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