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