fs/ocfs2/dlm/dlmrecovery.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmrecovery.c
   5  *
   6  * recovery stuff
   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
  28 #include <linux/module.h>
  29 #include <linux/fs.h>
  30 #include <linux/types.h>
  31 #include <linux/slab.h>
  32 #include <linux/highmem.h>
  33 #include <linux/init.h>
  34 #include <linux/sysctl.h>
  35 #include <linux/random.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/socket.h>
  38 #include <linux/inet.h>
  39 #include <linux/timer.h>
  40 #include <linux/kthread.h>
  41 #include <linux/delay.h>
  42
  43
  44 #include "cluster/heartbeat.h"
  45 #include "cluster/nodemanager.h"
  46 #include "cluster/tcp.h"
  47
  48 #include "dlmapi.h"
  49 #include "dlmcommon.h"
  50 #include "dlmdomain.h"
  51
  52 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY)
  53 #include "cluster/masklog.h"
  54
  55 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node);
  56
  57 static int dlm_recovery_thread(void *data);
  58 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
  59 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
  60 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm);
  61 static int dlm_do_recovery(struct dlm_ctxt *dlm);
  62
  63 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm);
  64 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node);
  65 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
  66 static int dlm_request_all_locks(struct dlm_ctxt *dlm,
  67                                  u8 request_from, u8 dead_node);
  68 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
  69
  70 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res);
  71 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
  72                                         const char *lockname, int namelen,
  73                                         int total_locks, u64 cookie,
  74                                         u8 flags, u8 master);
  75 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
  76                                     struct dlm_migratable_lockres *mres,
  77                                     u8 send_to,
  78                                     struct dlm_lock_resource *res,
  79                                     int total_locks);
  80 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
  81                                      struct dlm_lock_resource *res,
  82                                      struct dlm_migratable_lockres *mres);
  83 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm);
  84 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm,
  85                                  u8 dead_node, u8 send_to);
  86 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node);
  87 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
  88                                         struct list_head *list, u8 dead_node);
  89 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
  90                                               u8 dead_node, u8 new_master);
  91 static void dlm_reco_ast(void *astdata);
  92 static void dlm_reco_bast(void *astdata, int blocked_type);
  93 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st);
  94 static void dlm_request_all_locks_worker(struct dlm_work_item *item,
  95                                          void *data);
  96 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data);
  97 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
  98                                       struct dlm_lock_resource *res,
  99                                       u8 *real_master);
 100
 101 static u64 dlm_get_next_mig_cookie(void);
 102
 103 static DEFINE_SPINLOCK(dlm_reco_state_lock);
 104 static DEFINE_SPINLOCK(dlm_mig_cookie_lock);
 105 static u64 dlm_mig_cookie = 1;
 106
 107 static u64 dlm_get_next_mig_cookie(void)
 108 {
 109         u64 c;
 110         spin_lock(&dlm_mig_cookie_lock);
 111         c = dlm_mig_cookie;
 112         if (dlm_mig_cookie == (~0ULL))
 113                 dlm_mig_cookie = 1;
 114         else
 115                 dlm_mig_cookie++;
 116         spin_unlock(&dlm_mig_cookie_lock);
 117         return c;
 118 }
 119
 120 static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm,
 121                                           u8 dead_node)
 122 {
 123         assert_spin_locked(&dlm->spinlock);
 124         if (dlm->reco.dead_node != dead_node)
 125                 mlog(0, "%s: changing dead_node from %u to %u\n",
 126                      dlm->name, dlm->reco.dead_node, dead_node);
 127         dlm->reco.dead_node = dead_node;
 128 }
 129
 130 static inline void dlm_set_reco_master(struct dlm_ctxt *dlm,
 131                                        u8 master)
 132 {
 133         assert_spin_locked(&dlm->spinlock);
 134         mlog(0, "%s: changing new_master from %u to %u\n",
 135              dlm->name, dlm->reco.new_master, master);
 136         dlm->reco.new_master = master;
 137 }
 138
 139 static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm)
 140 {
 141         assert_spin_locked(&dlm->spinlock);
 142         clear_bit(dlm->reco.dead_node, dlm->recovery_map);
 143         dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
 144         dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
 145 }
 146
 147 static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
 148 {
 149         spin_lock(&dlm->spinlock);
 150         __dlm_reset_recovery(dlm);
 151         spin_unlock(&dlm->spinlock);
 152 }
 153
 154 /* Worker function used during recovery. */
 155 void dlm_dispatch_work(struct work_struct *work)
 156 {
 157         struct dlm_ctxt *dlm =
 158                 container_of(work, struct dlm_ctxt, dispatched_work);
 159         LIST_HEAD(tmp_list);
 160         struct dlm_work_item *item, *next;
 161         dlm_workfunc_t *workfunc;
 162         int tot=0;
 163
 164         spin_lock(&dlm->work_lock);
 165         list_splice_init(&dlm->work_list, &tmp_list);
 166         spin_unlock(&dlm->work_lock);
 167
 168         list_for_each_entry(item, &tmp_list, list) {
 169                 tot++;
 170         }
 171         mlog(0, "%s: work thread has %d work items\n", dlm->name, tot);
 172
 173         list_for_each_entry_safe(item, next, &tmp_list, list) {
 174                 workfunc = item->func;
 175                 list_del_init(&item->list);
 176
 177                 /* already have ref on dlm to avoid having
 178                  * it disappear.  just double-check. */
 179                 BUG_ON(item->dlm != dlm);
 180
 181                 /* this is allowed to sleep and
 182                  * call network stuff */
 183                 workfunc(item, item->data);
 184
 185                 dlm_put(dlm);
 186                 kfree(item);
 187         }
 188 }
 189
 190 /*
 191  * RECOVERY THREAD
 192  */
 193
 194 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
 195 {
 196         /* wake the recovery thread
 197          * this will wake the reco thread in one of three places
 198          * 1) sleeping with no recovery happening
 199          * 2) sleeping with recovery mastered elsewhere
 200          * 3) recovery mastered here, waiting on reco data */
 201
 202         wake_up(&dlm->dlm_reco_thread_wq);
 203 }
 204
 205 /* Launch the recovery thread */
 206 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
 207 {
 208         mlog(0, "starting dlm recovery thread...\n");
 209
 210         dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
 211                                                 "dlm_reco_thread");
 212         if (IS_ERR(dlm->dlm_reco_thread_task)) {
 213                 mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
 214                 dlm->dlm_reco_thread_task = NULL;
 215                 return -EINVAL;
 216         }
 217
 218         return 0;
 219 }
 220
 221 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
 222 {
 223         if (dlm->dlm_reco_thread_task) {
 224                 mlog(0, "waiting for dlm recovery thread to exit\n");
 225                 kthread_stop(dlm->dlm_reco_thread_task);
 226                 dlm->dlm_reco_thread_task = NULL;
 227         }
 228 }
 229
 230
 231
 232 /*
 233  * this is lame, but here's how recovery works...
 234  * 1) all recovery threads cluster wide will work on recovering
 235  *    ONE node at a time
 236  * 2) negotiate who will take over all the locks for the dead node.
 237  *    thats right... ALL the locks.
 238  * 3) once a new master is chosen, everyone scans all locks
 239  *    and moves aside those mastered by the dead guy
 240  * 4) each of these locks should be locked until recovery is done
 241  * 5) the new master collects up all of secondary lock queue info
 242  *    one lock at a time, forcing each node to communicate back
 243  *    before continuing
 244  * 6) each secondary lock queue responds with the full known lock info
 245  * 7) once the new master has run all its locks, it sends a ALLDONE!
 246  *    message to everyone
 247  * 8) upon receiving this message, the secondary queue node unlocks
 248  *    and responds to the ALLDONE
 249  * 9) once the new master gets responses from everyone, he unlocks
 250  *    everything and recovery for this dead node is done
 251  *10) go back to 2) while there are still dead nodes
 252  *
 253  */
 254
 255 static void dlm_print_reco_node_status(struct dlm_ctxt *dlm)
 256 {
 257         struct dlm_reco_node_data *ndata;
 258         struct dlm_lock_resource *res;
 259
 260         mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n",
 261              dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
 262              dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive",
 263              dlm->reco.dead_node, dlm->reco.new_master);
 264
 265         list_for_each_entry(ndata, &dlm->reco.node_data, list) {
 266                 char *st = "unknown";
 267                 switch (ndata->state) {
 268                         case DLM_RECO_NODE_DATA_INIT:
 269                                 st = "init";
 270                                 break;
 271                         case DLM_RECO_NODE_DATA_REQUESTING:
 272                                 st = "requesting";
 273                                 break;
 274                         case DLM_RECO_NODE_DATA_DEAD:
 275                                 st = "dead";
 276                                 break;
 277                         case DLM_RECO_NODE_DATA_RECEIVING:
 278                                 st = "receiving";
 279                                 break;
 280                         case DLM_RECO_NODE_DATA_REQUESTED:
 281                                 st = "requested";
 282                                 break;
 283                         case DLM_RECO_NODE_DATA_DONE:
 284                                 st = "done";
 285                                 break;
 286                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
 287                                 st = "finalize-sent";
 288                                 break;
 289                         default:
 290                                 st = "bad";
 291                                 break;
 292                 }
 293                 mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n",
 294                      dlm->name, ndata->node_num, st);
 295         }
 296         list_for_each_entry(res, &dlm->reco.resources, recovering) {
 297                 mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n",
 298                      dlm->name, res->lockname.len, res->lockname.name);
 299         }
 300 }
 301
 302 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
 303
 304 static int dlm_recovery_thread(void *data)
 305 {
 306         int status;
 307         struct dlm_ctxt *dlm = data;
 308         unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
 309
 310         mlog(0, "dlm thread running for %s...\n", dlm->name);
 311
 312         while (!kthread_should_stop()) {
 313                 if (dlm_domain_fully_joined(dlm)) {
 314                         status = dlm_do_recovery(dlm);
 315                         if (status == -EAGAIN) {
 316                                 /* do not sleep, recheck immediately. */
 317                                 continue;
 318                         }
 319                         if (status < 0)
 320                                 mlog_errno(status);
 321                 }
 322
 323                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
 324                                                  kthread_should_stop(),
 325                                                  timeout);
 326         }
 327
 328         mlog(0, "quitting DLM recovery thread\n");
 329         return 0;
 330 }
 331
 332 /* returns true when the recovery master has contacted us */
 333 static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
 334 {
 335         int ready;
 336         spin_lock(&dlm->spinlock);
 337         ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
 338         spin_unlock(&dlm->spinlock);
 339         return ready;
 340 }
 341
 342 /* returns true if node is no longer in the domain
 343  * could be dead or just not joined */
 344 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
 345 {
 346         int dead;
 347         spin_lock(&dlm->spinlock);
 348         dead = !test_bit(node, dlm->domain_map);
 349         spin_unlock(&dlm->spinlock);
 350         return dead;
 351 }
 352
 353 /* returns true if node is no longer in the domain
 354  * could be dead or just not joined */
 355 static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node)
 356 {
 357         int recovered;
 358         spin_lock(&dlm->spinlock);
 359         recovered = !test_bit(node, dlm->recovery_map);
 360         spin_unlock(&dlm->spinlock);
 361         return recovered;
 362 }
 363
 364
 365 void dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout)
 366 {
 367         if (dlm_is_node_dead(dlm, node))
 368                 return;
 369
 370         printk(KERN_NOTICE "o2dlm: Waiting on the death of node %u in "
 371                "domain %s\n", node, dlm->name);
 372
 373         if (timeout)
 374                 wait_event_timeout(dlm->dlm_reco_thread_wq,
 375                                    dlm_is_node_dead(dlm, node),
 376                                    msecs_to_jiffies(timeout));
 377         else
 378                 wait_event(dlm->dlm_reco_thread_wq,
 379                            dlm_is_node_dead(dlm, node));
 380 }
 381
 382 void dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout)
 383 {
 384         if (dlm_is_node_recovered(dlm, node))
 385                 return;
 386
 387         printk(KERN_NOTICE "o2dlm: Waiting on the recovery of node %u in "
 388                "domain %s\n", node, dlm->name);
 389
 390         if (timeout)
 391                 wait_event_timeout(dlm->dlm_reco_thread_wq,
 392                                    dlm_is_node_recovered(dlm, node),
 393                                    msecs_to_jiffies(timeout));
 394         else
 395                 wait_event(dlm->dlm_reco_thread_wq,
 396                            dlm_is_node_recovered(dlm, node));
 397 }
 398
 399 /* callers of the top-level api calls (dlmlock/dlmunlock) should
 400  * block on the dlm->reco.event when recovery is in progress.
 401  * the dlm recovery thread will set this state when it begins
 402  * recovering a dead node (as the new master or not) and clear
 403  * the state and wake as soon as all affected lock resources have
 404  * been marked with the RECOVERY flag */
 405 static int dlm_in_recovery(struct dlm_ctxt *dlm)
 406 {
 407         int in_recovery;
 408         spin_lock(&dlm->spinlock);
 409         in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
 410         spin_unlock(&dlm->spinlock);
 411         return in_recovery;
 412 }
 413
 414
 415 void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
 416 {
 417         if (dlm_in_recovery(dlm)) {
 418                 mlog(0, "%s: reco thread %d in recovery: "
 419                      "state=%d, master=%u, dead=%u\n",
 420                      dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
 421                      dlm->reco.state, dlm->reco.new_master,
 422                      dlm->reco.dead_node);
 423         }
 424         wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
 425 }
 426
 427 static void dlm_begin_recovery(struct dlm_ctxt *dlm)
 428 {
 429         spin_lock(&dlm->spinlock);
 430         BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
 431         printk(KERN_NOTICE "o2dlm: Begin recovery on domain %s for node %u\n",
 432                dlm->name, dlm->reco.dead_node);
 433         dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
 434         spin_unlock(&dlm->spinlock);
 435 }
 436
 437 static void dlm_end_recovery(struct dlm_ctxt *dlm)
 438 {
 439         spin_lock(&dlm->spinlock);
 440         BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
 441         dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
 442         spin_unlock(&dlm->spinlock);
 443         printk(KERN_NOTICE "o2dlm: End recovery on domain %s\n", dlm->name);
 444         wake_up(&dlm->reco.event);
 445 }
 446
 447 static void dlm_print_recovery_master(struct dlm_ctxt *dlm)
 448 {
 449         printk(KERN_NOTICE "o2dlm: Node %u (%s) is the Recovery Master for the "
 450                "dead node %u in domain %s\n", dlm->reco.new_master,
 451                (dlm->node_num == dlm->reco.new_master ? "me" : "he"),
 452                dlm->reco.dead_node, dlm->name);
 453 }
 454
 455 static int dlm_do_recovery(struct dlm_ctxt *dlm)
 456 {
 457         int status = 0;
 458         int ret;
 459
 460         spin_lock(&dlm->spinlock);
 461
 462         /* check to see if the new master has died */
 463         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
 464             test_bit(dlm->reco.new_master, dlm->recovery_map)) {
 465                 mlog(0, "new master %u died while recovering %u!\n",
 466                      dlm->reco.new_master, dlm->reco.dead_node);
 467                 /* unset the new_master, leave dead_node */
 468                 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
 469         }
 470
 471         /* select a target to recover */
 472         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
 473                 int bit;
 474
 475                 bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES, 0);
 476                 if (bit >= O2NM_MAX_NODES || bit < 0)
 477                         dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
 478                 else
 479                         dlm_set_reco_dead_node(dlm, bit);
 480         } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
 481                 /* BUG? */
 482                 mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
 483                      dlm->reco.dead_node);
 484                 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
 485         }
 486
 487         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
 488                 // mlog(0, "nothing to recover!  sleeping now!\n");
 489                 spin_unlock(&dlm->spinlock);
 490                 /* return to main thread loop and sleep. */
 491                 return 0;
 492         }
 493         mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n",
 494              dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
 495              dlm->reco.dead_node);
 496         spin_unlock(&dlm->spinlock);
 497
 498         /* take write barrier */
 499         /* (stops the list reshuffling thread, proxy ast handling) */
 500         dlm_begin_recovery(dlm);
 501
 502         if (dlm->reco.new_master == dlm->node_num)
 503                 goto master_here;
 504
 505         if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
 506                 /* choose a new master, returns 0 if this node
 507                  * is the master, -EEXIST if it's another node.
 508                  * this does not return until a new master is chosen
 509                  * or recovery completes entirely. */
 510                 ret = dlm_pick_recovery_master(dlm);
 511                 if (!ret) {
 512                         /* already notified everyone.  go. */
 513                         goto master_here;
 514                 }
 515                 mlog(0, "another node will master this recovery session.\n");
 516         }
 517
 518         dlm_print_recovery_master(dlm);
 519
 520         /* it is safe to start everything back up here
 521          * because all of the dead node's lock resources
 522          * have been marked as in-recovery */
 523         dlm_end_recovery(dlm);
 524
 525         /* sleep out in main dlm_recovery_thread loop. */
 526         return 0;
 527
 528 master_here:
 529         dlm_print_recovery_master(dlm);
 530
 531         status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
 532         if (status < 0) {
 533                 /* we should never hit this anymore */
 534                 mlog(ML_ERROR, "%s: Error %d remastering locks for node %u, "
 535                      "retrying.\n", dlm->name, status, dlm->reco.dead_node);
 536                 /* yield a bit to allow any final network messages
 537                  * to get handled on remaining nodes */
 538                 msleep(100);
 539         } else {
 540                 /* success!  see if any other nodes need recovery */
 541                 mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
 542                      dlm->name, dlm->reco.dead_node, dlm->node_num);
 543                 spin_lock(&dlm->spinlock);
 544                 __dlm_reset_recovery(dlm);
 545                 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
 546                 spin_unlock(&dlm->spinlock);
 547         }
 548         dlm_end_recovery(dlm);
 549
 550         /* continue and look for another dead node */
 551         return -EAGAIN;
 552 }
 553
 554 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
 555 {
 556         int status = 0;
 557         struct dlm_reco_node_data *ndata;
 558         int all_nodes_done;
 559         int destroy = 0;
 560         int pass = 0;
 561
 562         do {
 563                 /* we have become recovery master.  there is no escaping
 564                  * this, so just keep trying until we get it. */
 565                 status = dlm_init_recovery_area(dlm, dead_node);
 566                 if (status < 0) {
 567                         mlog(ML_ERROR, "%s: failed to alloc recovery area, "
 568                              "retrying\n", dlm->name);
 569                         msleep(1000);
 570                 }
 571         } while (status != 0);
 572
 573         /* safe to access the node data list without a lock, since this
 574          * process is the only one to change the list */
 575         list_for_each_entry(ndata, &dlm->reco.node_data, list) {
 576                 BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
 577                 ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
 578
 579                 mlog(0, "%s: Requesting lock info from node %u\n", dlm->name,
 580                      ndata->node_num);
 581
 582                 if (ndata->node_num == dlm->node_num) {
 583                         ndata->state = DLM_RECO_NODE_DATA_DONE;
 584                         continue;
 585                 }
 586
 587                 do {
 588                         status = dlm_request_all_locks(dlm, ndata->node_num,
 589                                                        dead_node);
 590                         if (status < 0) {
 591                                 mlog_errno(status);
 592                                 if (dlm_is_host_down(status)) {
 593                                         /* node died, ignore it for recovery */
 594                                         status = 0;
 595                                         ndata->state = DLM_RECO_NODE_DATA_DEAD;
 596                                         /* wait for the domain map to catch up
 597                                          * with the network state. */
 598                                         wait_event_timeout(dlm->dlm_reco_thread_wq,
 599                                                            dlm_is_node_dead(dlm,
 600                                                                 ndata->node_num),
 601                                                            msecs_to_jiffies(1000));
 602                                         mlog(0, "waited 1 sec for %u, "
 603                                              "dead? %s\n", ndata->node_num,
 604                                              dlm_is_node_dead(dlm, ndata->node_num) ?
 605                                              "yes" : "no");
 606                                 } else {
 607                                         /* -ENOMEM on the other node */
 608                                         mlog(0, "%s: node %u returned "
 609                                              "%d during recovery, retrying "
 610                                              "after a short wait\n",
 611                                              dlm->name, ndata->node_num,
 612                                              status);
 613                                         msleep(100);
 614                                 }
 615                         }
 616                 } while (status != 0);
 617
 618                 spin_lock(&dlm_reco_state_lock);
 619                 switch (ndata->state) {
 620                         case DLM_RECO_NODE_DATA_INIT:
 621                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
 622                         case DLM_RECO_NODE_DATA_REQUESTED:
 623                                 BUG();
 624                                 break;
 625                         case DLM_RECO_NODE_DATA_DEAD:
 626                                 mlog(0, "node %u died after requesting "
 627                                      "recovery info for node %u\n",
 628                                      ndata->node_num, dead_node);
 629                                 /* fine.  don't need this node's info.
 630                                  * continue without it. */
 631                                 break;
 632                         case DLM_RECO_NODE_DATA_REQUESTING:
 633                                 ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
 634                                 mlog(0, "now receiving recovery data from "
 635                                      "node %u for dead node %u\n",
 636                                      ndata->node_num, dead_node);
 637                                 break;
 638                         case DLM_RECO_NODE_DATA_RECEIVING:
 639                                 mlog(0, "already receiving recovery data from "
 640                                      "node %u for dead node %u\n",
 641                                      ndata->node_num, dead_node);
 642                                 break;
 643                         case DLM_RECO_NODE_DATA_DONE:
 644                                 mlog(0, "already DONE receiving recovery data "
 645                                      "from node %u for dead node %u\n",
 646                                      ndata->node_num, dead_node);
 647                                 break;
 648                 }
 649                 spin_unlock(&dlm_reco_state_lock);
 650         }
 651
 652         mlog(0, "%s: Done requesting all lock info\n", dlm->name);
 653
 654         /* nodes should be sending reco data now
 655          * just need to wait */
 656
 657         while (1) {
 658                 /* check all the nodes now to see if we are
 659                  * done, or if anyone died */
 660                 all_nodes_done = 1;
 661                 spin_lock(&dlm_reco_state_lock);
 662                 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
 663                         mlog(0, "checking recovery state of node %u\n",
 664                              ndata->node_num);
 665                         switch (ndata->state) {
 666                                 case DLM_RECO_NODE_DATA_INIT:
 667                                 case DLM_RECO_NODE_DATA_REQUESTING:
 668                                         mlog(ML_ERROR, "bad ndata state for "
 669                                              "node %u: state=%d\n",
 670                                              ndata->node_num, ndata->state);
 671                                         BUG();
 672                                         break;
 673                                 case DLM_RECO_NODE_DATA_DEAD:
 674                                         mlog(0, "node %u died after "
 675                                              "requesting recovery info for "
 676                                              "node %u\n", ndata->node_num,
 677                                              dead_node);
 678                                         break;
 679                                 case DLM_RECO_NODE_DATA_RECEIVING:
 680                                 case DLM_RECO_NODE_DATA_REQUESTED:
 681                                         mlog(0, "%s: node %u still in state %s\n",
 682                                              dlm->name, ndata->node_num,
 683                                              ndata->state==DLM_RECO_NODE_DATA_RECEIVING ?
 684                                              "receiving" : "requested");
 685                                         all_nodes_done = 0;
 686                                         break;
 687                                 case DLM_RECO_NODE_DATA_DONE:
 688                                         mlog(0, "%s: node %u state is done\n",
 689                                              dlm->name, ndata->node_num);
 690                                         break;
 691                                 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
 692                                         mlog(0, "%s: node %u state is finalize\n",
 693                                              dlm->name, ndata->node_num);
 694                                         break;
 695                         }
 696                 }
 697                 spin_unlock(&dlm_reco_state_lock);
 698
 699                 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
 700                      all_nodes_done?"yes":"no");
 701                 if (all_nodes_done) {
 702                         int ret;
 703
 704                         /* Set this flag on recovery master to avoid
 705                          * a new recovery for another dead node start
 706                          * before the recovery is not done. That may
 707                          * cause recovery hung.*/
 708                         spin_lock(&dlm->spinlock);
 709                         dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
 710                         spin_unlock(&dlm->spinlock);
 711
 712                         /* all nodes are now in DLM_RECO_NODE_DATA_DONE state
 713                          * just send a finalize message to everyone and
 714                          * clean up */
 715                         mlog(0, "all nodes are done! send finalize\n");
 716                         ret = dlm_send_finalize_reco_message(dlm);
 717                         if (ret < 0)
 718                                 mlog_errno(ret);
 719
 720                         spin_lock(&dlm->spinlock);
 721                         dlm_finish_local_lockres_recovery(dlm, dead_node,
 722                                                           dlm->node_num);
 723                         spin_unlock(&dlm->spinlock);
 724                         mlog(0, "should be done with recovery!\n");
 725
 726                         mlog(0, "finishing recovery of %s at %lu, "
 727                              "dead=%u, this=%u, new=%u\n", dlm->name,
 728                              jiffies, dlm->reco.dead_node,
 729                              dlm->node_num, dlm->reco.new_master);
 730                         destroy = 1;
 731                         status = 0;
 732                         /* rescan everything marked dirty along the way */
 733                         dlm_kick_thread(dlm, NULL);
 734                         break;
 735                 }
 736                 /* wait to be signalled, with periodic timeout
 737                  * to check for node death */
 738                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
 739                                          kthread_should_stop(),
 740                                          msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
 741
 742         }
 743
 744         if (destroy)
 745                 dlm_destroy_recovery_area(dlm, dead_node);
 746
 747         return status;
 748 }
 749
 750 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
 751 {
 752         int num=0;
 753         struct dlm_reco_node_data *ndata;
 754
 755         spin_lock(&dlm->spinlock);
 756         memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
 757         /* nodes can only be removed (by dying) after dropping
 758          * this lock, and death will be trapped later, so this should do */
 759         spin_unlock(&dlm->spinlock);
 760
 761         while (1) {
 762                 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
 763                 if (num >= O2NM_MAX_NODES) {
 764                         break;
 765                 }
 766                 BUG_ON(num == dead_node);
 767
 768                 ndata = kzalloc(sizeof(*ndata), GFP_NOFS);
 769                 if (!ndata) {
 770                         dlm_destroy_recovery_area(dlm, dead_node);
 771                         return -ENOMEM;
 772                 }
 773                 ndata->node_num = num;
 774                 ndata->state = DLM_RECO_NODE_DATA_INIT;
 775                 spin_lock(&dlm_reco_state_lock);
 776                 list_add_tail(&ndata->list, &dlm->reco.node_data);
 777                 spin_unlock(&dlm_reco_state_lock);
 778                 num++;
 779         }
 780
 781         return 0;
 782 }
 783
 784 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
 785 {
 786         struct dlm_reco_node_data *ndata, *next;
 787         LIST_HEAD(tmplist);
 788
 789         spin_lock(&dlm_reco_state_lock);
 790         list_splice_init(&dlm->reco.node_data, &tmplist);
 791         spin_unlock(&dlm_reco_state_lock);
 792
 793         list_for_each_entry_safe(ndata, next, &tmplist, list) {
 794                 list_del_init(&ndata->list);
 795                 kfree(ndata);
 796         }
 797 }
 798
 799 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
 800                                  u8 dead_node)
 801 {
 802         struct dlm_lock_request lr;
 803         enum dlm_status ret;
 804
 805         mlog(0, "\n");
 806
 807
 808         mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
 809                   "to %u\n", dead_node, request_from);
 810
 811         memset(&lr, 0, sizeof(lr));
 812         lr.node_idx = dlm->node_num;
 813         lr.dead_node = dead_node;
 814
 815         // send message
 816         ret = DLM_NOLOCKMGR;
 817         ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
 818                                  &lr, sizeof(lr), request_from, NULL);
 819
 820         /* negative status is handled by caller */
 821         if (ret < 0)
 822                 mlog(ML_ERROR, "%s: Error %d send LOCK_REQUEST to node %u "
 823                      "to recover dead node %u\n", dlm->name, ret,
 824                      request_from, dead_node);
 825         // return from here, then
 826         // sleep until all received or error
 827         return ret;
 828
 829 }
 830
 831 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
 832                                   void **ret_data)
 833 {
 834         struct dlm_ctxt *dlm = data;
 835         struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
 836         char *buf = NULL;
 837         struct dlm_work_item *item = NULL;
 838
 839         if (!dlm_grab(dlm))
 840                 return -EINVAL;
 841
 842         if (lr->dead_node != dlm->reco.dead_node) {
 843                 mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local "
 844                      "dead_node is %u\n", dlm->name, lr->node_idx,
 845                      lr->dead_node, dlm->reco.dead_node);
 846                 dlm_print_reco_node_status(dlm);
 847                 /* this is a hack */
 848                 dlm_put(dlm);
 849                 return -ENOMEM;
 850         }
 851         BUG_ON(lr->dead_node != dlm->reco.dead_node);
 852
 853         item = kzalloc(sizeof(*item), GFP_NOFS);
 854         if (!item) {
 855                 dlm_put(dlm);
 856                 return -ENOMEM;
 857         }
 858
 859         /* this will get freed by dlm_request_all_locks_worker */
 860         buf = (char *) __get_free_page(GFP_NOFS);
 861         if (!buf) {
 862                 kfree(item);
 863                 dlm_put(dlm);
 864                 return -ENOMEM;
 865         }
 866
 867         /* queue up work for dlm_request_all_locks_worker */
 868         dlm_grab(dlm);  /* get an extra ref for the work item */
 869         dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
 870         item->u.ral.reco_master = lr->node_idx;
 871         item->u.ral.dead_node = lr->dead_node;
 872         spin_lock(&dlm->work_lock);
 873         list_add_tail(&item->list, &dlm->work_list);
 874         spin_unlock(&dlm->work_lock);
 875         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
 876
 877         dlm_put(dlm);
 878         return 0;
 879 }
 880
 881 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
 882 {
 883         struct dlm_migratable_lockres *mres;
 884         struct dlm_lock_resource *res;
 885         struct dlm_ctxt *dlm;
 886         LIST_HEAD(resources);
 887         int ret;
 888         u8 dead_node, reco_master;
 889         int skip_all_done = 0;
 890
 891         dlm = item->dlm;
 892         dead_node = item->u.ral.dead_node;
 893         reco_master = item->u.ral.reco_master;
 894         mres = (struct dlm_migratable_lockres *)data;
 895
 896         mlog(0, "%s: recovery worker started, dead=%u, master=%u\n",
 897              dlm->name, dead_node, reco_master);
 898
 899         if (dead_node != dlm->reco.dead_node ||
 900             reco_master != dlm->reco.new_master) {
 901                 /* worker could have been created before the recovery master
 902                  * died.  if so, do not continue, but do not error. */
 903                 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
 904                         mlog(ML_NOTICE, "%s: will not send recovery state, "
 905                              "recovery master %u died, thread=(dead=%u,mas=%u)"
 906                              " current=(dead=%u,mas=%u)\n", dlm->name,
 907                              reco_master, dead_node, reco_master,
 908                              dlm->reco.dead_node, dlm->reco.new_master);
 909                 } else {
 910                         mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, "
 911                              "master=%u), request(dead=%u, master=%u)\n",
 912                              dlm->name, dlm->reco.dead_node,
 913                              dlm->reco.new_master, dead_node, reco_master);
 914                 }
 915                 goto leave;
 916         }
 917
 918         /* lock resources should have already been moved to the
 919          * dlm->reco.resources list.  now move items from that list
 920          * to a temp list if the dead owner matches.  note that the
 921          * whole cluster recovers only one node at a time, so we
 922          * can safely move UNKNOWN lock resources for each recovery
 923          * session. */
 924         dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
 925
 926         /* now we can begin blasting lockreses without the dlm lock */
 927
 928         /* any errors returned will be due to the new_master dying,
 929          * the dlm_reco_thread should detect this */
 930         list_for_each_entry(res, &resources, recovering) {
 931                 ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
 932                                         DLM_MRES_RECOVERY);
 933                 if (ret < 0) {
 934                         mlog(ML_ERROR, "%s: node %u went down while sending "
 935                              "recovery state for dead node %u, ret=%d\n", dlm->name,
 936                              reco_master, dead_node, ret);
 937                         skip_all_done = 1;
 938                         break;
 939                 }
 940         }
 941
 942         /* move the resources back to the list */
 943         spin_lock(&dlm->spinlock);
 944         list_splice_init(&resources, &dlm->reco.resources);
 945         spin_unlock(&dlm->spinlock);
 946
 947         if (!skip_all_done) {
 948                 ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
 949                 if (ret < 0) {
 950                         mlog(ML_ERROR, "%s: node %u went down while sending "
 951                              "recovery all-done for dead node %u, ret=%d\n",
 952                              dlm->name, reco_master, dead_node, ret);
 953                 }
 954         }
 955 leave:
 956         free_page((unsigned long)data);
 957 }
 958
 959
 960 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
 961 {
 962         int ret, tmpret;
 963         struct dlm_reco_data_done done_msg;
 964
 965         memset(&done_msg, 0, sizeof(done_msg));
 966         done_msg.node_idx = dlm->node_num;
 967         done_msg.dead_node = dead_node;
 968         mlog(0, "sending DATA DONE message to %u, "
 969              "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
 970              done_msg.dead_node);
 971
 972         ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
 973                                  sizeof(done_msg), send_to, &tmpret);
 974         if (ret < 0) {
 975                 mlog(ML_ERROR, "%s: Error %d send RECO_DATA_DONE to node %u "
 976                      "to recover dead node %u\n", dlm->name, ret, send_to,
 977                      dead_node);
 978                 if (!dlm_is_host_down(ret)) {
 979                         BUG();
 980                 }
 981         } else
 982                 ret = tmpret;
 983         return ret;
 984 }
 985
 986
 987 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
 988                                void **ret_data)
 989 {
 990         struct dlm_ctxt *dlm = data;
 991         struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
 992         struct dlm_reco_node_data *ndata = NULL;
 993         int ret = -EINVAL;
 994
 995         if (!dlm_grab(dlm))
 996                 return -EINVAL;
 997
 998         mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
 999              "node_idx=%u, this node=%u\n", done->dead_node,
1000              dlm->reco.dead_node, done->node_idx, dlm->node_num);
1001
1002         mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node),
1003                         "Got DATA DONE: dead_node=%u, reco.dead_node=%u, "
1004                         "node_idx=%u, this node=%u\n", done->dead_node,
1005                         dlm->reco.dead_node, done->node_idx, dlm->node_num);
1006
1007         spin_lock(&dlm_reco_state_lock);
1008         list_for_each_entry(ndata, &dlm->reco.node_data, list) {
1009                 if (ndata->node_num != done->node_idx)
1010                         continue;
1011
1012                 switch (ndata->state) {
1013                         /* should have moved beyond INIT but not to FINALIZE yet */
1014                         case DLM_RECO_NODE_DATA_INIT:
1015                         case DLM_RECO_NODE_DATA_DEAD:
1016                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
1017                                 mlog(ML_ERROR, "bad ndata state for node %u:"
1018                                      " state=%d\n", ndata->node_num,
1019                                      ndata->state);
1020                                 BUG();
1021                                 break;
1022                         /* these states are possible at this point, anywhere along
1023                          * the line of recovery */
1024                         case DLM_RECO_NODE_DATA_DONE:
1025                         case DLM_RECO_NODE_DATA_RECEIVING:
1026                         case DLM_RECO_NODE_DATA_REQUESTED:
1027                         case DLM_RECO_NODE_DATA_REQUESTING:
1028                                 mlog(0, "node %u is DONE sending "
1029                                           "recovery data!\n",
1030                                           ndata->node_num);
1031
1032                                 ndata->state = DLM_RECO_NODE_DATA_DONE;
1033                                 ret = 0;
1034                                 break;
1035                 }
1036         }
1037         spin_unlock(&dlm_reco_state_lock);
1038
1039         /* wake the recovery thread, some node is done */
1040         if (!ret)
1041                 dlm_kick_recovery_thread(dlm);
1042
1043         if (ret < 0)
1044                 mlog(ML_ERROR, "failed to find recovery node data for node "
1045                      "%u\n", done->node_idx);
1046         dlm_put(dlm);
1047
1048         mlog(0, "leaving reco data done handler, ret=%d\n", ret);
1049         return ret;
1050 }
1051
1052 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
1053                                         struct list_head *list,
1054                                         u8 dead_node)
1055 {
1056         struct dlm_lock_resource *res, *next;
1057         struct dlm_lock *lock;
1058
1059         spin_lock(&dlm->spinlock);
1060         list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
1061                 /* always prune any $RECOVERY entries for dead nodes,
1062                  * otherwise hangs can occur during later recovery */
1063                 if (dlm_is_recovery_lock(res->lockname.name,
1064                                          res->lockname.len)) {
1065                         spin_lock(&res->spinlock);
1066                         list_for_each_entry(lock, &res->granted, list) {
1067                                 if (lock->ml.node == dead_node) {
1068                                         mlog(0, "AHA! there was "
1069                                              "a $RECOVERY lock for dead "
1070                                              "node %u (%s)!\n",
1071                                              dead_node, dlm->name);
1072                                         list_del_init(&lock->list);
1073                                         dlm_lock_put(lock);
1074                                         break;
1075                                 }
1076                         }
1077                         spin_unlock(&res->spinlock);
1078                         continue;
1079                 }
1080
1081                 if (res->owner == dead_node) {
1082                         mlog(0, "found lockres owned by dead node while "
1083                                   "doing recovery for node %u. sending it.\n",
1084                                   dead_node);
1085                         list_move_tail(&res->recovering, list);
1086                 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1087                         mlog(0, "found UNKNOWN owner while doing recovery "
1088                                   "for node %u. sending it.\n", dead_node);
1089                         list_move_tail(&res->recovering, list);
1090                 }
1091         }
1092         spin_unlock(&dlm->spinlock);
1093 }
1094
1095 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
1096 {
1097         int total_locks = 0;
1098         struct list_head *iter, *queue = &res->granted;
1099         int i;
1100
1101         for (i=0; i<3; i++) {
1102                 list_for_each(iter, queue)
1103                         total_locks++;
1104                 queue++;
1105         }
1106         return total_locks;
1107 }
1108
1109
1110 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
1111                                       struct dlm_migratable_lockres *mres,
1112                                       u8 send_to,
1113                                       struct dlm_lock_resource *res,
1114                                       int total_locks)
1115 {
1116         u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
1117         int mres_total_locks = be32_to_cpu(mres->total_locks);
1118         int sz, ret = 0, status = 0;
1119         u8 orig_flags = mres->flags,
1120            orig_master = mres->master;
1121
1122         BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
1123         if (!mres->num_locks)
1124                 return 0;
1125
1126         sz = sizeof(struct dlm_migratable_lockres) +
1127                 (mres->num_locks * sizeof(struct dlm_migratable_lock));
1128
1129         /* add an all-done flag if we reached the last lock */
1130         orig_flags = mres->flags;
1131         BUG_ON(total_locks > mres_total_locks);
1132         if (total_locks == mres_total_locks)
1133                 mres->flags |= DLM_MRES_ALL_DONE;
1134
1135         mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n",
1136              dlm->name, res->lockname.len, res->lockname.name,
1137              orig_flags & DLM_MRES_MIGRATION ? "migration" : "recovery",
1138              send_to);
1139
1140         /* send it */
1141         ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
1142                                  sz, send_to, &status);
1143         if (ret < 0) {
1144                 /* XXX: negative status is not handled.
1145                  * this will end up killing this node. */
1146                 mlog(ML_ERROR, "%s: res %.*s, Error %d send MIG_LOCKRES to "
1147                      "node %u (%s)\n", dlm->name, mres->lockname_len,
1148                      mres->lockname, ret, send_to,
1149                      (orig_flags & DLM_MRES_MIGRATION ?
1150                       "migration" : "recovery"));
1151         } else {
1152                 /* might get an -ENOMEM back here */
1153                 ret = status;
1154                 if (ret < 0) {
1155                         mlog_errno(ret);
1156
1157                         if (ret == -EFAULT) {
1158                                 mlog(ML_ERROR, "node %u told me to kill "
1159                                      "myself!\n", send_to);
1160                                 BUG();
1161                         }
1162                 }
1163         }
1164
1165         /* zero and reinit the message buffer */
1166         dlm_init_migratable_lockres(mres, res->lockname.name,
1167                                     res->lockname.len, mres_total_locks,
1168                                     mig_cookie, orig_flags, orig_master);
1169         return ret;
1170 }
1171
1172 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
1173                                         const char *lockname, int namelen,
1174                                         int total_locks, u64 cookie,
1175                                         u8 flags, u8 master)
1176 {
1177         /* mres here is one full page */
1178         clear_page(mres);
1179         mres->lockname_len = namelen;
1180         memcpy(mres->lockname, lockname, namelen);
1181         mres->num_locks = 0;
1182         mres->total_locks = cpu_to_be32(total_locks);
1183         mres->mig_cookie = cpu_to_be64(cookie);
1184         mres->flags = flags;
1185         mres->master = master;
1186 }
1187
1188 static void dlm_prepare_lvb_for_migration(struct dlm_lock *lock,
1189                                           struct dlm_migratable_lockres *mres,
1190                                           int queue)
1191 {
1192         if (!lock->lksb)
1193                return;
1194
1195         /* Ignore lvb in all locks in the blocked list */
1196         if (queue == DLM_BLOCKED_LIST)
1197                 return;
1198
1199         /* Only consider lvbs in locks with granted EX or PR lock levels */
1200         if (lock->ml.type != LKM_EXMODE && lock->ml.type != LKM_PRMODE)
1201                 return;
1202
1203         if (dlm_lvb_is_empty(mres->lvb)) {
1204                 memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1205                 return;
1206         }
1207
1208         /* Ensure the lvb copied for migration matches in other valid locks */
1209         if (!memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))
1210                 return;
1211
1212         mlog(ML_ERROR, "Mismatched lvb in lock cookie=%u:%llu, name=%.*s, "
1213              "node=%u\n",
1214              dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
1215              dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
1216              lock->lockres->lockname.len, lock->lockres->lockname.name,
1217              lock->ml.node);
1218         dlm_print_one_lock_resource(lock->lockres);
1219         BUG();
1220 }
1221
1222 /* returns 1 if this lock fills the network structure,
1223  * 0 otherwise */
1224 static int dlm_add_lock_to_array(struct dlm_lock *lock,
1225                                  struct dlm_migratable_lockres *mres, int queue)
1226 {
1227         struct dlm_migratable_lock *ml;
1228         int lock_num = mres->num_locks;
1229
1230         ml = &(mres->ml[lock_num]);
1231         ml->cookie = lock->ml.cookie;
1232         ml->type = lock->ml.type;
1233         ml->convert_type = lock->ml.convert_type;
1234         ml->highest_blocked = lock->ml.highest_blocked;
1235         ml->list = queue;
1236         if (lock->lksb) {
1237                 ml->flags = lock->lksb->flags;
1238                 dlm_prepare_lvb_for_migration(lock, mres, queue);
1239         }
1240         ml->node = lock->ml.node;
1241         mres->num_locks++;
1242         /* we reached the max, send this network message */
1243         if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1244                 return 1;
1245         return 0;
1246 }
1247
1248 static void dlm_add_dummy_lock(struct dlm_ctxt *dlm,
1249                                struct dlm_migratable_lockres *mres)
1250 {
1251         struct dlm_lock dummy;
1252         memset(&dummy, 0, sizeof(dummy));
1253         dummy.ml.cookie = 0;
1254         dummy.ml.type = LKM_IVMODE;
1255         dummy.ml.convert_type = LKM_IVMODE;
1256         dummy.ml.highest_blocked = LKM_IVMODE;
1257         dummy.lksb = NULL;
1258         dummy.ml.node = dlm->node_num;
1259         dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST);
1260 }
1261
1262 static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm,
1263                                     struct dlm_migratable_lock *ml,
1264                                     u8 *nodenum)
1265 {
1266         if (unlikely(ml->cookie == 0 &&
1267             ml->type == LKM_IVMODE &&
1268             ml->convert_type == LKM_IVMODE &&
1269             ml->highest_blocked == LKM_IVMODE &&
1270             ml->list == DLM_BLOCKED_LIST)) {
1271                 *nodenum = ml->node;
1272                 return 1;
1273         }
1274         return 0;
1275 }
1276
1277 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1278                          struct dlm_migratable_lockres *mres,
1279                          u8 send_to, u8 flags)
1280 {
1281         struct list_head *queue;
1282         int total_locks, i;
1283         u64 mig_cookie = 0;
1284         struct dlm_lock *lock;
1285         int ret = 0;
1286
1287         BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1288
1289         mlog(0, "sending to %u\n", send_to);
1290
1291         total_locks = dlm_num_locks_in_lockres(res);
1292         if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1293                 /* rare, but possible */
1294                 mlog(0, "argh.  lockres has %d locks.  this will "
1295                           "require more than one network packet to "
1296                           "migrate\n", total_locks);
1297                 mig_cookie = dlm_get_next_mig_cookie();
1298         }
1299
1300         dlm_init_migratable_lockres(mres, res->lockname.name,
1301                                     res->lockname.len, total_locks,
1302                                     mig_cookie, flags, res->owner);
1303
1304         total_locks = 0;
1305         for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1306                 queue = dlm_list_idx_to_ptr(res, i);
1307                 list_for_each_entry(lock, queue, list) {
1308                         /* add another lock. */
1309                         total_locks++;
1310                         if (!dlm_add_lock_to_array(lock, mres, i))
1311                                 continue;
1312
1313                         /* this filled the lock message,
1314                          * we must send it immediately. */
1315                         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1316                                                        res, total_locks);
1317                         if (ret < 0)
1318                                 goto error;
1319                 }
1320         }
1321         if (total_locks == 0) {
1322                 /* send a dummy lock to indicate a mastery reference only */
1323                 mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n",
1324                      dlm->name, res->lockname.len, res->lockname.name,
1325                      send_to, flags & DLM_MRES_RECOVERY ? "recovery" :
1326                      "migration");
1327                 dlm_add_dummy_lock(dlm, mres);
1328         }
1329         /* flush any remaining locks */
1330         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1331         if (ret < 0)
1332                 goto error;
1333         return ret;
1334
1335 error:
1336         mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n",
1337              dlm->name, ret);
1338         if (!dlm_is_host_down(ret))
1339                 BUG();
1340         mlog(0, "%s: node %u went down while sending %s "
1341              "lockres %.*s\n", dlm->name, send_to,
1342              flags & DLM_MRES_RECOVERY ?  "recovery" : "migration",
1343              res->lockname.len, res->lockname.name);
1344         return ret;
1345 }
1346
1347
1348
1349 /*
1350  * this message will contain no more than one page worth of
1351  * recovery data, and it will work on only one lockres.
1352  * there may be many locks in this page, and we may need to wait
1353  * for additional packets to complete all the locks (rare, but
1354  * possible).
1355  */
1356 /*
1357  * NOTE: the allocation error cases here are scary
1358  * we really cannot afford to fail an alloc in recovery
1359  * do we spin?  returning an error only delays the problem really
1360  */
1361
1362 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
1363                             void **ret_data)
1364 {
1365         struct dlm_ctxt *dlm = data;
1366         struct dlm_migratable_lockres *mres =
1367                 (struct dlm_migratable_lockres *)msg->buf;
1368         int ret = 0;
1369         u8 real_master;
1370         u8 extra_refs = 0;
1371         char *buf = NULL;
1372         struct dlm_work_item *item = NULL;
1373         struct dlm_lock_resource *res = NULL;
1374
1375         if (!dlm_grab(dlm))
1376                 return -EINVAL;
1377
1378         BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1379
1380         real_master = mres->master;
1381         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1382                 /* cannot migrate a lockres with no master */
1383                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1384         }
1385
1386         mlog(0, "%s message received from node %u\n",
1387                   (mres->flags & DLM_MRES_RECOVERY) ?
1388                   "recovery" : "migration", mres->master);
1389         if (mres->flags & DLM_MRES_ALL_DONE)
1390                 mlog(0, "all done flag.  all lockres data received!\n");
1391
1392         ret = -ENOMEM;
1393         buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS);
1394         item = kzalloc(sizeof(*item), GFP_NOFS);
1395         if (!buf || !item)
1396                 goto leave;
1397
1398         /* lookup the lock to see if we have a secondary queue for this
1399          * already...  just add the locks in and this will have its owner
1400          * and RECOVERY flag changed when it completes. */
1401         res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len);
1402         if (res) {
1403                 /* this will get a ref on res */
1404                 /* mark it as recovering/migrating and hash it */
1405                 spin_lock(&res->spinlock);
1406                 if (mres->flags & DLM_MRES_RECOVERY) {
1407                         res->state |= DLM_LOCK_RES_RECOVERING;
1408                 } else {
1409                         if (res->state & DLM_LOCK_RES_MIGRATING) {
1410                                 /* this is at least the second
1411                                  * lockres message */
1412                                 mlog(0, "lock %.*s is already migrating\n",
1413                                           mres->lockname_len,
1414                                           mres->lockname);
1415                         } else if (res->state & DLM_LOCK_RES_RECOVERING) {
1416                                 /* caller should BUG */
1417                                 mlog(ML_ERROR, "node is attempting to migrate "
1418                                      "lock %.*s, but marked as recovering!\n",
1419                                      mres->lockname_len, mres->lockname);
1420                                 ret = -EFAULT;
1421                                 spin_unlock(&res->spinlock);
1422                                 goto leave;
1423                         }
1424                         res->state |= DLM_LOCK_RES_MIGRATING;
1425                 }
1426                 spin_unlock(&res->spinlock);
1427         } else {
1428                 /* need to allocate, just like if it was
1429                  * mastered here normally  */
1430                 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1431                 if (!res)
1432                         goto leave;
1433
1434                 /* to match the ref that we would have gotten if
1435                  * dlm_lookup_lockres had succeeded */
1436                 dlm_lockres_get(res);
1437
1438                 /* mark it as recovering/migrating and hash it */
1439                 if (mres->flags & DLM_MRES_RECOVERY)
1440                         res->state |= DLM_LOCK_RES_RECOVERING;
1441                 else
1442                         res->state |= DLM_LOCK_RES_MIGRATING;
1443
1444                 spin_lock(&dlm->spinlock);
1445                 __dlm_insert_lockres(dlm, res);
1446                 spin_unlock(&dlm->spinlock);
1447
1448                 /* Add an extra ref for this lock-less lockres lest the
1449                  * dlm_thread purges it before we get the chance to add
1450                  * locks to it */
1451                 dlm_lockres_get(res);
1452
1453                 /* There are three refs that need to be put.
1454                  * 1. Taken above.
1455                  * 2. kref_init in dlm_new_lockres()->dlm_init_lockres().
1456                  * 3. dlm_lookup_lockres()
1457                  * The first one is handled at the end of this function. The
1458                  * other two are handled in the worker thread after locks have
1459                  * been attached. Yes, we don't wait for purge time to match
1460                  * kref_init. The lockres will still have atleast one ref
1461                  * added because it is in the hash __dlm_insert_lockres() */
1462                 extra_refs++;
1463
1464                 /* now that the new lockres is inserted,
1465                  * make it usable by other processes */
1466                 spin_lock(&res->spinlock);
1467                 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1468                 spin_unlock(&res->spinlock);
1469                 wake_up(&res->wq);
1470         }
1471
1472         /* at this point we have allocated everything we need,
1473          * and we have a hashed lockres with an extra ref and
1474          * the proper res->state flags. */
1475         ret = 0;
1476         spin_lock(&res->spinlock);
1477         /* drop this either when master requery finds a different master
1478          * or when a lock is added by the recovery worker */
1479         dlm_lockres_grab_inflight_ref(dlm, res);
1480         if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1481                 /* migration cannot have an unknown master */
1482                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1483                 mlog(0, "recovery has passed me a lockres with an "
1484                           "unknown owner.. will need to requery: "
1485                           "%.*s\n", mres->lockname_len, mres->lockname);
1486         } else {
1487                 /* take a reference now to pin the lockres, drop it
1488                  * when locks are added in the worker */
1489                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
1490         }
1491         spin_unlock(&res->spinlock);
1492
1493         /* queue up work for dlm_mig_lockres_worker */
1494         dlm_grab(dlm);  /* get an extra ref for the work item */
1495         memcpy(buf, msg->buf, be16_to_cpu(msg->data_len));  /* copy the whole message */
1496         dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1497         item->u.ml.lockres = res; /* already have a ref */
1498         item->u.ml.real_master = real_master;
1499         item->u.ml.extra_ref = extra_refs;
1500         spin_lock(&dlm->work_lock);
1501         list_add_tail(&item->list, &dlm->work_list);
1502         spin_unlock(&dlm->work_lock);
1503         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
1504
1505 leave:
1506         /* One extra ref taken needs to be put here */
1507         if (extra_refs)
1508                 dlm_lockres_put(res);
1509
1510         dlm_put(dlm);
1511         if (ret < 0) {
1512                 if (buf)
1513                         kfree(buf);
1514                 if (item)
1515                         kfree(item);
1516                 mlog_errno(ret);
1517         }
1518
1519         return ret;
1520 }
1521
1522
1523 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1524 {
1525         struct dlm_ctxt *dlm;
1526         struct dlm_migratable_lockres *mres;
1527         int ret = 0;
1528         struct dlm_lock_resource *res;
1529         u8 real_master;
1530         u8 extra_ref;
1531
1532         dlm = item->dlm;
1533         mres = (struct dlm_migratable_lockres *)data;
1534
1535         res = item->u.ml.lockres;
1536         real_master = item->u.ml.real_master;
1537         extra_ref = item->u.ml.extra_ref;
1538
1539         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1540                 /* this case is super-rare. only occurs if
1541                  * node death happens during migration. */
1542 again:
1543                 ret = dlm_lockres_master_requery(dlm, res, &real_master);
1544                 if (ret < 0) {
1545                         mlog(0, "dlm_lockres_master_requery ret=%d\n",
1546                                   ret);
1547                         goto again;
1548                 }
1549                 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1550                         mlog(0, "lockres %.*s not claimed.  "
1551                                    "this node will take it.\n",
1552                                    res->lockname.len, res->lockname.name);
1553                 } else {
1554                         spin_lock(&res->spinlock);
1555                         dlm_lockres_drop_inflight_ref(dlm, res);
1556                         spin_unlock(&res->spinlock);
1557                         mlog(0, "master needs to respond to sender "
1558                                   "that node %u still owns %.*s\n",
1559                                   real_master, res->lockname.len,
1560                                   res->lockname.name);
1561                         /* cannot touch this lockres */
1562                         goto leave;
1563                 }
1564         }
1565
1566         ret = dlm_process_recovery_data(dlm, res, mres);
1567         if (ret < 0)
1568                 mlog(0, "dlm_process_recovery_data returned  %d\n", ret);
1569         else
1570                 mlog(0, "dlm_process_recovery_data succeeded\n");
1571
1572         if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1573                            (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1574                 ret = dlm_finish_migration(dlm, res, mres->master);
1575                 if (ret < 0)
1576                         mlog_errno(ret);
1577         }
1578
1579 leave:
1580         /* See comment in dlm_mig_lockres_handler() */
1581         if (res) {
1582                 if (extra_ref)
1583                         dlm_lockres_put(res);
1584                 dlm_lockres_put(res);
1585         }
1586         kfree(data);
1587 }
1588
1589
1590
1591 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1592                                       struct dlm_lock_resource *res,
1593                                       u8 *real_master)
1594 {
1595         struct dlm_node_iter iter;
1596         int nodenum;
1597         int ret = 0;
1598
1599         *real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1600
1601         /* we only reach here if one of the two nodes in a
1602          * migration died while the migration was in progress.
1603          * at this point we need to requery the master.  we
1604          * know that the new_master got as far as creating
1605          * an mle on at least one node, but we do not know
1606          * if any nodes had actually cleared the mle and set
1607          * the master to the new_master.  the old master
1608          * is supposed to set the owner to UNKNOWN in the
1609          * event of a new_master death, so the only possible
1610          * responses that we can get from nodes here are
1611          * that the master is new_master, or that the master
1612          * is UNKNOWN.
1613          * if all nodes come back with UNKNOWN then we know
1614          * the lock needs remastering here.
1615          * if any node comes back with a valid master, check
1616          * to see if that master is the one that we are
1617          * recovering.  if so, then the new_master died and
1618          * we need to remaster this lock.  if not, then the
1619          * new_master survived and that node will respond to
1620          * other nodes about the owner.
1621          * if there is an owner, this node needs to dump this
1622          * lockres and alert the sender that this lockres
1623          * was rejected. */
1624         spin_lock(&dlm->spinlock);
1625         dlm_node_iter_init(dlm->domain_map, &iter);
1626         spin_unlock(&dlm->spinlock);
1627
1628         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1629                 /* do not send to self */
1630                 if (nodenum == dlm->node_num)
1631                         continue;
1632                 ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1633                 if (ret < 0) {
1634                         mlog_errno(ret);
1635                         if (!dlm_is_host_down(ret))
1636                                 BUG();
1637                         /* host is down, so answer for that node would be
1638                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
1639                 }
1640                 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1641                         mlog(0, "lock master is %u\n", *real_master);
1642                         break;
1643                 }
1644         }
1645         return ret;
1646 }
1647
1648
1649 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1650                           u8 nodenum, u8 *real_master)
1651 {
1652         int ret = -EINVAL;
1653         struct dlm_master_requery req;
1654         int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1655
1656         memset(&req, 0, sizeof(req));
1657         req.node_idx = dlm->node_num;
1658         req.namelen = res->lockname.len;
1659         memcpy(req.name, res->lockname.name, res->lockname.len);
1660
1661         ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1662                                  &req, sizeof(req), nodenum, &status);
1663         /* XXX: negative status not handled properly here. */
1664         if (ret < 0)
1665                 mlog(ML_ERROR, "Error %d when sending message %u (key "
1666                      "0x%x) to node %u\n", ret, DLM_MASTER_REQUERY_MSG,
1667                      dlm->key, nodenum);
1668         else {
1669                 BUG_ON(status < 0);
1670                 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1671                 *real_master = (u8) (status & 0xff);
1672                 mlog(0, "node %u responded to master requery with %u\n",
1673                           nodenum, *real_master);
1674                 ret = 0;
1675         }
1676         return ret;
1677 }
1678
1679
1680 /* this function cannot error, so unless the sending
1681  * or receiving of the message failed, the owner can
1682  * be trusted */
1683 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
1684                                void **ret_data)
1685 {
1686         struct dlm_ctxt *dlm = data;
1687         struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1688         struct dlm_lock_resource *res = NULL;
1689         unsigned int hash;
1690         int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1691         u32 flags = DLM_ASSERT_MASTER_REQUERY;
1692         int dispatched = 0;
1693
1694         if (!dlm_grab(dlm)) {
1695                 /* since the domain has gone away on this
1696                  * node, the proper response is UNKNOWN */
1697                 return master;
1698         }
1699
1700         hash = dlm_lockid_hash(req->name, req->namelen);
1701
1702         spin_lock(&dlm->spinlock);
1703         res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash);
1704         if (res) {
1705                 spin_lock(&res->spinlock);
1706                 master = res->owner;
1707                 if (master == dlm->node_num) {
1708                         int ret = dlm_dispatch_assert_master(dlm, res,
1709                                                              0, 0, flags);
1710                         if (ret < 0) {
1711                                 mlog_errno(-ENOMEM);
1712                                 /* retry!? */
1713                                 BUG();
1714                         } else {
1715                                 dispatched = 1;
1716                         }
1717                 } else /* put.. incase we are not the master */
1718                         dlm_lockres_put(res);
1719                 spin_unlock(&res->spinlock);
1720         }
1721         spin_unlock(&dlm->spinlock);
1722
1723         if (!dispatched)
1724                 dlm_put(dlm);
1725         return master;
1726 }
1727
1728 static inline struct list_head *
1729 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1730 {
1731         struct list_head *ret;
1732         BUG_ON(list_num < 0);
1733         BUG_ON(list_num > 2);
1734         ret = &(res->granted);
1735         ret += list_num;
1736         return ret;
1737 }
1738 /* TODO: do ast flush business
1739  * TODO: do MIGRATING and RECOVERING spinning
1740  */
1741
1742 /*
1743 * NOTE about in-flight requests during migration:
1744 *
1745 * Before attempting the migrate, the master has marked the lockres as
1746 * MIGRATING and then flushed all of its pending ASTS.  So any in-flight
1747 * requests either got queued before the MIGRATING flag got set, in which
1748 * case the lock data will reflect the change and a return message is on
1749 * the way, or the request failed to get in before MIGRATING got set.  In
1750 * this case, the caller will be told to spin and wait for the MIGRATING
1751 * flag to be dropped, then recheck the master.
1752 * This holds true for the convert, cancel and unlock cases, and since lvb
1753 * updates are tied to these same messages, it applies to lvb updates as
1754 * well.  For the lock case, there is no way a lock can be on the master
1755 * queue and not be on the secondary queue since the lock is always added
1756 * locally first.  This means that the new target node will never be sent
1757 * a lock that he doesn't already have on the list.
1758 * In total, this means that the local lock is correct and should not be
1759 * updated to match the one sent by the master.  Any messages sent back
1760 * from the master before the MIGRATING flag will bring the lock properly
1761 * up-to-date, and the change will be ordered properly for the waiter.
1762 * We will *not* attempt to modify the lock underneath the waiter.
1763 */
1764
1765 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1766                                      struct dlm_lock_resource *res,
1767                                      struct dlm_migratable_lockres *mres)
1768 {
1769         struct dlm_migratable_lock *ml;
1770         struct list_head *queue, *iter;
1771         struct list_head *tmpq = NULL;
1772         struct dlm_lock *newlock = NULL;
1773         struct dlm_lockstatus *lksb = NULL;
1774         int ret = 0;
1775         int i, j, bad;
1776         struct dlm_lock *lock;
1777         u8 from = O2NM_MAX_NODES;
1778         unsigned int added = 0;
1779         __be64 c;
1780
1781         mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1782         for (i=0; i<mres->num_locks; i++) {
1783                 ml = &(mres->ml[i]);
1784
1785                 if (dlm_is_dummy_lock(dlm, ml, &from)) {
1786                         /* placeholder, just need to set the refmap bit */
1787                         BUG_ON(mres->num_locks != 1);
1788                         mlog(0, "%s:%.*s: dummy lock for %u\n",
1789                              dlm->name, mres->lockname_len, mres->lockname,
1790                              from);
1791                         spin_lock(&res->spinlock);
1792                         dlm_lockres_set_refmap_bit(dlm, res, from);
1793                         spin_unlock(&res->spinlock);
1794                         added++;
1795                         break;
1796                 }
1797                 BUG_ON(ml->highest_blocked != LKM_IVMODE);
1798                 newlock = NULL;
1799                 lksb = NULL;
1800
1801                 queue = dlm_list_num_to_pointer(res, ml->list);
1802                 tmpq = NULL;
1803
1804                 /* if the lock is for the local node it needs to
1805                  * be moved to the proper location within the queue.
1806                  * do not allocate a new lock structure. */
1807                 if (ml->node == dlm->node_num) {
1808                         /* MIGRATION ONLY! */
1809                         BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1810
1811                         lock = NULL;
1812                         spin_lock(&res->spinlock);
1813                         for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) {
1814                                 tmpq = dlm_list_idx_to_ptr(res, j);
1815                                 list_for_each(iter, tmpq) {
1816                                         lock = list_entry(iter,
1817                                                   struct dlm_lock, list);
1818                                         if (lock->ml.cookie == ml->cookie)
1819                                                 break;
1820                                         lock = NULL;
1821                                 }
1822                                 if (lock)
1823                                         break;
1824                         }
1825
1826                         /* lock is always created locally first, and
1827                          * destroyed locally last.  it must be on the list */
1828                         if (!lock) {
1829                                 c = ml->cookie;
1830                                 mlog(ML_ERROR, "Could not find local lock "
1831                                                "with cookie %u:%llu, node %u, "
1832                                                "list %u, flags 0x%x, type %d, "
1833                                                "conv %d, highest blocked %d\n",
1834                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1835                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1836                                      ml->node, ml->list, ml->flags, ml->type,
1837                                      ml->convert_type, ml->highest_blocked);
1838                                 __dlm_print_one_lock_resource(res);
1839                                 BUG();
1840                         }
1841
1842                         if (lock->ml.node != ml->node) {
1843                                 c = lock->ml.cookie;
1844                                 mlog(ML_ERROR, "Mismatched node# in lock "
1845                                      "cookie %u:%llu, name %.*s, node %u\n",
1846                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1847                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1848                                      res->lockname.len, res->lockname.name,
1849                                      lock->ml.node);
1850                                 c = ml->cookie;
1851                                 mlog(ML_ERROR, "Migrate lock cookie %u:%llu, "
1852                                      "node %u, list %u, flags 0x%x, type %d, "
1853                                      "conv %d, highest blocked %d\n",
1854                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1855                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1856                                      ml->node, ml->list, ml->flags, ml->type,
1857                                      ml->convert_type, ml->highest_blocked);
1858                                 __dlm_print_one_lock_resource(res);
1859                                 BUG();
1860                         }
1861
1862                         if (tmpq != queue) {
1863                                 c = ml->cookie;
1864                                 mlog(0, "Lock cookie %u:%llu was on list %u "
1865                                      "instead of list %u for %.*s\n",
1866                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1867                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1868                                      j, ml->list, res->lockname.len,
1869                                      res->lockname.name);
1870                                 __dlm_print_one_lock_resource(res);
1871                                 spin_unlock(&res->spinlock);
1872                                 continue;
1873                         }
1874
1875                         /* see NOTE above about why we do not update
1876                          * to match the master here */
1877
1878                         /* move the lock to its proper place */
1879                         /* do not alter lock refcount.  switching lists. */
1880                         list_move_tail(&lock->list, queue);
1881                         spin_unlock(&res->spinlock);
1882                         added++;
1883
1884                         mlog(0, "just reordered a local lock!\n");
1885                         continue;
1886                 }
1887
1888                 /* lock is for another node. */
1889                 newlock = dlm_new_lock(ml->type, ml->node,
1890                                        be64_to_cpu(ml->cookie), NULL);
1891                 if (!newlock) {
1892                         ret = -ENOMEM;
1893                         goto leave;
1894                 }
1895                 lksb = newlock->lksb;
1896                 dlm_lock_attach_lockres(newlock, res);
1897
1898                 if (ml->convert_type != LKM_IVMODE) {
1899                         BUG_ON(queue != &res->converting);
1900                         newlock->ml.convert_type = ml->convert_type;
1901                 }
1902                 lksb->flags |= (ml->flags &
1903                                 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1904
1905                 if (ml->type == LKM_NLMODE)
1906                         goto skip_lvb;
1907
1908                 if (!dlm_lvb_is_empty(mres->lvb)) {
1909                         if (lksb->flags & DLM_LKSB_PUT_LVB) {
1910                                 /* other node was trying to update
1911                                  * lvb when node died.  recreate the
1912                                  * lksb with the updated lvb. */
1913                                 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1914                                 /* the lock resource lvb update must happen
1915                                  * NOW, before the spinlock is dropped.
1916                                  * we no longer wait for the AST to update
1917                                  * the lvb. */
1918                                 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1919                         } else {
1920                                 /* otherwise, the node is sending its
1921                                  * most recent valid lvb info */
1922                                 BUG_ON(ml->type != LKM_EXMODE &&
1923                                        ml->type != LKM_PRMODE);
1924                                 if (!dlm_lvb_is_empty(res->lvb) &&
1925                                     (ml->type == LKM_EXMODE ||
1926                                      memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1927                                         int i;
1928                                         mlog(ML_ERROR, "%s:%.*s: received bad "
1929                                              "lvb! type=%d\n", dlm->name,
1930                                              res->lockname.len,
1931                                              res->lockname.name, ml->type);
1932                                         printk("lockres lvb=[");
1933                                         for (i=0; i<DLM_LVB_LEN; i++)
1934                                                 printk("%02x", res->lvb[i]);
1935                                         printk("]\nmigrated lvb=[");
1936                                         for (i=0; i<DLM_LVB_LEN; i++)
1937                                                 printk("%02x", mres->lvb[i]);
1938                                         printk("]\n");
1939                                         dlm_print_one_lock_resource(res);
1940                                         BUG();
1941                                 }
1942                                 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1943                         }
1944                 }
1945 skip_lvb:
1946
1947                 /* NOTE:
1948                  * wrt lock queue ordering and recovery:
1949                  *    1. order of locks on granted queue is
1950                  *       meaningless.
1951                  *    2. order of locks on converting queue is
1952                  *       LOST with the node death.  sorry charlie.
1953                  *    3. order of locks on the blocked queue is
1954                  *       also LOST.
1955                  * order of locks does not affect integrity, it
1956                  * just means that a lock request may get pushed
1957                  * back in line as a result of the node death.
1958                  * also note that for a given node the lock order
1959                  * for its secondary queue locks is preserved
1960                  * relative to each other, but clearly *not*
1961                  * preserved relative to locks from other nodes.
1962                  */
1963                 bad = 0;
1964                 spin_lock(&res->spinlock);
1965                 list_for_each_entry(lock, queue, list) {
1966                         if (lock->ml.cookie == ml->cookie) {
1967                                 c = lock->ml.cookie;
1968                                 mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already "
1969                                      "exists on this lockres!\n", dlm->name,
1970                                      res->lockname.len, res->lockname.name,
1971                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1972                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)));
1973
1974                                 mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, "
1975                                      "node=%u, cookie=%u:%llu, queue=%d\n",
1976                                      ml->type, ml->convert_type, ml->node,
1977                                      dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)),
1978                                      dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)),
1979                                      ml->list);
1980
1981                                 __dlm_print_one_lock_resource(res);
1982                                 bad = 1;
1983                                 break;
1984                         }
1985                 }
1986                 if (!bad) {
1987                         dlm_lock_get(newlock);
1988                         list_add_tail(&newlock->list, queue);
1989                         mlog(0, "%s:%.*s: added lock for node %u, "
1990                              "setting refmap bit\n", dlm->name,
1991                              res->lockname.len, res->lockname.name, ml->node);
1992                         dlm_lockres_set_refmap_bit(dlm, res, ml->node);
1993                         added++;
1994                 }
1995                 spin_unlock(&res->spinlock);
1996         }
1997         mlog(0, "done running all the locks\n");
1998
1999 leave:
2000         /* balance the ref taken when the work was queued */
2001         spin_lock(&res->spinlock);
2002         dlm_lockres_drop_inflight_ref(dlm, res);
2003         spin_unlock(&res->spinlock);
2004
2005         if (ret < 0) {
2006                 mlog_errno(ret);
2007                 if (newlock)
2008                         dlm_lock_put(newlock);
2009         }
2010
2011         return ret;
2012 }
2013
2014 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
2015                                        struct dlm_lock_resource *res)
2016 {
2017         int i;
2018         struct list_head *queue;
2019         struct dlm_lock *lock, *next;
2020
2021         assert_spin_locked(&dlm->spinlock);
2022         assert_spin_locked(&res->spinlock);
2023         res->state |= DLM_LOCK_RES_RECOVERING;
2024         if (!list_empty(&res->recovering)) {
2025                 mlog(0,
2026                      "Recovering res %s:%.*s, is already on recovery list!\n",
2027                      dlm->name, res->lockname.len, res->lockname.name);
2028                 list_del_init(&res->recovering);
2029                 dlm_lockres_put(res);
2030         }
2031         /* We need to hold a reference while on the recovery list */
2032         dlm_lockres_get(res);
2033         list_add_tail(&res->recovering, &dlm->reco.resources);
2034
2035         /* find any pending locks and put them back on proper list */
2036         for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
2037                 queue = dlm_list_idx_to_ptr(res, i);
2038                 list_for_each_entry_safe(lock, next, queue, list) {
2039                         dlm_lock_get(lock);
2040                         if (lock->convert_pending) {
2041                                 /* move converting lock back to granted */
2042                                 BUG_ON(i != DLM_CONVERTING_LIST);
2043                                 mlog(0, "node died with convert pending "
2044                                      "on %.*s. move back to granted list.\n",
2045                                      res->lockname.len, res->lockname.name);
2046                                 dlm_revert_pending_convert(res, lock);
2047                                 lock->convert_pending = 0;
2048                         } else if (lock->lock_pending) {
2049                                 /* remove pending lock requests completely */
2050                                 BUG_ON(i != DLM_BLOCKED_LIST);
2051                                 mlog(0, "node died with lock pending "
2052                                      "on %.*s. remove from blocked list and skip.\n",
2053                                      res->lockname.len, res->lockname.name);
2054                                 /* lock will be floating until ref in
2055                                  * dlmlock_remote is freed after the network
2056                                  * call returns.  ok for it to not be on any
2057                                  * list since no ast can be called
2058                                  * (the master is dead). */
2059                                 dlm_revert_pending_lock(res, lock);
2060                                 lock->lock_pending = 0;
2061                         } else if (lock->unlock_pending) {
2062                                 /* if an unlock was in progress, treat as
2063                                  * if this had completed successfully
2064                                  * before sending this lock state to the
2065                                  * new master.  note that the dlm_unlock
2066                                  * call is still responsible for calling
2067                                  * the unlockast.  that will happen after
2068                                  * the network call times out.  for now,
2069                                  * just move lists to prepare the new
2070                                  * recovery master.  */
2071                                 BUG_ON(i != DLM_GRANTED_LIST);
2072                                 mlog(0, "node died with unlock pending "
2073                                      "on %.*s. remove from blocked list and skip.\n",
2074                                      res->lockname.len, res->lockname.name);
2075                                 dlm_commit_pending_unlock(res, lock);
2076                                 lock->unlock_pending = 0;
2077                         } else if (lock->cancel_pending) {
2078                                 /* if a cancel was in progress, treat as
2079                                  * if this had completed successfully
2080                                  * before sending this lock state to the
2081                                  * new master */
2082                                 BUG_ON(i != DLM_CONVERTING_LIST);
2083                                 mlog(0, "node died with cancel pending "
2084                                      "on %.*s. move back to granted list.\n",
2085                                      res->lockname.len, res->lockname.name);
2086                                 dlm_commit_pending_cancel(res, lock);
2087                                 lock->cancel_pending = 0;
2088                         }
2089                         dlm_lock_put(lock);
2090                 }
2091         }
2092 }
2093
2094
2095
2096 /* removes all recovered locks from the recovery list.
2097  * sets the res->owner to the new master.
2098  * unsets the RECOVERY flag and wakes waiters. */
2099 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
2100                                               u8 dead_node, u8 new_master)
2101 {
2102         int i;
2103         struct hlist_node *hash_iter;
2104         struct hlist_head *bucket;
2105         struct dlm_lock_resource *res, *next;
2106
2107         assert_spin_locked(&dlm->spinlock);
2108
2109         list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
2110                 if (res->owner == dead_node) {
2111                         mlog(0, "%s: res %.*s, Changing owner from %u to %u\n",
2112                              dlm->name, res->lockname.len, res->lockname.name,
2113                              res->owner, new_master);
2114                         list_del_init(&res->recovering);
2115                         spin_lock(&res->spinlock);
2116                         /* new_master has our reference from
2117                          * the lock state sent during recovery */
2118                         dlm_change_lockres_owner(dlm, res, new_master);
2119                         res->state &= ~DLM_LOCK_RES_RECOVERING;
2120                         if (__dlm_lockres_has_locks(res))
2121                                 __dlm_dirty_lockres(dlm, res);
2122                         spin_unlock(&res->spinlock);
2123                         wake_up(&res->wq);
2124                         dlm_lockres_put(res);
2125                 }
2126         }
2127
2128         /* this will become unnecessary eventually, but
2129          * for now we need to run the whole hash, clear
2130          * the RECOVERING state and set the owner
2131          * if necessary */
2132         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2133                 bucket = dlm_lockres_hash(dlm, i);
2134                 hlist_for_each_entry(res, hash_iter, bucket, hash_node) {
2135                         if (!(res->state & DLM_LOCK_RES_RECOVERING))
2136                                 continue;
2137
2138                         if (res->owner != dead_node &&
2139                             res->owner != dlm->node_num)
2140                                 continue;
2141
2142                         if (!list_empty(&res->recovering)) {
2143                                 list_del_init(&res->recovering);
2144                                 dlm_lockres_put(res);
2145                         }
2146
2147                         /* new_master has our reference from
2148                          * the lock state sent during recovery */
2149                         mlog(0, "%s: res %.*s, Changing owner from %u to %u\n",
2150                              dlm->name, res->lockname.len, res->lockname.name,
2151                              res->owner, new_master);
2152                         spin_lock(&res->spinlock);
2153                         dlm_change_lockres_owner(dlm, res, new_master);
2154                         res->state &= ~DLM_LOCK_RES_RECOVERING;
2155                         if (__dlm_lockres_has_locks(res))
2156                                 __dlm_dirty_lockres(dlm, res);
2157                         spin_unlock(&res->spinlock);
2158                         wake_up(&res->wq);
2159                 }
2160         }
2161 }
2162
2163 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
2164 {
2165         if (local) {
2166                 if (lock->ml.type != LKM_EXMODE &&
2167                     lock->ml.type != LKM_PRMODE)
2168                         return 1;
2169         } else if (lock->ml.type == LKM_EXMODE)
2170                 return 1;
2171         return 0;
2172 }
2173
2174 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
2175                                struct dlm_lock_resource *res, u8 dead_node)
2176 {
2177         struct list_head *queue;
2178         struct dlm_lock *lock;
2179         int blank_lvb = 0, local = 0;
2180         int i;
2181         u8 search_node;
2182
2183         assert_spin_locked(&dlm->spinlock);
2184         assert_spin_locked(&res->spinlock);
2185
2186         if (res->owner == dlm->node_num)
2187                 /* if this node owned the lockres, and if the dead node
2188                  * had an EX when he died, blank out the lvb */
2189                 search_node = dead_node;
2190         else {
2191                 /* if this is a secondary lockres, and we had no EX or PR
2192                  * locks granted, we can no longer trust the lvb */
2193                 search_node = dlm->node_num;
2194                 local = 1;  /* check local state for valid lvb */
2195         }
2196
2197         for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
2198                 queue = dlm_list_idx_to_ptr(res, i);
2199                 list_for_each_entry(lock, queue, list) {
2200                         if (lock->ml.node == search_node) {
2201                                 if (dlm_lvb_needs_invalidation(lock, local)) {
2202                                         /* zero the lksb lvb and lockres lvb */
2203                                         blank_lvb = 1;
2204                                         memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
2205                                 }
2206                         }
2207                 }
2208         }
2209
2210         if (blank_lvb) {
2211                 mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
2212                      res->lockname.len, res->lockname.name, dead_node);
2213                 memset(res->lvb, 0, DLM_LVB_LEN);
2214         }
2215 }
2216
2217 static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
2218                                 struct dlm_lock_resource *res, u8 dead_node)
2219 {
2220         struct dlm_lock *lock, *next;
2221         unsigned int freed = 0;
2222
2223         /* this node is the lockres master:
2224          * 1) remove any stale locks for the dead node
2225          * 2) if the dead node had an EX when he died, blank out the lvb
2226          */
2227         assert_spin_locked(&dlm->spinlock);
2228         assert_spin_locked(&res->spinlock);
2229
2230         /* We do two dlm_lock_put(). One for removing from list and the other is
2231          * to force the DLM_UNLOCK_FREE_LOCK action so as to free the locks */
2232
2233         /* TODO: check pending_asts, pending_basts here */
2234         list_for_each_entry_safe(lock, next, &res->granted, list) {
2235                 if (lock->ml.node == dead_node) {
2236                         list_del_init(&lock->list);
2237                         dlm_lock_put(lock);
2238                         /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2239                         dlm_lock_put(lock);
2240                         freed++;
2241                 }
2242         }
2243         list_for_each_entry_safe(lock, next, &res->converting, list) {
2244                 if (lock->ml.node == dead_node) {
2245                         list_del_init(&lock->list);
2246                         dlm_lock_put(lock);
2247                         /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2248                         dlm_lock_put(lock);
2249                         freed++;
2250                 }
2251         }
2252         list_for_each_entry_safe(lock, next, &res->blocked, list) {
2253                 if (lock->ml.node == dead_node) {
2254                         list_del_init(&lock->list);
2255                         dlm_lock_put(lock);
2256                         /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2257                         dlm_lock_put(lock);
2258                         freed++;
2259                 }
2260         }
2261
2262         if (freed) {
2263                 mlog(0, "%s:%.*s: freed %u locks for dead node %u, "
2264                      "dropping ref from lockres\n", dlm->name,
2265                      res->lockname.len, res->lockname.name, freed, dead_node);
2266                 if(!test_bit(dead_node, res->refmap)) {
2267                         mlog(ML_ERROR, "%s:%.*s: freed %u locks for dead node %u, "
2268                              "but ref was not set\n", dlm->name,
2269                              res->lockname.len, res->lockname.name, freed, dead_node);
2270                         __dlm_print_one_lock_resource(res);
2271                 }
2272                 dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2273         } else if (test_bit(dead_node, res->refmap)) {
2274                 mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2275                      "no locks and had not purged before dying\n", dlm->name,
2276                      res->lockname.len, res->lockname.name, dead_node);
2277                 dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2278         }
2279
2280         /* do not kick thread yet */
2281         __dlm_dirty_lockres(dlm, res);
2282 }
2283
2284 /* if this node is the recovery master, and there are no
2285  * locks for a given lockres owned by this node that are in
2286  * either PR or EX mode, zero out the lvb before requesting.
2287  *
2288  */
2289
2290
2291 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
2292 {
2293         struct hlist_node *iter;
2294         struct dlm_lock_resource *res;
2295         int i;
2296         struct hlist_head *bucket;
2297         struct dlm_lock *lock;
2298
2299
2300         /* purge any stale mles */
2301         dlm_clean_master_list(dlm, dead_node);
2302
2303         /*
2304          * now clean up all lock resources.  there are two rules:
2305          *
2306          * 1) if the dead node was the master, move the lockres
2307          *    to the recovering list.  set the RECOVERING flag.
2308          *    this lockres needs to be cleaned up before it can
2309          *    be used further.
2310          *
2311          * 2) if this node was the master, remove all locks from
2312          *    each of the lockres queues that were owned by the
2313          *    dead node.  once recovery finishes, the dlm thread
2314          *    can be kicked again to see if any ASTs or BASTs
2315          *    need to be fired as a result.
2316          */
2317         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2318                 bucket = dlm_lockres_hash(dlm, i);
2319                 hlist_for_each_entry(res, iter, bucket, hash_node) {
2320                         /* always prune any $RECOVERY entries for dead nodes,
2321                          * otherwise hangs can occur during later recovery */
2322                         if (dlm_is_recovery_lock(res->lockname.name,
2323                                                  res->lockname.len)) {
2324                                 spin_lock(&res->spinlock);
2325                                 list_for_each_entry(lock, &res->granted, list) {
2326                                         if (lock->ml.node == dead_node) {
2327                                                 mlog(0, "AHA! there was "
2328                                                      "a $RECOVERY lock for dead "
2329                                                      "node %u (%s)!\n",
2330                                                      dead_node, dlm->name);
2331                                                 list_del_init(&lock->list);
2332                                                 dlm_lock_put(lock);
2333                                                 break;
2334                                         }
2335                                 }
2336                                 spin_unlock(&res->spinlock);
2337                                 continue;
2338                         }
2339                         spin_lock(&res->spinlock);
2340                         /* zero the lvb if necessary */
2341                         dlm_revalidate_lvb(dlm, res, dead_node);
2342                         if (res->owner == dead_node) {
2343                                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
2344                                         mlog(ML_NOTICE, "%s: res %.*s, Skip "
2345                                              "recovery as it is being freed\n",
2346                                              dlm->name, res->lockname.len,
2347                                              res->lockname.name);
2348                                 } else
2349                                         dlm_move_lockres_to_recovery_list(dlm,
2350                                                                           res);
2351
2352                         } else if (res->owner == dlm->node_num) {
2353                                 dlm_free_dead_locks(dlm, res, dead_node);
2354                                 __dlm_lockres_calc_usage(dlm, res);
2355                         }
2356                         spin_unlock(&res->spinlock);
2357                 }
2358         }
2359
2360 }
2361
2362 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
2363 {
2364         assert_spin_locked(&dlm->spinlock);
2365
2366         if (dlm->reco.new_master == idx) {
2367                 mlog(0, "%s: recovery master %d just died\n",
2368                      dlm->name, idx);
2369                 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2370                         /* finalize1 was reached, so it is safe to clear
2371                          * the new_master and dead_node.  that recovery
2372                          * is complete. */
2373                         mlog(0, "%s: dead master %d had reached "
2374                              "finalize1 state, clearing\n", dlm->name, idx);
2375                         dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2376                         __dlm_reset_recovery(dlm);
2377                 }
2378         }
2379
2380         /* Clean up join state on node death. */
2381         if (dlm->joining_node == idx) {
2382                 mlog(0, "Clearing join state for node %u\n", idx);
2383                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
2384         }
2385
2386         /* check to see if the node is already considered dead */
2387         if (!test_bit(idx, dlm->live_nodes_map)) {
2388                 mlog(0, "for domain %s, node %d is already dead. "
2389                      "another node likely did recovery already.\n",
2390                      dlm->name, idx);
2391                 return;
2392         }
2393
2394         /* check to see if we do not care about this node */
2395         if (!test_bit(idx, dlm->domain_map)) {
2396                 /* This also catches the case that we get a node down
2397                  * but haven't joined the domain yet. */
2398                 mlog(0, "node %u already removed from domain!\n", idx);
2399                 return;
2400         }
2401
2402         clear_bit(idx, dlm->live_nodes_map);
2403
2404         /* make sure local cleanup occurs before the heartbeat events */
2405         if (!test_bit(idx, dlm->recovery_map))
2406                 dlm_do_local_recovery_cleanup(dlm, idx);
2407
2408         /* notify anything attached to the heartbeat events */
2409         dlm_hb_event_notify_attached(dlm, idx, 0);
2410
2411         mlog(0, "node %u being removed from domain map!\n", idx);
2412         clear_bit(idx, dlm->domain_map);
2413         clear_bit(idx, dlm->exit_domain_map);
2414         /* wake up migration waiters if a node goes down.
2415          * perhaps later we can genericize this for other waiters. */
2416         wake_up(&dlm->migration_wq);
2417
2418         if (test_bit(idx, dlm->recovery_map))
2419                 mlog(0, "domain %s, node %u already added "
2420                      "to recovery map!\n", dlm->name, idx);
2421         else
2422                 set_bit(idx, dlm->recovery_map);
2423 }
2424
2425 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
2426 {
2427         struct dlm_ctxt *dlm = data;
2428
2429         if (!dlm_grab(dlm))
2430                 return;
2431
2432         /*
2433          * This will notify any dlm users that a node in our domain
2434          * went away without notifying us first.
2435          */
2436         if (test_bit(idx, dlm->domain_map))
2437                 dlm_fire_domain_eviction_callbacks(dlm, idx);
2438
2439         spin_lock(&dlm->spinlock);
2440         __dlm_hb_node_down(dlm, idx);
2441         spin_unlock(&dlm->spinlock);
2442
2443         dlm_put(dlm);
2444 }
2445
2446 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
2447 {
2448         struct dlm_ctxt *dlm = data;
2449
2450         if (!dlm_grab(dlm))
2451                 return;
2452
2453         spin_lock(&dlm->spinlock);
2454         set_bit(idx, dlm->live_nodes_map);
2455         /* do NOT notify mle attached to the heartbeat events.
2456          * new nodes are not interesting in mastery until joined. */
2457         spin_unlock(&dlm->spinlock);
2458
2459         dlm_put(dlm);
2460 }
2461
2462 static void dlm_reco_ast(void *astdata)
2463 {
2464         struct dlm_ctxt *dlm = astdata;
2465         mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2466              dlm->node_num, dlm->name);
2467 }
2468 static void dlm_reco_bast(void *astdata, int blocked_type)
2469 {
2470         struct dlm_ctxt *dlm = astdata;
2471         mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2472              dlm->node_num, dlm->name);
2473 }
2474 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
2475 {
2476         mlog(0, "unlockast for recovery lock fired!\n");
2477 }
2478
2479 /*
2480  * dlm_pick_recovery_master will continually attempt to use
2481  * dlmlock() on the special "$RECOVERY" lockres with the
2482  * LKM_NOQUEUE flag to get an EX.  every thread that enters
2483  * this function on each node racing to become the recovery
2484  * master will not stop attempting this until either:
2485  * a) this node gets the EX (and becomes the recovery master),
2486  * or b) dlm->reco.new_master gets set to some nodenum
2487  * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2488  * so each time a recovery master is needed, the entire cluster
2489  * will sync at this point.  if the new master dies, that will
2490  * be detected in dlm_do_recovery */
2491 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2492 {
2493         enum dlm_status ret;
2494         struct dlm_lockstatus lksb;
2495         int status = -EINVAL;
2496
2497         mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2498              dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2499 again:
2500         memset(&lksb, 0, sizeof(lksb));
2501
2502         ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
2503                       DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN,
2504                       dlm_reco_ast, dlm, dlm_reco_bast);
2505
2506         mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2507              dlm->name, ret, lksb.status);
2508
2509         if (ret == DLM_NORMAL) {
2510                 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2511                      dlm->name, dlm->node_num);
2512
2513                 /* got the EX lock.  check to see if another node
2514                  * just became the reco master */
2515                 if (dlm_reco_master_ready(dlm)) {
2516                         mlog(0, "%s: got reco EX lock, but %u will "
2517                              "do the recovery\n", dlm->name,
2518                              dlm->reco.new_master);
2519                         status = -EEXIST;
2520                 } else {
2521                         status = 0;
2522
2523                         /* see if recovery was already finished elsewhere */
2524                         spin_lock(&dlm->spinlock);
2525                         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
2526                                 status = -EINVAL;
2527                                 mlog(0, "%s: got reco EX lock, but "
2528                                      "node got recovered already\n", dlm->name);
2529                                 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2530                                         mlog(ML_ERROR, "%s: new master is %u "
2531                                              "but no dead node!\n",
2532                                              dlm->name, dlm->reco.new_master);
2533                                         BUG();
2534                                 }
2535                         }
2536                         spin_unlock(&dlm->spinlock);
2537                 }
2538
2539                 /* if this node has actually become the recovery master,
2540                  * set the master and send the messages to begin recovery */
2541                 if (!status) {
2542                         mlog(0, "%s: dead=%u, this=%u, sending "
2543                              "begin_reco now\n", dlm->name,
2544                              dlm->reco.dead_node, dlm->node_num);
2545                         status = dlm_send_begin_reco_message(dlm,
2546                                       dlm->reco.dead_node);
2547                         /* this always succeeds */
2548                         BUG_ON(status);
2549
2550                         /* set the new_master to this node */
2551                         spin_lock(&dlm->spinlock);
2552                         dlm_set_reco_master(dlm, dlm->node_num);
2553                         spin_unlock(&dlm->spinlock);
2554                 }
2555
2556                 /* recovery lock is a special case.  ast will not get fired,
2557                  * so just go ahead and unlock it. */
2558                 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
2559                 if (ret == DLM_DENIED) {
2560                         mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2561                         ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2562                 }
2563                 if (ret != DLM_NORMAL) {
2564                         /* this would really suck. this could only happen
2565                          * if there was a network error during the unlock
2566                          * because of node death.  this means the unlock
2567                          * is actually "done" and the lock structure is
2568                          * even freed.  we can continue, but only
2569                          * because this specific lock name is special. */
2570                         mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
2571                 }
2572         } else if (ret == DLM_NOTQUEUED) {
2573                 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2574                      dlm->name, dlm->node_num);
2575                 /* another node is master. wait on
2576                  * reco.new_master != O2NM_INVALID_NODE_NUM
2577                  * for at most one second */
2578                 wait_event_timeout(dlm->dlm_reco_thread_wq,
2579                                          dlm_reco_master_ready(dlm),
2580                                          msecs_to_jiffies(1000));
2581                 if (!dlm_reco_master_ready(dlm)) {
2582                         mlog(0, "%s: reco master taking awhile\n",
2583                              dlm->name);
2584                         goto again;
2585                 }
2586                 /* another node has informed this one that it is reco master */
2587                 mlog(0, "%s: reco master %u is ready to recover %u\n",
2588                      dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
2589                 status = -EEXIST;
2590         } else if (ret == DLM_RECOVERING) {
2591                 mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n",
2592                      dlm->name, dlm->node_num);
2593                 goto again;
2594         } else {
2595                 struct dlm_lock_resource *res;
2596
2597                 /* dlmlock returned something other than NOTQUEUED or NORMAL */
2598                 mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2599                      "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2600                      dlm_errname(lksb.status));
2601                 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2602                                          DLM_RECOVERY_LOCK_NAME_LEN);
2603                 if (res) {
2604                         dlm_print_one_lock_resource(res);
2605                         dlm_lockres_put(res);
2606                 } else {
2607                         mlog(ML_ERROR, "recovery lock not found\n");
2608                 }
2609                 BUG();
2610         }
2611
2612         return status;
2613 }
2614
2615 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2616 {
2617         struct dlm_begin_reco br;
2618         int ret = 0;
2619         struct dlm_node_iter iter;
2620         int nodenum;
2621         int status;
2622
2623         mlog(0, "%s: dead node is %u\n", dlm->name, dead_node);
2624
2625         spin_lock(&dlm->spinlock);
2626         dlm_node_iter_init(dlm->domain_map, &iter);
2627         spin_unlock(&dlm->spinlock);
2628
2629         clear_bit(dead_node, iter.node_map);
2630
2631         memset(&br, 0, sizeof(br));
2632         br.node_idx = dlm->node_num;
2633         br.dead_node = dead_node;
2634
2635         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2636                 ret = 0;
2637                 if (nodenum == dead_node) {
2638                         mlog(0, "not sending begin reco to dead node "
2639                                   "%u\n", dead_node);
2640                         continue;
2641                 }
2642                 if (nodenum == dlm->node_num) {
2643                         mlog(0, "not sending begin reco to self\n");
2644                         continue;
2645                 }
2646 retry:
2647                 ret = -EINVAL;
2648                 mlog(0, "attempting to send begin reco msg to %d\n",
2649                           nodenum);
2650                 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2651                                          &br, sizeof(br), nodenum, &status);
2652                 /* negative status is handled ok by caller here */
2653                 if (ret >= 0)
2654                         ret = status;
2655                 if (dlm_is_host_down(ret)) {
2656                         /* node is down.  not involved in recovery
2657                          * so just keep going */
2658                         mlog(ML_NOTICE, "%s: node %u was down when sending "
2659                              "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2660                         ret = 0;
2661                 }
2662
2663                 /*
2664                  * Prior to commit aad1b15310b9bcd59fa81ab8f2b1513b59553ea8,
2665                  * dlm_begin_reco_handler() returned EAGAIN and not -EAGAIN.
2666                  * We are handling both for compatibility reasons.
2667                  */
2668                 if (ret == -EAGAIN || ret == EAGAIN) {
2669                         mlog(0, "%s: trying to start recovery of node "
2670                              "%u, but node %u is waiting for last recovery "
2671                              "to complete, backoff for a bit\n", dlm->name,
2672                              dead_node, nodenum);
2673                         msleep(100);
2674                         goto retry;
2675                 }
2676                 if (ret < 0) {
2677                         struct dlm_lock_resource *res;
2678
2679                         /* this is now a serious problem, possibly ENOMEM
2680                          * in the network stack.  must retry */
2681                         mlog_errno(ret);
2682                         mlog(ML_ERROR, "begin reco of dlm %s to node %u "
2683                              "returned %d\n", dlm->name, nodenum, ret);
2684                         res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2685                                                  DLM_RECOVERY_LOCK_NAME_LEN);
2686                         if (res) {
2687                                 dlm_print_one_lock_resource(res);
2688                                 dlm_lockres_put(res);
2689                         } else {
2690                                 mlog(ML_ERROR, "recovery lock not found\n");
2691                         }
2692                         /* sleep for a bit in hopes that we can avoid
2693                          * another ENOMEM */
2694                         msleep(100);
2695                         goto retry;
2696                 }
2697         }
2698
2699         return ret;
2700 }
2701
2702 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2703                            void **ret_data)
2704 {
2705         struct dlm_ctxt *dlm = data;
2706         struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2707
2708         /* ok to return 0, domain has gone away */
2709         if (!dlm_grab(dlm))
2710                 return 0;
2711
2712         spin_lock(&dlm->spinlock);
2713         if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2714                 mlog(0, "%s: node %u wants to recover node %u (%u:%u) "
2715                      "but this node is in finalize state, waiting on finalize2\n",
2716                      dlm->name, br->node_idx, br->dead_node,
2717                      dlm->reco.dead_node, dlm->reco.new_master);
2718                 spin_unlock(&dlm->spinlock);
2719                 return -EAGAIN;
2720         }
2721         spin_unlock(&dlm->spinlock);
2722
2723         mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n",
2724              dlm->name, br->node_idx, br->dead_node,
2725              dlm->reco.dead_node, dlm->reco.new_master);
2726
2727         dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2728
2729         spin_lock(&dlm->spinlock);
2730         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2731                 if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2732                         mlog(0, "%s: new_master %u died, changing "
2733                              "to %u\n", dlm->name, dlm->reco.new_master,
2734                              br->node_idx);
2735                 } else {
2736                         mlog(0, "%s: new_master %u NOT DEAD, changing "
2737                              "to %u\n", dlm->name, dlm->reco.new_master,
2738                              br->node_idx);
2739                         /* may not have seen the new master as dead yet */
2740                 }
2741         }
2742         if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2743                 mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2744                      "node %u changing it to %u\n", dlm->name,
2745                      dlm->reco.dead_node, br->node_idx, br->dead_node);
2746         }
2747         dlm_set_reco_master(dlm, br->node_idx);
2748         dlm_set_reco_dead_node(dlm, br->dead_node);
2749         if (!test_bit(br->dead_node, dlm->recovery_map)) {
2750                 mlog(0, "recovery master %u sees %u as dead, but this "
2751                      "node has not yet.  marking %u as dead\n",
2752                      br->node_idx, br->dead_node, br->dead_node);
2753                 if (!test_bit(br->dead_node, dlm->domain_map) ||
2754                     !test_bit(br->dead_node, dlm->live_nodes_map))
2755                         mlog(0, "%u not in domain/live_nodes map "
2756                              "so setting it in reco map manually\n",
2757                              br->dead_node);
2758                 /* force the recovery cleanup in __dlm_hb_node_down
2759                  * both of these will be cleared in a moment */
2760                 set_bit(br->dead_node, dlm->domain_map);
2761                 set_bit(br->dead_node, dlm->live_nodes_map);
2762                 __dlm_hb_node_down(dlm, br->dead_node);
2763         }
2764         spin_unlock(&dlm->spinlock);
2765
2766         dlm_kick_recovery_thread(dlm);
2767
2768         mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n",
2769              dlm->name, br->node_idx, br->dead_node,
2770              dlm->reco.dead_node, dlm->reco.new_master);
2771
2772         dlm_put(dlm);
2773         return 0;
2774 }
2775
2776 #define DLM_FINALIZE_STAGE2  0x01
2777 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2778 {
2779         int ret = 0;
2780         struct dlm_finalize_reco fr;
2781         struct dlm_node_iter iter;
2782         int nodenum;
2783         int status;
2784         int stage = 1;
2785
2786         mlog(0, "finishing recovery for node %s:%u, "
2787              "stage %d\n", dlm->name, dlm->reco.dead_node, stage);
2788
2789         spin_lock(&dlm->spinlock);
2790         dlm_node_iter_init(dlm->domain_map, &iter);
2791         spin_unlock(&dlm->spinlock);
2792
2793 stage2:
2794         memset(&fr, 0, sizeof(fr));
2795         fr.node_idx = dlm->node_num;
2796         fr.dead_node = dlm->reco.dead_node;
2797         if (stage == 2)
2798                 fr.flags |= DLM_FINALIZE_STAGE2;
2799
2800         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2801                 if (nodenum == dlm->node_num)
2802                         continue;
2803                 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2804                                          &fr, sizeof(fr), nodenum, &status);
2805                 if (ret >= 0)
2806                         ret = status;
2807                 if (ret < 0) {
2808                         mlog(ML_ERROR, "Error %d when sending message %u (key "
2809                              "0x%x) to node %u\n", ret, DLM_FINALIZE_RECO_MSG,
2810                              dlm->key, nodenum);
2811                         if (dlm_is_host_down(ret)) {
2812                                 /* this has no effect on this recovery
2813                                  * session, so set the status to zero to
2814                                  * finish out the last recovery */
2815                                 mlog(ML_ERROR, "node %u went down after this "
2816                                      "node finished recovery.\n", nodenum);
2817                                 ret = 0;
2818                                 continue;
2819                         }
2820                         break;
2821                 }
2822         }
2823         if (stage == 1) {
2824                 /* reset the node_iter back to the top and send finalize2 */
2825                 iter.curnode = -1;
2826                 stage = 2;
2827                 goto stage2;
2828         }
2829
2830         return ret;
2831 }
2832
2833 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2834                               void **ret_data)
2835 {
2836         struct dlm_ctxt *dlm = data;
2837         struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2838         int stage = 1;
2839
2840         /* ok to return 0, domain has gone away */
2841         if (!dlm_grab(dlm))
2842                 return 0;
2843
2844         if (fr->flags & DLM_FINALIZE_STAGE2)
2845                 stage = 2;
2846
2847         mlog(0, "%s: node %u finalizing recovery stage%d of "
2848              "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage,
2849              fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master);
2850
2851         spin_lock(&dlm->spinlock);
2852
2853         if (dlm->reco.new_master != fr->node_idx) {
2854                 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2855                      "%u is supposed to be the new master, dead=%u\n",
2856                      fr->node_idx, dlm->reco.new_master, fr->dead_node);
2857                 BUG();
2858         }
2859         if (dlm->reco.dead_node != fr->dead_node) {
2860                 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2861                      "node %u, but node %u is supposed to be dead\n",
2862                      fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2863                 BUG();
2864         }
2865
2866         switch (stage) {
2867                 case 1:
2868                         dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2869                         if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2870                                 mlog(ML_ERROR, "%s: received finalize1 from "
2871                                      "new master %u for dead node %u, but "
2872                                      "this node has already received it!\n",
2873                                      dlm->name, fr->node_idx, fr->dead_node);
2874                                 dlm_print_reco_node_status(dlm);
2875                                 BUG();
2876                         }
2877                         dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
2878                         spin_unlock(&dlm->spinlock);
2879                         break;
2880                 case 2:
2881                         if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) {
2882                                 mlog(ML_ERROR, "%s: received finalize2 from "
2883                                      "new master %u for dead node %u, but "
2884                                      "this node did not have finalize1!\n",
2885                                      dlm->name, fr->node_idx, fr->dead_node);
2886                                 dlm_print_reco_node_status(dlm);
2887                                 BUG();
2888                         }
2889                         dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2890                         __dlm_reset_recovery(dlm);
2891                         spin_unlock(&dlm->spinlock);
2892                         dlm_kick_recovery_thread(dlm);
2893                         break;
2894                 default:
2895                         BUG();
2896         }
2897
2898         mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n",
2899              dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master);
2900
2901         dlm_put(dlm);
2902         return 0;
2903 }