ocfs2/trivial: Remove trailing whitespaces
[linux-2.6.git] / fs / ocfs2 / dlm / dlmdomain.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * dlmdomain.c
5  *
6  * defines domain join / leave apis
7  *
8  * Copyright (C) 2004 Oracle.  All rights reserved.
9  *
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.
14  *
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.
19  *
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.
24  *
25  */
26
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
37 #include "cluster/heartbeat.h"
38 #include "cluster/nodemanager.h"
39 #include "cluster/tcp.h"
40
41 #include "dlmapi.h"
42 #include "dlmcommon.h"
43 #include "dlmdomain.h"
44 #include "dlmdebug.h"
45
46 #include "dlmver.h"
47
48 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
49 #include "cluster/masklog.h"
50
51 /*
52  * ocfs2 node maps are array of long int, which limits to send them freely
53  * across the wire due to endianness issues. To workaround this, we convert
54  * long ints to byte arrays. Following 3 routines are helper functions to
55  * set/test/copy bits within those array of bytes
56  */
57 static inline void byte_set_bit(u8 nr, u8 map[])
58 {
59         map[nr >> 3] |= (1UL << (nr & 7));
60 }
61
62 static inline int byte_test_bit(u8 nr, u8 map[])
63 {
64         return ((1UL << (nr & 7)) & (map[nr >> 3])) != 0;
65 }
66
67 static inline void byte_copymap(u8 dmap[], unsigned long smap[],
68                         unsigned int sz)
69 {
70         unsigned int nn;
71
72         if (!sz)
73                 return;
74
75         memset(dmap, 0, ((sz + 7) >> 3));
76         for (nn = 0 ; nn < sz; nn++)
77                 if (test_bit(nn, smap))
78                         byte_set_bit(nn, dmap);
79 }
80
81 static void dlm_free_pagevec(void **vec, int pages)
82 {
83         while (pages--)
84                 free_page((unsigned long)vec[pages]);
85         kfree(vec);
86 }
87
88 static void **dlm_alloc_pagevec(int pages)
89 {
90         void **vec = kmalloc(pages * sizeof(void *), GFP_KERNEL);
91         int i;
92
93         if (!vec)
94                 return NULL;
95
96         for (i = 0; i < pages; i++)
97                 if (!(vec[i] = (void *)__get_free_page(GFP_KERNEL)))
98                         goto out_free;
99
100         mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n",
101              pages, (unsigned long)DLM_HASH_PAGES,
102              (unsigned long)DLM_BUCKETS_PER_PAGE);
103         return vec;
104 out_free:
105         dlm_free_pagevec(vec, i);
106         return NULL;
107 }
108
109 /*
110  *
111  * spinlock lock ordering: if multiple locks are needed, obey this ordering:
112  *    dlm_domain_lock
113  *    struct dlm_ctxt->spinlock
114  *    struct dlm_lock_resource->spinlock
115  *    struct dlm_ctxt->master_lock
116  *    struct dlm_ctxt->ast_lock
117  *    dlm_master_list_entry->spinlock
118  *    dlm_lock->spinlock
119  *
120  */
121
122 DEFINE_SPINLOCK(dlm_domain_lock);
123 LIST_HEAD(dlm_domains);
124 static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);
125
126 /*
127  * The supported protocol version for DLM communication.  Running domains
128  * will have a negotiated version with the same major number and a minor
129  * number equal or smaller.  The dlm_ctxt->dlm_locking_proto field should
130  * be used to determine what a running domain is actually using.
131  */
132 static const struct dlm_protocol_version dlm_protocol = {
133         .pv_major = 1,
134         .pv_minor = 0,
135 };
136
137 #define DLM_DOMAIN_BACKOFF_MS 200
138
139 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
140                                   void **ret_data);
141 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
142                                      void **ret_data);
143 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
144                                    void **ret_data);
145 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
146                                    void **ret_data);
147 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
148                                 struct dlm_protocol_version *request);
149
150 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
151
152 void __dlm_unhash_lockres(struct dlm_lock_resource *lockres)
153 {
154         if (!hlist_unhashed(&lockres->hash_node)) {
155                 hlist_del_init(&lockres->hash_node);
156                 dlm_lockres_put(lockres);
157         }
158 }
159
160 void __dlm_insert_lockres(struct dlm_ctxt *dlm,
161                        struct dlm_lock_resource *res)
162 {
163         struct hlist_head *bucket;
164         struct qstr *q;
165
166         assert_spin_locked(&dlm->spinlock);
167
168         q = &res->lockname;
169         bucket = dlm_lockres_hash(dlm, q->hash);
170
171         /* get a reference for our hashtable */
172         dlm_lockres_get(res);
173
174         hlist_add_head(&res->hash_node, bucket);
175 }
176
177 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
178                                                      const char *name,
179                                                      unsigned int len,
180                                                      unsigned int hash)
181 {
182         struct hlist_head *bucket;
183         struct hlist_node *list;
184
185         mlog_entry("%.*s\n", len, name);
186
187         assert_spin_locked(&dlm->spinlock);
188
189         bucket = dlm_lockres_hash(dlm, hash);
190
191         hlist_for_each(list, bucket) {
192                 struct dlm_lock_resource *res = hlist_entry(list,
193                         struct dlm_lock_resource, hash_node);
194                 if (res->lockname.name[0] != name[0])
195                         continue;
196                 if (unlikely(res->lockname.len != len))
197                         continue;
198                 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
199                         continue;
200                 dlm_lockres_get(res);
201                 return res;
202         }
203         return NULL;
204 }
205
206 /* intended to be called by functions which do not care about lock
207  * resources which are being purged (most net _handler functions).
208  * this will return NULL for any lock resource which is found but
209  * currently in the process of dropping its mastery reference.
210  * use __dlm_lookup_lockres_full when you need the lock resource
211  * regardless (e.g. dlm_get_lock_resource) */
212 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
213                                                 const char *name,
214                                                 unsigned int len,
215                                                 unsigned int hash)
216 {
217         struct dlm_lock_resource *res = NULL;
218
219         mlog_entry("%.*s\n", len, name);
220
221         assert_spin_locked(&dlm->spinlock);
222
223         res = __dlm_lookup_lockres_full(dlm, name, len, hash);
224         if (res) {
225                 spin_lock(&res->spinlock);
226                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
227                         spin_unlock(&res->spinlock);
228                         dlm_lockres_put(res);
229                         return NULL;
230                 }
231                 spin_unlock(&res->spinlock);
232         }
233
234         return res;
235 }
236
237 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
238                                     const char *name,
239                                     unsigned int len)
240 {
241         struct dlm_lock_resource *res;
242         unsigned int hash = dlm_lockid_hash(name, len);
243
244         spin_lock(&dlm->spinlock);
245         res = __dlm_lookup_lockres(dlm, name, len, hash);
246         spin_unlock(&dlm->spinlock);
247         return res;
248 }
249
250 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
251 {
252         struct dlm_ctxt *tmp = NULL;
253         struct list_head *iter;
254
255         assert_spin_locked(&dlm_domain_lock);
256
257         /* tmp->name here is always NULL terminated,
258          * but domain may not be! */
259         list_for_each(iter, &dlm_domains) {
260                 tmp = list_entry (iter, struct dlm_ctxt, list);
261                 if (strlen(tmp->name) == len &&
262                     memcmp(tmp->name, domain, len)==0)
263                         break;
264                 tmp = NULL;
265         }
266
267         return tmp;
268 }
269
270 /* For null terminated domain strings ONLY */
271 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
272 {
273         assert_spin_locked(&dlm_domain_lock);
274
275         return __dlm_lookup_domain_full(domain, strlen(domain));
276 }
277
278
279 /* returns true on one of two conditions:
280  * 1) the domain does not exist
281  * 2) the domain exists and it's state is "joined" */
282 static int dlm_wait_on_domain_helper(const char *domain)
283 {
284         int ret = 0;
285         struct dlm_ctxt *tmp = NULL;
286
287         spin_lock(&dlm_domain_lock);
288
289         tmp = __dlm_lookup_domain(domain);
290         if (!tmp)
291                 ret = 1;
292         else if (tmp->dlm_state == DLM_CTXT_JOINED)
293                 ret = 1;
294
295         spin_unlock(&dlm_domain_lock);
296         return ret;
297 }
298
299 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
300 {
301         dlm_destroy_debugfs_subroot(dlm);
302
303         if (dlm->lockres_hash)
304                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
305
306         if (dlm->master_hash)
307                 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
308
309         if (dlm->name)
310                 kfree(dlm->name);
311
312         kfree(dlm);
313 }
314
315 /* A little strange - this function will be called while holding
316  * dlm_domain_lock and is expected to be holding it on the way out. We
317  * will however drop and reacquire it multiple times */
318 static void dlm_ctxt_release(struct kref *kref)
319 {
320         struct dlm_ctxt *dlm;
321
322         dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
323
324         BUG_ON(dlm->num_joins);
325         BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
326
327         /* we may still be in the list if we hit an error during join. */
328         list_del_init(&dlm->list);
329
330         spin_unlock(&dlm_domain_lock);
331
332         mlog(0, "freeing memory from domain %s\n", dlm->name);
333
334         wake_up(&dlm_domain_events);
335
336         dlm_free_ctxt_mem(dlm);
337
338         spin_lock(&dlm_domain_lock);
339 }
340
341 void dlm_put(struct dlm_ctxt *dlm)
342 {
343         spin_lock(&dlm_domain_lock);
344         kref_put(&dlm->dlm_refs, dlm_ctxt_release);
345         spin_unlock(&dlm_domain_lock);
346 }
347
348 static void __dlm_get(struct dlm_ctxt *dlm)
349 {
350         kref_get(&dlm->dlm_refs);
351 }
352
353 /* given a questionable reference to a dlm object, gets a reference if
354  * it can find it in the list, otherwise returns NULL in which case
355  * you shouldn't trust your pointer. */
356 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
357 {
358         struct list_head *iter;
359         struct dlm_ctxt *target = NULL;
360
361         spin_lock(&dlm_domain_lock);
362
363         list_for_each(iter, &dlm_domains) {
364                 target = list_entry (iter, struct dlm_ctxt, list);
365
366                 if (target == dlm) {
367                         __dlm_get(target);
368                         break;
369                 }
370
371                 target = NULL;
372         }
373
374         spin_unlock(&dlm_domain_lock);
375
376         return target;
377 }
378
379 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
380 {
381         int ret;
382
383         spin_lock(&dlm_domain_lock);
384         ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
385                 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
386         spin_unlock(&dlm_domain_lock);
387
388         return ret;
389 }
390
391 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
392 {
393         if (dlm->dlm_worker) {
394                 flush_workqueue(dlm->dlm_worker);
395                 destroy_workqueue(dlm->dlm_worker);
396                 dlm->dlm_worker = NULL;
397         }
398 }
399
400 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
401 {
402         dlm_unregister_domain_handlers(dlm);
403         dlm_debug_shutdown(dlm);
404         dlm_complete_thread(dlm);
405         dlm_complete_recovery_thread(dlm);
406         dlm_destroy_dlm_worker(dlm);
407
408         /* We've left the domain. Now we can take ourselves out of the
409          * list and allow the kref stuff to help us free the
410          * memory. */
411         spin_lock(&dlm_domain_lock);
412         list_del_init(&dlm->list);
413         spin_unlock(&dlm_domain_lock);
414
415         /* Wake up anyone waiting for us to remove this domain */
416         wake_up(&dlm_domain_events);
417 }
418
419 static int dlm_migrate_all_locks(struct dlm_ctxt *dlm)
420 {
421         int i, num, n, ret = 0;
422         struct dlm_lock_resource *res;
423         struct hlist_node *iter;
424         struct hlist_head *bucket;
425         int dropped;
426
427         mlog(0, "Migrating locks from domain %s\n", dlm->name);
428
429         num = 0;
430         spin_lock(&dlm->spinlock);
431         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
432 redo_bucket:
433                 n = 0;
434                 bucket = dlm_lockres_hash(dlm, i);
435                 iter = bucket->first;
436                 while (iter) {
437                         n++;
438                         res = hlist_entry(iter, struct dlm_lock_resource,
439                                           hash_node);
440                         dlm_lockres_get(res);
441                         /* migrate, if necessary.  this will drop the dlm
442                          * spinlock and retake it if it does migration. */
443                         dropped = dlm_empty_lockres(dlm, res);
444
445                         spin_lock(&res->spinlock);
446                         __dlm_lockres_calc_usage(dlm, res);
447                         iter = res->hash_node.next;
448                         spin_unlock(&res->spinlock);
449
450                         dlm_lockres_put(res);
451
452                         if (dropped)
453                                 goto redo_bucket;
454                 }
455                 cond_resched_lock(&dlm->spinlock);
456                 num += n;
457                 mlog(0, "%s: touched %d lockreses in bucket %d "
458                      "(tot=%d)\n", dlm->name, n, i, num);
459         }
460         spin_unlock(&dlm->spinlock);
461         wake_up(&dlm->dlm_thread_wq);
462
463         /* let the dlm thread take care of purging, keep scanning until
464          * nothing remains in the hash */
465         if (num) {
466                 mlog(0, "%s: %d lock resources in hash last pass\n",
467                      dlm->name, num);
468                 ret = -EAGAIN;
469         }
470         mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
471         return ret;
472 }
473
474 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
475 {
476         int ret;
477
478         spin_lock(&dlm->spinlock);
479         ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
480         spin_unlock(&dlm->spinlock);
481
482         return ret;
483 }
484
485 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
486 {
487         /* Yikes, a double spinlock! I need domain_lock for the dlm
488          * state and the dlm spinlock for join state... Sorry! */
489 again:
490         spin_lock(&dlm_domain_lock);
491         spin_lock(&dlm->spinlock);
492
493         if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
494                 mlog(0, "Node %d is joining, we wait on it.\n",
495                           dlm->joining_node);
496                 spin_unlock(&dlm->spinlock);
497                 spin_unlock(&dlm_domain_lock);
498
499                 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
500                 goto again;
501         }
502
503         dlm->dlm_state = DLM_CTXT_LEAVING;
504         spin_unlock(&dlm->spinlock);
505         spin_unlock(&dlm_domain_lock);
506 }
507
508 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
509 {
510         int node = -1;
511
512         assert_spin_locked(&dlm->spinlock);
513
514         printk(KERN_INFO "ocfs2_dlm: Nodes in domain (\"%s\"): ", dlm->name);
515
516         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
517                                      node + 1)) < O2NM_MAX_NODES) {
518                 printk("%d ", node);
519         }
520         printk("\n");
521 }
522
523 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
524                                    void **ret_data)
525 {
526         struct dlm_ctxt *dlm = data;
527         unsigned int node;
528         struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
529
530         mlog_entry("%p %u %p", msg, len, data);
531
532         if (!dlm_grab(dlm))
533                 return 0;
534
535         node = exit_msg->node_idx;
536
537         printk(KERN_INFO "ocfs2_dlm: Node %u leaves domain %s\n", node, dlm->name);
538
539         spin_lock(&dlm->spinlock);
540         clear_bit(node, dlm->domain_map);
541         __dlm_print_nodes(dlm);
542
543         /* notify anything attached to the heartbeat events */
544         dlm_hb_event_notify_attached(dlm, node, 0);
545
546         spin_unlock(&dlm->spinlock);
547
548         dlm_put(dlm);
549
550         return 0;
551 }
552
553 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm,
554                                     unsigned int node)
555 {
556         int status;
557         struct dlm_exit_domain leave_msg;
558
559         mlog(0, "Asking node %u if we can leave the domain %s me = %u\n",
560                   node, dlm->name, dlm->node_num);
561
562         memset(&leave_msg, 0, sizeof(leave_msg));
563         leave_msg.node_idx = dlm->node_num;
564
565         status = o2net_send_message(DLM_EXIT_DOMAIN_MSG, dlm->key,
566                                     &leave_msg, sizeof(leave_msg), node,
567                                     NULL);
568
569         mlog(0, "status return %d from o2net_send_message\n", status);
570
571         return status;
572 }
573
574
575 static void dlm_leave_domain(struct dlm_ctxt *dlm)
576 {
577         int node, clear_node, status;
578
579         /* At this point we've migrated away all our locks and won't
580          * accept mastership of new ones. The dlm is responsible for
581          * almost nothing now. We make sure not to confuse any joining
582          * nodes and then commence shutdown procedure. */
583
584         spin_lock(&dlm->spinlock);
585         /* Clear ourselves from the domain map */
586         clear_bit(dlm->node_num, dlm->domain_map);
587         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
588                                      0)) < O2NM_MAX_NODES) {
589                 /* Drop the dlm spinlock. This is safe wrt the domain_map.
590                  * -nodes cannot be added now as the
591                  *   query_join_handlers knows to respond with OK_NO_MAP
592                  * -we catch the right network errors if a node is
593                  *   removed from the map while we're sending him the
594                  *   exit message. */
595                 spin_unlock(&dlm->spinlock);
596
597                 clear_node = 1;
598
599                 status = dlm_send_one_domain_exit(dlm, node);
600                 if (status < 0 &&
601                     status != -ENOPROTOOPT &&
602                     status != -ENOTCONN) {
603                         mlog(ML_NOTICE, "Error %d sending domain exit message "
604                              "to node %d\n", status, node);
605
606                         /* Not sure what to do here but lets sleep for
607                          * a bit in case this was a transient
608                          * error... */
609                         msleep(DLM_DOMAIN_BACKOFF_MS);
610                         clear_node = 0;
611                 }
612
613                 spin_lock(&dlm->spinlock);
614                 /* If we're not clearing the node bit then we intend
615                  * to loop back around to try again. */
616                 if (clear_node)
617                         clear_bit(node, dlm->domain_map);
618         }
619         spin_unlock(&dlm->spinlock);
620 }
621
622 int dlm_joined(struct dlm_ctxt *dlm)
623 {
624         int ret = 0;
625
626         spin_lock(&dlm_domain_lock);
627
628         if (dlm->dlm_state == DLM_CTXT_JOINED)
629                 ret = 1;
630
631         spin_unlock(&dlm_domain_lock);
632
633         return ret;
634 }
635
636 int dlm_shutting_down(struct dlm_ctxt *dlm)
637 {
638         int ret = 0;
639
640         spin_lock(&dlm_domain_lock);
641
642         if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
643                 ret = 1;
644
645         spin_unlock(&dlm_domain_lock);
646
647         return ret;
648 }
649
650 void dlm_unregister_domain(struct dlm_ctxt *dlm)
651 {
652         int leave = 0;
653         struct dlm_lock_resource *res;
654
655         spin_lock(&dlm_domain_lock);
656         BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
657         BUG_ON(!dlm->num_joins);
658
659         dlm->num_joins--;
660         if (!dlm->num_joins) {
661                 /* We mark it "in shutdown" now so new register
662                  * requests wait until we've completely left the
663                  * domain. Don't use DLM_CTXT_LEAVING yet as we still
664                  * want new domain joins to communicate with us at
665                  * least until we've completed migration of our
666                  * resources. */
667                 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
668                 leave = 1;
669         }
670         spin_unlock(&dlm_domain_lock);
671
672         if (leave) {
673                 mlog(0, "shutting down domain %s\n", dlm->name);
674
675                 /* We changed dlm state, notify the thread */
676                 dlm_kick_thread(dlm, NULL);
677
678                 while (dlm_migrate_all_locks(dlm)) {
679                         /* Give dlm_thread time to purge the lockres' */
680                         msleep(500);
681                         mlog(0, "%s: more migration to do\n", dlm->name);
682                 }
683
684                 /* This list should be empty. If not, print remaining lockres */
685                 if (!list_empty(&dlm->tracking_list)) {
686                         mlog(ML_ERROR, "Following lockres' are still on the "
687                              "tracking list:\n");
688                         list_for_each_entry(res, &dlm->tracking_list, tracking)
689                                 dlm_print_one_lock_resource(res);
690                 }
691
692                 dlm_mark_domain_leaving(dlm);
693                 dlm_leave_domain(dlm);
694                 dlm_complete_dlm_shutdown(dlm);
695         }
696         dlm_put(dlm);
697 }
698 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
699
700 static int dlm_query_join_proto_check(char *proto_type, int node,
701                                       struct dlm_protocol_version *ours,
702                                       struct dlm_protocol_version *request)
703 {
704         int rc;
705         struct dlm_protocol_version proto = *request;
706
707         if (!dlm_protocol_compare(ours, &proto)) {
708                 mlog(0,
709                      "node %u wanted to join with %s locking protocol "
710                      "%u.%u, we respond with %u.%u\n",
711                      node, proto_type,
712                      request->pv_major,
713                      request->pv_minor,
714                      proto.pv_major, proto.pv_minor);
715                 request->pv_minor = proto.pv_minor;
716                 rc = 0;
717         } else {
718                 mlog(ML_NOTICE,
719                      "Node %u wanted to join with %s locking "
720                      "protocol %u.%u, but we have %u.%u, disallowing\n",
721                      node, proto_type,
722                      request->pv_major,
723                      request->pv_minor,
724                      ours->pv_major,
725                      ours->pv_minor);
726                 rc = 1;
727         }
728
729         return rc;
730 }
731
732 /*
733  * struct dlm_query_join_packet is made up of four one-byte fields.  They
734  * are effectively in big-endian order already.  However, little-endian
735  * machines swap them before putting the packet on the wire (because
736  * query_join's response is a status, and that status is treated as a u32
737  * on the wire).  Thus, a big-endian and little-endian machines will treat
738  * this structure differently.
739  *
740  * The solution is to have little-endian machines swap the structure when
741  * converting from the structure to the u32 representation.  This will
742  * result in the structure having the correct format on the wire no matter
743  * the host endian format.
744  */
745 static void dlm_query_join_packet_to_wire(struct dlm_query_join_packet *packet,
746                                           u32 *wire)
747 {
748         union dlm_query_join_response response;
749
750         response.packet = *packet;
751         *wire = cpu_to_be32(response.intval);
752 }
753
754 static void dlm_query_join_wire_to_packet(u32 wire,
755                                           struct dlm_query_join_packet *packet)
756 {
757         union dlm_query_join_response response;
758
759         response.intval = cpu_to_be32(wire);
760         *packet = response.packet;
761 }
762
763 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
764                                   void **ret_data)
765 {
766         struct dlm_query_join_request *query;
767         struct dlm_query_join_packet packet = {
768                 .code = JOIN_DISALLOW,
769         };
770         struct dlm_ctxt *dlm = NULL;
771         u32 response;
772         u8 nodenum;
773
774         query = (struct dlm_query_join_request *) msg->buf;
775
776         mlog(0, "node %u wants to join domain %s\n", query->node_idx,
777                   query->domain);
778
779         /*
780          * If heartbeat doesn't consider the node live, tell it
781          * to back off and try again.  This gives heartbeat a chance
782          * to catch up.
783          */
784         if (!o2hb_check_node_heartbeating(query->node_idx)) {
785                 mlog(0, "node %u is not in our live map yet\n",
786                      query->node_idx);
787
788                 packet.code = JOIN_DISALLOW;
789                 goto respond;
790         }
791
792         packet.code = JOIN_OK_NO_MAP;
793
794         spin_lock(&dlm_domain_lock);
795         dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
796         if (!dlm)
797                 goto unlock_respond;
798
799         /*
800          * There is a small window where the joining node may not see the
801          * node(s) that just left but still part of the cluster. DISALLOW
802          * join request if joining node has different node map.
803          */
804         nodenum=0;
805         while (nodenum < O2NM_MAX_NODES) {
806                 if (test_bit(nodenum, dlm->domain_map)) {
807                         if (!byte_test_bit(nodenum, query->node_map)) {
808                                 mlog(0, "disallow join as node %u does not "
809                                      "have node %u in its nodemap\n",
810                                      query->node_idx, nodenum);
811                                 packet.code = JOIN_DISALLOW;
812                                 goto unlock_respond;
813                         }
814                 }
815                 nodenum++;
816         }
817
818         /* Once the dlm ctxt is marked as leaving then we don't want
819          * to be put in someone's domain map.
820          * Also, explicitly disallow joining at certain troublesome
821          * times (ie. during recovery). */
822         if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
823                 int bit = query->node_idx;
824                 spin_lock(&dlm->spinlock);
825
826                 if (dlm->dlm_state == DLM_CTXT_NEW &&
827                     dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
828                         /*If this is a brand new context and we
829                          * haven't started our join process yet, then
830                          * the other node won the race. */
831                         packet.code = JOIN_OK_NO_MAP;
832                 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
833                         /* Disallow parallel joins. */
834                         packet.code = JOIN_DISALLOW;
835                 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
836                         mlog(0, "node %u trying to join, but recovery "
837                              "is ongoing.\n", bit);
838                         packet.code = JOIN_DISALLOW;
839                 } else if (test_bit(bit, dlm->recovery_map)) {
840                         mlog(0, "node %u trying to join, but it "
841                              "still needs recovery.\n", bit);
842                         packet.code = JOIN_DISALLOW;
843                 } else if (test_bit(bit, dlm->domain_map)) {
844                         mlog(0, "node %u trying to join, but it "
845                              "is still in the domain! needs recovery?\n",
846                              bit);
847                         packet.code = JOIN_DISALLOW;
848                 } else {
849                         /* Alright we're fully a part of this domain
850                          * so we keep some state as to who's joining
851                          * and indicate to him that needs to be fixed
852                          * up. */
853
854                         /* Make sure we speak compatible locking protocols.  */
855                         if (dlm_query_join_proto_check("DLM", bit,
856                                                        &dlm->dlm_locking_proto,
857                                                        &query->dlm_proto)) {
858                                 packet.code = JOIN_PROTOCOL_MISMATCH;
859                         } else if (dlm_query_join_proto_check("fs", bit,
860                                                               &dlm->fs_locking_proto,
861                                                               &query->fs_proto)) {
862                                 packet.code = JOIN_PROTOCOL_MISMATCH;
863                         } else {
864                                 packet.dlm_minor = query->dlm_proto.pv_minor;
865                                 packet.fs_minor = query->fs_proto.pv_minor;
866                                 packet.code = JOIN_OK;
867                                 __dlm_set_joining_node(dlm, query->node_idx);
868                         }
869                 }
870
871                 spin_unlock(&dlm->spinlock);
872         }
873 unlock_respond:
874         spin_unlock(&dlm_domain_lock);
875
876 respond:
877         mlog(0, "We respond with %u\n", packet.code);
878
879         dlm_query_join_packet_to_wire(&packet, &response);
880         return response;
881 }
882
883 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
884                                      void **ret_data)
885 {
886         struct dlm_assert_joined *assert;
887         struct dlm_ctxt *dlm = NULL;
888
889         assert = (struct dlm_assert_joined *) msg->buf;
890
891         mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
892                   assert->domain);
893
894         spin_lock(&dlm_domain_lock);
895         dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
896         /* XXX should we consider no dlm ctxt an error? */
897         if (dlm) {
898                 spin_lock(&dlm->spinlock);
899
900                 /* Alright, this node has officially joined our
901                  * domain. Set him in the map and clean up our
902                  * leftover join state. */
903                 BUG_ON(dlm->joining_node != assert->node_idx);
904                 set_bit(assert->node_idx, dlm->domain_map);
905                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
906
907                 printk(KERN_INFO "ocfs2_dlm: Node %u joins domain %s\n",
908                        assert->node_idx, dlm->name);
909                 __dlm_print_nodes(dlm);
910
911                 /* notify anything attached to the heartbeat events */
912                 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
913
914                 spin_unlock(&dlm->spinlock);
915         }
916         spin_unlock(&dlm_domain_lock);
917
918         return 0;
919 }
920
921 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
922                                    void **ret_data)
923 {
924         struct dlm_cancel_join *cancel;
925         struct dlm_ctxt *dlm = NULL;
926
927         cancel = (struct dlm_cancel_join *) msg->buf;
928
929         mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
930                   cancel->domain);
931
932         spin_lock(&dlm_domain_lock);
933         dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
934
935         if (dlm) {
936                 spin_lock(&dlm->spinlock);
937
938                 /* Yikes, this guy wants to cancel his join. No
939                  * problem, we simply cleanup our join state. */
940                 BUG_ON(dlm->joining_node != cancel->node_idx);
941                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
942
943                 spin_unlock(&dlm->spinlock);
944         }
945         spin_unlock(&dlm_domain_lock);
946
947         return 0;
948 }
949
950 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
951                                     unsigned int node)
952 {
953         int status;
954         struct dlm_cancel_join cancel_msg;
955
956         memset(&cancel_msg, 0, sizeof(cancel_msg));
957         cancel_msg.node_idx = dlm->node_num;
958         cancel_msg.name_len = strlen(dlm->name);
959         memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
960
961         status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
962                                     &cancel_msg, sizeof(cancel_msg), node,
963                                     NULL);
964         if (status < 0) {
965                 mlog_errno(status);
966                 goto bail;
967         }
968
969 bail:
970         return status;
971 }
972
973 /* map_size should be in bytes. */
974 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
975                                  unsigned long *node_map,
976                                  unsigned int map_size)
977 {
978         int status, tmpstat;
979         unsigned int node;
980
981         if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
982                          sizeof(unsigned long))) {
983                 mlog(ML_ERROR,
984                      "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
985                      map_size, (unsigned)BITS_TO_LONGS(O2NM_MAX_NODES));
986                 return -EINVAL;
987         }
988
989         status = 0;
990         node = -1;
991         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
992                                      node + 1)) < O2NM_MAX_NODES) {
993                 if (node == dlm->node_num)
994                         continue;
995
996                 tmpstat = dlm_send_one_join_cancel(dlm, node);
997                 if (tmpstat) {
998                         mlog(ML_ERROR, "Error return %d cancelling join on "
999                              "node %d\n", tmpstat, node);
1000                         if (!status)
1001                                 status = tmpstat;
1002                 }
1003         }
1004
1005         if (status)
1006                 mlog_errno(status);
1007         return status;
1008 }
1009
1010 static int dlm_request_join(struct dlm_ctxt *dlm,
1011                             int node,
1012                             enum dlm_query_join_response_code *response)
1013 {
1014         int status;
1015         struct dlm_query_join_request join_msg;
1016         struct dlm_query_join_packet packet;
1017         u32 join_resp;
1018
1019         mlog(0, "querying node %d\n", node);
1020
1021         memset(&join_msg, 0, sizeof(join_msg));
1022         join_msg.node_idx = dlm->node_num;
1023         join_msg.name_len = strlen(dlm->name);
1024         memcpy(join_msg.domain, dlm->name, join_msg.name_len);
1025         join_msg.dlm_proto = dlm->dlm_locking_proto;
1026         join_msg.fs_proto = dlm->fs_locking_proto;
1027
1028         /* copy live node map to join message */
1029         byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES);
1030
1031         status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
1032                                     sizeof(join_msg), node,
1033                                     &join_resp);
1034         if (status < 0 && status != -ENOPROTOOPT) {
1035                 mlog_errno(status);
1036                 goto bail;
1037         }
1038         dlm_query_join_wire_to_packet(join_resp, &packet);
1039
1040         /* -ENOPROTOOPT from the net code means the other side isn't
1041             listening for our message type -- that's fine, it means
1042             his dlm isn't up, so we can consider him a 'yes' but not
1043             joined into the domain.  */
1044         if (status == -ENOPROTOOPT) {
1045                 status = 0;
1046                 *response = JOIN_OK_NO_MAP;
1047         } else if (packet.code == JOIN_DISALLOW ||
1048                    packet.code == JOIN_OK_NO_MAP) {
1049                 *response = packet.code;
1050         } else if (packet.code == JOIN_PROTOCOL_MISMATCH) {
1051                 mlog(ML_NOTICE,
1052                      "This node requested DLM locking protocol %u.%u and "
1053                      "filesystem locking protocol %u.%u.  At least one of "
1054                      "the protocol versions on node %d is not compatible, "
1055                      "disconnecting\n",
1056                      dlm->dlm_locking_proto.pv_major,
1057                      dlm->dlm_locking_proto.pv_minor,
1058                      dlm->fs_locking_proto.pv_major,
1059                      dlm->fs_locking_proto.pv_minor,
1060                      node);
1061                 status = -EPROTO;
1062                 *response = packet.code;
1063         } else if (packet.code == JOIN_OK) {
1064                 *response = packet.code;
1065                 /* Use the same locking protocol as the remote node */
1066                 dlm->dlm_locking_proto.pv_minor = packet.dlm_minor;
1067                 dlm->fs_locking_proto.pv_minor = packet.fs_minor;
1068                 mlog(0,
1069                      "Node %d responds JOIN_OK with DLM locking protocol "
1070                      "%u.%u and fs locking protocol %u.%u\n",
1071                      node,
1072                      dlm->dlm_locking_proto.pv_major,
1073                      dlm->dlm_locking_proto.pv_minor,
1074                      dlm->fs_locking_proto.pv_major,
1075                      dlm->fs_locking_proto.pv_minor);
1076         } else {
1077                 status = -EINVAL;
1078                 mlog(ML_ERROR, "invalid response %d from node %u\n",
1079                      packet.code, node);
1080         }
1081
1082         mlog(0, "status %d, node %d response is %d\n", status, node,
1083              *response);
1084
1085 bail:
1086         return status;
1087 }
1088
1089 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
1090                                     unsigned int node)
1091 {
1092         int status;
1093         struct dlm_assert_joined assert_msg;
1094
1095         mlog(0, "Sending join assert to node %u\n", node);
1096
1097         memset(&assert_msg, 0, sizeof(assert_msg));
1098         assert_msg.node_idx = dlm->node_num;
1099         assert_msg.name_len = strlen(dlm->name);
1100         memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
1101
1102         status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1103                                     &assert_msg, sizeof(assert_msg), node,
1104                                     NULL);
1105         if (status < 0)
1106                 mlog_errno(status);
1107
1108         return status;
1109 }
1110
1111 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
1112                                   unsigned long *node_map)
1113 {
1114         int status, node, live;
1115
1116         status = 0;
1117         node = -1;
1118         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1119                                      node + 1)) < O2NM_MAX_NODES) {
1120                 if (node == dlm->node_num)
1121                         continue;
1122
1123                 do {
1124                         /* It is very important that this message be
1125                          * received so we spin until either the node
1126                          * has died or it gets the message. */
1127                         status = dlm_send_one_join_assert(dlm, node);
1128
1129                         spin_lock(&dlm->spinlock);
1130                         live = test_bit(node, dlm->live_nodes_map);
1131                         spin_unlock(&dlm->spinlock);
1132
1133                         if (status) {
1134                                 mlog(ML_ERROR, "Error return %d asserting "
1135                                      "join on node %d\n", status, node);
1136
1137                                 /* give us some time between errors... */
1138                                 if (live)
1139                                         msleep(DLM_DOMAIN_BACKOFF_MS);
1140                         }
1141                 } while (status && live);
1142         }
1143 }
1144
1145 struct domain_join_ctxt {
1146         unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1147         unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1148 };
1149
1150 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
1151                                    struct domain_join_ctxt *ctxt,
1152                                    enum dlm_query_join_response_code response)
1153 {
1154         int ret;
1155
1156         if (response == JOIN_DISALLOW) {
1157                 mlog(0, "Latest response of disallow -- should restart\n");
1158                 return 1;
1159         }
1160
1161         spin_lock(&dlm->spinlock);
1162         /* For now, we restart the process if the node maps have
1163          * changed at all */
1164         ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
1165                      sizeof(dlm->live_nodes_map));
1166         spin_unlock(&dlm->spinlock);
1167
1168         if (ret)
1169                 mlog(0, "Node maps changed -- should restart\n");
1170
1171         return ret;
1172 }
1173
1174 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
1175 {
1176         int status = 0, tmpstat, node;
1177         struct domain_join_ctxt *ctxt;
1178         enum dlm_query_join_response_code response = JOIN_DISALLOW;
1179
1180         mlog_entry("%p", dlm);
1181
1182         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
1183         if (!ctxt) {
1184                 status = -ENOMEM;
1185                 mlog_errno(status);
1186                 goto bail;
1187         }
1188
1189         /* group sem locking should work for us here -- we're already
1190          * registered for heartbeat events so filling this should be
1191          * atomic wrt getting those handlers called. */
1192         o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
1193
1194         spin_lock(&dlm->spinlock);
1195         memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
1196
1197         __dlm_set_joining_node(dlm, dlm->node_num);
1198
1199         spin_unlock(&dlm->spinlock);
1200
1201         node = -1;
1202         while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
1203                                      node + 1)) < O2NM_MAX_NODES) {
1204                 if (node == dlm->node_num)
1205                         continue;
1206
1207                 status = dlm_request_join(dlm, node, &response);
1208                 if (status < 0) {
1209                         mlog_errno(status);
1210                         goto bail;
1211                 }
1212
1213                 /* Ok, either we got a response or the node doesn't have a
1214                  * dlm up. */
1215                 if (response == JOIN_OK)
1216                         set_bit(node, ctxt->yes_resp_map);
1217
1218                 if (dlm_should_restart_join(dlm, ctxt, response)) {
1219                         status = -EAGAIN;
1220                         goto bail;
1221                 }
1222         }
1223
1224         mlog(0, "Yay, done querying nodes!\n");
1225
1226         /* Yay, everyone agree's we can join the domain. My domain is
1227          * comprised of all nodes who were put in the
1228          * yes_resp_map. Copy that into our domain map and send a join
1229          * assert message to clean up everyone elses state. */
1230         spin_lock(&dlm->spinlock);
1231         memcpy(dlm->domain_map, ctxt->yes_resp_map,
1232                sizeof(ctxt->yes_resp_map));
1233         set_bit(dlm->node_num, dlm->domain_map);
1234         spin_unlock(&dlm->spinlock);
1235
1236         dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1237
1238         /* Joined state *must* be set before the joining node
1239          * information, otherwise the query_join handler may read no
1240          * current joiner but a state of NEW and tell joining nodes
1241          * we're not in the domain. */
1242         spin_lock(&dlm_domain_lock);
1243         dlm->dlm_state = DLM_CTXT_JOINED;
1244         dlm->num_joins++;
1245         spin_unlock(&dlm_domain_lock);
1246
1247 bail:
1248         spin_lock(&dlm->spinlock);
1249         __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1250         if (!status)
1251                 __dlm_print_nodes(dlm);
1252         spin_unlock(&dlm->spinlock);
1253
1254         if (ctxt) {
1255                 /* Do we need to send a cancel message to any nodes? */
1256                 if (status < 0) {
1257                         tmpstat = dlm_send_join_cancels(dlm,
1258                                                         ctxt->yes_resp_map,
1259                                                         sizeof(ctxt->yes_resp_map));
1260                         if (tmpstat < 0)
1261                                 mlog_errno(tmpstat);
1262                 }
1263                 kfree(ctxt);
1264         }
1265
1266         mlog(0, "returning %d\n", status);
1267         return status;
1268 }
1269
1270 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1271 {
1272         o2hb_unregister_callback(NULL, &dlm->dlm_hb_up);
1273         o2hb_unregister_callback(NULL, &dlm->dlm_hb_down);
1274         o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1275 }
1276
1277 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1278 {
1279         int status;
1280
1281         mlog(0, "registering handlers.\n");
1282
1283         o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1284                             dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1285         status = o2hb_register_callback(NULL, &dlm->dlm_hb_down);
1286         if (status)
1287                 goto bail;
1288
1289         o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1290                             dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1291         status = o2hb_register_callback(NULL, &dlm->dlm_hb_up);
1292         if (status)
1293                 goto bail;
1294
1295         status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1296                                         sizeof(struct dlm_master_request),
1297                                         dlm_master_request_handler,
1298                                         dlm, NULL, &dlm->dlm_domain_handlers);
1299         if (status)
1300                 goto bail;
1301
1302         status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1303                                         sizeof(struct dlm_assert_master),
1304                                         dlm_assert_master_handler,
1305                                         dlm, dlm_assert_master_post_handler,
1306                                         &dlm->dlm_domain_handlers);
1307         if (status)
1308                 goto bail;
1309
1310         status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1311                                         sizeof(struct dlm_create_lock),
1312                                         dlm_create_lock_handler,
1313                                         dlm, NULL, &dlm->dlm_domain_handlers);
1314         if (status)
1315                 goto bail;
1316
1317         status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1318                                         DLM_CONVERT_LOCK_MAX_LEN,
1319                                         dlm_convert_lock_handler,
1320                                         dlm, NULL, &dlm->dlm_domain_handlers);
1321         if (status)
1322                 goto bail;
1323
1324         status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1325                                         DLM_UNLOCK_LOCK_MAX_LEN,
1326                                         dlm_unlock_lock_handler,
1327                                         dlm, NULL, &dlm->dlm_domain_handlers);
1328         if (status)
1329                 goto bail;
1330
1331         status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1332                                         DLM_PROXY_AST_MAX_LEN,
1333                                         dlm_proxy_ast_handler,
1334                                         dlm, NULL, &dlm->dlm_domain_handlers);
1335         if (status)
1336                 goto bail;
1337
1338         status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1339                                         sizeof(struct dlm_exit_domain),
1340                                         dlm_exit_domain_handler,
1341                                         dlm, NULL, &dlm->dlm_domain_handlers);
1342         if (status)
1343                 goto bail;
1344
1345         status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1346                                         sizeof(struct dlm_deref_lockres),
1347                                         dlm_deref_lockres_handler,
1348                                         dlm, NULL, &dlm->dlm_domain_handlers);
1349         if (status)
1350                 goto bail;
1351
1352         status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1353                                         sizeof(struct dlm_migrate_request),
1354                                         dlm_migrate_request_handler,
1355                                         dlm, NULL, &dlm->dlm_domain_handlers);
1356         if (status)
1357                 goto bail;
1358
1359         status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1360                                         DLM_MIG_LOCKRES_MAX_LEN,
1361                                         dlm_mig_lockres_handler,
1362                                         dlm, NULL, &dlm->dlm_domain_handlers);
1363         if (status)
1364                 goto bail;
1365
1366         status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1367                                         sizeof(struct dlm_master_requery),
1368                                         dlm_master_requery_handler,
1369                                         dlm, NULL, &dlm->dlm_domain_handlers);
1370         if (status)
1371                 goto bail;
1372
1373         status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1374                                         sizeof(struct dlm_lock_request),
1375                                         dlm_request_all_locks_handler,
1376                                         dlm, NULL, &dlm->dlm_domain_handlers);
1377         if (status)
1378                 goto bail;
1379
1380         status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1381                                         sizeof(struct dlm_reco_data_done),
1382                                         dlm_reco_data_done_handler,
1383                                         dlm, NULL, &dlm->dlm_domain_handlers);
1384         if (status)
1385                 goto bail;
1386
1387         status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1388                                         sizeof(struct dlm_begin_reco),
1389                                         dlm_begin_reco_handler,
1390                                         dlm, NULL, &dlm->dlm_domain_handlers);
1391         if (status)
1392                 goto bail;
1393
1394         status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1395                                         sizeof(struct dlm_finalize_reco),
1396                                         dlm_finalize_reco_handler,
1397                                         dlm, NULL, &dlm->dlm_domain_handlers);
1398         if (status)
1399                 goto bail;
1400
1401 bail:
1402         if (status)
1403                 dlm_unregister_domain_handlers(dlm);
1404
1405         return status;
1406 }
1407
1408 static int dlm_join_domain(struct dlm_ctxt *dlm)
1409 {
1410         int status;
1411         unsigned int backoff;
1412         unsigned int total_backoff = 0;
1413
1414         BUG_ON(!dlm);
1415
1416         mlog(0, "Join domain %s\n", dlm->name);
1417
1418         status = dlm_register_domain_handlers(dlm);
1419         if (status) {
1420                 mlog_errno(status);
1421                 goto bail;
1422         }
1423
1424         status = dlm_debug_init(dlm);
1425         if (status < 0) {
1426                 mlog_errno(status);
1427                 goto bail;
1428         }
1429
1430         status = dlm_launch_thread(dlm);
1431         if (status < 0) {
1432                 mlog_errno(status);
1433                 goto bail;
1434         }
1435
1436         status = dlm_launch_recovery_thread(dlm);
1437         if (status < 0) {
1438                 mlog_errno(status);
1439                 goto bail;
1440         }
1441
1442         dlm->dlm_worker = create_singlethread_workqueue("dlm_wq");
1443         if (!dlm->dlm_worker) {
1444                 status = -ENOMEM;
1445                 mlog_errno(status);
1446                 goto bail;
1447         }
1448
1449         do {
1450                 status = dlm_try_to_join_domain(dlm);
1451
1452                 /* If we're racing another node to the join, then we
1453                  * need to back off temporarily and let them
1454                  * complete. */
1455 #define DLM_JOIN_TIMEOUT_MSECS  90000
1456                 if (status == -EAGAIN) {
1457                         if (signal_pending(current)) {
1458                                 status = -ERESTARTSYS;
1459                                 goto bail;
1460                         }
1461
1462                         if (total_backoff >
1463                             msecs_to_jiffies(DLM_JOIN_TIMEOUT_MSECS)) {
1464                                 status = -ERESTARTSYS;
1465                                 mlog(ML_NOTICE, "Timed out joining dlm domain "
1466                                      "%s after %u msecs\n", dlm->name,
1467                                      jiffies_to_msecs(total_backoff));
1468                                 goto bail;
1469                         }
1470
1471                         /*
1472                          * <chip> After you!
1473                          * <dale> No, after you!
1474                          * <chip> I insist!
1475                          * <dale> But you first!
1476                          * ...
1477                          */
1478                         backoff = (unsigned int)(jiffies & 0x3);
1479                         backoff *= DLM_DOMAIN_BACKOFF_MS;
1480                         total_backoff += backoff;
1481                         mlog(0, "backoff %d\n", backoff);
1482                         msleep(backoff);
1483                 }
1484         } while (status == -EAGAIN);
1485
1486         if (status < 0) {
1487                 mlog_errno(status);
1488                 goto bail;
1489         }
1490
1491         status = 0;
1492 bail:
1493         wake_up(&dlm_domain_events);
1494
1495         if (status) {
1496                 dlm_unregister_domain_handlers(dlm);
1497                 dlm_debug_shutdown(dlm);
1498                 dlm_complete_thread(dlm);
1499                 dlm_complete_recovery_thread(dlm);
1500                 dlm_destroy_dlm_worker(dlm);
1501         }
1502
1503         return status;
1504 }
1505
1506 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1507                                 u32 key)
1508 {
1509         int i;
1510         int ret;
1511         struct dlm_ctxt *dlm = NULL;
1512
1513         dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1514         if (!dlm) {
1515                 mlog_errno(-ENOMEM);
1516                 goto leave;
1517         }
1518
1519         dlm->name = kmalloc(strlen(domain) + 1, GFP_KERNEL);
1520         if (dlm->name == NULL) {
1521                 mlog_errno(-ENOMEM);
1522                 kfree(dlm);
1523                 dlm = NULL;
1524                 goto leave;
1525         }
1526
1527         dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1528         if (!dlm->lockres_hash) {
1529                 mlog_errno(-ENOMEM);
1530                 kfree(dlm->name);
1531                 kfree(dlm);
1532                 dlm = NULL;
1533                 goto leave;
1534         }
1535
1536         for (i = 0; i < DLM_HASH_BUCKETS; i++)
1537                 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1538
1539         dlm->master_hash = (struct hlist_head **)
1540                                 dlm_alloc_pagevec(DLM_HASH_PAGES);
1541         if (!dlm->master_hash) {
1542                 mlog_errno(-ENOMEM);
1543                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
1544                 kfree(dlm->name);
1545                 kfree(dlm);
1546                 dlm = NULL;
1547                 goto leave;
1548         }
1549
1550         for (i = 0; i < DLM_HASH_BUCKETS; i++)
1551                 INIT_HLIST_HEAD(dlm_master_hash(dlm, i));
1552
1553         strcpy(dlm->name, domain);
1554         dlm->key = key;
1555         dlm->node_num = o2nm_this_node();
1556
1557         ret = dlm_create_debugfs_subroot(dlm);
1558         if (ret < 0) {
1559                 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
1560                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
1561                 kfree(dlm->name);
1562                 kfree(dlm);
1563                 dlm = NULL;
1564                 goto leave;
1565         }
1566
1567         spin_lock_init(&dlm->spinlock);
1568         spin_lock_init(&dlm->master_lock);
1569         spin_lock_init(&dlm->ast_lock);
1570         spin_lock_init(&dlm->track_lock);
1571         INIT_LIST_HEAD(&dlm->list);
1572         INIT_LIST_HEAD(&dlm->dirty_list);
1573         INIT_LIST_HEAD(&dlm->reco.resources);
1574         INIT_LIST_HEAD(&dlm->reco.received);
1575         INIT_LIST_HEAD(&dlm->reco.node_data);
1576         INIT_LIST_HEAD(&dlm->purge_list);
1577         INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
1578         INIT_LIST_HEAD(&dlm->tracking_list);
1579         dlm->reco.state = 0;
1580
1581         INIT_LIST_HEAD(&dlm->pending_asts);
1582         INIT_LIST_HEAD(&dlm->pending_basts);
1583
1584         mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
1585                   dlm->recovery_map, &(dlm->recovery_map[0]));
1586
1587         memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
1588         memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
1589         memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
1590
1591         dlm->dlm_thread_task = NULL;
1592         dlm->dlm_reco_thread_task = NULL;
1593         dlm->dlm_worker = NULL;
1594         init_waitqueue_head(&dlm->dlm_thread_wq);
1595         init_waitqueue_head(&dlm->dlm_reco_thread_wq);
1596         init_waitqueue_head(&dlm->reco.event);
1597         init_waitqueue_head(&dlm->ast_wq);
1598         init_waitqueue_head(&dlm->migration_wq);
1599         INIT_LIST_HEAD(&dlm->mle_hb_events);
1600
1601         dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
1602         init_waitqueue_head(&dlm->dlm_join_events);
1603
1604         dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
1605         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
1606
1607         atomic_set(&dlm->res_tot_count, 0);
1608         atomic_set(&dlm->res_cur_count, 0);
1609         for (i = 0; i < DLM_MLE_NUM_TYPES; ++i) {
1610                 atomic_set(&dlm->mle_tot_count[i], 0);
1611                 atomic_set(&dlm->mle_cur_count[i], 0);
1612         }
1613
1614         spin_lock_init(&dlm->work_lock);
1615         INIT_LIST_HEAD(&dlm->work_list);
1616         INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
1617
1618         kref_init(&dlm->dlm_refs);
1619         dlm->dlm_state = DLM_CTXT_NEW;
1620
1621         INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
1622
1623         mlog(0, "context init: refcount %u\n",
1624                   atomic_read(&dlm->dlm_refs.refcount));
1625
1626 leave:
1627         return dlm;
1628 }
1629
1630 /*
1631  * Compare a requested locking protocol version against the current one.
1632  *
1633  * If the major numbers are different, they are incompatible.
1634  * If the current minor is greater than the request, they are incompatible.
1635  * If the current minor is less than or equal to the request, they are
1636  * compatible, and the requester should run at the current minor version.
1637  */
1638 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
1639                                 struct dlm_protocol_version *request)
1640 {
1641         if (existing->pv_major != request->pv_major)
1642                 return 1;
1643
1644         if (existing->pv_minor > request->pv_minor)
1645                 return 1;
1646
1647         if (existing->pv_minor < request->pv_minor)
1648                 request->pv_minor = existing->pv_minor;
1649
1650         return 0;
1651 }
1652
1653 /*
1654  * dlm_register_domain: one-time setup per "domain".
1655  *
1656  * The filesystem passes in the requested locking version via proto.
1657  * If registration was successful, proto will contain the negotiated
1658  * locking protocol.
1659  */
1660 struct dlm_ctxt * dlm_register_domain(const char *domain,
1661                                u32 key,
1662                                struct dlm_protocol_version *fs_proto)
1663 {
1664         int ret;
1665         struct dlm_ctxt *dlm = NULL;
1666         struct dlm_ctxt *new_ctxt = NULL;
1667
1668         if (strlen(domain) > O2NM_MAX_NAME_LEN) {
1669                 ret = -ENAMETOOLONG;
1670                 mlog(ML_ERROR, "domain name length too long\n");
1671                 goto leave;
1672         }
1673
1674         if (!o2hb_check_local_node_heartbeating()) {
1675                 mlog(ML_ERROR, "the local node has not been configured, or is "
1676                      "not heartbeating\n");
1677                 ret = -EPROTO;
1678                 goto leave;
1679         }
1680
1681         mlog(0, "register called for domain \"%s\"\n", domain);
1682
1683 retry:
1684         dlm = NULL;
1685         if (signal_pending(current)) {
1686                 ret = -ERESTARTSYS;
1687                 mlog_errno(ret);
1688                 goto leave;
1689         }
1690
1691         spin_lock(&dlm_domain_lock);
1692
1693         dlm = __dlm_lookup_domain(domain);
1694         if (dlm) {
1695                 if (dlm->dlm_state != DLM_CTXT_JOINED) {
1696                         spin_unlock(&dlm_domain_lock);
1697
1698                         mlog(0, "This ctxt is not joined yet!\n");
1699                         wait_event_interruptible(dlm_domain_events,
1700                                                  dlm_wait_on_domain_helper(
1701                                                          domain));
1702                         goto retry;
1703                 }
1704
1705                 if (dlm_protocol_compare(&dlm->fs_locking_proto, fs_proto)) {
1706                         mlog(ML_ERROR,
1707                              "Requested locking protocol version is not "
1708                              "compatible with already registered domain "
1709                              "\"%s\"\n", domain);
1710                         ret = -EPROTO;
1711                         goto leave;
1712                 }
1713
1714                 __dlm_get(dlm);
1715                 dlm->num_joins++;
1716
1717                 spin_unlock(&dlm_domain_lock);
1718
1719                 ret = 0;
1720                 goto leave;
1721         }
1722
1723         /* doesn't exist */
1724         if (!new_ctxt) {
1725                 spin_unlock(&dlm_domain_lock);
1726
1727                 new_ctxt = dlm_alloc_ctxt(domain, key);
1728                 if (new_ctxt)
1729                         goto retry;
1730
1731                 ret = -ENOMEM;
1732                 mlog_errno(ret);
1733                 goto leave;
1734         }
1735
1736         /* a little variable switch-a-roo here... */
1737         dlm = new_ctxt;
1738         new_ctxt = NULL;
1739
1740         /* add the new domain */
1741         list_add_tail(&dlm->list, &dlm_domains);
1742         spin_unlock(&dlm_domain_lock);
1743
1744         /*
1745          * Pass the locking protocol version into the join.  If the join
1746          * succeeds, it will have the negotiated protocol set.
1747          */
1748         dlm->dlm_locking_proto = dlm_protocol;
1749         dlm->fs_locking_proto = *fs_proto;
1750
1751         ret = dlm_join_domain(dlm);
1752         if (ret) {
1753                 mlog_errno(ret);
1754                 dlm_put(dlm);
1755                 goto leave;
1756         }
1757
1758         /* Tell the caller what locking protocol we negotiated */
1759         *fs_proto = dlm->fs_locking_proto;
1760
1761         ret = 0;
1762 leave:
1763         if (new_ctxt)
1764                 dlm_free_ctxt_mem(new_ctxt);
1765
1766         if (ret < 0)
1767                 dlm = ERR_PTR(ret);
1768
1769         return dlm;
1770 }
1771 EXPORT_SYMBOL_GPL(dlm_register_domain);
1772
1773 static LIST_HEAD(dlm_join_handlers);
1774
1775 static void dlm_unregister_net_handlers(void)
1776 {
1777         o2net_unregister_handler_list(&dlm_join_handlers);
1778 }
1779
1780 static int dlm_register_net_handlers(void)
1781 {
1782         int status = 0;
1783
1784         status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1785                                         sizeof(struct dlm_query_join_request),
1786                                         dlm_query_join_handler,
1787                                         NULL, NULL, &dlm_join_handlers);
1788         if (status)
1789                 goto bail;
1790
1791         status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1792                                         sizeof(struct dlm_assert_joined),
1793                                         dlm_assert_joined_handler,
1794                                         NULL, NULL, &dlm_join_handlers);
1795         if (status)
1796                 goto bail;
1797
1798         status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1799                                         sizeof(struct dlm_cancel_join),
1800                                         dlm_cancel_join_handler,
1801                                         NULL, NULL, &dlm_join_handlers);
1802
1803 bail:
1804         if (status < 0)
1805                 dlm_unregister_net_handlers();
1806
1807         return status;
1808 }
1809
1810 /* Domain eviction callback handling.
1811  *
1812  * The file system requires notification of node death *before* the
1813  * dlm completes it's recovery work, otherwise it may be able to
1814  * acquire locks on resources requiring recovery. Since the dlm can
1815  * evict a node from it's domain *before* heartbeat fires, a similar
1816  * mechanism is required. */
1817
1818 /* Eviction is not expected to happen often, so a per-domain lock is
1819  * not necessary. Eviction callbacks are allowed to sleep for short
1820  * periods of time. */
1821 static DECLARE_RWSEM(dlm_callback_sem);
1822
1823 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
1824                                         int node_num)
1825 {
1826         struct list_head *iter;
1827         struct dlm_eviction_cb *cb;
1828
1829         down_read(&dlm_callback_sem);
1830         list_for_each(iter, &dlm->dlm_eviction_callbacks) {
1831                 cb = list_entry(iter, struct dlm_eviction_cb, ec_item);
1832
1833                 cb->ec_func(node_num, cb->ec_data);
1834         }
1835         up_read(&dlm_callback_sem);
1836 }
1837
1838 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
1839                            dlm_eviction_func *f,
1840                            void *data)
1841 {
1842         INIT_LIST_HEAD(&cb->ec_item);
1843         cb->ec_func = f;
1844         cb->ec_data = data;
1845 }
1846 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
1847
1848 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
1849                               struct dlm_eviction_cb *cb)
1850 {
1851         down_write(&dlm_callback_sem);
1852         list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
1853         up_write(&dlm_callback_sem);
1854 }
1855 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
1856
1857 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
1858 {
1859         down_write(&dlm_callback_sem);
1860         list_del_init(&cb->ec_item);
1861         up_write(&dlm_callback_sem);
1862 }
1863 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
1864
1865 static int __init dlm_init(void)
1866 {
1867         int status;
1868
1869         dlm_print_version();
1870
1871         status = dlm_init_mle_cache();
1872         if (status) {
1873                 mlog(ML_ERROR, "Could not create o2dlm_mle slabcache\n");
1874                 goto error;
1875         }
1876
1877         status = dlm_init_master_caches();
1878         if (status) {
1879                 mlog(ML_ERROR, "Could not create o2dlm_lockres and "
1880                      "o2dlm_lockname slabcaches\n");
1881                 goto error;
1882         }
1883
1884         status = dlm_init_lock_cache();
1885         if (status) {
1886                 mlog(ML_ERROR, "Count not create o2dlm_lock slabcache\n");
1887                 goto error;
1888         }
1889
1890         status = dlm_register_net_handlers();
1891         if (status) {
1892                 mlog(ML_ERROR, "Unable to register network handlers\n");
1893                 goto error;
1894         }
1895
1896         status = dlm_create_debugfs_root();
1897         if (status)
1898                 goto error;
1899
1900         return 0;
1901 error:
1902         dlm_unregister_net_handlers();
1903         dlm_destroy_lock_cache();
1904         dlm_destroy_master_caches();
1905         dlm_destroy_mle_cache();
1906         return -1;
1907 }
1908
1909 static void __exit dlm_exit (void)
1910 {
1911         dlm_destroy_debugfs_root();
1912         dlm_unregister_net_handlers();
1913         dlm_destroy_lock_cache();
1914         dlm_destroy_master_caches();
1915         dlm_destroy_mle_cache();
1916 }
1917
1918 MODULE_AUTHOR("Oracle");
1919 MODULE_LICENSE("GPL");
1920
1921 module_init(dlm_init);
1922 module_exit(dlm_exit);