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/utsname.h>
  34 #include <linux/init.h>
  35 #include <linux/sysctl.h>
  36 #include <linux/random.h>
  37 #include <linux/blkdev.h>
  38 #include <linux/socket.h>
  39 #include <linux/inet.h>
  40 #include <linux/timer.h>
  41 #include <linux/kthread.h>
  42 #include <linux/delay.h>
  43
  44
  45 #include "cluster/heartbeat.h"
  46 #include "cluster/nodemanager.h"
  47 #include "cluster/tcp.h"
  48
  49 #include "dlmapi.h"
  50 #include "dlmcommon.h"
  51 #include "dlmdomain.h"
  52
  53 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY)
  54 #include "cluster/masklog.h"
  55
  56 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node);
  57
  58 static int dlm_recovery_thread(void *data);
  59 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
  60 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
  61 static void dlm_kick_recovery_thread(struct dlm_ctxt *dlm);
  62 static int dlm_do_recovery(struct dlm_ctxt *dlm);
  63
  64 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm);
  65 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node);
  66 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
  67 static int dlm_request_all_locks(struct dlm_ctxt *dlm,
  68                                  u8 request_from, u8 dead_node);
  69 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
  70
  71 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res);
  72 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
  73                                         const char *lockname, int namelen,
  74                                         int total_locks, u64 cookie,
  75                                         u8 flags, u8 master);
  76 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
  77                                     struct dlm_migratable_lockres *mres,
  78                                     u8 send_to,
  79                                     struct dlm_lock_resource *res,
  80                                     int total_locks);
  81 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
  82                                       struct dlm_lock_resource *res,
  83                                       u8 *real_master);
  84 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
  85                                      struct dlm_lock_resource *res,
  86                                      struct dlm_migratable_lockres *mres);
  87 static int dlm_do_master_requery(struct dlm_ctxt *dlm,
  88                                  struct dlm_lock_resource *res,
  89                                  u8 nodenum, u8 *real_master);
  90 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm);
  91 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm,
  92                                  u8 dead_node, u8 send_to);
  93 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node);
  94 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
  95                                         struct list_head *list, u8 dead_node);
  96 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
  97                                               u8 dead_node, u8 new_master);
  98 static void dlm_reco_ast(void *astdata);
  99 static void dlm_reco_bast(void *astdata, int blocked_type);
 100 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st);
 101 static void dlm_request_all_locks_worker(struct dlm_work_item *item,
 102                                          void *data);
 103 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data);
 104
 105 static u64 dlm_get_next_mig_cookie(void);
 106
 107 static spinlock_t dlm_reco_state_lock = SPIN_LOCK_UNLOCKED;
 108 static spinlock_t dlm_mig_cookie_lock = SPIN_LOCK_UNLOCKED;
 109 static u64 dlm_mig_cookie = 1;
 110
 111 static u64 dlm_get_next_mig_cookie(void)
 112 {
 113         u64 c;
 114         spin_lock(&dlm_mig_cookie_lock);
 115         c = dlm_mig_cookie;
 116         if (dlm_mig_cookie == (~0ULL))
 117                 dlm_mig_cookie = 1;
 118         else
 119                 dlm_mig_cookie++;
 120         spin_unlock(&dlm_mig_cookie_lock);
 121         return c;
 122 }
 123
 124 static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
 125 {
 126         spin_lock(&dlm->spinlock);
 127         clear_bit(dlm->reco.dead_node, dlm->recovery_map);
 128         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
 129         dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
 130         spin_unlock(&dlm->spinlock);
 131 }
 132
 133 /* Worker function used during recovery. */
 134 void dlm_dispatch_work(void *data)
 135 {
 136         struct dlm_ctxt *dlm = (struct dlm_ctxt *)data;
 137         LIST_HEAD(tmp_list);
 138         struct list_head *iter, *iter2;
 139         struct dlm_work_item *item;
 140         dlm_workfunc_t *workfunc;
 141
 142         spin_lock(&dlm->work_lock);
 143         list_splice_init(&dlm->work_list, &tmp_list);
 144         spin_unlock(&dlm->work_lock);
 145
 146         list_for_each_safe(iter, iter2, &tmp_list) {
 147                 item = list_entry(iter, struct dlm_work_item, list);
 148                 workfunc = item->func;
 149                 list_del_init(&item->list);
 150
 151                 /* already have ref on dlm to avoid having
 152                  * it disappear.  just double-check. */
 153                 BUG_ON(item->dlm != dlm);
 154
 155                 /* this is allowed to sleep and
 156                  * call network stuff */
 157                 workfunc(item, item->data);
 158
 159                 dlm_put(dlm);
 160                 kfree(item);
 161         }
 162 }
 163
 164 /*
 165  * RECOVERY THREAD
 166  */
 167
 168 static void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
 169 {
 170         /* wake the recovery thread
 171          * this will wake the reco thread in one of three places
 172          * 1) sleeping with no recovery happening
 173          * 2) sleeping with recovery mastered elsewhere
 174          * 3) recovery mastered here, waiting on reco data */
 175
 176         wake_up(&dlm->dlm_reco_thread_wq);
 177 }
 178
 179 /* Launch the recovery thread */
 180 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
 181 {
 182         mlog(0, "starting dlm recovery thread...\n");
 183
 184         dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
 185                                                 "dlm_reco_thread");
 186         if (IS_ERR(dlm->dlm_reco_thread_task)) {
 187                 mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
 188                 dlm->dlm_reco_thread_task = NULL;
 189                 return -EINVAL;
 190         }
 191
 192         return 0;
 193 }
 194
 195 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
 196 {
 197         if (dlm->dlm_reco_thread_task) {
 198                 mlog(0, "waiting for dlm recovery thread to exit\n");
 199                 kthread_stop(dlm->dlm_reco_thread_task);
 200                 dlm->dlm_reco_thread_task = NULL;
 201         }
 202 }
 203
 204
 205
 206 /*
 207  * this is lame, but here's how recovery works...
 208  * 1) all recovery threads cluster wide will work on recovering
 209  *    ONE node at a time
 210  * 2) negotiate who will take over all the locks for the dead node.
 211  *    thats right... ALL the locks.
 212  * 3) once a new master is chosen, everyone scans all locks
 213  *    and moves aside those mastered by the dead guy
 214  * 4) each of these locks should be locked until recovery is done
 215  * 5) the new master collects up all of secondary lock queue info
 216  *    one lock at a time, forcing each node to communicate back
 217  *    before continuing
 218  * 6) each secondary lock queue responds with the full known lock info
 219  * 7) once the new master has run all its locks, it sends a ALLDONE!
 220  *    message to everyone
 221  * 8) upon receiving this message, the secondary queue node unlocks
 222  *    and responds to the ALLDONE
 223  * 9) once the new master gets responses from everyone, he unlocks
 224  *    everything and recovery for this dead node is done
 225  *10) go back to 2) while there are still dead nodes
 226  *
 227  */
 228
 229
 230 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
 231
 232 static int dlm_recovery_thread(void *data)
 233 {
 234         int status;
 235         struct dlm_ctxt *dlm = data;
 236         unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
 237
 238         mlog(0, "dlm thread running for %s...\n", dlm->name);
 239
 240         while (!kthread_should_stop()) {
 241                 if (dlm_joined(dlm)) {
 242                         status = dlm_do_recovery(dlm);
 243                         if (status == -EAGAIN) {
 244                                 /* do not sleep, recheck immediately. */
 245                                 continue;
 246                         }
 247                         if (status < 0)
 248                                 mlog_errno(status);
 249                 }
 250
 251                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
 252                                                  kthread_should_stop(),
 253                                                  timeout);
 254         }
 255
 256         mlog(0, "quitting DLM recovery thread\n");
 257         return 0;
 258 }
 259
 260 /* returns true when the recovery master has contacted us */
 261 static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
 262 {
 263         int ready;
 264         spin_lock(&dlm->spinlock);
 265         ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
 266         spin_unlock(&dlm->spinlock);
 267         return ready;
 268 }
 269
 270 /* returns true if node is no longer in the domain
 271  * could be dead or just not joined */
 272 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
 273 {
 274         int dead;
 275         spin_lock(&dlm->spinlock);
 276         dead = test_bit(node, dlm->domain_map);
 277         spin_unlock(&dlm->spinlock);
 278         return dead;
 279 }
 280
 281 /* callers of the top-level api calls (dlmlock/dlmunlock) should
 282  * block on the dlm->reco.event when recovery is in progress.
 283  * the dlm recovery thread will set this state when it begins
 284  * recovering a dead node (as the new master or not) and clear
 285  * the state and wake as soon as all affected lock resources have
 286  * been marked with the RECOVERY flag */
 287 static int dlm_in_recovery(struct dlm_ctxt *dlm)
 288 {
 289         int in_recovery;
 290         spin_lock(&dlm->spinlock);
 291         in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
 292         spin_unlock(&dlm->spinlock);
 293         return in_recovery;
 294 }
 295
 296
 297 void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
 298 {
 299         wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
 300 }
 301
 302 static void dlm_begin_recovery(struct dlm_ctxt *dlm)
 303 {
 304         spin_lock(&dlm->spinlock);
 305         BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
 306         dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
 307         spin_unlock(&dlm->spinlock);
 308 }
 309
 310 static void dlm_end_recovery(struct dlm_ctxt *dlm)
 311 {
 312         spin_lock(&dlm->spinlock);
 313         BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
 314         dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
 315         spin_unlock(&dlm->spinlock);
 316         wake_up(&dlm->reco.event);
 317 }
 318
 319 static int dlm_do_recovery(struct dlm_ctxt *dlm)
 320 {
 321         int status = 0;
 322         int ret;
 323
 324         spin_lock(&dlm->spinlock);
 325
 326         /* check to see if the new master has died */
 327         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
 328             test_bit(dlm->reco.new_master, dlm->recovery_map)) {
 329                 mlog(0, "new master %u died while recovering %u!\n",
 330                      dlm->reco.new_master, dlm->reco.dead_node);
 331                 /* unset the new_master, leave dead_node */
 332                 dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
 333         }
 334
 335         /* select a target to recover */
 336         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
 337                 int bit;
 338
 339                 bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES+1, 0);
 340                 if (bit >= O2NM_MAX_NODES || bit < 0)
 341                         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
 342                 else
 343                         dlm->reco.dead_node = bit;
 344         } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
 345                 /* BUG? */
 346                 mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
 347                      dlm->reco.dead_node);
 348                 dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
 349         }
 350
 351         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
 352                 // mlog(0, "nothing to recover!  sleeping now!\n");
 353                 spin_unlock(&dlm->spinlock);
 354                 /* return to main thread loop and sleep. */
 355                 return 0;
 356         }
 357         mlog(0, "recovery thread found node %u in the recovery map!\n",
 358              dlm->reco.dead_node);
 359         spin_unlock(&dlm->spinlock);
 360
 361         /* take write barrier */
 362         /* (stops the list reshuffling thread, proxy ast handling) */
 363         dlm_begin_recovery(dlm);
 364
 365         if (dlm->reco.new_master == dlm->node_num)
 366                 goto master_here;
 367
 368         if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
 369                 /* choose a new master, returns 0 if this node
 370                  * is the master, -EEXIST if it's another node.
 371                  * this does not return until a new master is chosen
 372                  * or recovery completes entirely. */
 373                 ret = dlm_pick_recovery_master(dlm);
 374                 if (!ret) {
 375                         /* already notified everyone.  go. */
 376                         goto master_here;
 377                 }
 378                 mlog(0, "another node will master this recovery session.\n");
 379         }
 380         mlog(0, "dlm=%s, new_master=%u, this node=%u, dead_node=%u\n",
 381              dlm->name, dlm->reco.new_master,
 382              dlm->node_num, dlm->reco.dead_node);
 383
 384         /* it is safe to start everything back up here
 385          * because all of the dead node's lock resources
 386          * have been marked as in-recovery */
 387         dlm_end_recovery(dlm);
 388
 389         /* sleep out in main dlm_recovery_thread loop. */
 390         return 0;
 391
 392 master_here:
 393         mlog(0, "mastering recovery of %s:%u here(this=%u)!\n",
 394              dlm->name, dlm->reco.dead_node, dlm->node_num);
 395
 396         status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
 397         if (status < 0) {
 398                 mlog(ML_ERROR, "error %d remastering locks for node %u, "
 399                      "retrying.\n", status, dlm->reco.dead_node);
 400                 /* yield a bit to allow any final network messages
 401                  * to get handled on remaining nodes */
 402                 msleep(100);
 403         } else {
 404                 /* success!  see if any other nodes need recovery */
 405                 mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
 406                      dlm->name, dlm->reco.dead_node, dlm->node_num);
 407                 dlm_reset_recovery(dlm);
 408         }
 409         dlm_end_recovery(dlm);
 410
 411         /* continue and look for another dead node */
 412         return -EAGAIN;
 413 }
 414
 415 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
 416 {
 417         int status = 0;
 418         struct dlm_reco_node_data *ndata;
 419         struct list_head *iter;
 420         int all_nodes_done;
 421         int destroy = 0;
 422         int pass = 0;
 423
 424         status = dlm_init_recovery_area(dlm, dead_node);
 425         if (status < 0)
 426                 goto leave;
 427
 428         /* safe to access the node data list without a lock, since this
 429          * process is the only one to change the list */
 430         list_for_each(iter, &dlm->reco.node_data) {
 431                 ndata = list_entry (iter, struct dlm_reco_node_data, list);
 432                 BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
 433                 ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
 434
 435                 mlog(0, "requesting lock info from node %u\n",
 436                      ndata->node_num);
 437
 438                 if (ndata->node_num == dlm->node_num) {
 439                         ndata->state = DLM_RECO_NODE_DATA_DONE;
 440                         continue;
 441                 }
 442
 443                 status = dlm_request_all_locks(dlm, ndata->node_num, dead_node);
 444                 if (status < 0) {
 445                         mlog_errno(status);
 446                         if (dlm_is_host_down(status))
 447                                 ndata->state = DLM_RECO_NODE_DATA_DEAD;
 448                         else {
 449                                 destroy = 1;
 450                                 goto leave;
 451                         }
 452                 }
 453
 454                 switch (ndata->state) {
 455                         case DLM_RECO_NODE_DATA_INIT:
 456                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
 457                         case DLM_RECO_NODE_DATA_REQUESTED:
 458                                 BUG();
 459                                 break;
 460                         case DLM_RECO_NODE_DATA_DEAD:
 461                                 mlog(0, "node %u died after requesting "
 462                                      "recovery info for node %u\n",
 463                                      ndata->node_num, dead_node);
 464                                 // start all over
 465                                 destroy = 1;
 466                                 status = -EAGAIN;
 467                                 goto leave;
 468                         case DLM_RECO_NODE_DATA_REQUESTING:
 469                                 ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
 470                                 mlog(0, "now receiving recovery data from "
 471                                      "node %u for dead node %u\n",
 472                                      ndata->node_num, dead_node);
 473                                 break;
 474                         case DLM_RECO_NODE_DATA_RECEIVING:
 475                                 mlog(0, "already receiving recovery data from "
 476                                      "node %u for dead node %u\n",
 477                                      ndata->node_num, dead_node);
 478                                 break;
 479                         case DLM_RECO_NODE_DATA_DONE:
 480                                 mlog(0, "already DONE receiving recovery data "
 481                                      "from node %u for dead node %u\n",
 482                                      ndata->node_num, dead_node);
 483                                 break;
 484                 }
 485         }
 486
 487         mlog(0, "done requesting all lock info\n");
 488
 489         /* nodes should be sending reco data now
 490          * just need to wait */
 491
 492         while (1) {
 493                 /* check all the nodes now to see if we are
 494                  * done, or if anyone died */
 495                 all_nodes_done = 1;
 496                 spin_lock(&dlm_reco_state_lock);
 497                 list_for_each(iter, &dlm->reco.node_data) {
 498                         ndata = list_entry (iter, struct dlm_reco_node_data, list);
 499
 500                         mlog(0, "checking recovery state of node %u\n",
 501                              ndata->node_num);
 502                         switch (ndata->state) {
 503                                 case DLM_RECO_NODE_DATA_INIT:
 504                                 case DLM_RECO_NODE_DATA_REQUESTING:
 505                                         mlog(ML_ERROR, "bad ndata state for "
 506                                              "node %u: state=%d\n",
 507                                              ndata->node_num, ndata->state);
 508                                         BUG();
 509                                         break;
 510                                 case DLM_RECO_NODE_DATA_DEAD:
 511                                         mlog(ML_NOTICE, "node %u died after "
 512                                              "requesting recovery info for "
 513                                              "node %u\n", ndata->node_num,
 514                                              dead_node);
 515                                         spin_unlock(&dlm_reco_state_lock);
 516                                         // start all over
 517                                         destroy = 1;
 518                                         status = -EAGAIN;
 519                                         /* instead of spinning like crazy here,
 520                                          * wait for the domain map to catch up
 521                                          * with the network state.  otherwise this
 522                                          * can be hit hundreds of times before
 523                                          * the node is really seen as dead. */
 524                                         wait_event_timeout(dlm->dlm_reco_thread_wq,
 525                                                            dlm_is_node_dead(dlm,
 526                                                                 ndata->node_num),
 527                                                            msecs_to_jiffies(1000));
 528                                         mlog(0, "waited 1 sec for %u, "
 529                                              "dead? %s\n", ndata->node_num,
 530                                              dlm_is_node_dead(dlm, ndata->node_num) ?
 531                                              "yes" : "no");
 532                                         goto leave;
 533                                 case DLM_RECO_NODE_DATA_RECEIVING:
 534                                 case DLM_RECO_NODE_DATA_REQUESTED:
 535                                         all_nodes_done = 0;
 536                                         break;
 537                                 case DLM_RECO_NODE_DATA_DONE:
 538                                         break;
 539                                 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
 540                                         break;
 541                         }
 542                 }
 543                 spin_unlock(&dlm_reco_state_lock);
 544
 545                 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
 546                      all_nodes_done?"yes":"no");
 547                 if (all_nodes_done) {
 548                         int ret;
 549
 550                         /* all nodes are now in DLM_RECO_NODE_DATA_DONE state
 551                          * just send a finalize message to everyone and
 552                          * clean up */
 553                         mlog(0, "all nodes are done! send finalize\n");
 554                         ret = dlm_send_finalize_reco_message(dlm);
 555                         if (ret < 0)
 556                                 mlog_errno(ret);
 557
 558                         spin_lock(&dlm->spinlock);
 559                         dlm_finish_local_lockres_recovery(dlm, dead_node,
 560                                                           dlm->node_num);
 561                         spin_unlock(&dlm->spinlock);
 562                         mlog(0, "should be done with recovery!\n");
 563
 564                         mlog(0, "finishing recovery of %s at %lu, "
 565                              "dead=%u, this=%u, new=%u\n", dlm->name,
 566                              jiffies, dlm->reco.dead_node,
 567                              dlm->node_num, dlm->reco.new_master);
 568                         destroy = 1;
 569                         status = ret;
 570                         /* rescan everything marked dirty along the way */
 571                         dlm_kick_thread(dlm, NULL);
 572                         break;
 573                 }
 574                 /* wait to be signalled, with periodic timeout
 575                  * to check for node death */
 576                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
 577                                          kthread_should_stop(),
 578                                          msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
 579
 580         }
 581
 582 leave:
 583         if (destroy)
 584                 dlm_destroy_recovery_area(dlm, dead_node);
 585
 586         mlog_exit(status);
 587         return status;
 588 }
 589
 590 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
 591 {
 592         int num=0;
 593         struct dlm_reco_node_data *ndata;
 594
 595         spin_lock(&dlm->spinlock);
 596         memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
 597         /* nodes can only be removed (by dying) after dropping
 598          * this lock, and death will be trapped later, so this should do */
 599         spin_unlock(&dlm->spinlock);
 600
 601         while (1) {
 602                 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
 603                 if (num >= O2NM_MAX_NODES) {
 604                         break;
 605                 }
 606                 BUG_ON(num == dead_node);
 607
 608                 ndata = kcalloc(1, sizeof(*ndata), GFP_KERNEL);
 609                 if (!ndata) {
 610                         dlm_destroy_recovery_area(dlm, dead_node);
 611                         return -ENOMEM;
 612                 }
 613                 ndata->node_num = num;
 614                 ndata->state = DLM_RECO_NODE_DATA_INIT;
 615                 spin_lock(&dlm_reco_state_lock);
 616                 list_add_tail(&ndata->list, &dlm->reco.node_data);
 617                 spin_unlock(&dlm_reco_state_lock);
 618                 num++;
 619         }
 620
 621         return 0;
 622 }
 623
 624 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
 625 {
 626         struct list_head *iter, *iter2;
 627         struct dlm_reco_node_data *ndata;
 628         LIST_HEAD(tmplist);
 629
 630         spin_lock(&dlm_reco_state_lock);
 631         list_splice_init(&dlm->reco.node_data, &tmplist);
 632         spin_unlock(&dlm_reco_state_lock);
 633
 634         list_for_each_safe(iter, iter2, &tmplist) {
 635                 ndata = list_entry (iter, struct dlm_reco_node_data, list);
 636                 list_del_init(&ndata->list);
 637                 kfree(ndata);
 638         }
 639 }
 640
 641 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
 642                                  u8 dead_node)
 643 {
 644         struct dlm_lock_request lr;
 645         enum dlm_status ret;
 646
 647         mlog(0, "\n");
 648
 649
 650         mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
 651                   "to %u\n", dead_node, request_from);
 652
 653         memset(&lr, 0, sizeof(lr));
 654         lr.node_idx = dlm->node_num;
 655         lr.dead_node = dead_node;
 656
 657         // send message
 658         ret = DLM_NOLOCKMGR;
 659         ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
 660                                  &lr, sizeof(lr), request_from, NULL);
 661
 662         /* negative status is handled by caller */
 663         if (ret < 0)
 664                 mlog_errno(ret);
 665
 666         // return from here, then
 667         // sleep until all received or error
 668         return ret;
 669
 670 }
 671
 672 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data)
 673 {
 674         struct dlm_ctxt *dlm = data;
 675         struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
 676         char *buf = NULL;
 677         struct dlm_work_item *item = NULL;
 678
 679         if (!dlm_grab(dlm))
 680                 return -EINVAL;
 681
 682         BUG_ON(lr->dead_node != dlm->reco.dead_node);
 683
 684         item = kcalloc(1, sizeof(*item), GFP_KERNEL);
 685         if (!item) {
 686                 dlm_put(dlm);
 687                 return -ENOMEM;
 688         }
 689
 690         /* this will get freed by dlm_request_all_locks_worker */
 691         buf = (char *) __get_free_page(GFP_KERNEL);
 692         if (!buf) {
 693                 kfree(item);
 694                 dlm_put(dlm);
 695                 return -ENOMEM;
 696         }
 697
 698         /* queue up work for dlm_request_all_locks_worker */
 699         dlm_grab(dlm);  /* get an extra ref for the work item */
 700         dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
 701         item->u.ral.reco_master = lr->node_idx;
 702         item->u.ral.dead_node = lr->dead_node;
 703         spin_lock(&dlm->work_lock);
 704         list_add_tail(&item->list, &dlm->work_list);
 705         spin_unlock(&dlm->work_lock);
 706         schedule_work(&dlm->dispatched_work);
 707
 708         dlm_put(dlm);
 709         return 0;
 710 }
 711
 712 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
 713 {
 714         struct dlm_migratable_lockres *mres;
 715         struct dlm_lock_resource *res;
 716         struct dlm_ctxt *dlm;
 717         LIST_HEAD(resources);
 718         struct list_head *iter;
 719         int ret;
 720         u8 dead_node, reco_master;
 721
 722         dlm = item->dlm;
 723         dead_node = item->u.ral.dead_node;
 724         reco_master = item->u.ral.reco_master;
 725         mres = (struct dlm_migratable_lockres *)data;
 726
 727         if (dead_node != dlm->reco.dead_node ||
 728             reco_master != dlm->reco.new_master) {
 729                 /* show extra debug info if the recovery state is messed */
 730                 mlog(ML_ERROR, "%s: bad reco state: reco(dead=%u, master=%u), "
 731                      "request(dead=%u, master=%u)\n",
 732                      dlm->name, dlm->reco.dead_node, dlm->reco.new_master,
 733                      dead_node, reco_master);
 734                 mlog(ML_ERROR, "%s: name=%.*s master=%u locks=%u/%u flags=%u "
 735                      "entry[0]={c=%"MLFu64",l=%u,f=%u,t=%d,ct=%d,hb=%d,n=%u}\n",
 736                      dlm->name, mres->lockname_len, mres->lockname, mres->master,
 737                      mres->num_locks, mres->total_locks, mres->flags,
 738                      mres->ml[0].cookie, mres->ml[0].list, mres->ml[0].flags,
 739                      mres->ml[0].type, mres->ml[0].convert_type,
 740                      mres->ml[0].highest_blocked, mres->ml[0].node);
 741                 BUG();
 742         }
 743         BUG_ON(dead_node != dlm->reco.dead_node);
 744         BUG_ON(reco_master != dlm->reco.new_master);
 745
 746         /* lock resources should have already been moved to the
 747          * dlm->reco.resources list.  now move items from that list
 748          * to a temp list if the dead owner matches.  note that the
 749          * whole cluster recovers only one node at a time, so we
 750          * can safely move UNKNOWN lock resources for each recovery
 751          * session. */
 752         dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
 753
 754         /* now we can begin blasting lockreses without the dlm lock */
 755         list_for_each(iter, &resources) {
 756                 res = list_entry (iter, struct dlm_lock_resource, recovering);
 757                 ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
 758                                         DLM_MRES_RECOVERY);
 759                 if (ret < 0)
 760                         mlog_errno(ret);
 761         }
 762
 763         /* move the resources back to the list */
 764         spin_lock(&dlm->spinlock);
 765         list_splice_init(&resources, &dlm->reco.resources);
 766         spin_unlock(&dlm->spinlock);
 767
 768         ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
 769         if (ret < 0)
 770                 mlog_errno(ret);
 771
 772         free_page((unsigned long)data);
 773 }
 774
 775
 776 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
 777 {
 778         int ret, tmpret;
 779         struct dlm_reco_data_done done_msg;
 780
 781         memset(&done_msg, 0, sizeof(done_msg));
 782         done_msg.node_idx = dlm->node_num;
 783         done_msg.dead_node = dead_node;
 784         mlog(0, "sending DATA DONE message to %u, "
 785              "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
 786              done_msg.dead_node);
 787
 788         ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
 789                                  sizeof(done_msg), send_to, &tmpret);
 790         /* negative status is ignored by the caller */
 791         if (ret >= 0)
 792                 ret = tmpret;
 793         return ret;
 794 }
 795
 796
 797 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data)
 798 {
 799         struct dlm_ctxt *dlm = data;
 800         struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
 801         struct list_head *iter;
 802         struct dlm_reco_node_data *ndata = NULL;
 803         int ret = -EINVAL;
 804
 805         if (!dlm_grab(dlm))
 806                 return -EINVAL;
 807
 808         mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
 809              "node_idx=%u, this node=%u\n", done->dead_node,
 810              dlm->reco.dead_node, done->node_idx, dlm->node_num);
 811         BUG_ON(done->dead_node != dlm->reco.dead_node);
 812
 813         spin_lock(&dlm_reco_state_lock);
 814         list_for_each(iter, &dlm->reco.node_data) {
 815                 ndata = list_entry (iter, struct dlm_reco_node_data, list);
 816                 if (ndata->node_num != done->node_idx)
 817                         continue;
 818
 819                 switch (ndata->state) {
 820                         /* should have moved beyond INIT but not to FINALIZE yet */
 821                         case DLM_RECO_NODE_DATA_INIT:
 822                         case DLM_RECO_NODE_DATA_DEAD:
 823                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
 824                                 mlog(ML_ERROR, "bad ndata state for node %u:"
 825                                      " state=%d\n", ndata->node_num,
 826                                      ndata->state);
 827                                 BUG();
 828                                 break;
 829                         /* these states are possible at this point, anywhere along
 830                          * the line of recovery */
 831                         case DLM_RECO_NODE_DATA_DONE:
 832                         case DLM_RECO_NODE_DATA_RECEIVING:
 833                         case DLM_RECO_NODE_DATA_REQUESTED:
 834                         case DLM_RECO_NODE_DATA_REQUESTING:
 835                                 mlog(0, "node %u is DONE sending "
 836                                           "recovery data!\n",
 837                                           ndata->node_num);
 838
 839                                 ndata->state = DLM_RECO_NODE_DATA_DONE;
 840                                 ret = 0;
 841                                 break;
 842                 }
 843         }
 844         spin_unlock(&dlm_reco_state_lock);
 845
 846         /* wake the recovery thread, some node is done */
 847         if (!ret)
 848                 dlm_kick_recovery_thread(dlm);
 849
 850         if (ret < 0)
 851                 mlog(ML_ERROR, "failed to find recovery node data for node "
 852                      "%u\n", done->node_idx);
 853         dlm_put(dlm);
 854
 855         mlog(0, "leaving reco data done handler, ret=%d\n", ret);
 856         return ret;
 857 }
 858
 859 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
 860                                         struct list_head *list,
 861                                         u8 dead_node)
 862 {
 863         struct dlm_lock_resource *res;
 864         struct list_head *iter, *iter2;
 865         struct dlm_lock *lock;
 866
 867         spin_lock(&dlm->spinlock);
 868         list_for_each_safe(iter, iter2, &dlm->reco.resources) {
 869                 res = list_entry (iter, struct dlm_lock_resource, recovering);
 870                 /* always prune any $RECOVERY entries for dead nodes,
 871                  * otherwise hangs can occur during later recovery */
 872                 if (dlm_is_recovery_lock(res->lockname.name,
 873                                          res->lockname.len)) {
 874                         spin_lock(&res->spinlock);
 875                         list_for_each_entry(lock, &res->granted, list) {
 876                                 if (lock->ml.node == dead_node) {
 877                                         mlog(0, "AHA! there was "
 878                                              "a $RECOVERY lock for dead "
 879                                              "node %u (%s)!\n",
 880                                              dead_node, dlm->name);
 881                                         list_del_init(&lock->list);
 882                                         dlm_lock_put(lock);
 883                                         break;
 884                                 }
 885                         }
 886                         spin_unlock(&res->spinlock);
 887                         continue;
 888                 }
 889
 890                 if (res->owner == dead_node) {
 891                         mlog(0, "found lockres owned by dead node while "
 892                                   "doing recovery for node %u. sending it.\n",
 893                                   dead_node);
 894                         list_del_init(&res->recovering);
 895                         list_add_tail(&res->recovering, list);
 896                 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
 897                         mlog(0, "found UNKNOWN owner while doing recovery "
 898                                   "for node %u. sending it.\n", dead_node);
 899                         list_del_init(&res->recovering);
 900                         list_add_tail(&res->recovering, list);
 901                 }
 902         }
 903         spin_unlock(&dlm->spinlock);
 904 }
 905
 906 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
 907 {
 908         int total_locks = 0;
 909         struct list_head *iter, *queue = &res->granted;
 910         int i;
 911
 912         for (i=0; i<3; i++) {
 913                 list_for_each(iter, queue)
 914                         total_locks++;
 915                 queue++;
 916         }
 917         return total_locks;
 918 }
 919
 920
 921 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
 922                                       struct dlm_migratable_lockres *mres,
 923                                       u8 send_to,
 924                                       struct dlm_lock_resource *res,
 925                                       int total_locks)
 926 {
 927         u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
 928         int mres_total_locks = be32_to_cpu(mres->total_locks);
 929         int sz, ret = 0, status = 0;
 930         u8 orig_flags = mres->flags,
 931            orig_master = mres->master;
 932
 933         BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
 934         if (!mres->num_locks)
 935                 return 0;
 936
 937         sz = sizeof(struct dlm_migratable_lockres) +
 938                 (mres->num_locks * sizeof(struct dlm_migratable_lock));
 939
 940         /* add an all-done flag if we reached the last lock */
 941         orig_flags = mres->flags;
 942         BUG_ON(total_locks > mres_total_locks);
 943         if (total_locks == mres_total_locks)
 944                 mres->flags |= DLM_MRES_ALL_DONE;
 945
 946         /* send it */
 947         ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
 948                                  sz, send_to, &status);
 949         if (ret < 0) {
 950                 /* XXX: negative status is not handled.
 951                  * this will end up killing this node. */
 952                 mlog_errno(ret);
 953         } else {
 954                 /* might get an -ENOMEM back here */
 955                 ret = status;
 956                 if (ret < 0) {
 957                         mlog_errno(ret);
 958
 959                         if (ret == -EFAULT) {
 960                                 mlog(ML_ERROR, "node %u told me to kill "
 961                                      "myself!\n", send_to);
 962                                 BUG();
 963                         }
 964                 }
 965         }
 966
 967         /* zero and reinit the message buffer */
 968         dlm_init_migratable_lockres(mres, res->lockname.name,
 969                                     res->lockname.len, mres_total_locks,
 970                                     mig_cookie, orig_flags, orig_master);
 971         return ret;
 972 }
 973
 974 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
 975                                         const char *lockname, int namelen,
 976                                         int total_locks, u64 cookie,
 977                                         u8 flags, u8 master)
 978 {
 979         /* mres here is one full page */
 980         memset(mres, 0, PAGE_SIZE);
 981         mres->lockname_len = namelen;
 982         memcpy(mres->lockname, lockname, namelen);
 983         mres->num_locks = 0;
 984         mres->total_locks = cpu_to_be32(total_locks);
 985         mres->mig_cookie = cpu_to_be64(cookie);
 986         mres->flags = flags;
 987         mres->master = master;
 988 }
 989
 990
 991 /* returns 1 if this lock fills the network structure,
 992  * 0 otherwise */
 993 static int dlm_add_lock_to_array(struct dlm_lock *lock,
 994                                  struct dlm_migratable_lockres *mres, int queue)
 995 {
 996         struct dlm_migratable_lock *ml;
 997         int lock_num = mres->num_locks;
 998
 999         ml = &(mres->ml[lock_num]);
1000         ml->cookie = lock->ml.cookie;
1001         ml->type = lock->ml.type;
1002         ml->convert_type = lock->ml.convert_type;
1003         ml->highest_blocked = lock->ml.highest_blocked;
1004         ml->list = queue;
1005         if (lock->lksb) {
1006                 ml->flags = lock->lksb->flags;
1007                 /* send our current lvb */
1008                 if (ml->type == LKM_EXMODE ||
1009                     ml->type == LKM_PRMODE) {
1010                         /* if it is already set, this had better be a PR
1011                          * and it has to match */
1012                         if (mres->lvb[0] && (ml->type == LKM_EXMODE ||
1013                             memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))) {
1014                                 mlog(ML_ERROR, "mismatched lvbs!\n");
1015                                 __dlm_print_one_lock_resource(lock->lockres);
1016                                 BUG();
1017                         }
1018                         memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1019                 }
1020         }
1021         ml->node = lock->ml.node;
1022         mres->num_locks++;
1023         /* we reached the max, send this network message */
1024         if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1025                 return 1;
1026         return 0;
1027 }
1028
1029
1030 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1031                          struct dlm_migratable_lockres *mres,
1032                          u8 send_to, u8 flags)
1033 {
1034         struct list_head *queue, *iter;
1035         int total_locks, i;
1036         u64 mig_cookie = 0;
1037         struct dlm_lock *lock;
1038         int ret = 0;
1039
1040         BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1041
1042         mlog(0, "sending to %u\n", send_to);
1043
1044         total_locks = dlm_num_locks_in_lockres(res);
1045         if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1046                 /* rare, but possible */
1047                 mlog(0, "argh.  lockres has %d locks.  this will "
1048                           "require more than one network packet to "
1049                           "migrate\n", total_locks);
1050                 mig_cookie = dlm_get_next_mig_cookie();
1051         }
1052
1053         dlm_init_migratable_lockres(mres, res->lockname.name,
1054                                     res->lockname.len, total_locks,
1055                                     mig_cookie, flags, res->owner);
1056
1057         total_locks = 0;
1058         for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1059                 queue = dlm_list_idx_to_ptr(res, i);
1060                 list_for_each(iter, queue) {
1061                         lock = list_entry (iter, struct dlm_lock, list);
1062
1063                         /* add another lock. */
1064                         total_locks++;
1065                         if (!dlm_add_lock_to_array(lock, mres, i))
1066                                 continue;
1067
1068                         /* this filled the lock message,
1069                          * we must send it immediately. */
1070                         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1071                                                        res, total_locks);
1072                         if (ret < 0) {
1073                                 // TODO
1074                                 mlog(ML_ERROR, "dlm_send_mig_lockres_msg "
1075                                      "returned %d, TODO\n", ret);
1076                                 BUG();
1077                         }
1078                 }
1079         }
1080         /* flush any remaining locks */
1081         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1082         if (ret < 0) {
1083                 // TODO
1084                 mlog(ML_ERROR, "dlm_send_mig_lockres_msg returned %d, "
1085                      "TODO\n", ret);
1086                 BUG();
1087         }
1088         return ret;
1089 }
1090
1091
1092
1093 /*
1094  * this message will contain no more than one page worth of
1095  * recovery data, and it will work on only one lockres.
1096  * there may be many locks in this page, and we may need to wait
1097  * for additional packets to complete all the locks (rare, but
1098  * possible).
1099  */
1100 /*
1101  * NOTE: the allocation error cases here are scary
1102  * we really cannot afford to fail an alloc in recovery
1103  * do we spin?  returning an error only delays the problem really
1104  */
1105
1106 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data)
1107 {
1108         struct dlm_ctxt *dlm = data;
1109         struct dlm_migratable_lockres *mres =
1110                 (struct dlm_migratable_lockres *)msg->buf;
1111         int ret = 0;
1112         u8 real_master;
1113         char *buf = NULL;
1114         struct dlm_work_item *item = NULL;
1115         struct dlm_lock_resource *res = NULL;
1116
1117         if (!dlm_grab(dlm))
1118                 return -EINVAL;
1119
1120         BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1121
1122         real_master = mres->master;
1123         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1124                 /* cannot migrate a lockres with no master */
1125                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1126         }
1127
1128         mlog(0, "%s message received from node %u\n",
1129                   (mres->flags & DLM_MRES_RECOVERY) ?
1130                   "recovery" : "migration", mres->master);
1131         if (mres->flags & DLM_MRES_ALL_DONE)
1132                 mlog(0, "all done flag.  all lockres data received!\n");
1133
1134         ret = -ENOMEM;
1135         buf = kmalloc(be16_to_cpu(msg->data_len), GFP_KERNEL);
1136         item = kcalloc(1, sizeof(*item), GFP_KERNEL);
1137         if (!buf || !item)
1138                 goto leave;
1139
1140         /* lookup the lock to see if we have a secondary queue for this
1141          * already...  just add the locks in and this will have its owner
1142          * and RECOVERY flag changed when it completes. */
1143         res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len);
1144         if (res) {
1145                 /* this will get a ref on res */
1146                 /* mark it as recovering/migrating and hash it */
1147                 spin_lock(&res->spinlock);
1148                 if (mres->flags & DLM_MRES_RECOVERY) {
1149                         res->state |= DLM_LOCK_RES_RECOVERING;
1150                 } else {
1151                         if (res->state & DLM_LOCK_RES_MIGRATING) {
1152                                 /* this is at least the second
1153                                  * lockres message */
1154                                 mlog(0, "lock %.*s is already migrating\n",
1155                                           mres->lockname_len,
1156                                           mres->lockname);
1157                         } else if (res->state & DLM_LOCK_RES_RECOVERING) {
1158                                 /* caller should BUG */
1159                                 mlog(ML_ERROR, "node is attempting to migrate "
1160                                      "lock %.*s, but marked as recovering!\n",
1161                                      mres->lockname_len, mres->lockname);
1162                                 ret = -EFAULT;
1163                                 spin_unlock(&res->spinlock);
1164                                 goto leave;
1165                         }
1166                         res->state |= DLM_LOCK_RES_MIGRATING;
1167                 }
1168                 spin_unlock(&res->spinlock);
1169         } else {
1170                 /* need to allocate, just like if it was
1171                  * mastered here normally  */
1172                 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1173                 if (!res)
1174                         goto leave;
1175
1176                 /* to match the ref that we would have gotten if
1177                  * dlm_lookup_lockres had succeeded */
1178                 dlm_lockres_get(res);
1179
1180                 /* mark it as recovering/migrating and hash it */
1181                 if (mres->flags & DLM_MRES_RECOVERY)
1182                         res->state |= DLM_LOCK_RES_RECOVERING;
1183                 else
1184                         res->state |= DLM_LOCK_RES_MIGRATING;
1185
1186                 spin_lock(&dlm->spinlock);
1187                 __dlm_insert_lockres(dlm, res);
1188                 spin_unlock(&dlm->spinlock);
1189
1190                 /* now that the new lockres is inserted,
1191                  * make it usable by other processes */
1192                 spin_lock(&res->spinlock);
1193                 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1194                 spin_unlock(&res->spinlock);
1195
1196                 /* add an extra ref for just-allocated lockres
1197                  * otherwise the lockres will be purged immediately */
1198                 dlm_lockres_get(res);
1199
1200         }
1201
1202         /* at this point we have allocated everything we need,
1203          * and we have a hashed lockres with an extra ref and
1204          * the proper res->state flags. */
1205         ret = 0;
1206         if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1207                 /* migration cannot have an unknown master */
1208                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1209                 mlog(0, "recovery has passed me a lockres with an "
1210                           "unknown owner.. will need to requery: "
1211                           "%.*s\n", mres->lockname_len, mres->lockname);
1212         } else {
1213                 spin_lock(&res->spinlock);
1214                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
1215                 spin_unlock(&res->spinlock);
1216         }
1217
1218         /* queue up work for dlm_mig_lockres_worker */
1219         dlm_grab(dlm);  /* get an extra ref for the work item */
1220         memcpy(buf, msg->buf, be16_to_cpu(msg->data_len));  /* copy the whole message */
1221         dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1222         item->u.ml.lockres = res; /* already have a ref */
1223         item->u.ml.real_master = real_master;
1224         spin_lock(&dlm->work_lock);
1225         list_add_tail(&item->list, &dlm->work_list);
1226         spin_unlock(&dlm->work_lock);
1227         schedule_work(&dlm->dispatched_work);
1228
1229 leave:
1230         dlm_put(dlm);
1231         if (ret < 0) {
1232                 if (buf)
1233                         kfree(buf);
1234                 if (item)
1235                         kfree(item);
1236         }
1237
1238         mlog_exit(ret);
1239         return ret;
1240 }
1241
1242
1243 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1244 {
1245         struct dlm_ctxt *dlm = data;
1246         struct dlm_migratable_lockres *mres;
1247         int ret = 0;
1248         struct dlm_lock_resource *res;
1249         u8 real_master;
1250
1251         dlm = item->dlm;
1252         mres = (struct dlm_migratable_lockres *)data;
1253
1254         res = item->u.ml.lockres;
1255         real_master = item->u.ml.real_master;
1256
1257         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1258                 /* this case is super-rare. only occurs if
1259                  * node death happens during migration. */
1260 again:
1261                 ret = dlm_lockres_master_requery(dlm, res, &real_master);
1262                 if (ret < 0) {
1263                         mlog(0, "dlm_lockres_master_requery ret=%d\n",
1264                                   ret);
1265                         goto again;
1266                 }
1267                 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1268                         mlog(0, "lockres %.*s not claimed.  "
1269                                    "this node will take it.\n",
1270                                    res->lockname.len, res->lockname.name);
1271                 } else {
1272                         mlog(0, "master needs to respond to sender "
1273                                   "that node %u still owns %.*s\n",
1274                                   real_master, res->lockname.len,
1275                                   res->lockname.name);
1276                         /* cannot touch this lockres */
1277                         goto leave;
1278                 }
1279         }
1280
1281         ret = dlm_process_recovery_data(dlm, res, mres);
1282         if (ret < 0)
1283                 mlog(0, "dlm_process_recovery_data returned  %d\n", ret);
1284         else
1285                 mlog(0, "dlm_process_recovery_data succeeded\n");
1286
1287         if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1288                            (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1289                 ret = dlm_finish_migration(dlm, res, mres->master);
1290                 if (ret < 0)
1291                         mlog_errno(ret);
1292         }
1293
1294 leave:
1295         kfree(data);
1296         mlog_exit(ret);
1297 }
1298
1299
1300
1301 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1302                                       struct dlm_lock_resource *res,
1303                                       u8 *real_master)
1304 {
1305         struct dlm_node_iter iter;
1306         int nodenum;
1307         int ret = 0;
1308
1309         *real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1310
1311         /* we only reach here if one of the two nodes in a
1312          * migration died while the migration was in progress.
1313          * at this point we need to requery the master.  we
1314          * know that the new_master got as far as creating
1315          * an mle on at least one node, but we do not know
1316          * if any nodes had actually cleared the mle and set
1317          * the master to the new_master.  the old master
1318          * is supposed to set the owner to UNKNOWN in the
1319          * event of a new_master death, so the only possible
1320          * responses that we can get from nodes here are
1321          * that the master is new_master, or that the master
1322          * is UNKNOWN.
1323          * if all nodes come back with UNKNOWN then we know
1324          * the lock needs remastering here.
1325          * if any node comes back with a valid master, check
1326          * to see if that master is the one that we are
1327          * recovering.  if so, then the new_master died and
1328          * we need to remaster this lock.  if not, then the
1329          * new_master survived and that node will respond to
1330          * other nodes about the owner.
1331          * if there is an owner, this node needs to dump this
1332          * lockres and alert the sender that this lockres
1333          * was rejected. */
1334         spin_lock(&dlm->spinlock);
1335         dlm_node_iter_init(dlm->domain_map, &iter);
1336         spin_unlock(&dlm->spinlock);
1337
1338         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1339                 /* do not send to self */
1340                 if (nodenum == dlm->node_num)
1341                         continue;
1342                 ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1343                 if (ret < 0) {
1344                         mlog_errno(ret);
1345                         BUG();
1346                         /* TODO: need to figure a way to restart this */
1347                 }
1348                 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1349                         mlog(0, "lock master is %u\n", *real_master);
1350                         break;
1351                 }
1352         }
1353         return ret;
1354 }
1355
1356
1357 static int dlm_do_master_requery(struct dlm_ctxt *dlm,
1358                                  struct dlm_lock_resource *res,
1359                                  u8 nodenum, u8 *real_master)
1360 {
1361         int ret = -EINVAL;
1362         struct dlm_master_requery req;
1363         int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1364
1365         memset(&req, 0, sizeof(req));
1366         req.node_idx = dlm->node_num;
1367         req.namelen = res->lockname.len;
1368         memcpy(req.name, res->lockname.name, res->lockname.len);
1369
1370         ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1371                                  &req, sizeof(req), nodenum, &status);
1372         /* XXX: negative status not handled properly here. */
1373         if (ret < 0)
1374                 mlog_errno(ret);
1375         else {
1376                 BUG_ON(status < 0);
1377                 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1378                 *real_master = (u8) (status & 0xff);
1379                 mlog(0, "node %u responded to master requery with %u\n",
1380                           nodenum, *real_master);
1381                 ret = 0;
1382         }
1383         return ret;
1384 }
1385
1386
1387 /* this function cannot error, so unless the sending
1388  * or receiving of the message failed, the owner can
1389  * be trusted */
1390 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data)
1391 {
1392         struct dlm_ctxt *dlm = data;
1393         struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1394         struct dlm_lock_resource *res = NULL;
1395         int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1396         u32 flags = DLM_ASSERT_MASTER_REQUERY;
1397
1398         if (!dlm_grab(dlm)) {
1399                 /* since the domain has gone away on this
1400                  * node, the proper response is UNKNOWN */
1401                 return master;
1402         }
1403
1404         spin_lock(&dlm->spinlock);
1405         res = __dlm_lookup_lockres(dlm, req->name, req->namelen);
1406         if (res) {
1407                 spin_lock(&res->spinlock);
1408                 master = res->owner;
1409                 if (master == dlm->node_num) {
1410                         int ret = dlm_dispatch_assert_master(dlm, res,
1411                                                              0, 0, flags);
1412                         if (ret < 0) {
1413                                 mlog_errno(-ENOMEM);
1414                                 /* retry!? */
1415                                 BUG();
1416                         }
1417                 }
1418                 spin_unlock(&res->spinlock);
1419         }
1420         spin_unlock(&dlm->spinlock);
1421
1422         dlm_put(dlm);
1423         return master;
1424 }
1425
1426 static inline struct list_head *
1427 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1428 {
1429         struct list_head *ret;
1430         BUG_ON(list_num < 0);
1431         BUG_ON(list_num > 2);
1432         ret = &(res->granted);
1433         ret += list_num;
1434         return ret;
1435 }
1436 /* TODO: do ast flush business
1437  * TODO: do MIGRATING and RECOVERING spinning
1438  */
1439
1440 /*
1441 * NOTE about in-flight requests during migration:
1442 *
1443 * Before attempting the migrate, the master has marked the lockres as
1444 * MIGRATING and then flushed all of its pending ASTS.  So any in-flight
1445 * requests either got queued before the MIGRATING flag got set, in which
1446 * case the lock data will reflect the change and a return message is on
1447 * the way, or the request failed to get in before MIGRATING got set.  In
1448 * this case, the caller will be told to spin and wait for the MIGRATING
1449 * flag to be dropped, then recheck the master.
1450 * This holds true for the convert, cancel and unlock cases, and since lvb
1451 * updates are tied to these same messages, it applies to lvb updates as
1452 * well.  For the lock case, there is no way a lock can be on the master
1453 * queue and not be on the secondary queue since the lock is always added
1454 * locally first.  This means that the new target node will never be sent
1455 * a lock that he doesn't already have on the list.
1456 * In total, this means that the local lock is correct and should not be
1457 * updated to match the one sent by the master.  Any messages sent back
1458 * from the master before the MIGRATING flag will bring the lock properly
1459 * up-to-date, and the change will be ordered properly for the waiter.
1460 * We will *not* attempt to modify the lock underneath the waiter.
1461 */
1462
1463 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1464                                      struct dlm_lock_resource *res,
1465                                      struct dlm_migratable_lockres *mres)
1466 {
1467         struct dlm_migratable_lock *ml;
1468         struct list_head *queue;
1469         struct dlm_lock *newlock = NULL;
1470         struct dlm_lockstatus *lksb = NULL;
1471         int ret = 0;
1472         int i;
1473         struct list_head *iter;
1474         struct dlm_lock *lock = NULL;
1475
1476         mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1477         for (i=0; i<mres->num_locks; i++) {
1478                 ml = &(mres->ml[i]);
1479                 BUG_ON(ml->highest_blocked != LKM_IVMODE);
1480                 newlock = NULL;
1481                 lksb = NULL;
1482
1483                 queue = dlm_list_num_to_pointer(res, ml->list);
1484
1485                 /* if the lock is for the local node it needs to
1486                  * be moved to the proper location within the queue.
1487                  * do not allocate a new lock structure. */
1488                 if (ml->node == dlm->node_num) {
1489                         /* MIGRATION ONLY! */
1490                         BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1491
1492                         spin_lock(&res->spinlock);
1493                         list_for_each(iter, queue) {
1494                                 lock = list_entry (iter, struct dlm_lock, list);
1495                                 if (lock->ml.cookie != ml->cookie)
1496                                         lock = NULL;
1497                                 else
1498                                         break;
1499                         }
1500
1501                         /* lock is always created locally first, and
1502                          * destroyed locally last.  it must be on the list */
1503                         if (!lock) {
1504                                 mlog(ML_ERROR, "could not find local lock "
1505                                                "with cookie %"MLFu64"!\n",
1506                                      ml->cookie);
1507                                 BUG();
1508                         }
1509                         BUG_ON(lock->ml.node != ml->node);
1510
1511                         /* see NOTE above about why we do not update
1512                          * to match the master here */
1513
1514                         /* move the lock to its proper place */
1515                         /* do not alter lock refcount.  switching lists. */
1516                         list_del_init(&lock->list);
1517                         list_add_tail(&lock->list, queue);
1518                         spin_unlock(&res->spinlock);
1519
1520                         mlog(0, "just reordered a local lock!\n");
1521                         continue;
1522                 }
1523
1524                 /* lock is for another node. */
1525                 newlock = dlm_new_lock(ml->type, ml->node,
1526                                        be64_to_cpu(ml->cookie), NULL);
1527                 if (!newlock) {
1528                         ret = -ENOMEM;
1529                         goto leave;
1530                 }
1531                 lksb = newlock->lksb;
1532                 dlm_lock_attach_lockres(newlock, res);
1533
1534                 if (ml->convert_type != LKM_IVMODE) {
1535                         BUG_ON(queue != &res->converting);
1536                         newlock->ml.convert_type = ml->convert_type;
1537                 }
1538                 lksb->flags |= (ml->flags &
1539                                 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1540
1541                 if (mres->lvb[0]) {
1542                         if (lksb->flags & DLM_LKSB_PUT_LVB) {
1543                                 /* other node was trying to update
1544                                  * lvb when node died.  recreate the
1545                                  * lksb with the updated lvb. */
1546                                 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1547                         } else {
1548                                 /* otherwise, the node is sending its
1549                                  * most recent valid lvb info */
1550                                 BUG_ON(ml->type != LKM_EXMODE &&
1551                                        ml->type != LKM_PRMODE);
1552                                 if (res->lvb[0] && (ml->type == LKM_EXMODE ||
1553                                     memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1554                                         mlog(ML_ERROR, "received bad lvb!\n");
1555                                         __dlm_print_one_lock_resource(res);
1556                                         BUG();
1557                                 }
1558                                 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1559                         }
1560                 }
1561
1562
1563                 /* NOTE:
1564                  * wrt lock queue ordering and recovery:
1565                  *    1. order of locks on granted queue is
1566                  *       meaningless.
1567                  *    2. order of locks on converting queue is
1568                  *       LOST with the node death.  sorry charlie.
1569                  *    3. order of locks on the blocked queue is
1570                  *       also LOST.
1571                  * order of locks does not affect integrity, it
1572                  * just means that a lock request may get pushed
1573                  * back in line as a result of the node death.
1574                  * also note that for a given node the lock order
1575                  * for its secondary queue locks is preserved
1576                  * relative to each other, but clearly *not*
1577                  * preserved relative to locks from other nodes.
1578                  */
1579                 spin_lock(&res->spinlock);
1580                 dlm_lock_get(newlock);
1581                 list_add_tail(&newlock->list, queue);
1582                 spin_unlock(&res->spinlock);
1583         }
1584         mlog(0, "done running all the locks\n");
1585
1586 leave:
1587         if (ret < 0) {
1588                 mlog_errno(ret);
1589                 if (newlock)
1590                         dlm_lock_put(newlock);
1591         }
1592
1593         mlog_exit(ret);
1594         return ret;
1595 }
1596
1597 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
1598                                        struct dlm_lock_resource *res)
1599 {
1600         int i;
1601         struct list_head *queue, *iter, *iter2;
1602         struct dlm_lock *lock;
1603
1604         res->state |= DLM_LOCK_RES_RECOVERING;
1605         if (!list_empty(&res->recovering))
1606                 list_del_init(&res->recovering);
1607         list_add_tail(&res->recovering, &dlm->reco.resources);
1608
1609         /* find any pending locks and put them back on proper list */
1610         for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
1611                 queue = dlm_list_idx_to_ptr(res, i);
1612                 list_for_each_safe(iter, iter2, queue) {
1613                         lock = list_entry (iter, struct dlm_lock, list);
1614                         dlm_lock_get(lock);
1615                         if (lock->convert_pending) {
1616                                 /* move converting lock back to granted */
1617                                 BUG_ON(i != DLM_CONVERTING_LIST);
1618                                 mlog(0, "node died with convert pending "
1619                                      "on %.*s. move back to granted list.\n",
1620                                      res->lockname.len, res->lockname.name);
1621                                 dlm_revert_pending_convert(res, lock);
1622                                 lock->convert_pending = 0;
1623                         } else if (lock->lock_pending) {
1624                                 /* remove pending lock requests completely */
1625                                 BUG_ON(i != DLM_BLOCKED_LIST);
1626                                 mlog(0, "node died with lock pending "
1627                                      "on %.*s. remove from blocked list and skip.\n",
1628                                      res->lockname.len, res->lockname.name);
1629                                 /* lock will be floating until ref in
1630                                  * dlmlock_remote is freed after the network
1631                                  * call returns.  ok for it to not be on any
1632                                  * list since no ast can be called
1633                                  * (the master is dead). */
1634                                 dlm_revert_pending_lock(res, lock);
1635                                 lock->lock_pending = 0;
1636                         } else if (lock->unlock_pending) {
1637                                 /* if an unlock was in progress, treat as
1638                                  * if this had completed successfully
1639                                  * before sending this lock state to the
1640                                  * new master.  note that the dlm_unlock
1641                                  * call is still responsible for calling
1642                                  * the unlockast.  that will happen after
1643                                  * the network call times out.  for now,
1644                                  * just move lists to prepare the new
1645                                  * recovery master.  */
1646                                 BUG_ON(i != DLM_GRANTED_LIST);
1647                                 mlog(0, "node died with unlock pending "
1648                                      "on %.*s. remove from blocked list and skip.\n",
1649                                      res->lockname.len, res->lockname.name);
1650                                 dlm_commit_pending_unlock(res, lock);
1651                                 lock->unlock_pending = 0;
1652                         } else if (lock->cancel_pending) {
1653                                 /* if a cancel was in progress, treat as
1654                                  * if this had completed successfully
1655                                  * before sending this lock state to the
1656                                  * new master */
1657                                 BUG_ON(i != DLM_CONVERTING_LIST);
1658                                 mlog(0, "node died with cancel pending "
1659                                      "on %.*s. move back to granted list.\n",
1660                                      res->lockname.len, res->lockname.name);
1661                                 dlm_commit_pending_cancel(res, lock);
1662                                 lock->cancel_pending = 0;
1663                         }
1664                         dlm_lock_put(lock);
1665                 }
1666         }
1667 }
1668
1669
1670
1671 /* removes all recovered locks from the recovery list.
1672  * sets the res->owner to the new master.
1673  * unsets the RECOVERY flag and wakes waiters. */
1674 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
1675                                               u8 dead_node, u8 new_master)
1676 {
1677         int i;
1678         struct list_head *iter, *iter2, *bucket;
1679         struct dlm_lock_resource *res;
1680
1681         mlog_entry_void();
1682
1683         assert_spin_locked(&dlm->spinlock);
1684
1685         list_for_each_safe(iter, iter2, &dlm->reco.resources) {
1686                 res = list_entry (iter, struct dlm_lock_resource, recovering);
1687                 if (res->owner == dead_node) {
1688                         list_del_init(&res->recovering);
1689                         spin_lock(&res->spinlock);
1690                         dlm_change_lockres_owner(dlm, res, new_master);
1691                         res->state &= ~DLM_LOCK_RES_RECOVERING;
1692                         __dlm_dirty_lockres(dlm, res);
1693                         spin_unlock(&res->spinlock);
1694                         wake_up(&res->wq);
1695                 }
1696         }
1697
1698         /* this will become unnecessary eventually, but
1699          * for now we need to run the whole hash, clear
1700          * the RECOVERING state and set the owner
1701          * if necessary */
1702         for (i=0; i<DLM_HASH_SIZE; i++) {
1703                 bucket = &(dlm->resources[i]);
1704                 list_for_each(iter, bucket) {
1705                         res = list_entry (iter, struct dlm_lock_resource, list);
1706                         if (res->state & DLM_LOCK_RES_RECOVERING) {
1707                                 if (res->owner == dead_node) {
1708                                         mlog(0, "(this=%u) res %.*s owner=%u "
1709                                              "was not on recovering list, but "
1710                                              "clearing state anyway\n",
1711                                              dlm->node_num, res->lockname.len,
1712                                              res->lockname.name, new_master);
1713                                 } else if (res->owner == dlm->node_num) {
1714                                         mlog(0, "(this=%u) res %.*s owner=%u "
1715                                              "was not on recovering list, "
1716                                              "owner is THIS node, clearing\n",
1717                                              dlm->node_num, res->lockname.len,
1718                                              res->lockname.name, new_master);
1719                                 } else
1720                                         continue;
1721
1722                                 spin_lock(&res->spinlock);
1723                                 dlm_change_lockres_owner(dlm, res, new_master);
1724                                 res->state &= ~DLM_LOCK_RES_RECOVERING;
1725                                 __dlm_dirty_lockres(dlm, res);
1726                                 spin_unlock(&res->spinlock);
1727                                 wake_up(&res->wq);
1728                         }
1729                 }
1730         }
1731 }
1732
1733 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
1734 {
1735         if (local) {
1736                 if (lock->ml.type != LKM_EXMODE &&
1737                     lock->ml.type != LKM_PRMODE)
1738                         return 1;
1739         } else if (lock->ml.type == LKM_EXMODE)
1740                 return 1;
1741         return 0;
1742 }
1743
1744 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
1745                                struct dlm_lock_resource *res, u8 dead_node)
1746 {
1747         struct list_head *iter, *queue;
1748         struct dlm_lock *lock;
1749         int blank_lvb = 0, local = 0;
1750         int i;
1751         u8 search_node;
1752
1753         assert_spin_locked(&dlm->spinlock);
1754         assert_spin_locked(&res->spinlock);
1755
1756         if (res->owner == dlm->node_num)
1757                 /* if this node owned the lockres, and if the dead node
1758                  * had an EX when he died, blank out the lvb */
1759                 search_node = dead_node;
1760         else {
1761                 /* if this is a secondary lockres, and we had no EX or PR
1762                  * locks granted, we can no longer trust the lvb */
1763                 search_node = dlm->node_num;
1764                 local = 1;  /* check local state for valid lvb */
1765         }
1766
1767         for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
1768                 queue = dlm_list_idx_to_ptr(res, i);
1769                 list_for_each(iter, queue) {
1770                         lock = list_entry (iter, struct dlm_lock, list);
1771                         if (lock->ml.node == search_node) {
1772                                 if (dlm_lvb_needs_invalidation(lock, local)) {
1773                                         /* zero the lksb lvb and lockres lvb */
1774                                         blank_lvb = 1;
1775                                         memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
1776                                 }
1777                         }
1778                 }
1779         }
1780
1781         if (blank_lvb) {
1782                 mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
1783                      res->lockname.len, res->lockname.name, dead_node);
1784                 memset(res->lvb, 0, DLM_LVB_LEN);
1785         }
1786 }
1787
1788 static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
1789                                 struct dlm_lock_resource *res, u8 dead_node)
1790 {
1791         struct list_head *iter, *tmpiter;
1792         struct dlm_lock *lock;
1793
1794         /* this node is the lockres master:
1795          * 1) remove any stale locks for the dead node
1796          * 2) if the dead node had an EX when he died, blank out the lvb
1797          */
1798         assert_spin_locked(&dlm->spinlock);
1799         assert_spin_locked(&res->spinlock);
1800
1801         /* TODO: check pending_asts, pending_basts here */
1802         list_for_each_safe(iter, tmpiter, &res->granted) {
1803                 lock = list_entry (iter, struct dlm_lock, list);
1804                 if (lock->ml.node == dead_node) {
1805                         list_del_init(&lock->list);
1806                         dlm_lock_put(lock);
1807                 }
1808         }
1809         list_for_each_safe(iter, tmpiter, &res->converting) {
1810                 lock = list_entry (iter, struct dlm_lock, list);
1811                 if (lock->ml.node == dead_node) {
1812                         list_del_init(&lock->list);
1813                         dlm_lock_put(lock);
1814                 }
1815         }
1816         list_for_each_safe(iter, tmpiter, &res->blocked) {
1817                 lock = list_entry (iter, struct dlm_lock, list);
1818                 if (lock->ml.node == dead_node) {
1819                         list_del_init(&lock->list);
1820                         dlm_lock_put(lock);
1821                 }
1822         }
1823
1824         /* do not kick thread yet */
1825         __dlm_dirty_lockres(dlm, res);
1826 }
1827
1828 /* if this node is the recovery master, and there are no
1829  * locks for a given lockres owned by this node that are in
1830  * either PR or EX mode, zero out the lvb before requesting.
1831  *
1832  */
1833
1834
1835 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
1836 {
1837         struct list_head *iter;
1838         struct dlm_lock_resource *res;
1839         int i;
1840         struct list_head *bucket;
1841         struct dlm_lock *lock;
1842
1843
1844         /* purge any stale mles */
1845         dlm_clean_master_list(dlm, dead_node);
1846
1847         /*
1848          * now clean up all lock resources.  there are two rules:
1849          *
1850          * 1) if the dead node was the master, move the lockres
1851          *    to the recovering list.  set the RECOVERING flag.
1852          *    this lockres needs to be cleaned up before it can
1853          *    be used further.
1854          *
1855          * 2) if this node was the master, remove all locks from
1856          *    each of the lockres queues that were owned by the
1857          *    dead node.  once recovery finishes, the dlm thread
1858          *    can be kicked again to see if any ASTs or BASTs
1859          *    need to be fired as a result.
1860          */
1861         for (i=0; i<DLM_HASH_SIZE; i++) {
1862                 bucket = &(dlm->resources[i]);
1863                 list_for_each(iter, bucket) {
1864                         res = list_entry (iter, struct dlm_lock_resource, list);
1865                         /* always prune any $RECOVERY entries for dead nodes,
1866                          * otherwise hangs can occur during later recovery */
1867                         if (dlm_is_recovery_lock(res->lockname.name,
1868                                                  res->lockname.len)) {
1869                                 spin_lock(&res->spinlock);
1870                                 list_for_each_entry(lock, &res->granted, list) {
1871                                         if (lock->ml.node == dead_node) {
1872                                                 mlog(0, "AHA! there was "
1873                                                      "a $RECOVERY lock for dead "
1874                                                      "node %u (%s)!\n",
1875                                                      dead_node, dlm->name);
1876                                                 list_del_init(&lock->list);
1877                                                 dlm_lock_put(lock);
1878                                                 break;
1879                                         }
1880                                 }
1881                                 spin_unlock(&res->spinlock);
1882                                 continue;
1883                         }
1884                         spin_lock(&res->spinlock);
1885                         /* zero the lvb if necessary */
1886                         dlm_revalidate_lvb(dlm, res, dead_node);
1887                         if (res->owner == dead_node)
1888                                 dlm_move_lockres_to_recovery_list(dlm, res);
1889                         else if (res->owner == dlm->node_num) {
1890                                 dlm_free_dead_locks(dlm, res, dead_node);
1891                                 __dlm_lockres_calc_usage(dlm, res);
1892                         }
1893                         spin_unlock(&res->spinlock);
1894                 }
1895         }
1896
1897 }
1898
1899 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
1900 {
1901         assert_spin_locked(&dlm->spinlock);
1902
1903         /* check to see if the node is already considered dead */
1904         if (!test_bit(idx, dlm->live_nodes_map)) {
1905                 mlog(0, "for domain %s, node %d is already dead. "
1906                      "another node likely did recovery already.\n",
1907                      dlm->name, idx);
1908                 return;
1909         }
1910
1911         /* check to see if we do not care about this node */
1912         if (!test_bit(idx, dlm->domain_map)) {
1913                 /* This also catches the case that we get a node down
1914                  * but haven't joined the domain yet. */
1915                 mlog(0, "node %u already removed from domain!\n", idx);
1916                 return;
1917         }
1918
1919         clear_bit(idx, dlm->live_nodes_map);
1920
1921         /* Clean up join state on node death. */
1922         if (dlm->joining_node == idx) {
1923                 mlog(0, "Clearing join state for node %u\n", idx);
1924                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1925         }
1926
1927         /* make sure local cleanup occurs before the heartbeat events */
1928         if (!test_bit(idx, dlm->recovery_map))
1929                 dlm_do_local_recovery_cleanup(dlm, idx);
1930
1931         /* notify anything attached to the heartbeat events */
1932         dlm_hb_event_notify_attached(dlm, idx, 0);
1933
1934         mlog(0, "node %u being removed from domain map!\n", idx);
1935         clear_bit(idx, dlm->domain_map);
1936         /* wake up migration waiters if a node goes down.
1937          * perhaps later we can genericize this for other waiters. */
1938         wake_up(&dlm->migration_wq);
1939
1940         if (test_bit(idx, dlm->recovery_map))
1941                 mlog(0, "domain %s, node %u already added "
1942                      "to recovery map!\n", dlm->name, idx);
1943         else
1944                 set_bit(idx, dlm->recovery_map);
1945 }
1946
1947 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
1948 {
1949         struct dlm_ctxt *dlm = data;
1950
1951         if (!dlm_grab(dlm))
1952                 return;
1953
1954         spin_lock(&dlm->spinlock);
1955         __dlm_hb_node_down(dlm, idx);
1956         spin_unlock(&dlm->spinlock);
1957
1958         dlm_put(dlm);
1959 }
1960
1961 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
1962 {
1963         struct dlm_ctxt *dlm = data;
1964
1965         if (!dlm_grab(dlm))
1966                 return;
1967
1968         spin_lock(&dlm->spinlock);
1969         set_bit(idx, dlm->live_nodes_map);
1970         /* do NOT notify mle attached to the heartbeat events.
1971          * new nodes are not interesting in mastery until joined. */
1972         spin_unlock(&dlm->spinlock);
1973
1974         dlm_put(dlm);
1975 }
1976
1977 static void dlm_reco_ast(void *astdata)
1978 {
1979         struct dlm_ctxt *dlm = astdata;
1980         mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
1981              dlm->node_num, dlm->name);
1982 }
1983 static void dlm_reco_bast(void *astdata, int blocked_type)
1984 {
1985         struct dlm_ctxt *dlm = astdata;
1986         mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
1987              dlm->node_num, dlm->name);
1988 }
1989 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
1990 {
1991         mlog(0, "unlockast for recovery lock fired!\n");
1992 }
1993
1994 /*
1995  * dlm_pick_recovery_master will continually attempt to use
1996  * dlmlock() on the special "$RECOVERY" lockres with the
1997  * LKM_NOQUEUE flag to get an EX.  every thread that enters
1998  * this function on each node racing to become the recovery
1999  * master will not stop attempting this until either:
2000  * a) this node gets the EX (and becomes the recovery master),
2001  * or b) dlm->reco.new_master gets set to some nodenum
2002  * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2003  * so each time a recovery master is needed, the entire cluster
2004  * will sync at this point.  if the new master dies, that will
2005  * be detected in dlm_do_recovery */
2006 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2007 {
2008         enum dlm_status ret;
2009         struct dlm_lockstatus lksb;
2010         int status = -EINVAL;
2011
2012         mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2013              dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2014 again:
2015         memset(&lksb, 0, sizeof(lksb));
2016
2017         ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
2018                       DLM_RECOVERY_LOCK_NAME, dlm_reco_ast, dlm, dlm_reco_bast);
2019
2020         mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2021              dlm->name, ret, lksb.status);
2022
2023         if (ret == DLM_NORMAL) {
2024                 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2025                      dlm->name, dlm->node_num);
2026
2027                 /* got the EX lock.  check to see if another node
2028                  * just became the reco master */
2029                 if (dlm_reco_master_ready(dlm)) {
2030                         mlog(0, "%s: got reco EX lock, but %u will "
2031                              "do the recovery\n", dlm->name,
2032                              dlm->reco.new_master);
2033                         status = -EEXIST;
2034                 } else {
2035                         status = dlm_send_begin_reco_message(dlm,
2036                                       dlm->reco.dead_node);
2037                         /* this always succeeds */
2038                         BUG_ON(status);
2039
2040                         /* set the new_master to this node */
2041                         spin_lock(&dlm->spinlock);
2042                         dlm->reco.new_master = dlm->node_num;
2043                         spin_unlock(&dlm->spinlock);
2044                 }
2045
2046                 /* recovery lock is a special case.  ast will not get fired,
2047                  * so just go ahead and unlock it. */
2048                 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
2049                 if (ret == DLM_DENIED) {
2050                         mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2051                         ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2052                 }
2053                 if (ret != DLM_NORMAL) {
2054                         /* this would really suck. this could only happen
2055                          * if there was a network error during the unlock
2056                          * because of node death.  this means the unlock
2057                          * is actually "done" and the lock structure is
2058                          * even freed.  we can continue, but only
2059                          * because this specific lock name is special. */
2060                         mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
2061                 }
2062         } else if (ret == DLM_NOTQUEUED) {
2063                 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2064                      dlm->name, dlm->node_num);
2065                 /* another node is master. wait on
2066                  * reco.new_master != O2NM_INVALID_NODE_NUM
2067                  * for at most one second */
2068                 wait_event_timeout(dlm->dlm_reco_thread_wq,
2069                                          dlm_reco_master_ready(dlm),
2070                                          msecs_to_jiffies(1000));
2071                 if (!dlm_reco_master_ready(dlm)) {
2072                         mlog(0, "%s: reco master taking awhile\n",
2073                              dlm->name);
2074                         goto again;
2075                 }
2076                 /* another node has informed this one that it is reco master */
2077                 mlog(0, "%s: reco master %u is ready to recover %u\n",
2078                      dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
2079                 status = -EEXIST;
2080         } else {
2081                 struct dlm_lock_resource *res;
2082
2083                 /* dlmlock returned something other than NOTQUEUED or NORMAL */
2084                 mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2085                      "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2086                      dlm_errname(lksb.status));
2087                 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2088                                          DLM_RECOVERY_LOCK_NAME_LEN);
2089                 if (res) {
2090                         dlm_print_one_lock_resource(res);
2091                         dlm_lockres_put(res);
2092                 } else {
2093                         mlog(ML_ERROR, "recovery lock not found\n");
2094                 }
2095                 BUG();
2096         }
2097
2098         return status;
2099 }
2100
2101 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2102 {
2103         struct dlm_begin_reco br;
2104         int ret = 0;
2105         struct dlm_node_iter iter;
2106         int nodenum;
2107         int status;
2108
2109         mlog_entry("%u\n", dead_node);
2110
2111         mlog(0, "dead node is %u\n", dead_node);
2112
2113         spin_lock(&dlm->spinlock);
2114         dlm_node_iter_init(dlm->domain_map, &iter);
2115         spin_unlock(&dlm->spinlock);
2116
2117         clear_bit(dead_node, iter.node_map);
2118
2119         memset(&br, 0, sizeof(br));
2120         br.node_idx = dlm->node_num;
2121         br.dead_node = dead_node;
2122
2123         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2124                 ret = 0;
2125                 if (nodenum == dead_node) {
2126                         mlog(0, "not sending begin reco to dead node "
2127                                   "%u\n", dead_node);
2128                         continue;
2129                 }
2130                 if (nodenum == dlm->node_num) {
2131                         mlog(0, "not sending begin reco to self\n");
2132                         continue;
2133                 }
2134 retry:
2135                 ret = -EINVAL;
2136                 mlog(0, "attempting to send begin reco msg to %d\n",
2137                           nodenum);
2138                 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2139                                          &br, sizeof(br), nodenum, &status);
2140                 /* negative status is handled ok by caller here */
2141                 if (ret >= 0)
2142                         ret = status;
2143                 if (dlm_is_host_down(ret)) {
2144                         /* node is down.  not involved in recovery
2145                          * so just keep going */
2146                         mlog(0, "%s: node %u was down when sending "
2147                              "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2148                         ret = 0;
2149                 }
2150                 if (ret < 0) {
2151                         struct dlm_lock_resource *res;
2152                         /* this is now a serious problem, possibly ENOMEM
2153                          * in the network stack.  must retry */
2154                         mlog_errno(ret);
2155                         mlog(ML_ERROR, "begin reco of dlm %s to node %u "
2156                             " returned %d\n", dlm->name, nodenum, ret);
2157                         res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2158                                                  DLM_RECOVERY_LOCK_NAME_LEN);
2159                         if (res) {
2160                                 dlm_print_one_lock_resource(res);
2161                                 dlm_lockres_put(res);
2162                         } else {
2163                                 mlog(ML_ERROR, "recovery lock not found\n");
2164                         }
2165                         /* sleep for a bit in hopes that we can avoid
2166                          * another ENOMEM */
2167                         msleep(100);
2168                         goto retry;
2169                 }
2170         }
2171
2172         return ret;
2173 }
2174
2175 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data)
2176 {
2177         struct dlm_ctxt *dlm = data;
2178         struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2179
2180         /* ok to return 0, domain has gone away */
2181         if (!dlm_grab(dlm))
2182                 return 0;
2183
2184         mlog(0, "node %u wants to recover node %u\n",
2185                   br->node_idx, br->dead_node);
2186
2187         dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2188
2189         spin_lock(&dlm->spinlock);
2190         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2191                 if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2192                         mlog(0, "%s: new_master %u died, changing "
2193                              "to %u\n", dlm->name, dlm->reco.new_master,
2194                              br->node_idx);
2195                 } else {
2196                         mlog(0, "%s: new_master %u NOT DEAD, changing "
2197                              "to %u\n", dlm->name, dlm->reco.new_master,
2198                              br->node_idx);
2199                         /* may not have seen the new master as dead yet */
2200                 }
2201         }
2202         if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2203                 mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2204                      "node %u changing it to %u\n", dlm->name,
2205                      dlm->reco.dead_node, br->node_idx, br->dead_node);
2206         }
2207         dlm->reco.new_master = br->node_idx;
2208         dlm->reco.dead_node = br->dead_node;
2209         if (!test_bit(br->dead_node, dlm->recovery_map)) {
2210                 mlog(0, "recovery master %u sees %u as dead, but this "
2211                      "node has not yet.  marking %u as dead\n",
2212                      br->node_idx, br->dead_node, br->dead_node);
2213                 if (!test_bit(br->dead_node, dlm->domain_map) ||
2214                     !test_bit(br->dead_node, dlm->live_nodes_map))
2215                         mlog(0, "%u not in domain/live_nodes map "
2216                              "so setting it in reco map manually\n",
2217                              br->dead_node);
2218                 set_bit(br->dead_node, dlm->recovery_map);
2219                 __dlm_hb_node_down(dlm, br->dead_node);
2220         }
2221         spin_unlock(&dlm->spinlock);
2222
2223         dlm_kick_recovery_thread(dlm);
2224         dlm_put(dlm);
2225         return 0;
2226 }
2227
2228 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2229 {
2230         int ret = 0;
2231         struct dlm_finalize_reco fr;
2232         struct dlm_node_iter iter;
2233         int nodenum;
2234         int status;
2235
2236         mlog(0, "finishing recovery for node %s:%u\n",
2237              dlm->name, dlm->reco.dead_node);
2238
2239         spin_lock(&dlm->spinlock);
2240         dlm_node_iter_init(dlm->domain_map, &iter);
2241         spin_unlock(&dlm->spinlock);
2242
2243         memset(&fr, 0, sizeof(fr));
2244         fr.node_idx = dlm->node_num;
2245         fr.dead_node = dlm->reco.dead_node;
2246
2247         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2248                 if (nodenum == dlm->node_num)
2249                         continue;
2250                 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2251                                          &fr, sizeof(fr), nodenum, &status);
2252                 if (ret >= 0) {
2253                         ret = status;
2254                         if (dlm_is_host_down(ret)) {
2255                                 /* this has no effect on this recovery
2256                                  * session, so set the status to zero to
2257                                  * finish out the last recovery */
2258                                 mlog(ML_ERROR, "node %u went down after this "
2259                                      "node finished recovery.\n", nodenum);
2260                                 ret = 0;
2261                         }
2262                 }
2263                 if (ret < 0) {
2264                         mlog_errno(ret);
2265                         break;
2266                 }
2267         }
2268
2269         return ret;
2270 }
2271
2272 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data)
2273 {
2274         struct dlm_ctxt *dlm = data;
2275         struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2276
2277         /* ok to return 0, domain has gone away */
2278         if (!dlm_grab(dlm))
2279                 return 0;
2280
2281         mlog(0, "node %u finalizing recovery of node %u\n",
2282              fr->node_idx, fr->dead_node);
2283
2284         spin_lock(&dlm->spinlock);
2285
2286         if (dlm->reco.new_master != fr->node_idx) {
2287                 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2288                      "%u is supposed to be the new master, dead=%u\n",
2289                      fr->node_idx, dlm->reco.new_master, fr->dead_node);
2290                 BUG();
2291         }
2292         if (dlm->reco.dead_node != fr->dead_node) {
2293                 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2294                      "node %u, but node %u is supposed to be dead\n",
2295                      fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2296                 BUG();
2297         }
2298
2299         dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2300
2301         spin_unlock(&dlm->spinlock);
2302
2303         dlm_reset_recovery(dlm);
2304
2305         dlm_kick_recovery_thread(dlm);
2306         dlm_put(dlm);
2307         return 0;
2308 }