git.oblomov.eu Git - linux-2.6/blob - 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 #define DLM_DOMAIN_BACKOFF_MS 200
 127
 128 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
 129                                   void **ret_data);
 130 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 131                                      void **ret_data);
 132 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
 133                                    void **ret_data);
 134 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 135                                    void **ret_data);
 136
 137 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
 138
 139 void __dlm_unhash_lockres(struct dlm_lock_resource *lockres)
 140 {
 141         hlist_del_init(&lockres->hash_node);
 142         dlm_lockres_put(lockres);
 143 }
 144
 145 void __dlm_insert_lockres(struct dlm_ctxt *dlm,
 146                        struct dlm_lock_resource *res)
 147 {
 148         struct hlist_head *bucket;
 149         struct qstr *q;
 150
 151         assert_spin_locked(&dlm->spinlock);
 152
 153         q = &res->lockname;
 154         bucket = dlm_lockres_hash(dlm, q->hash);
 155
 156         /* get a reference for our hashtable */
 157         dlm_lockres_get(res);
 158
 159         hlist_add_head(&res->hash_node, bucket);
 160 }
 161
 162 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
 163                                                      const char *name,
 164                                                      unsigned int len,
 165                                                      unsigned int hash)
 166 {
 167         struct hlist_head *bucket;
 168         struct hlist_node *list;
 169
 170         mlog_entry("%.*s\n", len, name);
 171
 172         assert_spin_locked(&dlm->spinlock);
 173
 174         bucket = dlm_lockres_hash(dlm, hash);
 175
 176         hlist_for_each(list, bucket) {
 177                 struct dlm_lock_resource *res = hlist_entry(list,
 178                         struct dlm_lock_resource, hash_node);
 179                 if (res->lockname.name[0] != name[0])
 180                         continue;
 181                 if (unlikely(res->lockname.len != len))
 182                         continue;
 183                 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
 184                         continue;
 185                 dlm_lockres_get(res);
 186                 return res;
 187         }
 188         return NULL;
 189 }
 190
 191 /* intended to be called by functions which do not care about lock
 192  * resources which are being purged (most net _handler functions).
 193  * this will return NULL for any lock resource which is found but
 194  * currently in the process of dropping its mastery reference.
 195  * use __dlm_lookup_lockres_full when you need the lock resource
 196  * regardless (e.g. dlm_get_lock_resource) */
 197 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
 198                                                 const char *name,
 199                                                 unsigned int len,
 200                                                 unsigned int hash)
 201 {
 202         struct dlm_lock_resource *res = NULL;
 203
 204         mlog_entry("%.*s\n", len, name);
 205
 206         assert_spin_locked(&dlm->spinlock);
 207
 208         res = __dlm_lookup_lockres_full(dlm, name, len, hash);
 209         if (res) {
 210                 spin_lock(&res->spinlock);
 211                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
 212                         spin_unlock(&res->spinlock);
 213                         dlm_lockres_put(res);
 214                         return NULL;
 215                 }
 216                 spin_unlock(&res->spinlock);
 217         }
 218
 219         return res;
 220 }
 221
 222 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
 223                                     const char *name,
 224                                     unsigned int len)
 225 {
 226         struct dlm_lock_resource *res;
 227         unsigned int hash = dlm_lockid_hash(name, len);
 228
 229         spin_lock(&dlm->spinlock);
 230         res = __dlm_lookup_lockres(dlm, name, len, hash);
 231         spin_unlock(&dlm->spinlock);
 232         return res;
 233 }
 234
 235 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
 236 {
 237         struct dlm_ctxt *tmp = NULL;
 238         struct list_head *iter;
 239
 240         assert_spin_locked(&dlm_domain_lock);
 241
 242         /* tmp->name here is always NULL terminated,
 243          * but domain may not be! */
 244         list_for_each(iter, &dlm_domains) {
 245                 tmp = list_entry (iter, struct dlm_ctxt, list);
 246                 if (strlen(tmp->name) == len &&
 247                     memcmp(tmp->name, domain, len)==0)
 248                         break;
 249                 tmp = NULL;
 250         }
 251
 252         return tmp;
 253 }
 254
 255 /* For null terminated domain strings ONLY */
 256 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
 257 {
 258         assert_spin_locked(&dlm_domain_lock);
 259
 260         return __dlm_lookup_domain_full(domain, strlen(domain));
 261 }
 262
 263
 264 /* returns true on one of two conditions:
 265  * 1) the domain does not exist
 266  * 2) the domain exists and it's state is "joined" */
 267 static int dlm_wait_on_domain_helper(const char *domain)
 268 {
 269         int ret = 0;
 270         struct dlm_ctxt *tmp = NULL;
 271
 272         spin_lock(&dlm_domain_lock);
 273
 274         tmp = __dlm_lookup_domain(domain);
 275         if (!tmp)
 276                 ret = 1;
 277         else if (tmp->dlm_state == DLM_CTXT_JOINED)
 278                 ret = 1;
 279
 280         spin_unlock(&dlm_domain_lock);
 281         return ret;
 282 }
 283
 284 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
 285 {
 286         if (dlm->lockres_hash)
 287                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
 288
 289         if (dlm->name)
 290                 kfree(dlm->name);
 291
 292         kfree(dlm);
 293 }
 294
 295 /* A little strange - this function will be called while holding
 296  * dlm_domain_lock and is expected to be holding it on the way out. We
 297  * will however drop and reacquire it multiple times */
 298 static void dlm_ctxt_release(struct kref *kref)
 299 {
 300         struct dlm_ctxt *dlm;
 301
 302         dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
 303
 304         BUG_ON(dlm->num_joins);
 305         BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
 306
 307         /* we may still be in the list if we hit an error during join. */
 308         list_del_init(&dlm->list);
 309
 310         spin_unlock(&dlm_domain_lock);
 311
 312         mlog(0, "freeing memory from domain %s\n", dlm->name);
 313
 314         wake_up(&dlm_domain_events);
 315
 316         dlm_free_ctxt_mem(dlm);
 317
 318         spin_lock(&dlm_domain_lock);
 319 }
 320
 321 void dlm_put(struct dlm_ctxt *dlm)
 322 {
 323         spin_lock(&dlm_domain_lock);
 324         kref_put(&dlm->dlm_refs, dlm_ctxt_release);
 325         spin_unlock(&dlm_domain_lock);
 326 }
 327
 328 static void __dlm_get(struct dlm_ctxt *dlm)
 329 {
 330         kref_get(&dlm->dlm_refs);
 331 }
 332
 333 /* given a questionable reference to a dlm object, gets a reference if
 334  * it can find it in the list, otherwise returns NULL in which case
 335  * you shouldn't trust your pointer. */
 336 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
 337 {
 338         struct list_head *iter;
 339         struct dlm_ctxt *target = NULL;
 340
 341         spin_lock(&dlm_domain_lock);
 342
 343         list_for_each(iter, &dlm_domains) {
 344                 target = list_entry (iter, struct dlm_ctxt, list);
 345
 346                 if (target == dlm) {
 347                         __dlm_get(target);
 348                         break;
 349                 }
 350
 351                 target = NULL;
 352         }
 353
 354         spin_unlock(&dlm_domain_lock);
 355
 356         return target;
 357 }
 358
 359 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
 360 {
 361         int ret;
 362
 363         spin_lock(&dlm_domain_lock);
 364         ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
 365                 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
 366         spin_unlock(&dlm_domain_lock);
 367
 368         return ret;
 369 }
 370
 371 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
 372 {
 373         if (dlm->dlm_worker) {
 374                 flush_workqueue(dlm->dlm_worker);
 375                 destroy_workqueue(dlm->dlm_worker);
 376                 dlm->dlm_worker = NULL;
 377         }
 378 }
 379
 380 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
 381 {
 382         dlm_unregister_domain_handlers(dlm);
 383         dlm_complete_thread(dlm);
 384         dlm_complete_recovery_thread(dlm);
 385         dlm_destroy_dlm_worker(dlm);
 386
 387         /* We've left the domain. Now we can take ourselves out of the
 388          * list and allow the kref stuff to help us free the
 389          * memory. */
 390         spin_lock(&dlm_domain_lock);
 391         list_del_init(&dlm->list);
 392         spin_unlock(&dlm_domain_lock);
 393
 394         /* Wake up anyone waiting for us to remove this domain */
 395         wake_up(&dlm_domain_events);
 396 }
 397
 398 static int dlm_migrate_all_locks(struct dlm_ctxt *dlm)
 399 {
 400         int i, num, n, ret = 0;
 401         struct dlm_lock_resource *res;
 402         struct hlist_node *iter;
 403         struct hlist_head *bucket;
 404         int dropped;
 405
 406         mlog(0, "Migrating locks from domain %s\n", dlm->name);
 407
 408         num = 0;
 409         spin_lock(&dlm->spinlock);
 410         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
 411 redo_bucket:
 412                 n = 0;
 413                 bucket = dlm_lockres_hash(dlm, i);
 414                 iter = bucket->first;
 415                 while (iter) {
 416                         n++;
 417                         res = hlist_entry(iter, struct dlm_lock_resource,
 418                                           hash_node);
 419                         dlm_lockres_get(res);
 420                         /* migrate, if necessary.  this will drop the dlm
 421                          * spinlock and retake it if it does migration. */
 422                         dropped = dlm_empty_lockres(dlm, res);
 423
 424                         spin_lock(&res->spinlock);
 425                         __dlm_lockres_calc_usage(dlm, res);
 426                         iter = res->hash_node.next;
 427                         spin_unlock(&res->spinlock);
 428
 429                         dlm_lockres_put(res);
 430
 431                         cond_resched_lock(&dlm->spinlock);
 432
 433                         if (dropped)
 434                                 goto redo_bucket;
 435                 }
 436                 num += n;
 437                 mlog(0, "%s: touched %d lockreses in bucket %d "
 438                      "(tot=%d)\n", dlm->name, n, i, num);
 439         }
 440         spin_unlock(&dlm->spinlock);
 441         wake_up(&dlm->dlm_thread_wq);
 442
 443         /* let the dlm thread take care of purging, keep scanning until
 444          * nothing remains in the hash */
 445         if (num) {
 446                 mlog(0, "%s: %d lock resources in hash last pass\n",
 447                      dlm->name, num);
 448                 ret = -EAGAIN;
 449         }
 450         mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
 451         return ret;
 452 }
 453
 454 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
 455 {
 456         int ret;
 457
 458         spin_lock(&dlm->spinlock);
 459         ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
 460         spin_unlock(&dlm->spinlock);
 461
 462         return ret;
 463 }
 464
 465 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
 466 {
 467         /* Yikes, a double spinlock! I need domain_lock for the dlm
 468          * state and the dlm spinlock for join state... Sorry! */
 469 again:
 470         spin_lock(&dlm_domain_lock);
 471         spin_lock(&dlm->spinlock);
 472
 473         if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 474                 mlog(0, "Node %d is joining, we wait on it.\n",
 475                           dlm->joining_node);
 476                 spin_unlock(&dlm->spinlock);
 477                 spin_unlock(&dlm_domain_lock);
 478
 479                 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
 480                 goto again;
 481         }
 482
 483         dlm->dlm_state = DLM_CTXT_LEAVING;
 484         spin_unlock(&dlm->spinlock);
 485         spin_unlock(&dlm_domain_lock);
 486 }
 487
 488 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
 489 {
 490         int node = -1;
 491
 492         assert_spin_locked(&dlm->spinlock);
 493
 494         printk(KERN_INFO "ocfs2_dlm: Nodes in domain (\"%s\"): ", dlm->name);
 495
 496         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 497                                      node + 1)) < O2NM_MAX_NODES) {
 498                 printk("%d ", node);
 499         }
 500         printk("\n");
 501 }
 502
 503 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 504                                    void **ret_data)
 505 {
 506         struct dlm_ctxt *dlm = data;
 507         unsigned int node;
 508         struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
 509
 510         mlog_entry("%p %u %p", msg, len, data);
 511
 512         if (!dlm_grab(dlm))
 513                 return 0;
 514
 515         node = exit_msg->node_idx;
 516
 517         printk(KERN_INFO "ocfs2_dlm: Node %u leaves domain %s\n", node, dlm->name);
 518
 519         spin_lock(&dlm->spinlock);
 520         clear_bit(node, dlm->domain_map);
 521         __dlm_print_nodes(dlm);
 522
 523         /* notify anything attached to the heartbeat events */
 524         dlm_hb_event_notify_attached(dlm, node, 0);
 525
 526         spin_unlock(&dlm->spinlock);
 527
 528         dlm_put(dlm);
 529
 530         return 0;
 531 }
 532
 533 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm,
 534                                     unsigned int node)
 535 {
 536         int status;
 537         struct dlm_exit_domain leave_msg;
 538
 539         mlog(0, "Asking node %u if we can leave the domain %s me = %u\n",
 540                   node, dlm->name, dlm->node_num);
 541
 542         memset(&leave_msg, 0, sizeof(leave_msg));
 543         leave_msg.node_idx = dlm->node_num;
 544
 545         status = o2net_send_message(DLM_EXIT_DOMAIN_MSG, dlm->key,
 546                                     &leave_msg, sizeof(leave_msg), node,
 547                                     NULL);
 548
 549         mlog(0, "status return %d from o2net_send_message\n", status);
 550
 551         return status;
 552 }
 553
 554
 555 static void dlm_leave_domain(struct dlm_ctxt *dlm)
 556 {
 557         int node, clear_node, status;
 558
 559         /* At this point we've migrated away all our locks and won't
 560          * accept mastership of new ones. The dlm is responsible for
 561          * almost nothing now. We make sure not to confuse any joining
 562          * nodes and then commence shutdown procedure. */
 563
 564         spin_lock(&dlm->spinlock);
 565         /* Clear ourselves from the domain map */
 566         clear_bit(dlm->node_num, dlm->domain_map);
 567         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 568                                      0)) < O2NM_MAX_NODES) {
 569                 /* Drop the dlm spinlock. This is safe wrt the domain_map.
 570                  * -nodes cannot be added now as the
 571                  *   query_join_handlers knows to respond with OK_NO_MAP
 572                  * -we catch the right network errors if a node is
 573                  *   removed from the map while we're sending him the
 574                  *   exit message. */
 575                 spin_unlock(&dlm->spinlock);
 576
 577                 clear_node = 1;
 578
 579                 status = dlm_send_one_domain_exit(dlm, node);
 580                 if (status < 0 &&
 581                     status != -ENOPROTOOPT &&
 582                     status != -ENOTCONN) {
 583                         mlog(ML_NOTICE, "Error %d sending domain exit message "
 584                              "to node %d\n", status, node);
 585
 586                         /* Not sure what to do here but lets sleep for
 587                          * a bit in case this was a transient
 588                          * error... */
 589                         msleep(DLM_DOMAIN_BACKOFF_MS);
 590                         clear_node = 0;
 591                 }
 592
 593                 spin_lock(&dlm->spinlock);
 594                 /* If we're not clearing the node bit then we intend
 595                  * to loop back around to try again. */
 596                 if (clear_node)
 597                         clear_bit(node, dlm->domain_map);
 598         }
 599         spin_unlock(&dlm->spinlock);
 600 }
 601
 602 int dlm_joined(struct dlm_ctxt *dlm)
 603 {
 604         int ret = 0;
 605
 606         spin_lock(&dlm_domain_lock);
 607
 608         if (dlm->dlm_state == DLM_CTXT_JOINED)
 609                 ret = 1;
 610
 611         spin_unlock(&dlm_domain_lock);
 612
 613         return ret;
 614 }
 615
 616 int dlm_shutting_down(struct dlm_ctxt *dlm)
 617 {
 618         int ret = 0;
 619
 620         spin_lock(&dlm_domain_lock);
 621
 622         if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
 623                 ret = 1;
 624
 625         spin_unlock(&dlm_domain_lock);
 626
 627         return ret;
 628 }
 629
 630 void dlm_unregister_domain(struct dlm_ctxt *dlm)
 631 {
 632         int leave = 0;
 633
 634         spin_lock(&dlm_domain_lock);
 635         BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
 636         BUG_ON(!dlm->num_joins);
 637
 638         dlm->num_joins--;
 639         if (!dlm->num_joins) {
 640                 /* We mark it "in shutdown" now so new register
 641                  * requests wait until we've completely left the
 642                  * domain. Don't use DLM_CTXT_LEAVING yet as we still
 643                  * want new domain joins to communicate with us at
 644                  * least until we've completed migration of our
 645                  * resources. */
 646                 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
 647                 leave = 1;
 648         }
 649         spin_unlock(&dlm_domain_lock);
 650
 651         if (leave) {
 652                 mlog(0, "shutting down domain %s\n", dlm->name);
 653
 654                 /* We changed dlm state, notify the thread */
 655                 dlm_kick_thread(dlm, NULL);
 656
 657                 while (dlm_migrate_all_locks(dlm)) {
 658                         mlog(0, "%s: more migration to do\n", dlm->name);
 659                 }
 660                 dlm_mark_domain_leaving(dlm);
 661                 dlm_leave_domain(dlm);
 662                 dlm_complete_dlm_shutdown(dlm);
 663         }
 664         dlm_put(dlm);
 665 }
 666 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
 667
 668 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
 669                                   void **ret_data)
 670 {
 671         struct dlm_query_join_request *query;
 672         enum dlm_query_join_response response;
 673         struct dlm_ctxt *dlm = NULL;
 674         u8 nodenum;
 675
 676         query = (struct dlm_query_join_request *) msg->buf;
 677
 678         mlog(0, "node %u wants to join domain %s\n", query->node_idx,
 679                   query->domain);
 680
 681         /*
 682          * If heartbeat doesn't consider the node live, tell it
 683          * to back off and try again.  This gives heartbeat a chance
 684          * to catch up.
 685          */
 686         if (!o2hb_check_node_heartbeating(query->node_idx)) {
 687                 mlog(0, "node %u is not in our live map yet\n",
 688                      query->node_idx);
 689
 690                 response = JOIN_DISALLOW;
 691                 goto respond;
 692         }
 693
 694         response = JOIN_OK_NO_MAP;
 695
 696         spin_lock(&dlm_domain_lock);
 697         dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
 698         if (!dlm)
 699                 goto unlock_respond;
 700
 701         /*
 702          * There is a small window where the joining node may not see the
 703          * node(s) that just left but still part of the cluster. DISALLOW
 704          * join request if joining node has different node map.
 705          */
 706         nodenum=0;
 707         while (nodenum < O2NM_MAX_NODES) {
 708                 if (test_bit(nodenum, dlm->domain_map)) {
 709                         if (!byte_test_bit(nodenum, query->node_map)) {
 710                                 mlog(0, "disallow join as node %u does not "
 711                                      "have node %u in its nodemap\n",
 712                                      query->node_idx, nodenum);
 713                                 response = JOIN_DISALLOW;
 714                                 goto unlock_respond;
 715                         }
 716                 }
 717                 nodenum++;
 718         }
 719
 720         /* Once the dlm ctxt is marked as leaving then we don't want
 721          * to be put in someone's domain map.
 722          * Also, explicitly disallow joining at certain troublesome
 723          * times (ie. during recovery). */
 724         if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
 725                 int bit = query->node_idx;
 726                 spin_lock(&dlm->spinlock);
 727
 728                 if (dlm->dlm_state == DLM_CTXT_NEW &&
 729                     dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
 730                         /*If this is a brand new context and we
 731                          * haven't started our join process yet, then
 732                          * the other node won the race. */
 733                         response = JOIN_OK_NO_MAP;
 734                 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 735                         /* Disallow parallel joins. */
 736                         response = JOIN_DISALLOW;
 737                 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
 738                         mlog(0, "node %u trying to join, but recovery "
 739                              "is ongoing.\n", bit);
 740                         response = JOIN_DISALLOW;
 741                 } else if (test_bit(bit, dlm->recovery_map)) {
 742                         mlog(0, "node %u trying to join, but it "
 743                              "still needs recovery.\n", bit);
 744                         response = JOIN_DISALLOW;
 745                 } else if (test_bit(bit, dlm->domain_map)) {
 746                         mlog(0, "node %u trying to join, but it "
 747                              "is still in the domain! needs recovery?\n",
 748                              bit);
 749                         response = JOIN_DISALLOW;
 750                 } else {
 751                         /* Alright we're fully a part of this domain
 752                          * so we keep some state as to who's joining
 753                          * and indicate to him that needs to be fixed
 754                          * up. */
 755                         response = JOIN_OK;
 756                         __dlm_set_joining_node(dlm, query->node_idx);
 757                 }
 758
 759                 spin_unlock(&dlm->spinlock);
 760         }
 761 unlock_respond:
 762         spin_unlock(&dlm_domain_lock);
 763
 764 respond:
 765         mlog(0, "We respond with %u\n", response);
 766
 767         return response;
 768 }
 769
 770 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 771                                      void **ret_data)
 772 {
 773         struct dlm_assert_joined *assert;
 774         struct dlm_ctxt *dlm = NULL;
 775
 776         assert = (struct dlm_assert_joined *) msg->buf;
 777
 778         mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
 779                   assert->domain);
 780
 781         spin_lock(&dlm_domain_lock);
 782         dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
 783         /* XXX should we consider no dlm ctxt an error? */
 784         if (dlm) {
 785                 spin_lock(&dlm->spinlock);
 786
 787                 /* Alright, this node has officially joined our
 788                  * domain. Set him in the map and clean up our
 789                  * leftover join state. */
 790                 BUG_ON(dlm->joining_node != assert->node_idx);
 791                 set_bit(assert->node_idx, dlm->domain_map);
 792                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 793
 794                 printk(KERN_INFO "ocfs2_dlm: Node %u joins domain %s\n",
 795                        assert->node_idx, dlm->name);
 796                 __dlm_print_nodes(dlm);
 797
 798                 /* notify anything attached to the heartbeat events */
 799                 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
 800
 801                 spin_unlock(&dlm->spinlock);
 802         }
 803         spin_unlock(&dlm_domain_lock);
 804
 805         return 0;
 806 }
 807
 808 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
 809                                    void **ret_data)
 810 {
 811         struct dlm_cancel_join *cancel;
 812         struct dlm_ctxt *dlm = NULL;
 813
 814         cancel = (struct dlm_cancel_join *) msg->buf;
 815
 816         mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
 817                   cancel->domain);
 818
 819         spin_lock(&dlm_domain_lock);
 820         dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
 821
 822         if (dlm) {
 823                 spin_lock(&dlm->spinlock);
 824
 825                 /* Yikes, this guy wants to cancel his join. No
 826                  * problem, we simply cleanup our join state. */
 827                 BUG_ON(dlm->joining_node != cancel->node_idx);
 828                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 829
 830                 spin_unlock(&dlm->spinlock);
 831         }
 832         spin_unlock(&dlm_domain_lock);
 833
 834         return 0;
 835 }
 836
 837 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
 838                                     unsigned int node)
 839 {
 840         int status;
 841         struct dlm_cancel_join cancel_msg;
 842
 843         memset(&cancel_msg, 0, sizeof(cancel_msg));
 844         cancel_msg.node_idx = dlm->node_num;
 845         cancel_msg.name_len = strlen(dlm->name);
 846         memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
 847
 848         status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
 849                                     &cancel_msg, sizeof(cancel_msg), node,
 850                                     NULL);
 851         if (status < 0) {
 852                 mlog_errno(status);
 853                 goto bail;
 854         }
 855
 856 bail:
 857         return status;
 858 }
 859
 860 /* map_size should be in bytes. */
 861 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
 862                                  unsigned long *node_map,
 863                                  unsigned int map_size)
 864 {
 865         int status, tmpstat;
 866         unsigned int node;
 867
 868         if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
 869                          sizeof(unsigned long))) {
 870                 mlog(ML_ERROR,
 871                      "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
 872                      map_size, BITS_TO_LONGS(O2NM_MAX_NODES));
 873                 return -EINVAL;
 874         }
 875
 876         status = 0;
 877         node = -1;
 878         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
 879                                      node + 1)) < O2NM_MAX_NODES) {
 880                 if (node == dlm->node_num)
 881                         continue;
 882
 883                 tmpstat = dlm_send_one_join_cancel(dlm, node);
 884                 if (tmpstat) {
 885                         mlog(ML_ERROR, "Error return %d cancelling join on "
 886                              "node %d\n", tmpstat, node);
 887                         if (!status)
 888                                 status = tmpstat;
 889                 }
 890         }
 891
 892         if (status)
 893                 mlog_errno(status);
 894         return status;
 895 }
 896
 897 static int dlm_request_join(struct dlm_ctxt *dlm,
 898                             int node,
 899                             enum dlm_query_join_response *response)
 900 {
 901         int status, retval;
 902         struct dlm_query_join_request join_msg;
 903
 904         mlog(0, "querying node %d\n", node);
 905
 906         memset(&join_msg, 0, sizeof(join_msg));
 907         join_msg.node_idx = dlm->node_num;
 908         join_msg.name_len = strlen(dlm->name);
 909         memcpy(join_msg.domain, dlm->name, join_msg.name_len);
 910
 911         /* copy live node map to join message */
 912         byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES);
 913
 914         status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
 915                                     sizeof(join_msg), node, &retval);
 916         if (status < 0 && status != -ENOPROTOOPT) {
 917                 mlog_errno(status);
 918                 goto bail;
 919         }
 920
 921         /* -ENOPROTOOPT from the net code means the other side isn't
 922             listening for our message type -- that's fine, it means
 923             his dlm isn't up, so we can consider him a 'yes' but not
 924             joined into the domain.  */
 925         if (status == -ENOPROTOOPT) {
 926                 status = 0;
 927                 *response = JOIN_OK_NO_MAP;
 928         } else if (retval == JOIN_DISALLOW ||
 929                    retval == JOIN_OK ||
 930                    retval == JOIN_OK_NO_MAP) {
 931                 *response = retval;
 932         } else {
 933                 status = -EINVAL;
 934                 mlog(ML_ERROR, "invalid response %d from node %u\n", retval,
 935                      node);
 936         }
 937
 938         mlog(0, "status %d, node %d response is %d\n", status, node,
 939                   *response);
 940
 941 bail:
 942         return status;
 943 }
 944
 945 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
 946                                     unsigned int node)
 947 {
 948         int status;
 949         struct dlm_assert_joined assert_msg;
 950
 951         mlog(0, "Sending join assert to node %u\n", node);
 952
 953         memset(&assert_msg, 0, sizeof(assert_msg));
 954         assert_msg.node_idx = dlm->node_num;
 955         assert_msg.name_len = strlen(dlm->name);
 956         memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
 957
 958         status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
 959                                     &assert_msg, sizeof(assert_msg), node,
 960                                     NULL);
 961         if (status < 0)
 962                 mlog_errno(status);
 963
 964         return status;
 965 }
 966
 967 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
 968                                   unsigned long *node_map)
 969 {
 970         int status, node, live;
 971
 972         status = 0;
 973         node = -1;
 974         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
 975                                      node + 1)) < O2NM_MAX_NODES) {
 976                 if (node == dlm->node_num)
 977                         continue;
 978
 979                 do {
 980                         /* It is very important that this message be
 981                          * received so we spin until either the node
 982                          * has died or it gets the message. */
 983                         status = dlm_send_one_join_assert(dlm, node);
 984
 985                         spin_lock(&dlm->spinlock);
 986                         live = test_bit(node, dlm->live_nodes_map);
 987                         spin_unlock(&dlm->spinlock);
 988
 989                         if (status) {
 990                                 mlog(ML_ERROR, "Error return %d asserting "
 991                                      "join on node %d\n", status, node);
 992
 993                                 /* give us some time between errors... */
 994                                 if (live)
 995                                         msleep(DLM_DOMAIN_BACKOFF_MS);
 996                         }
 997                 } while (status && live);
 998         }
 999 }
1000
1001 struct domain_join_ctxt {
1002         unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1003         unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1004 };
1005
1006 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
1007                                    struct domain_join_ctxt *ctxt,
1008                                    enum dlm_query_join_response response)
1009 {
1010         int ret;
1011
1012         if (response == JOIN_DISALLOW) {
1013                 mlog(0, "Latest response of disallow -- should restart\n");
1014                 return 1;
1015         }
1016
1017         spin_lock(&dlm->spinlock);
1018         /* For now, we restart the process if the node maps have
1019          * changed at all */
1020         ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
1021                      sizeof(dlm->live_nodes_map));
1022         spin_unlock(&dlm->spinlock);
1023
1024         if (ret)
1025                 mlog(0, "Node maps changed -- should restart\n");
1026
1027         return ret;
1028 }
1029
1030 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
1031 {
1032         int status = 0, tmpstat, node;
1033         struct domain_join_ctxt *ctxt;
1034         enum dlm_query_join_response response;
1035
1036         mlog_entry("%p", dlm);
1037
1038         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
1039         if (!ctxt) {
1040                 status = -ENOMEM;
1041                 mlog_errno(status);
1042                 goto bail;
1043         }
1044
1045         /* group sem locking should work for us here -- we're already
1046          * registered for heartbeat events so filling this should be
1047          * atomic wrt getting those handlers called. */
1048         o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
1049
1050         spin_lock(&dlm->spinlock);
1051         memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
1052
1053         __dlm_set_joining_node(dlm, dlm->node_num);
1054
1055         spin_unlock(&dlm->spinlock);
1056
1057         node = -1;
1058         while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
1059                                      node + 1)) < O2NM_MAX_NODES) {
1060                 if (node == dlm->node_num)
1061                         continue;
1062
1063                 status = dlm_request_join(dlm, node, &response);
1064                 if (status < 0) {
1065                         mlog_errno(status);
1066                         goto bail;
1067                 }
1068
1069                 /* Ok, either we got a response or the node doesn't have a
1070                  * dlm up. */
1071                 if (response == JOIN_OK)
1072                         set_bit(node, ctxt->yes_resp_map);
1073
1074                 if (dlm_should_restart_join(dlm, ctxt, response)) {
1075                         status = -EAGAIN;
1076                         goto bail;
1077                 }
1078         }
1079
1080         mlog(0, "Yay, done querying nodes!\n");
1081
1082         /* Yay, everyone agree's we can join the domain. My domain is
1083          * comprised of all nodes who were put in the
1084          * yes_resp_map. Copy that into our domain map and send a join
1085          * assert message to clean up everyone elses state. */
1086         spin_lock(&dlm->spinlock);
1087         memcpy(dlm->domain_map, ctxt->yes_resp_map,
1088                sizeof(ctxt->yes_resp_map));
1089         set_bit(dlm->node_num, dlm->domain_map);
1090         spin_unlock(&dlm->spinlock);
1091
1092         dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1093
1094         /* Joined state *must* be set before the joining node
1095          * information, otherwise the query_join handler may read no
1096          * current joiner but a state of NEW and tell joining nodes
1097          * we're not in the domain. */
1098         spin_lock(&dlm_domain_lock);
1099         dlm->dlm_state = DLM_CTXT_JOINED;
1100         dlm->num_joins++;
1101         spin_unlock(&dlm_domain_lock);
1102
1103 bail:
1104         spin_lock(&dlm->spinlock);
1105         __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1106         if (!status)
1107                 __dlm_print_nodes(dlm);
1108         spin_unlock(&dlm->spinlock);
1109
1110         if (ctxt) {
1111                 /* Do we need to send a cancel message to any nodes? */
1112                 if (status < 0) {
1113                         tmpstat = dlm_send_join_cancels(dlm,
1114                                                         ctxt->yes_resp_map,
1115                                                         sizeof(ctxt->yes_resp_map));
1116                         if (tmpstat < 0)
1117                                 mlog_errno(tmpstat);
1118                 }
1119                 kfree(ctxt);
1120         }
1121
1122         mlog(0, "returning %d\n", status);
1123         return status;
1124 }
1125
1126 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1127 {
1128         o2hb_unregister_callback(&dlm->dlm_hb_up);
1129         o2hb_unregister_callback(&dlm->dlm_hb_down);
1130         o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1131 }
1132
1133 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1134 {
1135         int status;
1136
1137         mlog(0, "registering handlers.\n");
1138
1139         o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1140                             dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1141         status = o2hb_register_callback(&dlm->dlm_hb_down);
1142         if (status)
1143                 goto bail;
1144
1145         o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1146                             dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1147         status = o2hb_register_callback(&dlm->dlm_hb_up);
1148         if (status)
1149                 goto bail;
1150
1151         status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1152                                         sizeof(struct dlm_master_request),
1153                                         dlm_master_request_handler,
1154                                         dlm, NULL, &dlm->dlm_domain_handlers);
1155         if (status)
1156                 goto bail;
1157
1158         status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1159                                         sizeof(struct dlm_assert_master),
1160                                         dlm_assert_master_handler,
1161                                         dlm, dlm_assert_master_post_handler,
1162                                         &dlm->dlm_domain_handlers);
1163         if (status)
1164                 goto bail;
1165
1166         status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1167                                         sizeof(struct dlm_create_lock),
1168                                         dlm_create_lock_handler,
1169                                         dlm, NULL, &dlm->dlm_domain_handlers);
1170         if (status)
1171                 goto bail;
1172
1173         status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1174                                         DLM_CONVERT_LOCK_MAX_LEN,
1175                                         dlm_convert_lock_handler,
1176                                         dlm, NULL, &dlm->dlm_domain_handlers);
1177         if (status)
1178                 goto bail;
1179
1180         status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1181                                         DLM_UNLOCK_LOCK_MAX_LEN,
1182                                         dlm_unlock_lock_handler,
1183                                         dlm, NULL, &dlm->dlm_domain_handlers);
1184         if (status)
1185                 goto bail;
1186
1187         status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1188                                         DLM_PROXY_AST_MAX_LEN,
1189                                         dlm_proxy_ast_handler,
1190                                         dlm, NULL, &dlm->dlm_domain_handlers);
1191         if (status)
1192                 goto bail;
1193
1194         status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1195                                         sizeof(struct dlm_exit_domain),
1196                                         dlm_exit_domain_handler,
1197                                         dlm, NULL, &dlm->dlm_domain_handlers);
1198         if (status)
1199                 goto bail;
1200
1201         status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1202                                         sizeof(struct dlm_deref_lockres),
1203                                         dlm_deref_lockres_handler,
1204                                         dlm, NULL, &dlm->dlm_domain_handlers);
1205         if (status)
1206                 goto bail;
1207
1208         status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1209                                         sizeof(struct dlm_migrate_request),
1210                                         dlm_migrate_request_handler,
1211                                         dlm, NULL, &dlm->dlm_domain_handlers);
1212         if (status)
1213                 goto bail;
1214
1215         status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1216                                         DLM_MIG_LOCKRES_MAX_LEN,
1217                                         dlm_mig_lockres_handler,
1218                                         dlm, NULL, &dlm->dlm_domain_handlers);
1219         if (status)
1220                 goto bail;
1221
1222         status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1223                                         sizeof(struct dlm_master_requery),
1224                                         dlm_master_requery_handler,
1225                                         dlm, NULL, &dlm->dlm_domain_handlers);
1226         if (status)
1227                 goto bail;
1228
1229         status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1230                                         sizeof(struct dlm_lock_request),
1231                                         dlm_request_all_locks_handler,
1232                                         dlm, NULL, &dlm->dlm_domain_handlers);
1233         if (status)
1234                 goto bail;
1235
1236         status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1237                                         sizeof(struct dlm_reco_data_done),
1238                                         dlm_reco_data_done_handler,
1239                                         dlm, NULL, &dlm->dlm_domain_handlers);
1240         if (status)
1241                 goto bail;
1242
1243         status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1244                                         sizeof(struct dlm_begin_reco),
1245                                         dlm_begin_reco_handler,
1246                                         dlm, NULL, &dlm->dlm_domain_handlers);
1247         if (status)
1248                 goto bail;
1249
1250         status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1251                                         sizeof(struct dlm_finalize_reco),
1252                                         dlm_finalize_reco_handler,
1253                                         dlm, NULL, &dlm->dlm_domain_handlers);
1254         if (status)
1255                 goto bail;
1256
1257 bail:
1258         if (status)
1259                 dlm_unregister_domain_handlers(dlm);
1260
1261         return status;
1262 }
1263
1264 static int dlm_join_domain(struct dlm_ctxt *dlm)
1265 {
1266         int status;
1267         unsigned int backoff;
1268         unsigned int total_backoff = 0;
1269
1270         BUG_ON(!dlm);
1271
1272         mlog(0, "Join domain %s\n", dlm->name);
1273
1274         status = dlm_register_domain_handlers(dlm);
1275         if (status) {
1276                 mlog_errno(status);
1277                 goto bail;
1278         }
1279
1280         status = dlm_launch_thread(dlm);
1281         if (status < 0) {
1282                 mlog_errno(status);
1283                 goto bail;
1284         }
1285
1286         status = dlm_launch_recovery_thread(dlm);
1287         if (status < 0) {
1288                 mlog_errno(status);
1289                 goto bail;
1290         }
1291
1292         dlm->dlm_worker = create_singlethread_workqueue("dlm_wq");
1293         if (!dlm->dlm_worker) {
1294                 status = -ENOMEM;
1295                 mlog_errno(status);
1296                 goto bail;
1297         }
1298
1299         do {
1300                 status = dlm_try_to_join_domain(dlm);
1301
1302                 /* If we're racing another node to the join, then we
1303                  * need to back off temporarily and let them
1304                  * complete. */
1305 #define DLM_JOIN_TIMEOUT_MSECS  90000
1306                 if (status == -EAGAIN) {
1307                         if (signal_pending(current)) {
1308                                 status = -ERESTARTSYS;
1309                                 goto bail;
1310                         }
1311
1312                         if (total_backoff >
1313                             msecs_to_jiffies(DLM_JOIN_TIMEOUT_MSECS)) {
1314                                 status = -ERESTARTSYS;
1315                                 mlog(ML_NOTICE, "Timed out joining dlm domain "
1316                                      "%s after %u msecs\n", dlm->name,
1317                                      jiffies_to_msecs(total_backoff));
1318                                 goto bail;
1319                         }
1320
1321                         /*
1322                          * <chip> After you!
1323                          * <dale> No, after you!
1324                          * <chip> I insist!
1325                          * <dale> But you first!
1326                          * ...
1327                          */
1328                         backoff = (unsigned int)(jiffies & 0x3);
1329                         backoff *= DLM_DOMAIN_BACKOFF_MS;
1330                         total_backoff += backoff;
1331                         mlog(0, "backoff %d\n", backoff);
1332                         msleep(backoff);
1333                 }
1334         } while (status == -EAGAIN);
1335
1336         if (status < 0) {
1337                 mlog_errno(status);
1338                 goto bail;
1339         }
1340
1341         status = 0;
1342 bail:
1343         wake_up(&dlm_domain_events);
1344
1345         if (status) {
1346                 dlm_unregister_domain_handlers(dlm);
1347                 dlm_complete_thread(dlm);
1348                 dlm_complete_recovery_thread(dlm);
1349                 dlm_destroy_dlm_worker(dlm);
1350         }
1351
1352         return status;
1353 }
1354
1355 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1356                                 u32 key)
1357 {
1358         int i;
1359         struct dlm_ctxt *dlm = NULL;
1360
1361         dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1362         if (!dlm) {
1363                 mlog_errno(-ENOMEM);
1364                 goto leave;
1365         }
1366
1367         dlm->name = kmalloc(strlen(domain) + 1, GFP_KERNEL);
1368         if (dlm->name == NULL) {
1369                 mlog_errno(-ENOMEM);
1370                 kfree(dlm);
1371                 dlm = NULL;
1372                 goto leave;
1373         }
1374
1375         dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1376         if (!dlm->lockres_hash) {
1377                 mlog_errno(-ENOMEM);
1378                 kfree(dlm->name);
1379                 kfree(dlm);
1380                 dlm = NULL;
1381                 goto leave;
1382         }
1383
1384         for (i = 0; i < DLM_HASH_BUCKETS; i++)
1385                 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1386
1387         strcpy(dlm->name, domain);
1388         dlm->key = key;
1389         dlm->node_num = o2nm_this_node();
1390
1391         spin_lock_init(&dlm->spinlock);
1392         spin_lock_init(&dlm->master_lock);
1393         spin_lock_init(&dlm->ast_lock);
1394         INIT_LIST_HEAD(&dlm->list);
1395         INIT_LIST_HEAD(&dlm->dirty_list);
1396         INIT_LIST_HEAD(&dlm->reco.resources);
1397         INIT_LIST_HEAD(&dlm->reco.received);
1398         INIT_LIST_HEAD(&dlm->reco.node_data);
1399         INIT_LIST_HEAD(&dlm->purge_list);
1400         INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
1401         dlm->reco.state = 0;
1402
1403         INIT_LIST_HEAD(&dlm->pending_asts);
1404         INIT_LIST_HEAD(&dlm->pending_basts);
1405
1406         mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
1407                   dlm->recovery_map, &(dlm->recovery_map[0]));
1408
1409         memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
1410         memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
1411         memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
1412
1413         dlm->dlm_thread_task = NULL;
1414         dlm->dlm_reco_thread_task = NULL;
1415         dlm->dlm_worker = NULL;
1416         init_waitqueue_head(&dlm->dlm_thread_wq);
1417         init_waitqueue_head(&dlm->dlm_reco_thread_wq);
1418         init_waitqueue_head(&dlm->reco.event);
1419         init_waitqueue_head(&dlm->ast_wq);
1420         init_waitqueue_head(&dlm->migration_wq);
1421         INIT_LIST_HEAD(&dlm->master_list);
1422         INIT_LIST_HEAD(&dlm->mle_hb_events);
1423
1424         dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
1425         init_waitqueue_head(&dlm->dlm_join_events);
1426
1427         dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
1428         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
1429         atomic_set(&dlm->local_resources, 0);
1430         atomic_set(&dlm->remote_resources, 0);
1431         atomic_set(&dlm->unknown_resources, 0);
1432
1433         spin_lock_init(&dlm->work_lock);
1434         INIT_LIST_HEAD(&dlm->work_list);
1435         INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
1436
1437         kref_init(&dlm->dlm_refs);
1438         dlm->dlm_state = DLM_CTXT_NEW;
1439
1440         INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
1441
1442         mlog(0, "context init: refcount %u\n",
1443                   atomic_read(&dlm->dlm_refs.refcount));
1444
1445 leave:
1446         return dlm;
1447 }
1448
1449 /*
1450  * dlm_register_domain: one-time setup per "domain"
1451  */
1452 struct dlm_ctxt * dlm_register_domain(const char *domain,
1453                                u32 key)
1454 {
1455         int ret;
1456         struct dlm_ctxt *dlm = NULL;
1457         struct dlm_ctxt *new_ctxt = NULL;
1458
1459         if (strlen(domain) > O2NM_MAX_NAME_LEN) {
1460                 ret = -ENAMETOOLONG;
1461                 mlog(ML_ERROR, "domain name length too long\n");
1462                 goto leave;
1463         }
1464
1465         if (!o2hb_check_local_node_heartbeating()) {
1466                 mlog(ML_ERROR, "the local node has not been configured, or is "
1467                      "not heartbeating\n");
1468                 ret = -EPROTO;
1469                 goto leave;
1470         }
1471
1472         mlog(0, "register called for domain \"%s\"\n", domain);
1473
1474 retry:
1475         dlm = NULL;
1476         if (signal_pending(current)) {
1477                 ret = -ERESTARTSYS;
1478                 mlog_errno(ret);
1479                 goto leave;
1480         }
1481
1482         spin_lock(&dlm_domain_lock);
1483
1484         dlm = __dlm_lookup_domain(domain);
1485         if (dlm) {
1486                 if (dlm->dlm_state != DLM_CTXT_JOINED) {
1487                         spin_unlock(&dlm_domain_lock);
1488
1489                         mlog(0, "This ctxt is not joined yet!\n");
1490                         wait_event_interruptible(dlm_domain_events,
1491                                                  dlm_wait_on_domain_helper(
1492                                                          domain));
1493                         goto retry;
1494                 }
1495
1496                 __dlm_get(dlm);
1497                 dlm->num_joins++;
1498
1499                 spin_unlock(&dlm_domain_lock);
1500
1501                 ret = 0;
1502                 goto leave;
1503         }
1504
1505         /* doesn't exist */
1506         if (!new_ctxt) {
1507                 spin_unlock(&dlm_domain_lock);
1508
1509                 new_ctxt = dlm_alloc_ctxt(domain, key);
1510                 if (new_ctxt)
1511                         goto retry;
1512
1513                 ret = -ENOMEM;
1514                 mlog_errno(ret);
1515                 goto leave;
1516         }
1517
1518         /* a little variable switch-a-roo here... */
1519         dlm = new_ctxt;
1520         new_ctxt = NULL;
1521
1522         /* add the new domain */
1523         list_add_tail(&dlm->list, &dlm_domains);
1524         spin_unlock(&dlm_domain_lock);
1525
1526         ret = dlm_join_domain(dlm);
1527         if (ret) {
1528                 mlog_errno(ret);
1529                 dlm_put(dlm);
1530                 goto leave;
1531         }
1532
1533         ret = 0;
1534 leave:
1535         if (new_ctxt)
1536                 dlm_free_ctxt_mem(new_ctxt);
1537
1538         if (ret < 0)
1539                 dlm = ERR_PTR(ret);
1540
1541         return dlm;
1542 }
1543 EXPORT_SYMBOL_GPL(dlm_register_domain);
1544
1545 static LIST_HEAD(dlm_join_handlers);
1546
1547 static void dlm_unregister_net_handlers(void)
1548 {
1549         o2net_unregister_handler_list(&dlm_join_handlers);
1550 }
1551
1552 static int dlm_register_net_handlers(void)
1553 {
1554         int status = 0;
1555
1556         status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1557                                         sizeof(struct dlm_query_join_request),
1558                                         dlm_query_join_handler,
1559                                         NULL, NULL, &dlm_join_handlers);
1560         if (status)
1561                 goto bail;
1562
1563         status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1564                                         sizeof(struct dlm_assert_joined),
1565                                         dlm_assert_joined_handler,
1566                                         NULL, NULL, &dlm_join_handlers);
1567         if (status)
1568                 goto bail;
1569
1570         status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1571                                         sizeof(struct dlm_cancel_join),
1572                                         dlm_cancel_join_handler,
1573                                         NULL, NULL, &dlm_join_handlers);
1574
1575 bail:
1576         if (status < 0)
1577                 dlm_unregister_net_handlers();
1578
1579         return status;
1580 }
1581
1582 /* Domain eviction callback handling.
1583  *
1584  * The file system requires notification of node death *before* the
1585  * dlm completes it's recovery work, otherwise it may be able to
1586  * acquire locks on resources requiring recovery. Since the dlm can
1587  * evict a node from it's domain *before* heartbeat fires, a similar
1588  * mechanism is required. */
1589
1590 /* Eviction is not expected to happen often, so a per-domain lock is
1591  * not necessary. Eviction callbacks are allowed to sleep for short
1592  * periods of time. */
1593 static DECLARE_RWSEM(dlm_callback_sem);
1594
1595 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
1596                                         int node_num)
1597 {
1598         struct list_head *iter;
1599         struct dlm_eviction_cb *cb;
1600
1601         down_read(&dlm_callback_sem);
1602         list_for_each(iter, &dlm->dlm_eviction_callbacks) {
1603                 cb = list_entry(iter, struct dlm_eviction_cb, ec_item);
1604
1605                 cb->ec_func(node_num, cb->ec_data);
1606         }
1607         up_read(&dlm_callback_sem);
1608 }
1609
1610 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
1611                            dlm_eviction_func *f,
1612                            void *data)
1613 {
1614         INIT_LIST_HEAD(&cb->ec_item);
1615         cb->ec_func = f;
1616         cb->ec_data = data;
1617 }
1618 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
1619
1620 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
1621                               struct dlm_eviction_cb *cb)
1622 {
1623         down_write(&dlm_callback_sem);
1624         list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
1625         up_write(&dlm_callback_sem);
1626 }
1627 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
1628
1629 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
1630 {
1631         down_write(&dlm_callback_sem);
1632         list_del_init(&cb->ec_item);
1633         up_write(&dlm_callback_sem);
1634 }
1635 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
1636
1637 static int __init dlm_init(void)
1638 {
1639         int status;
1640
1641         dlm_print_version();
1642
1643         status = dlm_init_mle_cache();
1644         if (status)
1645                 return -1;
1646
1647         status = dlm_register_net_handlers();
1648         if (status) {
1649                 dlm_destroy_mle_cache();
1650                 return -1;
1651         }
1652
1653         return 0;
1654 }
1655
1656 static void __exit dlm_exit (void)
1657 {
1658         dlm_unregister_net_handlers();
1659         dlm_destroy_mle_cache();
1660 }
1661
1662 MODULE_AUTHOR("Oracle");
1663 MODULE_LICENSE("GPL");
1664
1665 module_init(dlm_init);
1666 module_exit(dlm_exit);