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/init.h>
32 #include <linux/spinlock.h>
33 #include <linux/delay.h>
34 #include <linux/err.h>
35 #include <linux/debugfs.h>
37 #include "cluster/heartbeat.h"
38 #include "cluster/nodemanager.h"
39 #include "cluster/tcp.h"
42 #include "dlmcommon.h"
43 #include "dlmdomain.h"
48 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
49 #include "cluster/masklog.h"
52 * ocfs2 node maps are array of long int, which limits to send them freely
53 * across the wire due to endianness issues. To workaround this, we convert
54 * long ints to byte arrays. Following 3 routines are helper functions to
55 * set/test/copy bits within those array of bytes
57 static inline void byte_set_bit(u8 nr, u8 map[])
59 map[nr >> 3] |= (1UL << (nr & 7));
62 static inline int byte_test_bit(u8 nr, u8 map[])
64 return ((1UL << (nr & 7)) & (map[nr >> 3])) != 0;
67 static inline void byte_copymap(u8 dmap[], unsigned long smap[],
75 memset(dmap, 0, ((sz + 7) >> 3));
76 for (nn = 0 ; nn < sz; nn++)
77 if (test_bit(nn, smap))
78 byte_set_bit(nn, dmap);
81 static void dlm_free_pagevec(void **vec, int pages)
84 free_page((unsigned long)vec[pages]);
88 static void **dlm_alloc_pagevec(int pages)
90 void **vec = kmalloc(pages * sizeof(void *), GFP_KERNEL);
96 for (i = 0; i < pages; i++)
97 if (!(vec[i] = (void *)__get_free_page(GFP_KERNEL)))
100 mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n",
101 pages, (unsigned long)DLM_HASH_PAGES,
102 (unsigned long)DLM_BUCKETS_PER_PAGE);
105 dlm_free_pagevec(vec, i);
111 * spinlock lock ordering: if multiple locks are needed, obey this ordering:
113 * struct dlm_ctxt->spinlock
114 * struct dlm_lock_resource->spinlock
115 * struct dlm_ctxt->master_lock
116 * struct dlm_ctxt->ast_lock
117 * dlm_master_list_entry->spinlock
122 DEFINE_SPINLOCK(dlm_domain_lock);
123 LIST_HEAD(dlm_domains);
124 static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);
127 * The supported protocol version for DLM communication. Running domains
128 * will have a negotiated version with the same major number and a minor
129 * number equal or smaller. The dlm_ctxt->dlm_locking_proto field should
130 * be used to determine what a running domain is actually using.
132 * New in version 1.1:
133 * - Message DLM_QUERY_REGION added to support global heartbeat
134 * - Message DLM_QUERY_NODEINFO added to allow online node removes
135 * New in version 1.2:
136 * - Message DLM_BEGIN_EXIT_DOMAIN_MSG added to mark start of exit domain
138 static const struct dlm_protocol_version dlm_protocol = {
143 #define DLM_DOMAIN_BACKOFF_MS 200
145 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
147 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
149 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
151 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
152 void *data, void **ret_data);
153 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
155 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
156 struct dlm_protocol_version *request);
158 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
160 void __dlm_unhash_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
162 if (hlist_unhashed(&res->hash_node))
165 mlog(0, "%s: Unhash res %.*s\n", dlm->name, res->lockname.len,
167 hlist_del_init(&res->hash_node);
168 dlm_lockres_put(res);
171 void __dlm_insert_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
173 struct hlist_head *bucket;
176 assert_spin_locked(&dlm->spinlock);
179 bucket = dlm_lockres_hash(dlm, q->hash);
181 /* get a reference for our hashtable */
182 dlm_lockres_get(res);
184 hlist_add_head(&res->hash_node, bucket);
186 mlog(0, "%s: Hash res %.*s\n", dlm->name, res->lockname.len,
190 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
195 struct hlist_head *bucket;
196 struct dlm_lock_resource *res;
198 mlog(0, "%.*s\n", len, name);
200 assert_spin_locked(&dlm->spinlock);
202 bucket = dlm_lockres_hash(dlm, hash);
204 hlist_for_each_entry(res, bucket, hash_node) {
205 if (res->lockname.name[0] != name[0])
207 if (unlikely(res->lockname.len != len))
209 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
211 dlm_lockres_get(res);
217 /* intended to be called by functions which do not care about lock
218 * resources which are being purged (most net _handler functions).
219 * this will return NULL for any lock resource which is found but
220 * currently in the process of dropping its mastery reference.
221 * use __dlm_lookup_lockres_full when you need the lock resource
222 * regardless (e.g. dlm_get_lock_resource) */
223 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
228 struct dlm_lock_resource *res = NULL;
230 mlog(0, "%.*s\n", len, name);
232 assert_spin_locked(&dlm->spinlock);
234 res = __dlm_lookup_lockres_full(dlm, name, len, hash);
236 spin_lock(&res->spinlock);
237 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
238 spin_unlock(&res->spinlock);
239 dlm_lockres_put(res);
242 spin_unlock(&res->spinlock);
248 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
252 struct dlm_lock_resource *res;
253 unsigned int hash = dlm_lockid_hash(name, len);
255 spin_lock(&dlm->spinlock);
256 res = __dlm_lookup_lockres(dlm, name, len, hash);
257 spin_unlock(&dlm->spinlock);
261 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
263 struct dlm_ctxt *tmp;
265 assert_spin_locked(&dlm_domain_lock);
267 /* tmp->name here is always NULL terminated,
268 * but domain may not be! */
269 list_for_each_entry(tmp, &dlm_domains, list) {
270 if (strlen(tmp->name) == len &&
271 memcmp(tmp->name, domain, len)==0)
278 /* For null terminated domain strings ONLY */
279 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
281 assert_spin_locked(&dlm_domain_lock);
283 return __dlm_lookup_domain_full(domain, strlen(domain));
287 /* returns true on one of two conditions:
288 * 1) the domain does not exist
289 * 2) the domain exists and it's state is "joined" */
290 static int dlm_wait_on_domain_helper(const char *domain)
293 struct dlm_ctxt *tmp = NULL;
295 spin_lock(&dlm_domain_lock);
297 tmp = __dlm_lookup_domain(domain);
300 else if (tmp->dlm_state == DLM_CTXT_JOINED)
303 spin_unlock(&dlm_domain_lock);
307 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
309 dlm_destroy_debugfs_subroot(dlm);
311 if (dlm->lockres_hash)
312 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
314 if (dlm->master_hash)
315 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
321 /* A little strange - this function will be called while holding
322 * dlm_domain_lock and is expected to be holding it on the way out. We
323 * will however drop and reacquire it multiple times */
324 static void dlm_ctxt_release(struct kref *kref)
326 struct dlm_ctxt *dlm;
328 dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
330 BUG_ON(dlm->num_joins);
331 BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
333 /* we may still be in the list if we hit an error during join. */
334 list_del_init(&dlm->list);
336 spin_unlock(&dlm_domain_lock);
338 mlog(0, "freeing memory from domain %s\n", dlm->name);
340 wake_up(&dlm_domain_events);
342 dlm_free_ctxt_mem(dlm);
344 spin_lock(&dlm_domain_lock);
347 void dlm_put(struct dlm_ctxt *dlm)
349 spin_lock(&dlm_domain_lock);
350 kref_put(&dlm->dlm_refs, dlm_ctxt_release);
351 spin_unlock(&dlm_domain_lock);
354 static void __dlm_get(struct dlm_ctxt *dlm)
356 kref_get(&dlm->dlm_refs);
359 /* given a questionable reference to a dlm object, gets a reference if
360 * it can find it in the list, otherwise returns NULL in which case
361 * you shouldn't trust your pointer. */
362 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
364 struct dlm_ctxt *target;
365 struct dlm_ctxt *ret = NULL;
367 spin_lock(&dlm_domain_lock);
369 list_for_each_entry(target, &dlm_domains, list) {
377 spin_unlock(&dlm_domain_lock);
382 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
386 spin_lock(&dlm_domain_lock);
387 ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
388 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
389 spin_unlock(&dlm_domain_lock);
394 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
396 if (dlm->dlm_worker) {
397 flush_workqueue(dlm->dlm_worker);
398 destroy_workqueue(dlm->dlm_worker);
399 dlm->dlm_worker = NULL;
403 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
405 dlm_unregister_domain_handlers(dlm);
406 dlm_debug_shutdown(dlm);
407 dlm_complete_thread(dlm);
408 dlm_complete_recovery_thread(dlm);
409 dlm_destroy_dlm_worker(dlm);
411 /* We've left the domain. Now we can take ourselves out of the
412 * list and allow the kref stuff to help us free the
414 spin_lock(&dlm_domain_lock);
415 list_del_init(&dlm->list);
416 spin_unlock(&dlm_domain_lock);
418 /* Wake up anyone waiting for us to remove this domain */
419 wake_up(&dlm_domain_events);
422 static int dlm_migrate_all_locks(struct dlm_ctxt *dlm)
424 int i, num, n, ret = 0;
425 struct dlm_lock_resource *res;
426 struct hlist_node *iter;
427 struct hlist_head *bucket;
430 mlog(0, "Migrating locks from domain %s\n", dlm->name);
433 spin_lock(&dlm->spinlock);
434 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
437 bucket = dlm_lockres_hash(dlm, i);
438 iter = bucket->first;
441 res = hlist_entry(iter, struct dlm_lock_resource,
443 dlm_lockres_get(res);
444 /* migrate, if necessary. this will drop the dlm
445 * spinlock and retake it if it does migration. */
446 dropped = dlm_empty_lockres(dlm, res);
448 spin_lock(&res->spinlock);
450 __dlm_lockres_calc_usage(dlm, res);
452 iter = res->hash_node.next;
453 spin_unlock(&res->spinlock);
455 dlm_lockres_put(res);
458 cond_resched_lock(&dlm->spinlock);
462 cond_resched_lock(&dlm->spinlock);
465 spin_unlock(&dlm->spinlock);
466 wake_up(&dlm->dlm_thread_wq);
468 /* let the dlm thread take care of purging, keep scanning until
469 * nothing remains in the hash */
471 mlog(0, "%s: %d lock resources in hash last pass\n",
475 mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
479 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
483 spin_lock(&dlm->spinlock);
484 ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
485 spin_unlock(&dlm->spinlock);
490 static int dlm_begin_exit_domain_handler(struct o2net_msg *msg, u32 len,
491 void *data, void **ret_data)
493 struct dlm_ctxt *dlm = data;
495 struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
500 node = exit_msg->node_idx;
501 mlog(0, "%s: Node %u sent a begin exit domain message\n", dlm->name, node);
503 spin_lock(&dlm->spinlock);
504 set_bit(node, dlm->exit_domain_map);
505 spin_unlock(&dlm->spinlock);
512 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
514 /* Yikes, a double spinlock! I need domain_lock for the dlm
515 * state and the dlm spinlock for join state... Sorry! */
517 spin_lock(&dlm_domain_lock);
518 spin_lock(&dlm->spinlock);
520 if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
521 mlog(0, "Node %d is joining, we wait on it.\n",
523 spin_unlock(&dlm->spinlock);
524 spin_unlock(&dlm_domain_lock);
526 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
530 dlm->dlm_state = DLM_CTXT_LEAVING;
531 spin_unlock(&dlm->spinlock);
532 spin_unlock(&dlm_domain_lock);
535 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
537 int node = -1, num = 0;
539 assert_spin_locked(&dlm->spinlock);
542 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
543 node + 1)) < O2NM_MAX_NODES) {
547 printk(") %u nodes\n", num);
550 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
553 struct dlm_ctxt *dlm = data;
555 struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
557 mlog(0, "%p %u %p", msg, len, data);
562 node = exit_msg->node_idx;
564 spin_lock(&dlm->spinlock);
565 clear_bit(node, dlm->domain_map);
566 clear_bit(node, dlm->exit_domain_map);
567 printk(KERN_NOTICE "o2dlm: Node %u leaves domain %s ", node, dlm->name);
568 __dlm_print_nodes(dlm);
570 /* notify anything attached to the heartbeat events */
571 dlm_hb_event_notify_attached(dlm, node, 0);
573 spin_unlock(&dlm->spinlock);
580 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm, u32 msg_type,
584 struct dlm_exit_domain leave_msg;
586 mlog(0, "%s: Sending domain exit message %u to node %u\n", dlm->name,
589 memset(&leave_msg, 0, sizeof(leave_msg));
590 leave_msg.node_idx = dlm->node_num;
592 status = o2net_send_message(msg_type, dlm->key, &leave_msg,
593 sizeof(leave_msg), node, NULL);
595 mlog(ML_ERROR, "Error %d sending domain exit message %u "
596 "to node %u on domain %s\n", status, msg_type, node,
602 static void dlm_begin_exit_domain(struct dlm_ctxt *dlm)
606 /* Support for begin exit domain was added in 1.2 */
607 if (dlm->dlm_locking_proto.pv_major == 1 &&
608 dlm->dlm_locking_proto.pv_minor < 2)
612 * Unlike DLM_EXIT_DOMAIN_MSG, DLM_BEGIN_EXIT_DOMAIN_MSG is purely
613 * informational. Meaning if a node does not receive the message,
616 spin_lock(&dlm->spinlock);
618 node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, node + 1);
619 if (node >= O2NM_MAX_NODES)
621 if (node == dlm->node_num)
624 spin_unlock(&dlm->spinlock);
625 dlm_send_one_domain_exit(dlm, DLM_BEGIN_EXIT_DOMAIN_MSG, node);
626 spin_lock(&dlm->spinlock);
628 spin_unlock(&dlm->spinlock);
631 static void dlm_leave_domain(struct dlm_ctxt *dlm)
633 int node, clear_node, status;
635 /* At this point we've migrated away all our locks and won't
636 * accept mastership of new ones. The dlm is responsible for
637 * almost nothing now. We make sure not to confuse any joining
638 * nodes and then commence shutdown procedure. */
640 spin_lock(&dlm->spinlock);
641 /* Clear ourselves from the domain map */
642 clear_bit(dlm->node_num, dlm->domain_map);
643 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
644 0)) < O2NM_MAX_NODES) {
645 /* Drop the dlm spinlock. This is safe wrt the domain_map.
646 * -nodes cannot be added now as the
647 * query_join_handlers knows to respond with OK_NO_MAP
648 * -we catch the right network errors if a node is
649 * removed from the map while we're sending him the
651 spin_unlock(&dlm->spinlock);
655 status = dlm_send_one_domain_exit(dlm, DLM_EXIT_DOMAIN_MSG,
658 status != -ENOPROTOOPT &&
659 status != -ENOTCONN) {
660 mlog(ML_NOTICE, "Error %d sending domain exit message "
661 "to node %d\n", status, node);
663 /* Not sure what to do here but lets sleep for
664 * a bit in case this was a transient
666 msleep(DLM_DOMAIN_BACKOFF_MS);
670 spin_lock(&dlm->spinlock);
671 /* If we're not clearing the node bit then we intend
672 * to loop back around to try again. */
674 clear_bit(node, dlm->domain_map);
676 spin_unlock(&dlm->spinlock);
679 int dlm_joined(struct dlm_ctxt *dlm)
683 spin_lock(&dlm_domain_lock);
685 if (dlm->dlm_state == DLM_CTXT_JOINED)
688 spin_unlock(&dlm_domain_lock);
693 int dlm_shutting_down(struct dlm_ctxt *dlm)
697 spin_lock(&dlm_domain_lock);
699 if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
702 spin_unlock(&dlm_domain_lock);
707 void dlm_unregister_domain(struct dlm_ctxt *dlm)
710 struct dlm_lock_resource *res;
712 spin_lock(&dlm_domain_lock);
713 BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
714 BUG_ON(!dlm->num_joins);
717 if (!dlm->num_joins) {
718 /* We mark it "in shutdown" now so new register
719 * requests wait until we've completely left the
720 * domain. Don't use DLM_CTXT_LEAVING yet as we still
721 * want new domain joins to communicate with us at
722 * least until we've completed migration of our
724 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
727 spin_unlock(&dlm_domain_lock);
730 mlog(0, "shutting down domain %s\n", dlm->name);
731 dlm_begin_exit_domain(dlm);
733 /* We changed dlm state, notify the thread */
734 dlm_kick_thread(dlm, NULL);
736 while (dlm_migrate_all_locks(dlm)) {
737 /* Give dlm_thread time to purge the lockres' */
739 mlog(0, "%s: more migration to do\n", dlm->name);
742 /* This list should be empty. If not, print remaining lockres */
743 if (!list_empty(&dlm->tracking_list)) {
744 mlog(ML_ERROR, "Following lockres' are still on the "
746 list_for_each_entry(res, &dlm->tracking_list, tracking)
747 dlm_print_one_lock_resource(res);
750 dlm_mark_domain_leaving(dlm);
751 dlm_leave_domain(dlm);
752 printk(KERN_NOTICE "o2dlm: Leaving domain %s\n", dlm->name);
753 dlm_force_free_mles(dlm);
754 dlm_complete_dlm_shutdown(dlm);
758 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
760 static int dlm_query_join_proto_check(char *proto_type, int node,
761 struct dlm_protocol_version *ours,
762 struct dlm_protocol_version *request)
765 struct dlm_protocol_version proto = *request;
767 if (!dlm_protocol_compare(ours, &proto)) {
769 "node %u wanted to join with %s locking protocol "
770 "%u.%u, we respond with %u.%u\n",
774 proto.pv_major, proto.pv_minor);
775 request->pv_minor = proto.pv_minor;
779 "Node %u wanted to join with %s locking "
780 "protocol %u.%u, but we have %u.%u, disallowing\n",
793 * struct dlm_query_join_packet is made up of four one-byte fields. They
794 * are effectively in big-endian order already. However, little-endian
795 * machines swap them before putting the packet on the wire (because
796 * query_join's response is a status, and that status is treated as a u32
797 * on the wire). Thus, a big-endian and little-endian machines will treat
798 * this structure differently.
800 * The solution is to have little-endian machines swap the structure when
801 * converting from the structure to the u32 representation. This will
802 * result in the structure having the correct format on the wire no matter
803 * the host endian format.
805 static void dlm_query_join_packet_to_wire(struct dlm_query_join_packet *packet,
808 union dlm_query_join_response response;
810 response.packet = *packet;
811 *wire = be32_to_cpu(response.intval);
814 static void dlm_query_join_wire_to_packet(u32 wire,
815 struct dlm_query_join_packet *packet)
817 union dlm_query_join_response response;
819 response.intval = cpu_to_be32(wire);
820 *packet = response.packet;
823 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
826 struct dlm_query_join_request *query;
827 struct dlm_query_join_packet packet = {
828 .code = JOIN_DISALLOW,
830 struct dlm_ctxt *dlm = NULL;
834 query = (struct dlm_query_join_request *) msg->buf;
836 mlog(0, "node %u wants to join domain %s\n", query->node_idx,
840 * If heartbeat doesn't consider the node live, tell it
841 * to back off and try again. This gives heartbeat a chance
844 if (!o2hb_check_node_heartbeating(query->node_idx)) {
845 mlog(0, "node %u is not in our live map yet\n",
848 packet.code = JOIN_DISALLOW;
852 packet.code = JOIN_OK_NO_MAP;
854 spin_lock(&dlm_domain_lock);
855 dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
860 * There is a small window where the joining node may not see the
861 * node(s) that just left but still part of the cluster. DISALLOW
862 * join request if joining node has different node map.
865 while (nodenum < O2NM_MAX_NODES) {
866 if (test_bit(nodenum, dlm->domain_map)) {
867 if (!byte_test_bit(nodenum, query->node_map)) {
868 mlog(0, "disallow join as node %u does not "
869 "have node %u in its nodemap\n",
870 query->node_idx, nodenum);
871 packet.code = JOIN_DISALLOW;
878 /* Once the dlm ctxt is marked as leaving then we don't want
879 * to be put in someone's domain map.
880 * Also, explicitly disallow joining at certain troublesome
881 * times (ie. during recovery). */
882 if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
883 int bit = query->node_idx;
884 spin_lock(&dlm->spinlock);
886 if (dlm->dlm_state == DLM_CTXT_NEW &&
887 dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
888 /*If this is a brand new context and we
889 * haven't started our join process yet, then
890 * the other node won the race. */
891 packet.code = JOIN_OK_NO_MAP;
892 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
893 /* Disallow parallel joins. */
894 packet.code = JOIN_DISALLOW;
895 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
896 mlog(0, "node %u trying to join, but recovery "
897 "is ongoing.\n", bit);
898 packet.code = JOIN_DISALLOW;
899 } else if (test_bit(bit, dlm->recovery_map)) {
900 mlog(0, "node %u trying to join, but it "
901 "still needs recovery.\n", bit);
902 packet.code = JOIN_DISALLOW;
903 } else if (test_bit(bit, dlm->domain_map)) {
904 mlog(0, "node %u trying to join, but it "
905 "is still in the domain! needs recovery?\n",
907 packet.code = JOIN_DISALLOW;
909 /* Alright we're fully a part of this domain
910 * so we keep some state as to who's joining
911 * and indicate to him that needs to be fixed
914 /* Make sure we speak compatible locking protocols. */
915 if (dlm_query_join_proto_check("DLM", bit,
916 &dlm->dlm_locking_proto,
917 &query->dlm_proto)) {
918 packet.code = JOIN_PROTOCOL_MISMATCH;
919 } else if (dlm_query_join_proto_check("fs", bit,
920 &dlm->fs_locking_proto,
922 packet.code = JOIN_PROTOCOL_MISMATCH;
924 packet.dlm_minor = query->dlm_proto.pv_minor;
925 packet.fs_minor = query->fs_proto.pv_minor;
926 packet.code = JOIN_OK;
927 __dlm_set_joining_node(dlm, query->node_idx);
931 spin_unlock(&dlm->spinlock);
934 spin_unlock(&dlm_domain_lock);
937 mlog(0, "We respond with %u\n", packet.code);
939 dlm_query_join_packet_to_wire(&packet, &response);
943 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
946 struct dlm_assert_joined *assert;
947 struct dlm_ctxt *dlm = NULL;
949 assert = (struct dlm_assert_joined *) msg->buf;
951 mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
954 spin_lock(&dlm_domain_lock);
955 dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
956 /* XXX should we consider no dlm ctxt an error? */
958 spin_lock(&dlm->spinlock);
960 /* Alright, this node has officially joined our
961 * domain. Set him in the map and clean up our
962 * leftover join state. */
963 BUG_ON(dlm->joining_node != assert->node_idx);
964 set_bit(assert->node_idx, dlm->domain_map);
965 clear_bit(assert->node_idx, dlm->exit_domain_map);
966 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
968 printk(KERN_NOTICE "o2dlm: Node %u joins domain %s ",
969 assert->node_idx, dlm->name);
970 __dlm_print_nodes(dlm);
972 /* notify anything attached to the heartbeat events */
973 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
975 spin_unlock(&dlm->spinlock);
977 spin_unlock(&dlm_domain_lock);
982 static int dlm_match_regions(struct dlm_ctxt *dlm,
983 struct dlm_query_region *qr,
984 char *local, int locallen)
986 char *remote = qr->qr_regions;
988 int localnr, i, j, foundit;
991 if (!o2hb_global_heartbeat_active()) {
992 if (qr->qr_numregions) {
993 mlog(ML_ERROR, "Domain %s: Joining node %d has global "
994 "heartbeat enabled but local node %d does not\n",
995 qr->qr_domain, qr->qr_node, dlm->node_num);
1001 if (o2hb_global_heartbeat_active() && !qr->qr_numregions) {
1002 mlog(ML_ERROR, "Domain %s: Local node %d has global "
1003 "heartbeat enabled but joining node %d does not\n",
1004 qr->qr_domain, dlm->node_num, qr->qr_node);
1010 for (i = 0; i < qr->qr_numregions; ++i) {
1011 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, r);
1012 r += O2HB_MAX_REGION_NAME_LEN;
1015 localnr = min(O2NM_MAX_REGIONS, locallen/O2HB_MAX_REGION_NAME_LEN);
1016 localnr = o2hb_get_all_regions(local, (u8)localnr);
1018 /* compare local regions with remote */
1020 for (i = 0; i < localnr; ++i) {
1023 for (j = 0; j <= qr->qr_numregions; ++j) {
1024 if (!memcmp(l, r, O2HB_MAX_REGION_NAME_LEN)) {
1028 r += O2HB_MAX_REGION_NAME_LEN;
1032 mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1033 "in local node %d but not in joining node %d\n",
1034 qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, l,
1035 dlm->node_num, qr->qr_node);
1038 l += O2HB_MAX_REGION_NAME_LEN;
1041 /* compare remote with local regions */
1043 for (i = 0; i < qr->qr_numregions; ++i) {
1046 for (j = 0; j < localnr; ++j) {
1047 if (!memcmp(r, l, O2HB_MAX_REGION_NAME_LEN)) {
1051 l += O2HB_MAX_REGION_NAME_LEN;
1055 mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1056 "in joining node %d but not in local node %d\n",
1057 qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, r,
1058 qr->qr_node, dlm->node_num);
1061 r += O2HB_MAX_REGION_NAME_LEN;
1068 static int dlm_send_regions(struct dlm_ctxt *dlm, unsigned long *node_map)
1070 struct dlm_query_region *qr = NULL;
1071 int status, ret = 0, i;
1074 if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1077 qr = kzalloc(sizeof(struct dlm_query_region), GFP_KERNEL);
1084 qr->qr_node = dlm->node_num;
1085 qr->qr_namelen = strlen(dlm->name);
1086 memcpy(qr->qr_domain, dlm->name, qr->qr_namelen);
1087 /* if local hb, the numregions will be zero */
1088 if (o2hb_global_heartbeat_active())
1089 qr->qr_numregions = o2hb_get_all_regions(qr->qr_regions,
1093 for (i = 0; i < qr->qr_numregions; ++i, p += O2HB_MAX_REGION_NAME_LEN)
1094 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, p);
1097 while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1098 i + 1)) < O2NM_MAX_NODES) {
1099 if (i == dlm->node_num)
1102 mlog(0, "Sending regions to node %d\n", i);
1104 ret = o2net_send_message(DLM_QUERY_REGION, DLM_MOD_KEY, qr,
1105 sizeof(struct dlm_query_region),
1110 mlog(ML_ERROR, "Region mismatch %d, node %d\n",
1121 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
1122 void *data, void **ret_data)
1124 struct dlm_query_region *qr;
1125 struct dlm_ctxt *dlm = NULL;
1130 qr = (struct dlm_query_region *) msg->buf;
1132 mlog(0, "Node %u queries hb regions on domain %s\n", qr->qr_node,
1135 /* buffer used in dlm_mast_regions() */
1136 local = kmalloc(sizeof(qr->qr_regions), GFP_KERNEL);
1144 spin_lock(&dlm_domain_lock);
1145 dlm = __dlm_lookup_domain_full(qr->qr_domain, qr->qr_namelen);
1147 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1148 "before join domain\n", qr->qr_node, qr->qr_domain);
1152 spin_lock(&dlm->spinlock);
1154 if (dlm->joining_node != qr->qr_node) {
1155 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1156 "but joining node is %d\n", qr->qr_node, qr->qr_domain,
1161 /* Support for global heartbeat was added in 1.1 */
1162 if (dlm->dlm_locking_proto.pv_major == 1 &&
1163 dlm->dlm_locking_proto.pv_minor == 0) {
1164 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1165 "but active dlm protocol is %d.%d\n", qr->qr_node,
1166 qr->qr_domain, dlm->dlm_locking_proto.pv_major,
1167 dlm->dlm_locking_proto.pv_minor);
1171 status = dlm_match_regions(dlm, qr, local, sizeof(qr->qr_regions));
1175 spin_unlock(&dlm->spinlock);
1176 spin_unlock(&dlm_domain_lock);
1183 static int dlm_match_nodes(struct dlm_ctxt *dlm, struct dlm_query_nodeinfo *qn)
1185 struct o2nm_node *local;
1186 struct dlm_node_info *remote;
1190 for (j = 0; j < qn->qn_numnodes; ++j)
1191 mlog(0, "Node %3d, %pI4:%u\n", qn->qn_nodes[j].ni_nodenum,
1192 &(qn->qn_nodes[j].ni_ipv4_address),
1193 ntohs(qn->qn_nodes[j].ni_ipv4_port));
1195 for (i = 0; i < O2NM_MAX_NODES && !status; ++i) {
1196 local = o2nm_get_node_by_num(i);
1198 for (j = 0; j < qn->qn_numnodes; ++j) {
1199 if (qn->qn_nodes[j].ni_nodenum == i) {
1200 remote = &(qn->qn_nodes[j]);
1205 if (!local && !remote)
1208 if ((local && !remote) || (!local && remote))
1212 ((remote->ni_nodenum != local->nd_num) ||
1213 (remote->ni_ipv4_port != local->nd_ipv4_port) ||
1214 (remote->ni_ipv4_address != local->nd_ipv4_address)))
1218 if (remote && !local)
1219 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1220 "registered in joining node %d but not in "
1221 "local node %d\n", qn->qn_domain,
1223 &(remote->ni_ipv4_address),
1224 ntohs(remote->ni_ipv4_port),
1225 qn->qn_nodenum, dlm->node_num);
1226 if (local && !remote)
1227 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1228 "registered in local node %d but not in "
1229 "joining node %d\n", qn->qn_domain,
1230 local->nd_num, &(local->nd_ipv4_address),
1231 ntohs(local->nd_ipv4_port),
1232 dlm->node_num, qn->qn_nodenum);
1233 BUG_ON((!local && !remote));
1237 o2nm_node_put(local);
1243 static int dlm_send_nodeinfo(struct dlm_ctxt *dlm, unsigned long *node_map)
1245 struct dlm_query_nodeinfo *qn = NULL;
1246 struct o2nm_node *node;
1247 int ret = 0, status, count, i;
1249 if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1252 qn = kzalloc(sizeof(struct dlm_query_nodeinfo), GFP_KERNEL);
1259 for (i = 0, count = 0; i < O2NM_MAX_NODES; ++i) {
1260 node = o2nm_get_node_by_num(i);
1263 qn->qn_nodes[count].ni_nodenum = node->nd_num;
1264 qn->qn_nodes[count].ni_ipv4_port = node->nd_ipv4_port;
1265 qn->qn_nodes[count].ni_ipv4_address = node->nd_ipv4_address;
1266 mlog(0, "Node %3d, %pI4:%u\n", node->nd_num,
1267 &(node->nd_ipv4_address), ntohs(node->nd_ipv4_port));
1269 o2nm_node_put(node);
1272 qn->qn_nodenum = dlm->node_num;
1273 qn->qn_numnodes = count;
1274 qn->qn_namelen = strlen(dlm->name);
1275 memcpy(qn->qn_domain, dlm->name, qn->qn_namelen);
1278 while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1279 i + 1)) < O2NM_MAX_NODES) {
1280 if (i == dlm->node_num)
1283 mlog(0, "Sending nodeinfo to node %d\n", i);
1285 ret = o2net_send_message(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
1286 qn, sizeof(struct dlm_query_nodeinfo),
1291 mlog(ML_ERROR, "node mismatch %d, node %d\n", ret, i);
1301 static int dlm_query_nodeinfo_handler(struct o2net_msg *msg, u32 len,
1302 void *data, void **ret_data)
1304 struct dlm_query_nodeinfo *qn;
1305 struct dlm_ctxt *dlm = NULL;
1306 int locked = 0, status = -EINVAL;
1308 qn = (struct dlm_query_nodeinfo *) msg->buf;
1310 mlog(0, "Node %u queries nodes on domain %s\n", qn->qn_nodenum,
1313 spin_lock(&dlm_domain_lock);
1314 dlm = __dlm_lookup_domain_full(qn->qn_domain, qn->qn_namelen);
1316 mlog(ML_ERROR, "Node %d queried nodes on domain %s before "
1317 "join domain\n", qn->qn_nodenum, qn->qn_domain);
1321 spin_lock(&dlm->spinlock);
1323 if (dlm->joining_node != qn->qn_nodenum) {
1324 mlog(ML_ERROR, "Node %d queried nodes on domain %s but "
1325 "joining node is %d\n", qn->qn_nodenum, qn->qn_domain,
1330 /* Support for node query was added in 1.1 */
1331 if (dlm->dlm_locking_proto.pv_major == 1 &&
1332 dlm->dlm_locking_proto.pv_minor == 0) {
1333 mlog(ML_ERROR, "Node %d queried nodes on domain %s "
1334 "but active dlm protocol is %d.%d\n", qn->qn_nodenum,
1335 qn->qn_domain, dlm->dlm_locking_proto.pv_major,
1336 dlm->dlm_locking_proto.pv_minor);
1340 status = dlm_match_nodes(dlm, qn);
1344 spin_unlock(&dlm->spinlock);
1345 spin_unlock(&dlm_domain_lock);
1350 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
1353 struct dlm_cancel_join *cancel;
1354 struct dlm_ctxt *dlm = NULL;
1356 cancel = (struct dlm_cancel_join *) msg->buf;
1358 mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
1361 spin_lock(&dlm_domain_lock);
1362 dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
1365 spin_lock(&dlm->spinlock);
1367 /* Yikes, this guy wants to cancel his join. No
1368 * problem, we simply cleanup our join state. */
1369 BUG_ON(dlm->joining_node != cancel->node_idx);
1370 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1372 spin_unlock(&dlm->spinlock);
1374 spin_unlock(&dlm_domain_lock);
1379 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
1383 struct dlm_cancel_join cancel_msg;
1385 memset(&cancel_msg, 0, sizeof(cancel_msg));
1386 cancel_msg.node_idx = dlm->node_num;
1387 cancel_msg.name_len = strlen(dlm->name);
1388 memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
1390 status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1391 &cancel_msg, sizeof(cancel_msg), node,
1394 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1395 "node %u\n", status, DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1404 /* map_size should be in bytes. */
1405 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
1406 unsigned long *node_map,
1407 unsigned int map_size)
1409 int status, tmpstat;
1412 if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
1413 sizeof(unsigned long))) {
1415 "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
1416 map_size, (unsigned)BITS_TO_LONGS(O2NM_MAX_NODES));
1422 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1423 node + 1)) < O2NM_MAX_NODES) {
1424 if (node == dlm->node_num)
1427 tmpstat = dlm_send_one_join_cancel(dlm, node);
1429 mlog(ML_ERROR, "Error return %d cancelling join on "
1430 "node %d\n", tmpstat, node);
1441 static int dlm_request_join(struct dlm_ctxt *dlm,
1443 enum dlm_query_join_response_code *response)
1446 struct dlm_query_join_request join_msg;
1447 struct dlm_query_join_packet packet;
1450 mlog(0, "querying node %d\n", node);
1452 memset(&join_msg, 0, sizeof(join_msg));
1453 join_msg.node_idx = dlm->node_num;
1454 join_msg.name_len = strlen(dlm->name);
1455 memcpy(join_msg.domain, dlm->name, join_msg.name_len);
1456 join_msg.dlm_proto = dlm->dlm_locking_proto;
1457 join_msg.fs_proto = dlm->fs_locking_proto;
1459 /* copy live node map to join message */
1460 byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES);
1462 status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
1463 sizeof(join_msg), node, &join_resp);
1464 if (status < 0 && status != -ENOPROTOOPT) {
1465 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1466 "node %u\n", status, DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1470 dlm_query_join_wire_to_packet(join_resp, &packet);
1472 /* -ENOPROTOOPT from the net code means the other side isn't
1473 listening for our message type -- that's fine, it means
1474 his dlm isn't up, so we can consider him a 'yes' but not
1475 joined into the domain. */
1476 if (status == -ENOPROTOOPT) {
1478 *response = JOIN_OK_NO_MAP;
1479 } else if (packet.code == JOIN_DISALLOW ||
1480 packet.code == JOIN_OK_NO_MAP) {
1481 *response = packet.code;
1482 } else if (packet.code == JOIN_PROTOCOL_MISMATCH) {
1484 "This node requested DLM locking protocol %u.%u and "
1485 "filesystem locking protocol %u.%u. At least one of "
1486 "the protocol versions on node %d is not compatible, "
1488 dlm->dlm_locking_proto.pv_major,
1489 dlm->dlm_locking_proto.pv_minor,
1490 dlm->fs_locking_proto.pv_major,
1491 dlm->fs_locking_proto.pv_minor,
1494 *response = packet.code;
1495 } else if (packet.code == JOIN_OK) {
1496 *response = packet.code;
1497 /* Use the same locking protocol as the remote node */
1498 dlm->dlm_locking_proto.pv_minor = packet.dlm_minor;
1499 dlm->fs_locking_proto.pv_minor = packet.fs_minor;
1501 "Node %d responds JOIN_OK with DLM locking protocol "
1502 "%u.%u and fs locking protocol %u.%u\n",
1504 dlm->dlm_locking_proto.pv_major,
1505 dlm->dlm_locking_proto.pv_minor,
1506 dlm->fs_locking_proto.pv_major,
1507 dlm->fs_locking_proto.pv_minor);
1510 mlog(ML_ERROR, "invalid response %d from node %u\n",
1514 mlog(0, "status %d, node %d response is %d\n", status, node,
1521 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
1525 struct dlm_assert_joined assert_msg;
1527 mlog(0, "Sending join assert to node %u\n", node);
1529 memset(&assert_msg, 0, sizeof(assert_msg));
1530 assert_msg.node_idx = dlm->node_num;
1531 assert_msg.name_len = strlen(dlm->name);
1532 memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
1534 status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1535 &assert_msg, sizeof(assert_msg), node,
1538 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1539 "node %u\n", status, DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1545 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
1546 unsigned long *node_map)
1548 int status, node, live;
1552 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1553 node + 1)) < O2NM_MAX_NODES) {
1554 if (node == dlm->node_num)
1558 /* It is very important that this message be
1559 * received so we spin until either the node
1560 * has died or it gets the message. */
1561 status = dlm_send_one_join_assert(dlm, node);
1563 spin_lock(&dlm->spinlock);
1564 live = test_bit(node, dlm->live_nodes_map);
1565 spin_unlock(&dlm->spinlock);
1568 mlog(ML_ERROR, "Error return %d asserting "
1569 "join on node %d\n", status, node);
1571 /* give us some time between errors... */
1573 msleep(DLM_DOMAIN_BACKOFF_MS);
1575 } while (status && live);
1579 struct domain_join_ctxt {
1580 unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1581 unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1584 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
1585 struct domain_join_ctxt *ctxt,
1586 enum dlm_query_join_response_code response)
1590 if (response == JOIN_DISALLOW) {
1591 mlog(0, "Latest response of disallow -- should restart\n");
1595 spin_lock(&dlm->spinlock);
1596 /* For now, we restart the process if the node maps have
1598 ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
1599 sizeof(dlm->live_nodes_map));
1600 spin_unlock(&dlm->spinlock);
1603 mlog(0, "Node maps changed -- should restart\n");
1608 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
1610 int status = 0, tmpstat, node;
1611 struct domain_join_ctxt *ctxt;
1612 enum dlm_query_join_response_code response = JOIN_DISALLOW;
1616 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
1623 /* group sem locking should work for us here -- we're already
1624 * registered for heartbeat events so filling this should be
1625 * atomic wrt getting those handlers called. */
1626 o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
1628 spin_lock(&dlm->spinlock);
1629 memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
1631 __dlm_set_joining_node(dlm, dlm->node_num);
1633 spin_unlock(&dlm->spinlock);
1636 while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
1637 node + 1)) < O2NM_MAX_NODES) {
1638 if (node == dlm->node_num)
1641 status = dlm_request_join(dlm, node, &response);
1647 /* Ok, either we got a response or the node doesn't have a
1649 if (response == JOIN_OK)
1650 set_bit(node, ctxt->yes_resp_map);
1652 if (dlm_should_restart_join(dlm, ctxt, response)) {
1658 mlog(0, "Yay, done querying nodes!\n");
1660 /* Yay, everyone agree's we can join the domain. My domain is
1661 * comprised of all nodes who were put in the
1662 * yes_resp_map. Copy that into our domain map and send a join
1663 * assert message to clean up everyone elses state. */
1664 spin_lock(&dlm->spinlock);
1665 memcpy(dlm->domain_map, ctxt->yes_resp_map,
1666 sizeof(ctxt->yes_resp_map));
1667 set_bit(dlm->node_num, dlm->domain_map);
1668 spin_unlock(&dlm->spinlock);
1670 /* Support for global heartbeat and node info was added in 1.1 */
1671 if (dlm->dlm_locking_proto.pv_major > 1 ||
1672 dlm->dlm_locking_proto.pv_minor > 0) {
1673 status = dlm_send_nodeinfo(dlm, ctxt->yes_resp_map);
1678 status = dlm_send_regions(dlm, ctxt->yes_resp_map);
1685 dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1687 /* Joined state *must* be set before the joining node
1688 * information, otherwise the query_join handler may read no
1689 * current joiner but a state of NEW and tell joining nodes
1690 * we're not in the domain. */
1691 spin_lock(&dlm_domain_lock);
1692 dlm->dlm_state = DLM_CTXT_JOINED;
1694 spin_unlock(&dlm_domain_lock);
1697 spin_lock(&dlm->spinlock);
1698 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1700 printk(KERN_NOTICE "o2dlm: Joining domain %s ", dlm->name);
1701 __dlm_print_nodes(dlm);
1703 spin_unlock(&dlm->spinlock);
1706 /* Do we need to send a cancel message to any nodes? */
1708 tmpstat = dlm_send_join_cancels(dlm,
1710 sizeof(ctxt->yes_resp_map));
1712 mlog_errno(tmpstat);
1717 mlog(0, "returning %d\n", status);
1721 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1723 o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_up);
1724 o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_down);
1725 o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1728 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1732 mlog(0, "registering handlers.\n");
1734 o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1735 dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1736 status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_down);
1740 o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1741 dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1742 status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_up);
1746 status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1747 sizeof(struct dlm_master_request),
1748 dlm_master_request_handler,
1749 dlm, NULL, &dlm->dlm_domain_handlers);
1753 status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1754 sizeof(struct dlm_assert_master),
1755 dlm_assert_master_handler,
1756 dlm, dlm_assert_master_post_handler,
1757 &dlm->dlm_domain_handlers);
1761 status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1762 sizeof(struct dlm_create_lock),
1763 dlm_create_lock_handler,
1764 dlm, NULL, &dlm->dlm_domain_handlers);
1768 status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1769 DLM_CONVERT_LOCK_MAX_LEN,
1770 dlm_convert_lock_handler,
1771 dlm, NULL, &dlm->dlm_domain_handlers);
1775 status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1776 DLM_UNLOCK_LOCK_MAX_LEN,
1777 dlm_unlock_lock_handler,
1778 dlm, NULL, &dlm->dlm_domain_handlers);
1782 status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1783 DLM_PROXY_AST_MAX_LEN,
1784 dlm_proxy_ast_handler,
1785 dlm, NULL, &dlm->dlm_domain_handlers);
1789 status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1790 sizeof(struct dlm_exit_domain),
1791 dlm_exit_domain_handler,
1792 dlm, NULL, &dlm->dlm_domain_handlers);
1796 status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1797 sizeof(struct dlm_deref_lockres),
1798 dlm_deref_lockres_handler,
1799 dlm, NULL, &dlm->dlm_domain_handlers);
1803 status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1804 sizeof(struct dlm_migrate_request),
1805 dlm_migrate_request_handler,
1806 dlm, NULL, &dlm->dlm_domain_handlers);
1810 status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1811 DLM_MIG_LOCKRES_MAX_LEN,
1812 dlm_mig_lockres_handler,
1813 dlm, NULL, &dlm->dlm_domain_handlers);
1817 status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1818 sizeof(struct dlm_master_requery),
1819 dlm_master_requery_handler,
1820 dlm, NULL, &dlm->dlm_domain_handlers);
1824 status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1825 sizeof(struct dlm_lock_request),
1826 dlm_request_all_locks_handler,
1827 dlm, NULL, &dlm->dlm_domain_handlers);
1831 status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1832 sizeof(struct dlm_reco_data_done),
1833 dlm_reco_data_done_handler,
1834 dlm, NULL, &dlm->dlm_domain_handlers);
1838 status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1839 sizeof(struct dlm_begin_reco),
1840 dlm_begin_reco_handler,
1841 dlm, NULL, &dlm->dlm_domain_handlers);
1845 status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1846 sizeof(struct dlm_finalize_reco),
1847 dlm_finalize_reco_handler,
1848 dlm, NULL, &dlm->dlm_domain_handlers);
1852 status = o2net_register_handler(DLM_BEGIN_EXIT_DOMAIN_MSG, dlm->key,
1853 sizeof(struct dlm_exit_domain),
1854 dlm_begin_exit_domain_handler,
1855 dlm, NULL, &dlm->dlm_domain_handlers);
1861 dlm_unregister_domain_handlers(dlm);
1866 static int dlm_join_domain(struct dlm_ctxt *dlm)
1869 unsigned int backoff;
1870 unsigned int total_backoff = 0;
1874 mlog(0, "Join domain %s\n", dlm->name);
1876 status = dlm_register_domain_handlers(dlm);
1882 status = dlm_debug_init(dlm);
1888 status = dlm_launch_thread(dlm);
1894 status = dlm_launch_recovery_thread(dlm);
1900 dlm->dlm_worker = create_singlethread_workqueue("dlm_wq");
1901 if (!dlm->dlm_worker) {
1908 status = dlm_try_to_join_domain(dlm);
1910 /* If we're racing another node to the join, then we
1911 * need to back off temporarily and let them
1913 #define DLM_JOIN_TIMEOUT_MSECS 90000
1914 if (status == -EAGAIN) {
1915 if (signal_pending(current)) {
1916 status = -ERESTARTSYS;
1921 msecs_to_jiffies(DLM_JOIN_TIMEOUT_MSECS)) {
1922 status = -ERESTARTSYS;
1923 mlog(ML_NOTICE, "Timed out joining dlm domain "
1924 "%s after %u msecs\n", dlm->name,
1925 jiffies_to_msecs(total_backoff));
1931 * <dale> No, after you!
1933 * <dale> But you first!
1936 backoff = (unsigned int)(jiffies & 0x3);
1937 backoff *= DLM_DOMAIN_BACKOFF_MS;
1938 total_backoff += backoff;
1939 mlog(0, "backoff %d\n", backoff);
1942 } while (status == -EAGAIN);
1951 wake_up(&dlm_domain_events);
1954 dlm_unregister_domain_handlers(dlm);
1955 dlm_debug_shutdown(dlm);
1956 dlm_complete_thread(dlm);
1957 dlm_complete_recovery_thread(dlm);
1958 dlm_destroy_dlm_worker(dlm);
1964 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1969 struct dlm_ctxt *dlm = NULL;
1971 dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1973 mlog_errno(-ENOMEM);
1977 dlm->name = kstrdup(domain, GFP_KERNEL);
1978 if (dlm->name == NULL) {
1979 mlog_errno(-ENOMEM);
1985 dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1986 if (!dlm->lockres_hash) {
1987 mlog_errno(-ENOMEM);
1994 for (i = 0; i < DLM_HASH_BUCKETS; i++)
1995 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1997 dlm->master_hash = (struct hlist_head **)
1998 dlm_alloc_pagevec(DLM_HASH_PAGES);
1999 if (!dlm->master_hash) {
2000 mlog_errno(-ENOMEM);
2001 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
2008 for (i = 0; i < DLM_HASH_BUCKETS; i++)
2009 INIT_HLIST_HEAD(dlm_master_hash(dlm, i));
2012 dlm->node_num = o2nm_this_node();
2014 ret = dlm_create_debugfs_subroot(dlm);
2016 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
2017 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
2024 spin_lock_init(&dlm->spinlock);
2025 spin_lock_init(&dlm->master_lock);
2026 spin_lock_init(&dlm->ast_lock);
2027 spin_lock_init(&dlm->track_lock);
2028 INIT_LIST_HEAD(&dlm->list);
2029 INIT_LIST_HEAD(&dlm->dirty_list);
2030 INIT_LIST_HEAD(&dlm->reco.resources);
2031 INIT_LIST_HEAD(&dlm->reco.received);
2032 INIT_LIST_HEAD(&dlm->reco.node_data);
2033 INIT_LIST_HEAD(&dlm->purge_list);
2034 INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
2035 INIT_LIST_HEAD(&dlm->tracking_list);
2036 dlm->reco.state = 0;
2038 INIT_LIST_HEAD(&dlm->pending_asts);
2039 INIT_LIST_HEAD(&dlm->pending_basts);
2041 mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
2042 dlm->recovery_map, &(dlm->recovery_map[0]));
2044 memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
2045 memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
2046 memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
2048 dlm->dlm_thread_task = NULL;
2049 dlm->dlm_reco_thread_task = NULL;
2050 dlm->dlm_worker = NULL;
2051 init_waitqueue_head(&dlm->dlm_thread_wq);
2052 init_waitqueue_head(&dlm->dlm_reco_thread_wq);
2053 init_waitqueue_head(&dlm->reco.event);
2054 init_waitqueue_head(&dlm->ast_wq);
2055 init_waitqueue_head(&dlm->migration_wq);
2056 INIT_LIST_HEAD(&dlm->mle_hb_events);
2058 dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
2059 init_waitqueue_head(&dlm->dlm_join_events);
2061 dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
2062 dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
2064 atomic_set(&dlm->res_tot_count, 0);
2065 atomic_set(&dlm->res_cur_count, 0);
2066 for (i = 0; i < DLM_MLE_NUM_TYPES; ++i) {
2067 atomic_set(&dlm->mle_tot_count[i], 0);
2068 atomic_set(&dlm->mle_cur_count[i], 0);
2071 spin_lock_init(&dlm->work_lock);
2072 INIT_LIST_HEAD(&dlm->work_list);
2073 INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
2075 kref_init(&dlm->dlm_refs);
2076 dlm->dlm_state = DLM_CTXT_NEW;
2078 INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
2080 mlog(0, "context init: refcount %u\n",
2081 atomic_read(&dlm->dlm_refs.refcount));
2088 * Compare a requested locking protocol version against the current one.
2090 * If the major numbers are different, they are incompatible.
2091 * If the current minor is greater than the request, they are incompatible.
2092 * If the current minor is less than or equal to the request, they are
2093 * compatible, and the requester should run at the current minor version.
2095 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
2096 struct dlm_protocol_version *request)
2098 if (existing->pv_major != request->pv_major)
2101 if (existing->pv_minor > request->pv_minor)
2104 if (existing->pv_minor < request->pv_minor)
2105 request->pv_minor = existing->pv_minor;
2111 * dlm_register_domain: one-time setup per "domain".
2113 * The filesystem passes in the requested locking version via proto.
2114 * If registration was successful, proto will contain the negotiated
2117 struct dlm_ctxt * dlm_register_domain(const char *domain,
2119 struct dlm_protocol_version *fs_proto)
2122 struct dlm_ctxt *dlm = NULL;
2123 struct dlm_ctxt *new_ctxt = NULL;
2125 if (strlen(domain) >= O2NM_MAX_NAME_LEN) {
2126 ret = -ENAMETOOLONG;
2127 mlog(ML_ERROR, "domain name length too long\n");
2131 mlog(0, "register called for domain \"%s\"\n", domain);
2135 if (signal_pending(current)) {
2141 spin_lock(&dlm_domain_lock);
2143 dlm = __dlm_lookup_domain(domain);
2145 if (dlm->dlm_state != DLM_CTXT_JOINED) {
2146 spin_unlock(&dlm_domain_lock);
2148 mlog(0, "This ctxt is not joined yet!\n");
2149 wait_event_interruptible(dlm_domain_events,
2150 dlm_wait_on_domain_helper(
2155 if (dlm_protocol_compare(&dlm->fs_locking_proto, fs_proto)) {
2156 spin_unlock(&dlm_domain_lock);
2158 "Requested locking protocol version is not "
2159 "compatible with already registered domain "
2160 "\"%s\"\n", domain);
2168 spin_unlock(&dlm_domain_lock);
2176 spin_unlock(&dlm_domain_lock);
2178 new_ctxt = dlm_alloc_ctxt(domain, key);
2187 /* a little variable switch-a-roo here... */
2191 /* add the new domain */
2192 list_add_tail(&dlm->list, &dlm_domains);
2193 spin_unlock(&dlm_domain_lock);
2196 * Pass the locking protocol version into the join. If the join
2197 * succeeds, it will have the negotiated protocol set.
2199 dlm->dlm_locking_proto = dlm_protocol;
2200 dlm->fs_locking_proto = *fs_proto;
2202 ret = dlm_join_domain(dlm);
2209 /* Tell the caller what locking protocol we negotiated */
2210 *fs_proto = dlm->fs_locking_proto;
2215 dlm_free_ctxt_mem(new_ctxt);
2222 EXPORT_SYMBOL_GPL(dlm_register_domain);
2224 static LIST_HEAD(dlm_join_handlers);
2226 static void dlm_unregister_net_handlers(void)
2228 o2net_unregister_handler_list(&dlm_join_handlers);
2231 static int dlm_register_net_handlers(void)
2235 status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
2236 sizeof(struct dlm_query_join_request),
2237 dlm_query_join_handler,
2238 NULL, NULL, &dlm_join_handlers);
2242 status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
2243 sizeof(struct dlm_assert_joined),
2244 dlm_assert_joined_handler,
2245 NULL, NULL, &dlm_join_handlers);
2249 status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
2250 sizeof(struct dlm_cancel_join),
2251 dlm_cancel_join_handler,
2252 NULL, NULL, &dlm_join_handlers);
2256 status = o2net_register_handler(DLM_QUERY_REGION, DLM_MOD_KEY,
2257 sizeof(struct dlm_query_region),
2258 dlm_query_region_handler,
2259 NULL, NULL, &dlm_join_handlers);
2264 status = o2net_register_handler(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
2265 sizeof(struct dlm_query_nodeinfo),
2266 dlm_query_nodeinfo_handler,
2267 NULL, NULL, &dlm_join_handlers);
2270 dlm_unregister_net_handlers();
2275 /* Domain eviction callback handling.
2277 * The file system requires notification of node death *before* the
2278 * dlm completes it's recovery work, otherwise it may be able to
2279 * acquire locks on resources requiring recovery. Since the dlm can
2280 * evict a node from it's domain *before* heartbeat fires, a similar
2281 * mechanism is required. */
2283 /* Eviction is not expected to happen often, so a per-domain lock is
2284 * not necessary. Eviction callbacks are allowed to sleep for short
2285 * periods of time. */
2286 static DECLARE_RWSEM(dlm_callback_sem);
2288 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
2291 struct dlm_eviction_cb *cb;
2293 down_read(&dlm_callback_sem);
2294 list_for_each_entry(cb, &dlm->dlm_eviction_callbacks, ec_item) {
2295 cb->ec_func(node_num, cb->ec_data);
2297 up_read(&dlm_callback_sem);
2300 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
2301 dlm_eviction_func *f,
2304 INIT_LIST_HEAD(&cb->ec_item);
2308 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
2310 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
2311 struct dlm_eviction_cb *cb)
2313 down_write(&dlm_callback_sem);
2314 list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
2315 up_write(&dlm_callback_sem);
2317 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
2319 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
2321 down_write(&dlm_callback_sem);
2322 list_del_init(&cb->ec_item);
2323 up_write(&dlm_callback_sem);
2325 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
2327 static int __init dlm_init(void)
2331 dlm_print_version();
2333 status = dlm_init_mle_cache();
2335 mlog(ML_ERROR, "Could not create o2dlm_mle slabcache\n");
2339 status = dlm_init_master_caches();
2341 mlog(ML_ERROR, "Could not create o2dlm_lockres and "
2342 "o2dlm_lockname slabcaches\n");
2346 status = dlm_init_lock_cache();
2348 mlog(ML_ERROR, "Count not create o2dlm_lock slabcache\n");
2352 status = dlm_register_net_handlers();
2354 mlog(ML_ERROR, "Unable to register network handlers\n");
2358 status = dlm_create_debugfs_root();
2364 dlm_unregister_net_handlers();
2365 dlm_destroy_lock_cache();
2366 dlm_destroy_master_caches();
2367 dlm_destroy_mle_cache();
2371 static void __exit dlm_exit (void)
2373 dlm_destroy_debugfs_root();
2374 dlm_unregister_net_handlers();
2375 dlm_destroy_lock_cache();
2376 dlm_destroy_master_caches();
2377 dlm_destroy_mle_cache();
2380 MODULE_AUTHOR("Oracle");
2381 MODULE_LICENSE("GPL");
2383 module_init(dlm_init);
2384 module_exit(dlm_exit);