fs/ocfs2/dlm/dlmmaster.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmmod.c
   5  *
   6  * standalone DLM module
   7  *
   8  * Copyright (C) 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  *
  25  */
  26
  27
  28 #include <linux/module.h>
  29 #include <linux/fs.h>
  30 #include <linux/types.h>
  31 #include <linux/slab.h>
  32 #include <linux/highmem.h>
  33 #include <linux/init.h>
  34 #include <linux/sysctl.h>
  35 #include <linux/random.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/socket.h>
  38 #include <linux/inet.h>
  39 #include <linux/spinlock.h>
  40 #include <linux/delay.h>
  41
  42
  43 #include "cluster/heartbeat.h"
  44 #include "cluster/nodemanager.h"
  45 #include "cluster/tcp.h"
  46
  47 #include "dlmapi.h"
  48 #include "dlmcommon.h"
  49 #include "dlmdomain.h"
  50 #include "dlmdebug.h"
  51
  52 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
  53 #include "cluster/masklog.h"
  54
  55 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
  56                               struct dlm_master_list_entry *mle,
  57                               struct o2nm_node *node,
  58                               int idx);
  59 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
  60                             struct dlm_master_list_entry *mle,
  61                             struct o2nm_node *node,
  62                             int idx);
  63
  64 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
  65 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
  66                                 struct dlm_lock_resource *res,
  67                                 void *nodemap, u32 flags);
  68 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
  69
  70 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
  71                                 struct dlm_master_list_entry *mle,
  72                                 const char *name,
  73                                 unsigned int namelen)
  74 {
  75         if (dlm != mle->dlm)
  76                 return 0;
  77
  78         if (namelen != mle->mnamelen ||
  79             memcmp(name, mle->mname, namelen) != 0)
  80                 return 0;
  81
  82         return 1;
  83 }
  84
  85 static struct kmem_cache *dlm_lockres_cache;
  86 static struct kmem_cache *dlm_lockname_cache;
  87 static struct kmem_cache *dlm_mle_cache;
  88
  89 static void dlm_mle_release(struct kref *kref);
  90 static void dlm_init_mle(struct dlm_master_list_entry *mle,
  91                         enum dlm_mle_type type,
  92                         struct dlm_ctxt *dlm,
  93                         struct dlm_lock_resource *res,
  94                         const char *name,
  95                         unsigned int namelen);
  96 static void dlm_put_mle(struct dlm_master_list_entry *mle);
  97 static void __dlm_put_mle(struct dlm_master_list_entry *mle);
  98 static int dlm_find_mle(struct dlm_ctxt *dlm,
  99                         struct dlm_master_list_entry **mle,
 100                         char *name, unsigned int namelen);
 101
 102 static int dlm_do_master_request(struct dlm_lock_resource *res,
 103                                  struct dlm_master_list_entry *mle, int to);
 104
 105
 106 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
 107                                      struct dlm_lock_resource *res,
 108                                      struct dlm_master_list_entry *mle,
 109                                      int *blocked);
 110 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
 111                                     struct dlm_lock_resource *res,
 112                                     struct dlm_master_list_entry *mle,
 113                                     int blocked);
 114 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
 115                                  struct dlm_lock_resource *res,
 116                                  struct dlm_master_list_entry *mle,
 117                                  struct dlm_master_list_entry **oldmle,
 118                                  const char *name, unsigned int namelen,
 119                                  u8 new_master, u8 master);
 120
 121 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
 122                                     struct dlm_lock_resource *res);
 123 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
 124                                       struct dlm_lock_resource *res);
 125 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
 126                                        struct dlm_lock_resource *res,
 127                                        u8 target);
 128 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
 129                                        struct dlm_lock_resource *res);
 130
 131
 132 int dlm_is_host_down(int errno)
 133 {
 134         switch (errno) {
 135                 case -EBADF:
 136                 case -ECONNREFUSED:
 137                 case -ENOTCONN:
 138                 case -ECONNRESET:
 139                 case -EPIPE:
 140                 case -EHOSTDOWN:
 141                 case -EHOSTUNREACH:
 142                 case -ETIMEDOUT:
 143                 case -ECONNABORTED:
 144                 case -ENETDOWN:
 145                 case -ENETUNREACH:
 146                 case -ENETRESET:
 147                 case -ESHUTDOWN:
 148                 case -ENOPROTOOPT:
 149                 case -EINVAL:   /* if returned from our tcp code,
 150                                    this means there is no socket */
 151                         return 1;
 152         }
 153         return 0;
 154 }
 155
 156
 157 /*
 158  * MASTER LIST FUNCTIONS
 159  */
 160
 161
 162 /*
 163  * regarding master list entries and heartbeat callbacks:
 164  *
 165  * in order to avoid sleeping and allocation that occurs in
 166  * heartbeat, master list entries are simply attached to the
 167  * dlm's established heartbeat callbacks.  the mle is attached
 168  * when it is created, and since the dlm->spinlock is held at
 169  * that time, any heartbeat event will be properly discovered
 170  * by the mle.  the mle needs to be detached from the
 171  * dlm->mle_hb_events list as soon as heartbeat events are no
 172  * longer useful to the mle, and before the mle is freed.
 173  *
 174  * as a general rule, heartbeat events are no longer needed by
 175  * the mle once an "answer" regarding the lock master has been
 176  * received.
 177  */
 178 static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
 179                                               struct dlm_master_list_entry *mle)
 180 {
 181         assert_spin_locked(&dlm->spinlock);
 182
 183         list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
 184 }
 185
 186
 187 static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 188                                               struct dlm_master_list_entry *mle)
 189 {
 190         if (!list_empty(&mle->hb_events))
 191                 list_del_init(&mle->hb_events);
 192 }
 193
 194
 195 static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 196                                             struct dlm_master_list_entry *mle)
 197 {
 198         spin_lock(&dlm->spinlock);
 199         __dlm_mle_detach_hb_events(dlm, mle);
 200         spin_unlock(&dlm->spinlock);
 201 }
 202
 203 static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
 204 {
 205         struct dlm_ctxt *dlm;
 206         dlm = mle->dlm;
 207
 208         assert_spin_locked(&dlm->spinlock);
 209         assert_spin_locked(&dlm->master_lock);
 210         mle->inuse++;
 211         kref_get(&mle->mle_refs);
 212 }
 213
 214 static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
 215 {
 216         struct dlm_ctxt *dlm;
 217         dlm = mle->dlm;
 218
 219         spin_lock(&dlm->spinlock);
 220         spin_lock(&dlm->master_lock);
 221         mle->inuse--;
 222         __dlm_put_mle(mle);
 223         spin_unlock(&dlm->master_lock);
 224         spin_unlock(&dlm->spinlock);
 225
 226 }
 227
 228 /* remove from list and free */
 229 static void __dlm_put_mle(struct dlm_master_list_entry *mle)
 230 {
 231         struct dlm_ctxt *dlm;
 232         dlm = mle->dlm;
 233
 234         assert_spin_locked(&dlm->spinlock);
 235         assert_spin_locked(&dlm->master_lock);
 236         if (!atomic_read(&mle->mle_refs.refcount)) {
 237                 /* this may or may not crash, but who cares.
 238                  * it's a BUG. */
 239                 mlog(ML_ERROR, "bad mle: %p\n", mle);
 240                 dlm_print_one_mle(mle);
 241                 BUG();
 242         } else
 243                 kref_put(&mle->mle_refs, dlm_mle_release);
 244 }
 245
 246
 247 /* must not have any spinlocks coming in */
 248 static void dlm_put_mle(struct dlm_master_list_entry *mle)
 249 {
 250         struct dlm_ctxt *dlm;
 251         dlm = mle->dlm;
 252
 253         spin_lock(&dlm->spinlock);
 254         spin_lock(&dlm->master_lock);
 255         __dlm_put_mle(mle);
 256         spin_unlock(&dlm->master_lock);
 257         spin_unlock(&dlm->spinlock);
 258 }
 259
 260 static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
 261 {
 262         kref_get(&mle->mle_refs);
 263 }
 264
 265 static void dlm_init_mle(struct dlm_master_list_entry *mle,
 266                         enum dlm_mle_type type,
 267                         struct dlm_ctxt *dlm,
 268                         struct dlm_lock_resource *res,
 269                         const char *name,
 270                         unsigned int namelen)
 271 {
 272         assert_spin_locked(&dlm->spinlock);
 273
 274         mle->dlm = dlm;
 275         mle->type = type;
 276         INIT_HLIST_NODE(&mle->master_hash_node);
 277         INIT_LIST_HEAD(&mle->hb_events);
 278         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
 279         spin_lock_init(&mle->spinlock);
 280         init_waitqueue_head(&mle->wq);
 281         atomic_set(&mle->woken, 0);
 282         kref_init(&mle->mle_refs);
 283         memset(mle->response_map, 0, sizeof(mle->response_map));
 284         mle->master = O2NM_MAX_NODES;
 285         mle->new_master = O2NM_MAX_NODES;
 286         mle->inuse = 0;
 287
 288         BUG_ON(mle->type != DLM_MLE_BLOCK &&
 289                mle->type != DLM_MLE_MASTER &&
 290                mle->type != DLM_MLE_MIGRATION);
 291
 292         if (mle->type == DLM_MLE_MASTER) {
 293                 BUG_ON(!res);
 294                 mle->mleres = res;
 295                 memcpy(mle->mname, res->lockname.name, res->lockname.len);
 296                 mle->mnamelen = res->lockname.len;
 297                 mle->mnamehash = res->lockname.hash;
 298         } else {
 299                 BUG_ON(!name);
 300                 mle->mleres = NULL;
 301                 memcpy(mle->mname, name, namelen);
 302                 mle->mnamelen = namelen;
 303                 mle->mnamehash = dlm_lockid_hash(name, namelen);
 304         }
 305
 306         atomic_inc(&dlm->mle_tot_count[mle->type]);
 307         atomic_inc(&dlm->mle_cur_count[mle->type]);
 308
 309         /* copy off the node_map and register hb callbacks on our copy */
 310         memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
 311         memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
 312         clear_bit(dlm->node_num, mle->vote_map);
 313         clear_bit(dlm->node_num, mle->node_map);
 314
 315         /* attach the mle to the domain node up/down events */
 316         __dlm_mle_attach_hb_events(dlm, mle);
 317 }
 318
 319 void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
 320 {
 321         assert_spin_locked(&dlm->spinlock);
 322         assert_spin_locked(&dlm->master_lock);
 323
 324         if (!hlist_unhashed(&mle->master_hash_node))
 325                 hlist_del_init(&mle->master_hash_node);
 326 }
 327
 328 void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
 329 {
 330         struct hlist_head *bucket;
 331
 332         assert_spin_locked(&dlm->master_lock);
 333
 334         bucket = dlm_master_hash(dlm, mle->mnamehash);
 335         hlist_add_head(&mle->master_hash_node, bucket);
 336 }
 337
 338 /* returns 1 if found, 0 if not */
 339 static int dlm_find_mle(struct dlm_ctxt *dlm,
 340                         struct dlm_master_list_entry **mle,
 341                         char *name, unsigned int namelen)
 342 {
 343         struct dlm_master_list_entry *tmpmle;
 344         struct hlist_head *bucket;
 345         unsigned int hash;
 346
 347         assert_spin_locked(&dlm->master_lock);
 348
 349         hash = dlm_lockid_hash(name, namelen);
 350         bucket = dlm_master_hash(dlm, hash);
 351         hlist_for_each_entry(tmpmle, bucket, master_hash_node) {
 352                 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
 353                         continue;
 354                 dlm_get_mle(tmpmle);
 355                 *mle = tmpmle;
 356                 return 1;
 357         }
 358         return 0;
 359 }
 360
 361 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
 362 {
 363         struct dlm_master_list_entry *mle;
 364
 365         assert_spin_locked(&dlm->spinlock);
 366
 367         list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
 368                 if (node_up)
 369                         dlm_mle_node_up(dlm, mle, NULL, idx);
 370                 else
 371                         dlm_mle_node_down(dlm, mle, NULL, idx);
 372         }
 373 }
 374
 375 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
 376                               struct dlm_master_list_entry *mle,
 377                               struct o2nm_node *node, int idx)
 378 {
 379         spin_lock(&mle->spinlock);
 380
 381         if (!test_bit(idx, mle->node_map))
 382                 mlog(0, "node %u already removed from nodemap!\n", idx);
 383         else
 384                 clear_bit(idx, mle->node_map);
 385
 386         spin_unlock(&mle->spinlock);
 387 }
 388
 389 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
 390                             struct dlm_master_list_entry *mle,
 391                             struct o2nm_node *node, int idx)
 392 {
 393         spin_lock(&mle->spinlock);
 394
 395         if (test_bit(idx, mle->node_map))
 396                 mlog(0, "node %u already in node map!\n", idx);
 397         else
 398                 set_bit(idx, mle->node_map);
 399
 400         spin_unlock(&mle->spinlock);
 401 }
 402
 403
 404 int dlm_init_mle_cache(void)
 405 {
 406         dlm_mle_cache = kmem_cache_create("o2dlm_mle",
 407                                           sizeof(struct dlm_master_list_entry),
 408                                           0, SLAB_HWCACHE_ALIGN,
 409                                           NULL);
 410         if (dlm_mle_cache == NULL)
 411                 return -ENOMEM;
 412         return 0;
 413 }
 414
 415 void dlm_destroy_mle_cache(void)
 416 {
 417         if (dlm_mle_cache)
 418                 kmem_cache_destroy(dlm_mle_cache);
 419 }
 420
 421 static void dlm_mle_release(struct kref *kref)
 422 {
 423         struct dlm_master_list_entry *mle;
 424         struct dlm_ctxt *dlm;
 425
 426         mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
 427         dlm = mle->dlm;
 428
 429         assert_spin_locked(&dlm->spinlock);
 430         assert_spin_locked(&dlm->master_lock);
 431
 432         mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
 433              mle->type);
 434
 435         /* remove from list if not already */
 436         __dlm_unlink_mle(dlm, mle);
 437
 438         /* detach the mle from the domain node up/down events */
 439         __dlm_mle_detach_hb_events(dlm, mle);
 440
 441         atomic_dec(&dlm->mle_cur_count[mle->type]);
 442
 443         /* NOTE: kfree under spinlock here.
 444          * if this is bad, we can move this to a freelist. */
 445         kmem_cache_free(dlm_mle_cache, mle);
 446 }
 447
 448
 449 /*
 450  * LOCK RESOURCE FUNCTIONS
 451  */
 452
 453 int dlm_init_master_caches(void)
 454 {
 455         dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
 456                                               sizeof(struct dlm_lock_resource),
 457                                               0, SLAB_HWCACHE_ALIGN, NULL);
 458         if (!dlm_lockres_cache)
 459                 goto bail;
 460
 461         dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
 462                                                DLM_LOCKID_NAME_MAX, 0,
 463                                                SLAB_HWCACHE_ALIGN, NULL);
 464         if (!dlm_lockname_cache)
 465                 goto bail;
 466
 467         return 0;
 468 bail:
 469         dlm_destroy_master_caches();
 470         return -ENOMEM;
 471 }
 472
 473 void dlm_destroy_master_caches(void)
 474 {
 475         if (dlm_lockname_cache) {
 476                 kmem_cache_destroy(dlm_lockname_cache);
 477                 dlm_lockname_cache = NULL;
 478         }
 479
 480         if (dlm_lockres_cache) {
 481                 kmem_cache_destroy(dlm_lockres_cache);
 482                 dlm_lockres_cache = NULL;
 483         }
 484 }
 485
 486 static void dlm_lockres_release(struct kref *kref)
 487 {
 488         struct dlm_lock_resource *res;
 489         struct dlm_ctxt *dlm;
 490
 491         res = container_of(kref, struct dlm_lock_resource, refs);
 492         dlm = res->dlm;
 493
 494         /* This should not happen -- all lockres' have a name
 495          * associated with them at init time. */
 496         BUG_ON(!res->lockname.name);
 497
 498         mlog(0, "destroying lockres %.*s\n", res->lockname.len,
 499              res->lockname.name);
 500
 501         spin_lock(&dlm->track_lock);
 502         if (!list_empty(&res->tracking))
 503                 list_del_init(&res->tracking);
 504         else {
 505                 mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n",
 506                      res->lockname.len, res->lockname.name);
 507                 dlm_print_one_lock_resource(res);
 508         }
 509         spin_unlock(&dlm->track_lock);
 510
 511         atomic_dec(&dlm->res_cur_count);
 512
 513         if (!hlist_unhashed(&res->hash_node) ||
 514             !list_empty(&res->granted) ||
 515             !list_empty(&res->converting) ||
 516             !list_empty(&res->blocked) ||
 517             !list_empty(&res->dirty) ||
 518             !list_empty(&res->recovering) ||
 519             !list_empty(&res->purge)) {
 520                 mlog(ML_ERROR,
 521                      "Going to BUG for resource %.*s."
 522                      "  We're on a list! [%c%c%c%c%c%c%c]\n",
 523                      res->lockname.len, res->lockname.name,
 524                      !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
 525                      !list_empty(&res->granted) ? 'G' : ' ',
 526                      !list_empty(&res->converting) ? 'C' : ' ',
 527                      !list_empty(&res->blocked) ? 'B' : ' ',
 528                      !list_empty(&res->dirty) ? 'D' : ' ',
 529                      !list_empty(&res->recovering) ? 'R' : ' ',
 530                      !list_empty(&res->purge) ? 'P' : ' ');
 531
 532                 dlm_print_one_lock_resource(res);
 533         }
 534
 535         /* By the time we're ready to blow this guy away, we shouldn't
 536          * be on any lists. */
 537         BUG_ON(!hlist_unhashed(&res->hash_node));
 538         BUG_ON(!list_empty(&res->granted));
 539         BUG_ON(!list_empty(&res->converting));
 540         BUG_ON(!list_empty(&res->blocked));
 541         BUG_ON(!list_empty(&res->dirty));
 542         BUG_ON(!list_empty(&res->recovering));
 543         BUG_ON(!list_empty(&res->purge));
 544
 545         kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
 546
 547         kmem_cache_free(dlm_lockres_cache, res);
 548 }
 549
 550 void dlm_lockres_put(struct dlm_lock_resource *res)
 551 {
 552         kref_put(&res->refs, dlm_lockres_release);
 553 }
 554
 555 static void dlm_init_lockres(struct dlm_ctxt *dlm,
 556                              struct dlm_lock_resource *res,
 557                              const char *name, unsigned int namelen)
 558 {
 559         char *qname;
 560
 561         /* If we memset here, we lose our reference to the kmalloc'd
 562          * res->lockname.name, so be sure to init every field
 563          * correctly! */
 564
 565         qname = (char *) res->lockname.name;
 566         memcpy(qname, name, namelen);
 567
 568         res->lockname.len = namelen;
 569         res->lockname.hash = dlm_lockid_hash(name, namelen);
 570
 571         init_waitqueue_head(&res->wq);
 572         spin_lock_init(&res->spinlock);
 573         INIT_HLIST_NODE(&res->hash_node);
 574         INIT_LIST_HEAD(&res->granted);
 575         INIT_LIST_HEAD(&res->converting);
 576         INIT_LIST_HEAD(&res->blocked);
 577         INIT_LIST_HEAD(&res->dirty);
 578         INIT_LIST_HEAD(&res->recovering);
 579         INIT_LIST_HEAD(&res->purge);
 580         INIT_LIST_HEAD(&res->tracking);
 581         atomic_set(&res->asts_reserved, 0);
 582         res->migration_pending = 0;
 583         res->inflight_locks = 0;
 584
 585         res->dlm = dlm;
 586
 587         kref_init(&res->refs);
 588
 589         atomic_inc(&dlm->res_tot_count);
 590         atomic_inc(&dlm->res_cur_count);
 591
 592         /* just for consistency */
 593         spin_lock(&res->spinlock);
 594         dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
 595         spin_unlock(&res->spinlock);
 596
 597         res->state = DLM_LOCK_RES_IN_PROGRESS;
 598
 599         res->last_used = 0;
 600
 601         spin_lock(&dlm->spinlock);
 602         list_add_tail(&res->tracking, &dlm->tracking_list);
 603         spin_unlock(&dlm->spinlock);
 604
 605         memset(res->lvb, 0, DLM_LVB_LEN);
 606         memset(res->refmap, 0, sizeof(res->refmap));
 607 }
 608
 609 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
 610                                    const char *name,
 611                                    unsigned int namelen)
 612 {
 613         struct dlm_lock_resource *res = NULL;
 614
 615         res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
 616         if (!res)
 617                 goto error;
 618
 619         res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
 620         if (!res->lockname.name)
 621                 goto error;
 622
 623         dlm_init_lockres(dlm, res, name, namelen);
 624         return res;
 625
 626 error:
 627         if (res && res->lockname.name)
 628                 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
 629
 630         if (res)
 631                 kmem_cache_free(dlm_lockres_cache, res);
 632         return NULL;
 633 }
 634
 635 void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm,
 636                                 struct dlm_lock_resource *res, int bit)
 637 {
 638         assert_spin_locked(&res->spinlock);
 639
 640         mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len,
 641              res->lockname.name, bit, __builtin_return_address(0));
 642
 643         set_bit(bit, res->refmap);
 644 }
 645
 646 void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm,
 647                                   struct dlm_lock_resource *res, int bit)
 648 {
 649         assert_spin_locked(&res->spinlock);
 650
 651         mlog(0, "res %.*s, clr node %u, %ps()\n", res->lockname.len,
 652              res->lockname.name, bit, __builtin_return_address(0));
 653
 654         clear_bit(bit, res->refmap);
 655 }
 656
 657
 658 void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
 659                                    struct dlm_lock_resource *res)
 660 {
 661         assert_spin_locked(&res->spinlock);
 662
 663         res->inflight_locks++;
 664
 665         mlog(0, "%s: res %.*s, inflight++: now %u, %ps()\n", dlm->name,
 666              res->lockname.len, res->lockname.name, res->inflight_locks,
 667              __builtin_return_address(0));
 668 }
 669
 670 void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
 671                                    struct dlm_lock_resource *res)
 672 {
 673         assert_spin_locked(&res->spinlock);
 674
 675         BUG_ON(res->inflight_locks == 0);
 676
 677         res->inflight_locks--;
 678
 679         mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm->name,
 680              res->lockname.len, res->lockname.name, res->inflight_locks,
 681              __builtin_return_address(0));
 682
 683         wake_up(&res->wq);
 684 }
 685
 686 /*
 687  * lookup a lock resource by name.
 688  * may already exist in the hashtable.
 689  * lockid is null terminated
 690  *
 691  * if not, allocate enough for the lockres and for
 692  * the temporary structure used in doing the mastering.
 693  *
 694  * also, do a lookup in the dlm->master_list to see
 695  * if another node has begun mastering the same lock.
 696  * if so, there should be a block entry in there
 697  * for this name, and we should *not* attempt to master
 698  * the lock here.   need to wait around for that node
 699  * to assert_master (or die).
 700  *
 701  */
 702 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
 703                                           const char *lockid,
 704                                           int namelen,
 705                                           int flags)
 706 {
 707         struct dlm_lock_resource *tmpres=NULL, *res=NULL;
 708         struct dlm_master_list_entry *mle = NULL;
 709         struct dlm_master_list_entry *alloc_mle = NULL;
 710         int blocked = 0;
 711         int ret, nodenum;
 712         struct dlm_node_iter iter;
 713         unsigned int hash;
 714         int tries = 0;
 715         int bit, wait_on_recovery = 0;
 716
 717         BUG_ON(!lockid);
 718
 719         hash = dlm_lockid_hash(lockid, namelen);
 720
 721         mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
 722
 723 lookup:
 724         spin_lock(&dlm->spinlock);
 725         tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
 726         if (tmpres) {
 727                 spin_unlock(&dlm->spinlock);
 728                 spin_lock(&tmpres->spinlock);
 729                 /* Wait on the thread that is mastering the resource */
 730                 if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
 731                         __dlm_wait_on_lockres(tmpres);
 732                         BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
 733                         spin_unlock(&tmpres->spinlock);
 734                         dlm_lockres_put(tmpres);
 735                         tmpres = NULL;
 736                         goto lookup;
 737                 }
 738
 739                 /* Wait on the resource purge to complete before continuing */
 740                 if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) {
 741                         BUG_ON(tmpres->owner == dlm->node_num);
 742                         __dlm_wait_on_lockres_flags(tmpres,
 743                                                     DLM_LOCK_RES_DROPPING_REF);
 744                         spin_unlock(&tmpres->spinlock);
 745                         dlm_lockres_put(tmpres);
 746                         tmpres = NULL;
 747                         goto lookup;
 748                 }
 749
 750                 /* Grab inflight ref to pin the resource */
 751                 dlm_lockres_grab_inflight_ref(dlm, tmpres);
 752
 753                 spin_unlock(&tmpres->spinlock);
 754                 if (res)
 755                         dlm_lockres_put(res);
 756                 res = tmpres;
 757                 goto leave;
 758         }
 759
 760         if (!res) {
 761                 spin_unlock(&dlm->spinlock);
 762                 mlog(0, "allocating a new resource\n");
 763                 /* nothing found and we need to allocate one. */
 764                 alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
 765                 if (!alloc_mle)
 766                         goto leave;
 767                 res = dlm_new_lockres(dlm, lockid, namelen);
 768                 if (!res)
 769                         goto leave;
 770                 goto lookup;
 771         }
 772
 773         mlog(0, "no lockres found, allocated our own: %p\n", res);
 774
 775         if (flags & LKM_LOCAL) {
 776                 /* caller knows it's safe to assume it's not mastered elsewhere
 777                  * DONE!  return right away */
 778                 spin_lock(&res->spinlock);
 779                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
 780                 __dlm_insert_lockres(dlm, res);
 781                 dlm_lockres_grab_inflight_ref(dlm, res);
 782                 spin_unlock(&res->spinlock);
 783                 spin_unlock(&dlm->spinlock);
 784                 /* lockres still marked IN_PROGRESS */
 785                 goto wake_waiters;
 786         }
 787
 788         /* check master list to see if another node has started mastering it */
 789         spin_lock(&dlm->master_lock);
 790
 791         /* if we found a block, wait for lock to be mastered by another node */
 792         blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
 793         if (blocked) {
 794                 int mig;
 795                 if (mle->type == DLM_MLE_MASTER) {
 796                         mlog(ML_ERROR, "master entry for nonexistent lock!\n");
 797                         BUG();
 798                 }
 799                 mig = (mle->type == DLM_MLE_MIGRATION);
 800                 /* if there is a migration in progress, let the migration
 801                  * finish before continuing.  we can wait for the absence
 802                  * of the MIGRATION mle: either the migrate finished or
 803                  * one of the nodes died and the mle was cleaned up.
 804                  * if there is a BLOCK here, but it already has a master
 805                  * set, we are too late.  the master does not have a ref
 806                  * for us in the refmap.  detach the mle and drop it.
 807                  * either way, go back to the top and start over. */
 808                 if (mig || mle->master != O2NM_MAX_NODES) {
 809                         BUG_ON(mig && mle->master == dlm->node_num);
 810                         /* we arrived too late.  the master does not
 811                          * have a ref for us. retry. */
 812                         mlog(0, "%s:%.*s: late on %s\n",
 813                              dlm->name, namelen, lockid,
 814                              mig ?  "MIGRATION" : "BLOCK");
 815                         spin_unlock(&dlm->master_lock);
 816                         spin_unlock(&dlm->spinlock);
 817
 818                         /* master is known, detach */
 819                         if (!mig)
 820                                 dlm_mle_detach_hb_events(dlm, mle);
 821                         dlm_put_mle(mle);
 822                         mle = NULL;
 823                         /* this is lame, but we can't wait on either
 824                          * the mle or lockres waitqueue here */
 825                         if (mig)
 826                                 msleep(100);
 827                         goto lookup;
 828                 }
 829         } else {
 830                 /* go ahead and try to master lock on this node */
 831                 mle = alloc_mle;
 832                 /* make sure this does not get freed below */
 833                 alloc_mle = NULL;
 834                 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
 835                 set_bit(dlm->node_num, mle->maybe_map);
 836                 __dlm_insert_mle(dlm, mle);
 837
 838                 /* still holding the dlm spinlock, check the recovery map
 839                  * to see if there are any nodes that still need to be
 840                  * considered.  these will not appear in the mle nodemap
 841                  * but they might own this lockres.  wait on them. */
 842                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 843                 if (bit < O2NM_MAX_NODES) {
 844                         mlog(0, "%s: res %.*s, At least one node (%d) "
 845                              "to recover before lock mastery can begin\n",
 846                              dlm->name, namelen, (char *)lockid, bit);
 847                         wait_on_recovery = 1;
 848                 }
 849         }
 850
 851         /* at this point there is either a DLM_MLE_BLOCK or a
 852          * DLM_MLE_MASTER on the master list, so it's safe to add the
 853          * lockres to the hashtable.  anyone who finds the lock will
 854          * still have to wait on the IN_PROGRESS. */
 855
 856         /* finally add the lockres to its hash bucket */
 857         __dlm_insert_lockres(dlm, res);
 858
 859         /* Grab inflight ref to pin the resource */
 860         spin_lock(&res->spinlock);
 861         dlm_lockres_grab_inflight_ref(dlm, res);
 862         spin_unlock(&res->spinlock);
 863
 864         /* get an extra ref on the mle in case this is a BLOCK
 865          * if so, the creator of the BLOCK may try to put the last
 866          * ref at this time in the assert master handler, so we
 867          * need an extra one to keep from a bad ptr deref. */
 868         dlm_get_mle_inuse(mle);
 869         spin_unlock(&dlm->master_lock);
 870         spin_unlock(&dlm->spinlock);
 871
 872 redo_request:
 873         while (wait_on_recovery) {
 874                 /* any cluster changes that occurred after dropping the
 875                  * dlm spinlock would be detectable be a change on the mle,
 876                  * so we only need to clear out the recovery map once. */
 877                 if (dlm_is_recovery_lock(lockid, namelen)) {
 878                         mlog(0, "%s: Recovery map is not empty, but must "
 879                              "master $RECOVERY lock now\n", dlm->name);
 880                         if (!dlm_pre_master_reco_lockres(dlm, res))
 881                                 wait_on_recovery = 0;
 882                         else {
 883                                 mlog(0, "%s: waiting 500ms for heartbeat state "
 884                                     "change\n", dlm->name);
 885                                 msleep(500);
 886                         }
 887                         continue;
 888                 }
 889
 890                 dlm_kick_recovery_thread(dlm);
 891                 msleep(1000);
 892                 dlm_wait_for_recovery(dlm);
 893
 894                 spin_lock(&dlm->spinlock);
 895                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 896                 if (bit < O2NM_MAX_NODES) {
 897                         mlog(0, "%s: res %.*s, At least one node (%d) "
 898                              "to recover before lock mastery can begin\n",
 899                              dlm->name, namelen, (char *)lockid, bit);
 900                         wait_on_recovery = 1;
 901                 } else
 902                         wait_on_recovery = 0;
 903                 spin_unlock(&dlm->spinlock);
 904
 905                 if (wait_on_recovery)
 906                         dlm_wait_for_node_recovery(dlm, bit, 10000);
 907         }
 908
 909         /* must wait for lock to be mastered elsewhere */
 910         if (blocked)
 911                 goto wait;
 912
 913         ret = -EINVAL;
 914         dlm_node_iter_init(mle->vote_map, &iter);
 915         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
 916                 ret = dlm_do_master_request(res, mle, nodenum);
 917                 if (ret < 0)
 918                         mlog_errno(ret);
 919                 if (mle->master != O2NM_MAX_NODES) {
 920                         /* found a master ! */
 921                         if (mle->master <= nodenum)
 922                                 break;
 923                         /* if our master request has not reached the master
 924                          * yet, keep going until it does.  this is how the
 925                          * master will know that asserts are needed back to
 926                          * the lower nodes. */
 927                         mlog(0, "%s: res %.*s, Requests only up to %u but "
 928                              "master is %u, keep going\n", dlm->name, namelen,
 929                              lockid, nodenum, mle->master);
 930                 }
 931         }
 932
 933 wait:
 934         /* keep going until the response map includes all nodes */
 935         ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
 936         if (ret < 0) {
 937                 wait_on_recovery = 1;
 938                 mlog(0, "%s: res %.*s, Node map changed, redo the master "
 939                      "request now, blocked=%d\n", dlm->name, res->lockname.len,
 940                      res->lockname.name, blocked);
 941                 if (++tries > 20) {
 942                         mlog(ML_ERROR, "%s: res %.*s, Spinning on "
 943                              "dlm_wait_for_lock_mastery, blocked = %d\n",
 944                              dlm->name, res->lockname.len,
 945                              res->lockname.name, blocked);
 946                         dlm_print_one_lock_resource(res);
 947                         dlm_print_one_mle(mle);
 948                         tries = 0;
 949                 }
 950                 goto redo_request;
 951         }
 952
 953         mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len,
 954              res->lockname.name, res->owner);
 955         /* make sure we never continue without this */
 956         BUG_ON(res->owner == O2NM_MAX_NODES);
 957
 958         /* master is known, detach if not already detached */
 959         dlm_mle_detach_hb_events(dlm, mle);
 960         dlm_put_mle(mle);
 961         /* put the extra ref */
 962         dlm_put_mle_inuse(mle);
 963
 964 wake_waiters:
 965         spin_lock(&res->spinlock);
 966         res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
 967         spin_unlock(&res->spinlock);
 968         wake_up(&res->wq);
 969
 970 leave:
 971         /* need to free the unused mle */
 972         if (alloc_mle)
 973                 kmem_cache_free(dlm_mle_cache, alloc_mle);
 974
 975         return res;
 976 }
 977
 978
 979 #define DLM_MASTERY_TIMEOUT_MS   5000
 980
 981 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
 982                                      struct dlm_lock_resource *res,
 983                                      struct dlm_master_list_entry *mle,
 984                                      int *blocked)
 985 {
 986         u8 m;
 987         int ret, bit;
 988         int map_changed, voting_done;
 989         int assert, sleep;
 990
 991 recheck:
 992         ret = 0;
 993         assert = 0;
 994
 995         /* check if another node has already become the owner */
 996         spin_lock(&res->spinlock);
 997         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
 998                 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
 999                      res->lockname.len, res->lockname.name, res->owner);
1000                 spin_unlock(&res->spinlock);
1001                 /* this will cause the master to re-assert across
1002                  * the whole cluster, freeing up mles */
1003                 if (res->owner != dlm->node_num) {
1004                         ret = dlm_do_master_request(res, mle, res->owner);
1005                         if (ret < 0) {
1006                                 /* give recovery a chance to run */
1007                                 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1008                                 msleep(500);
1009                                 goto recheck;
1010                         }
1011                 }
1012                 ret = 0;
1013                 goto leave;
1014         }
1015         spin_unlock(&res->spinlock);
1016
1017         spin_lock(&mle->spinlock);
1018         m = mle->master;
1019         map_changed = (memcmp(mle->vote_map, mle->node_map,
1020                               sizeof(mle->vote_map)) != 0);
1021         voting_done = (memcmp(mle->vote_map, mle->response_map,
1022                              sizeof(mle->vote_map)) == 0);
1023
1024         /* restart if we hit any errors */
1025         if (map_changed) {
1026                 int b;
1027                 mlog(0, "%s: %.*s: node map changed, restarting\n",
1028                      dlm->name, res->lockname.len, res->lockname.name);
1029                 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1030                 b = (mle->type == DLM_MLE_BLOCK);
1031                 if ((*blocked && !b) || (!*blocked && b)) {
1032                         mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1033                              dlm->name, res->lockname.len, res->lockname.name,
1034                              *blocked, b);
1035                         *blocked = b;
1036                 }
1037                 spin_unlock(&mle->spinlock);
1038                 if (ret < 0) {
1039                         mlog_errno(ret);
1040                         goto leave;
1041                 }
1042                 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1043                      "rechecking now\n", dlm->name, res->lockname.len,
1044                      res->lockname.name);
1045                 goto recheck;
1046         } else {
1047                 if (!voting_done) {
1048                         mlog(0, "map not changed and voting not done "
1049                              "for %s:%.*s\n", dlm->name, res->lockname.len,
1050                              res->lockname.name);
1051                 }
1052         }
1053
1054         if (m != O2NM_MAX_NODES) {
1055                 /* another node has done an assert!
1056                  * all done! */
1057                 sleep = 0;
1058         } else {
1059                 sleep = 1;
1060                 /* have all nodes responded? */
1061                 if (voting_done && !*blocked) {
1062                         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1063                         if (dlm->node_num <= bit) {
1064                                 /* my node number is lowest.
1065                                  * now tell other nodes that I am
1066                                  * mastering this. */
1067                                 mle->master = dlm->node_num;
1068                                 /* ref was grabbed in get_lock_resource
1069                                  * will be dropped in dlmlock_master */
1070                                 assert = 1;
1071                                 sleep = 0;
1072                         }
1073                         /* if voting is done, but we have not received
1074                          * an assert master yet, we must sleep */
1075                 }
1076         }
1077
1078         spin_unlock(&mle->spinlock);
1079
1080         /* sleep if we haven't finished voting yet */
1081         if (sleep) {
1082                 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1083
1084                 /*
1085                 if (atomic_read(&mle->mle_refs.refcount) < 2)
1086                         mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1087                         atomic_read(&mle->mle_refs.refcount),
1088                         res->lockname.len, res->lockname.name);
1089                 */
1090                 atomic_set(&mle->woken, 0);
1091                 (void)wait_event_timeout(mle->wq,
1092                                          (atomic_read(&mle->woken) == 1),
1093                                          timeo);
1094                 if (res->owner == O2NM_MAX_NODES) {
1095                         mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1096                              res->lockname.len, res->lockname.name);
1097                         goto recheck;
1098                 }
1099                 mlog(0, "done waiting, master is %u\n", res->owner);
1100                 ret = 0;
1101                 goto leave;
1102         }
1103
1104         ret = 0;   /* done */
1105         if (assert) {
1106                 m = dlm->node_num;
1107                 mlog(0, "about to master %.*s here, this=%u\n",
1108                      res->lockname.len, res->lockname.name, m);
1109                 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1110                 if (ret) {
1111                         /* This is a failure in the network path,
1112                          * not in the response to the assert_master
1113                          * (any nonzero response is a BUG on this node).
1114                          * Most likely a socket just got disconnected
1115                          * due to node death. */
1116                         mlog_errno(ret);
1117                 }
1118                 /* no longer need to restart lock mastery.
1119                  * all living nodes have been contacted. */
1120                 ret = 0;
1121         }
1122
1123         /* set the lockres owner */
1124         spin_lock(&res->spinlock);
1125         /* mastery reference obtained either during
1126          * assert_master_handler or in get_lock_resource */
1127         dlm_change_lockres_owner(dlm, res, m);
1128         spin_unlock(&res->spinlock);
1129
1130 leave:
1131         return ret;
1132 }
1133
1134 struct dlm_bitmap_diff_iter
1135 {
1136         int curnode;
1137         unsigned long *orig_bm;
1138         unsigned long *cur_bm;
1139         unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1140 };
1141
1142 enum dlm_node_state_change
1143 {
1144         NODE_DOWN = -1,
1145         NODE_NO_CHANGE = 0,
1146         NODE_UP
1147 };
1148
1149 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1150                                       unsigned long *orig_bm,
1151                                       unsigned long *cur_bm)
1152 {
1153         unsigned long p1, p2;
1154         int i;
1155
1156         iter->curnode = -1;
1157         iter->orig_bm = orig_bm;
1158         iter->cur_bm = cur_bm;
1159
1160         for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1161                 p1 = *(iter->orig_bm + i);
1162                 p2 = *(iter->cur_bm + i);
1163                 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1164         }
1165 }
1166
1167 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1168                                      enum dlm_node_state_change *state)
1169 {
1170         int bit;
1171
1172         if (iter->curnode >= O2NM_MAX_NODES)
1173                 return -ENOENT;
1174
1175         bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1176                             iter->curnode+1);
1177         if (bit >= O2NM_MAX_NODES) {
1178                 iter->curnode = O2NM_MAX_NODES;
1179                 return -ENOENT;
1180         }
1181
1182         /* if it was there in the original then this node died */
1183         if (test_bit(bit, iter->orig_bm))
1184                 *state = NODE_DOWN;
1185         else
1186                 *state = NODE_UP;
1187
1188         iter->curnode = bit;
1189         return bit;
1190 }
1191
1192
1193 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1194                                     struct dlm_lock_resource *res,
1195                                     struct dlm_master_list_entry *mle,
1196                                     int blocked)
1197 {
1198         struct dlm_bitmap_diff_iter bdi;
1199         enum dlm_node_state_change sc;
1200         int node;
1201         int ret = 0;
1202
1203         mlog(0, "something happened such that the "
1204              "master process may need to be restarted!\n");
1205
1206         assert_spin_locked(&mle->spinlock);
1207
1208         dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1209         node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1210         while (node >= 0) {
1211                 if (sc == NODE_UP) {
1212                         /* a node came up.  clear any old vote from
1213                          * the response map and set it in the vote map
1214                          * then restart the mastery. */
1215                         mlog(ML_NOTICE, "node %d up while restarting\n", node);
1216
1217                         /* redo the master request, but only for the new node */
1218                         mlog(0, "sending request to new node\n");
1219                         clear_bit(node, mle->response_map);
1220                         set_bit(node, mle->vote_map);
1221                 } else {
1222                         mlog(ML_ERROR, "node down! %d\n", node);
1223                         if (blocked) {
1224                                 int lowest = find_next_bit(mle->maybe_map,
1225                                                        O2NM_MAX_NODES, 0);
1226
1227                                 /* act like it was never there */
1228                                 clear_bit(node, mle->maybe_map);
1229
1230                                 if (node == lowest) {
1231                                         mlog(0, "expected master %u died"
1232                                             " while this node was blocked "
1233                                             "waiting on it!\n", node);
1234                                         lowest = find_next_bit(mle->maybe_map,
1235                                                         O2NM_MAX_NODES,
1236                                                         lowest+1);
1237                                         if (lowest < O2NM_MAX_NODES) {
1238                                                 mlog(0, "%s:%.*s:still "
1239                                                      "blocked. waiting on %u "
1240                                                      "now\n", dlm->name,
1241                                                      res->lockname.len,
1242                                                      res->lockname.name,
1243                                                      lowest);
1244                                         } else {
1245                                                 /* mle is an MLE_BLOCK, but
1246                                                  * there is now nothing left to
1247                                                  * block on.  we need to return
1248                                                  * all the way back out and try
1249                                                  * again with an MLE_MASTER.
1250                                                  * dlm_do_local_recovery_cleanup
1251                                                  * has already run, so the mle
1252                                                  * refcount is ok */
1253                                                 mlog(0, "%s:%.*s: no "
1254                                                      "longer blocking. try to "
1255                                                      "master this here\n",
1256                                                      dlm->name,
1257                                                      res->lockname.len,
1258                                                      res->lockname.name);
1259                                                 mle->type = DLM_MLE_MASTER;
1260                                                 mle->mleres = res;
1261                                         }
1262                                 }
1263                         }
1264
1265                         /* now blank out everything, as if we had never
1266                          * contacted anyone */
1267                         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1268                         memset(mle->response_map, 0, sizeof(mle->response_map));
1269                         /* reset the vote_map to the current node_map */
1270                         memcpy(mle->vote_map, mle->node_map,
1271                                sizeof(mle->node_map));
1272                         /* put myself into the maybe map */
1273                         if (mle->type != DLM_MLE_BLOCK)
1274                                 set_bit(dlm->node_num, mle->maybe_map);
1275                 }
1276                 ret = -EAGAIN;
1277                 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1278         }
1279         return ret;
1280 }
1281
1282
1283 /*
1284  * DLM_MASTER_REQUEST_MSG
1285  *
1286  * returns: 0 on success,
1287  *          -errno on a network error
1288  *
1289  * on error, the caller should assume the target node is "dead"
1290  *
1291  */
1292
1293 static int dlm_do_master_request(struct dlm_lock_resource *res,
1294                                  struct dlm_master_list_entry *mle, int to)
1295 {
1296         struct dlm_ctxt *dlm = mle->dlm;
1297         struct dlm_master_request request;
1298         int ret, response=0, resend;
1299
1300         memset(&request, 0, sizeof(request));
1301         request.node_idx = dlm->node_num;
1302
1303         BUG_ON(mle->type == DLM_MLE_MIGRATION);
1304
1305         request.namelen = (u8)mle->mnamelen;
1306         memcpy(request.name, mle->mname, request.namelen);
1307
1308 again:
1309         ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1310                                  sizeof(request), to, &response);
1311         if (ret < 0)  {
1312                 if (ret == -ESRCH) {
1313                         /* should never happen */
1314                         mlog(ML_ERROR, "TCP stack not ready!\n");
1315                         BUG();
1316                 } else if (ret == -EINVAL) {
1317                         mlog(ML_ERROR, "bad args passed to o2net!\n");
1318                         BUG();
1319                 } else if (ret == -ENOMEM) {
1320                         mlog(ML_ERROR, "out of memory while trying to send "
1321                              "network message!  retrying\n");
1322                         /* this is totally crude */
1323                         msleep(50);
1324                         goto again;
1325                 } else if (!dlm_is_host_down(ret)) {
1326                         /* not a network error. bad. */
1327                         mlog_errno(ret);
1328                         mlog(ML_ERROR, "unhandled error!");
1329                         BUG();
1330                 }
1331                 /* all other errors should be network errors,
1332                  * and likely indicate node death */
1333                 mlog(ML_ERROR, "link to %d went down!\n", to);
1334                 goto out;
1335         }
1336
1337         ret = 0;
1338         resend = 0;
1339         spin_lock(&mle->spinlock);
1340         switch (response) {
1341                 case DLM_MASTER_RESP_YES:
1342                         set_bit(to, mle->response_map);
1343                         mlog(0, "node %u is the master, response=YES\n", to);
1344                         mlog(0, "%s:%.*s: master node %u now knows I have a "
1345                              "reference\n", dlm->name, res->lockname.len,
1346                              res->lockname.name, to);
1347                         mle->master = to;
1348                         break;
1349                 case DLM_MASTER_RESP_NO:
1350                         mlog(0, "node %u not master, response=NO\n", to);
1351                         set_bit(to, mle->response_map);
1352                         break;
1353                 case DLM_MASTER_RESP_MAYBE:
1354                         mlog(0, "node %u not master, response=MAYBE\n", to);
1355                         set_bit(to, mle->response_map);
1356                         set_bit(to, mle->maybe_map);
1357                         break;
1358                 case DLM_MASTER_RESP_ERROR:
1359                         mlog(0, "node %u hit an error, resending\n", to);
1360                         resend = 1;
1361                         response = 0;
1362                         break;
1363                 default:
1364                         mlog(ML_ERROR, "bad response! %u\n", response);
1365                         BUG();
1366         }
1367         spin_unlock(&mle->spinlock);
1368         if (resend) {
1369                 /* this is also totally crude */
1370                 msleep(50);
1371                 goto again;
1372         }
1373
1374 out:
1375         return ret;
1376 }
1377
1378 /*
1379  * locks that can be taken here:
1380  * dlm->spinlock
1381  * res->spinlock
1382  * mle->spinlock
1383  * dlm->master_list
1384  *
1385  * if possible, TRIM THIS DOWN!!!
1386  */
1387 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1388                                void **ret_data)
1389 {
1390         u8 response = DLM_MASTER_RESP_MAYBE;
1391         struct dlm_ctxt *dlm = data;
1392         struct dlm_lock_resource *res = NULL;
1393         struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1394         struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1395         char *name;
1396         unsigned int namelen, hash;
1397         int found, ret;
1398         int set_maybe;
1399         int dispatch_assert = 0;
1400
1401         if (!dlm_grab(dlm))
1402                 return DLM_MASTER_RESP_NO;
1403
1404         if (!dlm_domain_fully_joined(dlm)) {
1405                 response = DLM_MASTER_RESP_NO;
1406                 goto send_response;
1407         }
1408
1409         name = request->name;
1410         namelen = request->namelen;
1411         hash = dlm_lockid_hash(name, namelen);
1412
1413         if (namelen > DLM_LOCKID_NAME_MAX) {
1414                 response = DLM_IVBUFLEN;
1415                 goto send_response;
1416         }
1417
1418 way_up_top:
1419         spin_lock(&dlm->spinlock);
1420         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1421         if (res) {
1422                 spin_unlock(&dlm->spinlock);
1423
1424                 /* take care of the easy cases up front */
1425                 spin_lock(&res->spinlock);
1426                 if (res->state & (DLM_LOCK_RES_RECOVERING|
1427                                   DLM_LOCK_RES_MIGRATING)) {
1428                         spin_unlock(&res->spinlock);
1429                         mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1430                              "being recovered/migrated\n");
1431                         response = DLM_MASTER_RESP_ERROR;
1432                         if (mle)
1433                                 kmem_cache_free(dlm_mle_cache, mle);
1434                         goto send_response;
1435                 }
1436
1437                 if (res->owner == dlm->node_num) {
1438                         dlm_lockres_set_refmap_bit(dlm, res, request->node_idx);
1439                         spin_unlock(&res->spinlock);
1440                         response = DLM_MASTER_RESP_YES;
1441                         if (mle)
1442                                 kmem_cache_free(dlm_mle_cache, mle);
1443
1444                         /* this node is the owner.
1445                          * there is some extra work that needs to
1446                          * happen now.  the requesting node has
1447                          * caused all nodes up to this one to
1448                          * create mles.  this node now needs to
1449                          * go back and clean those up. */
1450                         dispatch_assert = 1;
1451                         goto send_response;
1452                 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1453                         spin_unlock(&res->spinlock);
1454                         // mlog(0, "node %u is the master\n", res->owner);
1455                         response = DLM_MASTER_RESP_NO;
1456                         if (mle)
1457                                 kmem_cache_free(dlm_mle_cache, mle);
1458                         goto send_response;
1459                 }
1460
1461                 /* ok, there is no owner.  either this node is
1462                  * being blocked, or it is actively trying to
1463                  * master this lock. */
1464                 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1465                         mlog(ML_ERROR, "lock with no owner should be "
1466                              "in-progress!\n");
1467                         BUG();
1468                 }
1469
1470                 // mlog(0, "lockres is in progress...\n");
1471                 spin_lock(&dlm->master_lock);
1472                 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1473                 if (!found) {
1474                         mlog(ML_ERROR, "no mle found for this lock!\n");
1475                         BUG();
1476                 }
1477                 set_maybe = 1;
1478                 spin_lock(&tmpmle->spinlock);
1479                 if (tmpmle->type == DLM_MLE_BLOCK) {
1480                         // mlog(0, "this node is waiting for "
1481                         // "lockres to be mastered\n");
1482                         response = DLM_MASTER_RESP_NO;
1483                 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1484                         mlog(0, "node %u is master, but trying to migrate to "
1485                              "node %u.\n", tmpmle->master, tmpmle->new_master);
1486                         if (tmpmle->master == dlm->node_num) {
1487                                 mlog(ML_ERROR, "no owner on lockres, but this "
1488                                      "node is trying to migrate it to %u?!\n",
1489                                      tmpmle->new_master);
1490                                 BUG();
1491                         } else {
1492                                 /* the real master can respond on its own */
1493                                 response = DLM_MASTER_RESP_NO;
1494                         }
1495                 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1496                         set_maybe = 0;
1497                         if (tmpmle->master == dlm->node_num) {
1498                                 response = DLM_MASTER_RESP_YES;
1499                                 /* this node will be the owner.
1500                                  * go back and clean the mles on any
1501                                  * other nodes */
1502                                 dispatch_assert = 1;
1503                                 dlm_lockres_set_refmap_bit(dlm, res,
1504                                                            request->node_idx);
1505                         } else
1506                                 response = DLM_MASTER_RESP_NO;
1507                 } else {
1508                         // mlog(0, "this node is attempting to "
1509                         // "master lockres\n");
1510                         response = DLM_MASTER_RESP_MAYBE;
1511                 }
1512                 if (set_maybe)
1513                         set_bit(request->node_idx, tmpmle->maybe_map);
1514                 spin_unlock(&tmpmle->spinlock);
1515
1516                 spin_unlock(&dlm->master_lock);
1517                 spin_unlock(&res->spinlock);
1518
1519                 /* keep the mle attached to heartbeat events */
1520                 dlm_put_mle(tmpmle);
1521                 if (mle)
1522                         kmem_cache_free(dlm_mle_cache, mle);
1523                 goto send_response;
1524         }
1525
1526         /*
1527          * lockres doesn't exist on this node
1528          * if there is an MLE_BLOCK, return NO
1529          * if there is an MLE_MASTER, return MAYBE
1530          * otherwise, add an MLE_BLOCK, return NO
1531          */
1532         spin_lock(&dlm->master_lock);
1533         found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1534         if (!found) {
1535                 /* this lockid has never been seen on this node yet */
1536                 // mlog(0, "no mle found\n");
1537                 if (!mle) {
1538                         spin_unlock(&dlm->master_lock);
1539                         spin_unlock(&dlm->spinlock);
1540
1541                         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1542                         if (!mle) {
1543                                 response = DLM_MASTER_RESP_ERROR;
1544                                 mlog_errno(-ENOMEM);
1545                                 goto send_response;
1546                         }
1547                         goto way_up_top;
1548                 }
1549
1550                 // mlog(0, "this is second time thru, already allocated, "
1551                 // "add the block.\n");
1552                 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1553                 set_bit(request->node_idx, mle->maybe_map);
1554                 __dlm_insert_mle(dlm, mle);
1555                 response = DLM_MASTER_RESP_NO;
1556         } else {
1557                 // mlog(0, "mle was found\n");
1558                 set_maybe = 1;
1559                 spin_lock(&tmpmle->spinlock);
1560                 if (tmpmle->master == dlm->node_num) {
1561                         mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1562                         BUG();
1563                 }
1564                 if (tmpmle->type == DLM_MLE_BLOCK)
1565                         response = DLM_MASTER_RESP_NO;
1566                 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1567                         mlog(0, "migration mle was found (%u->%u)\n",
1568                              tmpmle->master, tmpmle->new_master);
1569                         /* real master can respond on its own */
1570                         response = DLM_MASTER_RESP_NO;
1571                 } else
1572                         response = DLM_MASTER_RESP_MAYBE;
1573                 if (set_maybe)
1574                         set_bit(request->node_idx, tmpmle->maybe_map);
1575                 spin_unlock(&tmpmle->spinlock);
1576         }
1577         spin_unlock(&dlm->master_lock);
1578         spin_unlock(&dlm->spinlock);
1579
1580         if (found) {
1581                 /* keep the mle attached to heartbeat events */
1582                 dlm_put_mle(tmpmle);
1583         }
1584 send_response:
1585         /*
1586          * __dlm_lookup_lockres() grabbed a reference to this lockres.
1587          * The reference is released by dlm_assert_master_worker() under
1588          * the call to dlm_dispatch_assert_master().  If
1589          * dlm_assert_master_worker() isn't called, we drop it here.
1590          */
1591         if (dispatch_assert) {
1592                 if (response != DLM_MASTER_RESP_YES)
1593                         mlog(ML_ERROR, "invalid response %d\n", response);
1594                 if (!res) {
1595                         mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1596                         BUG();
1597                 }
1598                 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1599                              dlm->node_num, res->lockname.len, res->lockname.name);
1600                 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1601                                                  DLM_ASSERT_MASTER_MLE_CLEANUP);
1602                 if (ret < 0) {
1603                         mlog(ML_ERROR, "failed to dispatch assert master work\n");
1604                         response = DLM_MASTER_RESP_ERROR;
1605                         dlm_lockres_put(res);
1606                 }
1607         } else {
1608                 if (res)
1609                         dlm_lockres_put(res);
1610         }
1611
1612         dlm_put(dlm);
1613         return response;
1614 }
1615
1616 /*
1617  * DLM_ASSERT_MASTER_MSG
1618  */
1619
1620
1621 /*
1622  * NOTE: this can be used for debugging
1623  * can periodically run all locks owned by this node
1624  * and re-assert across the cluster...
1625  */
1626 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1627                                 struct dlm_lock_resource *res,
1628                                 void *nodemap, u32 flags)
1629 {
1630         struct dlm_assert_master assert;
1631         int to, tmpret;
1632         struct dlm_node_iter iter;
1633         int ret = 0;
1634         int reassert;
1635         const char *lockname = res->lockname.name;
1636         unsigned int namelen = res->lockname.len;
1637
1638         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1639
1640         spin_lock(&res->spinlock);
1641         res->state |= DLM_LOCK_RES_SETREF_INPROG;
1642         spin_unlock(&res->spinlock);
1643
1644 again:
1645         reassert = 0;
1646
1647         /* note that if this nodemap is empty, it returns 0 */
1648         dlm_node_iter_init(nodemap, &iter);
1649         while ((to = dlm_node_iter_next(&iter)) >= 0) {
1650                 int r = 0;
1651                 struct dlm_master_list_entry *mle = NULL;
1652
1653                 mlog(0, "sending assert master to %d (%.*s)\n", to,
1654                      namelen, lockname);
1655                 memset(&assert, 0, sizeof(assert));
1656                 assert.node_idx = dlm->node_num;
1657                 assert.namelen = namelen;
1658                 memcpy(assert.name, lockname, namelen);
1659                 assert.flags = cpu_to_be32(flags);
1660
1661                 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1662                                             &assert, sizeof(assert), to, &r);
1663                 if (tmpret < 0) {
1664                         mlog(ML_ERROR, "Error %d when sending message %u (key "
1665                              "0x%x) to node %u\n", tmpret,
1666                              DLM_ASSERT_MASTER_MSG, dlm->key, to);
1667                         if (!dlm_is_host_down(tmpret)) {
1668                                 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1669                                 BUG();
1670                         }
1671                         /* a node died.  finish out the rest of the nodes. */
1672                         mlog(0, "link to %d went down!\n", to);
1673                         /* any nonzero status return will do */
1674                         ret = tmpret;
1675                         r = 0;
1676                 } else if (r < 0) {
1677                         /* ok, something horribly messed.  kill thyself. */
1678                         mlog(ML_ERROR,"during assert master of %.*s to %u, "
1679                              "got %d.\n", namelen, lockname, to, r);
1680                         spin_lock(&dlm->spinlock);
1681                         spin_lock(&dlm->master_lock);
1682                         if (dlm_find_mle(dlm, &mle, (char *)lockname,
1683                                          namelen)) {
1684                                 dlm_print_one_mle(mle);
1685                                 __dlm_put_mle(mle);
1686                         }
1687                         spin_unlock(&dlm->master_lock);
1688                         spin_unlock(&dlm->spinlock);
1689                         BUG();
1690                 }
1691
1692                 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1693                     !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1694                                 mlog(ML_ERROR, "%.*s: very strange, "
1695                                      "master MLE but no lockres on %u\n",
1696                                      namelen, lockname, to);
1697                 }
1698
1699                 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1700                         mlog(0, "%.*s: node %u create mles on other "
1701                              "nodes and requests a re-assert\n",
1702                              namelen, lockname, to);
1703                         reassert = 1;
1704                 }
1705                 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1706                         mlog(0, "%.*s: node %u has a reference to this "
1707                              "lockres, set the bit in the refmap\n",
1708                              namelen, lockname, to);
1709                         spin_lock(&res->spinlock);
1710                         dlm_lockres_set_refmap_bit(dlm, res, to);
1711                         spin_unlock(&res->spinlock);
1712                 }
1713         }
1714
1715         if (reassert)
1716                 goto again;
1717
1718         spin_lock(&res->spinlock);
1719         res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1720         spin_unlock(&res->spinlock);
1721         wake_up(&res->wq);
1722
1723         return ret;
1724 }
1725
1726 /*
1727  * locks that can be taken here:
1728  * dlm->spinlock
1729  * res->spinlock
1730  * mle->spinlock
1731  * dlm->master_list
1732  *
1733  * if possible, TRIM THIS DOWN!!!
1734  */
1735 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1736                               void **ret_data)
1737 {
1738         struct dlm_ctxt *dlm = data;
1739         struct dlm_master_list_entry *mle = NULL;
1740         struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1741         struct dlm_lock_resource *res = NULL;
1742         char *name;
1743         unsigned int namelen, hash;
1744         u32 flags;
1745         int master_request = 0, have_lockres_ref = 0;
1746         int ret = 0;
1747
1748         if (!dlm_grab(dlm))
1749                 return 0;
1750
1751         name = assert->name;
1752         namelen = assert->namelen;
1753         hash = dlm_lockid_hash(name, namelen);
1754         flags = be32_to_cpu(assert->flags);
1755
1756         if (namelen > DLM_LOCKID_NAME_MAX) {
1757                 mlog(ML_ERROR, "Invalid name length!");
1758                 goto done;
1759         }
1760
1761         spin_lock(&dlm->spinlock);
1762
1763         if (flags)
1764                 mlog(0, "assert_master with flags: %u\n", flags);
1765
1766         /* find the MLE */
1767         spin_lock(&dlm->master_lock);
1768         if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1769                 /* not an error, could be master just re-asserting */
1770                 mlog(0, "just got an assert_master from %u, but no "
1771                      "MLE for it! (%.*s)\n", assert->node_idx,
1772                      namelen, name);
1773         } else {
1774                 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1775                 if (bit >= O2NM_MAX_NODES) {
1776                         /* not necessarily an error, though less likely.
1777                          * could be master just re-asserting. */
1778                         mlog(0, "no bits set in the maybe_map, but %u "
1779                              "is asserting! (%.*s)\n", assert->node_idx,
1780                              namelen, name);
1781                 } else if (bit != assert->node_idx) {
1782                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1783                                 mlog(0, "master %u was found, %u should "
1784                                      "back off\n", assert->node_idx, bit);
1785                         } else {
1786                                 /* with the fix for bug 569, a higher node
1787                                  * number winning the mastery will respond
1788                                  * YES to mastery requests, but this node
1789                                  * had no way of knowing.  let it pass. */
1790                                 mlog(0, "%u is the lowest node, "
1791                                      "%u is asserting. (%.*s)  %u must "
1792                                      "have begun after %u won.\n", bit,
1793                                      assert->node_idx, namelen, name, bit,
1794                                      assert->node_idx);
1795                         }
1796                 }
1797                 if (mle->type == DLM_MLE_MIGRATION) {
1798                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1799                                 mlog(0, "%s:%.*s: got cleanup assert"
1800                                      " from %u for migration\n",
1801                                      dlm->name, namelen, name,
1802                                      assert->node_idx);
1803                         } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1804                                 mlog(0, "%s:%.*s: got unrelated assert"
1805                                      " from %u for migration, ignoring\n",
1806                                      dlm->name, namelen, name,
1807                                      assert->node_idx);
1808                                 __dlm_put_mle(mle);
1809                                 spin_unlock(&dlm->master_lock);
1810                                 spin_unlock(&dlm->spinlock);
1811                                 goto done;
1812                         }
1813                 }
1814         }
1815         spin_unlock(&dlm->master_lock);
1816
1817         /* ok everything checks out with the MLE
1818          * now check to see if there is a lockres */
1819         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1820         if (res) {
1821                 spin_lock(&res->spinlock);
1822                 if (res->state & DLM_LOCK_RES_RECOVERING)  {
1823                         mlog(ML_ERROR, "%u asserting but %.*s is "
1824                              "RECOVERING!\n", assert->node_idx, namelen, name);
1825                         goto kill;
1826                 }
1827                 if (!mle) {
1828                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1829                             res->owner != assert->node_idx) {
1830                                 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1831                                      "but current owner is %u! (%.*s)\n",
1832                                      assert->node_idx, res->owner, namelen,
1833                                      name);
1834                                 __dlm_print_one_lock_resource(res);
1835                                 BUG();
1836                         }
1837                 } else if (mle->type != DLM_MLE_MIGRATION) {
1838                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1839                                 /* owner is just re-asserting */
1840                                 if (res->owner == assert->node_idx) {
1841                                         mlog(0, "owner %u re-asserting on "
1842                                              "lock %.*s\n", assert->node_idx,
1843                                              namelen, name);
1844                                         goto ok;
1845                                 }
1846                                 mlog(ML_ERROR, "got assert_master from "
1847                                      "node %u, but %u is the owner! "
1848                                      "(%.*s)\n", assert->node_idx,
1849                                      res->owner, namelen, name);
1850                                 goto kill;
1851                         }
1852                         if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1853                                 mlog(ML_ERROR, "got assert from %u, but lock "
1854                                      "with no owner should be "
1855                                      "in-progress! (%.*s)\n",
1856                                      assert->node_idx,
1857                                      namelen, name);
1858                                 goto kill;
1859                         }
1860                 } else /* mle->type == DLM_MLE_MIGRATION */ {
1861                         /* should only be getting an assert from new master */
1862                         if (assert->node_idx != mle->new_master) {
1863                                 mlog(ML_ERROR, "got assert from %u, but "
1864                                      "new master is %u, and old master "
1865                                      "was %u (%.*s)\n",
1866                                      assert->node_idx, mle->new_master,
1867                                      mle->master, namelen, name);
1868                                 goto kill;
1869                         }
1870
1871                 }
1872 ok:
1873                 spin_unlock(&res->spinlock);
1874         }
1875
1876         // mlog(0, "woo!  got an assert_master from node %u!\n",
1877         //           assert->node_idx);
1878         if (mle) {
1879                 int extra_ref = 0;
1880                 int nn = -1;
1881                 int rr, err = 0;
1882
1883                 spin_lock(&mle->spinlock);
1884                 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1885                         extra_ref = 1;
1886                 else {
1887                         /* MASTER mle: if any bits set in the response map
1888                          * then the calling node needs to re-assert to clear
1889                          * up nodes that this node contacted */
1890                         while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1891                                                     nn+1)) < O2NM_MAX_NODES) {
1892                                 if (nn != dlm->node_num && nn != assert->node_idx) {
1893                                         master_request = 1;
1894                                         break;
1895                                 }
1896                         }
1897                 }
1898                 mle->master = assert->node_idx;
1899                 atomic_set(&mle->woken, 1);
1900                 wake_up(&mle->wq);
1901                 spin_unlock(&mle->spinlock);
1902
1903                 if (res) {
1904                         int wake = 0;
1905                         spin_lock(&res->spinlock);
1906                         if (mle->type == DLM_MLE_MIGRATION) {
1907                                 mlog(0, "finishing off migration of lockres %.*s, "
1908                                         "from %u to %u\n",
1909                                         res->lockname.len, res->lockname.name,
1910                                         dlm->node_num, mle->new_master);
1911                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
1912                                 wake = 1;
1913                                 dlm_change_lockres_owner(dlm, res, mle->new_master);
1914                                 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1915                         } else {
1916                                 dlm_change_lockres_owner(dlm, res, mle->master);
1917                         }
1918                         spin_unlock(&res->spinlock);
1919                         have_lockres_ref = 1;
1920                         if (wake)
1921                                 wake_up(&res->wq);
1922                 }
1923
1924                 /* master is known, detach if not already detached.
1925                  * ensures that only one assert_master call will happen
1926                  * on this mle. */
1927                 spin_lock(&dlm->master_lock);
1928
1929                 rr = atomic_read(&mle->mle_refs.refcount);
1930                 if (mle->inuse > 0) {
1931                         if (extra_ref && rr < 3)
1932                                 err = 1;
1933                         else if (!extra_ref && rr < 2)
1934                                 err = 1;
1935                 } else {
1936                         if (extra_ref && rr < 2)
1937                                 err = 1;
1938                         else if (!extra_ref && rr < 1)
1939                                 err = 1;
1940                 }
1941                 if (err) {
1942                         mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
1943                              "that will mess up this node, refs=%d, extra=%d, "
1944                              "inuse=%d\n", dlm->name, namelen, name,
1945                              assert->node_idx, rr, extra_ref, mle->inuse);
1946                         dlm_print_one_mle(mle);
1947                 }
1948                 __dlm_unlink_mle(dlm, mle);
1949                 __dlm_mle_detach_hb_events(dlm, mle);
1950                 __dlm_put_mle(mle);
1951                 if (extra_ref) {
1952                         /* the assert master message now balances the extra
1953                          * ref given by the master / migration request message.
1954                          * if this is the last put, it will be removed
1955                          * from the list. */
1956                         __dlm_put_mle(mle);
1957                 }
1958                 spin_unlock(&dlm->master_lock);
1959         } else if (res) {
1960                 if (res->owner != assert->node_idx) {
1961                         mlog(0, "assert_master from %u, but current "
1962                              "owner is %u (%.*s), no mle\n", assert->node_idx,
1963                              res->owner, namelen, name);
1964                 }
1965         }
1966         spin_unlock(&dlm->spinlock);
1967
1968 done:
1969         ret = 0;
1970         if (res) {
1971                 spin_lock(&res->spinlock);
1972                 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1973                 spin_unlock(&res->spinlock);
1974                 *ret_data = (void *)res;
1975         }
1976         dlm_put(dlm);
1977         if (master_request) {
1978                 mlog(0, "need to tell master to reassert\n");
1979                 /* positive. negative would shoot down the node. */
1980                 ret |= DLM_ASSERT_RESPONSE_REASSERT;
1981                 if (!have_lockres_ref) {
1982                         mlog(ML_ERROR, "strange, got assert from %u, MASTER "
1983                              "mle present here for %s:%.*s, but no lockres!\n",
1984                              assert->node_idx, dlm->name, namelen, name);
1985                 }
1986         }
1987         if (have_lockres_ref) {
1988                 /* let the master know we have a reference to the lockres */
1989                 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
1990                 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
1991                      dlm->name, namelen, name, assert->node_idx);
1992         }
1993         return ret;
1994
1995 kill:
1996         /* kill the caller! */
1997         mlog(ML_ERROR, "Bad message received from another node.  Dumping state "
1998              "and killing the other node now!  This node is OK and can continue.\n");
1999         __dlm_print_one_lock_resource(res);
2000         spin_unlock(&res->spinlock);
2001         spin_unlock(&dlm->spinlock);
2002         *ret_data = (void *)res;
2003         dlm_put(dlm);
2004         return -EINVAL;
2005 }
2006
2007 void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2008 {
2009         struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2010
2011         if (ret_data) {
2012                 spin_lock(&res->spinlock);
2013                 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2014                 spin_unlock(&res->spinlock);
2015                 wake_up(&res->wq);
2016                 dlm_lockres_put(res);
2017         }
2018         return;
2019 }
2020
2021 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2022                                struct dlm_lock_resource *res,
2023                                int ignore_higher, u8 request_from, u32 flags)
2024 {
2025         struct dlm_work_item *item;
2026         item = kzalloc(sizeof(*item), GFP_ATOMIC);
2027         if (!item)
2028                 return -ENOMEM;
2029
2030
2031         /* queue up work for dlm_assert_master_worker */
2032         dlm_grab(dlm);  /* get an extra ref for the work item */
2033         dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2034         item->u.am.lockres = res; /* already have a ref */
2035         /* can optionally ignore node numbers higher than this node */
2036         item->u.am.ignore_higher = ignore_higher;
2037         item->u.am.request_from = request_from;
2038         item->u.am.flags = flags;
2039
2040         if (ignore_higher)
2041                 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2042                      res->lockname.name);
2043
2044         spin_lock(&dlm->work_lock);
2045         list_add_tail(&item->list, &dlm->work_list);
2046         spin_unlock(&dlm->work_lock);
2047
2048         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2049         return 0;
2050 }
2051
2052 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2053 {
2054         struct dlm_ctxt *dlm = data;
2055         int ret = 0;
2056         struct dlm_lock_resource *res;
2057         unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2058         int ignore_higher;
2059         int bit;
2060         u8 request_from;
2061         u32 flags;
2062
2063         dlm = item->dlm;
2064         res = item->u.am.lockres;
2065         ignore_higher = item->u.am.ignore_higher;
2066         request_from = item->u.am.request_from;
2067         flags = item->u.am.flags;
2068
2069         spin_lock(&dlm->spinlock);
2070         memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2071         spin_unlock(&dlm->spinlock);
2072
2073         clear_bit(dlm->node_num, nodemap);
2074         if (ignore_higher) {
2075                 /* if is this just to clear up mles for nodes below
2076                  * this node, do not send the message to the original
2077                  * caller or any node number higher than this */
2078                 clear_bit(request_from, nodemap);
2079                 bit = dlm->node_num;
2080                 while (1) {
2081                         bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2082                                             bit+1);
2083                         if (bit >= O2NM_MAX_NODES)
2084                                 break;
2085                         clear_bit(bit, nodemap);
2086                 }
2087         }
2088
2089         /*
2090          * If we're migrating this lock to someone else, we are no
2091          * longer allowed to assert out own mastery.  OTOH, we need to
2092          * prevent migration from starting while we're still asserting
2093          * our dominance.  The reserved ast delays migration.
2094          */
2095         spin_lock(&res->spinlock);
2096         if (res->state & DLM_LOCK_RES_MIGRATING) {
2097                 mlog(0, "Someone asked us to assert mastery, but we're "
2098                      "in the middle of migration.  Skipping assert, "
2099                      "the new master will handle that.\n");
2100                 spin_unlock(&res->spinlock);
2101                 goto put;
2102         } else
2103                 __dlm_lockres_reserve_ast(res);
2104         spin_unlock(&res->spinlock);
2105
2106         /* this call now finishes out the nodemap
2107          * even if one or more nodes die */
2108         mlog(0, "worker about to master %.*s here, this=%u\n",
2109                      res->lockname.len, res->lockname.name, dlm->node_num);
2110         ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2111         if (ret < 0) {
2112                 /* no need to restart, we are done */
2113                 if (!dlm_is_host_down(ret))
2114                         mlog_errno(ret);
2115         }
2116
2117         /* Ok, we've asserted ourselves.  Let's let migration start. */
2118         dlm_lockres_release_ast(dlm, res);
2119
2120 put:
2121         dlm_lockres_put(res);
2122
2123         mlog(0, "finished with dlm_assert_master_worker\n");
2124 }
2125
2126 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2127  * We cannot wait for node recovery to complete to begin mastering this
2128  * lockres because this lockres is used to kick off recovery! ;-)
2129  * So, do a pre-check on all living nodes to see if any of those nodes
2130  * think that $RECOVERY is currently mastered by a dead node.  If so,
2131  * we wait a short time to allow that node to get notified by its own
2132  * heartbeat stack, then check again.  All $RECOVERY lock resources
2133  * mastered by dead nodes are purged when the hearbeat callback is
2134  * fired, so we can know for sure that it is safe to continue once
2135  * the node returns a live node or no node.  */
2136 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2137                                        struct dlm_lock_resource *res)
2138 {
2139         struct dlm_node_iter iter;
2140         int nodenum;
2141         int ret = 0;
2142         u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2143
2144         spin_lock(&dlm->spinlock);
2145         dlm_node_iter_init(dlm->domain_map, &iter);
2146         spin_unlock(&dlm->spinlock);
2147
2148         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2149                 /* do not send to self */
2150                 if (nodenum == dlm->node_num)
2151                         continue;
2152                 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2153                 if (ret < 0) {
2154                         mlog_errno(ret);
2155                         if (!dlm_is_host_down(ret))
2156                                 BUG();
2157                         /* host is down, so answer for that node would be
2158                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
2159                         ret = 0;
2160                 }
2161
2162                 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2163                         /* check to see if this master is in the recovery map */
2164                         spin_lock(&dlm->spinlock);
2165                         if (test_bit(master, dlm->recovery_map)) {
2166                                 mlog(ML_NOTICE, "%s: node %u has not seen "
2167                                      "node %u go down yet, and thinks the "
2168                                      "dead node is mastering the recovery "
2169                                      "lock.  must wait.\n", dlm->name,
2170                                      nodenum, master);
2171                                 ret = -EAGAIN;
2172                         }
2173                         spin_unlock(&dlm->spinlock);
2174                         mlog(0, "%s: reco lock master is %u\n", dlm->name,
2175                              master);
2176                         break;
2177                 }
2178         }
2179         return ret;
2180 }
2181
2182 /*
2183  * DLM_DEREF_LOCKRES_MSG
2184  */
2185
2186 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2187 {
2188         struct dlm_deref_lockres deref;
2189         int ret = 0, r;
2190         const char *lockname;
2191         unsigned int namelen;
2192
2193         lockname = res->lockname.name;
2194         namelen = res->lockname.len;
2195         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2196
2197         memset(&deref, 0, sizeof(deref));
2198         deref.node_idx = dlm->node_num;
2199         deref.namelen = namelen;
2200         memcpy(deref.name, lockname, namelen);
2201
2202         ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2203                                  &deref, sizeof(deref), res->owner, &r);
2204         if (ret < 0)
2205                 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n",
2206                      dlm->name, namelen, lockname, ret, res->owner);
2207         else if (r < 0) {
2208                 /* BAD.  other node says I did not have a ref. */
2209                 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2210                      dlm->name, namelen, lockname, res->owner, r);
2211                 dlm_print_one_lock_resource(res);
2212                 BUG();
2213         }
2214         return ret;
2215 }
2216
2217 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2218                               void **ret_data)
2219 {
2220         struct dlm_ctxt *dlm = data;
2221         struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2222         struct dlm_lock_resource *res = NULL;
2223         char *name;
2224         unsigned int namelen;
2225         int ret = -EINVAL;
2226         u8 node;
2227         unsigned int hash;
2228         struct dlm_work_item *item;
2229         int cleared = 0;
2230         int dispatch = 0;
2231
2232         if (!dlm_grab(dlm))
2233                 return 0;
2234
2235         name = deref->name;
2236         namelen = deref->namelen;
2237         node = deref->node_idx;
2238
2239         if (namelen > DLM_LOCKID_NAME_MAX) {
2240                 mlog(ML_ERROR, "Invalid name length!");
2241                 goto done;
2242         }
2243         if (deref->node_idx >= O2NM_MAX_NODES) {
2244                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2245                 goto done;
2246         }
2247
2248         hash = dlm_lockid_hash(name, namelen);
2249
2250         spin_lock(&dlm->spinlock);
2251         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2252         if (!res) {
2253                 spin_unlock(&dlm->spinlock);
2254                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2255                      dlm->name, namelen, name);
2256                 goto done;
2257         }
2258         spin_unlock(&dlm->spinlock);
2259
2260         spin_lock(&res->spinlock);
2261         if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2262                 dispatch = 1;
2263         else {
2264                 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2265                 if (test_bit(node, res->refmap)) {
2266                         dlm_lockres_clear_refmap_bit(dlm, res, node);
2267                         cleared = 1;
2268                 }
2269         }
2270         spin_unlock(&res->spinlock);
2271
2272         if (!dispatch) {
2273                 if (cleared)
2274                         dlm_lockres_calc_usage(dlm, res);
2275                 else {
2276                         mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2277                         "but it is already dropped!\n", dlm->name,
2278                         res->lockname.len, res->lockname.name, node);
2279                         dlm_print_one_lock_resource(res);
2280                 }
2281                 ret = 0;
2282                 goto done;
2283         }
2284
2285         item = kzalloc(sizeof(*item), GFP_NOFS);
2286         if (!item) {
2287                 ret = -ENOMEM;
2288                 mlog_errno(ret);
2289                 goto done;
2290         }
2291
2292         dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2293         item->u.dl.deref_res = res;
2294         item->u.dl.deref_node = node;
2295
2296         spin_lock(&dlm->work_lock);
2297         list_add_tail(&item->list, &dlm->work_list);
2298         spin_unlock(&dlm->work_lock);
2299
2300         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2301         return 0;
2302
2303 done:
2304         if (res)
2305                 dlm_lockres_put(res);
2306         dlm_put(dlm);
2307
2308         return ret;
2309 }
2310
2311 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2312 {
2313         struct dlm_ctxt *dlm;
2314         struct dlm_lock_resource *res;
2315         u8 node;
2316         u8 cleared = 0;
2317
2318         dlm = item->dlm;
2319         res = item->u.dl.deref_res;
2320         node = item->u.dl.deref_node;
2321
2322         spin_lock(&res->spinlock);
2323         BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2324         if (test_bit(node, res->refmap)) {
2325                 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2326                 dlm_lockres_clear_refmap_bit(dlm, res, node);
2327                 cleared = 1;
2328         }
2329         spin_unlock(&res->spinlock);
2330
2331         if (cleared) {
2332                 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2333                      dlm->name, res->lockname.len, res->lockname.name, node);
2334                 dlm_lockres_calc_usage(dlm, res);
2335         } else {
2336                 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2337                      "but it is already dropped!\n", dlm->name,
2338                      res->lockname.len, res->lockname.name, node);
2339                 dlm_print_one_lock_resource(res);
2340         }
2341
2342         dlm_lockres_put(res);
2343 }
2344
2345 /*
2346  * A migrateable resource is one that is :
2347  * 1. locally mastered, and,
2348  * 2. zero local locks, and,
2349  * 3. one or more non-local locks, or, one or more references
2350  * Returns 1 if yes, 0 if not.
2351  */
2352 static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
2353                                       struct dlm_lock_resource *res)
2354 {
2355         enum dlm_lockres_list idx;
2356         int nonlocal = 0, node_ref;
2357         struct list_head *queue;
2358         struct dlm_lock *lock;
2359         u64 cookie;
2360
2361         assert_spin_locked(&res->spinlock);
2362
2363         /* delay migration when the lockres is in MIGRATING state */
2364         if (res->state & DLM_LOCK_RES_MIGRATING)
2365                 return 0;
2366
2367         if (res->owner != dlm->node_num)
2368                 return 0;
2369
2370         for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2371                 queue = dlm_list_idx_to_ptr(res, idx);
2372                 list_for_each_entry(lock, queue, list) {
2373                         if (lock->ml.node != dlm->node_num) {
2374                                 nonlocal++;
2375                                 continue;
2376                         }
2377                         cookie = be64_to_cpu(lock->ml.cookie);
2378                         mlog(0, "%s: Not migrateable res %.*s, lock %u:%llu on "
2379                              "%s list\n", dlm->name, res->lockname.len,
2380                              res->lockname.name,
2381                              dlm_get_lock_cookie_node(cookie),
2382                              dlm_get_lock_cookie_seq(cookie),
2383                              dlm_list_in_text(idx));
2384                         return 0;
2385                 }
2386         }
2387
2388         if (!nonlocal) {
2389                 node_ref = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
2390                 if (node_ref >= O2NM_MAX_NODES)
2391                         return 0;
2392         }
2393
2394         mlog(0, "%s: res %.*s, Migrateable\n", dlm->name, res->lockname.len,
2395              res->lockname.name);
2396
2397         return 1;
2398 }
2399
2400 /*
2401  * DLM_MIGRATE_LOCKRES
2402  */
2403
2404
2405 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2406                                struct dlm_lock_resource *res, u8 target)
2407 {
2408         struct dlm_master_list_entry *mle = NULL;
2409         struct dlm_master_list_entry *oldmle = NULL;
2410         struct dlm_migratable_lockres *mres = NULL;
2411         int ret = 0;
2412         const char *name;
2413         unsigned int namelen;
2414         int mle_added = 0;
2415         int wake = 0;
2416
2417         if (!dlm_grab(dlm))
2418                 return -EINVAL;
2419
2420         BUG_ON(target == O2NM_MAX_NODES);
2421
2422         name = res->lockname.name;
2423         namelen = res->lockname.len;
2424
2425         mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name,
2426              target);
2427
2428         /* preallocate up front. if this fails, abort */
2429         ret = -ENOMEM;
2430         mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2431         if (!mres) {
2432                 mlog_errno(ret);
2433                 goto leave;
2434         }
2435
2436         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2437         if (!mle) {
2438                 mlog_errno(ret);
2439                 goto leave;
2440         }
2441         ret = 0;
2442
2443         /*
2444          * clear any existing master requests and
2445          * add the migration mle to the list
2446          */
2447         spin_lock(&dlm->spinlock);
2448         spin_lock(&dlm->master_lock);
2449         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2450                                     namelen, target, dlm->node_num);
2451         spin_unlock(&dlm->master_lock);
2452         spin_unlock(&dlm->spinlock);
2453
2454         if (ret == -EEXIST) {
2455                 mlog(0, "another process is already migrating it\n");
2456                 goto fail;
2457         }
2458         mle_added = 1;
2459
2460         /*
2461          * set the MIGRATING flag and flush asts
2462          * if we fail after this we need to re-dirty the lockres
2463          */
2464         if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2465                 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2466                      "the target went down.\n", res->lockname.len,
2467                      res->lockname.name, target);
2468                 spin_lock(&res->spinlock);
2469                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2470                 wake = 1;
2471                 spin_unlock(&res->spinlock);
2472                 ret = -EINVAL;
2473         }
2474
2475 fail:
2476         if (oldmle) {
2477                 /* master is known, detach if not already detached */
2478                 dlm_mle_detach_hb_events(dlm, oldmle);
2479                 dlm_put_mle(oldmle);
2480         }
2481
2482         if (ret < 0) {
2483                 if (mle_added) {
2484                         dlm_mle_detach_hb_events(dlm, mle);
2485                         dlm_put_mle(mle);
2486                 } else if (mle) {
2487                         kmem_cache_free(dlm_mle_cache, mle);
2488                         mle = NULL;
2489                 }
2490                 goto leave;
2491         }
2492
2493         /*
2494          * at this point, we have a migration target, an mle
2495          * in the master list, and the MIGRATING flag set on
2496          * the lockres
2497          */
2498
2499         /* now that remote nodes are spinning on the MIGRATING flag,
2500          * ensure that all assert_master work is flushed. */
2501         flush_workqueue(dlm->dlm_worker);
2502
2503         /* get an extra reference on the mle.
2504          * otherwise the assert_master from the new
2505          * master will destroy this.
2506          * also, make sure that all callers of dlm_get_mle
2507          * take both dlm->spinlock and dlm->master_lock */
2508         spin_lock(&dlm->spinlock);
2509         spin_lock(&dlm->master_lock);
2510         dlm_get_mle_inuse(mle);
2511         spin_unlock(&dlm->master_lock);
2512         spin_unlock(&dlm->spinlock);
2513
2514         /* notify new node and send all lock state */
2515         /* call send_one_lockres with migration flag.
2516          * this serves as notice to the target node that a
2517          * migration is starting. */
2518         ret = dlm_send_one_lockres(dlm, res, mres, target,
2519                                    DLM_MRES_MIGRATION);
2520
2521         if (ret < 0) {
2522                 mlog(0, "migration to node %u failed with %d\n",
2523                      target, ret);
2524                 /* migration failed, detach and clean up mle */
2525                 dlm_mle_detach_hb_events(dlm, mle);
2526                 dlm_put_mle(mle);
2527                 dlm_put_mle_inuse(mle);
2528                 spin_lock(&res->spinlock);
2529                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2530                 wake = 1;
2531                 spin_unlock(&res->spinlock);
2532                 if (dlm_is_host_down(ret))
2533                         dlm_wait_for_node_death(dlm, target,
2534                                                 DLM_NODE_DEATH_WAIT_MAX);
2535                 goto leave;
2536         }
2537
2538         /* at this point, the target sends a message to all nodes,
2539          * (using dlm_do_migrate_request).  this node is skipped since
2540          * we had to put an mle in the list to begin the process.  this
2541          * node now waits for target to do an assert master.  this node
2542          * will be the last one notified, ensuring that the migration
2543          * is complete everywhere.  if the target dies while this is
2544          * going on, some nodes could potentially see the target as the
2545          * master, so it is important that my recovery finds the migration
2546          * mle and sets the master to UNKNOWN. */
2547
2548
2549         /* wait for new node to assert master */
2550         while (1) {
2551                 ret = wait_event_interruptible_timeout(mle->wq,
2552                                         (atomic_read(&mle->woken) == 1),
2553                                         msecs_to_jiffies(5000));
2554
2555                 if (ret >= 0) {
2556                         if (atomic_read(&mle->woken) == 1 ||
2557                             res->owner == target)
2558                                 break;
2559
2560                         mlog(0, "%s:%.*s: timed out during migration\n",
2561                              dlm->name, res->lockname.len, res->lockname.name);
2562                         /* avoid hang during shutdown when migrating lockres
2563                          * to a node which also goes down */
2564                         if (dlm_is_node_dead(dlm, target)) {
2565                                 mlog(0, "%s:%.*s: expected migration "
2566                                      "target %u is no longer up, restarting\n",
2567                                      dlm->name, res->lockname.len,
2568                                      res->lockname.name, target);
2569                                 ret = -EINVAL;
2570                                 /* migration failed, detach and clean up mle */
2571                                 dlm_mle_detach_hb_events(dlm, mle);
2572                                 dlm_put_mle(mle);
2573                                 dlm_put_mle_inuse(mle);
2574                                 spin_lock(&res->spinlock);
2575                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2576                                 wake = 1;
2577                                 spin_unlock(&res->spinlock);
2578                                 goto leave;
2579                         }
2580                 } else
2581                         mlog(0, "%s:%.*s: caught signal during migration\n",
2582                              dlm->name, res->lockname.len, res->lockname.name);
2583         }
2584
2585         /* all done, set the owner, clear the flag */
2586         spin_lock(&res->spinlock);
2587         dlm_set_lockres_owner(dlm, res, target);
2588         res->state &= ~DLM_LOCK_RES_MIGRATING;
2589         dlm_remove_nonlocal_locks(dlm, res);
2590         spin_unlock(&res->spinlock);
2591         wake_up(&res->wq);
2592
2593         /* master is known, detach if not already detached */
2594         dlm_mle_detach_hb_events(dlm, mle);
2595         dlm_put_mle_inuse(mle);
2596         ret = 0;
2597
2598         dlm_lockres_calc_usage(dlm, res);
2599
2600 leave:
2601         /* re-dirty the lockres if we failed */
2602         if (ret < 0)
2603                 dlm_kick_thread(dlm, res);
2604
2605         /* wake up waiters if the MIGRATING flag got set
2606          * but migration failed */
2607         if (wake)
2608                 wake_up(&res->wq);
2609
2610         if (mres)
2611                 free_page((unsigned long)mres);
2612
2613         dlm_put(dlm);
2614
2615         mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen,
2616              name, target, ret);
2617         return ret;
2618 }
2619
2620 #define DLM_MIGRATION_RETRY_MS  100
2621
2622 /*
2623  * Should be called only after beginning the domain leave process.
2624  * There should not be any remaining locks on nonlocal lock resources,
2625  * and there should be no local locks left on locally mastered resources.
2626  *
2627  * Called with the dlm spinlock held, may drop it to do migration, but
2628  * will re-acquire before exit.
2629  *
2630  * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2631  */
2632 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2633 {
2634         int ret;
2635         int lock_dropped = 0;
2636         u8 target = O2NM_MAX_NODES;
2637
2638         assert_spin_locked(&dlm->spinlock);
2639
2640         spin_lock(&res->spinlock);
2641         if (dlm_is_lockres_migrateable(dlm, res))
2642                 target = dlm_pick_migration_target(dlm, res);
2643         spin_unlock(&res->spinlock);
2644
2645         if (target == O2NM_MAX_NODES)
2646                 goto leave;
2647
2648         /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2649         spin_unlock(&dlm->spinlock);
2650         lock_dropped = 1;
2651         ret = dlm_migrate_lockres(dlm, res, target);
2652         if (ret)
2653                 mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2654                      dlm->name, res->lockname.len, res->lockname.name,
2655                      target, ret);
2656         spin_lock(&dlm->spinlock);
2657 leave:
2658         return lock_dropped;
2659 }
2660
2661 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2662 {
2663         int ret;
2664         spin_lock(&dlm->ast_lock);
2665         spin_lock(&lock->spinlock);
2666         ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2667         spin_unlock(&lock->spinlock);
2668         spin_unlock(&dlm->ast_lock);
2669         return ret;
2670 }
2671
2672 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2673                                      struct dlm_lock_resource *res,
2674                                      u8 mig_target)
2675 {
2676         int can_proceed;
2677         spin_lock(&res->spinlock);
2678         can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2679         spin_unlock(&res->spinlock);
2680
2681         /* target has died, so make the caller break out of the
2682          * wait_event, but caller must recheck the domain_map */
2683         spin_lock(&dlm->spinlock);
2684         if (!test_bit(mig_target, dlm->domain_map))
2685                 can_proceed = 1;
2686         spin_unlock(&dlm->spinlock);
2687         return can_proceed;
2688 }
2689
2690 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2691                                 struct dlm_lock_resource *res)
2692 {
2693         int ret;
2694         spin_lock(&res->spinlock);
2695         ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2696         spin_unlock(&res->spinlock);
2697         return ret;
2698 }
2699
2700
2701 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2702                                        struct dlm_lock_resource *res,
2703                                        u8 target)
2704 {
2705         int ret = 0;
2706
2707         mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2708                res->lockname.len, res->lockname.name, dlm->node_num,
2709                target);
2710         /* need to set MIGRATING flag on lockres.  this is done by
2711          * ensuring that all asts have been flushed for this lockres. */
2712         spin_lock(&res->spinlock);
2713         BUG_ON(res->migration_pending);
2714         res->migration_pending = 1;
2715         /* strategy is to reserve an extra ast then release
2716          * it below, letting the release do all of the work */
2717         __dlm_lockres_reserve_ast(res);
2718         spin_unlock(&res->spinlock);
2719
2720         /* now flush all the pending asts */
2721         dlm_kick_thread(dlm, res);
2722         /* before waiting on DIRTY, block processes which may
2723          * try to dirty the lockres before MIGRATING is set */
2724         spin_lock(&res->spinlock);
2725         BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2726         res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2727         spin_unlock(&res->spinlock);
2728         /* now wait on any pending asts and the DIRTY state */
2729         wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2730         dlm_lockres_release_ast(dlm, res);
2731
2732         mlog(0, "about to wait on migration_wq, dirty=%s\n",
2733                res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2734         /* if the extra ref we just put was the final one, this
2735          * will pass thru immediately.  otherwise, we need to wait
2736          * for the last ast to finish. */
2737 again:
2738         ret = wait_event_interruptible_timeout(dlm->migration_wq,
2739                    dlm_migration_can_proceed(dlm, res, target),
2740                    msecs_to_jiffies(1000));
2741         if (ret < 0) {
2742                 mlog(0, "woken again: migrating? %s, dead? %s\n",
2743                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2744                        test_bit(target, dlm->domain_map) ? "no":"yes");
2745         } else {
2746                 mlog(0, "all is well: migrating? %s, dead? %s\n",
2747                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2748                        test_bit(target, dlm->domain_map) ? "no":"yes");
2749         }
2750         if (!dlm_migration_can_proceed(dlm, res, target)) {
2751                 mlog(0, "trying again...\n");
2752                 goto again;
2753         }
2754
2755         ret = 0;
2756         /* did the target go down or die? */
2757         spin_lock(&dlm->spinlock);
2758         if (!test_bit(target, dlm->domain_map)) {
2759                 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2760                      target);
2761                 ret = -EHOSTDOWN;
2762         }
2763         spin_unlock(&dlm->spinlock);
2764
2765         /*
2766          * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2767          * another try; otherwise, we are sure the MIGRATING state is there,
2768          * drop the unneded state which blocked threads trying to DIRTY
2769          */
2770         spin_lock(&res->spinlock);
2771         BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2772         res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2773         if (!ret)
2774                 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2775         spin_unlock(&res->spinlock);
2776
2777         /*
2778          * at this point:
2779          *
2780          *   o the DLM_LOCK_RES_MIGRATING flag is set if target not down
2781          *   o there are no pending asts on this lockres
2782          *   o all processes trying to reserve an ast on this
2783          *     lockres must wait for the MIGRATING flag to clear
2784          */
2785         return ret;
2786 }
2787
2788 /* last step in the migration process.
2789  * original master calls this to free all of the dlm_lock
2790  * structures that used to be for other nodes. */
2791 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2792                                       struct dlm_lock_resource *res)
2793 {
2794         struct list_head *queue = &res->granted;
2795         int i, bit;
2796         struct dlm_lock *lock, *next;
2797
2798         assert_spin_locked(&res->spinlock);
2799
2800         BUG_ON(res->owner == dlm->node_num);
2801
2802         for (i=0; i<3; i++) {
2803                 list_for_each_entry_safe(lock, next, queue, list) {
2804                         if (lock->ml.node != dlm->node_num) {
2805                                 mlog(0, "putting lock for node %u\n",
2806                                      lock->ml.node);
2807                                 /* be extra careful */
2808                                 BUG_ON(!list_empty(&lock->ast_list));
2809                                 BUG_ON(!list_empty(&lock->bast_list));
2810                                 BUG_ON(lock->ast_pending);
2811                                 BUG_ON(lock->bast_pending);
2812                                 dlm_lockres_clear_refmap_bit(dlm, res,
2813                                                              lock->ml.node);
2814                                 list_del_init(&lock->list);
2815                                 dlm_lock_put(lock);
2816                                 /* In a normal unlock, we would have added a
2817                                  * DLM_UNLOCK_FREE_LOCK action. Force it. */
2818                                 dlm_lock_put(lock);
2819                         }
2820                 }
2821                 queue++;
2822         }
2823         bit = 0;
2824         while (1) {
2825                 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2826                 if (bit >= O2NM_MAX_NODES)
2827                         break;
2828                 /* do not clear the local node reference, if there is a
2829                  * process holding this, let it drop the ref itself */
2830                 if (bit != dlm->node_num) {
2831                         mlog(0, "%s:%.*s: node %u had a ref to this "
2832                              "migrating lockres, clearing\n", dlm->name,
2833                              res->lockname.len, res->lockname.name, bit);
2834                         dlm_lockres_clear_refmap_bit(dlm, res, bit);
2835                 }
2836                 bit++;
2837         }
2838 }
2839
2840 /*
2841  * Pick a node to migrate the lock resource to. This function selects a
2842  * potential target based first on the locks and then on refmap. It skips
2843  * nodes that are in the process of exiting the domain.
2844  */
2845 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2846                                     struct dlm_lock_resource *res)
2847 {
2848         enum dlm_lockres_list idx;
2849         struct list_head *queue = &res->granted;
2850         struct dlm_lock *lock;
2851         int noderef;
2852         u8 nodenum = O2NM_MAX_NODES;
2853
2854         assert_spin_locked(&dlm->spinlock);
2855         assert_spin_locked(&res->spinlock);
2856
2857         /* Go through all the locks */
2858         for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2859                 queue = dlm_list_idx_to_ptr(res, idx);
2860                 list_for_each_entry(lock, queue, list) {
2861                         if (lock->ml.node == dlm->node_num)
2862                                 continue;
2863                         if (test_bit(lock->ml.node, dlm->exit_domain_map))
2864                                 continue;
2865                         nodenum = lock->ml.node;
2866                         goto bail;
2867                 }
2868         }
2869
2870         /* Go thru the refmap */
2871         noderef = -1;
2872         while (1) {
2873                 noderef = find_next_bit(res->refmap, O2NM_MAX_NODES,
2874                                         noderef + 1);
2875                 if (noderef >= O2NM_MAX_NODES)
2876                         break;
2877                 if (noderef == dlm->node_num)
2878                         continue;
2879                 if (test_bit(noderef, dlm->exit_domain_map))
2880                         continue;
2881                 nodenum = noderef;
2882                 goto bail;
2883         }
2884
2885 bail:
2886         return nodenum;
2887 }
2888
2889 /* this is called by the new master once all lockres
2890  * data has been received */
2891 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2892                                   struct dlm_lock_resource *res,
2893                                   u8 master, u8 new_master,
2894                                   struct dlm_node_iter *iter)
2895 {
2896         struct dlm_migrate_request migrate;
2897         int ret, skip, status = 0;
2898         int nodenum;
2899
2900         memset(&migrate, 0, sizeof(migrate));
2901         migrate.namelen = res->lockname.len;
2902         memcpy(migrate.name, res->lockname.name, migrate.namelen);
2903         migrate.new_master = new_master;
2904         migrate.master = master;
2905
2906         ret = 0;
2907
2908         /* send message to all nodes, except the master and myself */
2909         while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2910                 if (nodenum == master ||
2911                     nodenum == new_master)
2912                         continue;
2913
2914                 /* We could race exit domain. If exited, skip. */
2915                 spin_lock(&dlm->spinlock);
2916                 skip = (!test_bit(nodenum, dlm->domain_map));
2917                 spin_unlock(&dlm->spinlock);
2918                 if (skip) {
2919                         clear_bit(nodenum, iter->node_map);
2920                         continue;
2921                 }
2922
2923                 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2924                                          &migrate, sizeof(migrate), nodenum,
2925                                          &status);
2926                 if (ret < 0) {
2927                         mlog(ML_ERROR, "%s: res %.*s, Error %d send "
2928                              "MIGRATE_REQUEST to node %u\n", dlm->name,
2929                              migrate.namelen, migrate.name, ret, nodenum);
2930                         if (!dlm_is_host_down(ret)) {
2931                                 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
2932                                 BUG();
2933                         }
2934                         clear_bit(nodenum, iter->node_map);
2935                         ret = 0;
2936                 } else if (status < 0) {
2937                         mlog(0, "migrate request (node %u) returned %d!\n",
2938                              nodenum, status);
2939                         ret = status;
2940                 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
2941                         /* during the migration request we short-circuited
2942                          * the mastery of the lockres.  make sure we have
2943                          * a mastery ref for nodenum */
2944                         mlog(0, "%s:%.*s: need ref for node %u\n",
2945                              dlm->name, res->lockname.len, res->lockname.name,
2946                              nodenum);
2947                         spin_lock(&res->spinlock);
2948                         dlm_lockres_set_refmap_bit(dlm, res, nodenum);
2949                         spin_unlock(&res->spinlock);
2950                 }
2951         }
2952
2953         if (ret < 0)
2954                 mlog_errno(ret);
2955
2956         mlog(0, "returning ret=%d\n", ret);
2957         return ret;
2958 }
2959
2960
2961 /* if there is an existing mle for this lockres, we now know who the master is.
2962  * (the one who sent us *this* message) we can clear it up right away.
2963  * since the process that put the mle on the list still has a reference to it,
2964  * we can unhash it now, set the master and wake the process.  as a result,
2965  * we will have no mle in the list to start with.  now we can add an mle for
2966  * the migration and this should be the only one found for those scanning the
2967  * list.  */
2968 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
2969                                 void **ret_data)
2970 {
2971         struct dlm_ctxt *dlm = data;
2972         struct dlm_lock_resource *res = NULL;
2973         struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
2974         struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
2975         const char *name;
2976         unsigned int namelen, hash;
2977         int ret = 0;
2978
2979         if (!dlm_grab(dlm))
2980                 return -EINVAL;
2981
2982         name = migrate->name;
2983         namelen = migrate->namelen;
2984         hash = dlm_lockid_hash(name, namelen);
2985
2986         /* preallocate.. if this fails, abort */
2987         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2988
2989         if (!mle) {
2990                 ret = -ENOMEM;
2991                 goto leave;
2992         }
2993
2994         /* check for pre-existing lock */
2995         spin_lock(&dlm->spinlock);
2996         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
2997         if (res) {
2998                 spin_lock(&res->spinlock);
2999                 if (res->state & DLM_LOCK_RES_RECOVERING) {
3000                         /* if all is working ok, this can only mean that we got
3001                         * a migrate request from a node that we now see as
3002                         * dead.  what can we do here?  drop it to the floor? */
3003                         spin_unlock(&res->spinlock);
3004                         mlog(ML_ERROR, "Got a migrate request, but the "
3005                              "lockres is marked as recovering!");
3006                         kmem_cache_free(dlm_mle_cache, mle);
3007                         ret = -EINVAL; /* need a better solution */
3008                         goto unlock;
3009                 }
3010                 res->state |= DLM_LOCK_RES_MIGRATING;
3011                 spin_unlock(&res->spinlock);
3012         }
3013
3014         spin_lock(&dlm->master_lock);
3015         /* ignore status.  only nonzero status would BUG. */
3016         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3017                                     name, namelen,
3018                                     migrate->new_master,
3019                                     migrate->master);
3020
3021         spin_unlock(&dlm->master_lock);
3022 unlock:
3023         spin_unlock(&dlm->spinlock);
3024
3025         if (oldmle) {
3026                 /* master is known, detach if not already detached */
3027                 dlm_mle_detach_hb_events(dlm, oldmle);
3028                 dlm_put_mle(oldmle);
3029         }
3030
3031         if (res)
3032                 dlm_lockres_put(res);
3033 leave:
3034         dlm_put(dlm);
3035         return ret;
3036 }
3037
3038 /* must be holding dlm->spinlock and dlm->master_lock
3039  * when adding a migration mle, we can clear any other mles
3040  * in the master list because we know with certainty that
3041  * the master is "master".  so we remove any old mle from
3042  * the list after setting it's master field, and then add
3043  * the new migration mle.  this way we can hold with the rule
3044  * of having only one mle for a given lock name at all times. */
3045 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3046                                  struct dlm_lock_resource *res,
3047                                  struct dlm_master_list_entry *mle,
3048                                  struct dlm_master_list_entry **oldmle,
3049                                  const char *name, unsigned int namelen,
3050                                  u8 new_master, u8 master)
3051 {
3052         int found;
3053         int ret = 0;
3054
3055         *oldmle = NULL;
3056
3057         assert_spin_locked(&dlm->spinlock);
3058         assert_spin_locked(&dlm->master_lock);
3059
3060         /* caller is responsible for any ref taken here on oldmle */
3061         found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3062         if (found) {
3063                 struct dlm_master_list_entry *tmp = *oldmle;
3064                 spin_lock(&tmp->spinlock);
3065                 if (tmp->type == DLM_MLE_MIGRATION) {
3066                         if (master == dlm->node_num) {
3067                                 /* ah another process raced me to it */
3068                                 mlog(0, "tried to migrate %.*s, but some "
3069                                      "process beat me to it\n",
3070                                      namelen, name);
3071                                 ret = -EEXIST;
3072                         } else {
3073                                 /* bad.  2 NODES are trying to migrate! */
3074                                 mlog(ML_ERROR, "migration error  mle: "
3075                                      "master=%u new_master=%u // request: "
3076                                      "master=%u new_master=%u // "
3077                                      "lockres=%.*s\n",
3078                                      tmp->master, tmp->new_master,
3079                                      master, new_master,
3080                                      namelen, name);
3081                                 BUG();
3082                         }
3083                 } else {
3084                         /* this is essentially what assert_master does */
3085                         tmp->master = master;
3086                         atomic_set(&tmp->woken, 1);
3087                         wake_up(&tmp->wq);
3088                         /* remove it so that only one mle will be found */
3089                         __dlm_unlink_mle(dlm, tmp);
3090                         __dlm_mle_detach_hb_events(dlm, tmp);
3091                         ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3092                         mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3093                             "telling master to get ref for cleared out mle "
3094                             "during migration\n", dlm->name, namelen, name,
3095                             master, new_master);
3096                 }
3097                 spin_unlock(&tmp->spinlock);
3098         }
3099
3100         /* now add a migration mle to the tail of the list */
3101         dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3102         mle->new_master = new_master;
3103         /* the new master will be sending an assert master for this.
3104          * at that point we will get the refmap reference */
3105         mle->master = master;
3106         /* do this for consistency with other mle types */
3107         set_bit(new_master, mle->maybe_map);
3108         __dlm_insert_mle(dlm, mle);
3109
3110         return ret;
3111 }
3112
3113 /*
3114  * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3115  */
3116 static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3117                                         struct dlm_master_list_entry *mle)
3118 {
3119         struct dlm_lock_resource *res;
3120
3121         /* Find the lockres associated to the mle and set its owner to UNK */
3122         res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3123                                    mle->mnamehash);
3124         if (res) {
3125                 spin_unlock(&dlm->master_lock);
3126
3127                 /* move lockres onto recovery list */
3128                 spin_lock(&res->spinlock);
3129                 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3130                 dlm_move_lockres_to_recovery_list(dlm, res);
3131                 spin_unlock(&res->spinlock);
3132                 dlm_lockres_put(res);
3133
3134                 /* about to get rid of mle, detach from heartbeat */
3135                 __dlm_mle_detach_hb_events(dlm, mle);
3136
3137                 /* dump the mle */
3138                 spin_lock(&dlm->master_lock);
3139                 __dlm_put_mle(mle);
3140                 spin_unlock(&dlm->master_lock);
3141         }
3142
3143         return res;
3144 }
3145
3146 static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3147                                     struct dlm_master_list_entry *mle)
3148 {
3149         __dlm_mle_detach_hb_events(dlm, mle);
3150
3151         spin_lock(&mle->spinlock);
3152         __dlm_unlink_mle(dlm, mle);
3153         atomic_set(&mle->woken, 1);
3154         spin_unlock(&mle->spinlock);
3155
3156         wake_up(&mle->wq);
3157 }
3158
3159 static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3160                                 struct dlm_master_list_entry *mle, u8 dead_node)
3161 {
3162         int bit;
3163
3164         BUG_ON(mle->type != DLM_MLE_BLOCK);
3165
3166         spin_lock(&mle->spinlock);
3167         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3168         if (bit != dead_node) {
3169                 mlog(0, "mle found, but dead node %u would not have been "
3170                      "master\n", dead_node);
3171                 spin_unlock(&mle->spinlock);
3172         } else {
3173                 /* Must drop the refcount by one since the assert_master will
3174                  * never arrive. This may result in the mle being unlinked and
3175                  * freed, but there may still be a process waiting in the
3176                  * dlmlock path which is fine. */
3177                 mlog(0, "node %u was expected master\n", dead_node);
3178                 atomic_set(&mle->woken, 1);
3179                 spin_unlock(&mle->spinlock);
3180                 wake_up(&mle->wq);
3181
3182                 /* Do not need events any longer, so detach from heartbeat */
3183                 __dlm_mle_detach_hb_events(dlm, mle);
3184                 __dlm_put_mle(mle);
3185         }
3186 }
3187
3188 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3189 {
3190         struct dlm_master_list_entry *mle;
3191         struct dlm_lock_resource *res;
3192         struct hlist_head *bucket;
3193         struct hlist_node *tmp;
3194         unsigned int i;
3195
3196         mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
3197 top:
3198         assert_spin_locked(&dlm->spinlock);
3199
3200         /* clean the master list */
3201         spin_lock(&dlm->master_lock);
3202         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3203                 bucket = dlm_master_hash(dlm, i);
3204                 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3205                         BUG_ON(mle->type != DLM_MLE_BLOCK &&
3206                                mle->type != DLM_MLE_MASTER &&
3207                                mle->type != DLM_MLE_MIGRATION);
3208
3209                         /* MASTER mles are initiated locally. The waiting
3210                          * process will notice the node map change shortly.
3211                          * Let that happen as normal. */
3212                         if (mle->type == DLM_MLE_MASTER)
3213                                 continue;
3214
3215                         /* BLOCK mles are initiated by other nodes. Need to
3216                          * clean up if the dead node would have been the
3217                          * master. */
3218                         if (mle->type == DLM_MLE_BLOCK) {
3219                                 dlm_clean_block_mle(dlm, mle, dead_node);
3220                                 continue;
3221                         }
3222
3223                         /* Everything else is a MIGRATION mle */
3224
3225                         /* The rule for MIGRATION mles is that the master
3226                          * becomes UNKNOWN if *either* the original or the new
3227                          * master dies. All UNKNOWN lockres' are sent to
3228                          * whichever node becomes the recovery master. The new
3229                          * master is responsible for determining if there is
3230                          * still a master for this lockres, or if he needs to
3231                          * take over mastery. Either way, this node should
3232                          * expect another message to resolve this. */
3233
3234                         if (mle->master != dead_node &&
3235                             mle->new_master != dead_node)
3236                                 continue;
3237
3238                         /* If we have reached this point, this mle needs to be
3239                          * removed from the list and freed. */
3240                         dlm_clean_migration_mle(dlm, mle);
3241
3242                         mlog(0, "%s: node %u died during migration from "
3243                              "%u to %u!\n", dlm->name, dead_node, mle->master,
3244                              mle->new_master);
3245
3246                         /* If we find a lockres associated with the mle, we've
3247                          * hit this rare case that messes up our lock ordering.
3248                          * If so, we need to drop the master lock so that we can
3249                          * take the lockres lock, meaning that we will have to
3250                          * restart from the head of list. */
3251                         res = dlm_reset_mleres_owner(dlm, mle);
3252                         if (res)
3253                                 /* restart */
3254                                 goto top;
3255
3256                         /* This may be the last reference */
3257                         __dlm_put_mle(mle);
3258                 }
3259         }
3260         spin_unlock(&dlm->master_lock);
3261 }
3262
3263 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3264                          u8 old_master)
3265 {
3266         struct dlm_node_iter iter;
3267         int ret = 0;
3268
3269         spin_lock(&dlm->spinlock);
3270         dlm_node_iter_init(dlm->domain_map, &iter);
3271         clear_bit(old_master, iter.node_map);
3272         clear_bit(dlm->node_num, iter.node_map);
3273         spin_unlock(&dlm->spinlock);
3274
3275         /* ownership of the lockres is changing.  account for the
3276          * mastery reference here since old_master will briefly have
3277          * a reference after the migration completes */
3278         spin_lock(&res->spinlock);
3279         dlm_lockres_set_refmap_bit(dlm, res, old_master);
3280         spin_unlock(&res->spinlock);
3281
3282         mlog(0, "now time to do a migrate request to other nodes\n");
3283         ret = dlm_do_migrate_request(dlm, res, old_master,
3284                                      dlm->node_num, &iter);
3285         if (ret < 0) {
3286                 mlog_errno(ret);
3287                 goto leave;
3288         }
3289
3290         mlog(0, "doing assert master of %.*s to all except the original node\n",
3291              res->lockname.len, res->lockname.name);
3292         /* this call now finishes out the nodemap
3293          * even if one or more nodes die */
3294         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3295                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3296         if (ret < 0) {
3297                 /* no longer need to retry.  all living nodes contacted. */
3298                 mlog_errno(ret);
3299                 ret = 0;
3300         }
3301
3302         memset(iter.node_map, 0, sizeof(iter.node_map));
3303         set_bit(old_master, iter.node_map);
3304         mlog(0, "doing assert master of %.*s back to %u\n",
3305              res->lockname.len, res->lockname.name, old_master);
3306         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3307                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3308         if (ret < 0) {
3309                 mlog(0, "assert master to original master failed "
3310                      "with %d.\n", ret);
3311                 /* the only nonzero status here would be because of
3312                  * a dead original node.  we're done. */
3313                 ret = 0;
3314         }
3315
3316         /* all done, set the owner, clear the flag */
3317         spin_lock(&res->spinlock);
3318         dlm_set_lockres_owner(dlm, res, dlm->node_num);
3319         res->state &= ~DLM_LOCK_RES_MIGRATING;
3320         spin_unlock(&res->spinlock);
3321         /* re-dirty it on the new master */
3322         dlm_kick_thread(dlm, res);
3323         wake_up(&res->wq);
3324 leave:
3325         return ret;
3326 }
3327
3328 /*
3329  * LOCKRES AST REFCOUNT
3330  * this is integral to migration
3331  */
3332
3333 /* for future intent to call an ast, reserve one ahead of time.
3334  * this should be called only after waiting on the lockres
3335  * with dlm_wait_on_lockres, and while still holding the
3336  * spinlock after the call. */
3337 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3338 {
3339         assert_spin_locked(&res->spinlock);
3340         if (res->state & DLM_LOCK_RES_MIGRATING) {
3341                 __dlm_print_one_lock_resource(res);
3342         }
3343         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3344
3345         atomic_inc(&res->asts_reserved);
3346 }
3347
3348 /*
3349  * used to drop the reserved ast, either because it went unused,
3350  * or because the ast/bast was actually called.
3351  *
3352  * also, if there is a pending migration on this lockres,
3353  * and this was the last pending ast on the lockres,
3354  * atomically set the MIGRATING flag before we drop the lock.
3355  * this is how we ensure that migration can proceed with no
3356  * asts in progress.  note that it is ok if the state of the
3357  * queues is such that a lock should be granted in the future
3358  * or that a bast should be fired, because the new master will
3359  * shuffle the lists on this lockres as soon as it is migrated.
3360  */
3361 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3362                              struct dlm_lock_resource *res)
3363 {
3364         if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3365                 return;
3366
3367         if (!res->migration_pending) {
3368                 spin_unlock(&res->spinlock);
3369                 return;
3370         }
3371
3372         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3373         res->migration_pending = 0;
3374         res->state |= DLM_LOCK_RES_MIGRATING;
3375         spin_unlock(&res->spinlock);
3376         wake_up(&res->wq);
3377         wake_up(&dlm->migration_wq);
3378 }
3379
3380 void dlm_force_free_mles(struct dlm_ctxt *dlm)
3381 {
3382         int i;
3383         struct hlist_head *bucket;
3384         struct dlm_master_list_entry *mle;
3385         struct hlist_node *tmp;
3386
3387         /*
3388          * We notified all other nodes that we are exiting the domain and
3389          * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3390          * around we force free them and wake any processes that are waiting
3391          * on the mles
3392          */
3393         spin_lock(&dlm->spinlock);
3394         spin_lock(&dlm->master_lock);
3395
3396         BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3397         BUG_ON((find_next_bit(dlm->domain_map, O2NM_MAX_NODES, 0) < O2NM_MAX_NODES));
3398
3399         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3400                 bucket = dlm_master_hash(dlm, i);
3401                 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3402                         if (mle->type != DLM_MLE_BLOCK) {
3403                                 mlog(ML_ERROR, "bad mle: %p\n", mle);
3404                                 dlm_print_one_mle(mle);
3405                         }
3406                         atomic_set(&mle->woken, 1);
3407                         wake_up(&mle->wq);
3408
3409                         __dlm_unlink_mle(dlm, mle);
3410                         __dlm_mle_detach_hb_events(dlm, mle);
3411                         __dlm_put_mle(mle);
3412                 }
3413         }
3414         spin_unlock(&dlm->master_lock);
3415         spin_unlock(&dlm->spinlock);
3416 }