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