAD(&dlm->pending_asts);
1266 INIT_LIST_HEAD(&dlm->pending_basts);
1268 mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
1269 dlm->recovery_map, &(dlm->recovery_map[0]));
1271 memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
1272 memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
1273 memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
1275 dlm->dlm_thread_task = NULL;
1276 dlm->dlm_reco_thread_task = NULL;
1277 dlm->dlm_worker = NULL;
1278 init_waitqueue_head(&dlm->dlm_thread_wq);
1279 init_waitqueue_head(&dlm->dlm_reco_thread_wq);
1280 init_waitqueue_head(&dlm->reco.event);
1281 init_waitqueue_head(&dlm->ast_wq);
1282 init_waitqueue_head(&dlm->migration_wq);
1283 INIT_LIST_HEAD(&dlm->master_list);
1284 INIT_LIST_HEAD(&dlm->mle_hb_events);
1286 dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
1287 init_waitqueue_head(&dlm->dlm_join_events);
1289 dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
1290 dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
1291 atomic_set(&dlm->local_resources, 0);
1292 atomic_set(&dlm->remote_resources, 0);
1293 atomic_set(&dlm->unknown_resources, 0);
1295 spin_lock_init(&dlm->work_lock);
1296 INIT_LIST_HEAD(&dlm->work_list);
1297 INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work, dlm);
1299 kref_init(&dlm->dlm_refs);
1300 dlm->dlm_state = DLM_CTXT_NEW;
1302 INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
1304 mlog(0, "context init: refcount %u\n",
1305 atomic_read(&dlm->dlm_refs.refcount));
1312 * dlm_register_domain: one-time setup per "domain"
1314 struct dlm_ctxt * dlm_register_domain(const char *domain,
1318 struct dlm_ctxt *dlm = NULL;
1319 struct dlm_ctxt *new_ctxt = NULL;
1321 if (strlen(domain) > O2NM_MAX_NAME_LEN) {
1322 ret = -ENAMETOOLONG;
1323 mlog(ML_ERROR, "domain name length too long\n");
1327 if (!o2hb_check_local_node_heartbeating()) {
1328 mlog(ML_ERROR, "the local node has not been configured, or is "
1329 "not heartbeating\n");
1334 mlog(0, "register called for domain \"%s\"\n", domain);
1338 if (signal_pending(current)) {
1344 spin_lock(&dlm_domain_lock);
1346 dlm = __dlm_lookup_domain(domain);
1348 if (dlm->dlm_state != DLM_CTXT_JOINED) {
1349 spin_unlock(&dlm_domain_lock);
1351 mlog(0, "This ctxt is not joined yet!\n");
1352 wait_event_interruptible(dlm_domain_events,
1353 dlm_wait_on_domain_helper(
1361 spin_unlock(&dlm_domain_lock);
1369 spin_unlock(&dlm_domain_lock);
1371 new_ctxt = dlm_alloc_ctxt(domain, key);
1380 /* a little variable switch-a-roo here... */
1384 /* add the new domain */
1385 list_add_tail(&dlm->list, &dlm_domains);
1386 spin_unlock(&dlm_domain_lock);
1388 ret = dlm_join_domain(dlm);
1398 dlm_free_ctxt_mem(new_ctxt);
1405 EXPORT_SYMBOL_GPL(dlm_register_domain);
1407 static LIST_HEAD(dlm_join_handlers);
1409 static void dlm_unregister_net_handlers(void)
1411 o2net_unregister_handler_list(&dlm_join_handlers);
1414 static int dlm_register_net_handlers(void)
1418 status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1419 sizeof(struct dlm_query_join_request),
1420 dlm_query_join_handler,
1421 NULL, &dlm_join_handlers);
1425 status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1426 sizeof(struct dlm_assert_joined),
1427 dlm_assert_joined_handler,
1428 NULL, &dlm_join_handlers);
1432 status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1433 sizeof(struct dlm_cancel_join),
1434 dlm_cancel_join_handler,
1435 NULL, &dlm_join_handlers);
1439 dlm_unregister_net_handlers();
1444 /* Domain eviction callback handling.
1446 * The file system requires notification of node death *before* the
1447 * dlm completes it's recovery work, otherwise it may be able to
1448 * acquire locks on resources requiring recovery. Since the dlm can
1449 * evict a node from it's domain *before* heartbeat fires, a similar
1450 * mechanism is required. */
1452 /* Eviction is not expected to happen often, so a per-domain lock is
1453 * not necessary. Eviction callbacks are allowed to sleep for short
1454 * periods of time. */
1455 static DECLARE_RWSEM(dlm_callback_sem);
1457 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
1460 struct list_head *iter;
1461 struct dlm_eviction_cb *cb;
1463 down_read(&dlm_callback_sem);
1464 list_for_each(iter, &dlm->dlm_eviction_callbacks) {
1465 cb = list_entry(iter, struct dlm_eviction_cb, ec_item);
1467 cb->ec_func(node_num, cb->ec_data);
1469 up_read(&dlm_callback_sem);
1472 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
1473 dlm_eviction_func *f,
1476 INIT_LIST_HEAD(&cb->ec_item);
1480 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
1482 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
1483 struct dlm_eviction_cb *cb)
1485 down_write(&dlm_callback_sem);
1486 list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
1487 up_write(&dlm_callback_sem);
1489 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
1491 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
1493 down_write(&dlm_callback_sem);
1494 list_del_init(&cb->ec_item);
1495 up_write(&dlm_callback_sem);
1497 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
1499 static int __init dlm_init(void)
1503 dlm_print_version();
1505 status = dlm_init_mle_cache();
1509 status = dlm_register_net_handlers();
1511 dlm_destroy_mle_cache();
1518 static void __exit dlm_exit (void)
1520 dlm_unregister_net_handlers();
1521 dlm_destroy_mle_cache();
1524 MODULE_AUTHOR("Oracle");
1525 MODULE_LICENSE("GPL");
1527 module_init(dlm_init);
1528 module_exit(dlm_exit);