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 static int dlm_do_recovery(struct dlm_ctxt *dlm);
  59
  60 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm);
  61 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node);
  62 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
  63 static int dlm_request_all_locks(struct dlm_ctxt *dlm,
  64                                  u8 request_from, u8 dead_node);
  65 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
  66
  67 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res);
  68 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
  69                                         const char *lockname, int namelen,
  70                                         int total_locks, u64 cookie,
  71                                         u8 flags, u8 master);
  72 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
  73                                     struct dlm_migratable_lockres *mres,
  74                                     u8 send_to,
  75                                     struct dlm_lock_resource *res,
  76                                     int total_locks);
  77 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
  78                                      struct dlm_lock_resource *res,
  79                                      struct dlm_migratable_lockres *mres);
  80 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm);
  81 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm,
  82                                  u8 dead_node, u8 send_to);
  83 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node);
  84 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
  85                                         struct list_head *list, u8 dead_node);
  86 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
  87                                               u8 dead_node, u8 new_master);
  88 static void dlm_reco_ast(void *astdata);
  89 static void dlm_reco_bast(void *astdata, int blocked_type);
  90 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st);
  91 static void dlm_request_all_locks_worker(struct dlm_work_item *item,
  92                                          void *data);
  93 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data);
  94 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
  95                                       struct dlm_lock_resource *res,
  96                                       u8 *real_master);
  97
  98 static u64 dlm_get_next_mig_cookie(void);
  99
 100 static DEFINE_SPINLOCK(dlm_reco_state_lock);
 101 static DEFINE_SPINLOCK(dlm_mig_cookie_lock);
 102 static u64 dlm_mig_cookie = 1;
 103
 104 static u64 dlm_get_next_mig_cookie(void)
 105 {
 106         u64 c;
 107         spin_lock(&dlm_mig_cookie_lock);
 108         c = dlm_mig_cookie;
 109         if (dlm_mig_cookie == (~0ULL))
 110                 dlm_mig_cookie = 1;
 111         else
 112                 dlm_mig_cookie++;
 113         spin_unlock(&dlm_mig_cookie_lock);
 114         return c;
 115 }
 116
 117 static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm,
 118                                           u8 dead_node)
 119 {
 120         assert_spin_locked(&dlm->spinlock);
 121         if (dlm->reco.dead_node != dead_node)
 122                 mlog(0, "%s: changing dead_node from %u to %u\n",
 123                      dlm->name, dlm->reco.dead_node, dead_node);
 124         dlm->reco.dead_node = dead_node;
 125 }
 126
 127 static inline void dlm_set_reco_master(struct dlm_ctxt *dlm,
 128                                        u8 master)
 129 {
 130         assert_spin_locked(&dlm->spinlock);
 131         mlog(0, "%s: changing new_master from %u to %u\n",
 132              dlm->name, dlm->reco.new_master, master);
 133         dlm->reco.new_master = master;
 134 }
 135
 136 static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm)
 137 {
 138         assert_spin_locked(&dlm->spinlock);
 139         clear_bit(dlm->reco.dead_node, dlm->recovery_map);
 140         dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
 141         dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
 142 }
 143
 144 static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
 145 {
 146         spin_lock(&dlm->spinlock);
 147         __dlm_reset_recovery(dlm);
 148         spin_unlock(&dlm->spinlock);
 149 }
 150
 151 /* Worker function used during recovery. */
 152 void dlm_dispatch_work(struct work_struct *work)
 153 {
 154         struct dlm_ctxt *dlm =
 155                 container_of(work, struct dlm_ctxt, dispatched_work);
 156         LIST_HEAD(tmp_list);
 157         struct dlm_work_item *item, *next;
 158         dlm_workfunc_t *workfunc;
 159         int tot=0;
 160
 161         spin_lock(&dlm->work_lock);
 162         list_splice_init(&dlm->work_list, &tmp_list);
 163         spin_unlock(&dlm->work_lock);
 164
 165         list_for_each_entry(item, &tmp_list, list) {
 166                 tot++;
 167         }
 168         mlog(0, "%s: work thread has %d work items\n", dlm->name, tot);
 169
 170         list_for_each_entry_safe(item, next, &tmp_list, list) {
 171                 workfunc = item->func;
 172                 list_del_init(&item->list);
 173
 174                 /* already have ref on dlm to avoid having
 175                  * it disappear.  just double-check. */
 176                 BUG_ON(item->dlm != dlm);
 177
 178                 /* this is allowed to sleep and
 179                  * call network stuff */
 180                 workfunc(item, item->data);
 181
 182                 dlm_put(dlm);
 183                 kfree(item);
 184         }
 185 }
 186
 187 /*
 188  * RECOVERY THREAD
 189  */
 190
 191 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
 192 {
 193         /* wake the recovery thread
 194          * this will wake the reco thread in one of three places
 195          * 1) sleeping with no recovery happening
 196          * 2) sleeping with recovery mastered elsewhere
 197          * 3) recovery mastered here, waiting on reco data */
 198
 199         wake_up(&dlm->dlm_reco_thread_wq);
 200 }
 201
 202 /* Launch the recovery thread */
 203 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
 204 {
 205         mlog(0, "starting dlm recovery thread...\n");
 206
 207         dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
 208                         "dlm_reco-%s", dlm->name);
 209         if (IS_ERR(dlm->dlm_reco_thread_task)) {
 210                 mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
 211                 dlm->dlm_reco_thread_task = NULL;
 212                 return -EINVAL;
 213         }
 214
 215         return 0;
 216 }
 217
 218 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
 219 {
 220         if (dlm->dlm_reco_thread_task) {
 221                 mlog(0, "waiting for dlm recovery thread to exit\n");
 222                 kthread_stop(dlm->dlm_reco_thread_task);
 223                 dlm->dlm_reco_thread_task = NULL;
 224         }
 225 }
 226
 227
 228
 229 /*
 230  * this is lame, but here's how recovery works...
 231  * 1) all recovery threads cluster wide will work on recovering
 232  *    ONE node at a time
 233  * 2) negotiate who will take over all the locks for the dead node.
 234  *    thats right... ALL the locks.
 235  * 3) once a new master is chosen, everyone scans all locks
 236  *    and moves aside those mastered by the dead guy
 237  * 4) each of these locks should be locked until recovery is done
 238  * 5) the new master collects up all of secondary lock queue info
 239  *    one lock at a time, forcing each node to communicate back
 240  *    before continuing
 241  * 6) each secondary lock queue responds with the full known lock info
 242  * 7) once the new master has run all its locks, it sends a ALLDONE!
 243  *    message to everyone
 244  * 8) upon receiving this message, the secondary queue node unlocks
 245  *    and responds to the ALLDONE
 246  * 9) once the new master gets responses from everyone, he unlocks
 247  *    everything and recovery for this dead node is done
 248  *10) go back to 2) while there are still dead nodes
 249  *
 250  */
 251
 252 static void dlm_print_reco_node_status(struct dlm_ctxt *dlm)
 253 {
 254         struct dlm_reco_node_data *ndata;
 255         struct dlm_lock_resource *res;
 256
 257         mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n",
 258              dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
 259              dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive",
 260              dlm->reco.dead_node, dlm->reco.new_master);
 261
 262         list_for_each_entry(ndata, &dlm->reco.node_data, list) {
 263                 char *st = "unknown";
 264                 switch (ndata->state) {
 265                         case DLM_RECO_NODE_DATA_INIT:
 266                                 st = "init";
 267                                 break;
 268                         case DLM_RECO_NODE_DATA_REQUESTING:
 269                                 st = "requesting";
 270                                 break;
 271                         case DLM_RECO_NODE_DATA_DEAD:
 272                                 st = "dead";
 273                                 break;
 274                         case DLM_RECO_NODE_DATA_RECEIVING:
 275                                 st = "receiving";
 276                                 break;
 277                         case DLM_RECO_NODE_DATA_REQUESTED:
 278                                 st = "requested";
 279                                 break;
 280                         case DLM_RECO_NODE_DATA_DONE:
 281                                 st = "done";
 282                                 break;
 283                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
 284                                 st = "finalize-sent";
 285                                 break;
 286                         default:
 287                                 st = "bad";
 288                                 break;
 289                 }
 290                 mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n",
 291                      dlm->name, ndata->node_num, st);
 292         }
 293         list_for_each_entry(res, &dlm->reco.resources, recovering) {
 294                 mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n",
 295                      dlm->name, res->lockname.len, res->lockname.name);
 296         }
 297 }
 298
 299 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
 300
 301 static int dlm_recovery_thread(void *data)
 302 {
 303         int status;
 304         struct dlm_ctxt *dlm = data;
 305         unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
 306
 307         mlog(0, "dlm thread running for %s...\n", dlm->name);
 308
 309         while (!kthread_should_stop()) {
 310                 if (dlm_domain_fully_joined(dlm)) {
 311                         status = dlm_do_recovery(dlm);
 312                         if (status == -EAGAIN) {
 313                                 /* do not sleep, recheck immediately. */
 314                                 continue;
 315                         }
 316                         if (status < 0)
 317                                 mlog_errno(status);
 318                 }
 319
 320                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
 321                                                  kthread_should_stop(),
 322                                                  timeout);
 323         }
 324
 325         mlog(0, "quitting DLM recovery thread\n");
 326         return 0;
 327 }
 328
 329 /* returns true when the recovery master has contacted us */
 330 static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
 331 {
 332         int ready;
 333         spin_lock(&dlm->spinlock);
 334         ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
 335         spin_unlock(&dlm->spinlock);
 336         return ready;
 337 }
 338
 339 /* returns true if node is no longer in the domain
 340  * could be dead or just not joined */
 341 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
 342 {
 343         int dead;
 344         spin_lock(&dlm->spinlock);
 345         dead = !test_bit(node, dlm->domain_map);
 346         spin_unlock(&dlm->spinlock);
 347         return dead;
 348 }
 349
 350 /* returns true if node is no longer in the domain
 351  * could be dead or just not joined */
 352 static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node)
 353 {
 354         int recovered;
 355         spin_lock(&dlm->spinlock);
 356         recovered = !test_bit(node, dlm->recovery_map);
 357         spin_unlock(&dlm->spinlock);
 358         return recovered;
 359 }
 360
 361
 362 void dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout)
 363 {
 364         if (dlm_is_node_dead(dlm, node))
 365                 return;
 366
 367         printk(KERN_NOTICE "o2dlm: Waiting on the death of node %u in "
 368                "domain %s\n", node, dlm->name);
 369
 370         if (timeout)
 371                 wait_event_timeout(dlm->dlm_reco_thread_wq,
 372                                    dlm_is_node_dead(dlm, node),
 373                                    msecs_to_jiffies(timeout));
 374         else
 375                 wait_event(dlm->dlm_reco_thread_wq,
 376                            dlm_is_node_dead(dlm, node));
 377 }
 378
 379 void dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout)
 380 {
 381         if (dlm_is_node_recovered(dlm, node))
 382                 return;
 383
 384         printk(KERN_NOTICE "o2dlm: Waiting on the recovery of node %u in "
 385                "domain %s\n", node, dlm->name);
 386
 387         if (timeout)
 388                 wait_event_timeout(dlm->dlm_reco_thread_wq,
 389                                    dlm_is_node_recovered(dlm, node),
 390                                    msecs_to_jiffies(timeout));
 391         else
 392                 wait_event(dlm->dlm_reco_thread_wq,
 393                            dlm_is_node_recovered(dlm, node));
 394 }
 395
 396 /* callers of the top-level api calls (dlmlock/dlmunlock) should
 397  * block on the dlm->reco.event when recovery is in progress.
 398  * the dlm recovery thread will set this state when it begins
 399  * recovering a dead node (as the new master or not) and clear
 400  * the state and wake as soon as all affected lock resources have
 401  * been marked with the RECOVERY flag */
 402 static int dlm_in_recovery(struct dlm_ctxt *dlm)
 403 {
 404         int in_recovery;
 405         spin_lock(&dlm->spinlock);
 406         in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
 407         spin_unlock(&dlm->spinlock);
 408         return in_recovery;
 409 }
 410
 411
 412 void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
 413 {
 414         if (dlm_in_recovery(dlm)) {
 415                 mlog(0, "%s: reco thread %d in recovery: "
 416                      "state=%d, master=%u, dead=%u\n",
 417                      dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
 418                      dlm->reco.state, dlm->reco.new_master,
 419                      dlm->reco.dead_node);
 420         }
 421         wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
 422 }
 423
 424 static void dlm_begin_recovery(struct dlm_ctxt *dlm)
 425 {
 426         spin_lock(&dlm->spinlock);
 427         BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
 428         printk(KERN_NOTICE "o2dlm: Begin recovery on domain %s for node %u\n",
 429                dlm->name, dlm->reco.dead_node);
 430         dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
 431         spin_unlock(&dlm->spinlock);
 432 }
 433
 434 static void dlm_end_recovery(struct dlm_ctxt *dlm)
 435 {
 436         spin_lock(&dlm->spinlock);
 437         BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
 438         dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
 439         spin_unlock(&dlm->spinlock);
 440         printk(KERN_NOTICE "o2dlm: End recovery on domain %s\n", dlm->name);
 441         wake_up(&dlm->reco.event);
 442 }
 443
 444 static void dlm_print_recovery_master(struct dlm_ctxt *dlm)
 445 {
 446         printk(KERN_NOTICE "o2dlm: Node %u (%s) is the Recovery Master for the "
 447                "dead node %u in domain %s\n", dlm->reco.new_master,
 448                (dlm->node_num == dlm->reco.new_master ? "me" : "he"),
 449                dlm->reco.dead_node, dlm->name);
 450 }
 451
 452 static int dlm_do_recovery(struct dlm_ctxt *dlm)
 453 {
 454         int status = 0;
 455         int ret;
 456
 457         spin_lock(&dlm->spinlock);
 458
 459         /* check to see if the new master has died */
 460         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
 461             test_bit(dlm->reco.new_master, dlm->recovery_map)) {
 462                 mlog(0, "new master %u died while recovering %u!\n",
 463                      dlm->reco.new_master, dlm->reco.dead_node);
 464                 /* unset the new_master, leave dead_node */
 465                 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
 466         }
 467
 468         /* select a target to recover */
 469         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
 470                 int bit;
 471
 472                 bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES, 0);
 473                 if (bit >= O2NM_MAX_NODES || bit < 0)
 474                         dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
 475                 else
 476                         dlm_set_reco_dead_node(dlm, bit);
 477         } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
 478                 /* BUG? */
 479                 mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
 480                      dlm->reco.dead_node);
 481                 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
 482         }
 483
 484         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
 485                 // mlog(0, "nothing to recover!  sleeping now!\n");
 486                 spin_unlock(&dlm->spinlock);
 487                 /* return to main thread loop and sleep. */
 488                 return 0;
 489         }
 490         mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n",
 491              dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
 492              dlm->reco.dead_node);
 493         spin_unlock(&dlm->spinlock);
 494
 495         /* take write barrier */
 496         /* (stops the list reshuffling thread, proxy ast handling) */
 497         dlm_begin_recovery(dlm);
 498
 499         if (dlm->reco.new_master == dlm->node_num)
 500                 goto master_here;
 501
 502         if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
 503                 /* choose a new master, returns 0 if this node
 504                  * is the master, -EEXIST if it's another node.
 505                  * this does not return until a new master is chosen
 506                  * or recovery completes entirely. */
 507                 ret = dlm_pick_recovery_master(dlm);
 508                 if (!ret) {
 509                         /* already notified everyone.  go. */
 510                         goto master_here;
 511                 }
 512                 mlog(0, "another node will master this recovery session.\n");
 513         }
 514
 515         dlm_print_recovery_master(dlm);
 516
 517         /* it is safe to start everything back up here
 518          * because all of the dead node's lock resources
 519          * have been marked as in-recovery */
 520         dlm_end_recovery(dlm);
 521
 522         /* sleep out in main dlm_recovery_thread loop. */
 523         return 0;
 524
 525 master_here:
 526         dlm_print_recovery_master(dlm);
 527
 528         status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
 529         if (status < 0) {
 530                 /* we should never hit this anymore */
 531                 mlog(ML_ERROR, "%s: Error %d remastering locks for node %u, "
 532                      "retrying.\n", dlm->name, status, dlm->reco.dead_node);
 533                 /* yield a bit to allow any final network messages
 534                  * to get handled on remaining nodes */
 535                 msleep(100);
 536         } else {
 537                 /* success!  see if any other nodes need recovery */
 538                 mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
 539                      dlm->name, dlm->reco.dead_node, dlm->node_num);
 540                 spin_lock(&dlm->spinlock);
 541                 __dlm_reset_recovery(dlm);
 542                 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
 543                 spin_unlock(&dlm->spinlock);
 544         }
 545         dlm_end_recovery(dlm);
 546
 547         /* continue and look for another dead node */
 548         return -EAGAIN;
 549 }
 550
 551 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
 552 {
 553         int status = 0;
 554         struct dlm_reco_node_data *ndata;
 555         int all_nodes_done;
 556         int destroy = 0;
 557         int pass = 0;
 558
 559         do {
 560                 /* we have become recovery master.  there is no escaping
 561                  * this, so just keep trying until we get it. */
 562                 status = dlm_init_recovery_area(dlm, dead_node);
 563                 if (status < 0) {
 564                         mlog(ML_ERROR, "%s: failed to alloc recovery area, "
 565                              "retrying\n", dlm->name);
 566                         msleep(1000);
 567                 }
 568         } while (status != 0);
 569
 570         /* safe to access the node data list without a lock, since this
 571          * process is the only one to change the list */
 572         list_for_each_entry(ndata, &dlm->reco.node_data, list) {
 573                 BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
 574                 ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
 575
 576                 mlog(0, "%s: Requesting lock info from node %u\n", dlm->name,
 577                      ndata->node_num);
 578
 579                 if (ndata->node_num == dlm->node_num) {
 580                         ndata->state = DLM_RECO_NODE_DATA_DONE;
 581                         continue;
 582                 }
 583
 584                 do {
 585                         status = dlm_request_all_locks(dlm, ndata->node_num,
 586                                                        dead_node);
 587                         if (status < 0) {
 588                                 mlog_errno(status);
 589                                 if (dlm_is_host_down(status)) {
 590                                         /* node died, ignore it for recovery */
 591                                         status = 0;
 592                                         ndata->state = DLM_RECO_NODE_DATA_DEAD;
 593                                         /* wait for the domain map to catch up
 594                                          * with the network state. */
 595                                         wait_event_timeout(dlm->dlm_reco_thread_wq,
 596                                                            dlm_is_node_dead(dlm,
 597                                                                 ndata->node_num),
 598                                                            msecs_to_jiffies(1000));
 599                                         mlog(0, "waited 1 sec for %u, "
 600                                              "dead? %s\n", ndata->node_num,
 601                                              dlm_is_node_dead(dlm, ndata->node_num) ?
 602                                              "yes" : "no");
 603                                 } else {
 604                                         /* -ENOMEM on the other node */
 605                                         mlog(0, "%s: node %u returned "
 606                                              "%d during recovery, retrying "
 607                                              "after a short wait\n",
 608                                              dlm->name, ndata->node_num,
 609                                              status);
 610                                         msleep(100);
 611                                 }
 612                         }
 613                 } while (status != 0);
 614
 615                 spin_lock(&dlm_reco_state_lock);
 616                 switch (ndata->state) {
 617                         case DLM_RECO_NODE_DATA_INIT:
 618                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
 619                         case DLM_RECO_NODE_DATA_REQUESTED:
 620                                 BUG();
 621                                 break;
 622                         case DLM_RECO_NODE_DATA_DEAD:
 623                                 mlog(0, "node %u died after requesting "
 624                                      "recovery info for node %u\n",
 625                                      ndata->node_num, dead_node);
 626                                 /* fine.  don't need this node's info.
 627                                  * continue without it. */
 628                                 break;
 629                         case DLM_RECO_NODE_DATA_REQUESTING:
 630                                 ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
 631                                 mlog(0, "now receiving recovery data from "
 632                                      "node %u for dead node %u\n",
 633                                      ndata->node_num, dead_node);
 634                                 break;
 635                         case DLM_RECO_NODE_DATA_RECEIVING:
 636                                 mlog(0, "already receiving recovery data from "
 637                                      "node %u for dead node %u\n",
 638                                      ndata->node_num, dead_node);
 639                                 break;
 640                         case DLM_RECO_NODE_DATA_DONE:
 641                                 mlog(0, "already DONE receiving recovery data "
 642                                      "from node %u for dead node %u\n",
 643                                      ndata->node_num, dead_node);
 644                                 break;
 645                 }
 646                 spin_unlock(&dlm_reco_state_lock);
 647         }
 648
 649         mlog(0, "%s: Done requesting all lock info\n", dlm->name);
 650
 651         /* nodes should be sending reco data now
 652          * just need to wait */
 653
 654         while (1) {
 655                 /* check all the nodes now to see if we are
 656                  * done, or if anyone died */
 657                 all_nodes_done = 1;
 658                 spin_lock(&dlm_reco_state_lock);
 659                 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
 660                         mlog(0, "checking recovery state of node %u\n",
 661                              ndata->node_num);
 662                         switch (ndata->state) {
 663                                 case DLM_RECO_NODE_DATA_INIT:
 664                                 case DLM_RECO_NODE_DATA_REQUESTING:
 665                                         mlog(ML_ERROR, "bad ndata state for "
 666                                              "node %u: state=%d\n",
 667                                              ndata->node_num, ndata->state);
 668                                         BUG();
 669                                         break;
 670                                 case DLM_RECO_NODE_DATA_DEAD:
 671                                         mlog(0, "node %u died after "
 672                                              "requesting recovery info for "
 673                                              "node %u\n", ndata->node_num,
 674                                              dead_node);
 675                                         break;
 676                                 case DLM_RECO_NODE_DATA_RECEIVING:
 677                                 case DLM_RECO_NODE_DATA_REQUESTED:
 678                                         mlog(0, "%s: node %u still in state %s\n",
 679                                              dlm->name, ndata->node_num,
 680                                              ndata->state==DLM_RECO_NODE_DATA_RECEIVING ?
 681                                              "receiving" : "requested");
 682                                         all_nodes_done = 0;
 683                                         break;
 684                                 case DLM_RECO_NODE_DATA_DONE:
 685                                         mlog(0, "%s: node %u state is done\n",
 686                                              dlm->name, ndata->node_num);
 687                                         break;
 688                                 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
 689                                         mlog(0, "%s: node %u state is finalize\n",
 690                                              dlm->name, ndata->node_num);
 691                                         break;
 692                         }
 693                 }
 694                 spin_unlock(&dlm_reco_state_lock);
 695
 696                 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
 697                      all_nodes_done?"yes":"no");
 698                 if (all_nodes_done) {
 699                         int ret;
 700
 701                         /* Set this flag on recovery master to avoid
 702                          * a new recovery for another dead node start
 703                          * before the recovery is not done. That may
 704                          * cause recovery hung.*/
 705                         spin_lock(&dlm->spinlock);
 706                         dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
 707                         spin_unlock(&dlm->spinlock);
 708
 709                         /* all nodes are now in DLM_RECO_NODE_DATA_DONE state
 710                          * just send a finalize message to everyone and
 711                          * clean up */
 712                         mlog(0, "all nodes are done! send finalize\n");
 713                         ret = dlm_send_finalize_reco_message(dlm);
 714                         if (ret < 0)
 715                                 mlog_errno(ret);
 716
 717                         spin_lock(&dlm->spinlock);
 718                         dlm_finish_local_lockres_recovery(dlm, dead_node,
 719                                                           dlm->node_num);
 720                         spin_unlock(&dlm->spinlock);
 721                         mlog(0, "should be done with recovery!\n");
 722
 723                         mlog(0, "finishing recovery of %s at %lu, "
 724                              "dead=%u, this=%u, new=%u\n", dlm->name,
 725                              jiffies, dlm->reco.dead_node,
 726                              dlm->node_num, dlm->reco.new_master);
 727                         destroy = 1;
 728                         status = 0;
 729                         /* rescan everything marked dirty along the way */
 730                         dlm_kick_thread(dlm, NULL);
 731                         break;
 732                 }
 733                 /* wait to be signalled, with periodic timeout
 734                  * to check for node death */
 735                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
 736                                          kthread_should_stop(),
 737                                          msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
 738
 739         }
 740
 741         if (destroy)
 742                 dlm_destroy_recovery_area(dlm, dead_node);
 743
 744         return status;
 745 }
 746
 747 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
 748 {
 749         int num=0;
 750         struct dlm_reco_node_data *ndata;
 751
 752         spin_lock(&dlm->spinlock);
 753         memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
 754         /* nodes can only be removed (by dying) after dropping
 755          * this lock, and death will be trapped later, so this should do */
 756         spin_unlock(&dlm->spinlock);
 757
 758         while (1) {
 759                 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
 760                 if (num >= O2NM_MAX_NODES) {
 761                         break;
 762                 }
 763                 BUG_ON(num == dead_node);
 764
 765                 ndata = kzalloc(sizeof(*ndata), GFP_NOFS);
 766                 if (!ndata) {
 767                         dlm_destroy_recovery_area(dlm, dead_node);
 768                         return -ENOMEM;
 769                 }
 770                 ndata->node_num = num;
 771                 ndata->state = DLM_RECO_NODE_DATA_INIT;
 772                 spin_lock(&dlm_reco_state_lock);
 773                 list_add_tail(&ndata->list, &dlm->reco.node_data);
 774                 spin_unlock(&dlm_reco_state_lock);
 775                 num++;
 776         }
 777
 778         return 0;
 779 }
 780
 781 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
 782 {
 783         struct dlm_reco_node_data *ndata, *next;
 784         LIST_HEAD(tmplist);
 785
 786         spin_lock(&dlm_reco_state_lock);
 787         list_splice_init(&dlm->reco.node_data, &tmplist);
 788         spin_unlock(&dlm_reco_state_lock);
 789
 790         list_for_each_entry_safe(ndata, next, &tmplist, list) {
 791                 list_del_init(&ndata->list);
 792                 kfree(ndata);
 793         }
 794 }
 795
 796 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
 797                                  u8 dead_node)
 798 {
 799         struct dlm_lock_request lr;
 800         int ret;
 801         int status;
 802
 803         mlog(0, "\n");
 804
 805
 806         mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
 807                   "to %u\n", dead_node, request_from);
 808
 809         memset(&lr, 0, sizeof(lr));
 810         lr.node_idx = dlm->node_num;
 811         lr.dead_node = dead_node;
 812
 813         // send message
 814         ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
 815                                  &lr, sizeof(lr), request_from, &status);
 816
 817         /* negative status is handled by caller */
 818         if (ret < 0)
 819                 mlog(ML_ERROR, "%s: Error %d send LOCK_REQUEST to node %u "
 820                      "to recover dead node %u\n", dlm->name, ret,
 821                      request_from, dead_node);
 822         else
 823                 ret = status;
 824         // return from here, then
 825         // sleep until all received or error
 826         return ret;
 827
 828 }
 829
 830 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
 831                                   void **ret_data)
 832 {
 833         struct dlm_ctxt *dlm = data;
 834         struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
 835         char *buf = NULL;
 836         struct dlm_work_item *item = NULL;
 837
 838         if (!dlm_grab(dlm))
 839                 return -EINVAL;
 840
 841         if (lr->dead_node != dlm->reco.dead_node) {
 842                 mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local "
 843                      "dead_node is %u\n", dlm->name, lr->node_idx,
 844                      lr->dead_node, dlm->reco.dead_node);
 845                 dlm_print_reco_node_status(dlm);
 846                 /* this is a hack */
 847                 dlm_put(dlm);
 848                 return -ENOMEM;
 849         }
 850         BUG_ON(lr->dead_node != dlm->reco.dead_node);
 851
 852         item = kzalloc(sizeof(*item), GFP_NOFS);
 853         if (!item) {
 854                 dlm_put(dlm);
 855                 return -ENOMEM;
 856         }
 857
 858         /* this will get freed by dlm_request_all_locks_worker */
 859         buf = (char *) __get_free_page(GFP_NOFS);
 860         if (!buf) {
 861                 kfree(item);
 862                 dlm_put(dlm);
 863                 return -ENOMEM;
 864         }
 865
 866         /* queue up work for dlm_request_all_locks_worker */
 867         dlm_grab(dlm);  /* get an extra ref for the work item */
 868         dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
 869         item->u.ral.reco_master = lr->node_idx;
 870         item->u.ral.dead_node = lr->dead_node;
 871         spin_lock(&dlm->work_lock);
 872         list_add_tail(&item->list, &dlm->work_list);
 873         spin_unlock(&dlm->work_lock);
 874         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
 875
 876         dlm_put(dlm);
 877         return 0;
 878 }
 879
 880 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
 881 {
 882         struct dlm_migratable_lockres *mres;
 883         struct dlm_lock_resource *res;
 884         struct dlm_ctxt *dlm;
 885         LIST_HEAD(resources);
 886         int ret;
 887         u8 dead_node, reco_master;
 888         int skip_all_done = 0;
 889
 890         dlm = item->dlm;
 891         dead_node = item->u.ral.dead_node;
 892         reco_master = item->u.ral.reco_master;
 893         mres = (struct dlm_migratable_lockres *)data;
 894
 895         mlog(0, "%s: recovery worker started, dead=%u, master=%u\n",
 896              dlm->name, dead_node, reco_master);
 897
 898         if (dead_node != dlm->reco.dead_node ||
 899             reco_master != dlm->reco.new_master) {
 900                 /* worker could have been created before the recovery master
 901                  * died.  if so, do not continue, but do not error. */
 902                 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
 903                         mlog(ML_NOTICE, "%s: will not send recovery state, "
 904                              "recovery master %u died, thread=(dead=%u,mas=%u)"
 905                              " current=(dead=%u,mas=%u)\n", dlm->name,
 906                              reco_master, dead_node, reco_master,
 907                              dlm->reco.dead_node, dlm->reco.new_master);
 908                 } else {
 909                         mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, "
 910                              "master=%u), request(dead=%u, master=%u)\n",
 911                              dlm->name, dlm->reco.dead_node,
 912                              dlm->reco.new_master, dead_node, reco_master);
 913                 }
 914                 goto leave;
 915         }
 916
 917         /* lock resources should have already been moved to the
 918          * dlm->reco.resources list.  now move items from that list
 919          * to a temp list if the dead owner matches.  note that the
 920          * whole cluster recovers only one node at a time, so we
 921          * can safely move UNKNOWN lock resources for each recovery
 922          * session. */
 923         dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
 924
 925         /* now we can begin blasting lockreses without the dlm lock */
 926
 927         /* any errors returned will be due to the new_master dying,
 928          * the dlm_reco_thread should detect this */
 929         list_for_each_entry(res, &resources, recovering) {
 930                 ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
 931                                         DLM_MRES_RECOVERY);
 932                 if (ret < 0) {
 933                         mlog(ML_ERROR, "%s: node %u went down while sending "
 934                              "recovery state for dead node %u, ret=%d\n", dlm->name,
 935                              reco_master, dead_node, ret);
 936                         skip_all_done = 1;
 937                         break;
 938                 }
 939         }
 940
 941         /* move the resources back to the list */
 942         spin_lock(&dlm->spinlock);
 943         list_splice_init(&resources, &dlm->reco.resources);
 944         spin_unlock(&dlm->spinlock);
 945
 946         if (!skip_all_done) {
 947                 ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
 948                 if (ret < 0) {
 949                         mlog(ML_ERROR, "%s: node %u went down while sending "
 950                              "recovery all-done for dead node %u, ret=%d\n",
 951                              dlm->name, reco_master, dead_node, ret);
 952                 }
 953         }
 954 leave:
 955         free_page((unsigned long)data);
 956 }
 957
 958
 959 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
 960 {
 961         int ret, tmpret;
 962         struct dlm_reco_data_done done_msg;
 963
 964         memset(&done_msg, 0, sizeof(done_msg));
 965         done_msg.node_idx = dlm->node_num;
 966         done_msg.dead_node = dead_node;
 967         mlog(0, "sending DATA DONE message to %u, "
 968              "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
 969              done_msg.dead_node);
 970
 971         ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
 972                                  sizeof(done_msg), send_to, &tmpret);
 973         if (ret < 0) {
 974                 mlog(ML_ERROR, "%s: Error %d send RECO_DATA_DONE to node %u "
 975                      "to recover dead node %u\n", dlm->name, ret, send_to,
 976                      dead_node);
 977                 if (!dlm_is_host_down(ret)) {
 978                         BUG();
 979                 }
 980         } else
 981                 ret = tmpret;
 982         return ret;
 983 }
 984
 985
 986 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
 987                                void **ret_data)
 988 {
 989         struct dlm_ctxt *dlm = data;
 990         struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
 991         struct dlm_reco_node_data *ndata = NULL;
 992         int ret = -EINVAL;
 993
 994         if (!dlm_grab(dlm))
 995                 return -EINVAL;
 996
 997         mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
 998              "node_idx=%u, this node=%u\n", done->dead_node,
 999              dlm->reco.dead_node, done->node_idx, dlm->node_num);
1000
1001         mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node),
1002                         "Got DATA DONE: dead_node=%u, reco.dead_node=%u, "
1003                         "node_idx=%u, this node=%u\n", done->dead_node,
1004                         dlm->reco.dead_node, done->node_idx, dlm->node_num);
1005
1006         spin_lock(&dlm_reco_state_lock);
1007         list_for_each_entry(ndata, &dlm->reco.node_data, list) {
1008                 if (ndata->node_num != done->node_idx)
1009                         continue;
1010
1011                 switch (ndata->state) {
1012                         /* should have moved beyond INIT but not to FINALIZE yet */
1013                         case DLM_RECO_NODE_DATA_INIT:
1014                         case DLM_RECO_NODE_DATA_DEAD:
1015                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
1016                                 mlog(ML_ERROR, "bad ndata state for node %u:"
1017                                      " state=%d\n", ndata->node_num,
1018                                      ndata->state);
1019                                 BUG();
1020                                 break;
1021                         /* these states are possible at this point, anywhere along
1022                          * the line of recovery */
1023                         case DLM_RECO_NODE_DATA_DONE:
1024                         case DLM_RECO_NODE_DATA_RECEIVING:
1025                         case DLM_RECO_NODE_DATA_REQUESTED:
1026                         case DLM_RECO_NODE_DATA_REQUESTING:
1027                                 mlog(0, "node %u is DONE sending "
1028                                           "recovery data!\n",
1029                                           ndata->node_num);
1030
1031                                 ndata->state = DLM_RECO_NODE_DATA_DONE;
1032                                 ret = 0;
1033                                 break;
1034                 }
1035         }
1036         spin_unlock(&dlm_reco_state_lock);
1037
1038         /* wake the recovery thread, some node is done */
1039         if (!ret)
1040                 dlm_kick_recovery_thread(dlm);
1041
1042         if (ret < 0)
1043                 mlog(ML_ERROR, "failed to find recovery node data for node "
1044                      "%u\n", done->node_idx);
1045         dlm_put(dlm);
1046
1047         mlog(0, "leaving reco data done handler, ret=%d\n", ret);
1048         return ret;
1049 }
1050
1051 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
1052                                         struct list_head *list,
1053                                         u8 dead_node)
1054 {
1055         struct dlm_lock_resource *res, *next;
1056         struct dlm_lock *lock;
1057
1058         spin_lock(&dlm->spinlock);
1059         list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
1060                 /* always prune any $RECOVERY entries for dead nodes,
1061                  * otherwise hangs can occur during later recovery */
1062                 if (dlm_is_recovery_lock(res->lockname.name,
1063                                          res->lockname.len)) {
1064                         spin_lock(&res->spinlock);
1065                         list_for_each_entry(lock, &res->granted, list) {
1066                                 if (lock->ml.node == dead_node) {
1067                                         mlog(0, "AHA! there was "
1068                                              "a $RECOVERY lock for dead "
1069                                              "node %u (%s)!\n",
1070                                              dead_node, dlm->name);
1071                                         list_del_init(&lock->list);
1072                                         dlm_lock_put(lock);
1073                                         /* Can't schedule DLM_UNLOCK_FREE_LOCK
1074                                          * - do manually */
1075                                         dlm_lock_put(lock);
1076                                         break;
1077                                 }
1078                         }
1079                         spin_unlock(&res->spinlock);
1080                         continue;
1081                 }
1082
1083                 if (res->owner == dead_node) {
1084                         mlog(0, "found lockres owned by dead node while "
1085                                   "doing recovery for node %u. sending it.\n",
1086                                   dead_node);
1087                         list_move_tail(&res->recovering, list);
1088                 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1089                         mlog(0, "found UNKNOWN owner while doing recovery "
1090                                   "for node %u. sending it.\n", dead_node);
1091                         list_move_tail(&res->recovering, list);
1092                 }
1093         }
1094         spin_unlock(&dlm->spinlock);
1095 }
1096
1097 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
1098 {
1099         int total_locks = 0;
1100         struct list_head *iter, *queue = &res->granted;
1101         int i;
1102
1103         for (i=0; i<3; i++) {
1104                 list_for_each(iter, queue)
1105                         total_locks++;
1106                 queue++;
1107         }
1108         return total_locks;
1109 }
1110
1111
1112 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
1113                                       struct dlm_migratable_lockres *mres,
1114                                       u8 send_to,
1115                                       struct dlm_lock_resource *res,
1116                                       int total_locks)
1117 {
1118         u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
1119         int mres_total_locks = be32_to_cpu(mres->total_locks);
1120         int sz, ret = 0, status = 0;
1121         u8 orig_flags = mres->flags,
1122            orig_master = mres->master;
1123
1124         BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
1125         if (!mres->num_locks)
1126                 return 0;
1127
1128         sz = sizeof(struct dlm_migratable_lockres) +
1129                 (mres->num_locks * sizeof(struct dlm_migratable_lock));
1130
1131         /* add an all-done flag if we reached the last lock */
1132         orig_flags = mres->flags;
1133         BUG_ON(total_locks > mres_total_locks);
1134         if (total_locks == mres_total_locks)
1135                 mres->flags |= DLM_MRES_ALL_DONE;
1136
1137         mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n",
1138              dlm->name, res->lockname.len, res->lockname.name,
1139              orig_flags & DLM_MRES_MIGRATION ? "migration" : "recovery",
1140              send_to);
1141
1142         /* send it */
1143         ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
1144                                  sz, send_to, &status);
1145         if (ret < 0) {
1146                 /* XXX: negative status is not handled.
1147                  * this will end up killing this node. */
1148                 mlog(ML_ERROR, "%s: res %.*s, Error %d send MIG_LOCKRES to "
1149                      "node %u (%s)\n", dlm->name, mres->lockname_len,
1150                      mres->lockname, ret, send_to,
1151                      (orig_flags & DLM_MRES_MIGRATION ?
1152                       "migration" : "recovery"));
1153         } else {
1154                 /* might get an -ENOMEM back here */
1155                 ret = status;
1156                 if (ret < 0) {
1157                         mlog_errno(ret);
1158
1159                         if (ret == -EFAULT) {
1160                                 mlog(ML_ERROR, "node %u told me to kill "
1161                                      "myself!\n", send_to);
1162                                 BUG();
1163                         }
1164                 }
1165         }
1166
1167         /* zero and reinit the message buffer */
1168         dlm_init_migratable_lockres(mres, res->lockname.name,
1169                                     res->lockname.len, mres_total_locks,
1170                                     mig_cookie, orig_flags, orig_master);
1171         return ret;
1172 }
1173
1174 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
1175                                         const char *lockname, int namelen,
1176                                         int total_locks, u64 cookie,
1177                                         u8 flags, u8 master)
1178 {
1179         /* mres here is one full page */
1180         clear_page(mres);
1181         mres->lockname_len = namelen;
1182         memcpy(mres->lockname, lockname, namelen);
1183         mres->num_locks = 0;
1184         mres->total_locks = cpu_to_be32(total_locks);
1185         mres->mig_cookie = cpu_to_be64(cookie);
1186         mres->flags = flags;
1187         mres->master = master;
1188 }
1189
1190 static void dlm_prepare_lvb_for_migration(struct dlm_lock *lock,
1191                                           struct dlm_migratable_lockres *mres,
1192                                           int queue)
1193 {
1194         if (!lock->lksb)
1195                return;
1196
1197         /* Ignore lvb in all locks in the blocked list */
1198         if (queue == DLM_BLOCKED_LIST)
1199                 return;
1200
1201         /* Only consider lvbs in locks with granted EX or PR lock levels */
1202         if (lock->ml.type != LKM_EXMODE && lock->ml.type != LKM_PRMODE)
1203                 return;
1204
1205         if (dlm_lvb_is_empty(mres->lvb)) {
1206                 memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1207                 return;
1208         }
1209
1210         /* Ensure the lvb copied for migration matches in other valid locks */
1211         if (!memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))
1212                 return;
1213
1214         mlog(ML_ERROR, "Mismatched lvb in lock cookie=%u:%llu, name=%.*s, "
1215              "node=%u\n",
1216              dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
1217              dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
1218              lock->lockres->lockname.len, lock->lockres->lockname.name,
1219              lock->ml.node);
1220         dlm_print_one_lock_resource(lock->lockres);
1221         BUG();
1222 }
1223
1224 /* returns 1 if this lock fills the network structure,
1225  * 0 otherwise */
1226 static int dlm_add_lock_to_array(struct dlm_lock *lock,
1227                                  struct dlm_migratable_lockres *mres, int queue)
1228 {
1229         struct dlm_migratable_lock *ml;
1230         int lock_num = mres->num_locks;
1231
1232         ml = &(mres->ml[lock_num]);
1233         ml->cookie = lock->ml.cookie;
1234         ml->type = lock->ml.type;
1235         ml->convert_type = lock->ml.convert_type;
1236         ml->highest_blocked = lock->ml.highest_blocked;
1237         ml->list = queue;
1238         if (lock->lksb) {
1239                 ml->flags = lock->lksb->flags;
1240                 dlm_prepare_lvb_for_migration(lock, mres, queue);
1241         }
1242         ml->node = lock->ml.node;
1243         mres->num_locks++;
1244         /* we reached the max, send this network message */
1245         if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1246                 return 1;
1247         return 0;
1248 }
1249
1250 static void dlm_add_dummy_lock(struct dlm_ctxt *dlm,
1251                                struct dlm_migratable_lockres *mres)
1252 {
1253         struct dlm_lock dummy;
1254         memset(&dummy, 0, sizeof(dummy));
1255         dummy.ml.cookie = 0;
1256         dummy.ml.type = LKM_IVMODE;
1257         dummy.ml.convert_type = LKM_IVMODE;
1258         dummy.ml.highest_blocked = LKM_IVMODE;
1259         dummy.lksb = NULL;
1260         dummy.ml.node = dlm->node_num;
1261         dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST);
1262 }
1263
1264 static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm,
1265                                     struct dlm_migratable_lock *ml,
1266                                     u8 *nodenum)
1267 {
1268         if (unlikely(ml->cookie == 0 &&
1269             ml->type == LKM_IVMODE &&
1270             ml->convert_type == LKM_IVMODE &&
1271             ml->highest_blocked == LKM_IVMODE &&
1272             ml->list == DLM_BLOCKED_LIST)) {
1273                 *nodenum = ml->node;
1274                 return 1;
1275         }
1276         return 0;
1277 }
1278
1279 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1280                          struct dlm_migratable_lockres *mres,
1281                          u8 send_to, u8 flags)
1282 {
1283         struct list_head *queue;
1284         int total_locks, i;
1285         u64 mig_cookie = 0;
1286         struct dlm_lock *lock;
1287         int ret = 0;
1288
1289         BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1290
1291         mlog(0, "sending to %u\n", send_to);
1292
1293         total_locks = dlm_num_locks_in_lockres(res);
1294         if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1295                 /* rare, but possible */
1296                 mlog(0, "argh.  lockres has %d locks.  this will "
1297                           "require more than one network packet to "
1298                           "migrate\n", total_locks);
1299                 mig_cookie = dlm_get_next_mig_cookie();
1300         }
1301
1302         dlm_init_migratable_lockres(mres, res->lockname.name,
1303                                     res->lockname.len, total_locks,
1304                                     mig_cookie, flags, res->owner);
1305
1306         total_locks = 0;
1307         for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1308                 queue = dlm_list_idx_to_ptr(res, i);
1309                 list_for_each_entry(lock, queue, list) {
1310                         /* add another lock. */
1311                         total_locks++;
1312                         if (!dlm_add_lock_to_array(lock, mres, i))
1313                                 continue;
1314
1315                         /* this filled the lock message,
1316                          * we must send it immediately. */
1317                         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1318                                                        res, total_locks);
1319                         if (ret < 0)
1320                                 goto error;
1321                 }
1322         }
1323         if (total_locks == 0) {
1324                 /* send a dummy lock to indicate a mastery reference only */
1325                 mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n",
1326                      dlm->name, res->lockname.len, res->lockname.name,
1327                      send_to, flags & DLM_MRES_RECOVERY ? "recovery" :
1328                      "migration");
1329                 dlm_add_dummy_lock(dlm, mres);
1330         }
1331         /* flush any remaining locks */
1332         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1333         if (ret < 0)
1334                 goto error;
1335         return ret;
1336
1337 error:
1338         mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n",
1339              dlm->name, ret);
1340         if (!dlm_is_host_down(ret))
1341                 BUG();
1342         mlog(0, "%s: node %u went down while sending %s "
1343              "lockres %.*s\n", dlm->name, send_to,
1344              flags & DLM_MRES_RECOVERY ?  "recovery" : "migration",
1345              res->lockname.len, res->lockname.name);
1346         return ret;
1347 }
1348
1349
1350
1351 /*
1352  * this message will contain no more than one page worth of
1353  * recovery data, and it will work on only one lockres.
1354  * there may be many locks in this page, and we may need to wait
1355  * for additional packets to complete all the locks (rare, but
1356  * possible).
1357  */
1358 /*
1359  * NOTE: the allocation error cases here are scary
1360  * we really cannot afford to fail an alloc in recovery
1361  * do we spin?  returning an error only delays the problem really
1362  */
1363
1364 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
1365                             void **ret_data)
1366 {
1367         struct dlm_ctxt *dlm = data;
1368         struct dlm_migratable_lockres *mres =
1369                 (struct dlm_migratable_lockres *)msg->buf;
1370         int ret = 0;
1371         u8 real_master;
1372         u8 extra_refs = 0;
1373         char *buf = NULL;
1374         struct dlm_work_item *item = NULL;
1375         struct dlm_lock_resource *res = NULL;
1376         unsigned int hash;
1377
1378         if (!dlm_grab(dlm))
1379                 return -EINVAL;
1380
1381         BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1382
1383         real_master = mres->master;
1384         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1385                 /* cannot migrate a lockres with no master */
1386                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1387         }
1388
1389         mlog(0, "%s message received from node %u\n",
1390                   (mres->flags & DLM_MRES_RECOVERY) ?
1391                   "recovery" : "migration", mres->master);
1392         if (mres->flags & DLM_MRES_ALL_DONE)
1393                 mlog(0, "all done flag.  all lockres data received!\n");
1394
1395         ret = -ENOMEM;
1396         buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS);
1397         item = kzalloc(sizeof(*item), GFP_NOFS);
1398         if (!buf || !item)
1399                 goto leave;
1400
1401         /* lookup the lock to see if we have a secondary queue for this
1402          * already...  just add the locks in and this will have its owner
1403          * and RECOVERY flag changed when it completes. */
1404         hash = dlm_lockid_hash(mres->lockname, mres->lockname_len);
1405         spin_lock(&dlm->spinlock);
1406         res = __dlm_lookup_lockres_full(dlm, mres->lockname, mres->lockname_len,
1407                         hash);
1408         if (res) {
1409                 /* this will get a ref on res */
1410                 /* mark it as recovering/migrating and hash it */
1411                 spin_lock(&res->spinlock);
1412                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
1413                         mlog(0, "%s: node is attempting to migrate "
1414                                 "lockres %.*s, but marked as dropping "
1415                                 " ref!\n", dlm->name,
1416                                 mres->lockname_len, mres->lockname);
1417                         ret = -EINVAL;
1418                         spin_unlock(&res->spinlock);
1419                         spin_unlock(&dlm->spinlock);
1420                         dlm_lockres_put(res);
1421                         goto leave;
1422                 }
1423
1424                 if (mres->flags & DLM_MRES_RECOVERY) {
1425                         res->state |= DLM_LOCK_RES_RECOVERING;
1426                 } else {
1427                         if (res->state & DLM_LOCK_RES_MIGRATING) {
1428                                 /* this is at least the second
1429                                  * lockres message */
1430                                 mlog(0, "lock %.*s is already migrating\n",
1431                                           mres->lockname_len,
1432                                           mres->lockname);
1433                         } else if (res->state & DLM_LOCK_RES_RECOVERING) {
1434                                 /* caller should BUG */
1435                                 mlog(ML_ERROR, "node is attempting to migrate "
1436                                      "lock %.*s, but marked as recovering!\n",
1437                                      mres->lockname_len, mres->lockname);
1438                                 ret = -EFAULT;
1439                                 spin_unlock(&res->spinlock);
1440                                 spin_unlock(&dlm->spinlock);
1441                                 dlm_lockres_put(res);
1442                                 goto leave;
1443                         }
1444                         res->state |= DLM_LOCK_RES_MIGRATING;
1445                 }
1446                 spin_unlock(&res->spinlock);
1447                 spin_unlock(&dlm->spinlock);
1448         } else {
1449                 spin_unlock(&dlm->spinlock);
1450                 /* need to allocate, just like if it was
1451                  * mastered here normally  */
1452                 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1453                 if (!res)
1454                         goto leave;
1455
1456                 /* to match the ref that we would have gotten if
1457                  * dlm_lookup_lockres had succeeded */
1458                 dlm_lockres_get(res);
1459
1460                 /* mark it as recovering/migrating and hash it */
1461                 if (mres->flags & DLM_MRES_RECOVERY)
1462                         res->state |= DLM_LOCK_RES_RECOVERING;
1463                 else
1464                         res->state |= DLM_LOCK_RES_MIGRATING;
1465
1466                 spin_lock(&dlm->spinlock);
1467                 __dlm_insert_lockres(dlm, res);
1468                 spin_unlock(&dlm->spinlock);
1469
1470                 /* Add an extra ref for this lock-less lockres lest the
1471                  * dlm_thread purges it before we get the chance to add
1472                  * locks to it */
1473                 dlm_lockres_get(res);
1474
1475                 /* There are three refs that need to be put.
1476                  * 1. Taken above.
1477                  * 2. kref_init in dlm_new_lockres()->dlm_init_lockres().
1478                  * 3. dlm_lookup_lockres()
1479                  * The first one is handled at the end of this function. The
1480                  * other two are handled in the worker thread after locks have
1481                  * been attached. Yes, we don't wait for purge time to match
1482                  * kref_init. The lockres will still have atleast one ref
1483                  * added because it is in the hash __dlm_insert_lockres() */
1484                 extra_refs++;
1485
1486                 /* now that the new lockres is inserted,
1487                  * make it usable by other processes */
1488                 spin_lock(&res->spinlock);
1489                 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1490                 spin_unlock(&res->spinlock);
1491                 wake_up(&res->wq);
1492         }
1493
1494         /* at this point we have allocated everything we need,
1495          * and we have a hashed lockres with an extra ref and
1496          * the proper res->state flags. */
1497         ret = 0;
1498         spin_lock(&res->spinlock);
1499         /* drop this either when master requery finds a different master
1500          * or when a lock is added by the recovery worker */
1501         dlm_lockres_grab_inflight_ref(dlm, res);
1502         if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1503                 /* migration cannot have an unknown master */
1504                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1505                 mlog(0, "recovery has passed me a lockres with an "
1506                           "unknown owner.. will need to requery: "
1507                           "%.*s\n", mres->lockname_len, mres->lockname);
1508         } else {
1509                 /* take a reference now to pin the lockres, drop it
1510                  * when locks are added in the worker */
1511                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
1512         }
1513         spin_unlock(&res->spinlock);
1514
1515         /* queue up work for dlm_mig_lockres_worker */
1516         dlm_grab(dlm);  /* get an extra ref for the work item */
1517         memcpy(buf, msg->buf, be16_to_cpu(msg->data_len));  /* copy the whole message */
1518         dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1519         item->u.ml.lockres = res; /* already have a ref */
1520         item->u.ml.real_master = real_master;
1521         item->u.ml.extra_ref = extra_refs;
1522         spin_lock(&dlm->work_lock);
1523         list_add_tail(&item->list, &dlm->work_list);
1524         spin_unlock(&dlm->work_lock);
1525         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
1526
1527 leave:
1528         /* One extra ref taken needs to be put here */
1529         if (extra_refs)
1530                 dlm_lockres_put(res);
1531
1532         dlm_put(dlm);
1533         if (ret < 0) {
1534                 kfree(buf);
1535                 kfree(item);
1536                 mlog_errno(ret);
1537         }
1538
1539         return ret;
1540 }
1541
1542
1543 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1544 {
1545         struct dlm_ctxt *dlm;
1546         struct dlm_migratable_lockres *mres;
1547         int ret = 0;
1548         struct dlm_lock_resource *res;
1549         u8 real_master;
1550         u8 extra_ref;
1551
1552         dlm = item->dlm;
1553         mres = (struct dlm_migratable_lockres *)data;
1554
1555         res = item->u.ml.lockres;
1556         real_master = item->u.ml.real_master;
1557         extra_ref = item->u.ml.extra_ref;
1558
1559         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1560                 /* this case is super-rare. only occurs if
1561                  * node death happens during migration. */
1562 again:
1563                 ret = dlm_lockres_master_requery(dlm, res, &real_master);
1564                 if (ret < 0) {
1565                         mlog(0, "dlm_lockres_master_requery ret=%d\n",
1566                                   ret);
1567                         goto again;
1568                 }
1569                 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1570                         mlog(0, "lockres %.*s not claimed.  "
1571                                    "this node will take it.\n",
1572                                    res->lockname.len, res->lockname.name);
1573                 } else {
1574                         spin_lock(&res->spinlock);
1575                         dlm_lockres_drop_inflight_ref(dlm, res);
1576                         spin_unlock(&res->spinlock);
1577                         mlog(0, "master needs to respond to sender "
1578                                   "that node %u still owns %.*s\n",
1579                                   real_master, res->lockname.len,
1580                                   res->lockname.name);
1581                         /* cannot touch this lockres */
1582                         goto leave;
1583                 }
1584         }
1585
1586         ret = dlm_process_recovery_data(dlm, res, mres);
1587         if (ret < 0)
1588                 mlog(0, "dlm_process_recovery_data returned  %d\n", ret);
1589         else
1590                 mlog(0, "dlm_process_recovery_data succeeded\n");
1591
1592         if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1593                            (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1594                 ret = dlm_finish_migration(dlm, res, mres->master);
1595                 if (ret < 0)
1596                         mlog_errno(ret);
1597         }
1598
1599 leave:
1600         /* See comment in dlm_mig_lockres_handler() */
1601         if (res) {
1602                 if (extra_ref)
1603                         dlm_lockres_put(res);
1604                 dlm_lockres_put(res);
1605         }
1606         kfree(data);
1607 }
1608
1609
1610
1611 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1612                                       struct dlm_lock_resource *res,
1613                                       u8 *real_master)
1614 {
1615         struct dlm_node_iter iter;
1616         int nodenum;
1617         int ret = 0;
1618
1619         *real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1620
1621         /* we only reach here if one of the two nodes in a
1622          * migration died while the migration was in progress.
1623          * at this point we need to requery the master.  we
1624          * know that the new_master got as far as creating
1625          * an mle on at least one node, but we do not know
1626          * if any nodes had actually cleared the mle and set
1627          * the master to the new_master.  the old master
1628          * is supposed to set the owner to UNKNOWN in the
1629          * event of a new_master death, so the only possible
1630          * responses that we can get from nodes here are
1631          * that the master is new_master, or that the master
1632          * is UNKNOWN.
1633          * if all nodes come back with UNKNOWN then we know
1634          * the lock needs remastering here.
1635          * if any node comes back with a valid master, check
1636          * to see if that master is the one that we are
1637          * recovering.  if so, then the new_master died and
1638          * we need to remaster this lock.  if not, then the
1639          * new_master survived and that node will respond to
1640          * other nodes about the owner.
1641          * if there is an owner, this node needs to dump this
1642          * lockres and alert the sender that this lockres
1643          * was rejected. */
1644         spin_lock(&dlm->spinlock);
1645         dlm_node_iter_init(dlm->domain_map, &iter);
1646         spin_unlock(&dlm->spinlock);
1647
1648         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1649                 /* do not send to self */
1650                 if (nodenum == dlm->node_num)
1651                         continue;
1652                 ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1653                 if (ret < 0) {
1654                         mlog_errno(ret);
1655                         if (!dlm_is_host_down(ret))
1656                                 BUG();
1657                         /* host is down, so answer for that node would be
1658                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
1659                 }
1660                 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1661                         mlog(0, "lock master is %u\n", *real_master);
1662                         break;
1663                 }
1664         }
1665         return ret;
1666 }
1667
1668
1669 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1670                           u8 nodenum, u8 *real_master)
1671 {
1672         int ret = -EINVAL;
1673         struct dlm_master_requery req;
1674         int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1675
1676         memset(&req, 0, sizeof(req));
1677         req.node_idx = dlm->node_num;
1678         req.namelen = res->lockname.len;
1679         memcpy(req.name, res->lockname.name, res->lockname.len);
1680
1681 resend:
1682         ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1683                                  &req, sizeof(req), nodenum, &status);
1684         if (ret < 0)
1685                 mlog(ML_ERROR, "Error %d when sending message %u (key "
1686                      "0x%x) to node %u\n", ret, DLM_MASTER_REQUERY_MSG,
1687                      dlm->key, nodenum);
1688         else if (status == -ENOMEM) {
1689                 mlog_errno(status);
1690                 msleep(50);
1691                 goto resend;
1692         } else {
1693                 BUG_ON(status < 0);
1694                 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1695                 *real_master = (u8) (status & 0xff);
1696                 mlog(0, "node %u responded to master requery with %u\n",
1697                           nodenum, *real_master);
1698                 ret = 0;
1699         }
1700         return ret;
1701 }
1702
1703
1704 /* this function cannot error, so unless the sending
1705  * or receiving of the message failed, the owner can
1706  * be trusted */
1707 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
1708                                void **ret_data)
1709 {
1710         struct dlm_ctxt *dlm = data;
1711         struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1712         struct dlm_lock_resource *res = NULL;
1713         unsigned int hash;
1714         int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1715         u32 flags = DLM_ASSERT_MASTER_REQUERY;
1716         int dispatched = 0;
1717
1718         if (!dlm_grab(dlm)) {
1719                 /* since the domain has gone away on this
1720                  * node, the proper response is UNKNOWN */
1721                 return master;
1722         }
1723
1724         hash = dlm_lockid_hash(req->name, req->namelen);
1725
1726         spin_lock(&dlm->spinlock);
1727         res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash);
1728         if (res) {
1729                 spin_lock(&res->spinlock);
1730                 master = res->owner;
1731                 if (master == dlm->node_num) {
1732                         int ret = dlm_dispatch_assert_master(dlm, res,
1733                                                              0, 0, flags);
1734                         if (ret < 0) {
1735                                 mlog_errno(ret);
1736                                 spin_unlock(&res->spinlock);
1737                                 dlm_lockres_put(res);
1738                                 spin_unlock(&dlm->spinlock);
1739                                 dlm_put(dlm);
1740                                 /* sender will take care of this and retry */
1741                                 return ret;
1742                         } else {
1743                                 dispatched = 1;
1744                                 __dlm_lockres_grab_inflight_worker(dlm, res);
1745                                 spin_unlock(&res->spinlock);
1746                         }
1747                 } else {
1748                         /* put.. incase we are not the master */
1749                         spin_unlock(&res->spinlock);
1750                         dlm_lockres_put(res);
1751                 }
1752         }
1753         spin_unlock(&dlm->spinlock);
1754
1755         if (!dispatched)
1756                 dlm_put(dlm);
1757         return master;
1758 }
1759
1760 static inline struct list_head *
1761 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1762 {
1763         struct list_head *ret;
1764         BUG_ON(list_num < 0);
1765         BUG_ON(list_num > 2);
1766         ret = &(res->granted);
1767         ret += list_num;
1768         return ret;
1769 }
1770 /* TODO: do ast flush business
1771  * TODO: do MIGRATING and RECOVERING spinning
1772  */
1773
1774 /*
1775 * NOTE about in-flight requests during migration:
1776 *
1777 * Before attempting the migrate, the master has marked the lockres as
1778 * MIGRATING and then flushed all of its pending ASTS.  So any in-flight
1779 * requests either got queued before the MIGRATING flag got set, in which
1780 * case the lock data will reflect the change and a return message is on
1781 * the way, or the request failed to get in before MIGRATING got set.  In
1782 * this case, the caller will be told to spin and wait for the MIGRATING
1783 * flag to be dropped, then recheck the master.
1784 * This holds true for the convert, cancel and unlock cases, and since lvb
1785 * updates are tied to these same messages, it applies to lvb updates as
1786 * well.  For the lock case, there is no way a lock can be on the master
1787 * queue and not be on the secondary queue since the lock is always added
1788 * locally first.  This means that the new target node will never be sent
1789 * a lock that he doesn't already have on the list.
1790 * In total, this means that the local lock is correct and should not be
1791 * updated to match the one sent by the master.  Any messages sent back
1792 * from the master before the MIGRATING flag will bring the lock properly
1793 * up-to-date, and the change will be ordered properly for the waiter.
1794 * We will *not* attempt to modify the lock underneath the waiter.
1795 */
1796
1797 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1798                                      struct dlm_lock_resource *res,
1799                                      struct dlm_migratable_lockres *mres)
1800 {
1801         struct dlm_migratable_lock *ml;
1802         struct list_head *queue, *iter;
1803         struct list_head *tmpq = NULL;
1804         struct dlm_lock *newlock = NULL;
1805         struct dlm_lockstatus *lksb = NULL;
1806         int ret = 0;
1807         int i, j, bad;
1808         struct dlm_lock *lock;
1809         u8 from = O2NM_MAX_NODES;
1810         unsigned int added = 0;
1811         __be64 c;
1812
1813         mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1814         for (i=0; i<mres->num_locks; i++) {
1815                 ml = &(mres->ml[i]);
1816
1817                 if (dlm_is_dummy_lock(dlm, ml, &from)) {
1818                         /* placeholder, just need to set the refmap bit */
1819                         BUG_ON(mres->num_locks != 1);
1820                         mlog(0, "%s:%.*s: dummy lock for %u\n",
1821                              dlm->name, mres->lockname_len, mres->lockname,
1822                              from);
1823                         spin_lock(&res->spinlock);
1824                         dlm_lockres_set_refmap_bit(dlm, res, from);
1825                         spin_unlock(&res->spinlock);
1826                         added++;
1827                         break;
1828                 }
1829                 BUG_ON(ml->highest_blocked != LKM_IVMODE);
1830                 newlock = NULL;
1831                 lksb = NULL;
1832
1833                 queue = dlm_list_num_to_pointer(res, ml->list);
1834                 tmpq = NULL;
1835
1836                 /* if the lock is for the local node it needs to
1837                  * be moved to the proper location within the queue.
1838                  * do not allocate a new lock structure. */
1839                 if (ml->node == dlm->node_num) {
1840                         /* MIGRATION ONLY! */
1841                         BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1842
1843                         lock = NULL;
1844                         spin_lock(&res->spinlock);
1845                         for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) {
1846                                 tmpq = dlm_list_idx_to_ptr(res, j);
1847                                 list_for_each(iter, tmpq) {
1848                                         lock = list_entry(iter,
1849                                                   struct dlm_lock, list);
1850                                         if (lock->ml.cookie == ml->cookie)
1851                                                 break;
1852                                         lock = NULL;
1853                                 }
1854                                 if (lock)
1855                                         break;
1856                         }
1857
1858                         /* lock is always created locally first, and
1859                          * destroyed locally last.  it must be on the list */
1860                         if (!lock) {
1861                                 c = ml->cookie;
1862                                 mlog(ML_ERROR, "Could not find local lock "
1863                                                "with cookie %u:%llu, node %u, "
1864                                                "list %u, flags 0x%x, type %d, "
1865                                                "conv %d, highest blocked %d\n",
1866                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1867                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1868                                      ml->node, ml->list, ml->flags, ml->type,
1869                                      ml->convert_type, ml->highest_blocked);
1870                                 __dlm_print_one_lock_resource(res);
1871                                 BUG();
1872                         }
1873
1874                         if (lock->ml.node != ml->node) {
1875                                 c = lock->ml.cookie;
1876                                 mlog(ML_ERROR, "Mismatched node# in lock "
1877                                      "cookie %u:%llu, name %.*s, node %u\n",
1878                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1879                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1880                                      res->lockname.len, res->lockname.name,
1881                                      lock->ml.node);
1882                                 c = ml->cookie;
1883                                 mlog(ML_ERROR, "Migrate lock cookie %u:%llu, "
1884                                      "node %u, list %u, flags 0x%x, type %d, "
1885                                      "conv %d, highest blocked %d\n",
1886                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1887                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1888                                      ml->node, ml->list, ml->flags, ml->type,
1889                                      ml->convert_type, ml->highest_blocked);
1890                                 __dlm_print_one_lock_resource(res);
1891                                 BUG();
1892                         }
1893
1894                         if (tmpq != queue) {
1895                                 c = ml->cookie;
1896                                 mlog(0, "Lock cookie %u:%llu was on list %u "
1897                                      "instead of list %u for %.*s\n",
1898                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1899                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1900                                      j, ml->list, res->lockname.len,
1901                                      res->lockname.name);
1902                                 __dlm_print_one_lock_resource(res);
1903                                 spin_unlock(&res->spinlock);
1904                                 continue;
1905                         }
1906
1907                         /* see NOTE above about why we do not update
1908                          * to match the master here */
1909
1910                         /* move the lock to its proper place */
1911                         /* do not alter lock refcount.  switching lists. */
1912                         list_move_tail(&lock->list, queue);
1913                         spin_unlock(&res->spinlock);
1914                         added++;
1915
1916                         mlog(0, "just reordered a local lock!\n");
1917                         continue;
1918                 }
1919
1920                 /* lock is for another node. */
1921                 newlock = dlm_new_lock(ml->type, ml->node,
1922                                        be64_to_cpu(ml->cookie), NULL);
1923                 if (!newlock) {
1924                         ret = -ENOMEM;
1925                         goto leave;
1926                 }
1927                 lksb = newlock->lksb;
1928                 dlm_lock_attach_lockres(newlock, res);
1929
1930                 if (ml->convert_type != LKM_IVMODE) {
1931                         BUG_ON(queue != &res->converting);
1932                         newlock->ml.convert_type = ml->convert_type;
1933                 }
1934                 lksb->flags |= (ml->flags &
1935                                 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1936
1937                 if (ml->type == LKM_NLMODE)
1938                         goto skip_lvb;
1939
1940                 /*
1941                  * If the lock is in the blocked list it can't have a valid lvb,
1942                  * so skip it
1943                  */
1944                 if (ml->list == DLM_BLOCKED_LIST)
1945                         goto skip_lvb;
1946
1947                 if (!dlm_lvb_is_empty(mres->lvb)) {
1948                         if (lksb->flags & DLM_LKSB_PUT_LVB) {
1949                                 /* other node was trying to update
1950                                  * lvb when node died.  recreate the
1951                                  * lksb with the updated lvb. */
1952                                 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1953                                 /* the lock resource lvb update must happen
1954                                  * NOW, before the spinlock is dropped.
1955                                  * we no longer wait for the AST to update
1956                                  * the lvb. */
1957                                 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1958                         } else {
1959                                 /* otherwise, the node is sending its
1960                                  * most recent valid lvb info */
1961                                 BUG_ON(ml->type != LKM_EXMODE &&
1962                                        ml->type != LKM_PRMODE);
1963                                 if (!dlm_lvb_is_empty(res->lvb) &&
1964                                     (ml->type == LKM_EXMODE ||
1965                                      memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1966                                         int i;
1967                                         mlog(ML_ERROR, "%s:%.*s: received bad "
1968                                              "lvb! type=%d\n", dlm->name,
1969                                              res->lockname.len,
1970                                              res->lockname.name, ml->type);
1971                                         printk("lockres lvb=[");
1972                                         for (i=0; i<DLM_LVB_LEN; i++)
1973                                                 printk("%02x", res->lvb[i]);
1974                                         printk("]\nmigrated lvb=[");
1975                                         for (i=0; i<DLM_LVB_LEN; i++)
1976                                                 printk("%02x", mres->lvb[i]);
1977                                         printk("]\n");
1978                                         dlm_print_one_lock_resource(res);
1979                                         BUG();
1980                                 }
1981                                 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1982                         }
1983                 }
1984 skip_lvb:
1985
1986                 /* NOTE:
1987                  * wrt lock queue ordering and recovery:
1988                  *    1. order of locks on granted queue is
1989                  *       meaningless.
1990                  *    2. order of locks on converting queue is
1991                  *       LOST with the node death.  sorry charlie.
1992                  *    3. order of locks on the blocked queue is
1993                  *       also LOST.
1994                  * order of locks does not affect integrity, it
1995                  * just means that a lock request may get pushed
1996                  * back in line as a result of the node death.
1997                  * also note that for a given node the lock order
1998                  * for its secondary queue locks is preserved
1999                  * relative to each other, but clearly *not*
2000                  * preserved relative to locks from other nodes.
2001                  */
2002                 bad = 0;
2003                 spin_lock(&res->spinlock);
2004                 list_for_each_entry(lock, queue, list) {
2005                         if (lock->ml.cookie == ml->cookie) {
2006                                 c = lock->ml.cookie;
2007                                 mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already "
2008                                      "exists on this lockres!\n", dlm->name,
2009                                      res->lockname.len, res->lockname.name,
2010                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
2011                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)));
2012
2013                                 mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, "
2014                                      "node=%u, cookie=%u:%llu, queue=%d\n",
2015                                      ml->type, ml->convert_type, ml->node,
2016                                      dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)),
2017                                      dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)),
2018                                      ml->list);
2019
2020                                 __dlm_print_one_lock_resource(res);
2021                                 bad = 1;
2022                                 break;
2023                         }
2024                 }
2025                 if (!bad) {
2026                         dlm_lock_get(newlock);
2027                         if (mres->flags & DLM_MRES_RECOVERY &&
2028                                         ml->list == DLM_CONVERTING_LIST &&
2029                                         newlock->ml.type >
2030                                         newlock->ml.convert_type) {
2031                                 /* newlock is doing downconvert, add it to the
2032                                  * head of converting list */
2033                                 list_add(&newlock->list, queue);
2034                         } else
2035                                 list_add_tail(&newlock->list, queue);
2036                         mlog(0, "%s:%.*s: added lock for node %u, "
2037                              "setting refmap bit\n", dlm->name,
2038                              res->lockname.len, res->lockname.name, ml->node);
2039                         dlm_lockres_set_refmap_bit(dlm, res, ml->node);
2040                         added++;
2041                 }
2042                 spin_unlock(&res->spinlock);
2043         }
2044         mlog(0, "done running all the locks\n");
2045
2046 leave:
2047         /* balance the ref taken when the work was queued */
2048         spin_lock(&res->spinlock);
2049         dlm_lockres_drop_inflight_ref(dlm, res);
2050         spin_unlock(&res->spinlock);
2051
2052         if (ret < 0)
2053                 mlog_errno(ret);
2054
2055         return ret;
2056 }
2057
2058 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
2059                                        struct dlm_lock_resource *res)
2060 {
2061         int i;
2062         struct list_head *queue;
2063         struct dlm_lock *lock, *next;
2064
2065         assert_spin_locked(&dlm->spinlock);
2066         assert_spin_locked(&res->spinlock);
2067         res->state |= DLM_LOCK_RES_RECOVERING;
2068         if (!list_empty(&res->recovering)) {
2069                 mlog(0,
2070                      "Recovering res %s:%.*s, is already on recovery list!\n",
2071                      dlm->name, res->lockname.len, res->lockname.name);
2072                 list_del_init(&res->recovering);
2073                 dlm_lockres_put(res);
2074         }
2075         /* We need to hold a reference while on the recovery list */
2076         dlm_lockres_get(res);
2077         list_add_tail(&res->recovering, &dlm->reco.resources);
2078
2079         /* find any pending locks and put them back on proper list */
2080         for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
2081                 queue = dlm_list_idx_to_ptr(res, i);
2082                 list_for_each_entry_safe(lock, next, queue, list) {
2083                         dlm_lock_get(lock);
2084                         if (lock->convert_pending) {
2085                                 /* move converting lock back to granted */
2086                                 mlog(0, "node died with convert pending "
2087                                      "on %.*s. move back to granted list.\n",
2088                                      res->lockname.len, res->lockname.name);
2089                                 dlm_revert_pending_convert(res, lock);
2090                                 lock->convert_pending = 0;
2091                         } else if (lock->lock_pending) {
2092                                 /* remove pending lock requests completely */
2093                                 BUG_ON(i != DLM_BLOCKED_LIST);
2094                                 mlog(0, "node died with lock pending "
2095                                      "on %.*s. remove from blocked list and skip.\n",
2096                                      res->lockname.len, res->lockname.name);
2097                                 /* lock will be floating until ref in
2098                                  * dlmlock_remote is freed after the network
2099                                  * call returns.  ok for it to not be on any
2100                                  * list since no ast can be called
2101                                  * (the master is dead). */
2102                                 dlm_revert_pending_lock(res, lock);
2103                                 lock->lock_pending = 0;
2104                         } else if (lock->unlock_pending) {
2105                                 /* if an unlock was in progress, treat as
2106                                  * if this had completed successfully
2107                                  * before sending this lock state to the
2108                                  * new master.  note that the dlm_unlock
2109                                  * call is still responsible for calling
2110                                  * the unlockast.  that will happen after
2111                                  * the network call times out.  for now,
2112                                  * just move lists to prepare the new
2113                                  * recovery master.  */
2114                                 BUG_ON(i != DLM_GRANTED_LIST);
2115                                 mlog(0, "node died with unlock pending "
2116                                      "on %.*s. remove from blocked list and skip.\n",
2117                                      res->lockname.len, res->lockname.name);
2118                                 dlm_commit_pending_unlock(res, lock);
2119                                 lock->unlock_pending = 0;
2120                         } else if (lock->cancel_pending) {
2121                                 /* if a cancel was in progress, treat as
2122                                  * if this had completed successfully
2123                                  * before sending this lock state to the
2124                                  * new master */
2125                                 BUG_ON(i != DLM_CONVERTING_LIST);
2126                                 mlog(0, "node died with cancel pending "
2127                                      "on %.*s. move back to granted list.\n",
2128                                      res->lockname.len, res->lockname.name);
2129                                 dlm_commit_pending_cancel(res, lock);
2130                                 lock->cancel_pending = 0;
2131                         }
2132                         dlm_lock_put(lock);
2133                 }
2134         }
2135 }
2136
2137
2138
2139 /* removes all recovered locks from the recovery list.
2140  * sets the res->owner to the new master.
2141  * unsets the RECOVERY flag and wakes waiters. */
2142 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
2143                                               u8 dead_node, u8 new_master)
2144 {
2145         int i;
2146         struct hlist_head *bucket;
2147         struct dlm_lock_resource *res, *next;
2148
2149         assert_spin_locked(&dlm->spinlock);
2150
2151         list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
2152                 if (res->owner == dead_node) {
2153                         mlog(0, "%s: res %.*s, Changing owner from %u to %u\n",
2154                              dlm->name, res->lockname.len, res->lockname.name,
2155                              res->owner, new_master);
2156                         list_del_init(&res->recovering);
2157                         spin_lock(&res->spinlock);
2158                         /* new_master has our reference from
2159                          * the lock state sent during recovery */
2160                         dlm_change_lockres_owner(dlm, res, new_master);
2161                         res->state &= ~DLM_LOCK_RES_RECOVERING;
2162                         if (__dlm_lockres_has_locks(res))
2163                                 __dlm_dirty_lockres(dlm, res);
2164                         spin_unlock(&res->spinlock);
2165                         wake_up(&res->wq);
2166                         dlm_lockres_put(res);
2167                 }
2168         }
2169
2170         /* this will become unnecessary eventually, but
2171          * for now we need to run the whole hash, clear
2172          * the RECOVERING state and set the owner
2173          * if necessary */
2174         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2175                 bucket = dlm_lockres_hash(dlm, i);
2176                 hlist_for_each_entry(res, bucket, hash_node) {
2177                         if (res->state & DLM_LOCK_RES_RECOVERY_WAITING) {
2178                                 spin_lock(&res->spinlock);
2179                                 res->state &= ~DLM_LOCK_RES_RECOVERY_WAITING;
2180                                 spin_unlock(&res->spinlock);
2181                                 wake_up(&res->wq);
2182                         }
2183
2184                         if (!(res->state & DLM_LOCK_RES_RECOVERING))
2185                                 continue;
2186
2187                         if (res->owner != dead_node &&
2188                             res->owner != dlm->node_num)
2189                                 continue;
2190
2191                         if (!list_empty(&res->recovering)) {
2192                                 list_del_init(&res->recovering);
2193                                 dlm_lockres_put(res);
2194                         }
2195
2196                         /* new_master has our reference from
2197                          * the lock state sent during recovery */
2198                         mlog(0, "%s: res %.*s, Changing owner from %u to %u\n",
2199                              dlm->name, res->lockname.len, res->lockname.name,
2200                              res->owner, new_master);
2201                         spin_lock(&res->spinlock);
2202                         dlm_change_lockres_owner(dlm, res, new_master);
2203                         res->state &= ~DLM_LOCK_RES_RECOVERING;
2204                         if (__dlm_lockres_has_locks(res))
2205                                 __dlm_dirty_lockres(dlm, res);
2206                         spin_unlock(&res->spinlock);
2207                         wake_up(&res->wq);
2208                 }
2209         }
2210 }
2211
2212 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
2213 {
2214         if (local) {
2215                 if (lock->ml.type != LKM_EXMODE &&
2216                     lock->ml.type != LKM_PRMODE)
2217                         return 1;
2218         } else if (lock->ml.type == LKM_EXMODE)
2219                 return 1;
2220         return 0;
2221 }
2222
2223 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
2224                                struct dlm_lock_resource *res, u8 dead_node)
2225 {
2226         struct list_head *queue;
2227         struct dlm_lock *lock;
2228         int blank_lvb = 0, local = 0;
2229         int i;
2230         u8 search_node;
2231
2232         assert_spin_locked(&dlm->spinlock);
2233         assert_spin_locked(&res->spinlock);
2234
2235         if (res->owner == dlm->node_num)
2236                 /* if this node owned the lockres, and if the dead node
2237                  * had an EX when he died, blank out the lvb */
2238                 search_node = dead_node;
2239         else {
2240                 /* if this is a secondary lockres, and we had no EX or PR
2241                  * locks granted, we can no longer trust the lvb */
2242                 search_node = dlm->node_num;
2243                 local = 1;  /* check local state for valid lvb */
2244         }
2245
2246         for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
2247                 queue = dlm_list_idx_to_ptr(res, i);
2248                 list_for_each_entry(lock, queue, list) {
2249                         if (lock->ml.node == search_node) {
2250                                 if (dlm_lvb_needs_invalidation(lock, local)) {
2251                                         /* zero the lksb lvb and lockres lvb */
2252                                         blank_lvb = 1;
2253                                         memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
2254                                 }
2255                         }
2256                 }
2257         }
2258
2259         if (blank_lvb) {
2260                 mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
2261                      res->lockname.len, res->lockname.name, dead_node);
2262                 memset(res->lvb, 0, DLM_LVB_LEN);
2263         }
2264 }
2265
2266 static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
2267                                 struct dlm_lock_resource *res, u8 dead_node)
2268 {
2269         struct dlm_lock *lock, *next;
2270         unsigned int freed = 0;
2271
2272         /* this node is the lockres master:
2273          * 1) remove any stale locks for the dead node
2274          * 2) if the dead node had an EX when he died, blank out the lvb
2275          */
2276         assert_spin_locked(&dlm->spinlock);
2277         assert_spin_locked(&res->spinlock);
2278
2279         /* We do two dlm_lock_put(). One for removing from list and the other is
2280          * to force the DLM_UNLOCK_FREE_LOCK action so as to free the locks */
2281
2282         /* TODO: check pending_asts, pending_basts here */
2283         list_for_each_entry_safe(lock, next, &res->granted, list) {
2284                 if (lock->ml.node == dead_node) {
2285                         list_del_init(&lock->list);
2286                         dlm_lock_put(lock);
2287                         /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2288                         dlm_lock_put(lock);
2289                         freed++;
2290                 }
2291         }
2292         list_for_each_entry_safe(lock, next, &res->converting, list) {
2293                 if (lock->ml.node == dead_node) {
2294                         list_del_init(&lock->list);
2295                         dlm_lock_put(lock);
2296                         /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2297                         dlm_lock_put(lock);
2298                         freed++;
2299                 }
2300         }
2301         list_for_each_entry_safe(lock, next, &res->blocked, list) {
2302                 if (lock->ml.node == dead_node) {
2303                         list_del_init(&lock->list);
2304                         dlm_lock_put(lock);
2305                         /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2306                         dlm_lock_put(lock);
2307                         freed++;
2308                 }
2309         }
2310
2311         if (freed) {
2312                 mlog(0, "%s:%.*s: freed %u locks for dead node %u, "
2313                      "dropping ref from lockres\n", dlm->name,
2314                      res->lockname.len, res->lockname.name, freed, dead_node);
2315                 if(!test_bit(dead_node, res->refmap)) {
2316                         mlog(ML_ERROR, "%s:%.*s: freed %u locks for dead node %u, "
2317                              "but ref was not set\n", dlm->name,
2318                              res->lockname.len, res->lockname.name, freed, dead_node);
2319                         __dlm_print_one_lock_resource(res);
2320                 }
2321                 res->state |= DLM_LOCK_RES_RECOVERY_WAITING;
2322                 dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2323         } else if (test_bit(dead_node, res->refmap)) {
2324                 mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2325                      "no locks and had not purged before dying\n", dlm->name,
2326                      res->lockname.len, res->lockname.name, dead_node);
2327                 dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2328         }
2329
2330         /* do not kick thread yet */
2331         __dlm_dirty_lockres(dlm, res);
2332 }
2333
2334 /* if this node is the recovery master, and there are no
2335  * locks for a given lockres owned by this node that are in
2336  * either PR or EX mode, zero out the lvb before requesting.
2337  *
2338  */
2339
2340
2341 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
2342 {
2343         struct dlm_lock_resource *res;
2344         int i;
2345         struct hlist_head *bucket;
2346         struct hlist_node *tmp;
2347         struct dlm_lock *lock;
2348
2349
2350         /* purge any stale mles */
2351         dlm_clean_master_list(dlm, dead_node);
2352
2353         /*
2354          * now clean up all lock resources.  there are two rules:
2355          *
2356          * 1) if the dead node was the master, move the lockres
2357          *    to the recovering list.  set the RECOVERING flag.
2358          *    this lockres needs to be cleaned up before it can
2359          *    be used further.
2360          *
2361          * 2) if this node was the master, remove all locks from
2362          *    each of the lockres queues that were owned by the
2363          *    dead node.  once recovery finishes, the dlm thread
2364          *    can be kicked again to see if any ASTs or BASTs
2365          *    need to be fired as a result.
2366          */
2367         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2368                 bucket = dlm_lockres_hash(dlm, i);
2369                 hlist_for_each_entry_safe(res, tmp, bucket, hash_node) {
2370                         /* always prune any $RECOVERY entries for dead nodes,
2371                          * otherwise hangs can occur during later recovery */
2372                         if (dlm_is_recovery_lock(res->lockname.name,
2373                                                  res->lockname.len)) {
2374                                 spin_lock(&res->spinlock);
2375                                 list_for_each_entry(lock, &res->granted, list) {
2376                                         if (lock->ml.node == dead_node) {
2377                                                 mlog(0, "AHA! there was "
2378                                                      "a $RECOVERY lock for dead "
2379                                                      "node %u (%s)!\n",
2380                                                      dead_node, dlm->name);
2381                                                 list_del_init(&lock->list);
2382                                                 dlm_lock_put(lock);
2383                                                 /* Can't schedule
2384                                                  * DLM_UNLOCK_FREE_LOCK
2385                                                  * - do manually */
2386                                                 dlm_lock_put(lock);
2387                                                 break;
2388                                         }
2389                                 }
2390
2391                                 if ((res->owner == dead_node) &&
2392                                                         (res->state & DLM_LOCK_RES_DROPPING_REF)) {
2393                                         dlm_lockres_get(res);
2394                                         __dlm_do_purge_lockres(dlm, res);
2395                                         spin_unlock(&res->spinlock);
2396                                         wake_up(&res->wq);
2397                                         dlm_lockres_put(res);
2398                                         continue;
2399                                 } else if (res->owner == dlm->node_num)
2400                                         dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2401                                 spin_unlock(&res->spinlock);
2402                                 continue;
2403                         }
2404                         spin_lock(&res->spinlock);
2405                         /* zero the lvb if necessary */
2406                         dlm_revalidate_lvb(dlm, res, dead_node);
2407                         if (res->owner == dead_node) {
2408                                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
2409                                         mlog(0, "%s:%.*s: owned by "
2410                                                 "dead node %u, this node was "
2411                                                 "dropping its ref when master died. "
2412                                                 "continue, purging the lockres.\n",
2413                                                 dlm->name, res->lockname.len,
2414                                                 res->lockname.name, dead_node);
2415                                         dlm_lockres_get(res);
2416                                         __dlm_do_purge_lockres(dlm, res);
2417                                         spin_unlock(&res->spinlock);
2418                                         wake_up(&res->wq);
2419                                         dlm_lockres_put(res);
2420                                         continue;
2421                                 }
2422                         } else if (res->owner == dlm->node_num) {
2423                                 dlm_free_dead_locks(dlm, res, dead_node);
2424                                 __dlm_lockres_calc_usage(dlm, res);
2425                         } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
2426                                 if (test_bit(dead_node, res->refmap)) {
2427                                         mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2428                                                 "no locks and had not purged before dying\n",
2429                                                 dlm->name, res->lockname.len,
2430                                                 res->lockname.name, dead_node);
2431                                         dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2432                                 }
2433                         }
2434                         spin_unlock(&res->spinlock);
2435                 }
2436         }
2437
2438 }
2439
2440 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
2441 {
2442         assert_spin_locked(&dlm->spinlock);
2443
2444         if (dlm->reco.new_master == idx) {
2445                 mlog(0, "%s: recovery master %d just died\n",
2446                      dlm->name, idx);
2447                 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2448                         /* finalize1 was reached, so it is safe to clear
2449                          * the new_master and dead_node.  that recovery
2450                          * is complete. */
2451                         mlog(0, "%s: dead master %d had reached "
2452                              "finalize1 state, clearing\n", dlm->name, idx);
2453                         dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2454                         __dlm_reset_recovery(dlm);
2455                 }
2456         }
2457
2458         /* Clean up join state on node death. */
2459         if (dlm->joining_node == idx) {
2460                 mlog(0, "Clearing join state for node %u\n", idx);
2461                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
2462         }
2463
2464         /* check to see if the node is already considered dead */
2465         if (!test_bit(idx, dlm->live_nodes_map)) {
2466                 mlog(0, "for domain %s, node %d is already dead. "
2467                      "another node likely did recovery already.\n",
2468                      dlm->name, idx);
2469                 return;
2470         }
2471
2472         /* check to see if we do not care about this node */
2473         if (!test_bit(idx, dlm->domain_map)) {
2474                 /* This also catches the case that we get a node down
2475                  * but haven't joined the domain yet. */
2476                 mlog(0, "node %u already removed from domain!\n", idx);
2477                 return;
2478         }
2479
2480         clear_bit(idx, dlm->live_nodes_map);
2481
2482         /* make sure local cleanup occurs before the heartbeat events */
2483         if (!test_bit(idx, dlm->recovery_map))
2484                 dlm_do_local_recovery_cleanup(dlm, idx);
2485
2486         /* notify anything attached to the heartbeat events */
2487         dlm_hb_event_notify_attached(dlm, idx, 0);
2488
2489         mlog(0, "node %u being removed from domain map!\n", idx);
2490         clear_bit(idx, dlm->domain_map);
2491         clear_bit(idx, dlm->exit_domain_map);
2492         /* wake up migration waiters if a node goes down.
2493          * perhaps later we can genericize this for other waiters. */
2494         wake_up(&dlm->migration_wq);
2495
2496         set_bit(idx, dlm->recovery_map);
2497 }
2498
2499 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
2500 {
2501         struct dlm_ctxt *dlm = data;
2502
2503         if (!dlm_grab(dlm))
2504                 return;
2505
2506         /*
2507          * This will notify any dlm users that a node in our domain
2508          * went away without notifying us first.
2509          */
2510         if (test_bit(idx, dlm->domain_map))
2511                 dlm_fire_domain_eviction_callbacks(dlm, idx);
2512
2513         spin_lock(&dlm->spinlock);
2514         __dlm_hb_node_down(dlm, idx);
2515         spin_unlock(&dlm->spinlock);
2516
2517         dlm_put(dlm);
2518 }
2519
2520 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
2521 {
2522         struct dlm_ctxt *dlm = data;
2523
2524         if (!dlm_grab(dlm))
2525                 return;
2526
2527         spin_lock(&dlm->spinlock);
2528         set_bit(idx, dlm->live_nodes_map);
2529         /* do NOT notify mle attached to the heartbeat events.
2530          * new nodes are not interesting in mastery until joined. */
2531         spin_unlock(&dlm->spinlock);
2532
2533         dlm_put(dlm);
2534 }
2535
2536 static void dlm_reco_ast(void *astdata)
2537 {
2538         struct dlm_ctxt *dlm = astdata;
2539         mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2540              dlm->node_num, dlm->name);
2541 }
2542 static void dlm_reco_bast(void *astdata, int blocked_type)
2543 {
2544         struct dlm_ctxt *dlm = astdata;
2545         mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2546              dlm->node_num, dlm->name);
2547 }
2548 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
2549 {
2550         mlog(0, "unlockast for recovery lock fired!\n");
2551 }
2552
2553 /*
2554  * dlm_pick_recovery_master will continually attempt to use
2555  * dlmlock() on the special "$RECOVERY" lockres with the
2556  * LKM_NOQUEUE flag to get an EX.  every thread that enters
2557  * this function on each node racing to become the recovery
2558  * master will not stop attempting this until either:
2559  * a) this node gets the EX (and becomes the recovery master),
2560  * or b) dlm->reco.new_master gets set to some nodenum
2561  * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2562  * so each time a recovery master is needed, the entire cluster
2563  * will sync at this point.  if the new master dies, that will
2564  * be detected in dlm_do_recovery */
2565 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2566 {
2567         enum dlm_status ret;
2568         struct dlm_lockstatus lksb;
2569         int status = -EINVAL;
2570
2571         mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2572              dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2573 again:
2574         memset(&lksb, 0, sizeof(lksb));
2575
2576         ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
2577                       DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN,
2578                       dlm_reco_ast, dlm, dlm_reco_bast);
2579
2580         mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2581              dlm->name, ret, lksb.status);
2582
2583         if (ret == DLM_NORMAL) {
2584                 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2585                      dlm->name, dlm->node_num);
2586
2587                 /* got the EX lock.  check to see if another node
2588                  * just became the reco master */
2589                 if (dlm_reco_master_ready(dlm)) {
2590                         mlog(0, "%s: got reco EX lock, but %u will "
2591                              "do the recovery\n", dlm->name,
2592                              dlm->reco.new_master);
2593                         status = -EEXIST;
2594                 } else {
2595                         status = 0;
2596
2597                         /* see if recovery was already finished elsewhere */
2598                         spin_lock(&dlm->spinlock);
2599                         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
2600                                 status = -EINVAL;
2601                                 mlog(0, "%s: got reco EX lock, but "
2602                                      "node got recovered already\n", dlm->name);
2603                                 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2604                                         mlog(ML_ERROR, "%s: new master is %u "
2605                                              "but no dead node!\n",
2606                                              dlm->name, dlm->reco.new_master);
2607                                         BUG();
2608                                 }
2609                         }
2610                         spin_unlock(&dlm->spinlock);
2611                 }
2612
2613                 /* if this node has actually become the recovery master,
2614                  * set the master and send the messages to begin recovery */
2615                 if (!status) {
2616                         mlog(0, "%s: dead=%u, this=%u, sending "
2617                              "begin_reco now\n", dlm->name,
2618                              dlm->reco.dead_node, dlm->node_num);
2619                         status = dlm_send_begin_reco_message(dlm,
2620                                       dlm->reco.dead_node);
2621                         /* this always succeeds */
2622                         BUG_ON(status);
2623
2624                         /* set the new_master to this node */
2625                         spin_lock(&dlm->spinlock);
2626                         dlm_set_reco_master(dlm, dlm->node_num);
2627                         spin_unlock(&dlm->spinlock);
2628                 }
2629
2630                 /* recovery lock is a special case.  ast will not get fired,
2631                  * so just go ahead and unlock it. */
2632                 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
2633                 if (ret == DLM_DENIED) {
2634                         mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2635                         ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2636                 }
2637                 if (ret != DLM_NORMAL) {
2638                         /* this would really suck. this could only happen
2639                          * if there was a network error during the unlock
2640                          * because of node death.  this means the unlock
2641                          * is actually "done" and the lock structure is
2642                          * even freed.  we can continue, but only
2643                          * because this specific lock name is special. */
2644                         mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
2645                 }
2646         } else if (ret == DLM_NOTQUEUED) {
2647                 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2648                      dlm->name, dlm->node_num);
2649                 /* another node is master. wait on
2650                  * reco.new_master != O2NM_INVALID_NODE_NUM
2651                  * for at most one second */
2652                 wait_event_timeout(dlm->dlm_reco_thread_wq,
2653                                          dlm_reco_master_ready(dlm),
2654                                          msecs_to_jiffies(1000));
2655                 if (!dlm_reco_master_ready(dlm)) {
2656                         mlog(0, "%s: reco master taking awhile\n",
2657                              dlm->name);
2658                         goto again;
2659                 }
2660                 /* another node has informed this one that it is reco master */
2661                 mlog(0, "%s: reco master %u is ready to recover %u\n",
2662                      dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
2663                 status = -EEXIST;
2664         } else if (ret == DLM_RECOVERING) {
2665                 mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n",
2666                      dlm->name, dlm->node_num);
2667                 goto again;
2668         } else {
2669                 struct dlm_lock_resource *res;
2670
2671                 /* dlmlock returned something other than NOTQUEUED or NORMAL */
2672                 mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2673                      "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2674                      dlm_errname(lksb.status));
2675                 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2676                                          DLM_RECOVERY_LOCK_NAME_LEN);
2677                 if (res) {
2678                         dlm_print_one_lock_resource(res);
2679                         dlm_lockres_put(res);
2680                 } else {
2681                         mlog(ML_ERROR, "recovery lock not found\n");
2682                 }
2683                 BUG();
2684         }
2685
2686         return status;
2687 }
2688
2689 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2690 {
2691         struct dlm_begin_reco br;
2692         int ret = 0;
2693         struct dlm_node_iter iter;
2694         int nodenum;
2695         int status;
2696
2697         mlog(0, "%s: dead node is %u\n", dlm->name, dead_node);
2698
2699         spin_lock(&dlm->spinlock);
2700         dlm_node_iter_init(dlm->domain_map, &iter);
2701         spin_unlock(&dlm->spinlock);
2702
2703         clear_bit(dead_node, iter.node_map);
2704
2705         memset(&br, 0, sizeof(br));
2706         br.node_idx = dlm->node_num;
2707         br.dead_node = dead_node;
2708
2709         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2710                 ret = 0;
2711                 if (nodenum == dead_node) {
2712                         mlog(0, "not sending begin reco to dead node "
2713                                   "%u\n", dead_node);
2714                         continue;
2715                 }
2716                 if (nodenum == dlm->node_num) {
2717                         mlog(0, "not sending begin reco to self\n");
2718                         continue;
2719                 }
2720 retry:
2721                 ret = -EINVAL;
2722                 mlog(0, "attempting to send begin reco msg to %d\n",
2723                           nodenum);
2724                 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2725                                          &br, sizeof(br), nodenum, &status);
2726                 /* negative status is handled ok by caller here */
2727                 if (ret >= 0)
2728                         ret = status;
2729                 if (dlm_is_host_down(ret)) {
2730                         /* node is down.  not involved in recovery
2731                          * so just keep going */
2732                         mlog(ML_NOTICE, "%s: node %u was down when sending "
2733                              "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2734                         ret = 0;
2735                 }
2736
2737                 /*
2738                  * Prior to commit aad1b15310b9bcd59fa81ab8f2b1513b59553ea8,
2739                  * dlm_begin_reco_handler() returned EAGAIN and not -EAGAIN.
2740                  * We are handling both for compatibility reasons.
2741                  */
2742                 if (ret == -EAGAIN || ret == EAGAIN) {
2743                         mlog(0, "%s: trying to start recovery of node "
2744                              "%u, but node %u is waiting for last recovery "
2745                              "to complete, backoff for a bit\n", dlm->name,
2746                              dead_node, nodenum);
2747                         msleep(100);
2748                         goto retry;
2749                 }
2750                 if (ret < 0) {
2751                         struct dlm_lock_resource *res;
2752
2753                         /* this is now a serious problem, possibly ENOMEM
2754                          * in the network stack.  must retry */
2755                         mlog_errno(ret);
2756                         mlog(ML_ERROR, "begin reco of dlm %s to node %u "
2757                              "returned %d\n", dlm->name, nodenum, ret);
2758                         res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2759                                                  DLM_RECOVERY_LOCK_NAME_LEN);
2760                         if (res) {
2761                                 dlm_print_one_lock_resource(res);
2762                                 dlm_lockres_put(res);
2763                         } else {
2764                                 mlog(ML_ERROR, "recovery lock not found\n");
2765                         }
2766                         /* sleep for a bit in hopes that we can avoid
2767                          * another ENOMEM */
2768                         msleep(100);
2769                         goto retry;
2770                 }
2771         }
2772
2773         return ret;
2774 }
2775
2776 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2777                            void **ret_data)
2778 {
2779         struct dlm_ctxt *dlm = data;
2780         struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2781
2782         /* ok to return 0, domain has gone away */
2783         if (!dlm_grab(dlm))
2784                 return 0;
2785
2786         spin_lock(&dlm->spinlock);
2787         if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2788                 mlog(0, "%s: node %u wants to recover node %u (%u:%u) "
2789                      "but this node is in finalize state, waiting on finalize2\n",
2790                      dlm->name, br->node_idx, br->dead_node,
2791                      dlm->reco.dead_node, dlm->reco.new_master);
2792                 spin_unlock(&dlm->spinlock);
2793                 dlm_put(dlm);
2794                 return -EAGAIN;
2795         }
2796         spin_unlock(&dlm->spinlock);
2797
2798         mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n",
2799              dlm->name, br->node_idx, br->dead_node,
2800              dlm->reco.dead_node, dlm->reco.new_master);
2801
2802         dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2803
2804         spin_lock(&dlm->spinlock);
2805         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2806                 if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2807                         mlog(0, "%s: new_master %u died, changing "
2808                              "to %u\n", dlm->name, dlm->reco.new_master,
2809                              br->node_idx);
2810                 } else {
2811                         mlog(0, "%s: new_master %u NOT DEAD, changing "
2812                              "to %u\n", dlm->name, dlm->reco.new_master,
2813                              br->node_idx);
2814                         /* may not have seen the new master as dead yet */
2815                 }
2816         }
2817         if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2818                 mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2819                      "node %u changing it to %u\n", dlm->name,
2820                      dlm->reco.dead_node, br->node_idx, br->dead_node);
2821         }
2822         dlm_set_reco_master(dlm, br->node_idx);
2823         dlm_set_reco_dead_node(dlm, br->dead_node);
2824         if (!test_bit(br->dead_node, dlm->recovery_map)) {
2825                 mlog(0, "recovery master %u sees %u as dead, but this "
2826                      "node has not yet.  marking %u as dead\n",
2827                      br->node_idx, br->dead_node, br->dead_node);
2828                 if (!test_bit(br->dead_node, dlm->domain_map) ||
2829                     !test_bit(br->dead_node, dlm->live_nodes_map))
2830                         mlog(0, "%u not in domain/live_nodes map "
2831                              "so setting it in reco map manually\n",
2832                              br->dead_node);
2833                 /* force the recovery cleanup in __dlm_hb_node_down
2834                  * both of these will be cleared in a moment */
2835                 set_bit(br->dead_node, dlm->domain_map);
2836                 set_bit(br->dead_node, dlm->live_nodes_map);
2837                 __dlm_hb_node_down(dlm, br->dead_node);
2838         }
2839         spin_unlock(&dlm->spinlock);
2840
2841         dlm_kick_recovery_thread(dlm);
2842
2843         mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n",
2844              dlm->name, br->node_idx, br->dead_node,
2845              dlm->reco.dead_node, dlm->reco.new_master);
2846
2847         dlm_put(dlm);
2848         return 0;
2849 }
2850
2851 #define DLM_FINALIZE_STAGE2  0x01
2852 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2853 {
2854         int ret = 0;
2855         struct dlm_finalize_reco fr;
2856         struct dlm_node_iter iter;
2857         int nodenum;
2858         int status;
2859         int stage = 1;
2860
2861         mlog(0, "finishing recovery for node %s:%u, "
2862              "stage %d\n", dlm->name, dlm->reco.dead_node, stage);
2863
2864         spin_lock(&dlm->spinlock);
2865         dlm_node_iter_init(dlm->domain_map, &iter);
2866         spin_unlock(&dlm->spinlock);
2867
2868 stage2:
2869         memset(&fr, 0, sizeof(fr));
2870         fr.node_idx = dlm->node_num;
2871         fr.dead_node = dlm->reco.dead_node;
2872         if (stage == 2)
2873                 fr.flags |= DLM_FINALIZE_STAGE2;
2874
2875         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2876                 if (nodenum == dlm->node_num)
2877                         continue;
2878                 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2879                                          &fr, sizeof(fr), nodenum, &status);
2880                 if (ret >= 0)
2881                         ret = status;
2882                 if (ret < 0) {
2883                         mlog(ML_ERROR, "Error %d when sending message %u (key "
2884                              "0x%x) to node %u\n", ret, DLM_FINALIZE_RECO_MSG,
2885                              dlm->key, nodenum);
2886                         if (dlm_is_host_down(ret)) {
2887                                 /* this has no effect on this recovery
2888                                  * session, so set the status to zero to
2889                                  * finish out the last recovery */
2890                                 mlog(ML_ERROR, "node %u went down after this "
2891                                      "node finished recovery.\n", nodenum);
2892                                 ret = 0;
2893                                 continue;
2894                         }
2895                         break;
2896                 }
2897         }
2898         if (stage == 1) {
2899                 /* reset the node_iter back to the top and send finalize2 */
2900                 iter.curnode = -1;
2901                 stage = 2;
2902                 goto stage2;
2903         }
2904
2905         return ret;
2906 }
2907
2908 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2909                               void **ret_data)
2910 {
2911         struct dlm_ctxt *dlm = data;
2912         struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2913         int stage = 1;
2914
2915         /* ok to return 0, domain has gone away */
2916         if (!dlm_grab(dlm))
2917                 return 0;
2918
2919         if (fr->flags & DLM_FINALIZE_STAGE2)
2920                 stage = 2;
2921
2922         mlog(0, "%s: node %u finalizing recovery stage%d of "
2923              "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage,
2924              fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master);
2925
2926         spin_lock(&dlm->spinlock);
2927
2928         if (dlm->reco.new_master != fr->node_idx) {
2929                 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2930                      "%u is supposed to be the new master, dead=%u\n",
2931                      fr->node_idx, dlm->reco.new_master, fr->dead_node);
2932                 BUG();
2933         }
2934         if (dlm->reco.dead_node != fr->dead_node) {
2935                 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2936                      "node %u, but node %u is supposed to be dead\n",
2937                      fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2938                 BUG();
2939         }
2940
2941         switch (stage) {
2942                 case 1:
2943                         dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2944                         if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2945                                 mlog(ML_ERROR, "%s: received finalize1 from "
2946                                      "new master %u for dead node %u, but "
2947                                      "this node has already received it!\n",
2948                                      dlm->name, fr->node_idx, fr->dead_node);
2949                                 dlm_print_reco_node_status(dlm);
2950                                 BUG();
2951                         }
2952                         dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
2953                         spin_unlock(&dlm->spinlock);
2954                         break;
2955                 case 2:
2956                         if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) {
2957                                 mlog(ML_ERROR, "%s: received finalize2 from "
2958                                      "new master %u for dead node %u, but "
2959                                      "this node did not have finalize1!\n",
2960                                      dlm->name, fr->node_idx, fr->dead_node);
2961                                 dlm_print_reco_node_status(dlm);
2962                                 BUG();
2963                         }
2964                         dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2965                         __dlm_reset_recovery(dlm);
2966                         spin_unlock(&dlm->spinlock);
2967                         dlm_kick_recovery_thread(dlm);
2968                         break;
2969         }
2970
2971         mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n",
2972              dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master);
2973
2974         dlm_put(dlm);
2975         return 0;
2976 }