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>
36 #include <linux/sched/signal.h>
38 #include "cluster/heartbeat.h"
39 #include "cluster/nodemanager.h"
40 #include "cluster/tcp.h"
43 #include "dlmcommon.h"
44 #include "dlmdomain.h"
47 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
48 #include "cluster/masklog.h"
51 * ocfs2 node maps are array of long int, which limits to send them freely
52 * across the wire due to endianness issues. To workaround this, we convert
53 * long ints to byte arrays. Following 3 routines are helper functions to
54 * set/test/copy bits within those array of bytes
56 static inline void byte_set_bit(u8 nr, u8 map[])
58 map[nr >> 3] |= (1UL << (nr & 7));
61 static inline int byte_test_bit(u8 nr, u8 map[])
63 return ((1UL << (nr & 7)) & (map[nr >> 3])) != 0;
66 static inline void byte_copymap(u8 dmap[], unsigned long smap[],
74 memset(dmap, 0, ((sz + 7) >> 3));
75 for (nn = 0 ; nn < sz; nn++)
76 if (test_bit(nn, smap))
77 byte_set_bit(nn, dmap);
80 static void dlm_free_pagevec(void **vec, int pages)
83 free_page((unsigned long)vec[pages]);
87 static void **dlm_alloc_pagevec(int pages)
89 void **vec = kmalloc(pages * sizeof(void *), GFP_KERNEL);
95 for (i = 0; i < pages; i++)
96 if (!(vec[i] = (void *)__get_free_page(GFP_KERNEL)))
99 mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n",
100 pages, (unsigned long)DLM_HASH_PAGES,
101 (unsigned long)DLM_BUCKETS_PER_PAGE);
104 dlm_free_pagevec(vec, i);
110 * spinlock lock ordering: if multiple locks are needed, obey this ordering:
112 * struct dlm_ctxt->spinlock
113 * struct dlm_lock_resource->spinlock
114 * struct dlm_ctxt->master_lock
115 * struct dlm_ctxt->ast_lock
116 * dlm_master_list_entry->spinlock
121 DEFINE_SPINLOCK(dlm_domain_lock);
122 LIST_HEAD(dlm_domains);
123 static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);
126 * The supported protocol version for DLM communication. Running domains
127 * will have a negotiated version with the same major number and a minor
128 * number equal or smaller. The dlm_ctxt->dlm_locking_proto field should
129 * be used to determine what a running domain is actually using.
131 * New in version 1.1:
132 * - Message DLM_QUERY_REGION added to support global heartbeat
133 * - Message DLM_QUERY_NODEINFO added to allow online node removes
134 * New in version 1.2:
135 * - Message DLM_BEGIN_EXIT_DOMAIN_MSG added to mark start of exit domain
136 * New in version 1.3:
137 * - Message DLM_DEREF_LOCKRES_DONE added to inform non-master that the
140 static const struct dlm_protocol_version dlm_protocol = {
145 #define DLM_DOMAIN_BACKOFF_MS 200
147 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
149 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
151 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
153 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
154 void *data, void **ret_data);
155 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
157 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
158 struct dlm_protocol_version *request);
160 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
162 void __dlm_unhash_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
164 if (hlist_unhashed(&res->hash_node))
167 mlog(0, "%s: Unhash res %.*s\n", dlm->name, res->lockname.len,
169 hlist_del_init(&res->hash_node);
170 dlm_lockres_put(res);
173 void __dlm_insert_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
175 struct hlist_head *bucket;
177 assert_spin_locked(&dlm->spinlock);
179 bucket = dlm_lockres_hash(dlm, res->lockname.hash);
181 /* get a reference for our hashtable */
182 dlm_lockres_get(res);
184 hlist_add_head(&res->hash_node, bucket);
186 mlog(0, "%s: Hash res %.*s\n", dlm->name, res->lockname.len,
190 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
195 struct hlist_head *bucket;
196 struct dlm_lock_resource *res;
198 mlog(0, "%.*s\n", len, name);
200 assert_spin_locked(&dlm->spinlock);
202 bucket = dlm_lockres_hash(dlm, hash);
204 hlist_for_each_entry(res, bucket, hash_node) {
205 if (res->lockname.name[0] != name[0])
207 if (unlikely(res->lockname.len != len))
209 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
211 dlm_lockres_get(res);
217 /* intended to be called by functions which do not care about lock
218 * resources which are being purged (most net _handler functions).
219 * this will return NULL for any lock resource which is found but
220 * currently in the process of dropping its mastery reference.
221 * use __dlm_lookup_lockres_full when you need the lock resource
222 * regardless (e.g. dlm_get_lock_resource) */
223 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
228 struct dlm_lock_resource *res = NULL;
230 mlog(0, "%.*s\n", len, name);
232 assert_spin_locked(&dlm->spinlock);
234 res = __dlm_lookup_lockres_full(dlm, name, len, hash);
236 spin_lock(&res->spinlock);
237 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
238 spin_unlock(&res->spinlock);
239 dlm_lockres_put(res);
242 spin_unlock(&res->spinlock);
248 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
252 struct dlm_lock_resource *res;
253 unsigned int hash = dlm_lockid_hash(name, len);
255 spin_lock(&dlm->spinlock);
256 res = __dlm_lookup_lockres(dlm, name, len, hash);
257 spin_unlock(&dlm->spinlock);
261 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
263 struct dlm_ctxt *tmp;
265 assert_spin_locked(&dlm_domain_lock);
267 /* tmp->name here is always NULL terminated,
268 * but domain may not be! */
269 list_for_each_entry(tmp, &dlm_domains, list) {
270 if (strlen(tmp->name) == len &&
271 memcmp(tmp->name, domain, len)==0)
278 /* For null terminated domain strings ONLY */
279 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
281 assert_spin_locked(&dlm_domain_lock);
283 return __dlm_lookup_domain_full(domain, strlen(domain));
287 /* returns true on one of two conditions:
288 * 1) the domain does not exist
289 * 2) the domain exists and it's state is "joined" */
290 static int dlm_wait_on_domain_helper(const char *domain)
293 struct dlm_ctxt *tmp = NULL;
295 spin_lock(&dlm_domain_lock);
297 tmp = __dlm_lookup_domain(domain);
300 else if (tmp->dlm_state == DLM_CTXT_JOINED)
303 spin_unlock(&dlm_domain_lock);
307 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
309 dlm_destroy_debugfs_subroot(dlm);
311 if (dlm->lockres_hash)
312 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
314 if (dlm->master_hash)
315 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
321 /* A little strange - this function will be called while holding
322 * dlm_domain_lock and is expected to be holding it on the way out. We
323 * will however drop and reacquire it multiple times */
324 static void dlm_ctxt_release(struct kref *kref)
326 struct dlm_ctxt *dlm;
328 dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
330 BUG_ON(dlm->num_joins);
331 BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
333 /* we may still be in the list if we hit an error during join. */
334 list_del_init(&dlm->list);
336 spin_unlock(&dlm_domain_lock);
338 mlog(0, "freeing memory from domain %s\n", dlm->name);
340 wake_up(&dlm_domain_events);
342 dlm_free_ctxt_mem(dlm);
344 spin_lock(&dlm_domain_lock);
347 void dlm_put(struct dlm_ctxt *dlm)
349 spin_lock(&dlm_domain_lock);
350 kref_put(&dlm->dlm_refs, dlm_ctxt_release);
351 spin_unlock(&dlm_domain_lock);
354 static void __dlm_get(struct dlm_ctxt *dlm)
356 kref_get(&dlm->dlm_refs);
359 /* given a questionable reference to a dlm object, gets a reference if
360 * it can find it in the list, otherwise returns NULL in which case
361 * you shouldn't trust your pointer. */
362 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
364 struct dlm_ctxt *target;
365 struct dlm_ctxt *ret = NULL;
367 spin_lock(&dlm_domain_lock);
369 list_for_each_entry(target, &dlm_domains, list) {
377 spin_unlock(&dlm_domain_lock);
382 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
386 spin_lock(&dlm_domain_lock);
387 ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
388 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
389 spin_unlock(&dlm_domain_lock);
394 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
396 if (dlm->dlm_worker) {
397 flush_workqueue(dlm->dlm_worker);
398 destroy_workqueue(dlm->dlm_worker);
399 dlm->dlm_worker = NULL;
403 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
405 dlm_unregister_domain_handlers(dlm);
406 dlm_debug_shutdown(dlm);
407 dlm_complete_thread(dlm);
408 dlm_complete_recovery_thread(dlm);
409 dlm_destroy_dlm_worker(dlm);
411 /* We've left the domain. Now we can take ourselves out of the
412 * list and allow the kref stuff to help us free the
414 spin_lock(&dlm_domain_lock);
415 list_del_init(&dlm->list);
416 spin_unlock(&dlm_domain_lock);
418 /* Wake up anyone waiting for us to remove this domain */
419 wake_up(&dlm_domain_events);
422 static int dlm_migrate_all_locks(struct dlm_ctxt *dlm)
424 int i, num, n, ret = 0;
425 struct dlm_lock_resource *res;
426 struct hlist_node *iter;
427 struct hlist_head *bucket;
430 mlog(0, "Migrating locks from domain %s\n", dlm->name);
433 spin_lock(&dlm->spinlock);
434 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
437 bucket = dlm_lockres_hash(dlm, i);
438 iter = bucket->first;
441 res = hlist_entry(iter, struct dlm_lock_resource,
443 dlm_lockres_get(res);
444 /* migrate, if necessary. this will drop the dlm
445 * spinlock and retake it if it does migration. */
446 dropped = dlm_empty_lockres(dlm, res);
448 spin_lock(&res->spinlock);
450 __dlm_lockres_calc_usage(dlm, res);
452 iter = res->hash_node.next;
453 spin_unlock(&res->spinlock);
455 dlm_lockres_put(res);
458 cond_resched_lock(&dlm->spinlock);
462 cond_resched_lock(&dlm->spinlock);
465 spin_unlock(&dlm->spinlock);
466 wake_up(&dlm->dlm_thread_wq);
468 /* let the dlm thread take care of purging, keep scanning until
469 * nothing remains in the hash */
471 mlog(0, "%s: %d lock resources in hash last pass\n",
475 mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
479 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
483 spin_lock(&dlm->spinlock);
484 ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
485 spin_unlock(&dlm->spinlock);
490 static int dlm_begin_exit_domain_handler(struct o2net_msg *msg, u32 len,
491 void *data, void **ret_data)
493 struct dlm_ctxt *dlm = data;
495 struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
500 node = exit_msg->node_idx;
501 mlog(0, "%s: Node %u sent a begin exit domain message\n", dlm->name, node);
503 spin_lock(&dlm->spinlock);
504 set_bit(node, dlm->exit_domain_map);
505 spin_unlock(&dlm->spinlock);
512 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
514 /* Yikes, a double spinlock! I need domain_lock for the dlm
515 * state and the dlm spinlock for join state... Sorry! */
517 spin_lock(&dlm_domain_lock);
518 spin_lock(&dlm->spinlock);
520 if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
521 mlog(0, "Node %d is joining, we wait on it.\n",
523 spin_unlock(&dlm->spinlock);
524 spin_unlock(&dlm_domain_lock);
526 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
530 dlm->dlm_state = DLM_CTXT_LEAVING;
531 spin_unlock(&dlm->spinlock);
532 spin_unlock(&dlm_domain_lock);
535 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
537 int node = -1, num = 0;
539 assert_spin_locked(&dlm->spinlock);
542 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
543 node + 1)) < O2NM_MAX_NODES) {
547 printk(") %u nodes\n", num);
550 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
553 struct dlm_ctxt *dlm = data;
555 struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
557 mlog(0, "%p %u %p", msg, len, data);
562 node = exit_msg->node_idx;
564 spin_lock(&dlm->spinlock);
565 clear_bit(node, dlm->domain_map);
566 clear_bit(node, dlm->exit_domain_map);
567 printk(KERN_NOTICE "o2dlm: Node %u leaves domain %s ", node, dlm->name);
568 __dlm_print_nodes(dlm);
570 /* notify anything attached to the heartbeat events */
571 dlm_hb_event_notify_attached(dlm, node, 0);
573 spin_unlock(&dlm->spinlock);
580 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm, u32 msg_type,
584 struct dlm_exit_domain leave_msg;
586 mlog(0, "%s: Sending domain exit message %u to node %u\n", dlm->name,
589 memset(&leave_msg, 0, sizeof(leave_msg));
590 leave_msg.node_idx = dlm->node_num;
592 status = o2net_send_message(msg_type, dlm->key, &leave_msg,
593 sizeof(leave_msg), node, NULL);
595 mlog(ML_ERROR, "Error %d sending domain exit message %u "
596 "to node %u on domain %s\n", status, msg_type, node,
602 static void dlm_begin_exit_domain(struct dlm_ctxt *dlm)
606 /* Support for begin exit domain was added in 1.2 */
607 if (dlm->dlm_locking_proto.pv_major == 1 &&
608 dlm->dlm_locking_proto.pv_minor < 2)
612 * Unlike DLM_EXIT_DOMAIN_MSG, DLM_BEGIN_EXIT_DOMAIN_MSG is purely
613 * informational. Meaning if a node does not receive the message,
616 spin_lock(&dlm->spinlock);
618 node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, node + 1);
619 if (node >= O2NM_MAX_NODES)
621 if (node == dlm->node_num)
624 spin_unlock(&dlm->spinlock);
625 dlm_send_one_domain_exit(dlm, DLM_BEGIN_EXIT_DOMAIN_MSG, node);
626 spin_lock(&dlm->spinlock);
628 spin_unlock(&dlm->spinlock);
631 static void dlm_leave_domain(struct dlm_ctxt *dlm)
633 int node, clear_node, status;
635 /* At this point we've migrated away all our locks and won't
636 * accept mastership of new ones. The dlm is responsible for
637 * almost nothing now. We make sure not to confuse any joining
638 * nodes and then commence shutdown procedure. */
640 spin_lock(&dlm->spinlock);
641 /* Clear ourselves from the domain map */
642 clear_bit(dlm->node_num, dlm->domain_map);
643 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
644 0)) < O2NM_MAX_NODES) {
645 /* Drop the dlm spinlock. This is safe wrt the domain_map.
646 * -nodes cannot be added now as the
647 * query_join_handlers knows to respond with OK_NO_MAP
648 * -we catch the right network errors if a node is
649 * removed from the map while we're sending him the
651 spin_unlock(&dlm->spinlock);
655 status = dlm_send_one_domain_exit(dlm, DLM_EXIT_DOMAIN_MSG,
658 status != -ENOPROTOOPT &&
659 status != -ENOTCONN) {
660 mlog(ML_NOTICE, "Error %d sending domain exit message "
661 "to node %d\n", status, node);
663 /* Not sure what to do here but lets sleep for
664 * a bit in case this was a transient
666 msleep(DLM_DOMAIN_BACKOFF_MS);
670 spin_lock(&dlm->spinlock);
671 /* If we're not clearing the node bit then we intend
672 * to loop back around to try again. */
674 clear_bit(node, dlm->domain_map);
676 spin_unlock(&dlm->spinlock);
679 void dlm_unregister_domain(struct dlm_ctxt *dlm)
682 struct dlm_lock_resource *res;
684 spin_lock(&dlm_domain_lock);
685 BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
686 BUG_ON(!dlm->num_joins);
689 if (!dlm->num_joins) {
690 /* We mark it "in shutdown" now so new register
691 * requests wait until we've completely left the
692 * domain. Don't use DLM_CTXT_LEAVING yet as we still
693 * want new domain joins to communicate with us at
694 * least until we've completed migration of our
696 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
699 spin_unlock(&dlm_domain_lock);
702 mlog(0, "shutting down domain %s\n", dlm->name);
703 dlm_begin_exit_domain(dlm);
705 /* We changed dlm state, notify the thread */
706 dlm_kick_thread(dlm, NULL);
708 while (dlm_migrate_all_locks(dlm)) {
709 /* Give dlm_thread time to purge the lockres' */
711 mlog(0, "%s: more migration to do\n", dlm->name);
714 /* This list should be empty. If not, print remaining lockres */
715 if (!list_empty(&dlm->tracking_list)) {
716 mlog(ML_ERROR, "Following lockres' are still on the "
718 list_for_each_entry(res, &dlm->tracking_list, tracking)
719 dlm_print_one_lock_resource(res);
722 dlm_mark_domain_leaving(dlm);
723 dlm_leave_domain(dlm);
724 printk(KERN_NOTICE "o2dlm: Leaving domain %s\n", dlm->name);
725 dlm_force_free_mles(dlm);
726 dlm_complete_dlm_shutdown(dlm);
730 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
732 static int dlm_query_join_proto_check(char *proto_type, int node,
733 struct dlm_protocol_version *ours,
734 struct dlm_protocol_version *request)
737 struct dlm_protocol_version proto = *request;
739 if (!dlm_protocol_compare(ours, &proto)) {
741 "node %u wanted to join with %s locking protocol "
742 "%u.%u, we respond with %u.%u\n",
746 proto.pv_major, proto.pv_minor);
747 request->pv_minor = proto.pv_minor;
751 "Node %u wanted to join with %s locking "
752 "protocol %u.%u, but we have %u.%u, disallowing\n",
765 * struct dlm_query_join_packet is made up of four one-byte fields. They
766 * are effectively in big-endian order already. However, little-endian
767 * machines swap them before putting the packet on the wire (because
768 * query_join's response is a status, and that status is treated as a u32
769 * on the wire). Thus, a big-endian and little-endian machines will treat
770 * this structure differently.
772 * The solution is to have little-endian machines swap the structure when
773 * converting from the structure to the u32 representation. This will
774 * result in the structure having the correct format on the wire no matter
775 * the host endian format.
777 static void dlm_query_join_packet_to_wire(struct dlm_query_join_packet *packet,
780 union dlm_query_join_response response;
782 response.packet = *packet;
783 *wire = be32_to_cpu(response.intval);
786 static void dlm_query_join_wire_to_packet(u32 wire,
787 struct dlm_query_join_packet *packet)
789 union dlm_query_join_response response;
791 response.intval = cpu_to_be32(wire);
792 *packet = response.packet;
795 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
798 struct dlm_query_join_request *query;
799 struct dlm_query_join_packet packet = {
800 .code = JOIN_DISALLOW,
802 struct dlm_ctxt *dlm = NULL;
806 query = (struct dlm_query_join_request *) msg->buf;
808 mlog(0, "node %u wants to join domain %s\n", query->node_idx,
812 * If heartbeat doesn't consider the node live, tell it
813 * to back off and try again. This gives heartbeat a chance
816 if (!o2hb_check_node_heartbeating_no_sem(query->node_idx)) {
817 mlog(0, "node %u is not in our live map yet\n",
820 packet.code = JOIN_DISALLOW;
824 packet.code = JOIN_OK_NO_MAP;
826 spin_lock(&dlm_domain_lock);
827 dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
832 * There is a small window where the joining node may not see the
833 * node(s) that just left but still part of the cluster. DISALLOW
834 * join request if joining node has different node map.
837 while (nodenum < O2NM_MAX_NODES) {
838 if (test_bit(nodenum, dlm->domain_map)) {
839 if (!byte_test_bit(nodenum, query->node_map)) {
840 mlog(0, "disallow join as node %u does not "
841 "have node %u in its nodemap\n",
842 query->node_idx, nodenum);
843 packet.code = JOIN_DISALLOW;
850 /* Once the dlm ctxt is marked as leaving then we don't want
851 * to be put in someone's domain map.
852 * Also, explicitly disallow joining at certain troublesome
853 * times (ie. during recovery). */
854 if (dlm->dlm_state != DLM_CTXT_LEAVING) {
855 int bit = query->node_idx;
856 spin_lock(&dlm->spinlock);
858 if (dlm->dlm_state == DLM_CTXT_NEW &&
859 dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
860 /*If this is a brand new context and we
861 * haven't started our join process yet, then
862 * the other node won the race. */
863 packet.code = JOIN_OK_NO_MAP;
864 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
865 /* Disallow parallel joins. */
866 packet.code = JOIN_DISALLOW;
867 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
868 mlog(0, "node %u trying to join, but recovery "
869 "is ongoing.\n", bit);
870 packet.code = JOIN_DISALLOW;
871 } else if (test_bit(bit, dlm->recovery_map)) {
872 mlog(0, "node %u trying to join, but it "
873 "still needs recovery.\n", bit);
874 packet.code = JOIN_DISALLOW;
875 } else if (test_bit(bit, dlm->domain_map)) {
876 mlog(0, "node %u trying to join, but it "
877 "is still in the domain! needs recovery?\n",
879 packet.code = JOIN_DISALLOW;
881 /* Alright we're fully a part of this domain
882 * so we keep some state as to who's joining
883 * and indicate to him that needs to be fixed
886 /* Make sure we speak compatible locking protocols. */
887 if (dlm_query_join_proto_check("DLM", bit,
888 &dlm->dlm_locking_proto,
889 &query->dlm_proto)) {
890 packet.code = JOIN_PROTOCOL_MISMATCH;
891 } else if (dlm_query_join_proto_check("fs", bit,
892 &dlm->fs_locking_proto,
894 packet.code = JOIN_PROTOCOL_MISMATCH;
896 packet.dlm_minor = query->dlm_proto.pv_minor;
897 packet.fs_minor = query->fs_proto.pv_minor;
898 packet.code = JOIN_OK;
899 __dlm_set_joining_node(dlm, query->node_idx);
903 spin_unlock(&dlm->spinlock);
906 spin_unlock(&dlm_domain_lock);
909 mlog(0, "We respond with %u\n", packet.code);
911 dlm_query_join_packet_to_wire(&packet, &response);
915 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
918 struct dlm_assert_joined *assert;
919 struct dlm_ctxt *dlm = NULL;
921 assert = (struct dlm_assert_joined *) msg->buf;
923 mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
926 spin_lock(&dlm_domain_lock);
927 dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
928 /* XXX should we consider no dlm ctxt an error? */
930 spin_lock(&dlm->spinlock);
932 /* Alright, this node has officially joined our
933 * domain. Set him in the map and clean up our
934 * leftover join state. */
935 BUG_ON(dlm->joining_node != assert->node_idx);
937 if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
938 mlog(0, "dlm recovery is ongoing, disallow join\n");
939 spin_unlock(&dlm->spinlock);
940 spin_unlock(&dlm_domain_lock);
944 set_bit(assert->node_idx, dlm->domain_map);
945 clear_bit(assert->node_idx, dlm->exit_domain_map);
946 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
948 printk(KERN_NOTICE "o2dlm: Node %u joins domain %s ",
949 assert->node_idx, dlm->name);
950 __dlm_print_nodes(dlm);
952 /* notify anything attached to the heartbeat events */
953 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
955 spin_unlock(&dlm->spinlock);
957 spin_unlock(&dlm_domain_lock);
962 static int dlm_match_regions(struct dlm_ctxt *dlm,
963 struct dlm_query_region *qr,
964 char *local, int locallen)
966 char *remote = qr->qr_regions;
968 int localnr, i, j, foundit;
971 if (!o2hb_global_heartbeat_active()) {
972 if (qr->qr_numregions) {
973 mlog(ML_ERROR, "Domain %s: Joining node %d has global "
974 "heartbeat enabled but local node %d does not\n",
975 qr->qr_domain, qr->qr_node, dlm->node_num);
981 if (o2hb_global_heartbeat_active() && !qr->qr_numregions) {
982 mlog(ML_ERROR, "Domain %s: Local node %d has global "
983 "heartbeat enabled but joining node %d does not\n",
984 qr->qr_domain, dlm->node_num, qr->qr_node);
990 for (i = 0; i < qr->qr_numregions; ++i) {
991 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, r);
992 r += O2HB_MAX_REGION_NAME_LEN;
995 localnr = min(O2NM_MAX_REGIONS, locallen/O2HB_MAX_REGION_NAME_LEN);
996 localnr = o2hb_get_all_regions(local, (u8)localnr);
998 /* compare local regions with remote */
1000 for (i = 0; i < localnr; ++i) {
1003 for (j = 0; j <= qr->qr_numregions; ++j) {
1004 if (!memcmp(l, r, O2HB_MAX_REGION_NAME_LEN)) {
1008 r += O2HB_MAX_REGION_NAME_LEN;
1012 mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1013 "in local node %d but not in joining node %d\n",
1014 qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, l,
1015 dlm->node_num, qr->qr_node);
1018 l += O2HB_MAX_REGION_NAME_LEN;
1021 /* compare remote with local regions */
1023 for (i = 0; i < qr->qr_numregions; ++i) {
1026 for (j = 0; j < localnr; ++j) {
1027 if (!memcmp(r, l, O2HB_MAX_REGION_NAME_LEN)) {
1031 l += O2HB_MAX_REGION_NAME_LEN;
1035 mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1036 "in joining node %d but not in local node %d\n",
1037 qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, r,
1038 qr->qr_node, dlm->node_num);
1041 r += O2HB_MAX_REGION_NAME_LEN;
1048 static int dlm_send_regions(struct dlm_ctxt *dlm, unsigned long *node_map)
1050 struct dlm_query_region *qr = NULL;
1051 int status, ret = 0, i;
1054 if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1057 qr = kzalloc(sizeof(struct dlm_query_region), GFP_KERNEL);
1064 qr->qr_node = dlm->node_num;
1065 qr->qr_namelen = strlen(dlm->name);
1066 memcpy(qr->qr_domain, dlm->name, qr->qr_namelen);
1067 /* if local hb, the numregions will be zero */
1068 if (o2hb_global_heartbeat_active())
1069 qr->qr_numregions = o2hb_get_all_regions(qr->qr_regions,
1073 for (i = 0; i < qr->qr_numregions; ++i, p += O2HB_MAX_REGION_NAME_LEN)
1074 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, p);
1077 while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1078 i + 1)) < O2NM_MAX_NODES) {
1079 if (i == dlm->node_num)
1082 mlog(0, "Sending regions to node %d\n", i);
1084 ret = o2net_send_message(DLM_QUERY_REGION, DLM_MOD_KEY, qr,
1085 sizeof(struct dlm_query_region),
1090 mlog(ML_ERROR, "Region mismatch %d, node %d\n",
1101 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
1102 void *data, void **ret_data)
1104 struct dlm_query_region *qr;
1105 struct dlm_ctxt *dlm = NULL;
1109 qr = (struct dlm_query_region *) msg->buf;
1111 mlog(0, "Node %u queries hb regions on domain %s\n", qr->qr_node,
1114 /* buffer used in dlm_mast_regions() */
1115 local = kmalloc(sizeof(qr->qr_regions), GFP_KERNEL);
1121 spin_lock(&dlm_domain_lock);
1122 dlm = __dlm_lookup_domain_full(qr->qr_domain, qr->qr_namelen);
1124 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1125 "before join domain\n", qr->qr_node, qr->qr_domain);
1126 goto out_domain_lock;
1129 spin_lock(&dlm->spinlock);
1130 if (dlm->joining_node != qr->qr_node) {
1131 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1132 "but joining node is %d\n", qr->qr_node, qr->qr_domain,
1137 /* Support for global heartbeat was added in 1.1 */
1138 if (dlm->dlm_locking_proto.pv_major == 1 &&
1139 dlm->dlm_locking_proto.pv_minor == 0) {
1140 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1141 "but active dlm protocol is %d.%d\n", qr->qr_node,
1142 qr->qr_domain, dlm->dlm_locking_proto.pv_major,
1143 dlm->dlm_locking_proto.pv_minor);
1147 status = dlm_match_regions(dlm, qr, local, sizeof(qr->qr_regions));
1150 spin_unlock(&dlm->spinlock);
1153 spin_unlock(&dlm_domain_lock);
1160 static int dlm_match_nodes(struct dlm_ctxt *dlm, struct dlm_query_nodeinfo *qn)
1162 struct o2nm_node *local;
1163 struct dlm_node_info *remote;
1167 for (j = 0; j < qn->qn_numnodes; ++j)
1168 mlog(0, "Node %3d, %pI4:%u\n", qn->qn_nodes[j].ni_nodenum,
1169 &(qn->qn_nodes[j].ni_ipv4_address),
1170 ntohs(qn->qn_nodes[j].ni_ipv4_port));
1172 for (i = 0; i < O2NM_MAX_NODES && !status; ++i) {
1173 local = o2nm_get_node_by_num(i);
1175 for (j = 0; j < qn->qn_numnodes; ++j) {
1176 if (qn->qn_nodes[j].ni_nodenum == i) {
1177 remote = &(qn->qn_nodes[j]);
1182 if (!local && !remote)
1185 if ((local && !remote) || (!local && remote))
1189 ((remote->ni_nodenum != local->nd_num) ||
1190 (remote->ni_ipv4_port != local->nd_ipv4_port) ||
1191 (remote->ni_ipv4_address != local->nd_ipv4_address)))
1195 if (remote && !local)
1196 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1197 "registered in joining node %d but not in "
1198 "local node %d\n", qn->qn_domain,
1200 &(remote->ni_ipv4_address),
1201 ntohs(remote->ni_ipv4_port),
1202 qn->qn_nodenum, dlm->node_num);
1203 if (local && !remote)
1204 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1205 "registered in local node %d but not in "
1206 "joining node %d\n", qn->qn_domain,
1207 local->nd_num, &(local->nd_ipv4_address),
1208 ntohs(local->nd_ipv4_port),
1209 dlm->node_num, qn->qn_nodenum);
1210 BUG_ON((!local && !remote));
1214 o2nm_node_put(local);
1220 static int dlm_send_nodeinfo(struct dlm_ctxt *dlm, unsigned long *node_map)
1222 struct dlm_query_nodeinfo *qn = NULL;
1223 struct o2nm_node *node;
1224 int ret = 0, status, count, i;
1226 if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1229 qn = kzalloc(sizeof(struct dlm_query_nodeinfo), GFP_KERNEL);
1236 for (i = 0, count = 0; i < O2NM_MAX_NODES; ++i) {
1237 node = o2nm_get_node_by_num(i);
1240 qn->qn_nodes[count].ni_nodenum = node->nd_num;
1241 qn->qn_nodes[count].ni_ipv4_port = node->nd_ipv4_port;
1242 qn->qn_nodes[count].ni_ipv4_address = node->nd_ipv4_address;
1243 mlog(0, "Node %3d, %pI4:%u\n", node->nd_num,
1244 &(node->nd_ipv4_address), ntohs(node->nd_ipv4_port));
1246 o2nm_node_put(node);
1249 qn->qn_nodenum = dlm->node_num;
1250 qn->qn_numnodes = count;
1251 qn->qn_namelen = strlen(dlm->name);
1252 memcpy(qn->qn_domain, dlm->name, qn->qn_namelen);
1255 while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1256 i + 1)) < O2NM_MAX_NODES) {
1257 if (i == dlm->node_num)
1260 mlog(0, "Sending nodeinfo to node %d\n", i);
1262 ret = o2net_send_message(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
1263 qn, sizeof(struct dlm_query_nodeinfo),
1268 mlog(ML_ERROR, "node mismatch %d, node %d\n", ret, i);
1278 static int dlm_query_nodeinfo_handler(struct o2net_msg *msg, u32 len,
1279 void *data, void **ret_data)
1281 struct dlm_query_nodeinfo *qn;
1282 struct dlm_ctxt *dlm = NULL;
1283 int locked = 0, status = -EINVAL;
1285 qn = (struct dlm_query_nodeinfo *) msg->buf;
1287 mlog(0, "Node %u queries nodes on domain %s\n", qn->qn_nodenum,
1290 spin_lock(&dlm_domain_lock);
1291 dlm = __dlm_lookup_domain_full(qn->qn_domain, qn->qn_namelen);
1293 mlog(ML_ERROR, "Node %d queried nodes on domain %s before "
1294 "join domain\n", qn->qn_nodenum, qn->qn_domain);
1298 spin_lock(&dlm->spinlock);
1300 if (dlm->joining_node != qn->qn_nodenum) {
1301 mlog(ML_ERROR, "Node %d queried nodes on domain %s but "
1302 "joining node is %d\n", qn->qn_nodenum, qn->qn_domain,
1307 /* Support for node query was added in 1.1 */
1308 if (dlm->dlm_locking_proto.pv_major == 1 &&
1309 dlm->dlm_locking_proto.pv_minor == 0) {
1310 mlog(ML_ERROR, "Node %d queried nodes on domain %s "
1311 "but active dlm protocol is %d.%d\n", qn->qn_nodenum,
1312 qn->qn_domain, dlm->dlm_locking_proto.pv_major,
1313 dlm->dlm_locking_proto.pv_minor);
1317 status = dlm_match_nodes(dlm, qn);
1321 spin_unlock(&dlm->spinlock);
1322 spin_unlock(&dlm_domain_lock);
1327 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
1330 struct dlm_cancel_join *cancel;
1331 struct dlm_ctxt *dlm = NULL;
1333 cancel = (struct dlm_cancel_join *) msg->buf;
1335 mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
1338 spin_lock(&dlm_domain_lock);
1339 dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
1342 spin_lock(&dlm->spinlock);
1344 /* Yikes, this guy wants to cancel his join. No
1345 * problem, we simply cleanup our join state. */
1346 BUG_ON(dlm->joining_node != cancel->node_idx);
1347 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1349 spin_unlock(&dlm->spinlock);
1351 spin_unlock(&dlm_domain_lock);
1356 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
1360 struct dlm_cancel_join cancel_msg;
1362 memset(&cancel_msg, 0, sizeof(cancel_msg));
1363 cancel_msg.node_idx = dlm->node_num;
1364 cancel_msg.name_len = strlen(dlm->name);
1365 memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
1367 status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1368 &cancel_msg, sizeof(cancel_msg), node,
1371 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1372 "node %u\n", status, DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1381 /* map_size should be in bytes. */
1382 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
1383 unsigned long *node_map,
1384 unsigned int map_size)
1386 int status, tmpstat;
1389 if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
1390 sizeof(unsigned long))) {
1392 "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
1393 map_size, (unsigned)BITS_TO_LONGS(O2NM_MAX_NODES));
1399 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1400 node + 1)) < O2NM_MAX_NODES) {
1401 if (node == dlm->node_num)
1404 tmpstat = dlm_send_one_join_cancel(dlm, node);
1406 mlog(ML_ERROR, "Error return %d cancelling join on "
1407 "node %d\n", tmpstat, node);
1418 static int dlm_request_join(struct dlm_ctxt *dlm,
1420 enum dlm_query_join_response_code *response)
1423 struct dlm_query_join_request join_msg;
1424 struct dlm_query_join_packet packet;
1427 mlog(0, "querying node %d\n", node);
1429 memset(&join_msg, 0, sizeof(join_msg));
1430 join_msg.node_idx = dlm->node_num;
1431 join_msg.name_len = strlen(dlm->name);
1432 memcpy(join_msg.domain, dlm->name, join_msg.name_len);
1433 join_msg.dlm_proto = dlm->dlm_locking_proto;
1434 join_msg.fs_proto = dlm->fs_locking_proto;
1436 /* copy live node map to join message */
1437 byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES);
1439 status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
1440 sizeof(join_msg), node, &join_resp);
1441 if (status < 0 && status != -ENOPROTOOPT) {
1442 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1443 "node %u\n", status, DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1447 dlm_query_join_wire_to_packet(join_resp, &packet);
1449 /* -ENOPROTOOPT from the net code means the other side isn't
1450 listening for our message type -- that's fine, it means
1451 his dlm isn't up, so we can consider him a 'yes' but not
1452 joined into the domain. */
1453 if (status == -ENOPROTOOPT) {
1455 *response = JOIN_OK_NO_MAP;
1457 *response = packet.code;
1458 switch (packet.code) {
1460 case JOIN_OK_NO_MAP:
1462 case JOIN_PROTOCOL_MISMATCH:
1464 "This node requested DLM locking protocol %u.%u and "
1465 "filesystem locking protocol %u.%u. At least one of "
1466 "the protocol versions on node %d is not compatible, "
1468 dlm->dlm_locking_proto.pv_major,
1469 dlm->dlm_locking_proto.pv_minor,
1470 dlm->fs_locking_proto.pv_major,
1471 dlm->fs_locking_proto.pv_minor,
1476 /* Use the same locking protocol as the remote node */
1477 dlm->dlm_locking_proto.pv_minor = packet.dlm_minor;
1478 dlm->fs_locking_proto.pv_minor = packet.fs_minor;
1480 "Node %d responds JOIN_OK with DLM locking protocol "
1481 "%u.%u and fs locking protocol %u.%u\n",
1483 dlm->dlm_locking_proto.pv_major,
1484 dlm->dlm_locking_proto.pv_minor,
1485 dlm->fs_locking_proto.pv_major,
1486 dlm->fs_locking_proto.pv_minor);
1490 mlog(ML_ERROR, "invalid response %d from node %u\n",
1492 /* Reset response to JOIN_DISALLOW */
1493 *response = JOIN_DISALLOW;
1498 mlog(0, "status %d, node %d response is %d\n", status, node,
1505 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
1510 struct dlm_assert_joined assert_msg;
1512 mlog(0, "Sending join assert to node %u\n", node);
1514 memset(&assert_msg, 0, sizeof(assert_msg));
1515 assert_msg.node_idx = dlm->node_num;
1516 assert_msg.name_len = strlen(dlm->name);
1517 memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
1519 status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1520 &assert_msg, sizeof(assert_msg), node,
1523 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1524 "node %u\n", status, DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1532 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
1533 unsigned long *node_map)
1535 int status, node, live;
1539 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1540 node + 1)) < O2NM_MAX_NODES) {
1541 if (node == dlm->node_num)
1545 /* It is very important that this message be
1546 * received so we spin until either the node
1547 * has died or it gets the message. */
1548 status = dlm_send_one_join_assert(dlm, node);
1550 spin_lock(&dlm->spinlock);
1551 live = test_bit(node, dlm->live_nodes_map);
1552 spin_unlock(&dlm->spinlock);
1555 mlog(ML_ERROR, "Error return %d asserting "
1556 "join on node %d\n", status, node);
1558 /* give us some time between errors... */
1560 msleep(DLM_DOMAIN_BACKOFF_MS);
1562 } while (status && live);
1566 struct domain_join_ctxt {
1567 unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1568 unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1571 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
1572 struct domain_join_ctxt *ctxt,
1573 enum dlm_query_join_response_code response)
1577 if (response == JOIN_DISALLOW) {
1578 mlog(0, "Latest response of disallow -- should restart\n");
1582 spin_lock(&dlm->spinlock);
1583 /* For now, we restart the process if the node maps have
1585 ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
1586 sizeof(dlm->live_nodes_map));
1587 spin_unlock(&dlm->spinlock);
1590 mlog(0, "Node maps changed -- should restart\n");
1595 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
1597 int status = 0, tmpstat, node;
1598 struct domain_join_ctxt *ctxt;
1599 enum dlm_query_join_response_code response = JOIN_DISALLOW;
1603 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
1610 /* group sem locking should work for us here -- we're already
1611 * registered for heartbeat events so filling this should be
1612 * atomic wrt getting those handlers called. */
1613 o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
1615 spin_lock(&dlm->spinlock);
1616 memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
1618 __dlm_set_joining_node(dlm, dlm->node_num);
1620 spin_unlock(&dlm->spinlock);
1623 while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
1624 node + 1)) < O2NM_MAX_NODES) {
1625 if (node == dlm->node_num)
1628 status = dlm_request_join(dlm, node, &response);
1634 /* Ok, either we got a response or the node doesn't have a
1636 if (response == JOIN_OK)
1637 set_bit(node, ctxt->yes_resp_map);
1639 if (dlm_should_restart_join(dlm, ctxt, response)) {
1645 mlog(0, "Yay, done querying nodes!\n");
1647 /* Yay, everyone agree's we can join the domain. My domain is
1648 * comprised of all nodes who were put in the
1649 * yes_resp_map. Copy that into our domain map and send a join
1650 * assert message to clean up everyone elses state. */
1651 spin_lock(&dlm->spinlock);
1652 memcpy(dlm->domain_map, ctxt->yes_resp_map,
1653 sizeof(ctxt->yes_resp_map));
1654 set_bit(dlm->node_num, dlm->domain_map);
1655 spin_unlock(&dlm->spinlock);
1657 /* Support for global heartbeat and node info was added in 1.1 */
1658 if (dlm->dlm_locking_proto.pv_major > 1 ||
1659 dlm->dlm_locking_proto.pv_minor > 0) {
1660 status = dlm_send_nodeinfo(dlm, ctxt->yes_resp_map);
1665 status = dlm_send_regions(dlm, ctxt->yes_resp_map);
1672 dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1674 /* Joined state *must* be set before the joining node
1675 * information, otherwise the query_join handler may read no
1676 * current joiner but a state of NEW and tell joining nodes
1677 * we're not in the domain. */
1678 spin_lock(&dlm_domain_lock);
1679 dlm->dlm_state = DLM_CTXT_JOINED;
1681 spin_unlock(&dlm_domain_lock);
1684 spin_lock(&dlm->spinlock);
1685 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1687 printk(KERN_NOTICE "o2dlm: Joining domain %s ", dlm->name);
1688 __dlm_print_nodes(dlm);
1690 spin_unlock(&dlm->spinlock);
1693 /* Do we need to send a cancel message to any nodes? */
1695 tmpstat = dlm_send_join_cancels(dlm,
1697 sizeof(ctxt->yes_resp_map));
1699 mlog_errno(tmpstat);
1704 mlog(0, "returning %d\n", status);
1708 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1710 o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_up);
1711 o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_down);
1712 o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1715 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1719 mlog(0, "registering handlers.\n");
1721 o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1722 dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1723 o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1724 dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1726 status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_down);
1730 status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_up);
1734 status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1735 sizeof(struct dlm_master_request),
1736 dlm_master_request_handler,
1737 dlm, NULL, &dlm->dlm_domain_handlers);
1741 status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1742 sizeof(struct dlm_assert_master),
1743 dlm_assert_master_handler,
1744 dlm, dlm_assert_master_post_handler,
1745 &dlm->dlm_domain_handlers);
1749 status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1750 sizeof(struct dlm_create_lock),
1751 dlm_create_lock_handler,
1752 dlm, NULL, &dlm->dlm_domain_handlers);
1756 status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1757 DLM_CONVERT_LOCK_MAX_LEN,
1758 dlm_convert_lock_handler,
1759 dlm, NULL, &dlm->dlm_domain_handlers);
1763 status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1764 DLM_UNLOCK_LOCK_MAX_LEN,
1765 dlm_unlock_lock_handler,
1766 dlm, NULL, &dlm->dlm_domain_handlers);
1770 status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1771 DLM_PROXY_AST_MAX_LEN,
1772 dlm_proxy_ast_handler,
1773 dlm, NULL, &dlm->dlm_domain_handlers);
1777 status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1778 sizeof(struct dlm_exit_domain),
1779 dlm_exit_domain_handler,
1780 dlm, NULL, &dlm->dlm_domain_handlers);
1784 status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1785 sizeof(struct dlm_deref_lockres),
1786 dlm_deref_lockres_handler,
1787 dlm, NULL, &dlm->dlm_domain_handlers);
1791 status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1792 sizeof(struct dlm_migrate_request),
1793 dlm_migrate_request_handler,
1794 dlm, NULL, &dlm->dlm_domain_handlers);
1798 status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1799 DLM_MIG_LOCKRES_MAX_LEN,
1800 dlm_mig_lockres_handler,
1801 dlm, NULL, &dlm->dlm_domain_handlers);
1805 status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1806 sizeof(struct dlm_master_requery),
1807 dlm_master_requery_handler,
1808 dlm, NULL, &dlm->dlm_domain_handlers);
1812 status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1813 sizeof(struct dlm_lock_request),
1814 dlm_request_all_locks_handler,
1815 dlm, NULL, &dlm->dlm_domain_handlers);
1819 status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1820 sizeof(struct dlm_reco_data_done),
1821 dlm_reco_data_done_handler,
1822 dlm, NULL, &dlm->dlm_domain_handlers);
1826 status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1827 sizeof(struct dlm_begin_reco),
1828 dlm_begin_reco_handler,
1829 dlm, NULL, &dlm->dlm_domain_handlers);
1833 status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1834 sizeof(struct dlm_finalize_reco),
1835 dlm_finalize_reco_handler,
1836 dlm, NULL, &dlm->dlm_domain_handlers);
1840 status = o2net_register_handler(DLM_BEGIN_EXIT_DOMAIN_MSG, dlm->key,
1841 sizeof(struct dlm_exit_domain),
1842 dlm_begin_exit_domain_handler,
1843 dlm, NULL, &dlm->dlm_domain_handlers);
1847 status = o2net_register_handler(DLM_DEREF_LOCKRES_DONE, dlm->key,
1848 sizeof(struct dlm_deref_lockres_done),
1849 dlm_deref_lockres_done_handler,
1850 dlm, NULL, &dlm->dlm_domain_handlers);
1853 dlm_unregister_domain_handlers(dlm);
1858 static int dlm_join_domain(struct dlm_ctxt *dlm)
1861 unsigned int backoff;
1862 unsigned int total_backoff = 0;
1863 char wq_name[O2NM_MAX_NAME_LEN];
1867 mlog(0, "Join domain %s\n", dlm->name);
1869 status = dlm_register_domain_handlers(dlm);
1875 status = dlm_launch_thread(dlm);
1881 status = dlm_launch_recovery_thread(dlm);
1887 status = dlm_debug_init(dlm);
1893 snprintf(wq_name, O2NM_MAX_NAME_LEN, "dlm_wq-%s", dlm->name);
1894 dlm->dlm_worker = alloc_workqueue(wq_name, WQ_MEM_RECLAIM, 0);
1895 if (!dlm->dlm_worker) {
1902 status = dlm_try_to_join_domain(dlm);
1904 /* If we're racing another node to the join, then we
1905 * need to back off temporarily and let them
1907 #define DLM_JOIN_TIMEOUT_MSECS 90000
1908 if (status == -EAGAIN) {
1909 if (signal_pending(current)) {
1910 status = -ERESTARTSYS;
1914 if (total_backoff > DLM_JOIN_TIMEOUT_MSECS) {
1915 status = -ERESTARTSYS;
1916 mlog(ML_NOTICE, "Timed out joining dlm domain "
1917 "%s after %u msecs\n", dlm->name,
1924 * <dale> No, after you!
1926 * <dale> But you first!
1929 backoff = (unsigned int)(jiffies & 0x3);
1930 backoff *= DLM_DOMAIN_BACKOFF_MS;
1931 total_backoff += backoff;
1932 mlog(0, "backoff %d\n", backoff);
1935 } while (status == -EAGAIN);
1944 wake_up(&dlm_domain_events);
1947 dlm_unregister_domain_handlers(dlm);
1948 dlm_debug_shutdown(dlm);
1949 dlm_complete_thread(dlm);
1950 dlm_complete_recovery_thread(dlm);
1951 dlm_destroy_dlm_worker(dlm);
1957 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1962 struct dlm_ctxt *dlm = NULL;
1964 dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1971 dlm->name = kstrdup(domain, GFP_KERNEL);
1972 if (dlm->name == NULL) {
1978 dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1979 if (!dlm->lockres_hash) {
1985 for (i = 0; i < DLM_HASH_BUCKETS; i++)
1986 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1988 dlm->master_hash = (struct hlist_head **)
1989 dlm_alloc_pagevec(DLM_HASH_PAGES);
1990 if (!dlm->master_hash) {
1996 for (i = 0; i < DLM_HASH_BUCKETS; i++)
1997 INIT_HLIST_HEAD(dlm_master_hash(dlm, i));
2000 dlm->node_num = o2nm_this_node();
2002 ret = dlm_create_debugfs_subroot(dlm);
2006 spin_lock_init(&dlm->spinlock);
2007 spin_lock_init(&dlm->master_lock);
2008 spin_lock_init(&dlm->ast_lock);
2009 spin_lock_init(&dlm->track_lock);
2010 INIT_LIST_HEAD(&dlm->list);
2011 INIT_LIST_HEAD(&dlm->dirty_list);
2012 INIT_LIST_HEAD(&dlm->reco.resources);
2013 INIT_LIST_HEAD(&dlm->reco.node_data);
2014 INIT_LIST_HEAD(&dlm->purge_list);
2015 INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
2016 INIT_LIST_HEAD(&dlm->tracking_list);
2017 dlm->reco.state = 0;
2019 INIT_LIST_HEAD(&dlm->pending_asts);
2020 INIT_LIST_HEAD(&dlm->pending_basts);
2022 mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
2023 dlm->recovery_map, &(dlm->recovery_map[0]));
2025 memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
2026 memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
2027 memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
2029 dlm->dlm_thread_task = NULL;
2030 dlm->dlm_reco_thread_task = NULL;
2031 dlm->dlm_worker = NULL;
2032 init_waitqueue_head(&dlm->dlm_thread_wq);
2033 init_waitqueue_head(&dlm->dlm_reco_thread_wq);
2034 init_waitqueue_head(&dlm->reco.event);
2035 init_waitqueue_head(&dlm->ast_wq);
2036 init_waitqueue_head(&dlm->migration_wq);
2037 INIT_LIST_HEAD(&dlm->mle_hb_events);
2039 dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
2040 init_waitqueue_head(&dlm->dlm_join_events);
2042 dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
2043 dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
2045 atomic_set(&dlm->res_tot_count, 0);
2046 atomic_set(&dlm->res_cur_count, 0);
2047 for (i = 0; i < DLM_MLE_NUM_TYPES; ++i) {
2048 atomic_set(&dlm->mle_tot_count[i], 0);
2049 atomic_set(&dlm->mle_cur_count[i], 0);
2052 spin_lock_init(&dlm->work_lock);
2053 INIT_LIST_HEAD(&dlm->work_list);
2054 INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
2056 kref_init(&dlm->dlm_refs);
2057 dlm->dlm_state = DLM_CTXT_NEW;
2059 INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
2061 mlog(0, "context init: refcount %u\n",
2062 kref_read(&dlm->dlm_refs));
2065 if (ret < 0 && dlm) {
2066 if (dlm->master_hash)
2067 dlm_free_pagevec((void **)dlm->master_hash,
2070 if (dlm->lockres_hash)
2071 dlm_free_pagevec((void **)dlm->lockres_hash,
2082 * Compare a requested locking protocol version against the current one.
2084 * If the major numbers are different, they are incompatible.
2085 * If the current minor is greater than the request, they are incompatible.
2086 * If the current minor is less than or equal to the request, they are
2087 * compatible, and the requester should run at the current minor version.
2089 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
2090 struct dlm_protocol_version *request)
2092 if (existing->pv_major != request->pv_major)
2095 if (existing->pv_minor > request->pv_minor)
2098 if (existing->pv_minor < request->pv_minor)
2099 request->pv_minor = existing->pv_minor;
2105 * dlm_register_domain: one-time setup per "domain".
2107 * The filesystem passes in the requested locking version via proto.
2108 * If registration was successful, proto will contain the negotiated
2111 struct dlm_ctxt * dlm_register_domain(const char *domain,
2113 struct dlm_protocol_version *fs_proto)
2116 struct dlm_ctxt *dlm = NULL;
2117 struct dlm_ctxt *new_ctxt = NULL;
2119 if (strlen(domain) >= O2NM_MAX_NAME_LEN) {
2120 ret = -ENAMETOOLONG;
2121 mlog(ML_ERROR, "domain name length too long\n");
2125 mlog(0, "register called for domain \"%s\"\n", domain);
2129 if (signal_pending(current)) {
2135 spin_lock(&dlm_domain_lock);
2137 dlm = __dlm_lookup_domain(domain);
2139 if (dlm->dlm_state != DLM_CTXT_JOINED) {
2140 spin_unlock(&dlm_domain_lock);
2142 mlog(0, "This ctxt is not joined yet!\n");
2143 wait_event_interruptible(dlm_domain_events,
2144 dlm_wait_on_domain_helper(
2149 if (dlm_protocol_compare(&dlm->fs_locking_proto, fs_proto)) {
2150 spin_unlock(&dlm_domain_lock);
2152 "Requested locking protocol version is not "
2153 "compatible with already registered domain "
2154 "\"%s\"\n", domain);
2162 spin_unlock(&dlm_domain_lock);
2170 spin_unlock(&dlm_domain_lock);
2172 new_ctxt = dlm_alloc_ctxt(domain, key);
2181 /* a little variable switch-a-roo here... */
2185 /* add the new domain */
2186 list_add_tail(&dlm->list, &dlm_domains);
2187 spin_unlock(&dlm_domain_lock);
2190 * Pass the locking protocol version into the join. If the join
2191 * succeeds, it will have the negotiated protocol set.
2193 dlm->dlm_locking_proto = dlm_protocol;
2194 dlm->fs_locking_proto = *fs_proto;
2196 ret = dlm_join_domain(dlm);
2203 /* Tell the caller what locking protocol we negotiated */
2204 *fs_proto = dlm->fs_locking_proto;
2209 dlm_free_ctxt_mem(new_ctxt);
2216 EXPORT_SYMBOL_GPL(dlm_register_domain);
2218 static LIST_HEAD(dlm_join_handlers);
2220 static void dlm_unregister_net_handlers(void)
2222 o2net_unregister_handler_list(&dlm_join_handlers);
2225 static int dlm_register_net_handlers(void)
2229 status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
2230 sizeof(struct dlm_query_join_request),
2231 dlm_query_join_handler,
2232 NULL, NULL, &dlm_join_handlers);
2236 status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
2237 sizeof(struct dlm_assert_joined),
2238 dlm_assert_joined_handler,
2239 NULL, NULL, &dlm_join_handlers);
2243 status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
2244 sizeof(struct dlm_cancel_join),
2245 dlm_cancel_join_handler,
2246 NULL, NULL, &dlm_join_handlers);
2250 status = o2net_register_handler(DLM_QUERY_REGION, DLM_MOD_KEY,
2251 sizeof(struct dlm_query_region),
2252 dlm_query_region_handler,
2253 NULL, NULL, &dlm_join_handlers);
2258 status = o2net_register_handler(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
2259 sizeof(struct dlm_query_nodeinfo),
2260 dlm_query_nodeinfo_handler,
2261 NULL, NULL, &dlm_join_handlers);
2264 dlm_unregister_net_handlers();
2269 /* Domain eviction callback handling.
2271 * The file system requires notification of node death *before* the
2272 * dlm completes it's recovery work, otherwise it may be able to
2273 * acquire locks on resources requiring recovery. Since the dlm can
2274 * evict a node from it's domain *before* heartbeat fires, a similar
2275 * mechanism is required. */
2277 /* Eviction is not expected to happen often, so a per-domain lock is
2278 * not necessary. Eviction callbacks are allowed to sleep for short
2279 * periods of time. */
2280 static DECLARE_RWSEM(dlm_callback_sem);
2282 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
2285 struct dlm_eviction_cb *cb;
2287 down_read(&dlm_callback_sem);
2288 list_for_each_entry(cb, &dlm->dlm_eviction_callbacks, ec_item) {
2289 cb->ec_func(node_num, cb->ec_data);
2291 up_read(&dlm_callback_sem);
2294 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
2295 dlm_eviction_func *f,
2298 INIT_LIST_HEAD(&cb->ec_item);
2302 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
2304 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
2305 struct dlm_eviction_cb *cb)
2307 down_write(&dlm_callback_sem);
2308 list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
2309 up_write(&dlm_callback_sem);
2311 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
2313 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
2315 down_write(&dlm_callback_sem);
2316 list_del_init(&cb->ec_item);
2317 up_write(&dlm_callback_sem);
2319 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
2321 static int __init dlm_init(void)
2325 status = dlm_init_mle_cache();
2327 mlog(ML_ERROR, "Could not create o2dlm_mle slabcache\n");
2331 status = dlm_init_master_caches();
2333 mlog(ML_ERROR, "Could not create o2dlm_lockres and "
2334 "o2dlm_lockname slabcaches\n");
2338 status = dlm_init_lock_cache();
2340 mlog(ML_ERROR, "Count not create o2dlm_lock slabcache\n");
2344 status = dlm_register_net_handlers();
2346 mlog(ML_ERROR, "Unable to register network handlers\n");
2350 status = dlm_create_debugfs_root();
2356 dlm_unregister_net_handlers();
2357 dlm_destroy_lock_cache();
2358 dlm_destroy_master_caches();
2359 dlm_destroy_mle_cache();
2363 static void __exit dlm_exit (void)
2365 dlm_destroy_debugfs_root();
2366 dlm_unregister_net_handlers();
2367 dlm_destroy_lock_cache();
2368 dlm_destroy_master_caches();
2369 dlm_destroy_mle_cache();
2372 MODULE_AUTHOR("Oracle");
2373 MODULE_LICENSE("GPL");
2374 MODULE_DESCRIPTION("OCFS2 Distributed Lock Management");
2376 module_init(dlm_init);
2377 module_exit(dlm_exit);