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"
46 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
47 #include "cluster/masklog.h"
50 * ocfs2 node maps are array of long int, which limits to send them freely
51 * across the wire due to endianness issues. To workaround this, we convert
52 * long ints to byte arrays. Following 3 routines are helper functions to
53 * set/test/copy bits within those array of bytes
55 static inline void byte_set_bit(u8 nr, u8 map[])
57 map[nr >> 3] |= (1UL << (nr & 7));
60 static inline int byte_test_bit(u8 nr, u8 map[])
62 return ((1UL << (nr & 7)) & (map[nr >> 3])) != 0;
65 static inline void byte_copymap(u8 dmap[], unsigned long smap[],
73 memset(dmap, 0, ((sz + 7) >> 3));
74 for (nn = 0 ; nn < sz; nn++)
75 if (test_bit(nn, smap))
76 byte_set_bit(nn, dmap);
79 static void dlm_free_pagevec(void **vec, int pages)
82 free_page((unsigned long)vec[pages]);
86 static void **dlm_alloc_pagevec(int pages)
88 void **vec = kmalloc(pages * sizeof(void *), GFP_KERNEL);
94 for (i = 0; i < pages; i++)
95 if (!(vec[i] = (void *)__get_free_page(GFP_KERNEL)))
98 mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n",
99 pages, (unsigned long)DLM_HASH_PAGES,
100 (unsigned long)DLM_BUCKETS_PER_PAGE);
103 dlm_free_pagevec(vec, i);
109 * spinlock lock ordering: if multiple locks are needed, obey this ordering:
111 * struct dlm_ctxt->spinlock
112 * struct dlm_lock_resource->spinlock
113 * struct dlm_ctxt->master_lock
114 * struct dlm_ctxt->ast_lock
115 * dlm_master_list_entry->spinlock
120 DEFINE_SPINLOCK(dlm_domain_lock);
121 LIST_HEAD(dlm_domains);
122 static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);
125 * The supported protocol version for DLM communication. Running domains
126 * will have a negotiated version with the same major number and a minor
127 * number equal or smaller. The dlm_ctxt->dlm_locking_proto field should
128 * be used to determine what a running domain is actually using.
130 * New in version 1.1:
131 * - Message DLM_QUERY_REGION added to support global heartbeat
132 * - Message DLM_QUERY_NODEINFO added to allow online node removes
133 * New in version 1.2:
134 * - Message DLM_BEGIN_EXIT_DOMAIN_MSG added to mark start of exit domain
135 * New in version 1.3:
136 * - Message DLM_DEREF_LOCKRES_DONE added to inform non-master that the
139 static const struct dlm_protocol_version dlm_protocol = {
144 #define DLM_DOMAIN_BACKOFF_MS 200
146 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
148 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
150 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
152 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
153 void *data, void **ret_data);
154 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
156 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
157 struct dlm_protocol_version *request);
159 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
161 void __dlm_unhash_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
163 if (hlist_unhashed(&res->hash_node))
166 mlog(0, "%s: Unhash res %.*s\n", dlm->name, res->lockname.len,
168 hlist_del_init(&res->hash_node);
169 dlm_lockres_put(res);
172 void __dlm_insert_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
174 struct hlist_head *bucket;
176 assert_spin_locked(&dlm->spinlock);
178 bucket = dlm_lockres_hash(dlm, res->lockname.hash);
180 /* get a reference for our hashtable */
181 dlm_lockres_get(res);
183 hlist_add_head(&res->hash_node, bucket);
185 mlog(0, "%s: Hash res %.*s\n", dlm->name, res->lockname.len,
189 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
194 struct hlist_head *bucket;
195 struct dlm_lock_resource *res;
197 mlog(0, "%.*s\n", len, name);
199 assert_spin_locked(&dlm->spinlock);
201 bucket = dlm_lockres_hash(dlm, hash);
203 hlist_for_each_entry(res, bucket, hash_node) {
204 if (res->lockname.name[0] != name[0])
206 if (unlikely(res->lockname.len != len))
208 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
210 dlm_lockres_get(res);
216 /* intended to be called by functions which do not care about lock
217 * resources which are being purged (most net _handler functions).
218 * this will return NULL for any lock resource which is found but
219 * currently in the process of dropping its mastery reference.
220 * use __dlm_lookup_lockres_full when you need the lock resource
221 * regardless (e.g. dlm_get_lock_resource) */
222 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
227 struct dlm_lock_resource *res = NULL;
229 mlog(0, "%.*s\n", len, name);
231 assert_spin_locked(&dlm->spinlock);
233 res = __dlm_lookup_lockres_full(dlm, name, len, hash);
235 spin_lock(&res->spinlock);
236 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
237 spin_unlock(&res->spinlock);
238 dlm_lockres_put(res);
241 spin_unlock(&res->spinlock);
247 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
251 struct dlm_lock_resource *res;
252 unsigned int hash = dlm_lockid_hash(name, len);
254 spin_lock(&dlm->spinlock);
255 res = __dlm_lookup_lockres(dlm, name, len, hash);
256 spin_unlock(&dlm->spinlock);
260 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
262 struct dlm_ctxt *tmp;
264 assert_spin_locked(&dlm_domain_lock);
266 /* tmp->name here is always NULL terminated,
267 * but domain may not be! */
268 list_for_each_entry(tmp, &dlm_domains, list) {
269 if (strlen(tmp->name) == len &&
270 memcmp(tmp->name, domain, len)==0)
277 /* For null terminated domain strings ONLY */
278 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
280 assert_spin_locked(&dlm_domain_lock);
282 return __dlm_lookup_domain_full(domain, strlen(domain));
286 /* returns true on one of two conditions:
287 * 1) the domain does not exist
288 * 2) the domain exists and it's state is "joined" */
289 static int dlm_wait_on_domain_helper(const char *domain)
292 struct dlm_ctxt *tmp = NULL;
294 spin_lock(&dlm_domain_lock);
296 tmp = __dlm_lookup_domain(domain);
299 else if (tmp->dlm_state == DLM_CTXT_JOINED)
302 spin_unlock(&dlm_domain_lock);
306 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
308 dlm_destroy_debugfs_subroot(dlm);
310 if (dlm->lockres_hash)
311 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
313 if (dlm->master_hash)
314 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
320 /* A little strange - this function will be called while holding
321 * dlm_domain_lock and is expected to be holding it on the way out. We
322 * will however drop and reacquire it multiple times */
323 static void dlm_ctxt_release(struct kref *kref)
325 struct dlm_ctxt *dlm;
327 dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
329 BUG_ON(dlm->num_joins);
330 BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
332 /* we may still be in the list if we hit an error during join. */
333 list_del_init(&dlm->list);
335 spin_unlock(&dlm_domain_lock);
337 mlog(0, "freeing memory from domain %s\n", dlm->name);
339 wake_up(&dlm_domain_events);
341 dlm_free_ctxt_mem(dlm);
343 spin_lock(&dlm_domain_lock);
346 void dlm_put(struct dlm_ctxt *dlm)
348 spin_lock(&dlm_domain_lock);
349 kref_put(&dlm->dlm_refs, dlm_ctxt_release);
350 spin_unlock(&dlm_domain_lock);
353 static void __dlm_get(struct dlm_ctxt *dlm)
355 kref_get(&dlm->dlm_refs);
358 /* given a questionable reference to a dlm object, gets a reference if
359 * it can find it in the list, otherwise returns NULL in which case
360 * you shouldn't trust your pointer. */
361 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
363 struct dlm_ctxt *target;
364 struct dlm_ctxt *ret = NULL;
366 spin_lock(&dlm_domain_lock);
368 list_for_each_entry(target, &dlm_domains, list) {
376 spin_unlock(&dlm_domain_lock);
381 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
385 spin_lock(&dlm_domain_lock);
386 ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
387 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
388 spin_unlock(&dlm_domain_lock);
393 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
395 if (dlm->dlm_worker) {
396 flush_workqueue(dlm->dlm_worker);
397 destroy_workqueue(dlm->dlm_worker);
398 dlm->dlm_worker = NULL;
402 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
404 dlm_unregister_domain_handlers(dlm);
405 dlm_debug_shutdown(dlm);
406 dlm_complete_thread(dlm);
407 dlm_complete_recovery_thread(dlm);
408 dlm_destroy_dlm_worker(dlm);
410 /* We've left the domain. Now we can take ourselves out of the
411 * list and allow the kref stuff to help us free the
413 spin_lock(&dlm_domain_lock);
414 list_del_init(&dlm->list);
415 spin_unlock(&dlm_domain_lock);
417 /* Wake up anyone waiting for us to remove this domain */
418 wake_up(&dlm_domain_events);
421 static int dlm_migrate_all_locks(struct dlm_ctxt *dlm)
423 int i, num, n, ret = 0;
424 struct dlm_lock_resource *res;
425 struct hlist_node *iter;
426 struct hlist_head *bucket;
429 mlog(0, "Migrating locks from domain %s\n", dlm->name);
432 spin_lock(&dlm->spinlock);
433 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
436 bucket = dlm_lockres_hash(dlm, i);
437 iter = bucket->first;
440 res = hlist_entry(iter, struct dlm_lock_resource,
442 dlm_lockres_get(res);
443 /* migrate, if necessary. this will drop the dlm
444 * spinlock and retake it if it does migration. */
445 dropped = dlm_empty_lockres(dlm, res);
447 spin_lock(&res->spinlock);
449 __dlm_lockres_calc_usage(dlm, res);
451 iter = res->hash_node.next;
452 spin_unlock(&res->spinlock);
454 dlm_lockres_put(res);
457 cond_resched_lock(&dlm->spinlock);
461 cond_resched_lock(&dlm->spinlock);
464 spin_unlock(&dlm->spinlock);
465 wake_up(&dlm->dlm_thread_wq);
467 /* let the dlm thread take care of purging, keep scanning until
468 * nothing remains in the hash */
470 mlog(0, "%s: %d lock resources in hash last pass\n",
474 mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
478 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
482 spin_lock(&dlm->spinlock);
483 ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
484 spin_unlock(&dlm->spinlock);
489 static int dlm_begin_exit_domain_handler(struct o2net_msg *msg, u32 len,
490 void *data, void **ret_data)
492 struct dlm_ctxt *dlm = data;
494 struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
499 node = exit_msg->node_idx;
500 mlog(0, "%s: Node %u sent a begin exit domain message\n", dlm->name, node);
502 spin_lock(&dlm->spinlock);
503 set_bit(node, dlm->exit_domain_map);
504 spin_unlock(&dlm->spinlock);
511 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
513 /* Yikes, a double spinlock! I need domain_lock for the dlm
514 * state and the dlm spinlock for join state... Sorry! */
516 spin_lock(&dlm_domain_lock);
517 spin_lock(&dlm->spinlock);
519 if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
520 mlog(0, "Node %d is joining, we wait on it.\n",
522 spin_unlock(&dlm->spinlock);
523 spin_unlock(&dlm_domain_lock);
525 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
529 dlm->dlm_state = DLM_CTXT_LEAVING;
530 spin_unlock(&dlm->spinlock);
531 spin_unlock(&dlm_domain_lock);
534 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
536 int node = -1, num = 0;
538 assert_spin_locked(&dlm->spinlock);
541 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
542 node + 1)) < O2NM_MAX_NODES) {
546 printk(") %u nodes\n", num);
549 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
552 struct dlm_ctxt *dlm = data;
554 struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
556 mlog(0, "%p %u %p", msg, len, data);
561 node = exit_msg->node_idx;
563 spin_lock(&dlm->spinlock);
564 clear_bit(node, dlm->domain_map);
565 clear_bit(node, dlm->exit_domain_map);
566 printk(KERN_NOTICE "o2dlm: Node %u leaves domain %s ", node, dlm->name);
567 __dlm_print_nodes(dlm);
569 /* notify anything attached to the heartbeat events */
570 dlm_hb_event_notify_attached(dlm, node, 0);
572 spin_unlock(&dlm->spinlock);
579 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm, u32 msg_type,
583 struct dlm_exit_domain leave_msg;
585 mlog(0, "%s: Sending domain exit message %u to node %u\n", dlm->name,
588 memset(&leave_msg, 0, sizeof(leave_msg));
589 leave_msg.node_idx = dlm->node_num;
591 status = o2net_send_message(msg_type, dlm->key, &leave_msg,
592 sizeof(leave_msg), node, NULL);
594 mlog(ML_ERROR, "Error %d sending domain exit message %u "
595 "to node %u on domain %s\n", status, msg_type, node,
601 static void dlm_begin_exit_domain(struct dlm_ctxt *dlm)
605 /* Support for begin exit domain was added in 1.2 */
606 if (dlm->dlm_locking_proto.pv_major == 1 &&
607 dlm->dlm_locking_proto.pv_minor < 2)
611 * Unlike DLM_EXIT_DOMAIN_MSG, DLM_BEGIN_EXIT_DOMAIN_MSG is purely
612 * informational. Meaning if a node does not receive the message,
615 spin_lock(&dlm->spinlock);
617 node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, node + 1);
618 if (node >= O2NM_MAX_NODES)
620 if (node == dlm->node_num)
623 spin_unlock(&dlm->spinlock);
624 dlm_send_one_domain_exit(dlm, DLM_BEGIN_EXIT_DOMAIN_MSG, node);
625 spin_lock(&dlm->spinlock);
627 spin_unlock(&dlm->spinlock);
630 static void dlm_leave_domain(struct dlm_ctxt *dlm)
632 int node, clear_node, status;
634 /* At this point we've migrated away all our locks and won't
635 * accept mastership of new ones. The dlm is responsible for
636 * almost nothing now. We make sure not to confuse any joining
637 * nodes and then commence shutdown procedure. */
639 spin_lock(&dlm->spinlock);
640 /* Clear ourselves from the domain map */
641 clear_bit(dlm->node_num, dlm->domain_map);
642 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
643 0)) < O2NM_MAX_NODES) {
644 /* Drop the dlm spinlock. This is safe wrt the domain_map.
645 * -nodes cannot be added now as the
646 * query_join_handlers knows to respond with OK_NO_MAP
647 * -we catch the right network errors if a node is
648 * removed from the map while we're sending him the
650 spin_unlock(&dlm->spinlock);
654 status = dlm_send_one_domain_exit(dlm, DLM_EXIT_DOMAIN_MSG,
657 status != -ENOPROTOOPT &&
658 status != -ENOTCONN) {
659 mlog(ML_NOTICE, "Error %d sending domain exit message "
660 "to node %d\n", status, node);
662 /* Not sure what to do here but lets sleep for
663 * a bit in case this was a transient
665 msleep(DLM_DOMAIN_BACKOFF_MS);
669 spin_lock(&dlm->spinlock);
670 /* If we're not clearing the node bit then we intend
671 * to loop back around to try again. */
673 clear_bit(node, dlm->domain_map);
675 spin_unlock(&dlm->spinlock);
678 void dlm_unregister_domain(struct dlm_ctxt *dlm)
681 struct dlm_lock_resource *res;
683 spin_lock(&dlm_domain_lock);
684 BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
685 BUG_ON(!dlm->num_joins);
688 if (!dlm->num_joins) {
689 /* We mark it "in shutdown" now so new register
690 * requests wait until we've completely left the
691 * domain. Don't use DLM_CTXT_LEAVING yet as we still
692 * want new domain joins to communicate with us at
693 * least until we've completed migration of our
695 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
698 spin_unlock(&dlm_domain_lock);
701 mlog(0, "shutting down domain %s\n", dlm->name);
702 dlm_begin_exit_domain(dlm);
704 /* We changed dlm state, notify the thread */
705 dlm_kick_thread(dlm, NULL);
707 while (dlm_migrate_all_locks(dlm)) {
708 /* Give dlm_thread time to purge the lockres' */
710 mlog(0, "%s: more migration to do\n", dlm->name);
713 /* This list should be empty. If not, print remaining lockres */
714 if (!list_empty(&dlm->tracking_list)) {
715 mlog(ML_ERROR, "Following lockres' are still on the "
717 list_for_each_entry(res, &dlm->tracking_list, tracking)
718 dlm_print_one_lock_resource(res);
721 dlm_mark_domain_leaving(dlm);
722 dlm_leave_domain(dlm);
723 printk(KERN_NOTICE "o2dlm: Leaving domain %s\n", dlm->name);
724 dlm_force_free_mles(dlm);
725 dlm_complete_dlm_shutdown(dlm);
729 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
731 static int dlm_query_join_proto_check(char *proto_type, int node,
732 struct dlm_protocol_version *ours,
733 struct dlm_protocol_version *request)
736 struct dlm_protocol_version proto = *request;
738 if (!dlm_protocol_compare(ours, &proto)) {
740 "node %u wanted to join with %s locking protocol "
741 "%u.%u, we respond with %u.%u\n",
745 proto.pv_major, proto.pv_minor);
746 request->pv_minor = proto.pv_minor;
750 "Node %u wanted to join with %s locking "
751 "protocol %u.%u, but we have %u.%u, disallowing\n",
764 * struct dlm_query_join_packet is made up of four one-byte fields. They
765 * are effectively in big-endian order already. However, little-endian
766 * machines swap them before putting the packet on the wire (because
767 * query_join's response is a status, and that status is treated as a u32
768 * on the wire). Thus, a big-endian and little-endian machines will treat
769 * this structure differently.
771 * The solution is to have little-endian machines swap the structure when
772 * converting from the structure to the u32 representation. This will
773 * result in the structure having the correct format on the wire no matter
774 * the host endian format.
776 static void dlm_query_join_packet_to_wire(struct dlm_query_join_packet *packet,
779 union dlm_query_join_response response;
781 response.packet = *packet;
782 *wire = be32_to_cpu(response.intval);
785 static void dlm_query_join_wire_to_packet(u32 wire,
786 struct dlm_query_join_packet *packet)
788 union dlm_query_join_response response;
790 response.intval = cpu_to_be32(wire);
791 *packet = response.packet;
794 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
797 struct dlm_query_join_request *query;
798 struct dlm_query_join_packet packet = {
799 .code = JOIN_DISALLOW,
801 struct dlm_ctxt *dlm = NULL;
805 query = (struct dlm_query_join_request *) msg->buf;
807 mlog(0, "node %u wants to join domain %s\n", query->node_idx,
811 * If heartbeat doesn't consider the node live, tell it
812 * to back off and try again. This gives heartbeat a chance
815 if (!o2hb_check_node_heartbeating_no_sem(query->node_idx)) {
816 mlog(0, "node %u is not in our live map yet\n",
819 packet.code = JOIN_DISALLOW;
823 packet.code = JOIN_OK_NO_MAP;
825 spin_lock(&dlm_domain_lock);
826 dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
831 * There is a small window where the joining node may not see the
832 * node(s) that just left but still part of the cluster. DISALLOW
833 * join request if joining node has different node map.
836 while (nodenum < O2NM_MAX_NODES) {
837 if (test_bit(nodenum, dlm->domain_map)) {
838 if (!byte_test_bit(nodenum, query->node_map)) {
839 mlog(0, "disallow join as node %u does not "
840 "have node %u in its nodemap\n",
841 query->node_idx, nodenum);
842 packet.code = JOIN_DISALLOW;
849 /* Once the dlm ctxt is marked as leaving then we don't want
850 * to be put in someone's domain map.
851 * Also, explicitly disallow joining at certain troublesome
852 * times (ie. during recovery). */
853 if (dlm->dlm_state != DLM_CTXT_LEAVING) {
854 int bit = query->node_idx;
855 spin_lock(&dlm->spinlock);
857 if (dlm->dlm_state == DLM_CTXT_NEW &&
858 dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
859 /*If this is a brand new context and we
860 * haven't started our join process yet, then
861 * the other node won the race. */
862 packet.code = JOIN_OK_NO_MAP;
863 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
864 /* Disallow parallel joins. */
865 packet.code = JOIN_DISALLOW;
866 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
867 mlog(0, "node %u trying to join, but recovery "
868 "is ongoing.\n", bit);
869 packet.code = JOIN_DISALLOW;
870 } else if (test_bit(bit, dlm->recovery_map)) {
871 mlog(0, "node %u trying to join, but it "
872 "still needs recovery.\n", bit);
873 packet.code = JOIN_DISALLOW;
874 } else if (test_bit(bit, dlm->domain_map)) {
875 mlog(0, "node %u trying to join, but it "
876 "is still in the domain! needs recovery?\n",
878 packet.code = JOIN_DISALLOW;
880 /* Alright we're fully a part of this domain
881 * so we keep some state as to who's joining
882 * and indicate to him that needs to be fixed
885 /* Make sure we speak compatible locking protocols. */
886 if (dlm_query_join_proto_check("DLM", bit,
887 &dlm->dlm_locking_proto,
888 &query->dlm_proto)) {
889 packet.code = JOIN_PROTOCOL_MISMATCH;
890 } else if (dlm_query_join_proto_check("fs", bit,
891 &dlm->fs_locking_proto,
893 packet.code = JOIN_PROTOCOL_MISMATCH;
895 packet.dlm_minor = query->dlm_proto.pv_minor;
896 packet.fs_minor = query->fs_proto.pv_minor;
897 packet.code = JOIN_OK;
898 __dlm_set_joining_node(dlm, query->node_idx);
902 spin_unlock(&dlm->spinlock);
905 spin_unlock(&dlm_domain_lock);
908 mlog(0, "We respond with %u\n", packet.code);
910 dlm_query_join_packet_to_wire(&packet, &response);
914 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
917 struct dlm_assert_joined *assert;
918 struct dlm_ctxt *dlm = NULL;
920 assert = (struct dlm_assert_joined *) msg->buf;
922 mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
925 spin_lock(&dlm_domain_lock);
926 dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
927 /* XXX should we consider no dlm ctxt an error? */
929 spin_lock(&dlm->spinlock);
931 /* Alright, this node has officially joined our
932 * domain. Set him in the map and clean up our
933 * leftover join state. */
934 BUG_ON(dlm->joining_node != assert->node_idx);
936 if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
937 mlog(0, "dlm recovery is ongoing, disallow join\n");
938 spin_unlock(&dlm->spinlock);
939 spin_unlock(&dlm_domain_lock);
943 set_bit(assert->node_idx, dlm->domain_map);
944 clear_bit(assert->node_idx, dlm->exit_domain_map);
945 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
947 printk(KERN_NOTICE "o2dlm: Node %u joins domain %s ",
948 assert->node_idx, dlm->name);
949 __dlm_print_nodes(dlm);
951 /* notify anything attached to the heartbeat events */
952 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
954 spin_unlock(&dlm->spinlock);
956 spin_unlock(&dlm_domain_lock);
961 static int dlm_match_regions(struct dlm_ctxt *dlm,
962 struct dlm_query_region *qr,
963 char *local, int locallen)
965 char *remote = qr->qr_regions;
967 int localnr, i, j, foundit;
970 if (!o2hb_global_heartbeat_active()) {
971 if (qr->qr_numregions) {
972 mlog(ML_ERROR, "Domain %s: Joining node %d has global "
973 "heartbeat enabled but local node %d does not\n",
974 qr->qr_domain, qr->qr_node, dlm->node_num);
980 if (o2hb_global_heartbeat_active() && !qr->qr_numregions) {
981 mlog(ML_ERROR, "Domain %s: Local node %d has global "
982 "heartbeat enabled but joining node %d does not\n",
983 qr->qr_domain, dlm->node_num, qr->qr_node);
989 for (i = 0; i < qr->qr_numregions; ++i) {
990 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, r);
991 r += O2HB_MAX_REGION_NAME_LEN;
994 localnr = min(O2NM_MAX_REGIONS, locallen/O2HB_MAX_REGION_NAME_LEN);
995 localnr = o2hb_get_all_regions(local, (u8)localnr);
997 /* compare local regions with remote */
999 for (i = 0; i < localnr; ++i) {
1002 for (j = 0; j <= qr->qr_numregions; ++j) {
1003 if (!memcmp(l, r, O2HB_MAX_REGION_NAME_LEN)) {
1007 r += O2HB_MAX_REGION_NAME_LEN;
1011 mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1012 "in local node %d but not in joining node %d\n",
1013 qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, l,
1014 dlm->node_num, qr->qr_node);
1017 l += O2HB_MAX_REGION_NAME_LEN;
1020 /* compare remote with local regions */
1022 for (i = 0; i < qr->qr_numregions; ++i) {
1025 for (j = 0; j < localnr; ++j) {
1026 if (!memcmp(r, l, O2HB_MAX_REGION_NAME_LEN)) {
1030 l += O2HB_MAX_REGION_NAME_LEN;
1034 mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1035 "in joining node %d but not in local node %d\n",
1036 qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, r,
1037 qr->qr_node, dlm->node_num);
1040 r += O2HB_MAX_REGION_NAME_LEN;
1047 static int dlm_send_regions(struct dlm_ctxt *dlm, unsigned long *node_map)
1049 struct dlm_query_region *qr = NULL;
1050 int status, ret = 0, i;
1053 if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1056 qr = kzalloc(sizeof(struct dlm_query_region), GFP_KERNEL);
1063 qr->qr_node = dlm->node_num;
1064 qr->qr_namelen = strlen(dlm->name);
1065 memcpy(qr->qr_domain, dlm->name, qr->qr_namelen);
1066 /* if local hb, the numregions will be zero */
1067 if (o2hb_global_heartbeat_active())
1068 qr->qr_numregions = o2hb_get_all_regions(qr->qr_regions,
1072 for (i = 0; i < qr->qr_numregions; ++i, p += O2HB_MAX_REGION_NAME_LEN)
1073 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, p);
1076 while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1077 i + 1)) < O2NM_MAX_NODES) {
1078 if (i == dlm->node_num)
1081 mlog(0, "Sending regions to node %d\n", i);
1083 ret = o2net_send_message(DLM_QUERY_REGION, DLM_MOD_KEY, qr,
1084 sizeof(struct dlm_query_region),
1089 mlog(ML_ERROR, "Region mismatch %d, node %d\n",
1100 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
1101 void *data, void **ret_data)
1103 struct dlm_query_region *qr;
1104 struct dlm_ctxt *dlm = NULL;
1108 qr = (struct dlm_query_region *) msg->buf;
1110 mlog(0, "Node %u queries hb regions on domain %s\n", qr->qr_node,
1113 /* buffer used in dlm_mast_regions() */
1114 local = kmalloc(sizeof(qr->qr_regions), GFP_KERNEL);
1120 spin_lock(&dlm_domain_lock);
1121 dlm = __dlm_lookup_domain_full(qr->qr_domain, qr->qr_namelen);
1123 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1124 "before join domain\n", qr->qr_node, qr->qr_domain);
1125 goto out_domain_lock;
1128 spin_lock(&dlm->spinlock);
1129 if (dlm->joining_node != qr->qr_node) {
1130 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1131 "but joining node is %d\n", qr->qr_node, qr->qr_domain,
1136 /* Support for global heartbeat was added in 1.1 */
1137 if (dlm->dlm_locking_proto.pv_major == 1 &&
1138 dlm->dlm_locking_proto.pv_minor == 0) {
1139 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1140 "but active dlm protocol is %d.%d\n", qr->qr_node,
1141 qr->qr_domain, dlm->dlm_locking_proto.pv_major,
1142 dlm->dlm_locking_proto.pv_minor);
1146 status = dlm_match_regions(dlm, qr, local, sizeof(qr->qr_regions));
1149 spin_unlock(&dlm->spinlock);
1152 spin_unlock(&dlm_domain_lock);
1159 static int dlm_match_nodes(struct dlm_ctxt *dlm, struct dlm_query_nodeinfo *qn)
1161 struct o2nm_node *local;
1162 struct dlm_node_info *remote;
1166 for (j = 0; j < qn->qn_numnodes; ++j)
1167 mlog(0, "Node %3d, %pI4:%u\n", qn->qn_nodes[j].ni_nodenum,
1168 &(qn->qn_nodes[j].ni_ipv4_address),
1169 ntohs(qn->qn_nodes[j].ni_ipv4_port));
1171 for (i = 0; i < O2NM_MAX_NODES && !status; ++i) {
1172 local = o2nm_get_node_by_num(i);
1174 for (j = 0; j < qn->qn_numnodes; ++j) {
1175 if (qn->qn_nodes[j].ni_nodenum == i) {
1176 remote = &(qn->qn_nodes[j]);
1181 if (!local && !remote)
1184 if ((local && !remote) || (!local && remote))
1188 ((remote->ni_nodenum != local->nd_num) ||
1189 (remote->ni_ipv4_port != local->nd_ipv4_port) ||
1190 (remote->ni_ipv4_address != local->nd_ipv4_address)))
1194 if (remote && !local)
1195 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1196 "registered in joining node %d but not in "
1197 "local node %d\n", qn->qn_domain,
1199 &(remote->ni_ipv4_address),
1200 ntohs(remote->ni_ipv4_port),
1201 qn->qn_nodenum, dlm->node_num);
1202 if (local && !remote)
1203 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1204 "registered in local node %d but not in "
1205 "joining node %d\n", qn->qn_domain,
1206 local->nd_num, &(local->nd_ipv4_address),
1207 ntohs(local->nd_ipv4_port),
1208 dlm->node_num, qn->qn_nodenum);
1209 BUG_ON((!local && !remote));
1213 o2nm_node_put(local);
1219 static int dlm_send_nodeinfo(struct dlm_ctxt *dlm, unsigned long *node_map)
1221 struct dlm_query_nodeinfo *qn = NULL;
1222 struct o2nm_node *node;
1223 int ret = 0, status, count, i;
1225 if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1228 qn = kzalloc(sizeof(struct dlm_query_nodeinfo), GFP_KERNEL);
1235 for (i = 0, count = 0; i < O2NM_MAX_NODES; ++i) {
1236 node = o2nm_get_node_by_num(i);
1239 qn->qn_nodes[count].ni_nodenum = node->nd_num;
1240 qn->qn_nodes[count].ni_ipv4_port = node->nd_ipv4_port;
1241 qn->qn_nodes[count].ni_ipv4_address = node->nd_ipv4_address;
1242 mlog(0, "Node %3d, %pI4:%u\n", node->nd_num,
1243 &(node->nd_ipv4_address), ntohs(node->nd_ipv4_port));
1245 o2nm_node_put(node);
1248 qn->qn_nodenum = dlm->node_num;
1249 qn->qn_numnodes = count;
1250 qn->qn_namelen = strlen(dlm->name);
1251 memcpy(qn->qn_domain, dlm->name, qn->qn_namelen);
1254 while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1255 i + 1)) < O2NM_MAX_NODES) {
1256 if (i == dlm->node_num)
1259 mlog(0, "Sending nodeinfo to node %d\n", i);
1261 ret = o2net_send_message(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
1262 qn, sizeof(struct dlm_query_nodeinfo),
1267 mlog(ML_ERROR, "node mismatch %d, node %d\n", ret, i);
1277 static int dlm_query_nodeinfo_handler(struct o2net_msg *msg, u32 len,
1278 void *data, void **ret_data)
1280 struct dlm_query_nodeinfo *qn;
1281 struct dlm_ctxt *dlm = NULL;
1282 int locked = 0, status = -EINVAL;
1284 qn = (struct dlm_query_nodeinfo *) msg->buf;
1286 mlog(0, "Node %u queries nodes on domain %s\n", qn->qn_nodenum,
1289 spin_lock(&dlm_domain_lock);
1290 dlm = __dlm_lookup_domain_full(qn->qn_domain, qn->qn_namelen);
1292 mlog(ML_ERROR, "Node %d queried nodes on domain %s before "
1293 "join domain\n", qn->qn_nodenum, qn->qn_domain);
1297 spin_lock(&dlm->spinlock);
1299 if (dlm->joining_node != qn->qn_nodenum) {
1300 mlog(ML_ERROR, "Node %d queried nodes on domain %s but "
1301 "joining node is %d\n", qn->qn_nodenum, qn->qn_domain,
1306 /* Support for node query was added in 1.1 */
1307 if (dlm->dlm_locking_proto.pv_major == 1 &&
1308 dlm->dlm_locking_proto.pv_minor == 0) {
1309 mlog(ML_ERROR, "Node %d queried nodes on domain %s "
1310 "but active dlm protocol is %d.%d\n", qn->qn_nodenum,
1311 qn->qn_domain, dlm->dlm_locking_proto.pv_major,
1312 dlm->dlm_locking_proto.pv_minor);
1316 status = dlm_match_nodes(dlm, qn);
1320 spin_unlock(&dlm->spinlock);
1321 spin_unlock(&dlm_domain_lock);
1326 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
1329 struct dlm_cancel_join *cancel;
1330 struct dlm_ctxt *dlm = NULL;
1332 cancel = (struct dlm_cancel_join *) msg->buf;
1334 mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
1337 spin_lock(&dlm_domain_lock);
1338 dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
1341 spin_lock(&dlm->spinlock);
1343 /* Yikes, this guy wants to cancel his join. No
1344 * problem, we simply cleanup our join state. */
1345 BUG_ON(dlm->joining_node != cancel->node_idx);
1346 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1348 spin_unlock(&dlm->spinlock);
1350 spin_unlock(&dlm_domain_lock);
1355 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
1359 struct dlm_cancel_join cancel_msg;
1361 memset(&cancel_msg, 0, sizeof(cancel_msg));
1362 cancel_msg.node_idx = dlm->node_num;
1363 cancel_msg.name_len = strlen(dlm->name);
1364 memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
1366 status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1367 &cancel_msg, sizeof(cancel_msg), node,
1370 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1371 "node %u\n", status, DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1380 /* map_size should be in bytes. */
1381 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
1382 unsigned long *node_map,
1383 unsigned int map_size)
1385 int status, tmpstat;
1388 if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
1389 sizeof(unsigned long))) {
1391 "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
1392 map_size, (unsigned)BITS_TO_LONGS(O2NM_MAX_NODES));
1398 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1399 node + 1)) < O2NM_MAX_NODES) {
1400 if (node == dlm->node_num)
1403 tmpstat = dlm_send_one_join_cancel(dlm, node);
1405 mlog(ML_ERROR, "Error return %d cancelling join on "
1406 "node %d\n", tmpstat, node);
1417 static int dlm_request_join(struct dlm_ctxt *dlm,
1419 enum dlm_query_join_response_code *response)
1422 struct dlm_query_join_request join_msg;
1423 struct dlm_query_join_packet packet;
1426 mlog(0, "querying node %d\n", node);
1428 memset(&join_msg, 0, sizeof(join_msg));
1429 join_msg.node_idx = dlm->node_num;
1430 join_msg.name_len = strlen(dlm->name);
1431 memcpy(join_msg.domain, dlm->name, join_msg.name_len);
1432 join_msg.dlm_proto = dlm->dlm_locking_proto;
1433 join_msg.fs_proto = dlm->fs_locking_proto;
1435 /* copy live node map to join message */
1436 byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES);
1438 status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
1439 sizeof(join_msg), node, &join_resp);
1440 if (status < 0 && status != -ENOPROTOOPT) {
1441 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1442 "node %u\n", status, DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1446 dlm_query_join_wire_to_packet(join_resp, &packet);
1448 /* -ENOPROTOOPT from the net code means the other side isn't
1449 listening for our message type -- that's fine, it means
1450 his dlm isn't up, so we can consider him a 'yes' but not
1451 joined into the domain. */
1452 if (status == -ENOPROTOOPT) {
1454 *response = JOIN_OK_NO_MAP;
1456 *response = packet.code;
1457 switch (packet.code) {
1459 case JOIN_OK_NO_MAP:
1461 case JOIN_PROTOCOL_MISMATCH:
1463 "This node requested DLM locking protocol %u.%u and "
1464 "filesystem locking protocol %u.%u. At least one of "
1465 "the protocol versions on node %d is not compatible, "
1467 dlm->dlm_locking_proto.pv_major,
1468 dlm->dlm_locking_proto.pv_minor,
1469 dlm->fs_locking_proto.pv_major,
1470 dlm->fs_locking_proto.pv_minor,
1475 /* Use the same locking protocol as the remote node */
1476 dlm->dlm_locking_proto.pv_minor = packet.dlm_minor;
1477 dlm->fs_locking_proto.pv_minor = packet.fs_minor;
1479 "Node %d responds JOIN_OK with DLM locking protocol "
1480 "%u.%u and fs locking protocol %u.%u\n",
1482 dlm->dlm_locking_proto.pv_major,
1483 dlm->dlm_locking_proto.pv_minor,
1484 dlm->fs_locking_proto.pv_major,
1485 dlm->fs_locking_proto.pv_minor);
1489 mlog(ML_ERROR, "invalid response %d from node %u\n",
1491 /* Reset response to JOIN_DISALLOW */
1492 *response = JOIN_DISALLOW;
1497 mlog(0, "status %d, node %d response is %d\n", status, node,
1504 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
1509 struct dlm_assert_joined assert_msg;
1511 mlog(0, "Sending join assert to node %u\n", node);
1513 memset(&assert_msg, 0, sizeof(assert_msg));
1514 assert_msg.node_idx = dlm->node_num;
1515 assert_msg.name_len = strlen(dlm->name);
1516 memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
1518 status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1519 &assert_msg, sizeof(assert_msg), node,
1522 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1523 "node %u\n", status, DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1531 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
1532 unsigned long *node_map)
1534 int status, node, live;
1538 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1539 node + 1)) < O2NM_MAX_NODES) {
1540 if (node == dlm->node_num)
1544 /* It is very important that this message be
1545 * received so we spin until either the node
1546 * has died or it gets the message. */
1547 status = dlm_send_one_join_assert(dlm, node);
1549 spin_lock(&dlm->spinlock);
1550 live = test_bit(node, dlm->live_nodes_map);
1551 spin_unlock(&dlm->spinlock);
1554 mlog(ML_ERROR, "Error return %d asserting "
1555 "join on node %d\n", status, node);
1557 /* give us some time between errors... */
1559 msleep(DLM_DOMAIN_BACKOFF_MS);
1561 } while (status && live);
1565 struct domain_join_ctxt {
1566 unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1567 unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1570 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
1571 struct domain_join_ctxt *ctxt,
1572 enum dlm_query_join_response_code response)
1576 if (response == JOIN_DISALLOW) {
1577 mlog(0, "Latest response of disallow -- should restart\n");
1581 spin_lock(&dlm->spinlock);
1582 /* For now, we restart the process if the node maps have
1584 ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
1585 sizeof(dlm->live_nodes_map));
1586 spin_unlock(&dlm->spinlock);
1589 mlog(0, "Node maps changed -- should restart\n");
1594 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
1596 int status = 0, tmpstat, node;
1597 struct domain_join_ctxt *ctxt;
1598 enum dlm_query_join_response_code response = JOIN_DISALLOW;
1602 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
1609 /* group sem locking should work for us here -- we're already
1610 * registered for heartbeat events so filling this should be
1611 * atomic wrt getting those handlers called. */
1612 o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
1614 spin_lock(&dlm->spinlock);
1615 memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
1617 __dlm_set_joining_node(dlm, dlm->node_num);
1619 spin_unlock(&dlm->spinlock);
1622 while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
1623 node + 1)) < O2NM_MAX_NODES) {
1624 if (node == dlm->node_num)
1627 status = dlm_request_join(dlm, node, &response);
1633 /* Ok, either we got a response or the node doesn't have a
1635 if (response == JOIN_OK)
1636 set_bit(node, ctxt->yes_resp_map);
1638 if (dlm_should_restart_join(dlm, ctxt, response)) {
1644 mlog(0, "Yay, done querying nodes!\n");
1646 /* Yay, everyone agree's we can join the domain. My domain is
1647 * comprised of all nodes who were put in the
1648 * yes_resp_map. Copy that into our domain map and send a join
1649 * assert message to clean up everyone elses state. */
1650 spin_lock(&dlm->spinlock);
1651 memcpy(dlm->domain_map, ctxt->yes_resp_map,
1652 sizeof(ctxt->yes_resp_map));
1653 set_bit(dlm->node_num, dlm->domain_map);
1654 spin_unlock(&dlm->spinlock);
1656 /* Support for global heartbeat and node info was added in 1.1 */
1657 if (dlm->dlm_locking_proto.pv_major > 1 ||
1658 dlm->dlm_locking_proto.pv_minor > 0) {
1659 status = dlm_send_nodeinfo(dlm, ctxt->yes_resp_map);
1664 status = dlm_send_regions(dlm, ctxt->yes_resp_map);
1671 dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1673 /* Joined state *must* be set before the joining node
1674 * information, otherwise the query_join handler may read no
1675 * current joiner but a state of NEW and tell joining nodes
1676 * we're not in the domain. */
1677 spin_lock(&dlm_domain_lock);
1678 dlm->dlm_state = DLM_CTXT_JOINED;
1680 spin_unlock(&dlm_domain_lock);
1683 spin_lock(&dlm->spinlock);
1684 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1686 printk(KERN_NOTICE "o2dlm: Joining domain %s ", dlm->name);
1687 __dlm_print_nodes(dlm);
1689 spin_unlock(&dlm->spinlock);
1692 /* Do we need to send a cancel message to any nodes? */
1694 tmpstat = dlm_send_join_cancels(dlm,
1696 sizeof(ctxt->yes_resp_map));
1698 mlog_errno(tmpstat);
1703 mlog(0, "returning %d\n", status);
1707 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1709 o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_up);
1710 o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_down);
1711 o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1714 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1718 mlog(0, "registering handlers.\n");
1720 o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1721 dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1722 o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1723 dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1725 status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_down);
1729 status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_up);
1733 status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1734 sizeof(struct dlm_master_request),
1735 dlm_master_request_handler,
1736 dlm, NULL, &dlm->dlm_domain_handlers);
1740 status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1741 sizeof(struct dlm_assert_master),
1742 dlm_assert_master_handler,
1743 dlm, dlm_assert_master_post_handler,
1744 &dlm->dlm_domain_handlers);
1748 status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1749 sizeof(struct dlm_create_lock),
1750 dlm_create_lock_handler,
1751 dlm, NULL, &dlm->dlm_domain_handlers);
1755 status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1756 DLM_CONVERT_LOCK_MAX_LEN,
1757 dlm_convert_lock_handler,
1758 dlm, NULL, &dlm->dlm_domain_handlers);
1762 status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1763 DLM_UNLOCK_LOCK_MAX_LEN,
1764 dlm_unlock_lock_handler,
1765 dlm, NULL, &dlm->dlm_domain_handlers);
1769 status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1770 DLM_PROXY_AST_MAX_LEN,
1771 dlm_proxy_ast_handler,
1772 dlm, NULL, &dlm->dlm_domain_handlers);
1776 status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1777 sizeof(struct dlm_exit_domain),
1778 dlm_exit_domain_handler,
1779 dlm, NULL, &dlm->dlm_domain_handlers);
1783 status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1784 sizeof(struct dlm_deref_lockres),
1785 dlm_deref_lockres_handler,
1786 dlm, NULL, &dlm->dlm_domain_handlers);
1790 status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1791 sizeof(struct dlm_migrate_request),
1792 dlm_migrate_request_handler,
1793 dlm, NULL, &dlm->dlm_domain_handlers);
1797 status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1798 DLM_MIG_LOCKRES_MAX_LEN,
1799 dlm_mig_lockres_handler,
1800 dlm, NULL, &dlm->dlm_domain_handlers);
1804 status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1805 sizeof(struct dlm_master_requery),
1806 dlm_master_requery_handler,
1807 dlm, NULL, &dlm->dlm_domain_handlers);
1811 status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1812 sizeof(struct dlm_lock_request),
1813 dlm_request_all_locks_handler,
1814 dlm, NULL, &dlm->dlm_domain_handlers);
1818 status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1819 sizeof(struct dlm_reco_data_done),
1820 dlm_reco_data_done_handler,
1821 dlm, NULL, &dlm->dlm_domain_handlers);
1825 status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1826 sizeof(struct dlm_begin_reco),
1827 dlm_begin_reco_handler,
1828 dlm, NULL, &dlm->dlm_domain_handlers);
1832 status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1833 sizeof(struct dlm_finalize_reco),
1834 dlm_finalize_reco_handler,
1835 dlm, NULL, &dlm->dlm_domain_handlers);
1839 status = o2net_register_handler(DLM_BEGIN_EXIT_DOMAIN_MSG, dlm->key,
1840 sizeof(struct dlm_exit_domain),
1841 dlm_begin_exit_domain_handler,
1842 dlm, NULL, &dlm->dlm_domain_handlers);
1846 status = o2net_register_handler(DLM_DEREF_LOCKRES_DONE, dlm->key,
1847 sizeof(struct dlm_deref_lockres_done),
1848 dlm_deref_lockres_done_handler,
1849 dlm, NULL, &dlm->dlm_domain_handlers);
1852 dlm_unregister_domain_handlers(dlm);
1857 static int dlm_join_domain(struct dlm_ctxt *dlm)
1860 unsigned int backoff;
1861 unsigned int total_backoff = 0;
1862 char wq_name[O2NM_MAX_NAME_LEN];
1866 mlog(0, "Join domain %s\n", dlm->name);
1868 status = dlm_register_domain_handlers(dlm);
1874 status = dlm_launch_thread(dlm);
1880 status = dlm_launch_recovery_thread(dlm);
1886 status = dlm_debug_init(dlm);
1892 snprintf(wq_name, O2NM_MAX_NAME_LEN, "dlm_wq-%s", dlm->name);
1893 dlm->dlm_worker = alloc_workqueue(wq_name, WQ_MEM_RECLAIM, 0);
1894 if (!dlm->dlm_worker) {
1901 status = dlm_try_to_join_domain(dlm);
1903 /* If we're racing another node to the join, then we
1904 * need to back off temporarily and let them
1906 #define DLM_JOIN_TIMEOUT_MSECS 90000
1907 if (status == -EAGAIN) {
1908 if (signal_pending(current)) {
1909 status = -ERESTARTSYS;
1913 if (total_backoff > DLM_JOIN_TIMEOUT_MSECS) {
1914 status = -ERESTARTSYS;
1915 mlog(ML_NOTICE, "Timed out joining dlm domain "
1916 "%s after %u msecs\n", dlm->name,
1923 * <dale> No, after you!
1925 * <dale> But you first!
1928 backoff = (unsigned int)(jiffies & 0x3);
1929 backoff *= DLM_DOMAIN_BACKOFF_MS;
1930 total_backoff += backoff;
1931 mlog(0, "backoff %d\n", backoff);
1934 } while (status == -EAGAIN);
1943 wake_up(&dlm_domain_events);
1946 dlm_unregister_domain_handlers(dlm);
1947 dlm_debug_shutdown(dlm);
1948 dlm_complete_thread(dlm);
1949 dlm_complete_recovery_thread(dlm);
1950 dlm_destroy_dlm_worker(dlm);
1956 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1961 struct dlm_ctxt *dlm = NULL;
1963 dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1970 dlm->name = kstrdup(domain, GFP_KERNEL);
1971 if (dlm->name == NULL) {
1977 dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1978 if (!dlm->lockres_hash) {
1984 for (i = 0; i < DLM_HASH_BUCKETS; i++)
1985 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1987 dlm->master_hash = (struct hlist_head **)
1988 dlm_alloc_pagevec(DLM_HASH_PAGES);
1989 if (!dlm->master_hash) {
1995 for (i = 0; i < DLM_HASH_BUCKETS; i++)
1996 INIT_HLIST_HEAD(dlm_master_hash(dlm, i));
1999 dlm->node_num = o2nm_this_node();
2001 ret = dlm_create_debugfs_subroot(dlm);
2005 spin_lock_init(&dlm->spinlock);
2006 spin_lock_init(&dlm->master_lock);
2007 spin_lock_init(&dlm->ast_lock);
2008 spin_lock_init(&dlm->track_lock);
2009 INIT_LIST_HEAD(&dlm->list);
2010 INIT_LIST_HEAD(&dlm->dirty_list);
2011 INIT_LIST_HEAD(&dlm->reco.resources);
2012 INIT_LIST_HEAD(&dlm->reco.node_data);
2013 INIT_LIST_HEAD(&dlm->purge_list);
2014 INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
2015 INIT_LIST_HEAD(&dlm->tracking_list);
2016 dlm->reco.state = 0;
2018 INIT_LIST_HEAD(&dlm->pending_asts);
2019 INIT_LIST_HEAD(&dlm->pending_basts);
2021 mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
2022 dlm->recovery_map, &(dlm->recovery_map[0]));
2024 memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
2025 memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
2026 memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
2028 dlm->dlm_thread_task = NULL;
2029 dlm->dlm_reco_thread_task = NULL;
2030 dlm->dlm_worker = NULL;
2031 init_waitqueue_head(&dlm->dlm_thread_wq);
2032 init_waitqueue_head(&dlm->dlm_reco_thread_wq);
2033 init_waitqueue_head(&dlm->reco.event);
2034 init_waitqueue_head(&dlm->ast_wq);
2035 init_waitqueue_head(&dlm->migration_wq);
2036 INIT_LIST_HEAD(&dlm->mle_hb_events);
2038 dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
2039 init_waitqueue_head(&dlm->dlm_join_events);
2041 dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
2042 dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
2044 atomic_set(&dlm->res_tot_count, 0);
2045 atomic_set(&dlm->res_cur_count, 0);
2046 for (i = 0; i < DLM_MLE_NUM_TYPES; ++i) {
2047 atomic_set(&dlm->mle_tot_count[i], 0);
2048 atomic_set(&dlm->mle_cur_count[i], 0);
2051 spin_lock_init(&dlm->work_lock);
2052 INIT_LIST_HEAD(&dlm->work_list);
2053 INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
2055 kref_init(&dlm->dlm_refs);
2056 dlm->dlm_state = DLM_CTXT_NEW;
2058 INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
2060 mlog(0, "context init: refcount %u\n",
2061 atomic_read(&dlm->dlm_refs.refcount));
2064 if (ret < 0 && dlm) {
2065 if (dlm->master_hash)
2066 dlm_free_pagevec((void **)dlm->master_hash,
2069 if (dlm->lockres_hash)
2070 dlm_free_pagevec((void **)dlm->lockres_hash,
2081 * Compare a requested locking protocol version against the current one.
2083 * If the major numbers are different, they are incompatible.
2084 * If the current minor is greater than the request, they are incompatible.
2085 * If the current minor is less than or equal to the request, they are
2086 * compatible, and the requester should run at the current minor version.
2088 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
2089 struct dlm_protocol_version *request)
2091 if (existing->pv_major != request->pv_major)
2094 if (existing->pv_minor > request->pv_minor)
2097 if (existing->pv_minor < request->pv_minor)
2098 request->pv_minor = existing->pv_minor;
2104 * dlm_register_domain: one-time setup per "domain".
2106 * The filesystem passes in the requested locking version via proto.
2107 * If registration was successful, proto will contain the negotiated
2110 struct dlm_ctxt * dlm_register_domain(const char *domain,
2112 struct dlm_protocol_version *fs_proto)
2115 struct dlm_ctxt *dlm = NULL;
2116 struct dlm_ctxt *new_ctxt = NULL;
2118 if (strlen(domain) >= O2NM_MAX_NAME_LEN) {
2119 ret = -ENAMETOOLONG;
2120 mlog(ML_ERROR, "domain name length too long\n");
2124 mlog(0, "register called for domain \"%s\"\n", domain);
2128 if (signal_pending(current)) {
2134 spin_lock(&dlm_domain_lock);
2136 dlm = __dlm_lookup_domain(domain);
2138 if (dlm->dlm_state != DLM_CTXT_JOINED) {
2139 spin_unlock(&dlm_domain_lock);
2141 mlog(0, "This ctxt is not joined yet!\n");
2142 wait_event_interruptible(dlm_domain_events,
2143 dlm_wait_on_domain_helper(
2148 if (dlm_protocol_compare(&dlm->fs_locking_proto, fs_proto)) {
2149 spin_unlock(&dlm_domain_lock);
2151 "Requested locking protocol version is not "
2152 "compatible with already registered domain "
2153 "\"%s\"\n", domain);
2161 spin_unlock(&dlm_domain_lock);
2169 spin_unlock(&dlm_domain_lock);
2171 new_ctxt = dlm_alloc_ctxt(domain, key);
2180 /* a little variable switch-a-roo here... */
2184 /* add the new domain */
2185 list_add_tail(&dlm->list, &dlm_domains);
2186 spin_unlock(&dlm_domain_lock);
2189 * Pass the locking protocol version into the join. If the join
2190 * succeeds, it will have the negotiated protocol set.
2192 dlm->dlm_locking_proto = dlm_protocol;
2193 dlm->fs_locking_proto = *fs_proto;
2195 ret = dlm_join_domain(dlm);
2202 /* Tell the caller what locking protocol we negotiated */
2203 *fs_proto = dlm->fs_locking_proto;
2208 dlm_free_ctxt_mem(new_ctxt);
2215 EXPORT_SYMBOL_GPL(dlm_register_domain);
2217 static LIST_HEAD(dlm_join_handlers);
2219 static void dlm_unregister_net_handlers(void)
2221 o2net_unregister_handler_list(&dlm_join_handlers);
2224 static int dlm_register_net_handlers(void)
2228 status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
2229 sizeof(struct dlm_query_join_request),
2230 dlm_query_join_handler,
2231 NULL, NULL, &dlm_join_handlers);
2235 status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
2236 sizeof(struct dlm_assert_joined),
2237 dlm_assert_joined_handler,
2238 NULL, NULL, &dlm_join_handlers);
2242 status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
2243 sizeof(struct dlm_cancel_join),
2244 dlm_cancel_join_handler,
2245 NULL, NULL, &dlm_join_handlers);
2249 status = o2net_register_handler(DLM_QUERY_REGION, DLM_MOD_KEY,
2250 sizeof(struct dlm_query_region),
2251 dlm_query_region_handler,
2252 NULL, NULL, &dlm_join_handlers);
2257 status = o2net_register_handler(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
2258 sizeof(struct dlm_query_nodeinfo),
2259 dlm_query_nodeinfo_handler,
2260 NULL, NULL, &dlm_join_handlers);
2263 dlm_unregister_net_handlers();
2268 /* Domain eviction callback handling.
2270 * The file system requires notification of node death *before* the
2271 * dlm completes it's recovery work, otherwise it may be able to
2272 * acquire locks on resources requiring recovery. Since the dlm can
2273 * evict a node from it's domain *before* heartbeat fires, a similar
2274 * mechanism is required. */
2276 /* Eviction is not expected to happen often, so a per-domain lock is
2277 * not necessary. Eviction callbacks are allowed to sleep for short
2278 * periods of time. */
2279 static DECLARE_RWSEM(dlm_callback_sem);
2281 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
2284 struct dlm_eviction_cb *cb;
2286 down_read(&dlm_callback_sem);
2287 list_for_each_entry(cb, &dlm->dlm_eviction_callbacks, ec_item) {
2288 cb->ec_func(node_num, cb->ec_data);
2290 up_read(&dlm_callback_sem);
2293 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
2294 dlm_eviction_func *f,
2297 INIT_LIST_HEAD(&cb->ec_item);
2301 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
2303 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
2304 struct dlm_eviction_cb *cb)
2306 down_write(&dlm_callback_sem);
2307 list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
2308 up_write(&dlm_callback_sem);
2310 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
2312 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
2314 down_write(&dlm_callback_sem);
2315 list_del_init(&cb->ec_item);
2316 up_write(&dlm_callback_sem);
2318 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
2320 static int __init dlm_init(void)
2324 status = dlm_init_mle_cache();
2326 mlog(ML_ERROR, "Could not create o2dlm_mle slabcache\n");
2330 status = dlm_init_master_caches();
2332 mlog(ML_ERROR, "Could not create o2dlm_lockres and "
2333 "o2dlm_lockname slabcaches\n");
2337 status = dlm_init_lock_cache();
2339 mlog(ML_ERROR, "Count not create o2dlm_lock slabcache\n");
2343 status = dlm_register_net_handlers();
2345 mlog(ML_ERROR, "Unable to register network handlers\n");
2349 status = dlm_create_debugfs_root();
2355 dlm_unregister_net_handlers();
2356 dlm_destroy_lock_cache();
2357 dlm_destroy_master_caches();
2358 dlm_destroy_mle_cache();
2362 static void __exit dlm_exit (void)
2364 dlm_destroy_debugfs_root();
2365 dlm_unregister_net_handlers();
2366 dlm_destroy_lock_cache();
2367 dlm_destroy_master_caches();
2368 dlm_destroy_mle_cache();
2371 MODULE_AUTHOR("Oracle");
2372 MODULE_LICENSE("GPL");
2373 MODULE_DESCRIPTION("OCFS2 Distributed Lock Management");
2375 module_init(dlm_init);
2376 module_exit(dlm_exit);