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  * New in version 1.1:
 133  *      - Message DLM_QUERY_REGION added to support global heartbeat
 134  *      - Message DLM_QUERY_NODEINFO added to allow online node removes
 135  * New in version 1.2:
 136  *      - Message DLM_BEGIN_EXIT_DOMAIN_MSG added to mark start of exit domain
 137  */
 138 static const struct dlm_protocol_version dlm_protocol = {
 139         .pv_major = 1,
 140         .pv_minor = 2,
 141 };
 142
 143 #define DLM_DOMAIN_BACKOFF_MS 200
 144
 145 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
 146                                   void **ret_data);
 147 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 148                                      void **ret_data);
 149 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
 150                                    void **ret_data);
 151 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
 152                                     void *data, void **ret_data);
 153 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 154                                    void **ret_data);
 155 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
 156                                 struct dlm_protocol_version *request);
 157
 158 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
 159
 160 void __dlm_unhash_lockres(struct dlm_lock_resource *lockres)
 161 {
 162         if (!hlist_unhashed(&lockres->hash_node)) {
 163                 hlist_del_init(&lockres->hash_node);
 164                 dlm_lockres_put(lockres);
 165         }
 166 }
 167
 168 void __dlm_insert_lockres(struct dlm_ctxt *dlm,
 169                        struct dlm_lock_resource *res)
 170 {
 171         struct hlist_head *bucket;
 172         struct qstr *q;
 173
 174         assert_spin_locked(&dlm->spinlock);
 175
 176         q = &res->lockname;
 177         bucket = dlm_lockres_hash(dlm, q->hash);
 178
 179         /* get a reference for our hashtable */
 180         dlm_lockres_get(res);
 181
 182         hlist_add_head(&res->hash_node, bucket);
 183 }
 184
 185 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
 186                                                      const char *name,
 187                                                      unsigned int len,
 188                                                      unsigned int hash)
 189 {
 190         struct hlist_head *bucket;
 191         struct hlist_node *list;
 192
 193         mlog(0, "%.*s\n", len, name);
 194
 195         assert_spin_locked(&dlm->spinlock);
 196
 197         bucket = dlm_lockres_hash(dlm, hash);
 198
 199         hlist_for_each(list, bucket) {
 200                 struct dlm_lock_resource *res = hlist_entry(list,
 201                         struct dlm_lock_resource, hash_node);
 202                 if (res->lockname.name[0] != name[0])
 203                         continue;
 204                 if (unlikely(res->lockname.len != len))
 205                         continue;
 206                 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
 207                         continue;
 208                 dlm_lockres_get(res);
 209                 return res;
 210         }
 211         return NULL;
 212 }
 213
 214 /* intended to be called by functions which do not care about lock
 215  * resources which are being purged (most net _handler functions).
 216  * this will return NULL for any lock resource which is found but
 217  * currently in the process of dropping its mastery reference.
 218  * use __dlm_lookup_lockres_full when you need the lock resource
 219  * regardless (e.g. dlm_get_lock_resource) */
 220 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
 221                                                 const char *name,
 222                                                 unsigned int len,
 223                                                 unsigned int hash)
 224 {
 225         struct dlm_lock_resource *res = NULL;
 226
 227         mlog(0, "%.*s\n", len, name);
 228
 229         assert_spin_locked(&dlm->spinlock);
 230
 231         res = __dlm_lookup_lockres_full(dlm, name, len, hash);
 232         if (res) {
 233                 spin_lock(&res->spinlock);
 234                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
 235                         spin_unlock(&res->spinlock);
 236                         dlm_lockres_put(res);
 237                         return NULL;
 238                 }
 239                 spin_unlock(&res->spinlock);
 240         }
 241
 242         return res;
 243 }
 244
 245 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
 246                                     const char *name,
 247                                     unsigned int len)
 248 {
 249         struct dlm_lock_resource *res;
 250         unsigned int hash = dlm_lockid_hash(name, len);
 251
 252         spin_lock(&dlm->spinlock);
 253         res = __dlm_lookup_lockres(dlm, name, len, hash);
 254         spin_unlock(&dlm->spinlock);
 255         return res;
 256 }
 257
 258 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
 259 {
 260         struct dlm_ctxt *tmp = NULL;
 261         struct list_head *iter;
 262
 263         assert_spin_locked(&dlm_domain_lock);
 264
 265         /* tmp->name here is always NULL terminated,
 266          * but domain may not be! */
 267         list_for_each(iter, &dlm_domains) {
 268                 tmp = list_entry (iter, struct dlm_ctxt, list);
 269                 if (strlen(tmp->name) == len &&
 270                     memcmp(tmp->name, domain, len)==0)
 271                         break;
 272                 tmp = NULL;
 273         }
 274
 275         return tmp;
 276 }
 277
 278 /* For null terminated domain strings ONLY */
 279 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
 280 {
 281         assert_spin_locked(&dlm_domain_lock);
 282
 283         return __dlm_lookup_domain_full(domain, strlen(domain));
 284 }
 285
 286
 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)
 291 {
 292         int ret = 0;
 293         struct dlm_ctxt *tmp = NULL;
 294
 295         spin_lock(&dlm_domain_lock);
 296
 297         tmp = __dlm_lookup_domain(domain);
 298         if (!tmp)
 299                 ret = 1;
 300         else if (tmp->dlm_state == DLM_CTXT_JOINED)
 301                 ret = 1;
 302
 303         spin_unlock(&dlm_domain_lock);
 304         return ret;
 305 }
 306
 307 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
 308 {
 309         dlm_destroy_debugfs_subroot(dlm);
 310
 311         if (dlm->lockres_hash)
 312                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
 313
 314         if (dlm->master_hash)
 315                 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
 316
 317         if (dlm->name)
 318                 kfree(dlm->name);
 319
 320         kfree(dlm);
 321 }
 322
 323 /* A little strange - this function will be called while holding
 324  * dlm_domain_lock and is expected to be holding it on the way out. We
 325  * will however drop and reacquire it multiple times */
 326 static void dlm_ctxt_release(struct kref *kref)
 327 {
 328         struct dlm_ctxt *dlm;
 329
 330         dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
 331
 332         BUG_ON(dlm->num_joins);
 333         BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
 334
 335         /* we may still be in the list if we hit an error during join. */
 336         list_del_init(&dlm->list);
 337
 338         spin_unlock(&dlm_domain_lock);
 339
 340         mlog(0, "freeing memory from domain %s\n", dlm->name);
 341
 342         wake_up(&dlm_domain_events);
 343
 344         dlm_free_ctxt_mem(dlm);
 345
 346         spin_lock(&dlm_domain_lock);
 347 }
 348
 349 void dlm_put(struct dlm_ctxt *dlm)
 350 {
 351         spin_lock(&dlm_domain_lock);
 352         kref_put(&dlm->dlm_refs, dlm_ctxt_release);
 353         spin_unlock(&dlm_domain_lock);
 354 }
 355
 356 static void __dlm_get(struct dlm_ctxt *dlm)
 357 {
 358         kref_get(&dlm->dlm_refs);
 359 }
 360
 361 /* given a questionable reference to a dlm object, gets a reference if
 362  * it can find it in the list, otherwise returns NULL in which case
 363  * you shouldn't trust your pointer. */
 364 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
 365 {
 366         struct list_head *iter;
 367         struct dlm_ctxt *target = NULL;
 368
 369         spin_lock(&dlm_domain_lock);
 370
 371         list_for_each(iter, &dlm_domains) {
 372                 target = list_entry (iter, struct dlm_ctxt, list);
 373
 374                 if (target == dlm) {
 375                         __dlm_get(target);
 376                         break;
 377                 }
 378
 379                 target = NULL;
 380         }
 381
 382         spin_unlock(&dlm_domain_lock);
 383
 384         return target;
 385 }
 386
 387 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
 388 {
 389         int ret;
 390
 391         spin_lock(&dlm_domain_lock);
 392         ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
 393                 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
 394         spin_unlock(&dlm_domain_lock);
 395
 396         return ret;
 397 }
 398
 399 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
 400 {
 401         if (dlm->dlm_worker) {
 402                 flush_workqueue(dlm->dlm_worker);
 403                 destroy_workqueue(dlm->dlm_worker);
 404                 dlm->dlm_worker = NULL;
 405         }
 406 }
 407
 408 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
 409 {
 410         dlm_unregister_domain_handlers(dlm);
 411         dlm_debug_shutdown(dlm);
 412         dlm_complete_thread(dlm);
 413         dlm_complete_recovery_thread(dlm);
 414         dlm_destroy_dlm_worker(dlm);
 415
 416         /* We've left the domain. Now we can take ourselves out of the
 417          * list and allow the kref stuff to help us free the
 418          * memory. */
 419         spin_lock(&dlm_domain_lock);
 420         list_del_init(&dlm->list);
 421         spin_unlock(&dlm_domain_lock);
 422
 423         /* Wake up anyone waiting for us to remove this domain */
 424         wake_up(&dlm_domain_events);
 425 }
 426
 427 static int dlm_migrate_all_locks(struct dlm_ctxt *dlm)
 428 {
 429         int i, num, n, ret = 0;
 430         struct dlm_lock_resource *res;
 431         struct hlist_node *iter;
 432         struct hlist_head *bucket;
 433         int dropped;
 434
 435         mlog(0, "Migrating locks from domain %s\n", dlm->name);
 436
 437         num = 0;
 438         spin_lock(&dlm->spinlock);
 439         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
 440 redo_bucket:
 441                 n = 0;
 442                 bucket = dlm_lockres_hash(dlm, i);
 443                 iter = bucket->first;
 444                 while (iter) {
 445                         n++;
 446                         res = hlist_entry(iter, struct dlm_lock_resource,
 447                                           hash_node);
 448                         dlm_lockres_get(res);
 449                         /* migrate, if necessary.  this will drop the dlm
 450                          * spinlock and retake it if it does migration. */
 451                         dropped = dlm_empty_lockres(dlm, res);
 452
 453                         spin_lock(&res->spinlock);
 454                         if (dropped)
 455                                 __dlm_lockres_calc_usage(dlm, res);
 456                         else
 457                                 iter = res->hash_node.next;
 458                         spin_unlock(&res->spinlock);
 459
 460                         dlm_lockres_put(res);
 461
 462                         if (dropped) {
 463                                 cond_resched_lock(&dlm->spinlock);
 464                                 goto redo_bucket;
 465                         }
 466                 }
 467                 cond_resched_lock(&dlm->spinlock);
 468                 num += n;
 469         }
 470         spin_unlock(&dlm->spinlock);
 471         wake_up(&dlm->dlm_thread_wq);
 472
 473         /* let the dlm thread take care of purging, keep scanning until
 474          * nothing remains in the hash */
 475         if (num) {
 476                 mlog(0, "%s: %d lock resources in hash last pass\n",
 477                      dlm->name, num);
 478                 ret = -EAGAIN;
 479         }
 480         mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
 481         return ret;
 482 }
 483
 484 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
 485 {
 486         int ret;
 487
 488         spin_lock(&dlm->spinlock);
 489         ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
 490         spin_unlock(&dlm->spinlock);
 491
 492         return ret;
 493 }
 494
 495 static int dlm_begin_exit_domain_handler(struct o2net_msg *msg, u32 len,
 496                                          void *data, void **ret_data)
 497 {
 498         struct dlm_ctxt *dlm = data;
 499         unsigned int node;
 500         struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
 501
 502         if (!dlm_grab(dlm))
 503                 return 0;
 504
 505         node = exit_msg->node_idx;
 506         mlog(0, "%s: Node %u sent a begin exit domain message\n", dlm->name, node);
 507
 508         spin_lock(&dlm->spinlock);
 509         set_bit(node, dlm->exit_domain_map);
 510         spin_unlock(&dlm->spinlock);
 511
 512         dlm_put(dlm);
 513
 514         return 0;
 515 }
 516
 517 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
 518 {
 519         /* Yikes, a double spinlock! I need domain_lock for the dlm
 520          * state and the dlm spinlock for join state... Sorry! */
 521 again:
 522         spin_lock(&dlm_domain_lock);
 523         spin_lock(&dlm->spinlock);
 524
 525         if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 526                 mlog(0, "Node %d is joining, we wait on it.\n",
 527                           dlm->joining_node);
 528                 spin_unlock(&dlm->spinlock);
 529                 spin_unlock(&dlm_domain_lock);
 530
 531                 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
 532                 goto again;
 533         }
 534
 535         dlm->dlm_state = DLM_CTXT_LEAVING;
 536         spin_unlock(&dlm->spinlock);
 537         spin_unlock(&dlm_domain_lock);
 538 }
 539
 540 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
 541 {
 542         int node = -1;
 543
 544         assert_spin_locked(&dlm->spinlock);
 545
 546         printk(KERN_NOTICE "o2dlm: Nodes in domain %s: ", dlm->name);
 547
 548         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 549                                      node + 1)) < O2NM_MAX_NODES) {
 550                 printk("%d ", node);
 551         }
 552         printk("\n");
 553 }
 554
 555 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 556                                    void **ret_data)
 557 {
 558         struct dlm_ctxt *dlm = data;
 559         unsigned int node;
 560         struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
 561
 562         mlog(0, "%p %u %p", msg, len, data);
 563
 564         if (!dlm_grab(dlm))
 565                 return 0;
 566
 567         node = exit_msg->node_idx;
 568
 569         printk(KERN_NOTICE "o2dlm: Node %u leaves domain %s\n", node, dlm->name);
 570
 571         spin_lock(&dlm->spinlock);
 572         clear_bit(node, dlm->domain_map);
 573         clear_bit(node, dlm->exit_domain_map);
 574         __dlm_print_nodes(dlm);
 575
 576         /* notify anything attached to the heartbeat events */
 577         dlm_hb_event_notify_attached(dlm, node, 0);
 578
 579         spin_unlock(&dlm->spinlock);
 580
 581         dlm_put(dlm);
 582
 583         return 0;
 584 }
 585
 586 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm, u32 msg_type,
 587                                     unsigned int node)
 588 {
 589         int status;
 590         struct dlm_exit_domain leave_msg;
 591
 592         mlog(0, "%s: Sending domain exit message %u to node %u\n", dlm->name,
 593              msg_type, node);
 594
 595         memset(&leave_msg, 0, sizeof(leave_msg));
 596         leave_msg.node_idx = dlm->node_num;
 597
 598         status = o2net_send_message(msg_type, dlm->key, &leave_msg,
 599                                     sizeof(leave_msg), node, NULL);
 600         if (status < 0)
 601                 mlog(ML_ERROR, "Error %d sending domain exit message %u "
 602                      "to node %u on domain %s\n", status, msg_type, node,
 603                      dlm->name);
 604
 605         return status;
 606 }
 607
 608 static void dlm_begin_exit_domain(struct dlm_ctxt *dlm)
 609 {
 610         int node = -1;
 611
 612         /* Support for begin exit domain was added in 1.2 */
 613         if (dlm->dlm_locking_proto.pv_major == 1 &&
 614             dlm->dlm_locking_proto.pv_minor < 2)
 615                 return;
 616
 617         /*
 618          * Unlike DLM_EXIT_DOMAIN_MSG, DLM_BEGIN_EXIT_DOMAIN_MSG is purely
 619          * informational. Meaning if a node does not receive the message,
 620          * so be it.
 621          */
 622         spin_lock(&dlm->spinlock);
 623         while (1) {
 624                 node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, node + 1);
 625                 if (node >= O2NM_MAX_NODES)
 626                         break;
 627                 if (node == dlm->node_num)
 628                         continue;
 629
 630                 spin_unlock(&dlm->spinlock);
 631                 dlm_send_one_domain_exit(dlm, DLM_BEGIN_EXIT_DOMAIN_MSG, node);
 632                 spin_lock(&dlm->spinlock);
 633         }
 634         spin_unlock(&dlm->spinlock);
 635 }
 636
 637 static void dlm_leave_domain(struct dlm_ctxt *dlm)
 638 {
 639         int node, clear_node, status;
 640
 641         /* At this point we've migrated away all our locks and won't
 642          * accept mastership of new ones. The dlm is responsible for
 643          * almost nothing now. We make sure not to confuse any joining
 644          * nodes and then commence shutdown procedure. */
 645
 646         spin_lock(&dlm->spinlock);
 647         /* Clear ourselves from the domain map */
 648         clear_bit(dlm->node_num, dlm->domain_map);
 649         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 650                                      0)) < O2NM_MAX_NODES) {
 651                 /* Drop the dlm spinlock. This is safe wrt the domain_map.
 652                  * -nodes cannot be added now as the
 653                  *   query_join_handlers knows to respond with OK_NO_MAP
 654                  * -we catch the right network errors if a node is
 655                  *   removed from the map while we're sending him the
 656                  *   exit message. */
 657                 spin_unlock(&dlm->spinlock);
 658
 659                 clear_node = 1;
 660
 661                 status = dlm_send_one_domain_exit(dlm, DLM_EXIT_DOMAIN_MSG,
 662                                                   node);
 663                 if (status < 0 &&
 664                     status != -ENOPROTOOPT &&
 665                     status != -ENOTCONN) {
 666                         mlog(ML_NOTICE, "Error %d sending domain exit message "
 667                              "to node %d\n", status, node);
 668
 669                         /* Not sure what to do here but lets sleep for
 670                          * a bit in case this was a transient
 671                          * error... */
 672                         msleep(DLM_DOMAIN_BACKOFF_MS);
 673                         clear_node = 0;
 674                 }
 675
 676                 spin_lock(&dlm->spinlock);
 677                 /* If we're not clearing the node bit then we intend
 678                  * to loop back around to try again. */
 679                 if (clear_node)
 680                         clear_bit(node, dlm->domain_map);
 681         }
 682         spin_unlock(&dlm->spinlock);
 683 }
 684
 685 int dlm_joined(struct dlm_ctxt *dlm)
 686 {
 687         int ret = 0;
 688
 689         spin_lock(&dlm_domain_lock);
 690
 691         if (dlm->dlm_state == DLM_CTXT_JOINED)
 692                 ret = 1;
 693
 694         spin_unlock(&dlm_domain_lock);
 695
 696         return ret;
 697 }
 698
 699 int dlm_shutting_down(struct dlm_ctxt *dlm)
 700 {
 701         int ret = 0;
 702
 703         spin_lock(&dlm_domain_lock);
 704
 705         if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
 706                 ret = 1;
 707
 708         spin_unlock(&dlm_domain_lock);
 709
 710         return ret;
 711 }
 712
 713 void dlm_unregister_domain(struct dlm_ctxt *dlm)
 714 {
 715         int leave = 0;
 716         struct dlm_lock_resource *res;
 717
 718         spin_lock(&dlm_domain_lock);
 719         BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
 720         BUG_ON(!dlm->num_joins);
 721
 722         dlm->num_joins--;
 723         if (!dlm->num_joins) {
 724                 /* We mark it "in shutdown" now so new register
 725                  * requests wait until we've completely left the
 726                  * domain. Don't use DLM_CTXT_LEAVING yet as we still
 727                  * want new domain joins to communicate with us at
 728                  * least until we've completed migration of our
 729                  * resources. */
 730                 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
 731                 leave = 1;
 732         }
 733         spin_unlock(&dlm_domain_lock);
 734
 735         if (leave) {
 736                 mlog(0, "shutting down domain %s\n", dlm->name);
 737                 dlm_begin_exit_domain(dlm);
 738
 739                 /* We changed dlm state, notify the thread */
 740                 dlm_kick_thread(dlm, NULL);
 741
 742                 while (dlm_migrate_all_locks(dlm)) {
 743                         /* Give dlm_thread time to purge the lockres' */
 744                         msleep(500);
 745                         mlog(0, "%s: more migration to do\n", dlm->name);
 746                 }
 747
 748                 /* This list should be empty. If not, print remaining lockres */
 749                 if (!list_empty(&dlm->tracking_list)) {
 750                         mlog(ML_ERROR, "Following lockres' are still on the "
 751                              "tracking list:\n");
 752                         list_for_each_entry(res, &dlm->tracking_list, tracking)
 753                                 dlm_print_one_lock_resource(res);
 754                 }
 755
 756                 dlm_mark_domain_leaving(dlm);
 757                 dlm_leave_domain(dlm);
 758                 dlm_force_free_mles(dlm);
 759                 dlm_complete_dlm_shutdown(dlm);
 760         }
 761         dlm_put(dlm);
 762 }
 763 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
 764
 765 static int dlm_query_join_proto_check(char *proto_type, int node,
 766                                       struct dlm_protocol_version *ours,
 767                                       struct dlm_protocol_version *request)
 768 {
 769         int rc;
 770         struct dlm_protocol_version proto = *request;
 771
 772         if (!dlm_protocol_compare(ours, &proto)) {
 773                 mlog(0,
 774                      "node %u wanted to join with %s locking protocol "
 775                      "%u.%u, we respond with %u.%u\n",
 776                      node, proto_type,
 777                      request->pv_major,
 778                      request->pv_minor,
 779                      proto.pv_major, proto.pv_minor);
 780                 request->pv_minor = proto.pv_minor;
 781                 rc = 0;
 782         } else {
 783                 mlog(ML_NOTICE,
 784                      "Node %u wanted to join with %s locking "
 785                      "protocol %u.%u, but we have %u.%u, disallowing\n",
 786                      node, proto_type,
 787                      request->pv_major,
 788                      request->pv_minor,
 789                      ours->pv_major,
 790                      ours->pv_minor);
 791                 rc = 1;
 792         }
 793
 794         return rc;
 795 }
 796
 797 /*
 798  * struct dlm_query_join_packet is made up of four one-byte fields.  They
 799  * are effectively in big-endian order already.  However, little-endian
 800  * machines swap them before putting the packet on the wire (because
 801  * query_join's response is a status, and that status is treated as a u32
 802  * on the wire).  Thus, a big-endian and little-endian machines will treat
 803  * this structure differently.
 804  *
 805  * The solution is to have little-endian machines swap the structure when
 806  * converting from the structure to the u32 representation.  This will
 807  * result in the structure having the correct format on the wire no matter
 808  * the host endian format.
 809  */
 810 static void dlm_query_join_packet_to_wire(struct dlm_query_join_packet *packet,
 811                                           u32 *wire)
 812 {
 813         union dlm_query_join_response response;
 814
 815         response.packet = *packet;
 816         *wire = cpu_to_be32(response.intval);
 817 }
 818
 819 static void dlm_query_join_wire_to_packet(u32 wire,
 820                                           struct dlm_query_join_packet *packet)
 821 {
 822         union dlm_query_join_response response;
 823
 824         response.intval = cpu_to_be32(wire);
 825         *packet = response.packet;
 826 }
 827
 828 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
 829                                   void **ret_data)
 830 {
 831         struct dlm_query_join_request *query;
 832         struct dlm_query_join_packet packet = {
 833                 .code = JOIN_DISALLOW,
 834         };
 835         struct dlm_ctxt *dlm = NULL;
 836         u32 response;
 837         u8 nodenum;
 838
 839         query = (struct dlm_query_join_request *) msg->buf;
 840
 841         mlog(0, "node %u wants to join domain %s\n", query->node_idx,
 842                   query->domain);
 843
 844         /*
 845          * If heartbeat doesn't consider the node live, tell it
 846          * to back off and try again.  This gives heartbeat a chance
 847          * to catch up.
 848          */
 849         if (!o2hb_check_node_heartbeating(query->node_idx)) {
 850                 mlog(0, "node %u is not in our live map yet\n",
 851                      query->node_idx);
 852
 853                 packet.code = JOIN_DISALLOW;
 854                 goto respond;
 855         }
 856
 857         packet.code = JOIN_OK_NO_MAP;
 858
 859         spin_lock(&dlm_domain_lock);
 860         dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
 861         if (!dlm)
 862                 goto unlock_respond;
 863
 864         /*
 865          * There is a small window where the joining node may not see the
 866          * node(s) that just left but still part of the cluster. DISALLOW
 867          * join request if joining node has different node map.
 868          */
 869         nodenum=0;
 870         while (nodenum < O2NM_MAX_NODES) {
 871                 if (test_bit(nodenum, dlm->domain_map)) {
 872                         if (!byte_test_bit(nodenum, query->node_map)) {
 873                                 mlog(0, "disallow join as node %u does not "
 874                                      "have node %u in its nodemap\n",
 875                                      query->node_idx, nodenum);
 876                                 packet.code = JOIN_DISALLOW;
 877                                 goto unlock_respond;
 878                         }
 879                 }
 880                 nodenum++;
 881         }
 882
 883         /* Once the dlm ctxt is marked as leaving then we don't want
 884          * to be put in someone's domain map.
 885          * Also, explicitly disallow joining at certain troublesome
 886          * times (ie. during recovery). */
 887         if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
 888                 int bit = query->node_idx;
 889                 spin_lock(&dlm->spinlock);
 890
 891                 if (dlm->dlm_state == DLM_CTXT_NEW &&
 892                     dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
 893                         /*If this is a brand new context and we
 894                          * haven't started our join process yet, then
 895                          * the other node won the race. */
 896                         packet.code = JOIN_OK_NO_MAP;
 897                 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 898                         /* Disallow parallel joins. */
 899                         packet.code = JOIN_DISALLOW;
 900                 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
 901                         mlog(0, "node %u trying to join, but recovery "
 902                              "is ongoing.\n", bit);
 903                         packet.code = JOIN_DISALLOW;
 904                 } else if (test_bit(bit, dlm->recovery_map)) {
 905                         mlog(0, "node %u trying to join, but it "
 906                              "still needs recovery.\n", bit);
 907                         packet.code = JOIN_DISALLOW;
 908                 } else if (test_bit(bit, dlm->domain_map)) {
 909                         mlog(0, "node %u trying to join, but it "
 910                              "is still in the domain! needs recovery?\n",
 911                              bit);
 912                         packet.code = JOIN_DISALLOW;
 913                 } else {
 914                         /* Alright we're fully a part of this domain
 915                          * so we keep some state as to who's joining
 916                          * and indicate to him that needs to be fixed
 917                          * up. */
 918
 919                         /* Make sure we speak compatible locking protocols.  */
 920                         if (dlm_query_join_proto_check("DLM", bit,
 921                                                        &dlm->dlm_locking_proto,
 922                                                        &query->dlm_proto)) {
 923                                 packet.code = JOIN_PROTOCOL_MISMATCH;
 924                         } else if (dlm_query_join_proto_check("fs", bit,
 925                                                               &dlm->fs_locking_proto,
 926                                                               &query->fs_proto)) {
 927                                 packet.code = JOIN_PROTOCOL_MISMATCH;
 928                         } else {
 929                                 packet.dlm_minor = query->dlm_proto.pv_minor;
 930                                 packet.fs_minor = query->fs_proto.pv_minor;
 931                                 packet.code = JOIN_OK;
 932                                 __dlm_set_joining_node(dlm, query->node_idx);
 933                         }
 934                 }
 935
 936                 spin_unlock(&dlm->spinlock);
 937         }
 938 unlock_respond:
 939         spin_unlock(&dlm_domain_lock);
 940
 941 respond:
 942         mlog(0, "We respond with %u\n", packet.code);
 943
 944         dlm_query_join_packet_to_wire(&packet, &response);
 945         return response;
 946 }
 947
 948 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 949                                      void **ret_data)
 950 {
 951         struct dlm_assert_joined *assert;
 952         struct dlm_ctxt *dlm = NULL;
 953
 954         assert = (struct dlm_assert_joined *) msg->buf;
 955
 956         mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
 957                   assert->domain);
 958
 959         spin_lock(&dlm_domain_lock);
 960         dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
 961         /* XXX should we consider no dlm ctxt an error? */
 962         if (dlm) {
 963                 spin_lock(&dlm->spinlock);
 964
 965                 /* Alright, this node has officially joined our
 966                  * domain. Set him in the map and clean up our
 967                  * leftover join state. */
 968                 BUG_ON(dlm->joining_node != assert->node_idx);
 969                 set_bit(assert->node_idx, dlm->domain_map);
 970                 clear_bit(assert->node_idx, dlm->exit_domain_map);
 971                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 972
 973                 printk(KERN_NOTICE "o2dlm: Node %u joins domain %s\n",
 974                        assert->node_idx, dlm->name);
 975                 __dlm_print_nodes(dlm);
 976
 977                 /* notify anything attached to the heartbeat events */
 978                 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
 979
 980                 spin_unlock(&dlm->spinlock);
 981         }
 982         spin_unlock(&dlm_domain_lock);
 983
 984         return 0;
 985 }
 986
 987 static int dlm_match_regions(struct dlm_ctxt *dlm,
 988                              struct dlm_query_region *qr,
 989                              char *local, int locallen)
 990 {
 991         char *remote = qr->qr_regions;
 992         char *l, *r;
 993         int localnr, i, j, foundit;
 994         int status = 0;
 995
 996         if (!o2hb_global_heartbeat_active()) {
 997                 if (qr->qr_numregions) {
 998                         mlog(ML_ERROR, "Domain %s: Joining node %d has global "
 999                              "heartbeat enabled but local node %d does not\n",
1000                              qr->qr_domain, qr->qr_node, dlm->node_num);
1001                         status = -EINVAL;
1002                 }
1003                 goto bail;
1004         }
1005
1006         if (o2hb_global_heartbeat_active() && !qr->qr_numregions) {
1007                 mlog(ML_ERROR, "Domain %s: Local node %d has global "
1008                      "heartbeat enabled but joining node %d does not\n",
1009                      qr->qr_domain, dlm->node_num, qr->qr_node);
1010                 status = -EINVAL;
1011                 goto bail;
1012         }
1013
1014         r = remote;
1015         for (i = 0; i < qr->qr_numregions; ++i) {
1016                 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, r);
1017                 r += O2HB_MAX_REGION_NAME_LEN;
1018         }
1019
1020         localnr = min(O2NM_MAX_REGIONS, locallen/O2HB_MAX_REGION_NAME_LEN);
1021         localnr = o2hb_get_all_regions(local, (u8)localnr);
1022
1023         /* compare local regions with remote */
1024         l = local;
1025         for (i = 0; i < localnr; ++i) {
1026                 foundit = 0;
1027                 r = remote;
1028                 for (j = 0; j <= qr->qr_numregions; ++j) {
1029                         if (!memcmp(l, r, O2HB_MAX_REGION_NAME_LEN)) {
1030                                 foundit = 1;
1031                                 break;
1032                         }
1033                         r += O2HB_MAX_REGION_NAME_LEN;
1034                 }
1035                 if (!foundit) {
1036                         status = -EINVAL;
1037                         mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1038                              "in local node %d but not in joining node %d\n",
1039                              qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, l,
1040                              dlm->node_num, qr->qr_node);
1041                         goto bail;
1042                 }
1043                 l += O2HB_MAX_REGION_NAME_LEN;
1044         }
1045
1046         /* compare remote with local regions */
1047         r = remote;
1048         for (i = 0; i < qr->qr_numregions; ++i) {
1049                 foundit = 0;
1050                 l = local;
1051                 for (j = 0; j < localnr; ++j) {
1052                         if (!memcmp(r, l, O2HB_MAX_REGION_NAME_LEN)) {
1053                                 foundit = 1;
1054                                 break;
1055                         }
1056                         l += O2HB_MAX_REGION_NAME_LEN;
1057                 }
1058                 if (!foundit) {
1059                         status = -EINVAL;
1060                         mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1061                              "in joining node %d but not in local node %d\n",
1062                              qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, r,
1063                              qr->qr_node, dlm->node_num);
1064                         goto bail;
1065                 }
1066                 r += O2HB_MAX_REGION_NAME_LEN;
1067         }
1068
1069 bail:
1070         return status;
1071 }
1072
1073 static int dlm_send_regions(struct dlm_ctxt *dlm, unsigned long *node_map)
1074 {
1075         struct dlm_query_region *qr = NULL;
1076         int status, ret = 0, i;
1077         char *p;
1078
1079         if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1080                 goto bail;
1081
1082         qr = kzalloc(sizeof(struct dlm_query_region), GFP_KERNEL);
1083         if (!qr) {
1084                 ret = -ENOMEM;
1085                 mlog_errno(ret);
1086                 goto bail;
1087         }
1088
1089         qr->qr_node = dlm->node_num;
1090         qr->qr_namelen = strlen(dlm->name);
1091         memcpy(qr->qr_domain, dlm->name, qr->qr_namelen);
1092         /* if local hb, the numregions will be zero */
1093         if (o2hb_global_heartbeat_active())
1094                 qr->qr_numregions = o2hb_get_all_regions(qr->qr_regions,
1095                                                          O2NM_MAX_REGIONS);
1096
1097         p = qr->qr_regions;
1098         for (i = 0; i < qr->qr_numregions; ++i, p += O2HB_MAX_REGION_NAME_LEN)
1099                 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, p);
1100
1101         i = -1;
1102         while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1103                                   i + 1)) < O2NM_MAX_NODES) {
1104                 if (i == dlm->node_num)
1105                         continue;
1106
1107                 mlog(0, "Sending regions to node %d\n", i);
1108
1109                 ret = o2net_send_message(DLM_QUERY_REGION, DLM_MOD_KEY, qr,
1110                                          sizeof(struct dlm_query_region),
1111                                          i, &status);
1112                 if (ret >= 0)
1113                         ret = status;
1114                 if (ret) {
1115                         mlog(ML_ERROR, "Region mismatch %d, node %d\n",
1116                              ret, i);
1117                         break;
1118                 }
1119         }
1120
1121 bail:
1122         kfree(qr);
1123         return ret;
1124 }
1125
1126 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
1127                                     void *data, void **ret_data)
1128 {
1129         struct dlm_query_region *qr;
1130         struct dlm_ctxt *dlm = NULL;
1131         char *local = NULL;
1132         int status = 0;
1133         int locked = 0;
1134
1135         qr = (struct dlm_query_region *) msg->buf;
1136
1137         mlog(0, "Node %u queries hb regions on domain %s\n", qr->qr_node,
1138              qr->qr_domain);
1139
1140         /* buffer used in dlm_mast_regions() */
1141         local = kmalloc(sizeof(qr->qr_regions), GFP_KERNEL);
1142         if (!local) {
1143                 status = -ENOMEM;
1144                 goto bail;
1145         }
1146
1147         status = -EINVAL;
1148
1149         spin_lock(&dlm_domain_lock);
1150         dlm = __dlm_lookup_domain_full(qr->qr_domain, qr->qr_namelen);
1151         if (!dlm) {
1152                 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1153                      "before join domain\n", qr->qr_node, qr->qr_domain);
1154                 goto bail;
1155         }
1156
1157         spin_lock(&dlm->spinlock);
1158         locked = 1;
1159         if (dlm->joining_node != qr->qr_node) {
1160                 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1161                      "but joining node is %d\n", qr->qr_node, qr->qr_domain,
1162                      dlm->joining_node);
1163                 goto bail;
1164         }
1165
1166         /* Support for global heartbeat was added in 1.1 */
1167         if (dlm->dlm_locking_proto.pv_major == 1 &&
1168             dlm->dlm_locking_proto.pv_minor == 0) {
1169                 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1170                      "but active dlm protocol is %d.%d\n", qr->qr_node,
1171                      qr->qr_domain, dlm->dlm_locking_proto.pv_major,
1172                      dlm->dlm_locking_proto.pv_minor);
1173                 goto bail;
1174         }
1175
1176         status = dlm_match_regions(dlm, qr, local, sizeof(qr->qr_regions));
1177
1178 bail:
1179         if (locked)
1180                 spin_unlock(&dlm->spinlock);
1181         spin_unlock(&dlm_domain_lock);
1182
1183         kfree(local);
1184
1185         return status;
1186 }
1187
1188 static int dlm_match_nodes(struct dlm_ctxt *dlm, struct dlm_query_nodeinfo *qn)
1189 {
1190         struct o2nm_node *local;
1191         struct dlm_node_info *remote;
1192         int i, j;
1193         int status = 0;
1194
1195         for (j = 0; j < qn->qn_numnodes; ++j)
1196                 mlog(0, "Node %3d, %pI4:%u\n", qn->qn_nodes[j].ni_nodenum,
1197                      &(qn->qn_nodes[j].ni_ipv4_address),
1198                      ntohs(qn->qn_nodes[j].ni_ipv4_port));
1199
1200         for (i = 0; i < O2NM_MAX_NODES && !status; ++i) {
1201                 local = o2nm_get_node_by_num(i);
1202                 remote = NULL;
1203                 for (j = 0; j < qn->qn_numnodes; ++j) {
1204                         if (qn->qn_nodes[j].ni_nodenum == i) {
1205                                 remote = &(qn->qn_nodes[j]);
1206                                 break;
1207                         }
1208                 }
1209
1210                 if (!local && !remote)
1211                         continue;
1212
1213                 if ((local && !remote) || (!local && remote))
1214                         status = -EINVAL;
1215
1216                 if (!status &&
1217                     ((remote->ni_nodenum != local->nd_num) ||
1218                      (remote->ni_ipv4_port != local->nd_ipv4_port) ||
1219                      (remote->ni_ipv4_address != local->nd_ipv4_address)))
1220                         status = -EINVAL;
1221
1222                 if (status) {
1223                         if (remote && !local)
1224                                 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1225                                      "registered in joining node %d but not in "
1226                                      "local node %d\n", qn->qn_domain,
1227                                      remote->ni_nodenum,
1228                                      &(remote->ni_ipv4_address),
1229                                      ntohs(remote->ni_ipv4_port),
1230                                      qn->qn_nodenum, dlm->node_num);
1231                         if (local && !remote)
1232                                 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1233                                      "registered in local node %d but not in "
1234                                      "joining node %d\n", qn->qn_domain,
1235                                      local->nd_num, &(local->nd_ipv4_address),
1236                                      ntohs(local->nd_ipv4_port),
1237                                      dlm->node_num, qn->qn_nodenum);
1238                         BUG_ON((!local && !remote));
1239                 }
1240
1241                 if (local)
1242                         o2nm_node_put(local);
1243         }
1244
1245         return status;
1246 }
1247
1248 static int dlm_send_nodeinfo(struct dlm_ctxt *dlm, unsigned long *node_map)
1249 {
1250         struct dlm_query_nodeinfo *qn = NULL;
1251         struct o2nm_node *node;
1252         int ret = 0, status, count, i;
1253
1254         if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1255                 goto bail;
1256
1257         qn = kzalloc(sizeof(struct dlm_query_nodeinfo), GFP_KERNEL);
1258         if (!qn) {
1259                 ret = -ENOMEM;
1260                 mlog_errno(ret);
1261                 goto bail;
1262         }
1263
1264         for (i = 0, count = 0; i < O2NM_MAX_NODES; ++i) {
1265                 node = o2nm_get_node_by_num(i);
1266                 if (!node)
1267                         continue;
1268                 qn->qn_nodes[count].ni_nodenum = node->nd_num;
1269                 qn->qn_nodes[count].ni_ipv4_port = node->nd_ipv4_port;
1270                 qn->qn_nodes[count].ni_ipv4_address = node->nd_ipv4_address;
1271                 mlog(0, "Node %3d, %pI4:%u\n", node->nd_num,
1272                      &(node->nd_ipv4_address), ntohs(node->nd_ipv4_port));
1273                 ++count;
1274                 o2nm_node_put(node);
1275         }
1276
1277         qn->qn_nodenum = dlm->node_num;
1278         qn->qn_numnodes = count;
1279         qn->qn_namelen = strlen(dlm->name);
1280         memcpy(qn->qn_domain, dlm->name, qn->qn_namelen);
1281
1282         i = -1;
1283         while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1284                                   i + 1)) < O2NM_MAX_NODES) {
1285                 if (i == dlm->node_num)
1286                         continue;
1287
1288                 mlog(0, "Sending nodeinfo to node %d\n", i);
1289
1290                 ret = o2net_send_message(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
1291                                          qn, sizeof(struct dlm_query_nodeinfo),
1292                                          i, &status);
1293                 if (ret >= 0)
1294                         ret = status;
1295                 if (ret) {
1296                         mlog(ML_ERROR, "node mismatch %d, node %d\n", ret, i);
1297                         break;
1298                 }
1299         }
1300
1301 bail:
1302         kfree(qn);
1303         return ret;
1304 }
1305
1306 static int dlm_query_nodeinfo_handler(struct o2net_msg *msg, u32 len,
1307                                       void *data, void **ret_data)
1308 {
1309         struct dlm_query_nodeinfo *qn;
1310         struct dlm_ctxt *dlm = NULL;
1311         int locked = 0, status = -EINVAL;
1312
1313         qn = (struct dlm_query_nodeinfo *) msg->buf;
1314
1315         mlog(0, "Node %u queries nodes on domain %s\n", qn->qn_nodenum,
1316              qn->qn_domain);
1317
1318         spin_lock(&dlm_domain_lock);
1319         dlm = __dlm_lookup_domain_full(qn->qn_domain, qn->qn_namelen);
1320         if (!dlm) {
1321                 mlog(ML_ERROR, "Node %d queried nodes on domain %s before "
1322                      "join domain\n", qn->qn_nodenum, qn->qn_domain);
1323                 goto bail;
1324         }
1325
1326         spin_lock(&dlm->spinlock);
1327         locked = 1;
1328         if (dlm->joining_node != qn->qn_nodenum) {
1329                 mlog(ML_ERROR, "Node %d queried nodes on domain %s but "
1330                      "joining node is %d\n", qn->qn_nodenum, qn->qn_domain,
1331                      dlm->joining_node);
1332                 goto bail;
1333         }
1334
1335         /* Support for node query was added in 1.1 */
1336         if (dlm->dlm_locking_proto.pv_major == 1 &&
1337             dlm->dlm_locking_proto.pv_minor == 0) {
1338                 mlog(ML_ERROR, "Node %d queried nodes on domain %s "
1339                      "but active dlm protocol is %d.%d\n", qn->qn_nodenum,
1340                      qn->qn_domain, dlm->dlm_locking_proto.pv_major,
1341                      dlm->dlm_locking_proto.pv_minor);
1342                 goto bail;
1343         }
1344
1345         status = dlm_match_nodes(dlm, qn);
1346
1347 bail:
1348         if (locked)
1349                 spin_unlock(&dlm->spinlock);
1350         spin_unlock(&dlm_domain_lock);
1351
1352         return status;
1353 }
1354
1355 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
1356                                    void **ret_data)
1357 {
1358         struct dlm_cancel_join *cancel;
1359         struct dlm_ctxt *dlm = NULL;
1360
1361         cancel = (struct dlm_cancel_join *) msg->buf;
1362
1363         mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
1364                   cancel->domain);
1365
1366         spin_lock(&dlm_domain_lock);
1367         dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
1368
1369         if (dlm) {
1370                 spin_lock(&dlm->spinlock);
1371
1372                 /* Yikes, this guy wants to cancel his join. No
1373                  * problem, we simply cleanup our join state. */
1374                 BUG_ON(dlm->joining_node != cancel->node_idx);
1375                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1376
1377                 spin_unlock(&dlm->spinlock);
1378         }
1379         spin_unlock(&dlm_domain_lock);
1380
1381         return 0;
1382 }
1383
1384 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
1385                                     unsigned int node)
1386 {
1387         int status;
1388         struct dlm_cancel_join cancel_msg;
1389
1390         memset(&cancel_msg, 0, sizeof(cancel_msg));
1391         cancel_msg.node_idx = dlm->node_num;
1392         cancel_msg.name_len = strlen(dlm->name);
1393         memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
1394
1395         status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1396                                     &cancel_msg, sizeof(cancel_msg), node,
1397                                     NULL);
1398         if (status < 0) {
1399                 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1400                      "node %u\n", status, DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1401                      node);
1402                 goto bail;
1403         }
1404
1405 bail:
1406         return status;
1407 }
1408
1409 /* map_size should be in bytes. */
1410 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
1411                                  unsigned long *node_map,
1412                                  unsigned int map_size)
1413 {
1414         int status, tmpstat;
1415         unsigned int node;
1416
1417         if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
1418                          sizeof(unsigned long))) {
1419                 mlog(ML_ERROR,
1420                      "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
1421                      map_size, (unsigned)BITS_TO_LONGS(O2NM_MAX_NODES));
1422                 return -EINVAL;
1423         }
1424
1425         status = 0;
1426         node = -1;
1427         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1428                                      node + 1)) < O2NM_MAX_NODES) {
1429                 if (node == dlm->node_num)
1430                         continue;
1431
1432                 tmpstat = dlm_send_one_join_cancel(dlm, node);
1433                 if (tmpstat) {
1434                         mlog(ML_ERROR, "Error return %d cancelling join on "
1435                              "node %d\n", tmpstat, node);
1436                         if (!status)
1437                                 status = tmpstat;
1438                 }
1439         }
1440
1441         if (status)
1442                 mlog_errno(status);
1443         return status;
1444 }
1445
1446 static int dlm_request_join(struct dlm_ctxt *dlm,
1447                             int node,
1448                             enum dlm_query_join_response_code *response)
1449 {
1450         int status;
1451         struct dlm_query_join_request join_msg;
1452         struct dlm_query_join_packet packet;
1453         u32 join_resp;
1454
1455         mlog(0, "querying node %d\n", node);
1456
1457         memset(&join_msg, 0, sizeof(join_msg));
1458         join_msg.node_idx = dlm->node_num;
1459         join_msg.name_len = strlen(dlm->name);
1460         memcpy(join_msg.domain, dlm->name, join_msg.name_len);
1461         join_msg.dlm_proto = dlm->dlm_locking_proto;
1462         join_msg.fs_proto = dlm->fs_locking_proto;
1463
1464         /* copy live node map to join message */
1465         byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES);
1466
1467         status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
1468                                     sizeof(join_msg), node, &join_resp);
1469         if (status < 0 && status != -ENOPROTOOPT) {
1470                 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1471                      "node %u\n", status, DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1472                      node);
1473                 goto bail;
1474         }
1475         dlm_query_join_wire_to_packet(join_resp, &packet);
1476
1477         /* -ENOPROTOOPT from the net code means the other side isn't
1478             listening for our message type -- that's fine, it means
1479             his dlm isn't up, so we can consider him a 'yes' but not
1480             joined into the domain.  */
1481         if (status == -ENOPROTOOPT) {
1482                 status = 0;
1483                 *response = JOIN_OK_NO_MAP;
1484         } else if (packet.code == JOIN_DISALLOW ||
1485                    packet.code == JOIN_OK_NO_MAP) {
1486                 *response = packet.code;
1487         } else if (packet.code == JOIN_PROTOCOL_MISMATCH) {
1488                 mlog(ML_NOTICE,
1489                      "This node requested DLM locking protocol %u.%u and "
1490                      "filesystem locking protocol %u.%u.  At least one of "
1491                      "the protocol versions on node %d is not compatible, "
1492                      "disconnecting\n",
1493                      dlm->dlm_locking_proto.pv_major,
1494                      dlm->dlm_locking_proto.pv_minor,
1495                      dlm->fs_locking_proto.pv_major,
1496                      dlm->fs_locking_proto.pv_minor,
1497                      node);
1498                 status = -EPROTO;
1499                 *response = packet.code;
1500         } else if (packet.code == JOIN_OK) {
1501                 *response = packet.code;
1502                 /* Use the same locking protocol as the remote node */
1503                 dlm->dlm_locking_proto.pv_minor = packet.dlm_minor;
1504                 dlm->fs_locking_proto.pv_minor = packet.fs_minor;
1505                 mlog(0,
1506                      "Node %d responds JOIN_OK with DLM locking protocol "
1507                      "%u.%u and fs locking protocol %u.%u\n",
1508                      node,
1509                      dlm->dlm_locking_proto.pv_major,
1510                      dlm->dlm_locking_proto.pv_minor,
1511                      dlm->fs_locking_proto.pv_major,
1512                      dlm->fs_locking_proto.pv_minor);
1513         } else {
1514                 status = -EINVAL;
1515                 mlog(ML_ERROR, "invalid response %d from node %u\n",
1516                      packet.code, node);
1517         }
1518
1519         mlog(0, "status %d, node %d response is %d\n", status, node,
1520              *response);
1521
1522 bail:
1523         return status;
1524 }
1525
1526 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
1527                                     unsigned int node)
1528 {
1529         int status;
1530         struct dlm_assert_joined assert_msg;
1531
1532         mlog(0, "Sending join assert to node %u\n", node);
1533
1534         memset(&assert_msg, 0, sizeof(assert_msg));
1535         assert_msg.node_idx = dlm->node_num;
1536         assert_msg.name_len = strlen(dlm->name);
1537         memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
1538
1539         status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1540                                     &assert_msg, sizeof(assert_msg), node,
1541                                     NULL);
1542         if (status < 0)
1543                 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1544                      "node %u\n", status, DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1545                      node);
1546
1547         return status;
1548 }
1549
1550 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
1551                                   unsigned long *node_map)
1552 {
1553         int status, node, live;
1554
1555         status = 0;
1556         node = -1;
1557         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1558                                      node + 1)) < O2NM_MAX_NODES) {
1559                 if (node == dlm->node_num)
1560                         continue;
1561
1562                 do {
1563                         /* It is very important that this message be
1564                          * received so we spin until either the node
1565                          * has died or it gets the message. */
1566                         status = dlm_send_one_join_assert(dlm, node);
1567
1568                         spin_lock(&dlm->spinlock);
1569                         live = test_bit(node, dlm->live_nodes_map);
1570                         spin_unlock(&dlm->spinlock);
1571
1572                         if (status) {
1573                                 mlog(ML_ERROR, "Error return %d asserting "
1574                                      "join on node %d\n", status, node);
1575
1576                                 /* give us some time between errors... */
1577                                 if (live)
1578                                         msleep(DLM_DOMAIN_BACKOFF_MS);
1579                         }
1580                 } while (status && live);
1581         }
1582 }
1583
1584 struct domain_join_ctxt {
1585         unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1586         unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1587 };
1588
1589 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
1590                                    struct domain_join_ctxt *ctxt,
1591                                    enum dlm_query_join_response_code response)
1592 {
1593         int ret;
1594
1595         if (response == JOIN_DISALLOW) {
1596                 mlog(0, "Latest response of disallow -- should restart\n");
1597                 return 1;
1598         }
1599
1600         spin_lock(&dlm->spinlock);
1601         /* For now, we restart the process if the node maps have
1602          * changed at all */
1603         ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
1604                      sizeof(dlm->live_nodes_map));
1605         spin_unlock(&dlm->spinlock);
1606
1607         if (ret)
1608                 mlog(0, "Node maps changed -- should restart\n");
1609
1610         return ret;
1611 }
1612
1613 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
1614 {
1615         int status = 0, tmpstat, node;
1616         struct domain_join_ctxt *ctxt;
1617         enum dlm_query_join_response_code response = JOIN_DISALLOW;
1618
1619         mlog(0, "%p", dlm);
1620
1621         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
1622         if (!ctxt) {
1623                 status = -ENOMEM;
1624                 mlog_errno(status);
1625                 goto bail;
1626         }
1627
1628         /* group sem locking should work for us here -- we're already
1629          * registered for heartbeat events so filling this should be
1630          * atomic wrt getting those handlers called. */
1631         o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
1632
1633         spin_lock(&dlm->spinlock);
1634         memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
1635
1636         __dlm_set_joining_node(dlm, dlm->node_num);
1637
1638         spin_unlock(&dlm->spinlock);
1639
1640         node = -1;
1641         while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
1642                                      node + 1)) < O2NM_MAX_NODES) {
1643                 if (node == dlm->node_num)
1644                         continue;
1645
1646                 status = dlm_request_join(dlm, node, &response);
1647                 if (status < 0) {
1648                         mlog_errno(status);
1649                         goto bail;
1650                 }
1651
1652                 /* Ok, either we got a response or the node doesn't have a
1653                  * dlm up. */
1654                 if (response == JOIN_OK)
1655                         set_bit(node, ctxt->yes_resp_map);
1656
1657                 if (dlm_should_restart_join(dlm, ctxt, response)) {
1658                         status = -EAGAIN;
1659                         goto bail;
1660                 }
1661         }
1662
1663         mlog(0, "Yay, done querying nodes!\n");
1664
1665         /* Yay, everyone agree's we can join the domain. My domain is
1666          * comprised of all nodes who were put in the
1667          * yes_resp_map. Copy that into our domain map and send a join
1668          * assert message to clean up everyone elses state. */
1669         spin_lock(&dlm->spinlock);
1670         memcpy(dlm->domain_map, ctxt->yes_resp_map,
1671                sizeof(ctxt->yes_resp_map));
1672         set_bit(dlm->node_num, dlm->domain_map);
1673         spin_unlock(&dlm->spinlock);
1674
1675         /* Support for global heartbeat and node info was added in 1.1 */
1676         if (dlm->dlm_locking_proto.pv_major > 1 ||
1677             dlm->dlm_locking_proto.pv_minor > 0) {
1678                 status = dlm_send_nodeinfo(dlm, ctxt->yes_resp_map);
1679                 if (status) {
1680                         mlog_errno(status);
1681                         goto bail;
1682                 }
1683                 status = dlm_send_regions(dlm, ctxt->yes_resp_map);
1684                 if (status) {
1685                         mlog_errno(status);
1686                         goto bail;
1687                 }
1688         }
1689
1690         dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1691
1692         /* Joined state *must* be set before the joining node
1693          * information, otherwise the query_join handler may read no
1694          * current joiner but a state of NEW and tell joining nodes
1695          * we're not in the domain. */
1696         spin_lock(&dlm_domain_lock);
1697         dlm->dlm_state = DLM_CTXT_JOINED;
1698         dlm->num_joins++;
1699         spin_unlock(&dlm_domain_lock);
1700
1701 bail:
1702         spin_lock(&dlm->spinlock);
1703         __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1704         if (!status)
1705                 __dlm_print_nodes(dlm);
1706         spin_unlock(&dlm->spinlock);
1707
1708         if (ctxt) {
1709                 /* Do we need to send a cancel message to any nodes? */
1710                 if (status < 0) {
1711                         tmpstat = dlm_send_join_cancels(dlm,
1712                                                         ctxt->yes_resp_map,
1713                                                         sizeof(ctxt->yes_resp_map));
1714                         if (tmpstat < 0)
1715                                 mlog_errno(tmpstat);
1716                 }
1717                 kfree(ctxt);
1718         }
1719
1720         mlog(0, "returning %d\n", status);
1721         return status;
1722 }
1723
1724 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1725 {
1726         o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_up);
1727         o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_down);
1728         o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1729 }
1730
1731 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1732 {
1733         int status;
1734
1735         mlog(0, "registering handlers.\n");
1736
1737         o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1738                             dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1739         status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_down);
1740         if (status)
1741                 goto bail;
1742
1743         o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1744                             dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1745         status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_up);
1746         if (status)
1747                 goto bail;
1748
1749         status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1750                                         sizeof(struct dlm_master_request),
1751                                         dlm_master_request_handler,
1752                                         dlm, NULL, &dlm->dlm_domain_handlers);
1753         if (status)
1754                 goto bail;
1755
1756         status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1757                                         sizeof(struct dlm_assert_master),
1758                                         dlm_assert_master_handler,
1759                                         dlm, dlm_assert_master_post_handler,
1760                                         &dlm->dlm_domain_handlers);
1761         if (status)
1762                 goto bail;
1763
1764         status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1765                                         sizeof(struct dlm_create_lock),
1766                                         dlm_create_lock_handler,
1767                                         dlm, NULL, &dlm->dlm_domain_handlers);
1768         if (status)
1769                 goto bail;
1770
1771         status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1772                                         DLM_CONVERT_LOCK_MAX_LEN,
1773                                         dlm_convert_lock_handler,
1774                                         dlm, NULL, &dlm->dlm_domain_handlers);
1775         if (status)
1776                 goto bail;
1777
1778         status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1779                                         DLM_UNLOCK_LOCK_MAX_LEN,
1780                                         dlm_unlock_lock_handler,
1781                                         dlm, NULL, &dlm->dlm_domain_handlers);
1782         if (status)
1783                 goto bail;
1784
1785         status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1786                                         DLM_PROXY_AST_MAX_LEN,
1787                                         dlm_proxy_ast_handler,
1788                                         dlm, NULL, &dlm->dlm_domain_handlers);
1789         if (status)
1790                 goto bail;
1791
1792         status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1793                                         sizeof(struct dlm_exit_domain),
1794                                         dlm_exit_domain_handler,
1795                                         dlm, NULL, &dlm->dlm_domain_handlers);
1796         if (status)
1797                 goto bail;
1798
1799         status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1800                                         sizeof(struct dlm_deref_lockres),
1801                                         dlm_deref_lockres_handler,
1802                                         dlm, NULL, &dlm->dlm_domain_handlers);
1803         if (status)
1804                 goto bail;
1805
1806         status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1807                                         sizeof(struct dlm_migrate_request),
1808                                         dlm_migrate_request_handler,
1809                                         dlm, NULL, &dlm->dlm_domain_handlers);
1810         if (status)
1811                 goto bail;
1812
1813         status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1814                                         DLM_MIG_LOCKRES_MAX_LEN,
1815                                         dlm_mig_lockres_handler,
1816                                         dlm, NULL, &dlm->dlm_domain_handlers);
1817         if (status)
1818                 goto bail;
1819
1820         status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1821                                         sizeof(struct dlm_master_requery),
1822                                         dlm_master_requery_handler,
1823                                         dlm, NULL, &dlm->dlm_domain_handlers);
1824         if (status)
1825                 goto bail;
1826
1827         status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1828                                         sizeof(struct dlm_lock_request),
1829                                         dlm_request_all_locks_handler,
1830                                         dlm, NULL, &dlm->dlm_domain_handlers);
1831         if (status)
1832                 goto bail;
1833
1834         status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1835                                         sizeof(struct dlm_reco_data_done),
1836                                         dlm_reco_data_done_handler,
1837                                         dlm, NULL, &dlm->dlm_domain_handlers);
1838         if (status)
1839                 goto bail;
1840
1841         status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1842                                         sizeof(struct dlm_begin_reco),
1843                                         dlm_begin_reco_handler,
1844                                         dlm, NULL, &dlm->dlm_domain_handlers);
1845         if (status)
1846                 goto bail;
1847
1848         status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1849                                         sizeof(struct dlm_finalize_reco),
1850                                         dlm_finalize_reco_handler,
1851                                         dlm, NULL, &dlm->dlm_domain_handlers);
1852         if (status)
1853                 goto bail;
1854
1855         status = o2net_register_handler(DLM_BEGIN_EXIT_DOMAIN_MSG, dlm->key,
1856                                         sizeof(struct dlm_exit_domain),
1857                                         dlm_begin_exit_domain_handler,
1858                                         dlm, NULL, &dlm->dlm_domain_handlers);
1859         if (status)
1860                 goto bail;
1861
1862 bail:
1863         if (status)
1864                 dlm_unregister_domain_handlers(dlm);
1865
1866         return status;
1867 }
1868
1869 static int dlm_join_domain(struct dlm_ctxt *dlm)
1870 {
1871         int status;
1872         unsigned int backoff;
1873         unsigned int total_backoff = 0;
1874
1875         BUG_ON(!dlm);
1876
1877         mlog(0, "Join domain %s\n", dlm->name);
1878
1879         status = dlm_register_domain_handlers(dlm);
1880         if (status) {
1881                 mlog_errno(status);
1882                 goto bail;
1883         }
1884
1885         status = dlm_debug_init(dlm);
1886         if (status < 0) {
1887                 mlog_errno(status);
1888                 goto bail;
1889         }
1890
1891         status = dlm_launch_thread(dlm);
1892         if (status < 0) {
1893                 mlog_errno(status);
1894                 goto bail;
1895         }
1896
1897         status = dlm_launch_recovery_thread(dlm);
1898         if (status < 0) {
1899                 mlog_errno(status);
1900                 goto bail;
1901         }
1902
1903         dlm->dlm_worker = create_singlethread_workqueue("dlm_wq");
1904         if (!dlm->dlm_worker) {
1905                 status = -ENOMEM;
1906                 mlog_errno(status);
1907                 goto bail;
1908         }
1909
1910         do {
1911                 status = dlm_try_to_join_domain(dlm);
1912
1913                 /* If we're racing another node to the join, then we
1914                  * need to back off temporarily and let them
1915                  * complete. */
1916 #define DLM_JOIN_TIMEOUT_MSECS  90000
1917                 if (status == -EAGAIN) {
1918                         if (signal_pending(current)) {
1919                                 status = -ERESTARTSYS;
1920                                 goto bail;
1921                         }
1922
1923                         if (total_backoff >
1924                             msecs_to_jiffies(DLM_JOIN_TIMEOUT_MSECS)) {
1925                                 status = -ERESTARTSYS;
1926                                 mlog(ML_NOTICE, "Timed out joining dlm domain "
1927                                      "%s after %u msecs\n", dlm->name,
1928                                      jiffies_to_msecs(total_backoff));
1929                                 goto bail;
1930                         }
1931
1932                         /*
1933                          * <chip> After you!
1934                          * <dale> No, after you!
1935                          * <chip> I insist!
1936                          * <dale> But you first!
1937                          * ...
1938                          */
1939                         backoff = (unsigned int)(jiffies & 0x3);
1940                         backoff *= DLM_DOMAIN_BACKOFF_MS;
1941                         total_backoff += backoff;
1942                         mlog(0, "backoff %d\n", backoff);
1943                         msleep(backoff);
1944                 }
1945         } while (status == -EAGAIN);
1946
1947         if (status < 0) {
1948                 mlog_errno(status);
1949                 goto bail;
1950         }
1951
1952         status = 0;
1953 bail:
1954         wake_up(&dlm_domain_events);
1955
1956         if (status) {
1957                 dlm_unregister_domain_handlers(dlm);
1958                 dlm_debug_shutdown(dlm);
1959                 dlm_complete_thread(dlm);
1960                 dlm_complete_recovery_thread(dlm);
1961                 dlm_destroy_dlm_worker(dlm);
1962         }
1963
1964         return status;
1965 }
1966
1967 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1968                                 u32 key)
1969 {
1970         int i;
1971         int ret;
1972         struct dlm_ctxt *dlm = NULL;
1973
1974         dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1975         if (!dlm) {
1976                 mlog_errno(-ENOMEM);
1977                 goto leave;
1978         }
1979
1980         dlm->name = kstrdup(domain, GFP_KERNEL);
1981         if (dlm->name == NULL) {
1982                 mlog_errno(-ENOMEM);
1983                 kfree(dlm);
1984                 dlm = NULL;
1985                 goto leave;
1986         }
1987
1988         dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1989         if (!dlm->lockres_hash) {
1990                 mlog_errno(-ENOMEM);
1991                 kfree(dlm->name);
1992                 kfree(dlm);
1993                 dlm = NULL;
1994                 goto leave;
1995         }
1996
1997         for (i = 0; i < DLM_HASH_BUCKETS; i++)
1998                 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1999
2000         dlm->master_hash = (struct hlist_head **)
2001                                 dlm_alloc_pagevec(DLM_HASH_PAGES);
2002         if (!dlm->master_hash) {
2003                 mlog_errno(-ENOMEM);
2004                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
2005                 kfree(dlm->name);
2006                 kfree(dlm);
2007                 dlm = NULL;
2008                 goto leave;
2009         }
2010
2011         for (i = 0; i < DLM_HASH_BUCKETS; i++)
2012                 INIT_HLIST_HEAD(dlm_master_hash(dlm, i));
2013
2014         dlm->key = key;
2015         dlm->node_num = o2nm_this_node();
2016
2017         ret = dlm_create_debugfs_subroot(dlm);
2018         if (ret < 0) {
2019                 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
2020                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
2021                 kfree(dlm->name);
2022                 kfree(dlm);
2023                 dlm = NULL;
2024                 goto leave;
2025         }
2026
2027         spin_lock_init(&dlm->spinlock);
2028         spin_lock_init(&dlm->master_lock);
2029         spin_lock_init(&dlm->ast_lock);
2030         spin_lock_init(&dlm->track_lock);
2031         INIT_LIST_HEAD(&dlm->list);
2032         INIT_LIST_HEAD(&dlm->dirty_list);
2033         INIT_LIST_HEAD(&dlm->reco.resources);
2034         INIT_LIST_HEAD(&dlm->reco.received);
2035         INIT_LIST_HEAD(&dlm->reco.node_data);
2036         INIT_LIST_HEAD(&dlm->purge_list);
2037         INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
2038         INIT_LIST_HEAD(&dlm->tracking_list);
2039         dlm->reco.state = 0;
2040
2041         INIT_LIST_HEAD(&dlm->pending_asts);
2042         INIT_LIST_HEAD(&dlm->pending_basts);
2043
2044         mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
2045                   dlm->recovery_map, &(dlm->recovery_map[0]));
2046
2047         memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
2048         memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
2049         memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
2050
2051         dlm->dlm_thread_task = NULL;
2052         dlm->dlm_reco_thread_task = NULL;
2053         dlm->dlm_worker = NULL;
2054         init_waitqueue_head(&dlm->dlm_thread_wq);
2055         init_waitqueue_head(&dlm->dlm_reco_thread_wq);
2056         init_waitqueue_head(&dlm->reco.event);
2057         init_waitqueue_head(&dlm->ast_wq);
2058         init_waitqueue_head(&dlm->migration_wq);
2059         INIT_LIST_HEAD(&dlm->mle_hb_events);
2060
2061         dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
2062         init_waitqueue_head(&dlm->dlm_join_events);
2063
2064         dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
2065         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
2066
2067         atomic_set(&dlm->res_tot_count, 0);
2068         atomic_set(&dlm->res_cur_count, 0);
2069         for (i = 0; i < DLM_MLE_NUM_TYPES; ++i) {
2070                 atomic_set(&dlm->mle_tot_count[i], 0);
2071                 atomic_set(&dlm->mle_cur_count[i], 0);
2072         }
2073
2074         spin_lock_init(&dlm->work_lock);
2075         INIT_LIST_HEAD(&dlm->work_list);
2076         INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
2077
2078         kref_init(&dlm->dlm_refs);
2079         dlm->dlm_state = DLM_CTXT_NEW;
2080
2081         INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
2082
2083         mlog(0, "context init: refcount %u\n",
2084                   atomic_read(&dlm->dlm_refs.refcount));
2085
2086 leave:
2087         return dlm;
2088 }
2089
2090 /*
2091  * Compare a requested locking protocol version against the current one.
2092  *
2093  * If the major numbers are different, they are incompatible.
2094  * If the current minor is greater than the request, they are incompatible.
2095  * If the current minor is less than or equal to the request, they are
2096  * compatible, and the requester should run at the current minor version.
2097  */
2098 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
2099                                 struct dlm_protocol_version *request)
2100 {
2101         if (existing->pv_major != request->pv_major)
2102                 return 1;
2103
2104         if (existing->pv_minor > request->pv_minor)
2105                 return 1;
2106
2107         if (existing->pv_minor < request->pv_minor)
2108                 request->pv_minor = existing->pv_minor;
2109
2110         return 0;
2111 }
2112
2113 /*
2114  * dlm_register_domain: one-time setup per "domain".
2115  *
2116  * The filesystem passes in the requested locking version via proto.
2117  * If registration was successful, proto will contain the negotiated
2118  * locking protocol.
2119  */
2120 struct dlm_ctxt * dlm_register_domain(const char *domain,
2121                                u32 key,
2122                                struct dlm_protocol_version *fs_proto)
2123 {
2124         int ret;
2125         struct dlm_ctxt *dlm = NULL;
2126         struct dlm_ctxt *new_ctxt = NULL;
2127
2128         if (strlen(domain) >= O2NM_MAX_NAME_LEN) {
2129                 ret = -ENAMETOOLONG;
2130                 mlog(ML_ERROR, "domain name length too long\n");
2131                 goto leave;
2132         }
2133
2134         if (!o2hb_check_local_node_heartbeating()) {
2135                 mlog(ML_ERROR, "the local node has not been configured, or is "
2136                      "not heartbeating\n");
2137                 ret = -EPROTO;
2138                 goto leave;
2139         }
2140
2141         mlog(0, "register called for domain \"%s\"\n", domain);
2142
2143 retry:
2144         dlm = NULL;
2145         if (signal_pending(current)) {
2146                 ret = -ERESTARTSYS;
2147                 mlog_errno(ret);
2148                 goto leave;
2149         }
2150
2151         spin_lock(&dlm_domain_lock);
2152
2153         dlm = __dlm_lookup_domain(domain);
2154         if (dlm) {
2155                 if (dlm->dlm_state != DLM_CTXT_JOINED) {
2156                         spin_unlock(&dlm_domain_lock);
2157
2158                         mlog(0, "This ctxt is not joined yet!\n");
2159                         wait_event_interruptible(dlm_domain_events,
2160                                                  dlm_wait_on_domain_helper(
2161                                                          domain));
2162                         goto retry;
2163                 }
2164
2165                 if (dlm_protocol_compare(&dlm->fs_locking_proto, fs_proto)) {
2166                         spin_unlock(&dlm_domain_lock);
2167                         mlog(ML_ERROR,
2168                              "Requested locking protocol version is not "
2169                              "compatible with already registered domain "
2170                              "\"%s\"\n", domain);
2171                         ret = -EPROTO;
2172                         goto leave;
2173                 }
2174
2175                 __dlm_get(dlm);
2176                 dlm->num_joins++;
2177
2178                 spin_unlock(&dlm_domain_lock);
2179
2180                 ret = 0;
2181                 goto leave;
2182         }
2183
2184         /* doesn't exist */
2185         if (!new_ctxt) {
2186                 spin_unlock(&dlm_domain_lock);
2187
2188                 new_ctxt = dlm_alloc_ctxt(domain, key);
2189                 if (new_ctxt)
2190                         goto retry;
2191
2192                 ret = -ENOMEM;
2193                 mlog_errno(ret);
2194                 goto leave;
2195         }
2196
2197         /* a little variable switch-a-roo here... */
2198         dlm = new_ctxt;
2199         new_ctxt = NULL;
2200
2201         /* add the new domain */
2202         list_add_tail(&dlm->list, &dlm_domains);
2203         spin_unlock(&dlm_domain_lock);
2204
2205         /*
2206          * Pass the locking protocol version into the join.  If the join
2207          * succeeds, it will have the negotiated protocol set.
2208          */
2209         dlm->dlm_locking_proto = dlm_protocol;
2210         dlm->fs_locking_proto = *fs_proto;
2211
2212         ret = dlm_join_domain(dlm);
2213         if (ret) {
2214                 mlog_errno(ret);
2215                 dlm_put(dlm);
2216                 goto leave;
2217         }
2218
2219         /* Tell the caller what locking protocol we negotiated */
2220         *fs_proto = dlm->fs_locking_proto;
2221
2222         ret = 0;
2223 leave:
2224         if (new_ctxt)
2225                 dlm_free_ctxt_mem(new_ctxt);
2226
2227         if (ret < 0)
2228                 dlm = ERR_PTR(ret);
2229
2230         return dlm;
2231 }
2232 EXPORT_SYMBOL_GPL(dlm_register_domain);
2233
2234 static LIST_HEAD(dlm_join_handlers);
2235
2236 static void dlm_unregister_net_handlers(void)
2237 {
2238         o2net_unregister_handler_list(&dlm_join_handlers);
2239 }
2240
2241 static int dlm_register_net_handlers(void)
2242 {
2243         int status = 0;
2244
2245         status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
2246                                         sizeof(struct dlm_query_join_request),
2247                                         dlm_query_join_handler,
2248                                         NULL, NULL, &dlm_join_handlers);
2249         if (status)
2250                 goto bail;
2251
2252         status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
2253                                         sizeof(struct dlm_assert_joined),
2254                                         dlm_assert_joined_handler,
2255                                         NULL, NULL, &dlm_join_handlers);
2256         if (status)
2257                 goto bail;
2258
2259         status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
2260                                         sizeof(struct dlm_cancel_join),
2261                                         dlm_cancel_join_handler,
2262                                         NULL, NULL, &dlm_join_handlers);
2263         if (status)
2264                 goto bail;
2265
2266         status = o2net_register_handler(DLM_QUERY_REGION, DLM_MOD_KEY,
2267                                         sizeof(struct dlm_query_region),
2268                                         dlm_query_region_handler,
2269                                         NULL, NULL, &dlm_join_handlers);
2270
2271         if (status)
2272                 goto bail;
2273
2274         status = o2net_register_handler(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
2275                                         sizeof(struct dlm_query_nodeinfo),
2276                                         dlm_query_nodeinfo_handler,
2277                                         NULL, NULL, &dlm_join_handlers);
2278 bail:
2279         if (status < 0)
2280                 dlm_unregister_net_handlers();
2281
2282         return status;
2283 }
2284
2285 /* Domain eviction callback handling.
2286  *
2287  * The file system requires notification of node death *before* the
2288  * dlm completes it's recovery work, otherwise it may be able to
2289  * acquire locks on resources requiring recovery. Since the dlm can
2290  * evict a node from it's domain *before* heartbeat fires, a similar
2291  * mechanism is required. */
2292
2293 /* Eviction is not expected to happen often, so a per-domain lock is
2294  * not necessary. Eviction callbacks are allowed to sleep for short
2295  * periods of time. */
2296 static DECLARE_RWSEM(dlm_callback_sem);
2297
2298 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
2299                                         int node_num)
2300 {
2301         struct list_head *iter;
2302         struct dlm_eviction_cb *cb;
2303
2304         down_read(&dlm_callback_sem);
2305         list_for_each(iter, &dlm->dlm_eviction_callbacks) {
2306                 cb = list_entry(iter, struct dlm_eviction_cb, ec_item);
2307
2308                 cb->ec_func(node_num, cb->ec_data);
2309         }
2310         up_read(&dlm_callback_sem);
2311 }
2312
2313 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
2314                            dlm_eviction_func *f,
2315                            void *data)
2316 {
2317         INIT_LIST_HEAD(&cb->ec_item);
2318         cb->ec_func = f;
2319         cb->ec_data = data;
2320 }
2321 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
2322
2323 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
2324                               struct dlm_eviction_cb *cb)
2325 {
2326         down_write(&dlm_callback_sem);
2327         list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
2328         up_write(&dlm_callback_sem);
2329 }
2330 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
2331
2332 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
2333 {
2334         down_write(&dlm_callback_sem);
2335         list_del_init(&cb->ec_item);
2336         up_write(&dlm_callback_sem);
2337 }
2338 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
2339
2340 static int __init dlm_init(void)
2341 {
2342         int status;
2343
2344         dlm_print_version();
2345
2346         status = dlm_init_mle_cache();
2347         if (status) {
2348                 mlog(ML_ERROR, "Could not create o2dlm_mle slabcache\n");
2349                 goto error;
2350         }
2351
2352         status = dlm_init_master_caches();
2353         if (status) {
2354                 mlog(ML_ERROR, "Could not create o2dlm_lockres and "
2355                      "o2dlm_lockname slabcaches\n");
2356                 goto error;
2357         }
2358
2359         status = dlm_init_lock_cache();
2360         if (status) {
2361                 mlog(ML_ERROR, "Count not create o2dlm_lock slabcache\n");
2362                 goto error;
2363         }
2364
2365         status = dlm_register_net_handlers();
2366         if (status) {
2367                 mlog(ML_ERROR, "Unable to register network handlers\n");
2368                 goto error;
2369         }
2370
2371         status = dlm_create_debugfs_root();
2372         if (status)
2373                 goto error;
2374
2375         return 0;
2376 error:
2377         dlm_unregister_net_handlers();
2378         dlm_destroy_lock_cache();
2379         dlm_destroy_master_caches();
2380         dlm_destroy_mle_cache();
2381         return -1;
2382 }
2383
2384 static void __exit dlm_exit (void)
2385 {
2386         dlm_destroy_debugfs_root();
2387         dlm_unregister_net_handlers();
2388         dlm_destroy_lock_cache();
2389         dlm_destroy_master_caches();
2390         dlm_destroy_mle_cache();
2391 }
2392
2393 MODULE_AUTHOR("Oracle");
2394 MODULE_LICENSE("GPL");
2395
2396 module_init(dlm_init);
2397 module_exit(dlm_exit);