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