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                                 response = JOIN_DISALLOW;
 711                                 goto unlock_respond;
 712                         }
 713                 }
 714                 nodenum++;
 715         }
 716
 717         /* Once the dlm ctxt is marked as leaving then we don't want
 718          * to be put in someone's domain map.
 719          * Also, explicitly disallow joining at certain troublesome
 720          * times (ie. during recovery). */
 721         if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
 722                 int bit = query->node_idx;
 723                 spin_lock(&dlm->spinlock);
 724
 725                 if (dlm->dlm_state == DLM_CTXT_NEW &&
 726                     dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
 727                         /*If this is a brand new context and we
 728                          * haven't started our join process yet, then
 729                          * the other node won the race. */
 730                         response = JOIN_OK_NO_MAP;
 731                 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 732                         /* Disallow parallel joins. */
 733                         response = JOIN_DISALLOW;
 734                 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
 735                         mlog(ML_NOTICE, "node %u trying to join, but recovery "
 736                              "is ongoing.\n", bit);
 737                         response = JOIN_DISALLOW;
 738                 } else if (test_bit(bit, dlm->recovery_map)) {
 739                         mlog(ML_NOTICE, "node %u trying to join, but it "
 740                              "still needs recovery.\n", bit);
 741                         response = JOIN_DISALLOW;
 742                 } else if (test_bit(bit, dlm->domain_map)) {
 743                         mlog(ML_NOTICE, "node %u trying to join, but it "
 744                              "is still in the domain! needs recovery?\n",
 745                              bit);
 746                         response = JOIN_DISALLOW;
 747                 } else {
 748                         /* Alright we're fully a part of this domain
 749                          * so we keep some state as to who's joining
 750                          * and indicate to him that needs to be fixed
 751                          * up. */
 752                         response = JOIN_OK;
 753                         __dlm_set_joining_node(dlm, query->node_idx);
 754                 }
 755
 756                 spin_unlock(&dlm->spinlock);
 757         }
 758 unlock_respond:
 759         spin_unlock(&dlm_domain_lock);
 760
 761 respond:
 762         mlog(0, "We respond with %u\n", response);
 763
 764         return response;
 765 }
 766
 767 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 768                                      void **ret_data)
 769 {
 770         struct dlm_assert_joined *assert;
 771         struct dlm_ctxt *dlm = NULL;
 772
 773         assert = (struct dlm_assert_joined *) msg->buf;
 774
 775         mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
 776                   assert->domain);
 777
 778         spin_lock(&dlm_domain_lock);
 779         dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
 780         /* XXX should we consider no dlm ctxt an error? */
 781         if (dlm) {
 782                 spin_lock(&dlm->spinlock);
 783
 784                 /* Alright, this node has officially joined our
 785                  * domain. Set him in the map and clean up our
 786                  * leftover join state. */
 787                 BUG_ON(dlm->joining_node != assert->node_idx);
 788                 set_bit(assert->node_idx, dlm->domain_map);
 789                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 790
 791                 printk(KERN_INFO "ocfs2_dlm: Node %u joins domain %s\n",
 792                        assert->node_idx, dlm->name);
 793                 __dlm_print_nodes(dlm);
 794
 795                 /* notify anything attached to the heartbeat events */
 796                 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
 797
 798                 spin_unlock(&dlm->spinlock);
 799         }
 800         spin_unlock(&dlm_domain_lock);
 801
 802         return 0;
 803 }
 804
 805 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
 806                                    void **ret_data)
 807 {
 808         struct dlm_cancel_join *cancel;
 809         struct dlm_ctxt *dlm = NULL;
 810
 811         cancel = (struct dlm_cancel_join *) msg->buf;
 812
 813         mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
 814                   cancel->domain);
 815
 816         spin_lock(&dlm_domain_lock);
 817         dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
 818
 819         if (dlm) {
 820                 spin_lock(&dlm->spinlock);
 821
 822                 /* Yikes, this guy wants to cancel his join. No
 823                  * problem, we simply cleanup our join state. */
 824                 BUG_ON(dlm->joining_node != cancel->node_idx);
 825                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 826
 827                 spin_unlock(&dlm->spinlock);
 828         }
 829         spin_unlock(&dlm_domain_lock);
 830
 831         return 0;
 832 }
 833
 834 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
 835                                     unsigned int node)
 836 {
 837         int status;
 838         struct dlm_cancel_join cancel_msg;
 839
 840         memset(&cancel_msg, 0, sizeof(cancel_msg));
 841         cancel_msg.node_idx = dlm->node_num;
 842         cancel_msg.name_len = strlen(dlm->name);
 843         memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
 844
 845         status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
 846                                     &cancel_msg, sizeof(cancel_msg), node,
 847                                     NULL);
 848         if (status < 0) {
 849                 mlog_errno(status);
 850                 goto bail;
 851         }
 852
 853 bail:
 854         return status;
 855 }
 856
 857 /* map_size should be in bytes. */
 858 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
 859                                  unsigned long *node_map,
 860                                  unsigned int map_size)
 861 {
 862         int status, tmpstat;
 863         unsigned int node;
 864
 865         if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
 866                          sizeof(unsigned long))) {
 867                 mlog(ML_ERROR,
 868                      "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
 869                      map_size, BITS_TO_LONGS(O2NM_MAX_NODES));
 870                 return -EINVAL;
 871         }
 872
 873         status = 0;
 874         node = -1;
 875         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
 876                                      node + 1)) < O2NM_MAX_NODES) {
 877                 if (node == dlm->node_num)
 878                         continue;
 879
 880                 tmpstat = dlm_send_one_join_cancel(dlm, node);
 881                 if (tmpstat) {
 882                         mlog(ML_ERROR, "Error return %d cancelling join on "
 883                              "node %d\n", tmpstat, node);
 884                         if (!status)
 885                                 status = tmpstat;
 886                 }
 887         }
 888
 889         if (status)
 890                 mlog_errno(status);
 891         return status;
 892 }
 893
 894 static int dlm_request_join(struct dlm_ctxt *dlm,
 895                             int node,
 896                             enum dlm_query_join_response *response)
 897 {
 898         int status, retval;
 899         struct dlm_query_join_request join_msg;
 900
 901         mlog(0, "querying node %d\n", node);
 902
 903         memset(&join_msg, 0, sizeof(join_msg));
 904         join_msg.node_idx = dlm->node_num;
 905         join_msg.name_len = strlen(dlm->name);
 906         memcpy(join_msg.domain, dlm->name, join_msg.name_len);
 907
 908         /* copy live node map to join message */
 909         byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES);
 910
 911         status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
 912                                     sizeof(join_msg), node, &retval);
 913         if (status < 0 && status != -ENOPROTOOPT) {
 914                 mlog_errno(status);
 915                 goto bail;
 916         }
 917
 918         /* -ENOPROTOOPT from the net code means the other side isn't
 919             listening for our message type -- that's fine, it means
 920             his dlm isn't up, so we can consider him a 'yes' but not
 921             joined into the domain.  */
 922         if (status == -ENOPROTOOPT) {
 923                 status = 0;
 924                 *response = JOIN_OK_NO_MAP;
 925         } else if (retval == JOIN_DISALLOW ||
 926                    retval == JOIN_OK ||
 927                    retval == JOIN_OK_NO_MAP) {
 928                 *response = retval;
 929         } else {
 930                 status = -EINVAL;
 931                 mlog(ML_ERROR, "invalid response %d from node %u\n", retval,
 932                      node);
 933         }
 934
 935         mlog(0, "status %d, node %d response is %d\n", status, node,
 936                   *response);
 937
 938 bail:
 939         return status;
 940 }
 941
 942 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
 943                                     unsigned int node)
 944 {
 945         int status;
 946         struct dlm_assert_joined assert_msg;
 947
 948         mlog(0, "Sending join assert to node %u\n", node);
 949
 950         memset(&assert_msg, 0, sizeof(assert_msg));
 951         assert_msg.node_idx = dlm->node_num;
 952         assert_msg.name_len = strlen(dlm->name);
 953         memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
 954
 955         status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
 956                                     &assert_msg, sizeof(assert_msg), node,
 957                                     NULL);
 958         if (status < 0)
 959                 mlog_errno(status);
 960
 961         return status;
 962 }
 963
 964 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
 965                                   unsigned long *node_map)
 966 {
 967         int status, node, live;
 968
 969         status = 0;
 970         node = -1;
 971         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
 972                                      node + 1)) < O2NM_MAX_NODES) {
 973                 if (node == dlm->node_num)
 974                         continue;
 975
 976                 do {
 977                         /* It is very important that this message be
 978                          * received so we spin until either the node
 979                          * has died or it gets the message. */
 980                         status = dlm_send_one_join_assert(dlm, node);
 981
 982                         spin_lock(&dlm->spinlock);
 983                         live = test_bit(node, dlm->live_nodes_map);
 984                         spin_unlock(&dlm->spinlock);
 985
 986                         if (status) {
 987                                 mlog(ML_ERROR, "Error return %d asserting "
 988                                      "join on node %d\n", status, node);
 989
 990                                 /* give us some time between errors... */
 991                                 if (live)
 992                                         msleep(DLM_DOMAIN_BACKOFF_MS);
 993                         }
 994                 } while (status && live);
 995         }
 996 }
 997
 998 struct domain_join_ctxt {
 999         unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1000         unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1001 };
1002
1003 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
1004                                    struct domain_join_ctxt *ctxt,
1005                                    enum dlm_query_join_response response)
1006 {
1007         int ret;
1008
1009         if (response == JOIN_DISALLOW) {
1010                 mlog(0, "Latest response of disallow -- should restart\n");
1011                 return 1;
1012         }
1013
1014         spin_lock(&dlm->spinlock);
1015         /* For now, we restart the process if the node maps have
1016          * changed at all */
1017         ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
1018                      sizeof(dlm->live_nodes_map));
1019         spin_unlock(&dlm->spinlock);
1020
1021         if (ret)
1022                 mlog(0, "Node maps changed -- should restart\n");
1023
1024         return ret;
1025 }
1026
1027 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
1028 {
1029         int status = 0, tmpstat, node;
1030         struct domain_join_ctxt *ctxt;
1031         enum dlm_query_join_response response;
1032
1033         mlog_entry("%p", dlm);
1034
1035         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
1036         if (!ctxt) {
1037                 status = -ENOMEM;
1038                 mlog_errno(status);
1039                 goto bail;
1040         }
1041
1042         /* group sem locking should work for us here -- we're already
1043          * registered for heartbeat events so filling this should be
1044          * atomic wrt getting those handlers called. */
1045         o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
1046
1047         spin_lock(&dlm->spinlock);
1048         memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
1049
1050         __dlm_set_joining_node(dlm, dlm->node_num);
1051
1052         spin_unlock(&dlm->spinlock);
1053
1054         node = -1;
1055         while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
1056                                      node + 1)) < O2NM_MAX_NODES) {
1057                 if (node == dlm->node_num)
1058                         continue;
1059
1060                 status = dlm_request_join(dlm, node, &response);
1061                 if (status < 0) {
1062                         mlog_errno(status);
1063                         goto bail;
1064                 }
1065
1066                 /* Ok, either we got a response or the node doesn't have a
1067                  * dlm up. */
1068                 if (response == JOIN_OK)
1069                         set_bit(node, ctxt->yes_resp_map);
1070
1071                 if (dlm_should_restart_join(dlm, ctxt, response)) {
1072                         status = -EAGAIN;
1073                         goto bail;
1074                 }
1075         }
1076
1077         mlog(0, "Yay, done querying nodes!\n");
1078
1079         /* Yay, everyone agree's we can join the domain. My domain is
1080          * comprised of all nodes who were put in the
1081          * yes_resp_map. Copy that into our domain map and send a join
1082          * assert message to clean up everyone elses state. */
1083         spin_lock(&dlm->spinlock);
1084         memcpy(dlm->domain_map, ctxt->yes_resp_map,
1085                sizeof(ctxt->yes_resp_map));
1086         set_bit(dlm->node_num, dlm->domain_map);
1087         spin_unlock(&dlm->spinlock);
1088
1089         dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1090
1091         /* Joined state *must* be set before the joining node
1092          * information, otherwise the query_join handler may read no
1093          * current joiner but a state of NEW and tell joining nodes
1094          * we're not in the domain. */
1095         spin_lock(&dlm_domain_lock);
1096         dlm->dlm_state = DLM_CTXT_JOINED;
1097         dlm->num_joins++;
1098         spin_unlock(&dlm_domain_lock);
1099
1100 bail:
1101         spin_lock(&dlm->spinlock);
1102         __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1103         if (!status)
1104                 __dlm_print_nodes(dlm);
1105         spin_unlock(&dlm->spinlock);
1106
1107         if (ctxt) {
1108                 /* Do we need to send a cancel message to any nodes? */
1109                 if (status < 0) {
1110                         tmpstat = dlm_send_join_cancels(dlm,
1111                                                         ctxt->yes_resp_map,
1112                                                         sizeof(ctxt->yes_resp_map));
1113                         if (tmpstat < 0)
1114                                 mlog_errno(tmpstat);
1115                 }
1116                 kfree(ctxt);
1117         }
1118
1119         mlog(0, "returning %d\n", status);
1120         return status;
1121 }
1122
1123 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1124 {
1125         o2hb_unregister_callback(&dlm->dlm_hb_up);
1126         o2hb_unregister_callback(&dlm->dlm_hb_down);
1127         o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1128 }
1129
1130 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1131 {
1132         int status;
1133
1134         mlog(0, "registering handlers.\n");
1135
1136         o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1137                             dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1138         status = o2hb_register_callback(&dlm->dlm_hb_down);
1139         if (status)
1140                 goto bail;
1141
1142         o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1143                             dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1144         status = o2hb_register_callback(&dlm->dlm_hb_up);
1145         if (status)
1146                 goto bail;
1147
1148         status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1149                                         sizeof(struct dlm_master_request),
1150                                         dlm_master_request_handler,
1151                                         dlm, NULL, &dlm->dlm_domain_handlers);
1152         if (status)
1153                 goto bail;
1154
1155         status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1156                                         sizeof(struct dlm_assert_master),
1157                                         dlm_assert_master_handler,
1158                                         dlm, dlm_assert_master_post_handler,
1159                                         &dlm->dlm_domain_handlers);
1160         if (status)
1161                 goto bail;
1162
1163         status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1164                                         sizeof(struct dlm_create_lock),
1165                                         dlm_create_lock_handler,
1166                                         dlm, NULL, &dlm->dlm_domain_handlers);
1167         if (status)
1168                 goto bail;
1169
1170         status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1171                                         DLM_CONVERT_LOCK_MAX_LEN,
1172                                         dlm_convert_lock_handler,
1173                                         dlm, NULL, &dlm->dlm_domain_handlers);
1174         if (status)
1175                 goto bail;
1176
1177         status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1178                                         DLM_UNLOCK_LOCK_MAX_LEN,
1179                                         dlm_unlock_lock_handler,
1180                                         dlm, NULL, &dlm->dlm_domain_handlers);
1181         if (status)
1182                 goto bail;
1183
1184         status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1185                                         DLM_PROXY_AST_MAX_LEN,
1186                                         dlm_proxy_ast_handler,
1187                                         dlm, NULL, &dlm->dlm_domain_handlers);
1188         if (status)
1189                 goto bail;
1190
1191         status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1192                                         sizeof(struct dlm_exit_domain),
1193                                         dlm_exit_domain_handler,
1194                                         dlm, NULL, &dlm->dlm_domain_handlers);
1195         if (status)
1196                 goto bail;
1197
1198         status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1199                                         sizeof(struct dlm_deref_lockres),
1200                                         dlm_deref_lockres_handler,
1201                                         dlm, NULL, &dlm->dlm_domain_handlers);
1202         if (status)
1203                 goto bail;
1204
1205         status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1206                                         sizeof(struct dlm_migrate_request),
1207                                         dlm_migrate_request_handler,
1208                                         dlm, NULL, &dlm->dlm_domain_handlers);
1209         if (status)
1210                 goto bail;
1211
1212         status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1213                                         DLM_MIG_LOCKRES_MAX_LEN,
1214                                         dlm_mig_lockres_handler,
1215                                         dlm, NULL, &dlm->dlm_domain_handlers);
1216         if (status)
1217                 goto bail;
1218
1219         status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1220                                         sizeof(struct dlm_master_requery),
1221                                         dlm_master_requery_handler,
1222                                         dlm, NULL, &dlm->dlm_domain_handlers);
1223         if (status)
1224                 goto bail;
1225
1226         status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1227                                         sizeof(struct dlm_lock_request),
1228                                         dlm_request_all_locks_handler,
1229                                         dlm, NULL, &dlm->dlm_domain_handlers);
1230         if (status)
1231                 goto bail;
1232
1233         status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1234                                         sizeof(struct dlm_reco_data_done),
1235                                         dlm_reco_data_done_handler,
1236                                         dlm, NULL, &dlm->dlm_domain_handlers);
1237         if (status)
1238                 goto bail;
1239
1240         status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1241                                         sizeof(struct dlm_begin_reco),
1242                                         dlm_begin_reco_handler,
1243                                         dlm, NULL, &dlm->dlm_domain_handlers);
1244         if (status)
1245                 goto bail;
1246
1247         status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1248                                         sizeof(struct dlm_finalize_reco),
1249                                         dlm_finalize_reco_handler,
1250                                         dlm, NULL, &dlm->dlm_domain_handlers);
1251         if (status)
1252                 goto bail;
1253
1254 bail:
1255         if (status)
1256                 dlm_unregister_domain_handlers(dlm);
1257
1258         return status;
1259 }
1260
1261 static int dlm_join_domain(struct dlm_ctxt *dlm)
1262 {
1263         int status;
1264
1265         BUG_ON(!dlm);
1266
1267         mlog(0, "Join domain %s\n", dlm->name);
1268
1269         status = dlm_register_domain_handlers(dlm);
1270         if (status) {
1271                 mlog_errno(status);
1272                 goto bail;
1273         }
1274
1275         status = dlm_launch_thread(dlm);
1276         if (status < 0) {
1277                 mlog_errno(status);
1278                 goto bail;
1279         }
1280
1281         status = dlm_launch_recovery_thread(dlm);
1282         if (status < 0) {
1283                 mlog_errno(status);
1284                 goto bail;
1285         }
1286
1287         dlm->dlm_worker = create_singlethread_workqueue("dlm_wq");
1288         if (!dlm->dlm_worker) {
1289                 status = -ENOMEM;
1290                 mlog_errno(status);
1291                 goto bail;
1292         }
1293
1294         do {
1295                 unsigned int backoff;
1296                 status = dlm_try_to_join_domain(dlm);
1297
1298                 /* If we're racing another node to the join, then we
1299                  * need to back off temporarily and let them
1300                  * complete. */
1301                 if (status == -EAGAIN) {
1302                         if (signal_pending(current)) {
1303                                 status = -ERESTARTSYS;
1304                                 goto bail;
1305                         }
1306
1307                         /*
1308                          * <chip> After you!
1309                          * <dale> No, after you!
1310                          * <chip> I insist!
1311                          * <dale> But you first!
1312                          * ...
1313                          */
1314                         backoff = (unsigned int)(jiffies & 0x3);
1315                         backoff *= DLM_DOMAIN_BACKOFF_MS;
1316                         mlog(0, "backoff %d\n", backoff);
1317                         msleep(backoff);
1318                 }
1319         } while (status == -EAGAIN);
1320
1321         if (status < 0) {
1322                 mlog_errno(status);
1323                 goto bail;
1324         }
1325
1326         status = 0;
1327 bail:
1328         wake_up(&dlm_domain_events);
1329
1330         if (status) {
1331                 dlm_unregister_domain_handlers(dlm);
1332                 dlm_complete_thread(dlm);
1333                 dlm_complete_recovery_thread(dlm);
1334                 dlm_destroy_dlm_worker(dlm);
1335         }
1336
1337         return status;
1338 }
1339
1340 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1341                                 u32 key)
1342 {
1343         int i;
1344         struct dlm_ctxt *dlm = NULL;
1345
1346         dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1347         if (!dlm) {
1348                 mlog_errno(-ENOMEM);
1349                 goto leave;
1350         }
1351
1352         dlm->name = kmalloc(strlen(domain) + 1, GFP_KERNEL);
1353         if (dlm->name == NULL) {
1354                 mlog_errno(-ENOMEM);
1355                 kfree(dlm);
1356                 dlm = NULL;
1357                 goto leave;
1358         }
1359
1360         dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1361         if (!dlm->lockres_hash) {
1362                 mlog_errno(-ENOMEM);
1363                 kfree(dlm->name);
1364                 kfree(dlm);
1365                 dlm = NULL;
1366                 goto leave;
1367         }
1368
1369         for (i = 0; i < DLM_HASH_BUCKETS; i++)
1370                 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1371
1372         strcpy(dlm->name, domain);
1373         dlm->key = key;
1374         dlm->node_num = o2nm_this_node();
1375
1376         spin_lock_init(&dlm->spinlock);
1377         spin_lock_init(&dlm->master_lock);
1378         spin_lock_init(&dlm->ast_lock);
1379         INIT_LIST_HEAD(&dlm->list);
1380         INIT_LIST_HEAD(&dlm->dirty_list);
1381         INIT_LIST_HEAD(&dlm->reco.resources);
1382         INIT_LIST_HEAD(&dlm->reco.received);
1383         INIT_LIST_HEAD(&dlm->reco.node_data);
1384         INIT_LIST_HEAD(&dlm->purge_list);
1385         INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
1386         dlm->reco.state = 0;
1387
1388         INIT_LIST_HEAD(&dlm->pending_asts);
1389         INIT_LIST_HEAD(&dlm->pending_basts);
1390
1391         mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
1392                   dlm->recovery_map, &(dlm->recovery_map[0]));
1393
1394         memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
1395         memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
1396         memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
1397
1398         dlm->dlm_thread_task = NULL;
1399         dlm->dlm_reco_thread_task = NULL;
1400         dlm->dlm_worker = NULL;
1401         init_waitqueue_head(&dlm->dlm_thread_wq);
1402         init_waitqueue_head(&dlm->dlm_reco_thread_wq);
1403         init_waitqueue_head(&dlm->reco.event);
1404         init_waitqueue_head(&dlm->ast_wq);
1405         init_waitqueue_head(&dlm->migration_wq);
1406         INIT_LIST_HEAD(&dlm->master_list);
1407         INIT_LIST_HEAD(&dlm->mle_hb_events);
1408
1409         dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
1410         init_waitqueue_head(&dlm->dlm_join_events);
1411
1412         dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
1413         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
1414         atomic_set(&dlm->local_resources, 0);
1415         atomic_set(&dlm->remote_resources, 0);
1416         atomic_set(&dlm->unknown_resources, 0);
1417
1418         spin_lock_init(&dlm->work_lock);
1419         INIT_LIST_HEAD(&dlm->work_list);
1420         INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
1421
1422         kref_init(&dlm->dlm_refs);
1423         dlm->dlm_state = DLM_CTXT_NEW;
1424
1425         INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
1426
1427         mlog(0, "context init: refcount %u\n",
1428                   atomic_read(&dlm->dlm_refs.refcount));
1429
1430 leave:
1431         return dlm;
1432 }
1433
1434 /*
1435  * dlm_register_domain: one-time setup per "domain"
1436  */
1437 struct dlm_ctxt * dlm_register_domain(const char *domain,
1438                                u32 key)
1439 {
1440         int ret;
1441         struct dlm_ctxt *dlm = NULL;
1442         struct dlm_ctxt *new_ctxt = NULL;
1443
1444         if (strlen(domain) > O2NM_MAX_NAME_LEN) {
1445                 ret = -ENAMETOOLONG;
1446                 mlog(ML_ERROR, "domain name length too long\n");
1447                 goto leave;
1448         }
1449
1450         if (!o2hb_check_local_node_heartbeating()) {
1451                 mlog(ML_ERROR, "the local node has not been configured, or is "
1452                      "not heartbeating\n");
1453                 ret = -EPROTO;
1454                 goto leave;
1455         }
1456
1457         mlog(0, "register called for domain \"%s\"\n", domain);
1458
1459 retry:
1460         dlm = NULL;
1461         if (signal_pending(current)) {
1462                 ret = -ERESTARTSYS;
1463                 mlog_errno(ret);
1464                 goto leave;
1465         }
1466
1467         spin_lock(&dlm_domain_lock);
1468
1469         dlm = __dlm_lookup_domain(domain);
1470         if (dlm) {
1471                 if (dlm->dlm_state != DLM_CTXT_JOINED) {
1472                         spin_unlock(&dlm_domain_lock);
1473
1474                         mlog(0, "This ctxt is not joined yet!\n");
1475                         wait_event_interruptible(dlm_domain_events,
1476                                                  dlm_wait_on_domain_helper(
1477                                                          domain));
1478                         goto retry;
1479                 }
1480
1481                 __dlm_get(dlm);
1482                 dlm->num_joins++;
1483
1484                 spin_unlock(&dlm_domain_lock);
1485
1486                 ret = 0;
1487                 goto leave;
1488         }
1489
1490         /* doesn't exist */
1491         if (!new_ctxt) {
1492                 spin_unlock(&dlm_domain_lock);
1493
1494                 new_ctxt = dlm_alloc_ctxt(domain, key);
1495                 if (new_ctxt)
1496                         goto retry;
1497
1498                 ret = -ENOMEM;
1499                 mlog_errno(ret);
1500                 goto leave;
1501         }
1502
1503         /* a little variable switch-a-roo here... */
1504         dlm = new_ctxt;
1505         new_ctxt = NULL;
1506
1507         /* add the new domain */
1508         list_add_tail(&dlm->list, &dlm_domains);
1509         spin_unlock(&dlm_domain_lock);
1510
1511         ret = dlm_join_domain(dlm);
1512         if (ret) {
1513                 mlog_errno(ret);
1514                 dlm_put(dlm);
1515                 goto leave;
1516         }
1517
1518         ret = 0;
1519 leave:
1520         if (new_ctxt)
1521                 dlm_free_ctxt_mem(new_ctxt);
1522
1523         if (ret < 0)
1524                 dlm = ERR_PTR(ret);
1525
1526         return dlm;
1527 }
1528 EXPORT_SYMBOL_GPL(dlm_register_domain);
1529
1530 static LIST_HEAD(dlm_join_handlers);
1531
1532 static void dlm_unregister_net_handlers(void)
1533 {
1534         o2net_unregister_handler_list(&dlm_join_handlers);
1535 }
1536
1537 static int dlm_register_net_handlers(void)
1538 {
1539         int status = 0;
1540
1541         status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1542                                         sizeof(struct dlm_query_join_request),
1543                                         dlm_query_join_handler,
1544                                         NULL, NULL, &dlm_join_handlers);
1545         if (status)
1546                 goto bail;
1547
1548         status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1549                                         sizeof(struct dlm_assert_joined),
1550                                         dlm_assert_joined_handler,
1551                                         NULL, NULL, &dlm_join_handlers);
1552         if (status)
1553                 goto bail;
1554
1555         status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1556                                         sizeof(struct dlm_cancel_join),
1557                                         dlm_cancel_join_handler,
1558                                         NULL, NULL, &dlm_join_handlers);
1559
1560 bail:
1561         if (status < 0)
1562                 dlm_unregister_net_handlers();
1563
1564         return status;
1565 }
1566
1567 /* Domain eviction callback handling.
1568  *
1569  * The file system requires notification of node death *before* the
1570  * dlm completes it's recovery work, otherwise it may be able to
1571  * acquire locks on resources requiring recovery. Since the dlm can
1572  * evict a node from it's domain *before* heartbeat fires, a similar
1573  * mechanism is required. */
1574
1575 /* Eviction is not expected to happen often, so a per-domain lock is
1576  * not necessary. Eviction callbacks are allowed to sleep for short
1577  * periods of time. */
1578 static DECLARE_RWSEM(dlm_callback_sem);
1579
1580 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
1581                                         int node_num)
1582 {
1583         struct list_head *iter;
1584         struct dlm_eviction_cb *cb;
1585
1586         down_read(&dlm_callback_sem);
1587         list_for_each(iter, &dlm->dlm_eviction_callbacks) {
1588                 cb = list_entry(iter, struct dlm_eviction_cb, ec_item);
1589
1590                 cb->ec_func(node_num, cb->ec_data);
1591         }
1592         up_read(&dlm_callback_sem);
1593 }
1594
1595 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
1596                            dlm_eviction_func *f,
1597                            void *data)
1598 {
1599         INIT_LIST_HEAD(&cb->ec_item);
1600         cb->ec_func = f;
1601         cb->ec_data = data;
1602 }
1603 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
1604
1605 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
1606                               struct dlm_eviction_cb *cb)
1607 {
1608         down_write(&dlm_callback_sem);
1609         list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
1610         up_write(&dlm_callback_sem);
1611 }
1612 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
1613
1614 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
1615 {
1616         down_write(&dlm_callback_sem);
1617         list_del_init(&cb->ec_item);
1618         up_write(&dlm_callback_sem);
1619 }
1620 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
1621
1622 static int __init dlm_init(void)
1623 {
1624         int status;
1625
1626         dlm_print_version();
1627
1628         status = dlm_init_mle_cache();
1629         if (status)
1630                 return -1;
1631
1632         status = dlm_register_net_handlers();
1633         if (status) {
1634                 dlm_destroy_mle_cache();
1635                 return -1;
1636         }
1637
1638         return 0;
1639 }
1640
1641 static void __exit dlm_exit (void)
1642 {
1643         dlm_unregister_net_handlers();
1644         dlm_destroy_mle_cache();
1645 }
1646
1647 MODULE_AUTHOR("Oracle");
1648 MODULE_LICENSE("GPL");
1649
1650 module_init(dlm_init);
1651 module_exit(dlm_exit);