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