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 hlist_node *list;
198 mlog(0, "%.*s\n", len, name);
200 assert_spin_locked(&dlm->spinlock);
202 bucket = dlm_lockres_hash(dlm, hash);
204 hlist_for_each(list, bucket) {
205 struct dlm_lock_resource *res = hlist_entry(list,
206 struct dlm_lock_resource, hash_node);
207 if (res->lockname.name[0] != name[0])
209 if (unlikely(res->lockname.len != len))
211 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
213 dlm_lockres_get(res);
219 /* intended to be called by functions which do not care about lock
220 * resources which are being purged (most net _handler functions).
221 * this will return NULL for any lock resource which is found but
222 * currently in the process of dropping its mastery reference.
223 * use __dlm_lookup_lockres_full when you need the lock resource
224 * regardless (e.g. dlm_get_lock_resource) */
225 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
230 struct dlm_lock_resource *res = NULL;
232 mlog(0, "%.*s\n", len, name);
234 assert_spin_locked(&dlm->spinlock);
236 res = __dlm_lookup_lockres_full(dlm, name, len, hash);
238 spin_lock(&res->spinlock);
239 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
240 spin_unlock(&res->spinlock);
241 dlm_lockres_put(res);
244 spin_unlock(&res->spinlock);
250 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
254 struct dlm_lock_resource *res;
255 unsigned int hash = dlm_lockid_hash(name, len);
257 spin_lock(&dlm->spinlock);
258 res = __dlm_lookup_lockres(dlm, name, len, hash);
259 spin_unlock(&dlm->spinlock);
263 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
265 struct dlm_ctxt *tmp = NULL;
266 struct list_head *iter;
268 assert_spin_locked(&dlm_domain_lock);
270 /* tmp->name here is always NULL terminated,
271 * but domain may not be! */
272 list_for_each(iter, &dlm_domains) {
273 tmp = list_entry (iter, struct dlm_ctxt, list);
274 if (strlen(tmp->name) == len &&
275 memcmp(tmp->name, domain, len)==0)
283 /* For null terminated domain strings ONLY */
284 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
286 assert_spin_locked(&dlm_domain_lock);
288 return __dlm_lookup_domain_full(domain, strlen(domain));
292 /* returns true on one of two conditions:
293 * 1) the domain does not exist
294 * 2) the domain exists and it's state is "joined" */
295 static int dlm_wait_on_domain_helper(const char *domain)
298 struct dlm_ctxt *tmp = NULL;
300 spin_lock(&dlm_domain_lock);
302 tmp = __dlm_lookup_domain(domain);
305 else if (tmp->dlm_state == DLM_CTXT_JOINED)
308 spin_unlock(&dlm_domain_lock);
312 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
314 dlm_destroy_debugfs_subroot(dlm);
316 if (dlm->lockres_hash)
317 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
319 if (dlm->master_hash)
320 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
328 /* A little strange - this function will be called while holding
329 * dlm_domain_lock and is expected to be holding it on the way out. We
330 * will however drop and reacquire it multiple times */
331 static void dlm_ctxt_release(struct kref *kref)
333 struct dlm_ctxt *dlm;
335 dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
337 BUG_ON(dlm->num_joins);
338 BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
340 /* we may still be in the list if we hit an error during join. */
341 list_del_init(&dlm->list);
343 spin_unlock(&dlm_domain_lock);
345 mlog(0, "freeing memory from domain %s\n", dlm->name);
347 wake_up(&dlm_domain_events);
349 dlm_free_ctxt_mem(dlm);
351 spin_lock(&dlm_domain_lock);
354 void dlm_put(struct dlm_ctxt *dlm)
356 spin_lock(&dlm_domain_lock);
357 kref_put(&dlm->dlm_refs, dlm_ctxt_release);
358 spin_unlock(&dlm_domain_lock);
361 static void __dlm_get(struct dlm_ctxt *dlm)
363 kref_get(&dlm->dlm_refs);
366 /* given a questionable reference to a dlm object, gets a reference if
367 * it can find it in the list, otherwise returns NULL in which case
368 * you shouldn't trust your pointer. */
369 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
371 struct list_head *iter;
372 struct dlm_ctxt *target = NULL;
374 spin_lock(&dlm_domain_lock);
376 list_for_each(iter, &dlm_domains) {
377 target = list_entry (iter, struct dlm_ctxt, list);
387 spin_unlock(&dlm_domain_lock);
392 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
396 spin_lock(&dlm_domain_lock);
397 ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
398 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
399 spin_unlock(&dlm_domain_lock);
404 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
406 if (dlm->dlm_worker) {
407 flush_workqueue(dlm->dlm_worker);
408 destroy_workqueue(dlm->dlm_worker);
409 dlm->dlm_worker = NULL;
413 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
415 dlm_unregister_domain_handlers(dlm);
416 dlm_debug_shutdown(dlm);
417 dlm_complete_thread(dlm);
418 dlm_complete_recovery_thread(dlm);
419 dlm_destroy_dlm_worker(dlm);
421 /* We've left the domain. Now we can take ourselves out of the
422 * list and allow the kref stuff to help us free the
424 spin_lock(&dlm_domain_lock);
425 list_del_init(&dlm->list);
426 spin_unlock(&dlm_domain_lock);
428 /* Wake up anyone waiting for us to remove this domain */
429 wake_up(&dlm_domain_events);
432 static int dlm_migrate_all_locks(struct dlm_ctxt *dlm)
434 int i, num, n, ret = 0;
435 struct dlm_lock_resource *res;
436 struct hlist_node *iter;
437 struct hlist_head *bucket;
440 mlog(0, "Migrating locks from domain %s\n", dlm->name);
443 spin_lock(&dlm->spinlock);
444 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
447 bucket = dlm_lockres_hash(dlm, i);
448 iter = bucket->first;
451 res = hlist_entry(iter, struct dlm_lock_resource,
453 dlm_lockres_get(res);
454 /* migrate, if necessary. this will drop the dlm
455 * spinlock and retake it if it does migration. */
456 dropped = dlm_empty_lockres(dlm, res);
458 spin_lock(&res->spinlock);
460 __dlm_lockres_calc_usage(dlm, res);
462 iter = res->hash_node.next;
463 spin_unlock(&res->spinlock);
465 dlm_lockres_put(res);
468 cond_resched_lock(&dlm->spinlock);
472 cond_resched_lock(&dlm->spinlock);
475 spin_unlock(&dlm->spinlock);
476 wake_up(&dlm->dlm_thread_wq);
478 /* let the dlm thread take care of purging, keep scanning until
479 * nothing remains in the hash */
481 mlog(0, "%s: %d lock resources in hash last pass\n",
485 mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
489 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
493 spin_lock(&dlm->spinlock);
494 ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
495 spin_unlock(&dlm->spinlock);
500 static int dlm_begin_exit_domain_handler(struct o2net_msg *msg, u32 len,
501 void *data, void **ret_data)
503 struct dlm_ctxt *dlm = data;
505 struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
510 node = exit_msg->node_idx;
511 mlog(0, "%s: Node %u sent a begin exit domain message\n", dlm->name, node);
513 spin_lock(&dlm->spinlock);
514 set_bit(node, dlm->exit_domain_map);
515 spin_unlock(&dlm->spinlock);
522 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
524 /* Yikes, a double spinlock! I need domain_lock for the dlm
525 * state and the dlm spinlock for join state... Sorry! */
527 spin_lock(&dlm_domain_lock);
528 spin_lock(&dlm->spinlock);
530 if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
531 mlog(0, "Node %d is joining, we wait on it.\n",
533 spin_unlock(&dlm->spinlock);
534 spin_unlock(&dlm_domain_lock);
536 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
540 dlm->dlm_state = DLM_CTXT_LEAVING;
541 spin_unlock(&dlm->spinlock);
542 spin_unlock(&dlm_domain_lock);
545 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
547 int node = -1, num = 0;
549 assert_spin_locked(&dlm->spinlock);
552 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
553 node + 1)) < O2NM_MAX_NODES) {
557 printk(") %u nodes\n", num);
560 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
563 struct dlm_ctxt *dlm = data;
565 struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
567 mlog(0, "%p %u %p", msg, len, data);
572 node = exit_msg->node_idx;
574 spin_lock(&dlm->spinlock);
575 clear_bit(node, dlm->domain_map);
576 clear_bit(node, dlm->exit_domain_map);
577 printk(KERN_NOTICE "o2dlm: Node %u leaves domain %s ", node, dlm->name);
578 __dlm_print_nodes(dlm);
580 /* notify anything attached to the heartbeat events */
581 dlm_hb_event_notify_attached(dlm, node, 0);
583 spin_unlock(&dlm->spinlock);
590 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm, u32 msg_type,
594 struct dlm_exit_domain leave_msg;
596 mlog(0, "%s: Sending domain exit message %u to node %u\n", dlm->name,
599 memset(&leave_msg, 0, sizeof(leave_msg));
600 leave_msg.node_idx = dlm->node_num;
602 status = o2net_send_message(msg_type, dlm->key, &leave_msg,
603 sizeof(leave_msg), node, NULL);
605 mlog(ML_ERROR, "Error %d sending domain exit message %u "
606 "to node %u on domain %s\n", status, msg_type, node,
612 static void dlm_begin_exit_domain(struct dlm_ctxt *dlm)
616 /* Support for begin exit domain was added in 1.2 */
617 if (dlm->dlm_locking_proto.pv_major == 1 &&
618 dlm->dlm_locking_proto.pv_minor < 2)
622 * Unlike DLM_EXIT_DOMAIN_MSG, DLM_BEGIN_EXIT_DOMAIN_MSG is purely
623 * informational. Meaning if a node does not receive the message,
626 spin_lock(&dlm->spinlock);
628 node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, node + 1);
629 if (node >= O2NM_MAX_NODES)
631 if (node == dlm->node_num)
634 spin_unlock(&dlm->spinlock);
635 dlm_send_one_domain_exit(dlm, DLM_BEGIN_EXIT_DOMAIN_MSG, node);
636 spin_lock(&dlm->spinlock);
638 spin_unlock(&dlm->spinlock);
641 static void dlm_leave_domain(struct dlm_ctxt *dlm)
643 int node, clear_node, status;
645 /* At this point we've migrated away all our locks and won't
646 * accept mastership of new ones. The dlm is responsible for
647 * almost nothing now. We make sure not to confuse any joining
648 * nodes and then commence shutdown procedure. */
650 spin_lock(&dlm->spinlock);
651 /* Clear ourselves from the domain map */
652 clear_bit(dlm->node_num, dlm->domain_map);
653 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
654 0)) < O2NM_MAX_NODES) {
655 /* Drop the dlm spinlock. This is safe wrt the domain_map.
656 * -nodes cannot be added now as the
657 * query_join_handlers knows to respond with OK_NO_MAP
658 * -we catch the right network errors if a node is
659 * removed from the map while we're sending him the
661 spin_unlock(&dlm->spinlock);
665 status = dlm_send_one_domain_exit(dlm, DLM_EXIT_DOMAIN_MSG,
668 status != -ENOPROTOOPT &&
669 status != -ENOTCONN) {
670 mlog(ML_NOTICE, "Error %d sending domain exit message "
671 "to node %d\n", status, node);
673 /* Not sure what to do here but lets sleep for
674 * a bit in case this was a transient
676 msleep(DLM_DOMAIN_BACKOFF_MS);
680 spin_lock(&dlm->spinlock);
681 /* If we're not clearing the node bit then we intend
682 * to loop back around to try again. */
684 clear_bit(node, dlm->domain_map);
686 spin_unlock(&dlm->spinlock);
689 int dlm_joined(struct dlm_ctxt *dlm)
693 spin_lock(&dlm_domain_lock);
695 if (dlm->dlm_state == DLM_CTXT_JOINED)
698 spin_unlock(&dlm_domain_lock);
703 int dlm_shutting_down(struct dlm_ctxt *dlm)
707 spin_lock(&dlm_domain_lock);
709 if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
712 spin_unlock(&dlm_domain_lock);
717 void dlm_unregister_domain(struct dlm_ctxt *dlm)
720 struct dlm_lock_resource *res;
722 spin_lock(&dlm_domain_lock);
723 BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
724 BUG_ON(!dlm->num_joins);
727 if (!dlm->num_joins) {
728 /* We mark it "in shutdown" now so new register
729 * requests wait until we've completely left the
730 * domain. Don't use DLM_CTXT_LEAVING yet as we still
731 * want new domain joins to communicate with us at
732 * least until we've completed migration of our
734 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
737 spin_unlock(&dlm_domain_lock);
740 mlog(0, "shutting down domain %s\n", dlm->name);
741 dlm_begin_exit_domain(dlm);
743 /* We changed dlm state, notify the thread */
744 dlm_kick_thread(dlm, NULL);
746 while (dlm_migrate_all_locks(dlm)) {
747 /* Give dlm_thread time to purge the lockres' */
749 mlog(0, "%s: more migration to do\n", dlm->name);
752 /* This list should be empty. If not, print remaining lockres */
753 if (!list_empty(&dlm->tracking_list)) {
754 mlog(ML_ERROR, "Following lockres' are still on the "
756 list_for_each_entry(res, &dlm->tracking_list, tracking)
757 dlm_print_one_lock_resource(res);
760 dlm_mark_domain_leaving(dlm);
761 dlm_leave_domain(dlm);
762 printk(KERN_NOTICE "o2dlm: Leaving domain %s\n", dlm->name);
763 dlm_force_free_mles(dlm);
764 dlm_complete_dlm_shutdown(dlm);
768 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
770 static int dlm_query_join_proto_check(char *proto_type, int node,
771 struct dlm_protocol_version *ours,
772 struct dlm_protocol_version *request)
775 struct dlm_protocol_version proto = *request;
777 if (!dlm_protocol_compare(ours, &proto)) {
779 "node %u wanted to join with %s locking protocol "
780 "%u.%u, we respond with %u.%u\n",
784 proto.pv_major, proto.pv_minor);
785 request->pv_minor = proto.pv_minor;
789 "Node %u wanted to join with %s locking "
790 "protocol %u.%u, but we have %u.%u, disallowing\n",
803 * struct dlm_query_join_packet is made up of four one-byte fields. They
804 * are effectively in big-endian order already. However, little-endian
805 * machines swap them before putting the packet on the wire (because
806 * query_join's response is a status, and that status is treated as a u32
807 * on the wire). Thus, a big-endian and little-endian machines will treat
808 * this structure differently.
810 * The solution is to have little-endian machines swap the structure when
811 * converting from the structure to the u32 representation. This will
812 * result in the structure having the correct format on the wire no matter
813 * the host endian format.
815 static void dlm_query_join_packet_to_wire(struct dlm_query_join_packet *packet,
818 union dlm_query_join_response response;
820 response.packet = *packet;
821 *wire = be32_to_cpu(response.intval);
824 static void dlm_query_join_wire_to_packet(u32 wire,
825 struct dlm_query_join_packet *packet)
827 union dlm_query_join_response response;
829 response.intval = cpu_to_be32(wire);
830 *packet = response.packet;
833 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
836 struct dlm_query_join_request *query;
837 struct dlm_query_join_packet packet = {
838 .code = JOIN_DISALLOW,
840 struct dlm_ctxt *dlm = NULL;
844 query = (struct dlm_query_join_request *) msg->buf;
846 mlog(0, "node %u wants to join domain %s\n", query->node_idx,
850 * If heartbeat doesn't consider the node live, tell it
851 * to back off and try again. This gives heartbeat a chance
854 if (!o2hb_check_node_heartbeating(query->node_idx)) {
855 mlog(0, "node %u is not in our live map yet\n",
858 packet.code = JOIN_DISALLOW;
862 packet.code = JOIN_OK_NO_MAP;
864 spin_lock(&dlm_domain_lock);
865 dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
870 * There is a small window where the joining node may not see the
871 * node(s) that just left but still part of the cluster. DISALLOW
872 * join request if joining node has different node map.
875 while (nodenum < O2NM_MAX_NODES) {
876 if (test_bit(nodenum, dlm->domain_map)) {
877 if (!byte_test_bit(nodenum, query->node_map)) {
878 mlog(0, "disallow join as node %u does not "
879 "have node %u in its nodemap\n",
880 query->node_idx, nodenum);
881 packet.code = JOIN_DISALLOW;
888 /* Once the dlm ctxt is marked as leaving then we don't want
889 * to be put in someone's domain map.
890 * Also, explicitly disallow joining at certain troublesome
891 * times (ie. during recovery). */
892 if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
893 int bit = query->node_idx;
894 spin_lock(&dlm->spinlock);
896 if (dlm->dlm_state == DLM_CTXT_NEW &&
897 dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
898 /*If this is a brand new context and we
899 * haven't started our join process yet, then
900 * the other node won the race. */
901 packet.code = JOIN_OK_NO_MAP;
902 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
903 /* Disallow parallel joins. */
904 packet.code = JOIN_DISALLOW;
905 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
906 mlog(0, "node %u trying to join, but recovery "
907 "is ongoing.\n", bit);
908 packet.code = JOIN_DISALLOW;
909 } else if (test_bit(bit, dlm->recovery_map)) {
910 mlog(0, "node %u trying to join, but it "
911 "still needs recovery.\n", bit);
912 packet.code = JOIN_DISALLOW;
913 } else if (test_bit(bit, dlm->domain_map)) {
914 mlog(0, "node %u trying to join, but it "
915 "is still in the domain! needs recovery?\n",
917 packet.code = JOIN_DISALLOW;
919 /* Alright we're fully a part of this domain
920 * so we keep some state as to who's joining
921 * and indicate to him that needs to be fixed
924 /* Make sure we speak compatible locking protocols. */
925 if (dlm_query_join_proto_check("DLM", bit,
926 &dlm->dlm_locking_proto,
927 &query->dlm_proto)) {
928 packet.code = JOIN_PROTOCOL_MISMATCH;
929 } else if (dlm_query_join_proto_check("fs", bit,
930 &dlm->fs_locking_proto,
932 packet.code = JOIN_PROTOCOL_MISMATCH;
934 packet.dlm_minor = query->dlm_proto.pv_minor;
935 packet.fs_minor = query->fs_proto.pv_minor;
936 packet.code = JOIN_OK;
937 __dlm_set_joining_node(dlm, query->node_idx);
941 spin_unlock(&dlm->spinlock);
944 spin_unlock(&dlm_domain_lock);
947 mlog(0, "We respond with %u\n", packet.code);
949 dlm_query_join_packet_to_wire(&packet, &response);
953 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
956 struct dlm_assert_joined *assert;
957 struct dlm_ctxt *dlm = NULL;
959 assert = (struct dlm_assert_joined *) msg->buf;
961 mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
964 spin_lock(&dlm_domain_lock);
965 dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
966 /* XXX should we consider no dlm ctxt an error? */
968 spin_lock(&dlm->spinlock);
970 /* Alright, this node has officially joined our
971 * domain. Set him in the map and clean up our
972 * leftover join state. */
973 BUG_ON(dlm->joining_node != assert->node_idx);
974 set_bit(assert->node_idx, dlm->domain_map);
975 clear_bit(assert->node_idx, dlm->exit_domain_map);
976 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
978 printk(KERN_NOTICE "o2dlm: Node %u joins domain %s ",
979 assert->node_idx, dlm->name);
980 __dlm_print_nodes(dlm);
982 /* notify anything attached to the heartbeat events */
983 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
985 spin_unlock(&dlm->spinlock);
987 spin_unlock(&dlm_domain_lock);
992 static int dlm_match_regions(struct dlm_ctxt *dlm,
993 struct dlm_query_region *qr,
994 char *local, int locallen)
996 char *remote = qr->qr_regions;
998 int localnr, i, j, foundit;
1001 if (!o2hb_global_heartbeat_active()) {
1002 if (qr->qr_numregions) {
1003 mlog(ML_ERROR, "Domain %s: Joining node %d has global "
1004 "heartbeat enabled but local node %d does not\n",
1005 qr->qr_domain, qr->qr_node, dlm->node_num);
1011 if (o2hb_global_heartbeat_active() && !qr->qr_numregions) {
1012 mlog(ML_ERROR, "Domain %s: Local node %d has global "
1013 "heartbeat enabled but joining node %d does not\n",
1014 qr->qr_domain, dlm->node_num, qr->qr_node);
1020 for (i = 0; i < qr->qr_numregions; ++i) {
1021 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, r);
1022 r += O2HB_MAX_REGION_NAME_LEN;
1025 localnr = min(O2NM_MAX_REGIONS, locallen/O2HB_MAX_REGION_NAME_LEN);
1026 localnr = o2hb_get_all_regions(local, (u8)localnr);
1028 /* compare local regions with remote */
1030 for (i = 0; i < localnr; ++i) {
1033 for (j = 0; j <= qr->qr_numregions; ++j) {
1034 if (!memcmp(l, r, O2HB_MAX_REGION_NAME_LEN)) {
1038 r += O2HB_MAX_REGION_NAME_LEN;
1042 mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1043 "in local node %d but not in joining node %d\n",
1044 qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, l,
1045 dlm->node_num, qr->qr_node);
1048 l += O2HB_MAX_REGION_NAME_LEN;
1051 /* compare remote with local regions */
1053 for (i = 0; i < qr->qr_numregions; ++i) {
1056 for (j = 0; j < localnr; ++j) {
1057 if (!memcmp(r, l, O2HB_MAX_REGION_NAME_LEN)) {
1061 l += O2HB_MAX_REGION_NAME_LEN;
1065 mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1066 "in joining node %d but not in local node %d\n",
1067 qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, r,
1068 qr->qr_node, dlm->node_num);
1071 r += O2HB_MAX_REGION_NAME_LEN;
1078 static int dlm_send_regions(struct dlm_ctxt *dlm, unsigned long *node_map)
1080 struct dlm_query_region *qr = NULL;
1081 int status, ret = 0, i;
1084 if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1087 qr = kzalloc(sizeof(struct dlm_query_region), GFP_KERNEL);
1094 qr->qr_node = dlm->node_num;
1095 qr->qr_namelen = strlen(dlm->name);
1096 memcpy(qr->qr_domain, dlm->name, qr->qr_namelen);
1097 /* if local hb, the numregions will be zero */
1098 if (o2hb_global_heartbeat_active())
1099 qr->qr_numregions = o2hb_get_all_regions(qr->qr_regions,
1103 for (i = 0; i < qr->qr_numregions; ++i, p += O2HB_MAX_REGION_NAME_LEN)
1104 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, p);
1107 while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1108 i + 1)) < O2NM_MAX_NODES) {
1109 if (i == dlm->node_num)
1112 mlog(0, "Sending regions to node %d\n", i);
1114 ret = o2net_send_message(DLM_QUERY_REGION, DLM_MOD_KEY, qr,
1115 sizeof(struct dlm_query_region),
1120 mlog(ML_ERROR, "Region mismatch %d, node %d\n",
1131 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
1132 void *data, void **ret_data)
1134 struct dlm_query_region *qr;
1135 struct dlm_ctxt *dlm = NULL;
1140 qr = (struct dlm_query_region *) msg->buf;
1142 mlog(0, "Node %u queries hb regions on domain %s\n", qr->qr_node,
1145 /* buffer used in dlm_mast_regions() */
1146 local = kmalloc(sizeof(qr->qr_regions), GFP_KERNEL);
1154 spin_lock(&dlm_domain_lock);
1155 dlm = __dlm_lookup_domain_full(qr->qr_domain, qr->qr_namelen);
1157 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1158 "before join domain\n", qr->qr_node, qr->qr_domain);
1162 spin_lock(&dlm->spinlock);
1164 if (dlm->joining_node != qr->qr_node) {
1165 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1166 "but joining node is %d\n", qr->qr_node, qr->qr_domain,
1171 /* Support for global heartbeat was added in 1.1 */
1172 if (dlm->dlm_locking_proto.pv_major == 1 &&
1173 dlm->dlm_locking_proto.pv_minor == 0) {
1174 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1175 "but active dlm protocol is %d.%d\n", qr->qr_node,
1176 qr->qr_domain, dlm->dlm_locking_proto.pv_major,
1177 dlm->dlm_locking_proto.pv_minor);
1181 status = dlm_match_regions(dlm, qr, local, sizeof(qr->qr_regions));
1185 spin_unlock(&dlm->spinlock);
1186 spin_unlock(&dlm_domain_lock);
1193 static int dlm_match_nodes(struct dlm_ctxt *dlm, struct dlm_query_nodeinfo *qn)
1195 struct o2nm_node *local;
1196 struct dlm_node_info *remote;
1200 for (j = 0; j < qn->qn_numnodes; ++j)
1201 mlog(0, "Node %3d, %pI4:%u\n", qn->qn_nodes[j].ni_nodenum,
1202 &(qn->qn_nodes[j].ni_ipv4_address),
1203 ntohs(qn->qn_nodes[j].ni_ipv4_port));
1205 for (i = 0; i < O2NM_MAX_NODES && !status; ++i) {
1206 local = o2nm_get_node_by_num(i);
1208 for (j = 0; j < qn->qn_numnodes; ++j) {
1209 if (qn->qn_nodes[j].ni_nodenum == i) {
1210 remote = &(qn->qn_nodes[j]);
1215 if (!local && !remote)
1218 if ((local && !remote) || (!local && remote))
1222 ((remote->ni_nodenum != local->nd_num) ||
1223 (remote->ni_ipv4_port != local->nd_ipv4_port) ||
1224 (remote->ni_ipv4_address != local->nd_ipv4_address)))
1228 if (remote && !local)
1229 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1230 "registered in joining node %d but not in "
1231 "local node %d\n", qn->qn_domain,
1233 &(remote->ni_ipv4_address),
1234 ntohs(remote->ni_ipv4_port),
1235 qn->qn_nodenum, dlm->node_num);
1236 if (local && !remote)
1237 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1238 "registered in local node %d but not in "
1239 "joining node %d\n", qn->qn_domain,
1240 local->nd_num, &(local->nd_ipv4_address),
1241 ntohs(local->nd_ipv4_port),
1242 dlm->node_num, qn->qn_nodenum);
1243 BUG_ON((!local && !remote));
1247 o2nm_node_put(local);
1253 static int dlm_send_nodeinfo(struct dlm_ctxt *dlm, unsigned long *node_map)
1255 struct dlm_query_nodeinfo *qn = NULL;
1256 struct o2nm_node *node;
1257 int ret = 0, status, count, i;
1259 if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1262 qn = kzalloc(sizeof(struct dlm_query_nodeinfo), GFP_KERNEL);
1269 for (i = 0, count = 0; i < O2NM_MAX_NODES; ++i) {
1270 node = o2nm_get_node_by_num(i);
1273 qn->qn_nodes[count].ni_nodenum = node->nd_num;
1274 qn->qn_nodes[count].ni_ipv4_port = node->nd_ipv4_port;
1275 qn->qn_nodes[count].ni_ipv4_address = node->nd_ipv4_address;
1276 mlog(0, "Node %3d, %pI4:%u\n", node->nd_num,
1277 &(node->nd_ipv4_address), ntohs(node->nd_ipv4_port));
1279 o2nm_node_put(node);
1282 qn->qn_nodenum = dlm->node_num;
1283 qn->qn_numnodes = count;
1284 qn->qn_namelen = strlen(dlm->name);
1285 memcpy(qn->qn_domain, dlm->name, qn->qn_namelen);
1288 while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1289 i + 1)) < O2NM_MAX_NODES) {
1290 if (i == dlm->node_num)
1293 mlog(0, "Sending nodeinfo to node %d\n", i);
1295 ret = o2net_send_message(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
1296 qn, sizeof(struct dlm_query_nodeinfo),
1301 mlog(ML_ERROR, "node mismatch %d, node %d\n", ret, i);
1311 static int dlm_query_nodeinfo_handler(struct o2net_msg *msg, u32 len,
1312 void *data, void **ret_data)
1314 struct dlm_query_nodeinfo *qn;
1315 struct dlm_ctxt *dlm = NULL;
1316 int locked = 0, status = -EINVAL;
1318 qn = (struct dlm_query_nodeinfo *) msg->buf;
1320 mlog(0, "Node %u queries nodes on domain %s\n", qn->qn_nodenum,
1323 spin_lock(&dlm_domain_lock);
1324 dlm = __dlm_lookup_domain_full(qn->qn_domain, qn->qn_namelen);
1326 mlog(ML_ERROR, "Node %d queried nodes on domain %s before "
1327 "join domain\n", qn->qn_nodenum, qn->qn_domain);
1331 spin_lock(&dlm->spinlock);
1333 if (dlm->joining_node != qn->qn_nodenum) {
1334 mlog(ML_ERROR, "Node %d queried nodes on domain %s but "
1335 "joining node is %d\n", qn->qn_nodenum, qn->qn_domain,
1340 /* Support for node query was added in 1.1 */
1341 if (dlm->dlm_locking_proto.pv_major == 1 &&
1342 dlm->dlm_locking_proto.pv_minor == 0) {
1343 mlog(ML_ERROR, "Node %d queried nodes on domain %s "
1344 "but active dlm protocol is %d.%d\n", qn->qn_nodenum,
1345 qn->qn_domain, dlm->dlm_locking_proto.pv_major,
1346 dlm->dlm_locking_proto.pv_minor);
1350 status = dlm_match_nodes(dlm, qn);
1354 spin_unlock(&dlm->spinlock);
1355 spin_unlock(&dlm_domain_lock);
1360 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
1363 struct dlm_cancel_join *cancel;
1364 struct dlm_ctxt *dlm = NULL;
1366 cancel = (struct dlm_cancel_join *) msg->buf;
1368 mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
1371 spin_lock(&dlm_domain_lock);
1372 dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
1375 spin_lock(&dlm->spinlock);
1377 /* Yikes, this guy wants to cancel his join. No
1378 * problem, we simply cleanup our join state. */
1379 BUG_ON(dlm->joining_node != cancel->node_idx);
1380 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1382 spin_unlock(&dlm->spinlock);
1384 spin_unlock(&dlm_domain_lock);
1389 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
1393 struct dlm_cancel_join cancel_msg;
1395 memset(&cancel_msg, 0, sizeof(cancel_msg));
1396 cancel_msg.node_idx = dlm->node_num;
1397 cancel_msg.name_len = strlen(dlm->name);
1398 memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
1400 status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1401 &cancel_msg, sizeof(cancel_msg), node,
1404 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1405 "node %u\n", status, DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1414 /* map_size should be in bytes. */
1415 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
1416 unsigned long *node_map,
1417 unsigned int map_size)
1419 int status, tmpstat;
1422 if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
1423 sizeof(unsigned long))) {
1425 "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
1426 map_size, (unsigned)BITS_TO_LONGS(O2NM_MAX_NODES));
1432 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1433 node + 1)) < O2NM_MAX_NODES) {
1434 if (node == dlm->node_num)
1437 tmpstat = dlm_send_one_join_cancel(dlm, node);
1439 mlog(ML_ERROR, "Error return %d cancelling join on "
1440 "node %d\n", tmpstat, node);
1451 static int dlm_request_join(struct dlm_ctxt *dlm,
1453 enum dlm_query_join_response_code *response)
1456 struct dlm_query_join_request join_msg;
1457 struct dlm_query_join_packet packet;
1460 mlog(0, "querying node %d\n", node);
1462 memset(&join_msg, 0, sizeof(join_msg));
1463 join_msg.node_idx = dlm->node_num;
1464 join_msg.name_len = strlen(dlm->name);
1465 memcpy(join_msg.domain, dlm->name, join_msg.name_len);
1466 join_msg.dlm_proto = dlm->dlm_locking_proto;
1467 join_msg.fs_proto = dlm->fs_locking_proto;
1469 /* copy live node map to join message */
1470 byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES);
1472 status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
1473 sizeof(join_msg), node, &join_resp);
1474 if (status < 0 && status != -ENOPROTOOPT) {
1475 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1476 "node %u\n", status, DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1480 dlm_query_join_wire_to_packet(join_resp, &packet);
1482 /* -ENOPROTOOPT from the net code means the other side isn't
1483 listening for our message type -- that's fine, it means
1484 his dlm isn't up, so we can consider him a 'yes' but not
1485 joined into the domain. */
1486 if (status == -ENOPROTOOPT) {
1488 *response = JOIN_OK_NO_MAP;
1489 } else if (packet.code == JOIN_DISALLOW ||
1490 packet.code == JOIN_OK_NO_MAP) {
1491 *response = packet.code;
1492 } else if (packet.code == JOIN_PROTOCOL_MISMATCH) {
1494 "This node requested DLM locking protocol %u.%u and "
1495 "filesystem locking protocol %u.%u. At least one of "
1496 "the protocol versions on node %d is not compatible, "
1498 dlm->dlm_locking_proto.pv_major,
1499 dlm->dlm_locking_proto.pv_minor,
1500 dlm->fs_locking_proto.pv_major,
1501 dlm->fs_locking_proto.pv_minor,
1504 *response = packet.code;
1505 } else if (packet.code == JOIN_OK) {
1506 *response = packet.code;
1507 /* Use the same locking protocol as the remote node */
1508 dlm->dlm_locking_proto.pv_minor = packet.dlm_minor;
1509 dlm->fs_locking_proto.pv_minor = packet.fs_minor;
1511 "Node %d responds JOIN_OK with DLM locking protocol "
1512 "%u.%u and fs locking protocol %u.%u\n",
1514 dlm->dlm_locking_proto.pv_major,
1515 dlm->dlm_locking_proto.pv_minor,
1516 dlm->fs_locking_proto.pv_major,
1517 dlm->fs_locking_proto.pv_minor);
1520 mlog(ML_ERROR, "invalid response %d from node %u\n",
1524 mlog(0, "status %d, node %d response is %d\n", status, node,
1531 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
1535 struct dlm_assert_joined assert_msg;
1537 mlog(0, "Sending join assert to node %u\n", node);
1539 memset(&assert_msg, 0, sizeof(assert_msg));
1540 assert_msg.node_idx = dlm->node_num;
1541 assert_msg.name_len = strlen(dlm->name);
1542 memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
1544 status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1545 &assert_msg, sizeof(assert_msg), node,
1548 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1549 "node %u\n", status, DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1555 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
1556 unsigned long *node_map)
1558 int status, node, live;
1562 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1563 node + 1)) < O2NM_MAX_NODES) {
1564 if (node == dlm->node_num)
1568 /* It is very important that this message be
1569 * received so we spin until either the node
1570 * has died or it gets the message. */
1571 status = dlm_send_one_join_assert(dlm, node);
1573 spin_lock(&dlm->spinlock);
1574 live = test_bit(node, dlm->live_nodes_map);
1575 spin_unlock(&dlm->spinlock);
1578 mlog(ML_ERROR, "Error return %d asserting "
1579 "join on node %d\n", status, node);
1581 /* give us some time between errors... */
1583 msleep(DLM_DOMAIN_BACKOFF_MS);
1585 } while (status && live);
1589 struct domain_join_ctxt {
1590 unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1591 unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1594 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
1595 struct domain_join_ctxt *ctxt,
1596 enum dlm_query_join_response_code response)
1600 if (response == JOIN_DISALLOW) {
1601 mlog(0, "Latest response of disallow -- should restart\n");
1605 spin_lock(&dlm->spinlock);
1606 /* For now, we restart the process if the node maps have
1608 ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
1609 sizeof(dlm->live_nodes_map));
1610 spin_unlock(&dlm->spinlock);
1613 mlog(0, "Node maps changed -- should restart\n");
1618 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
1620 int status = 0, tmpstat, node;
1621 struct domain_join_ctxt *ctxt;
1622 enum dlm_query_join_response_code response = JOIN_DISALLOW;
1626 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
1633 /* group sem locking should work for us here -- we're already
1634 * registered for heartbeat events so filling this should be
1635 * atomic wrt getting those handlers called. */
1636 o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
1638 spin_lock(&dlm->spinlock);
1639 memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
1641 __dlm_set_joining_node(dlm, dlm->node_num);
1643 spin_unlock(&dlm->spinlock);
1646 while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
1647 node + 1)) < O2NM_MAX_NODES) {
1648 if (node == dlm->node_num)
1651 status = dlm_request_join(dlm, node, &response);
1657 /* Ok, either we got a response or the node doesn't have a
1659 if (response == JOIN_OK)
1660 set_bit(node, ctxt->yes_resp_map);
1662 if (dlm_should_restart_join(dlm, ctxt, response)) {
1668 mlog(0, "Yay, done querying nodes!\n");
1670 /* Yay, everyone agree's we can join the domain. My domain is
1671 * comprised of all nodes who were put in the
1672 * yes_resp_map. Copy that into our domain map and send a join
1673 * assert message to clean up everyone elses state. */
1674 spin_lock(&dlm->spinlock);
1675 memcpy(dlm->domain_map, ctxt->yes_resp_map,
1676 sizeof(ctxt->yes_resp_map));
1677 set_bit(dlm->node_num, dlm->domain_map);
1678 spin_unlock(&dlm->spinlock);
1680 /* Support for global heartbeat and node info was added in 1.1 */
1681 if (dlm->dlm_locking_proto.pv_major > 1 ||
1682 dlm->dlm_locking_proto.pv_minor > 0) {
1683 status = dlm_send_nodeinfo(dlm, ctxt->yes_resp_map);
1688 status = dlm_send_regions(dlm, ctxt->yes_resp_map);
1695 dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1697 /* Joined state *must* be set before the joining node
1698 * information, otherwise the query_join handler may read no
1699 * current joiner but a state of NEW and tell joining nodes
1700 * we're not in the domain. */
1701 spin_lock(&dlm_domain_lock);
1702 dlm->dlm_state = DLM_CTXT_JOINED;
1704 spin_unlock(&dlm_domain_lock);
1707 spin_lock(&dlm->spinlock);
1708 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1710 printk(KERN_NOTICE "o2dlm: Joining domain %s ", dlm->name);
1711 __dlm_print_nodes(dlm);
1713 spin_unlock(&dlm->spinlock);
1716 /* Do we need to send a cancel message to any nodes? */
1718 tmpstat = dlm_send_join_cancels(dlm,
1720 sizeof(ctxt->yes_resp_map));
1722 mlog_errno(tmpstat);
1727 mlog(0, "returning %d\n", status);
1731 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1733 o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_up);
1734 o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_down);
1735 o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1738 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1742 mlog(0, "registering handlers.\n");
1744 o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1745 dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1746 status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_down);
1750 o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1751 dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1752 status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_up);
1756 status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1757 sizeof(struct dlm_master_request),
1758 dlm_master_request_handler,
1759 dlm, NULL, &dlm->dlm_domain_handlers);
1763 status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1764 sizeof(struct dlm_assert_master),
1765 dlm_assert_master_handler,
1766 dlm, dlm_assert_master_post_handler,
1767 &dlm->dlm_domain_handlers);
1771 status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1772 sizeof(struct dlm_create_lock),
1773 dlm_create_lock_handler,
1774 dlm, NULL, &dlm->dlm_domain_handlers);
1778 status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1779 DLM_CONVERT_LOCK_MAX_LEN,
1780 dlm_convert_lock_handler,
1781 dlm, NULL, &dlm->dlm_domain_handlers);
1785 status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1786 DLM_UNLOCK_LOCK_MAX_LEN,
1787 dlm_unlock_lock_handler,
1788 dlm, NULL, &dlm->dlm_domain_handlers);
1792 status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1793 DLM_PROXY_AST_MAX_LEN,
1794 dlm_proxy_ast_handler,
1795 dlm, NULL, &dlm->dlm_domain_handlers);
1799 status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1800 sizeof(struct dlm_exit_domain),
1801 dlm_exit_domain_handler,
1802 dlm, NULL, &dlm->dlm_domain_handlers);
1806 status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1807 sizeof(struct dlm_deref_lockres),
1808 dlm_deref_lockres_handler,
1809 dlm, NULL, &dlm->dlm_domain_handlers);
1813 status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1814 sizeof(struct dlm_migrate_request),
1815 dlm_migrate_request_handler,
1816 dlm, NULL, &dlm->dlm_domain_handlers);
1820 status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1821 DLM_MIG_LOCKRES_MAX_LEN,
1822 dlm_mig_lockres_handler,
1823 dlm, NULL, &dlm->dlm_domain_handlers);
1827 status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1828 sizeof(struct dlm_master_requery),
1829 dlm_master_requery_handler,
1830 dlm, NULL, &dlm->dlm_domain_handlers);
1834 status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1835 sizeof(struct dlm_lock_request),
1836 dlm_request_all_locks_handler,
1837 dlm, NULL, &dlm->dlm_domain_handlers);
1841 status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1842 sizeof(struct dlm_reco_data_done),
1843 dlm_reco_data_done_handler,
1844 dlm, NULL, &dlm->dlm_domain_handlers);
1848 status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1849 sizeof(struct dlm_begin_reco),
1850 dlm_begin_reco_handler,
1851 dlm, NULL, &dlm->dlm_domain_handlers);
1855 status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1856 sizeof(struct dlm_finalize_reco),
1857 dlm_finalize_reco_handler,
1858 dlm, NULL, &dlm->dlm_domain_handlers);
1862 status = o2net_register_handler(DLM_BEGIN_EXIT_DOMAIN_MSG, dlm->key,
1863 sizeof(struct dlm_exit_domain),
1864 dlm_begin_exit_domain_handler,
1865 dlm, NULL, &dlm->dlm_domain_handlers);
1871 dlm_unregister_domain_handlers(dlm);
1876 static int dlm_join_domain(struct dlm_ctxt *dlm)
1879 unsigned int backoff;
1880 unsigned int total_backoff = 0;
1884 mlog(0, "Join domain %s\n", dlm->name);
1886 status = dlm_register_domain_handlers(dlm);
1892 status = dlm_debug_init(dlm);
1898 status = dlm_launch_thread(dlm);
1904 status = dlm_launch_recovery_thread(dlm);
1910 dlm->dlm_worker = create_singlethread_workqueue("dlm_wq");
1911 if (!dlm->dlm_worker) {
1918 status = dlm_try_to_join_domain(dlm);
1920 /* If we're racing another node to the join, then we
1921 * need to back off temporarily and let them
1923 #define DLM_JOIN_TIMEOUT_MSECS 90000
1924 if (status == -EAGAIN) {
1925 if (signal_pending(current)) {
1926 status = -ERESTARTSYS;
1931 msecs_to_jiffies(DLM_JOIN_TIMEOUT_MSECS)) {
1932 status = -ERESTARTSYS;
1933 mlog(ML_NOTICE, "Timed out joining dlm domain "
1934 "%s after %u msecs\n", dlm->name,
1935 jiffies_to_msecs(total_backoff));
1941 * <dale> No, after you!
1943 * <dale> But you first!
1946 backoff = (unsigned int)(jiffies & 0x3);
1947 backoff *= DLM_DOMAIN_BACKOFF_MS;
1948 total_backoff += backoff;
1949 mlog(0, "backoff %d\n", backoff);
1952 } while (status == -EAGAIN);
1961 wake_up(&dlm_domain_events);
1964 dlm_unregister_domain_handlers(dlm);
1965 dlm_debug_shutdown(dlm);
1966 dlm_complete_thread(dlm);
1967 dlm_complete_recovery_thread(dlm);
1968 dlm_destroy_dlm_worker(dlm);
1974 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1979 struct dlm_ctxt *dlm = NULL;
1981 dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1983 mlog_errno(-ENOMEM);
1987 dlm->name = kstrdup(domain, GFP_KERNEL);
1988 if (dlm->name == NULL) {
1989 mlog_errno(-ENOMEM);
1995 dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1996 if (!dlm->lockres_hash) {
1997 mlog_errno(-ENOMEM);
2004 for (i = 0; i < DLM_HASH_BUCKETS; i++)
2005 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
2007 dlm->master_hash = (struct hlist_head **)
2008 dlm_alloc_pagevec(DLM_HASH_PAGES);
2009 if (!dlm->master_hash) {
2010 mlog_errno(-ENOMEM);
2011 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
2018 for (i = 0; i < DLM_HASH_BUCKETS; i++)
2019 INIT_HLIST_HEAD(dlm_master_hash(dlm, i));
2022 dlm->node_num = o2nm_this_node();
2024 ret = dlm_create_debugfs_subroot(dlm);
2026 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
2027 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
2034 spin_lock_init(&dlm->spinlock);
2035 spin_lock_init(&dlm->master_lock);
2036 spin_lock_init(&dlm->ast_lock);
2037 spin_lock_init(&dlm->track_lock);
2038 INIT_LIST_HEAD(&dlm->list);
2039 INIT_LIST_HEAD(&dlm->dirty_list);
2040 INIT_LIST_HEAD(&dlm->reco.resources);
2041 INIT_LIST_HEAD(&dlm->reco.received);
2042 INIT_LIST_HEAD(&dlm->reco.node_data);
2043 INIT_LIST_HEAD(&dlm->purge_list);
2044 INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
2045 INIT_LIST_HEAD(&dlm->tracking_list);
2046 dlm->reco.state = 0;
2048 INIT_LIST_HEAD(&dlm->pending_asts);
2049 INIT_LIST_HEAD(&dlm->pending_basts);
2051 mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
2052 dlm->recovery_map, &(dlm->recovery_map[0]));
2054 memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
2055 memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
2056 memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
2058 dlm->dlm_thread_task = NULL;
2059 dlm->dlm_reco_thread_task = NULL;
2060 dlm->dlm_worker = NULL;
2061 init_waitqueue_head(&dlm->dlm_thread_wq);
2062 init_waitqueue_head(&dlm->dlm_reco_thread_wq);
2063 init_waitqueue_head(&dlm->reco.event);
2064 init_waitqueue_head(&dlm->ast_wq);
2065 init_waitqueue_head(&dlm->migration_wq);
2066 INIT_LIST_HEAD(&dlm->mle_hb_events);
2068 dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
2069 init_waitqueue_head(&dlm->dlm_join_events);
2071 dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
2072 dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
2074 atomic_set(&dlm->res_tot_count, 0);
2075 atomic_set(&dlm->res_cur_count, 0);
2076 for (i = 0; i < DLM_MLE_NUM_TYPES; ++i) {
2077 atomic_set(&dlm->mle_tot_count[i], 0);
2078 atomic_set(&dlm->mle_cur_count[i], 0);
2081 spin_lock_init(&dlm->work_lock);
2082 INIT_LIST_HEAD(&dlm->work_list);
2083 INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
2085 kref_init(&dlm->dlm_refs);
2086 dlm->dlm_state = DLM_CTXT_NEW;
2088 INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
2090 mlog(0, "context init: refcount %u\n",
2091 atomic_read(&dlm->dlm_refs.refcount));
2098 * Compare a requested locking protocol version against the current one.
2100 * If the major numbers are different, they are incompatible.
2101 * If the current minor is greater than the request, they are incompatible.
2102 * If the current minor is less than or equal to the request, they are
2103 * compatible, and the requester should run at the current minor version.
2105 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
2106 struct dlm_protocol_version *request)
2108 if (existing->pv_major != request->pv_major)
2111 if (existing->pv_minor > request->pv_minor)
2114 if (existing->pv_minor < request->pv_minor)
2115 request->pv_minor = existing->pv_minor;
2121 * dlm_register_domain: one-time setup per "domain".
2123 * The filesystem passes in the requested locking version via proto.
2124 * If registration was successful, proto will contain the negotiated
2127 struct dlm_ctxt * dlm_register_domain(const char *domain,
2129 struct dlm_protocol_version *fs_proto)
2132 struct dlm_ctxt *dlm = NULL;
2133 struct dlm_ctxt *new_ctxt = NULL;
2135 if (strlen(domain) >= O2NM_MAX_NAME_LEN) {
2136 ret = -ENAMETOOLONG;
2137 mlog(ML_ERROR, "domain name length too long\n");
2141 mlog(0, "register called for domain \"%s\"\n", domain);
2145 if (signal_pending(current)) {
2151 spin_lock(&dlm_domain_lock);
2153 dlm = __dlm_lookup_domain(domain);
2155 if (dlm->dlm_state != DLM_CTXT_JOINED) {
2156 spin_unlock(&dlm_domain_lock);
2158 mlog(0, "This ctxt is not joined yet!\n");
2159 wait_event_interruptible(dlm_domain_events,
2160 dlm_wait_on_domain_helper(
2165 if (dlm_protocol_compare(&dlm->fs_locking_proto, fs_proto)) {
2166 spin_unlock(&dlm_domain_lock);
2168 "Requested locking protocol version is not "
2169 "compatible with already registered domain "
2170 "\"%s\"\n", domain);
2178 spin_unlock(&dlm_domain_lock);
2186 spin_unlock(&dlm_domain_lock);
2188 new_ctxt = dlm_alloc_ctxt(domain, key);
2197 /* a little variable switch-a-roo here... */
2201 /* add the new domain */
2202 list_add_tail(&dlm->list, &dlm_domains);
2203 spin_unlock(&dlm_domain_lock);
2206 * Pass the locking protocol version into the join. If the join
2207 * succeeds, it will have the negotiated protocol set.
2209 dlm->dlm_locking_proto = dlm_protocol;
2210 dlm->fs_locking_proto = *fs_proto;
2212 ret = dlm_join_domain(dlm);
2219 /* Tell the caller what locking protocol we negotiated */
2220 *fs_proto = dlm->fs_locking_proto;
2225 dlm_free_ctxt_mem(new_ctxt);
2232 EXPORT_SYMBOL_GPL(dlm_register_domain);
2234 static LIST_HEAD(dlm_join_handlers);
2236 static void dlm_unregister_net_handlers(void)
2238 o2net_unregister_handler_list(&dlm_join_handlers);
2241 static int dlm_register_net_handlers(void)
2245 status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
2246 sizeof(struct dlm_query_join_request),
2247 dlm_query_join_handler,
2248 NULL, NULL, &dlm_join_handlers);
2252 status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
2253 sizeof(struct dlm_assert_joined),
2254 dlm_assert_joined_handler,
2255 NULL, NULL, &dlm_join_handlers);
2259 status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
2260 sizeof(struct dlm_cancel_join),
2261 dlm_cancel_join_handler,
2262 NULL, NULL, &dlm_join_handlers);
2266 status = o2net_register_handler(DLM_QUERY_REGION, DLM_MOD_KEY,
2267 sizeof(struct dlm_query_region),
2268 dlm_query_region_handler,
2269 NULL, NULL, &dlm_join_handlers);
2274 status = o2net_register_handler(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
2275 sizeof(struct dlm_query_nodeinfo),
2276 dlm_query_nodeinfo_handler,
2277 NULL, NULL, &dlm_join_handlers);
2280 dlm_unregister_net_handlers();
2285 /* Domain eviction callback handling.
2287 * The file system requires notification of node death *before* the
2288 * dlm completes it's recovery work, otherwise it may be able to
2289 * acquire locks on resources requiring recovery. Since the dlm can
2290 * evict a node from it's domain *before* heartbeat fires, a similar
2291 * mechanism is required. */
2293 /* Eviction is not expected to happen often, so a per-domain lock is
2294 * not necessary. Eviction callbacks are allowed to sleep for short
2295 * periods of time. */
2296 static DECLARE_RWSEM(dlm_callback_sem);
2298 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
2301 struct list_head *iter;
2302 struct dlm_eviction_cb *cb;
2304 down_read(&dlm_callback_sem);
2305 list_for_each(iter, &dlm->dlm_eviction_callbacks) {
2306 cb = list_entry(iter, struct dlm_eviction_cb, ec_item);
2308 cb->ec_func(node_num, cb->ec_data);
2310 up_read(&dlm_callback_sem);
2313 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
2314 dlm_eviction_func *f,
2317 INIT_LIST_HEAD(&cb->ec_item);
2321 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
2323 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
2324 struct dlm_eviction_cb *cb)
2326 down_write(&dlm_callback_sem);
2327 list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
2328 up_write(&dlm_callback_sem);
2330 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
2332 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
2334 down_write(&dlm_callback_sem);
2335 list_del_init(&cb->ec_item);
2336 up_write(&dlm_callback_sem);
2338 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
2340 static int __init dlm_init(void)
2344 dlm_print_version();
2346 status = dlm_init_mle_cache();
2348 mlog(ML_ERROR, "Could not create o2dlm_mle slabcache\n");
2352 status = dlm_init_master_caches();
2354 mlog(ML_ERROR, "Could not create o2dlm_lockres and "
2355 "o2dlm_lockname slabcaches\n");
2359 status = dlm_init_lock_cache();
2361 mlog(ML_ERROR, "Count not create o2dlm_lock slabcache\n");
2365 status = dlm_register_net_handlers();
2367 mlog(ML_ERROR, "Unable to register network handlers\n");
2371 status = dlm_create_debugfs_root();
2377 dlm_unregister_net_handlers();
2378 dlm_destroy_lock_cache();
2379 dlm_destroy_master_caches();
2380 dlm_destroy_mle_cache();
2384 static void __exit dlm_exit (void)
2386 dlm_destroy_debugfs_root();
2387 dlm_unregister_net_handlers();
2388 dlm_destroy_lock_cache();
2389 dlm_destroy_master_caches();
2390 dlm_destroy_mle_cache();
2393 MODULE_AUTHOR("Oracle");
2394 MODULE_LICENSE("GPL");
2396 module_init(dlm_init);
2397 module_exit(dlm_exit);