1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* -*- mode: c; c-basic-offset: 8; -*-
3 * vim: noexpandtab sw=8 ts=8 sts=0:
7 * defines domain join / leave apis
9 * Copyright (C) 2004 Oracle. All rights reserved.
12 #include <linux/module.h>
13 #include <linux/types.h>
14 #include <linux/slab.h>
15 #include <linux/highmem.h>
16 #include <linux/init.h>
17 #include <linux/spinlock.h>
18 #include <linux/delay.h>
19 #include <linux/err.h>
20 #include <linux/debugfs.h>
21 #include <linux/sched/signal.h>
23 #include "cluster/heartbeat.h"
24 #include "cluster/nodemanager.h"
25 #include "cluster/tcp.h"
28 #include "dlmcommon.h"
29 #include "dlmdomain.h"
32 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
33 #include "cluster/masklog.h"
36 * ocfs2 node maps are array of long int, which limits to send them freely
37 * across the wire due to endianness issues. To workaround this, we convert
38 * long ints to byte arrays. Following 3 routines are helper functions to
39 * set/test/copy bits within those array of bytes
41 static inline void byte_set_bit(u8 nr, u8 map[])
43 map[nr >> 3] |= (1UL << (nr & 7));
46 static inline int byte_test_bit(u8 nr, u8 map[])
48 return ((1UL << (nr & 7)) & (map[nr >> 3])) != 0;
51 static inline void byte_copymap(u8 dmap[], unsigned long smap[],
59 memset(dmap, 0, ((sz + 7) >> 3));
60 for (nn = 0 ; nn < sz; nn++)
61 if (test_bit(nn, smap))
62 byte_set_bit(nn, dmap);
65 static void dlm_free_pagevec(void **vec, int pages)
68 free_page((unsigned long)vec[pages]);
72 static void **dlm_alloc_pagevec(int pages)
74 void **vec = kmalloc_array(pages, sizeof(void *), GFP_KERNEL);
80 for (i = 0; i < pages; i++)
81 if (!(vec[i] = (void *)__get_free_page(GFP_KERNEL)))
84 mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n",
85 pages, (unsigned long)DLM_HASH_PAGES,
86 (unsigned long)DLM_BUCKETS_PER_PAGE);
89 dlm_free_pagevec(vec, i);
95 * spinlock lock ordering: if multiple locks are needed, obey this ordering:
97 * struct dlm_ctxt->spinlock
98 * struct dlm_lock_resource->spinlock
99 * struct dlm_ctxt->master_lock
100 * struct dlm_ctxt->ast_lock
101 * dlm_master_list_entry->spinlock
106 DEFINE_SPINLOCK(dlm_domain_lock);
107 LIST_HEAD(dlm_domains);
108 static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);
111 * The supported protocol version for DLM communication. Running domains
112 * will have a negotiated version with the same major number and a minor
113 * number equal or smaller. The dlm_ctxt->dlm_locking_proto field should
114 * be used to determine what a running domain is actually using.
116 * New in version 1.1:
117 * - Message DLM_QUERY_REGION added to support global heartbeat
118 * - Message DLM_QUERY_NODEINFO added to allow online node removes
119 * New in version 1.2:
120 * - Message DLM_BEGIN_EXIT_DOMAIN_MSG added to mark start of exit domain
121 * New in version 1.3:
122 * - Message DLM_DEREF_LOCKRES_DONE added to inform non-master that the
125 static const struct dlm_protocol_version dlm_protocol = {
130 #define DLM_DOMAIN_BACKOFF_MS 200
132 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
134 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
136 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
138 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
139 void *data, void **ret_data);
140 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
142 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
143 struct dlm_protocol_version *request);
145 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
147 void __dlm_unhash_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
149 if (hlist_unhashed(&res->hash_node))
152 mlog(0, "%s: Unhash res %.*s\n", dlm->name, res->lockname.len,
154 hlist_del_init(&res->hash_node);
155 dlm_lockres_put(res);
158 void __dlm_insert_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
160 struct hlist_head *bucket;
162 assert_spin_locked(&dlm->spinlock);
164 bucket = dlm_lockres_hash(dlm, res->lockname.hash);
166 /* get a reference for our hashtable */
167 dlm_lockres_get(res);
169 hlist_add_head(&res->hash_node, bucket);
171 mlog(0, "%s: Hash res %.*s\n", dlm->name, res->lockname.len,
175 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
180 struct hlist_head *bucket;
181 struct dlm_lock_resource *res;
183 mlog(0, "%.*s\n", len, name);
185 assert_spin_locked(&dlm->spinlock);
187 bucket = dlm_lockres_hash(dlm, hash);
189 hlist_for_each_entry(res, bucket, hash_node) {
190 if (res->lockname.name[0] != name[0])
192 if (unlikely(res->lockname.len != len))
194 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
196 dlm_lockres_get(res);
202 /* intended to be called by functions which do not care about lock
203 * resources which are being purged (most net _handler functions).
204 * this will return NULL for any lock resource which is found but
205 * currently in the process of dropping its mastery reference.
206 * use __dlm_lookup_lockres_full when you need the lock resource
207 * regardless (e.g. dlm_get_lock_resource) */
208 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
213 struct dlm_lock_resource *res = NULL;
215 mlog(0, "%.*s\n", len, name);
217 assert_spin_locked(&dlm->spinlock);
219 res = __dlm_lookup_lockres_full(dlm, name, len, hash);
221 spin_lock(&res->spinlock);
222 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
223 spin_unlock(&res->spinlock);
224 dlm_lockres_put(res);
227 spin_unlock(&res->spinlock);
233 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
237 struct dlm_lock_resource *res;
238 unsigned int hash = dlm_lockid_hash(name, len);
240 spin_lock(&dlm->spinlock);
241 res = __dlm_lookup_lockres(dlm, name, len, hash);
242 spin_unlock(&dlm->spinlock);
246 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
248 struct dlm_ctxt *tmp;
250 assert_spin_locked(&dlm_domain_lock);
252 /* tmp->name here is always NULL terminated,
253 * but domain may not be! */
254 list_for_each_entry(tmp, &dlm_domains, list) {
255 if (strlen(tmp->name) == len &&
256 memcmp(tmp->name, domain, len)==0)
263 /* For null terminated domain strings ONLY */
264 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
266 assert_spin_locked(&dlm_domain_lock);
268 return __dlm_lookup_domain_full(domain, strlen(domain));
272 /* returns true on one of two conditions:
273 * 1) the domain does not exist
274 * 2) the domain exists and it's state is "joined" */
275 static int dlm_wait_on_domain_helper(const char *domain)
278 struct dlm_ctxt *tmp = NULL;
280 spin_lock(&dlm_domain_lock);
282 tmp = __dlm_lookup_domain(domain);
285 else if (tmp->dlm_state == DLM_CTXT_JOINED)
288 spin_unlock(&dlm_domain_lock);
292 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
294 dlm_destroy_debugfs_subroot(dlm);
296 if (dlm->lockres_hash)
297 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
299 if (dlm->master_hash)
300 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
306 /* A little strange - this function will be called while holding
307 * dlm_domain_lock and is expected to be holding it on the way out. We
308 * will however drop and reacquire it multiple times */
309 static void dlm_ctxt_release(struct kref *kref)
311 struct dlm_ctxt *dlm;
313 dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
315 BUG_ON(dlm->num_joins);
316 BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
318 /* we may still be in the list if we hit an error during join. */
319 list_del_init(&dlm->list);
321 spin_unlock(&dlm_domain_lock);
323 mlog(0, "freeing memory from domain %s\n", dlm->name);
325 wake_up(&dlm_domain_events);
327 dlm_free_ctxt_mem(dlm);
329 spin_lock(&dlm_domain_lock);
332 void dlm_put(struct dlm_ctxt *dlm)
334 spin_lock(&dlm_domain_lock);
335 kref_put(&dlm->dlm_refs, dlm_ctxt_release);
336 spin_unlock(&dlm_domain_lock);
339 static void __dlm_get(struct dlm_ctxt *dlm)
341 kref_get(&dlm->dlm_refs);
344 /* given a questionable reference to a dlm object, gets a reference if
345 * it can find it in the list, otherwise returns NULL in which case
346 * you shouldn't trust your pointer. */
347 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
349 struct dlm_ctxt *target;
350 struct dlm_ctxt *ret = NULL;
352 spin_lock(&dlm_domain_lock);
354 list_for_each_entry(target, &dlm_domains, list) {
362 spin_unlock(&dlm_domain_lock);
367 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
371 spin_lock(&dlm_domain_lock);
372 ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
373 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
374 spin_unlock(&dlm_domain_lock);
379 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
381 if (dlm->dlm_worker) {
382 destroy_workqueue(dlm->dlm_worker);
383 dlm->dlm_worker = NULL;
387 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
389 dlm_unregister_domain_handlers(dlm);
390 dlm_debug_shutdown(dlm);
391 dlm_complete_thread(dlm);
392 dlm_complete_recovery_thread(dlm);
393 dlm_destroy_dlm_worker(dlm);
395 /* We've left the domain. Now we can take ourselves out of the
396 * list and allow the kref stuff to help us free the
398 spin_lock(&dlm_domain_lock);
399 list_del_init(&dlm->list);
400 spin_unlock(&dlm_domain_lock);
402 /* Wake up anyone waiting for us to remove this domain */
403 wake_up(&dlm_domain_events);
406 static int dlm_migrate_all_locks(struct dlm_ctxt *dlm)
408 int i, num, n, ret = 0;
409 struct dlm_lock_resource *res;
410 struct hlist_node *iter;
411 struct hlist_head *bucket;
414 mlog(0, "Migrating locks from domain %s\n", dlm->name);
417 spin_lock(&dlm->spinlock);
418 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
421 bucket = dlm_lockres_hash(dlm, i);
422 iter = bucket->first;
425 res = hlist_entry(iter, struct dlm_lock_resource,
427 dlm_lockres_get(res);
428 /* migrate, if necessary. this will drop the dlm
429 * spinlock and retake it if it does migration. */
430 dropped = dlm_empty_lockres(dlm, res);
432 spin_lock(&res->spinlock);
434 __dlm_lockres_calc_usage(dlm, res);
436 iter = res->hash_node.next;
437 spin_unlock(&res->spinlock);
439 dlm_lockres_put(res);
442 cond_resched_lock(&dlm->spinlock);
446 cond_resched_lock(&dlm->spinlock);
451 if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
452 mlog(0, "%s: perhaps there are more lock resources "
453 "need to be migrated after dlm recovery\n", dlm->name);
456 mlog(0, "%s: we won't do dlm recovery after migrating "
457 "all lock resources\n", dlm->name);
458 dlm->migrate_done = 1;
462 spin_unlock(&dlm->spinlock);
463 wake_up(&dlm->dlm_thread_wq);
465 /* let the dlm thread take care of purging, keep scanning until
466 * nothing remains in the hash */
468 mlog(0, "%s: %d lock resources in hash last pass\n",
472 mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
476 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
480 spin_lock(&dlm->spinlock);
481 ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
482 spin_unlock(&dlm->spinlock);
487 static int dlm_begin_exit_domain_handler(struct o2net_msg *msg, u32 len,
488 void *data, void **ret_data)
490 struct dlm_ctxt *dlm = data;
492 struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
497 node = exit_msg->node_idx;
498 mlog(0, "%s: Node %u sent a begin exit domain message\n", dlm->name, node);
500 spin_lock(&dlm->spinlock);
501 set_bit(node, dlm->exit_domain_map);
502 spin_unlock(&dlm->spinlock);
509 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
511 /* Yikes, a double spinlock! I need domain_lock for the dlm
512 * state and the dlm spinlock for join state... Sorry! */
514 spin_lock(&dlm_domain_lock);
515 spin_lock(&dlm->spinlock);
517 if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
518 mlog(0, "Node %d is joining, we wait on it.\n",
520 spin_unlock(&dlm->spinlock);
521 spin_unlock(&dlm_domain_lock);
523 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
527 dlm->dlm_state = DLM_CTXT_LEAVING;
528 spin_unlock(&dlm->spinlock);
529 spin_unlock(&dlm_domain_lock);
532 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
534 int node = -1, num = 0;
536 assert_spin_locked(&dlm->spinlock);
539 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
540 node + 1)) < O2NM_MAX_NODES) {
544 printk(") %u nodes\n", num);
547 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
550 struct dlm_ctxt *dlm = data;
552 struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
554 mlog(0, "%p %u %p", msg, len, data);
559 node = exit_msg->node_idx;
561 spin_lock(&dlm->spinlock);
562 clear_bit(node, dlm->domain_map);
563 clear_bit(node, dlm->exit_domain_map);
564 printk(KERN_NOTICE "o2dlm: Node %u leaves domain %s ", node, dlm->name);
565 __dlm_print_nodes(dlm);
567 /* notify anything attached to the heartbeat events */
568 dlm_hb_event_notify_attached(dlm, node, 0);
570 spin_unlock(&dlm->spinlock);
577 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm, u32 msg_type,
581 struct dlm_exit_domain leave_msg;
583 mlog(0, "%s: Sending domain exit message %u to node %u\n", dlm->name,
586 memset(&leave_msg, 0, sizeof(leave_msg));
587 leave_msg.node_idx = dlm->node_num;
589 status = o2net_send_message(msg_type, dlm->key, &leave_msg,
590 sizeof(leave_msg), node, NULL);
592 mlog(ML_ERROR, "Error %d sending domain exit message %u "
593 "to node %u on domain %s\n", status, msg_type, node,
599 static void dlm_begin_exit_domain(struct dlm_ctxt *dlm)
603 /* Support for begin exit domain was added in 1.2 */
604 if (dlm->dlm_locking_proto.pv_major == 1 &&
605 dlm->dlm_locking_proto.pv_minor < 2)
609 * Unlike DLM_EXIT_DOMAIN_MSG, DLM_BEGIN_EXIT_DOMAIN_MSG is purely
610 * informational. Meaning if a node does not receive the message,
613 spin_lock(&dlm->spinlock);
615 node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, node + 1);
616 if (node >= O2NM_MAX_NODES)
618 if (node == dlm->node_num)
621 spin_unlock(&dlm->spinlock);
622 dlm_send_one_domain_exit(dlm, DLM_BEGIN_EXIT_DOMAIN_MSG, node);
623 spin_lock(&dlm->spinlock);
625 spin_unlock(&dlm->spinlock);
628 static void dlm_leave_domain(struct dlm_ctxt *dlm)
630 int node, clear_node, status;
632 /* At this point we've migrated away all our locks and won't
633 * accept mastership of new ones. The dlm is responsible for
634 * almost nothing now. We make sure not to confuse any joining
635 * nodes and then commence shutdown procedure. */
637 spin_lock(&dlm->spinlock);
638 /* Clear ourselves from the domain map */
639 clear_bit(dlm->node_num, dlm->domain_map);
640 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
641 0)) < O2NM_MAX_NODES) {
642 /* Drop the dlm spinlock. This is safe wrt the domain_map.
643 * -nodes cannot be added now as the
644 * query_join_handlers knows to respond with OK_NO_MAP
645 * -we catch the right network errors if a node is
646 * removed from the map while we're sending him the
648 spin_unlock(&dlm->spinlock);
652 status = dlm_send_one_domain_exit(dlm, DLM_EXIT_DOMAIN_MSG,
655 status != -ENOPROTOOPT &&
656 status != -ENOTCONN) {
657 mlog(ML_NOTICE, "Error %d sending domain exit message "
658 "to node %d\n", status, node);
660 /* Not sure what to do here but lets sleep for
661 * a bit in case this was a transient
663 msleep(DLM_DOMAIN_BACKOFF_MS);
667 spin_lock(&dlm->spinlock);
668 /* If we're not clearing the node bit then we intend
669 * to loop back around to try again. */
671 clear_bit(node, dlm->domain_map);
673 spin_unlock(&dlm->spinlock);
676 void dlm_unregister_domain(struct dlm_ctxt *dlm)
679 struct dlm_lock_resource *res;
681 spin_lock(&dlm_domain_lock);
682 BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
683 BUG_ON(!dlm->num_joins);
686 if (!dlm->num_joins) {
687 /* We mark it "in shutdown" now so new register
688 * requests wait until we've completely left the
689 * domain. Don't use DLM_CTXT_LEAVING yet as we still
690 * want new domain joins to communicate with us at
691 * least until we've completed migration of our
693 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
696 spin_unlock(&dlm_domain_lock);
699 mlog(0, "shutting down domain %s\n", dlm->name);
700 dlm_begin_exit_domain(dlm);
702 /* We changed dlm state, notify the thread */
703 dlm_kick_thread(dlm, NULL);
705 while (dlm_migrate_all_locks(dlm)) {
706 /* Give dlm_thread time to purge the lockres' */
708 mlog(0, "%s: more migration to do\n", dlm->name);
711 /* This list should be empty. If not, print remaining lockres */
712 if (!list_empty(&dlm->tracking_list)) {
713 mlog(ML_ERROR, "Following lockres' are still on the "
715 list_for_each_entry(res, &dlm->tracking_list, tracking)
716 dlm_print_one_lock_resource(res);
719 dlm_mark_domain_leaving(dlm);
720 dlm_leave_domain(dlm);
721 printk(KERN_NOTICE "o2dlm: Leaving domain %s\n", dlm->name);
722 dlm_force_free_mles(dlm);
723 dlm_complete_dlm_shutdown(dlm);
727 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
729 static int dlm_query_join_proto_check(char *proto_type, int node,
730 struct dlm_protocol_version *ours,
731 struct dlm_protocol_version *request)
734 struct dlm_protocol_version proto = *request;
736 if (!dlm_protocol_compare(ours, &proto)) {
738 "node %u wanted to join with %s locking protocol "
739 "%u.%u, we respond with %u.%u\n",
743 proto.pv_major, proto.pv_minor);
744 request->pv_minor = proto.pv_minor;
748 "Node %u wanted to join with %s locking "
749 "protocol %u.%u, but we have %u.%u, disallowing\n",
762 * struct dlm_query_join_packet is made up of four one-byte fields. They
763 * are effectively in big-endian order already. However, little-endian
764 * machines swap them before putting the packet on the wire (because
765 * query_join's response is a status, and that status is treated as a u32
766 * on the wire). Thus, a big-endian and little-endian machines will treat
767 * this structure differently.
769 * The solution is to have little-endian machines swap the structure when
770 * converting from the structure to the u32 representation. This will
771 * result in the structure having the correct format on the wire no matter
772 * the host endian format.
774 static void dlm_query_join_packet_to_wire(struct dlm_query_join_packet *packet,
777 union dlm_query_join_response response;
779 response.packet = *packet;
780 *wire = be32_to_cpu(response.intval);
783 static void dlm_query_join_wire_to_packet(u32 wire,
784 struct dlm_query_join_packet *packet)
786 union dlm_query_join_response response;
788 response.intval = cpu_to_be32(wire);
789 *packet = response.packet;
792 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
795 struct dlm_query_join_request *query;
796 struct dlm_query_join_packet packet = {
797 .code = JOIN_DISALLOW,
799 struct dlm_ctxt *dlm = NULL;
803 query = (struct dlm_query_join_request *) msg->buf;
805 mlog(0, "node %u wants to join domain %s\n", query->node_idx,
809 * If heartbeat doesn't consider the node live, tell it
810 * to back off and try again. This gives heartbeat a chance
813 if (!o2hb_check_node_heartbeating_no_sem(query->node_idx)) {
814 mlog(0, "node %u is not in our live map yet\n",
817 packet.code = JOIN_DISALLOW;
821 packet.code = JOIN_OK_NO_MAP;
823 spin_lock(&dlm_domain_lock);
824 dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
829 * There is a small window where the joining node may not see the
830 * node(s) that just left but still part of the cluster. DISALLOW
831 * join request if joining node has different node map.
834 while (nodenum < O2NM_MAX_NODES) {
835 if (test_bit(nodenum, dlm->domain_map)) {
836 if (!byte_test_bit(nodenum, query->node_map)) {
837 mlog(0, "disallow join as node %u does not "
838 "have node %u in its nodemap\n",
839 query->node_idx, nodenum);
840 packet.code = JOIN_DISALLOW;
847 /* Once the dlm ctxt is marked as leaving then we don't want
848 * to be put in someone's domain map.
849 * Also, explicitly disallow joining at certain troublesome
850 * times (ie. during recovery). */
851 if (dlm->dlm_state != DLM_CTXT_LEAVING) {
852 int bit = query->node_idx;
853 spin_lock(&dlm->spinlock);
855 if (dlm->dlm_state == DLM_CTXT_NEW &&
856 dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
857 /*If this is a brand new context and we
858 * haven't started our join process yet, then
859 * the other node won the race. */
860 packet.code = JOIN_OK_NO_MAP;
861 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
862 /* Disallow parallel joins. */
863 packet.code = JOIN_DISALLOW;
864 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
865 mlog(0, "node %u trying to join, but recovery "
866 "is ongoing.\n", bit);
867 packet.code = JOIN_DISALLOW;
868 } else if (test_bit(bit, dlm->recovery_map)) {
869 mlog(0, "node %u trying to join, but it "
870 "still needs recovery.\n", bit);
871 packet.code = JOIN_DISALLOW;
872 } else if (test_bit(bit, dlm->domain_map)) {
873 mlog(0, "node %u trying to join, but it "
874 "is still in the domain! needs recovery?\n",
876 packet.code = JOIN_DISALLOW;
878 /* Alright we're fully a part of this domain
879 * so we keep some state as to who's joining
880 * and indicate to him that needs to be fixed
883 /* Make sure we speak compatible locking protocols. */
884 if (dlm_query_join_proto_check("DLM", bit,
885 &dlm->dlm_locking_proto,
886 &query->dlm_proto)) {
887 packet.code = JOIN_PROTOCOL_MISMATCH;
888 } else if (dlm_query_join_proto_check("fs", bit,
889 &dlm->fs_locking_proto,
891 packet.code = JOIN_PROTOCOL_MISMATCH;
893 packet.dlm_minor = query->dlm_proto.pv_minor;
894 packet.fs_minor = query->fs_proto.pv_minor;
895 packet.code = JOIN_OK;
896 __dlm_set_joining_node(dlm, query->node_idx);
900 spin_unlock(&dlm->spinlock);
903 spin_unlock(&dlm_domain_lock);
906 mlog(0, "We respond with %u\n", packet.code);
908 dlm_query_join_packet_to_wire(&packet, &response);
912 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
915 struct dlm_assert_joined *assert;
916 struct dlm_ctxt *dlm = NULL;
918 assert = (struct dlm_assert_joined *) msg->buf;
920 mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
923 spin_lock(&dlm_domain_lock);
924 dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
925 /* XXX should we consider no dlm ctxt an error? */
927 spin_lock(&dlm->spinlock);
929 /* Alright, this node has officially joined our
930 * domain. Set him in the map and clean up our
931 * leftover join state. */
932 BUG_ON(dlm->joining_node != assert->node_idx);
934 if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
935 mlog(0, "dlm recovery is ongoing, disallow join\n");
936 spin_unlock(&dlm->spinlock);
937 spin_unlock(&dlm_domain_lock);
941 set_bit(assert->node_idx, dlm->domain_map);
942 clear_bit(assert->node_idx, dlm->exit_domain_map);
943 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
945 printk(KERN_NOTICE "o2dlm: Node %u joins domain %s ",
946 assert->node_idx, dlm->name);
947 __dlm_print_nodes(dlm);
949 /* notify anything attached to the heartbeat events */
950 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
952 spin_unlock(&dlm->spinlock);
954 spin_unlock(&dlm_domain_lock);
959 static int dlm_match_regions(struct dlm_ctxt *dlm,
960 struct dlm_query_region *qr,
961 char *local, int locallen)
963 char *remote = qr->qr_regions;
965 int localnr, i, j, foundit;
968 if (!o2hb_global_heartbeat_active()) {
969 if (qr->qr_numregions) {
970 mlog(ML_ERROR, "Domain %s: Joining node %d has global "
971 "heartbeat enabled but local node %d does not\n",
972 qr->qr_domain, qr->qr_node, dlm->node_num);
978 if (o2hb_global_heartbeat_active() && !qr->qr_numregions) {
979 mlog(ML_ERROR, "Domain %s: Local node %d has global "
980 "heartbeat enabled but joining node %d does not\n",
981 qr->qr_domain, dlm->node_num, qr->qr_node);
987 for (i = 0; i < qr->qr_numregions; ++i) {
988 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, r);
989 r += O2HB_MAX_REGION_NAME_LEN;
992 localnr = min(O2NM_MAX_REGIONS, locallen/O2HB_MAX_REGION_NAME_LEN);
993 localnr = o2hb_get_all_regions(local, (u8)localnr);
995 /* compare local regions with remote */
997 for (i = 0; i < localnr; ++i) {
1000 for (j = 0; j <= qr->qr_numregions; ++j) {
1001 if (!memcmp(l, r, O2HB_MAX_REGION_NAME_LEN)) {
1005 r += O2HB_MAX_REGION_NAME_LEN;
1009 mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1010 "in local node %d but not in joining node %d\n",
1011 qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, l,
1012 dlm->node_num, qr->qr_node);
1015 l += O2HB_MAX_REGION_NAME_LEN;
1018 /* compare remote with local regions */
1020 for (i = 0; i < qr->qr_numregions; ++i) {
1023 for (j = 0; j < localnr; ++j) {
1024 if (!memcmp(r, l, O2HB_MAX_REGION_NAME_LEN)) {
1028 l += O2HB_MAX_REGION_NAME_LEN;
1032 mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1033 "in joining node %d but not in local node %d\n",
1034 qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, r,
1035 qr->qr_node, dlm->node_num);
1038 r += O2HB_MAX_REGION_NAME_LEN;
1045 static int dlm_send_regions(struct dlm_ctxt *dlm, unsigned long *node_map)
1047 struct dlm_query_region *qr = NULL;
1048 int status, ret = 0, i;
1051 if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1054 qr = kzalloc(sizeof(struct dlm_query_region), GFP_KERNEL);
1061 qr->qr_node = dlm->node_num;
1062 qr->qr_namelen = strlen(dlm->name);
1063 memcpy(qr->qr_domain, dlm->name, qr->qr_namelen);
1064 /* if local hb, the numregions will be zero */
1065 if (o2hb_global_heartbeat_active())
1066 qr->qr_numregions = o2hb_get_all_regions(qr->qr_regions,
1070 for (i = 0; i < qr->qr_numregions; ++i, p += O2HB_MAX_REGION_NAME_LEN)
1071 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, p);
1074 while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1075 i + 1)) < O2NM_MAX_NODES) {
1076 if (i == dlm->node_num)
1079 mlog(0, "Sending regions to node %d\n", i);
1081 ret = o2net_send_message(DLM_QUERY_REGION, DLM_MOD_KEY, qr,
1082 sizeof(struct dlm_query_region),
1087 mlog(ML_ERROR, "Region mismatch %d, node %d\n",
1098 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
1099 void *data, void **ret_data)
1101 struct dlm_query_region *qr;
1102 struct dlm_ctxt *dlm = NULL;
1106 qr = (struct dlm_query_region *) msg->buf;
1108 mlog(0, "Node %u queries hb regions on domain %s\n", qr->qr_node,
1111 /* buffer used in dlm_mast_regions() */
1112 local = kmalloc(sizeof(qr->qr_regions), GFP_KERNEL);
1118 spin_lock(&dlm_domain_lock);
1119 dlm = __dlm_lookup_domain_full(qr->qr_domain, qr->qr_namelen);
1121 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1122 "before join domain\n", qr->qr_node, qr->qr_domain);
1123 goto out_domain_lock;
1126 spin_lock(&dlm->spinlock);
1127 if (dlm->joining_node != qr->qr_node) {
1128 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1129 "but joining node is %d\n", qr->qr_node, qr->qr_domain,
1134 /* Support for global heartbeat was added in 1.1 */
1135 if (dlm->dlm_locking_proto.pv_major == 1 &&
1136 dlm->dlm_locking_proto.pv_minor == 0) {
1137 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1138 "but active dlm protocol is %d.%d\n", qr->qr_node,
1139 qr->qr_domain, dlm->dlm_locking_proto.pv_major,
1140 dlm->dlm_locking_proto.pv_minor);
1144 status = dlm_match_regions(dlm, qr, local, sizeof(qr->qr_regions));
1147 spin_unlock(&dlm->spinlock);
1150 spin_unlock(&dlm_domain_lock);
1157 static int dlm_match_nodes(struct dlm_ctxt *dlm, struct dlm_query_nodeinfo *qn)
1159 struct o2nm_node *local;
1160 struct dlm_node_info *remote;
1164 for (j = 0; j < qn->qn_numnodes; ++j)
1165 mlog(0, "Node %3d, %pI4:%u\n", qn->qn_nodes[j].ni_nodenum,
1166 &(qn->qn_nodes[j].ni_ipv4_address),
1167 ntohs(qn->qn_nodes[j].ni_ipv4_port));
1169 for (i = 0; i < O2NM_MAX_NODES && !status; ++i) {
1170 local = o2nm_get_node_by_num(i);
1172 for (j = 0; j < qn->qn_numnodes; ++j) {
1173 if (qn->qn_nodes[j].ni_nodenum == i) {
1174 remote = &(qn->qn_nodes[j]);
1179 if (!local && !remote)
1182 if ((local && !remote) || (!local && remote))
1186 ((remote->ni_nodenum != local->nd_num) ||
1187 (remote->ni_ipv4_port != local->nd_ipv4_port) ||
1188 (remote->ni_ipv4_address != local->nd_ipv4_address)))
1192 if (remote && !local)
1193 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1194 "registered in joining node %d but not in "
1195 "local node %d\n", qn->qn_domain,
1197 &(remote->ni_ipv4_address),
1198 ntohs(remote->ni_ipv4_port),
1199 qn->qn_nodenum, dlm->node_num);
1200 if (local && !remote)
1201 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1202 "registered in local node %d but not in "
1203 "joining node %d\n", qn->qn_domain,
1204 local->nd_num, &(local->nd_ipv4_address),
1205 ntohs(local->nd_ipv4_port),
1206 dlm->node_num, qn->qn_nodenum);
1207 BUG_ON((!local && !remote));
1211 o2nm_node_put(local);
1217 static int dlm_send_nodeinfo(struct dlm_ctxt *dlm, unsigned long *node_map)
1219 struct dlm_query_nodeinfo *qn = NULL;
1220 struct o2nm_node *node;
1221 int ret = 0, status, count, i;
1223 if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1226 qn = kzalloc(sizeof(struct dlm_query_nodeinfo), GFP_KERNEL);
1233 for (i = 0, count = 0; i < O2NM_MAX_NODES; ++i) {
1234 node = o2nm_get_node_by_num(i);
1237 qn->qn_nodes[count].ni_nodenum = node->nd_num;
1238 qn->qn_nodes[count].ni_ipv4_port = node->nd_ipv4_port;
1239 qn->qn_nodes[count].ni_ipv4_address = node->nd_ipv4_address;
1240 mlog(0, "Node %3d, %pI4:%u\n", node->nd_num,
1241 &(node->nd_ipv4_address), ntohs(node->nd_ipv4_port));
1243 o2nm_node_put(node);
1246 qn->qn_nodenum = dlm->node_num;
1247 qn->qn_numnodes = count;
1248 qn->qn_namelen = strlen(dlm->name);
1249 memcpy(qn->qn_domain, dlm->name, qn->qn_namelen);
1252 while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1253 i + 1)) < O2NM_MAX_NODES) {
1254 if (i == dlm->node_num)
1257 mlog(0, "Sending nodeinfo to node %d\n", i);
1259 ret = o2net_send_message(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
1260 qn, sizeof(struct dlm_query_nodeinfo),
1265 mlog(ML_ERROR, "node mismatch %d, node %d\n", ret, i);
1275 static int dlm_query_nodeinfo_handler(struct o2net_msg *msg, u32 len,
1276 void *data, void **ret_data)
1278 struct dlm_query_nodeinfo *qn;
1279 struct dlm_ctxt *dlm = NULL;
1280 int locked = 0, status = -EINVAL;
1282 qn = (struct dlm_query_nodeinfo *) msg->buf;
1284 mlog(0, "Node %u queries nodes on domain %s\n", qn->qn_nodenum,
1287 spin_lock(&dlm_domain_lock);
1288 dlm = __dlm_lookup_domain_full(qn->qn_domain, qn->qn_namelen);
1290 mlog(ML_ERROR, "Node %d queried nodes on domain %s before "
1291 "join domain\n", qn->qn_nodenum, qn->qn_domain);
1295 spin_lock(&dlm->spinlock);
1297 if (dlm->joining_node != qn->qn_nodenum) {
1298 mlog(ML_ERROR, "Node %d queried nodes on domain %s but "
1299 "joining node is %d\n", qn->qn_nodenum, qn->qn_domain,
1304 /* Support for node query was added in 1.1 */
1305 if (dlm->dlm_locking_proto.pv_major == 1 &&
1306 dlm->dlm_locking_proto.pv_minor == 0) {
1307 mlog(ML_ERROR, "Node %d queried nodes on domain %s "
1308 "but active dlm protocol is %d.%d\n", qn->qn_nodenum,
1309 qn->qn_domain, dlm->dlm_locking_proto.pv_major,
1310 dlm->dlm_locking_proto.pv_minor);
1314 status = dlm_match_nodes(dlm, qn);
1318 spin_unlock(&dlm->spinlock);
1319 spin_unlock(&dlm_domain_lock);
1324 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
1327 struct dlm_cancel_join *cancel;
1328 struct dlm_ctxt *dlm = NULL;
1330 cancel = (struct dlm_cancel_join *) msg->buf;
1332 mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
1335 spin_lock(&dlm_domain_lock);
1336 dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
1339 spin_lock(&dlm->spinlock);
1341 /* Yikes, this guy wants to cancel his join. No
1342 * problem, we simply cleanup our join state. */
1343 BUG_ON(dlm->joining_node != cancel->node_idx);
1344 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1346 spin_unlock(&dlm->spinlock);
1348 spin_unlock(&dlm_domain_lock);
1353 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
1357 struct dlm_cancel_join cancel_msg;
1359 memset(&cancel_msg, 0, sizeof(cancel_msg));
1360 cancel_msg.node_idx = dlm->node_num;
1361 cancel_msg.name_len = strlen(dlm->name);
1362 memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
1364 status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1365 &cancel_msg, sizeof(cancel_msg), node,
1368 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1369 "node %u\n", status, DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1378 /* map_size should be in bytes. */
1379 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
1380 unsigned long *node_map,
1381 unsigned int map_size)
1383 int status, tmpstat;
1386 if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
1387 sizeof(unsigned long))) {
1389 "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
1390 map_size, (unsigned)BITS_TO_LONGS(O2NM_MAX_NODES));
1396 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1397 node + 1)) < O2NM_MAX_NODES) {
1398 if (node == dlm->node_num)
1401 tmpstat = dlm_send_one_join_cancel(dlm, node);
1403 mlog(ML_ERROR, "Error return %d cancelling join on "
1404 "node %d\n", tmpstat, node);
1415 static int dlm_request_join(struct dlm_ctxt *dlm,
1417 enum dlm_query_join_response_code *response)
1420 struct dlm_query_join_request join_msg;
1421 struct dlm_query_join_packet packet;
1424 mlog(0, "querying node %d\n", node);
1426 memset(&join_msg, 0, sizeof(join_msg));
1427 join_msg.node_idx = dlm->node_num;
1428 join_msg.name_len = strlen(dlm->name);
1429 memcpy(join_msg.domain, dlm->name, join_msg.name_len);
1430 join_msg.dlm_proto = dlm->dlm_locking_proto;
1431 join_msg.fs_proto = dlm->fs_locking_proto;
1433 /* copy live node map to join message */
1434 byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES);
1436 status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
1437 sizeof(join_msg), node, &join_resp);
1438 if (status < 0 && status != -ENOPROTOOPT) {
1439 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1440 "node %u\n", status, DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1444 dlm_query_join_wire_to_packet(join_resp, &packet);
1446 /* -ENOPROTOOPT from the net code means the other side isn't
1447 listening for our message type -- that's fine, it means
1448 his dlm isn't up, so we can consider him a 'yes' but not
1449 joined into the domain. */
1450 if (status == -ENOPROTOOPT) {
1452 *response = JOIN_OK_NO_MAP;
1454 *response = packet.code;
1455 switch (packet.code) {
1457 case JOIN_OK_NO_MAP:
1459 case JOIN_PROTOCOL_MISMATCH:
1461 "This node requested DLM locking protocol %u.%u and "
1462 "filesystem locking protocol %u.%u. At least one of "
1463 "the protocol versions on node %d is not compatible, "
1465 dlm->dlm_locking_proto.pv_major,
1466 dlm->dlm_locking_proto.pv_minor,
1467 dlm->fs_locking_proto.pv_major,
1468 dlm->fs_locking_proto.pv_minor,
1473 /* Use the same locking protocol as the remote node */
1474 dlm->dlm_locking_proto.pv_minor = packet.dlm_minor;
1475 dlm->fs_locking_proto.pv_minor = packet.fs_minor;
1477 "Node %d responds JOIN_OK with DLM locking protocol "
1478 "%u.%u and fs locking protocol %u.%u\n",
1480 dlm->dlm_locking_proto.pv_major,
1481 dlm->dlm_locking_proto.pv_minor,
1482 dlm->fs_locking_proto.pv_major,
1483 dlm->fs_locking_proto.pv_minor);
1487 mlog(ML_ERROR, "invalid response %d from node %u\n",
1489 /* Reset response to JOIN_DISALLOW */
1490 *response = JOIN_DISALLOW;
1495 mlog(0, "status %d, node %d response is %d\n", status, node,
1502 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
1507 struct dlm_assert_joined assert_msg;
1509 mlog(0, "Sending join assert to node %u\n", node);
1511 memset(&assert_msg, 0, sizeof(assert_msg));
1512 assert_msg.node_idx = dlm->node_num;
1513 assert_msg.name_len = strlen(dlm->name);
1514 memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
1516 status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1517 &assert_msg, sizeof(assert_msg), node,
1520 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1521 "node %u\n", status, DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1529 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
1530 unsigned long *node_map)
1532 int status, node, live;
1536 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1537 node + 1)) < O2NM_MAX_NODES) {
1538 if (node == dlm->node_num)
1542 /* It is very important that this message be
1543 * received so we spin until either the node
1544 * has died or it gets the message. */
1545 status = dlm_send_one_join_assert(dlm, node);
1547 spin_lock(&dlm->spinlock);
1548 live = test_bit(node, dlm->live_nodes_map);
1549 spin_unlock(&dlm->spinlock);
1552 mlog(ML_ERROR, "Error return %d asserting "
1553 "join on node %d\n", status, node);
1555 /* give us some time between errors... */
1557 msleep(DLM_DOMAIN_BACKOFF_MS);
1559 } while (status && live);
1563 struct domain_join_ctxt {
1564 unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1565 unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1568 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
1569 struct domain_join_ctxt *ctxt,
1570 enum dlm_query_join_response_code response)
1574 if (response == JOIN_DISALLOW) {
1575 mlog(0, "Latest response of disallow -- should restart\n");
1579 spin_lock(&dlm->spinlock);
1580 /* For now, we restart the process if the node maps have
1582 ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
1583 sizeof(dlm->live_nodes_map));
1584 spin_unlock(&dlm->spinlock);
1587 mlog(0, "Node maps changed -- should restart\n");
1592 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
1594 int status = 0, tmpstat, node;
1595 struct domain_join_ctxt *ctxt;
1596 enum dlm_query_join_response_code response = JOIN_DISALLOW;
1600 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
1607 /* group sem locking should work for us here -- we're already
1608 * registered for heartbeat events so filling this should be
1609 * atomic wrt getting those handlers called. */
1610 o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
1612 spin_lock(&dlm->spinlock);
1613 memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
1615 __dlm_set_joining_node(dlm, dlm->node_num);
1617 spin_unlock(&dlm->spinlock);
1620 while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
1621 node + 1)) < O2NM_MAX_NODES) {
1622 if (node == dlm->node_num)
1625 status = dlm_request_join(dlm, node, &response);
1631 /* Ok, either we got a response or the node doesn't have a
1633 if (response == JOIN_OK)
1634 set_bit(node, ctxt->yes_resp_map);
1636 if (dlm_should_restart_join(dlm, ctxt, response)) {
1642 mlog(0, "Yay, done querying nodes!\n");
1644 /* Yay, everyone agree's we can join the domain. My domain is
1645 * comprised of all nodes who were put in the
1646 * yes_resp_map. Copy that into our domain map and send a join
1647 * assert message to clean up everyone elses state. */
1648 spin_lock(&dlm->spinlock);
1649 memcpy(dlm->domain_map, ctxt->yes_resp_map,
1650 sizeof(ctxt->yes_resp_map));
1651 set_bit(dlm->node_num, dlm->domain_map);
1652 spin_unlock(&dlm->spinlock);
1654 /* Support for global heartbeat and node info was added in 1.1 */
1655 if (dlm->dlm_locking_proto.pv_major > 1 ||
1656 dlm->dlm_locking_proto.pv_minor > 0) {
1657 status = dlm_send_nodeinfo(dlm, ctxt->yes_resp_map);
1662 status = dlm_send_regions(dlm, ctxt->yes_resp_map);
1669 dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1671 /* Joined state *must* be set before the joining node
1672 * information, otherwise the query_join handler may read no
1673 * current joiner but a state of NEW and tell joining nodes
1674 * we're not in the domain. */
1675 spin_lock(&dlm_domain_lock);
1676 dlm->dlm_state = DLM_CTXT_JOINED;
1678 spin_unlock(&dlm_domain_lock);
1681 spin_lock(&dlm->spinlock);
1682 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1684 printk(KERN_NOTICE "o2dlm: Joining domain %s ", dlm->name);
1685 __dlm_print_nodes(dlm);
1687 spin_unlock(&dlm->spinlock);
1690 /* Do we need to send a cancel message to any nodes? */
1692 tmpstat = dlm_send_join_cancels(dlm,
1694 sizeof(ctxt->yes_resp_map));
1696 mlog_errno(tmpstat);
1701 mlog(0, "returning %d\n", status);
1705 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1707 o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_up);
1708 o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_down);
1709 o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1712 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1716 mlog(0, "registering handlers.\n");
1718 o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1719 dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1720 o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1721 dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1723 status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_down);
1727 status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_up);
1731 status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1732 sizeof(struct dlm_master_request),
1733 dlm_master_request_handler,
1734 dlm, NULL, &dlm->dlm_domain_handlers);
1738 status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1739 sizeof(struct dlm_assert_master),
1740 dlm_assert_master_handler,
1741 dlm, dlm_assert_master_post_handler,
1742 &dlm->dlm_domain_handlers);
1746 status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1747 sizeof(struct dlm_create_lock),
1748 dlm_create_lock_handler,
1749 dlm, NULL, &dlm->dlm_domain_handlers);
1753 status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1754 DLM_CONVERT_LOCK_MAX_LEN,
1755 dlm_convert_lock_handler,
1756 dlm, NULL, &dlm->dlm_domain_handlers);
1760 status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1761 DLM_UNLOCK_LOCK_MAX_LEN,
1762 dlm_unlock_lock_handler,
1763 dlm, NULL, &dlm->dlm_domain_handlers);
1767 status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1768 DLM_PROXY_AST_MAX_LEN,
1769 dlm_proxy_ast_handler,
1770 dlm, NULL, &dlm->dlm_domain_handlers);
1774 status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1775 sizeof(struct dlm_exit_domain),
1776 dlm_exit_domain_handler,
1777 dlm, NULL, &dlm->dlm_domain_handlers);
1781 status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1782 sizeof(struct dlm_deref_lockres),
1783 dlm_deref_lockres_handler,
1784 dlm, NULL, &dlm->dlm_domain_handlers);
1788 status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1789 sizeof(struct dlm_migrate_request),
1790 dlm_migrate_request_handler,
1791 dlm, NULL, &dlm->dlm_domain_handlers);
1795 status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1796 DLM_MIG_LOCKRES_MAX_LEN,
1797 dlm_mig_lockres_handler,
1798 dlm, NULL, &dlm->dlm_domain_handlers);
1802 status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1803 sizeof(struct dlm_master_requery),
1804 dlm_master_requery_handler,
1805 dlm, NULL, &dlm->dlm_domain_handlers);
1809 status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1810 sizeof(struct dlm_lock_request),
1811 dlm_request_all_locks_handler,
1812 dlm, NULL, &dlm->dlm_domain_handlers);
1816 status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1817 sizeof(struct dlm_reco_data_done),
1818 dlm_reco_data_done_handler,
1819 dlm, NULL, &dlm->dlm_domain_handlers);
1823 status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1824 sizeof(struct dlm_begin_reco),
1825 dlm_begin_reco_handler,
1826 dlm, NULL, &dlm->dlm_domain_handlers);
1830 status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1831 sizeof(struct dlm_finalize_reco),
1832 dlm_finalize_reco_handler,
1833 dlm, NULL, &dlm->dlm_domain_handlers);
1837 status = o2net_register_handler(DLM_BEGIN_EXIT_DOMAIN_MSG, dlm->key,
1838 sizeof(struct dlm_exit_domain),
1839 dlm_begin_exit_domain_handler,
1840 dlm, NULL, &dlm->dlm_domain_handlers);
1844 status = o2net_register_handler(DLM_DEREF_LOCKRES_DONE, dlm->key,
1845 sizeof(struct dlm_deref_lockres_done),
1846 dlm_deref_lockres_done_handler,
1847 dlm, NULL, &dlm->dlm_domain_handlers);
1850 dlm_unregister_domain_handlers(dlm);
1855 static int dlm_join_domain(struct dlm_ctxt *dlm)
1858 unsigned int backoff;
1859 unsigned int total_backoff = 0;
1860 char wq_name[O2NM_MAX_NAME_LEN];
1864 mlog(0, "Join domain %s\n", dlm->name);
1866 status = dlm_register_domain_handlers(dlm);
1872 status = dlm_launch_thread(dlm);
1878 status = dlm_launch_recovery_thread(dlm);
1884 dlm_debug_init(dlm);
1886 snprintf(wq_name, O2NM_MAX_NAME_LEN, "dlm_wq-%s", dlm->name);
1887 dlm->dlm_worker = alloc_workqueue(wq_name, WQ_MEM_RECLAIM, 0);
1888 if (!dlm->dlm_worker) {
1895 status = dlm_try_to_join_domain(dlm);
1897 /* If we're racing another node to the join, then we
1898 * need to back off temporarily and let them
1900 #define DLM_JOIN_TIMEOUT_MSECS 90000
1901 if (status == -EAGAIN) {
1902 if (signal_pending(current)) {
1903 status = -ERESTARTSYS;
1907 if (total_backoff > DLM_JOIN_TIMEOUT_MSECS) {
1908 status = -ERESTARTSYS;
1909 mlog(ML_NOTICE, "Timed out joining dlm domain "
1910 "%s after %u msecs\n", dlm->name,
1917 * <dale> No, after you!
1919 * <dale> But you first!
1922 backoff = (unsigned int)(jiffies & 0x3);
1923 backoff *= DLM_DOMAIN_BACKOFF_MS;
1924 total_backoff += backoff;
1925 mlog(0, "backoff %d\n", backoff);
1928 } while (status == -EAGAIN);
1937 wake_up(&dlm_domain_events);
1940 dlm_unregister_domain_handlers(dlm);
1941 dlm_debug_shutdown(dlm);
1942 dlm_complete_thread(dlm);
1943 dlm_complete_recovery_thread(dlm);
1944 dlm_destroy_dlm_worker(dlm);
1950 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1955 struct dlm_ctxt *dlm = NULL;
1957 dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1964 dlm->name = kstrdup(domain, GFP_KERNEL);
1965 if (dlm->name == NULL) {
1971 dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1972 if (!dlm->lockres_hash) {
1978 for (i = 0; i < DLM_HASH_BUCKETS; i++)
1979 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1981 dlm->master_hash = (struct hlist_head **)
1982 dlm_alloc_pagevec(DLM_HASH_PAGES);
1983 if (!dlm->master_hash) {
1989 for (i = 0; i < DLM_HASH_BUCKETS; i++)
1990 INIT_HLIST_HEAD(dlm_master_hash(dlm, i));
1993 dlm->node_num = o2nm_this_node();
1995 ret = dlm_create_debugfs_subroot(dlm);
1999 spin_lock_init(&dlm->spinlock);
2000 spin_lock_init(&dlm->master_lock);
2001 spin_lock_init(&dlm->ast_lock);
2002 spin_lock_init(&dlm->track_lock);
2003 INIT_LIST_HEAD(&dlm->list);
2004 INIT_LIST_HEAD(&dlm->dirty_list);
2005 INIT_LIST_HEAD(&dlm->reco.resources);
2006 INIT_LIST_HEAD(&dlm->reco.node_data);
2007 INIT_LIST_HEAD(&dlm->purge_list);
2008 INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
2009 INIT_LIST_HEAD(&dlm->tracking_list);
2010 dlm->reco.state = 0;
2012 INIT_LIST_HEAD(&dlm->pending_asts);
2013 INIT_LIST_HEAD(&dlm->pending_basts);
2015 mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
2016 dlm->recovery_map, &(dlm->recovery_map[0]));
2018 memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
2019 memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
2020 memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
2022 dlm->dlm_thread_task = NULL;
2023 dlm->dlm_reco_thread_task = NULL;
2024 dlm->dlm_worker = NULL;
2025 init_waitqueue_head(&dlm->dlm_thread_wq);
2026 init_waitqueue_head(&dlm->dlm_reco_thread_wq);
2027 init_waitqueue_head(&dlm->reco.event);
2028 init_waitqueue_head(&dlm->ast_wq);
2029 init_waitqueue_head(&dlm->migration_wq);
2030 INIT_LIST_HEAD(&dlm->mle_hb_events);
2032 dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
2033 init_waitqueue_head(&dlm->dlm_join_events);
2035 dlm->migrate_done = 0;
2037 dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
2038 dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
2040 atomic_set(&dlm->res_tot_count, 0);
2041 atomic_set(&dlm->res_cur_count, 0);
2042 for (i = 0; i < DLM_MLE_NUM_TYPES; ++i) {
2043 atomic_set(&dlm->mle_tot_count[i], 0);
2044 atomic_set(&dlm->mle_cur_count[i], 0);
2047 spin_lock_init(&dlm->work_lock);
2048 INIT_LIST_HEAD(&dlm->work_list);
2049 INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
2051 kref_init(&dlm->dlm_refs);
2052 dlm->dlm_state = DLM_CTXT_NEW;
2054 INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
2056 mlog(0, "context init: refcount %u\n",
2057 kref_read(&dlm->dlm_refs));
2060 if (ret < 0 && dlm) {
2061 if (dlm->master_hash)
2062 dlm_free_pagevec((void **)dlm->master_hash,
2065 if (dlm->lockres_hash)
2066 dlm_free_pagevec((void **)dlm->lockres_hash,
2077 * Compare a requested locking protocol version against the current one.
2079 * If the major numbers are different, they are incompatible.
2080 * If the current minor is greater than the request, they are incompatible.
2081 * If the current minor is less than or equal to the request, they are
2082 * compatible, and the requester should run at the current minor version.
2084 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
2085 struct dlm_protocol_version *request)
2087 if (existing->pv_major != request->pv_major)
2090 if (existing->pv_minor > request->pv_minor)
2093 if (existing->pv_minor < request->pv_minor)
2094 request->pv_minor = existing->pv_minor;
2100 * dlm_register_domain: one-time setup per "domain".
2102 * The filesystem passes in the requested locking version via proto.
2103 * If registration was successful, proto will contain the negotiated
2106 struct dlm_ctxt * dlm_register_domain(const char *domain,
2108 struct dlm_protocol_version *fs_proto)
2111 struct dlm_ctxt *dlm = NULL;
2112 struct dlm_ctxt *new_ctxt = NULL;
2114 if (strlen(domain) >= O2NM_MAX_NAME_LEN) {
2115 ret = -ENAMETOOLONG;
2116 mlog(ML_ERROR, "domain name length too long\n");
2120 mlog(0, "register called for domain \"%s\"\n", domain);
2124 if (signal_pending(current)) {
2130 spin_lock(&dlm_domain_lock);
2132 dlm = __dlm_lookup_domain(domain);
2134 if (dlm->dlm_state != DLM_CTXT_JOINED) {
2135 spin_unlock(&dlm_domain_lock);
2137 mlog(0, "This ctxt is not joined yet!\n");
2138 wait_event_interruptible(dlm_domain_events,
2139 dlm_wait_on_domain_helper(
2144 if (dlm_protocol_compare(&dlm->fs_locking_proto, fs_proto)) {
2145 spin_unlock(&dlm_domain_lock);
2147 "Requested locking protocol version is not "
2148 "compatible with already registered domain "
2149 "\"%s\"\n", domain);
2157 spin_unlock(&dlm_domain_lock);
2165 spin_unlock(&dlm_domain_lock);
2167 new_ctxt = dlm_alloc_ctxt(domain, key);
2176 /* a little variable switch-a-roo here... */
2180 /* add the new domain */
2181 list_add_tail(&dlm->list, &dlm_domains);
2182 spin_unlock(&dlm_domain_lock);
2185 * Pass the locking protocol version into the join. If the join
2186 * succeeds, it will have the negotiated protocol set.
2188 dlm->dlm_locking_proto = dlm_protocol;
2189 dlm->fs_locking_proto = *fs_proto;
2191 ret = dlm_join_domain(dlm);
2198 /* Tell the caller what locking protocol we negotiated */
2199 *fs_proto = dlm->fs_locking_proto;
2204 dlm_free_ctxt_mem(new_ctxt);
2211 EXPORT_SYMBOL_GPL(dlm_register_domain);
2213 static LIST_HEAD(dlm_join_handlers);
2215 static void dlm_unregister_net_handlers(void)
2217 o2net_unregister_handler_list(&dlm_join_handlers);
2220 static int dlm_register_net_handlers(void)
2224 status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
2225 sizeof(struct dlm_query_join_request),
2226 dlm_query_join_handler,
2227 NULL, NULL, &dlm_join_handlers);
2231 status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
2232 sizeof(struct dlm_assert_joined),
2233 dlm_assert_joined_handler,
2234 NULL, NULL, &dlm_join_handlers);
2238 status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
2239 sizeof(struct dlm_cancel_join),
2240 dlm_cancel_join_handler,
2241 NULL, NULL, &dlm_join_handlers);
2245 status = o2net_register_handler(DLM_QUERY_REGION, DLM_MOD_KEY,
2246 sizeof(struct dlm_query_region),
2247 dlm_query_region_handler,
2248 NULL, NULL, &dlm_join_handlers);
2253 status = o2net_register_handler(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
2254 sizeof(struct dlm_query_nodeinfo),
2255 dlm_query_nodeinfo_handler,
2256 NULL, NULL, &dlm_join_handlers);
2259 dlm_unregister_net_handlers();
2264 /* Domain eviction callback handling.
2266 * The file system requires notification of node death *before* the
2267 * dlm completes it's recovery work, otherwise it may be able to
2268 * acquire locks on resources requiring recovery. Since the dlm can
2269 * evict a node from it's domain *before* heartbeat fires, a similar
2270 * mechanism is required. */
2272 /* Eviction is not expected to happen often, so a per-domain lock is
2273 * not necessary. Eviction callbacks are allowed to sleep for short
2274 * periods of time. */
2275 static DECLARE_RWSEM(dlm_callback_sem);
2277 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
2280 struct dlm_eviction_cb *cb;
2282 down_read(&dlm_callback_sem);
2283 list_for_each_entry(cb, &dlm->dlm_eviction_callbacks, ec_item) {
2284 cb->ec_func(node_num, cb->ec_data);
2286 up_read(&dlm_callback_sem);
2289 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
2290 dlm_eviction_func *f,
2293 INIT_LIST_HEAD(&cb->ec_item);
2297 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
2299 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
2300 struct dlm_eviction_cb *cb)
2302 down_write(&dlm_callback_sem);
2303 list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
2304 up_write(&dlm_callback_sem);
2306 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
2308 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
2310 down_write(&dlm_callback_sem);
2311 list_del_init(&cb->ec_item);
2312 up_write(&dlm_callback_sem);
2314 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
2316 static int __init dlm_init(void)
2320 status = dlm_init_mle_cache();
2322 mlog(ML_ERROR, "Could not create o2dlm_mle slabcache\n");
2326 status = dlm_init_master_caches();
2328 mlog(ML_ERROR, "Could not create o2dlm_lockres and "
2329 "o2dlm_lockname slabcaches\n");
2333 status = dlm_init_lock_cache();
2335 mlog(ML_ERROR, "Count not create o2dlm_lock slabcache\n");
2339 status = dlm_register_net_handlers();
2341 mlog(ML_ERROR, "Unable to register network handlers\n");
2345 dlm_create_debugfs_root();
2349 dlm_unregister_net_handlers();
2350 dlm_destroy_lock_cache();
2351 dlm_destroy_master_caches();
2352 dlm_destroy_mle_cache();
2356 static void __exit dlm_exit (void)
2358 dlm_destroy_debugfs_root();
2359 dlm_unregister_net_handlers();
2360 dlm_destroy_lock_cache();
2361 dlm_destroy_master_caches();
2362 dlm_destroy_mle_cache();
2365 MODULE_AUTHOR("Oracle");
2366 MODULE_LICENSE("GPL");
2367 MODULE_DESCRIPTION("OCFS2 Distributed Lock Management");
2369 module_init(dlm_init);
2370 module_exit(dlm_exit);