fs/ocfs2/dlm/dlmmaster.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmmod.c
   5  *
   6  * standalone DLM module
   7  *
   8  * Copyright (C) 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  *
  25  */
  26
  27
  28 #include <linux/module.h>
  29 #include <linux/fs.h>
  30 #include <linux/types.h>
  31 #include <linux/slab.h>
  32 #include <linux/highmem.h>
  33 #include <linux/init.h>
  34 #include <linux/sysctl.h>
  35 #include <linux/random.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/socket.h>
  38 #include <linux/inet.h>
  39 #include <linux/spinlock.h>
  40 #include <linux/delay.h>
  41
  42
  43 #include "cluster/heartbeat.h"
  44 #include "cluster/nodemanager.h"
  45 #include "cluster/tcp.h"
  46
  47 #include "dlmapi.h"
  48 #include "dlmcommon.h"
  49 #include "dlmdomain.h"
  50 #include "dlmdebug.h"
  51
  52 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
  53 #include "cluster/masklog.h"
  54
  55 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
  56                               struct dlm_master_list_entry *mle,
  57                               struct o2nm_node *node,
  58                               int idx);
  59 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
  60                             struct dlm_master_list_entry *mle,
  61                             struct o2nm_node *node,
  62                             int idx);
  63
  64 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
  65 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
  66                                 struct dlm_lock_resource *res,
  67                                 void *nodemap, u32 flags);
  68 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
  69
  70 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
  71                                 struct dlm_master_list_entry *mle,
  72                                 const char *name,
  73                                 unsigned int namelen)
  74 {
  75         if (dlm != mle->dlm)
  76                 return 0;
  77
  78         if (namelen != mle->mnamelen ||
  79             memcmp(name, mle->mname, namelen) != 0)
  80                 return 0;
  81
  82         return 1;
  83 }
  84
  85 static struct kmem_cache *dlm_lockres_cache;
  86 static struct kmem_cache *dlm_lockname_cache;
  87 static struct kmem_cache *dlm_mle_cache;
  88
  89 static void dlm_mle_release(struct kref *kref);
  90 static void dlm_init_mle(struct dlm_master_list_entry *mle,
  91                         enum dlm_mle_type type,
  92                         struct dlm_ctxt *dlm,
  93                         struct dlm_lock_resource *res,
  94                         const char *name,
  95                         unsigned int namelen);
  96 static void dlm_put_mle(struct dlm_master_list_entry *mle);
  97 static void __dlm_put_mle(struct dlm_master_list_entry *mle);
  98 static int dlm_find_mle(struct dlm_ctxt *dlm,
  99                         struct dlm_master_list_entry **mle,
 100                         char *name, unsigned int namelen);
 101
 102 static int dlm_do_master_request(struct dlm_lock_resource *res,
 103                                  struct dlm_master_list_entry *mle, int to);
 104
 105
 106 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
 107                                      struct dlm_lock_resource *res,
 108                                      struct dlm_master_list_entry *mle,
 109                                      int *blocked);
 110 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
 111                                     struct dlm_lock_resource *res,
 112                                     struct dlm_master_list_entry *mle,
 113                                     int blocked);
 114 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
 115                                  struct dlm_lock_resource *res,
 116                                  struct dlm_master_list_entry *mle,
 117                                  struct dlm_master_list_entry **oldmle,
 118                                  const char *name, unsigned int namelen,
 119                                  u8 new_master, u8 master);
 120
 121 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
 122                                     struct dlm_lock_resource *res);
 123 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
 124                                       struct dlm_lock_resource *res);
 125 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
 126                                        struct dlm_lock_resource *res,
 127                                        u8 target);
 128 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
 129                                        struct dlm_lock_resource *res);
 130
 131
 132 int dlm_is_host_down(int errno)
 133 {
 134         switch (errno) {
 135                 case -EBADF:
 136                 case -ECONNREFUSED:
 137                 case -ENOTCONN:
 138                 case -ECONNRESET:
 139                 case -EPIPE:
 140                 case -EHOSTDOWN:
 141                 case -EHOSTUNREACH:
 142                 case -ETIMEDOUT:
 143                 case -ECONNABORTED:
 144                 case -ENETDOWN:
 145                 case -ENETUNREACH:
 146                 case -ENETRESET:
 147                 case -ESHUTDOWN:
 148                 case -ENOPROTOOPT:
 149                 case -EINVAL:   /* if returned from our tcp code,
 150                                    this means there is no socket */
 151                         return 1;
 152         }
 153         return 0;
 154 }
 155
 156
 157 /*
 158  * MASTER LIST FUNCTIONS
 159  */
 160
 161
 162 /*
 163  * regarding master list entries and heartbeat callbacks:
 164  *
 165  * in order to avoid sleeping and allocation that occurs in
 166  * heartbeat, master list entries are simply attached to the
 167  * dlm's established heartbeat callbacks.  the mle is attached
 168  * when it is created, and since the dlm->spinlock is held at
 169  * that time, any heartbeat event will be properly discovered
 170  * by the mle.  the mle needs to be detached from the
 171  * dlm->mle_hb_events list as soon as heartbeat events are no
 172  * longer useful to the mle, and before the mle is freed.
 173  *
 174  * as a general rule, heartbeat events are no longer needed by
 175  * the mle once an "answer" regarding the lock master has been
 176  * received.
 177  */
 178 static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
 179                                               struct dlm_master_list_entry *mle)
 180 {
 181         assert_spin_locked(&dlm->spinlock);
 182
 183         list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
 184 }
 185
 186
 187 static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 188                                               struct dlm_master_list_entry *mle)
 189 {
 190         if (!list_empty(&mle->hb_events))
 191                 list_del_init(&mle->hb_events);
 192 }
 193
 194
 195 static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 196                                             struct dlm_master_list_entry *mle)
 197 {
 198         spin_lock(&dlm->spinlock);
 199         __dlm_mle_detach_hb_events(dlm, mle);
 200         spin_unlock(&dlm->spinlock);
 201 }
 202
 203 static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
 204 {
 205         struct dlm_ctxt *dlm;
 206         dlm = mle->dlm;
 207
 208         assert_spin_locked(&dlm->spinlock);
 209         assert_spin_locked(&dlm->master_lock);
 210         mle->inuse++;
 211         kref_get(&mle->mle_refs);
 212 }
 213
 214 static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
 215 {
 216         struct dlm_ctxt *dlm;
 217         dlm = mle->dlm;
 218
 219         spin_lock(&dlm->spinlock);
 220         spin_lock(&dlm->master_lock);
 221         mle->inuse--;
 222         __dlm_put_mle(mle);
 223         spin_unlock(&dlm->master_lock);
 224         spin_unlock(&dlm->spinlock);
 225
 226 }
 227
 228 /* remove from list and free */
 229 static void __dlm_put_mle(struct dlm_master_list_entry *mle)
 230 {
 231         struct dlm_ctxt *dlm;
 232         dlm = mle->dlm;
 233
 234         assert_spin_locked(&dlm->spinlock);
 235         assert_spin_locked(&dlm->master_lock);
 236         if (!kref_read(&mle->mle_refs)) {
 237                 /* this may or may not crash, but who cares.
 238                  * it's a BUG. */
 239                 mlog(ML_ERROR, "bad mle: %p\n", mle);
 240                 dlm_print_one_mle(mle);
 241                 BUG();
 242         } else
 243                 kref_put(&mle->mle_refs, dlm_mle_release);
 244 }
 245
 246
 247 /* must not have any spinlocks coming in */
 248 static void dlm_put_mle(struct dlm_master_list_entry *mle)
 249 {
 250         struct dlm_ctxt *dlm;
 251         dlm = mle->dlm;
 252
 253         spin_lock(&dlm->spinlock);
 254         spin_lock(&dlm->master_lock);
 255         __dlm_put_mle(mle);
 256         spin_unlock(&dlm->master_lock);
 257         spin_unlock(&dlm->spinlock);
 258 }
 259
 260 static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
 261 {
 262         kref_get(&mle->mle_refs);
 263 }
 264
 265 static void dlm_init_mle(struct dlm_master_list_entry *mle,
 266                         enum dlm_mle_type type,
 267                         struct dlm_ctxt *dlm,
 268                         struct dlm_lock_resource *res,
 269                         const char *name,
 270                         unsigned int namelen)
 271 {
 272         assert_spin_locked(&dlm->spinlock);
 273
 274         mle->dlm = dlm;
 275         mle->type = type;
 276         INIT_HLIST_NODE(&mle->master_hash_node);
 277         INIT_LIST_HEAD(&mle->hb_events);
 278         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
 279         spin_lock_init(&mle->spinlock);
 280         init_waitqueue_head(&mle->wq);
 281         atomic_set(&mle->woken, 0);
 282         kref_init(&mle->mle_refs);
 283         memset(mle->response_map, 0, sizeof(mle->response_map));
 284         mle->master = O2NM_MAX_NODES;
 285         mle->new_master = O2NM_MAX_NODES;
 286         mle->inuse = 0;
 287
 288         BUG_ON(mle->type != DLM_MLE_BLOCK &&
 289                mle->type != DLM_MLE_MASTER &&
 290                mle->type != DLM_MLE_MIGRATION);
 291
 292         if (mle->type == DLM_MLE_MASTER) {
 293                 BUG_ON(!res);
 294                 mle->mleres = res;
 295                 memcpy(mle->mname, res->lockname.name, res->lockname.len);
 296                 mle->mnamelen = res->lockname.len;
 297                 mle->mnamehash = res->lockname.hash;
 298         } else {
 299                 BUG_ON(!name);
 300                 mle->mleres = NULL;
 301                 memcpy(mle->mname, name, namelen);
 302                 mle->mnamelen = namelen;
 303                 mle->mnamehash = dlm_lockid_hash(name, namelen);
 304         }
 305
 306         atomic_inc(&dlm->mle_tot_count[mle->type]);
 307         atomic_inc(&dlm->mle_cur_count[mle->type]);
 308
 309         /* copy off the node_map and register hb callbacks on our copy */
 310         memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
 311         memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
 312         clear_bit(dlm->node_num, mle->vote_map);
 313         clear_bit(dlm->node_num, mle->node_map);
 314
 315         /* attach the mle to the domain node up/down events */
 316         __dlm_mle_attach_hb_events(dlm, mle);
 317 }
 318
 319 void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
 320 {
 321         assert_spin_locked(&dlm->spinlock);
 322         assert_spin_locked(&dlm->master_lock);
 323
 324         if (!hlist_unhashed(&mle->master_hash_node))
 325                 hlist_del_init(&mle->master_hash_node);
 326 }
 327
 328 void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
 329 {
 330         struct hlist_head *bucket;
 331
 332         assert_spin_locked(&dlm->master_lock);
 333
 334         bucket = dlm_master_hash(dlm, mle->mnamehash);
 335         hlist_add_head(&mle->master_hash_node, bucket);
 336 }
 337
 338 /* returns 1 if found, 0 if not */
 339 static int dlm_find_mle(struct dlm_ctxt *dlm,
 340                         struct dlm_master_list_entry **mle,
 341                         char *name, unsigned int namelen)
 342 {
 343         struct dlm_master_list_entry *tmpmle;
 344         struct hlist_head *bucket;
 345         unsigned int hash;
 346
 347         assert_spin_locked(&dlm->master_lock);
 348
 349         hash = dlm_lockid_hash(name, namelen);
 350         bucket = dlm_master_hash(dlm, hash);
 351         hlist_for_each_entry(tmpmle, bucket, master_hash_node) {
 352                 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
 353                         continue;
 354                 dlm_get_mle(tmpmle);
 355                 *mle = tmpmle;
 356                 return 1;
 357         }
 358         return 0;
 359 }
 360
 361 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
 362 {
 363         struct dlm_master_list_entry *mle;
 364
 365         assert_spin_locked(&dlm->spinlock);
 366
 367         list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
 368                 if (node_up)
 369                         dlm_mle_node_up(dlm, mle, NULL, idx);
 370                 else
 371                         dlm_mle_node_down(dlm, mle, NULL, idx);
 372         }
 373 }
 374
 375 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
 376                               struct dlm_master_list_entry *mle,
 377                               struct o2nm_node *node, int idx)
 378 {
 379         spin_lock(&mle->spinlock);
 380
 381         if (!test_bit(idx, mle->node_map))
 382                 mlog(0, "node %u already removed from nodemap!\n", idx);
 383         else
 384                 clear_bit(idx, mle->node_map);
 385
 386         spin_unlock(&mle->spinlock);
 387 }
 388
 389 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
 390                             struct dlm_master_list_entry *mle,
 391                             struct o2nm_node *node, int idx)
 392 {
 393         spin_lock(&mle->spinlock);
 394
 395         if (test_bit(idx, mle->node_map))
 396                 mlog(0, "node %u already in node map!\n", idx);
 397         else
 398                 set_bit(idx, mle->node_map);
 399
 400         spin_unlock(&mle->spinlock);
 401 }
 402
 403
 404 int dlm_init_mle_cache(void)
 405 {
 406         dlm_mle_cache = kmem_cache_create("o2dlm_mle",
 407                                           sizeof(struct dlm_master_list_entry),
 408                                           0, SLAB_HWCACHE_ALIGN,
 409                                           NULL);
 410         if (dlm_mle_cache == NULL)
 411                 return -ENOMEM;
 412         return 0;
 413 }
 414
 415 void dlm_destroy_mle_cache(void)
 416 {
 417         if (dlm_mle_cache)
 418                 kmem_cache_destroy(dlm_mle_cache);
 419 }
 420
 421 static void dlm_mle_release(struct kref *kref)
 422 {
 423         struct dlm_master_list_entry *mle;
 424         struct dlm_ctxt *dlm;
 425
 426         mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
 427         dlm = mle->dlm;
 428
 429         assert_spin_locked(&dlm->spinlock);
 430         assert_spin_locked(&dlm->master_lock);
 431
 432         mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
 433              mle->type);
 434
 435         /* remove from list if not already */
 436         __dlm_unlink_mle(dlm, mle);
 437
 438         /* detach the mle from the domain node up/down events */
 439         __dlm_mle_detach_hb_events(dlm, mle);
 440
 441         atomic_dec(&dlm->mle_cur_count[mle->type]);
 442
 443         /* NOTE: kfree under spinlock here.
 444          * if this is bad, we can move this to a freelist. */
 445         kmem_cache_free(dlm_mle_cache, mle);
 446 }
 447
 448
 449 /*
 450  * LOCK RESOURCE FUNCTIONS
 451  */
 452
 453 int dlm_init_master_caches(void)
 454 {
 455         dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
 456                                               sizeof(struct dlm_lock_resource),
 457                                               0, SLAB_HWCACHE_ALIGN, NULL);
 458         if (!dlm_lockres_cache)
 459                 goto bail;
 460
 461         dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
 462                                                DLM_LOCKID_NAME_MAX, 0,
 463                                                SLAB_HWCACHE_ALIGN, NULL);
 464         if (!dlm_lockname_cache)
 465                 goto bail;
 466
 467         return 0;
 468 bail:
 469         dlm_destroy_master_caches();
 470         return -ENOMEM;
 471 }
 472
 473 void dlm_destroy_master_caches(void)
 474 {
 475         if (dlm_lockname_cache) {
 476                 kmem_cache_destroy(dlm_lockname_cache);
 477                 dlm_lockname_cache = NULL;
 478         }
 479
 480         if (dlm_lockres_cache) {
 481                 kmem_cache_destroy(dlm_lockres_cache);
 482                 dlm_lockres_cache = NULL;
 483         }
 484 }
 485
 486 static void dlm_lockres_release(struct kref *kref)
 487 {
 488         struct dlm_lock_resource *res;
 489         struct dlm_ctxt *dlm;
 490
 491         res = container_of(kref, struct dlm_lock_resource, refs);
 492         dlm = res->dlm;
 493
 494         /* This should not happen -- all lockres' have a name
 495          * associated with them at init time. */
 496         BUG_ON(!res->lockname.name);
 497
 498         mlog(0, "destroying lockres %.*s\n", res->lockname.len,
 499              res->lockname.name);
 500
 501         atomic_dec(&dlm->res_cur_count);
 502
 503         if (!hlist_unhashed(&res->hash_node) ||
 504             !list_empty(&res->granted) ||
 505             !list_empty(&res->converting) ||
 506             !list_empty(&res->blocked) ||
 507             !list_empty(&res->dirty) ||
 508             !list_empty(&res->recovering) ||
 509             !list_empty(&res->purge)) {
 510                 mlog(ML_ERROR,
 511                      "Going to BUG for resource %.*s."
 512                      "  We're on a list! [%c%c%c%c%c%c%c]\n",
 513                      res->lockname.len, res->lockname.name,
 514                      !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
 515                      !list_empty(&res->granted) ? 'G' : ' ',
 516                      !list_empty(&res->converting) ? 'C' : ' ',
 517                      !list_empty(&res->blocked) ? 'B' : ' ',
 518                      !list_empty(&res->dirty) ? 'D' : ' ',
 519                      !list_empty(&res->recovering) ? 'R' : ' ',
 520                      !list_empty(&res->purge) ? 'P' : ' ');
 521
 522                 dlm_print_one_lock_resource(res);
 523         }
 524
 525         /* By the time we're ready to blow this guy away, we shouldn't
 526          * be on any lists. */
 527         BUG_ON(!hlist_unhashed(&res->hash_node));
 528         BUG_ON(!list_empty(&res->granted));
 529         BUG_ON(!list_empty(&res->converting));
 530         BUG_ON(!list_empty(&res->blocked));
 531         BUG_ON(!list_empty(&res->dirty));
 532         BUG_ON(!list_empty(&res->recovering));
 533         BUG_ON(!list_empty(&res->purge));
 534
 535         kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
 536
 537         kmem_cache_free(dlm_lockres_cache, res);
 538 }
 539
 540 void dlm_lockres_put(struct dlm_lock_resource *res)
 541 {
 542         kref_put(&res->refs, dlm_lockres_release);
 543 }
 544
 545 static void dlm_init_lockres(struct dlm_ctxt *dlm,
 546                              struct dlm_lock_resource *res,
 547                              const char *name, unsigned int namelen)
 548 {
 549         char *qname;
 550
 551         /* If we memset here, we lose our reference to the kmalloc'd
 552          * res->lockname.name, so be sure to init every field
 553          * correctly! */
 554
 555         qname = (char *) res->lockname.name;
 556         memcpy(qname, name, namelen);
 557
 558         res->lockname.len = namelen;
 559         res->lockname.hash = dlm_lockid_hash(name, namelen);
 560
 561         init_waitqueue_head(&res->wq);
 562         spin_lock_init(&res->spinlock);
 563         INIT_HLIST_NODE(&res->hash_node);
 564         INIT_LIST_HEAD(&res->granted);
 565         INIT_LIST_HEAD(&res->converting);
 566         INIT_LIST_HEAD(&res->blocked);
 567         INIT_LIST_HEAD(&res->dirty);
 568         INIT_LIST_HEAD(&res->recovering);
 569         INIT_LIST_HEAD(&res->purge);
 570         INIT_LIST_HEAD(&res->tracking);
 571         atomic_set(&res->asts_reserved, 0);
 572         res->migration_pending = 0;
 573         res->inflight_locks = 0;
 574         res->inflight_assert_workers = 0;
 575
 576         res->dlm = dlm;
 577
 578         kref_init(&res->refs);
 579
 580         atomic_inc(&dlm->res_tot_count);
 581         atomic_inc(&dlm->res_cur_count);
 582
 583         /* just for consistency */
 584         spin_lock(&res->spinlock);
 585         dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
 586         spin_unlock(&res->spinlock);
 587
 588         res->state = DLM_LOCK_RES_IN_PROGRESS;
 589
 590         res->last_used = 0;
 591
 592         spin_lock(&dlm->track_lock);
 593         list_add_tail(&res->tracking, &dlm->tracking_list);
 594         spin_unlock(&dlm->track_lock);
 595
 596         memset(res->lvb, 0, DLM_LVB_LEN);
 597         memset(res->refmap, 0, sizeof(res->refmap));
 598 }
 599
 600 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
 601                                    const char *name,
 602                                    unsigned int namelen)
 603 {
 604         struct dlm_lock_resource *res = NULL;
 605
 606         res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
 607         if (!res)
 608                 goto error;
 609
 610         res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
 611         if (!res->lockname.name)
 612                 goto error;
 613
 614         dlm_init_lockres(dlm, res, name, namelen);
 615         return res;
 616
 617 error:
 618         if (res)
 619                 kmem_cache_free(dlm_lockres_cache, res);
 620         return NULL;
 621 }
 622
 623 void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm,
 624                                 struct dlm_lock_resource *res, int bit)
 625 {
 626         assert_spin_locked(&res->spinlock);
 627
 628         mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len,
 629              res->lockname.name, bit, __builtin_return_address(0));
 630
 631         set_bit(bit, res->refmap);
 632 }
 633
 634 void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm,
 635                                   struct dlm_lock_resource *res, int bit)
 636 {
 637         assert_spin_locked(&res->spinlock);
 638
 639         mlog(0, "res %.*s, clr node %u, %ps()\n", res->lockname.len,
 640              res->lockname.name, bit, __builtin_return_address(0));
 641
 642         clear_bit(bit, res->refmap);
 643 }
 644
 645 static void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
 646                                    struct dlm_lock_resource *res)
 647 {
 648         res->inflight_locks++;
 649
 650         mlog(0, "%s: res %.*s, inflight++: now %u, %ps()\n", dlm->name,
 651              res->lockname.len, res->lockname.name, res->inflight_locks,
 652              __builtin_return_address(0));
 653 }
 654
 655 void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
 656                                    struct dlm_lock_resource *res)
 657 {
 658         assert_spin_locked(&res->spinlock);
 659         __dlm_lockres_grab_inflight_ref(dlm, res);
 660 }
 661
 662 void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
 663                                    struct dlm_lock_resource *res)
 664 {
 665         assert_spin_locked(&res->spinlock);
 666
 667         BUG_ON(res->inflight_locks == 0);
 668
 669         res->inflight_locks--;
 670
 671         mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm->name,
 672              res->lockname.len, res->lockname.name, res->inflight_locks,
 673              __builtin_return_address(0));
 674
 675         wake_up(&res->wq);
 676 }
 677
 678 void __dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm,
 679                 struct dlm_lock_resource *res)
 680 {
 681         assert_spin_locked(&res->spinlock);
 682         res->inflight_assert_workers++;
 683         mlog(0, "%s:%.*s: inflight assert worker++: now %u\n",
 684                         dlm->name, res->lockname.len, res->lockname.name,
 685                         res->inflight_assert_workers);
 686 }
 687
 688 static void __dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
 689                 struct dlm_lock_resource *res)
 690 {
 691         assert_spin_locked(&res->spinlock);
 692         BUG_ON(res->inflight_assert_workers == 0);
 693         res->inflight_assert_workers--;
 694         mlog(0, "%s:%.*s: inflight assert worker--: now %u\n",
 695                         dlm->name, res->lockname.len, res->lockname.name,
 696                         res->inflight_assert_workers);
 697 }
 698
 699 static void dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
 700                 struct dlm_lock_resource *res)
 701 {
 702         spin_lock(&res->spinlock);
 703         __dlm_lockres_drop_inflight_worker(dlm, res);
 704         spin_unlock(&res->spinlock);
 705 }
 706
 707 /*
 708  * lookup a lock resource by name.
 709  * may already exist in the hashtable.
 710  * lockid is null terminated
 711  *
 712  * if not, allocate enough for the lockres and for
 713  * the temporary structure used in doing the mastering.
 714  *
 715  * also, do a lookup in the dlm->master_list to see
 716  * if another node has begun mastering the same lock.
 717  * if so, there should be a block entry in there
 718  * for this name, and we should *not* attempt to master
 719  * the lock here.   need to wait around for that node
 720  * to assert_master (or die).
 721  *
 722  */
 723 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
 724                                           const char *lockid,
 725                                           int namelen,
 726                                           int flags)
 727 {
 728         struct dlm_lock_resource *tmpres=NULL, *res=NULL;
 729         struct dlm_master_list_entry *mle = NULL;
 730         struct dlm_master_list_entry *alloc_mle = NULL;
 731         int blocked = 0;
 732         int ret, nodenum;
 733         struct dlm_node_iter iter;
 734         unsigned int hash;
 735         int tries = 0;
 736         int bit, wait_on_recovery = 0;
 737
 738         BUG_ON(!lockid);
 739
 740         hash = dlm_lockid_hash(lockid, namelen);
 741
 742         mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
 743
 744 lookup:
 745         spin_lock(&dlm->spinlock);
 746         tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
 747         if (tmpres) {
 748                 spin_unlock(&dlm->spinlock);
 749                 spin_lock(&tmpres->spinlock);
 750
 751                 /*
 752                  * Right after dlm spinlock was released, dlm_thread could have
 753                  * purged the lockres. Check if lockres got unhashed. If so
 754                  * start over.
 755                  */
 756                 if (hlist_unhashed(&tmpres->hash_node)) {
 757                         spin_unlock(&tmpres->spinlock);
 758                         dlm_lockres_put(tmpres);
 759                         tmpres = NULL;
 760                         goto lookup;
 761                 }
 762
 763                 /* Wait on the thread that is mastering the resource */
 764                 if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
 765                         __dlm_wait_on_lockres(tmpres);
 766                         BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
 767                         spin_unlock(&tmpres->spinlock);
 768                         dlm_lockres_put(tmpres);
 769                         tmpres = NULL;
 770                         goto lookup;
 771                 }
 772
 773                 /* Wait on the resource purge to complete before continuing */
 774                 if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) {
 775                         BUG_ON(tmpres->owner == dlm->node_num);
 776                         __dlm_wait_on_lockres_flags(tmpres,
 777                                                     DLM_LOCK_RES_DROPPING_REF);
 778                         spin_unlock(&tmpres->spinlock);
 779                         dlm_lockres_put(tmpres);
 780                         tmpres = NULL;
 781                         goto lookup;
 782                 }
 783
 784                 /* Grab inflight ref to pin the resource */
 785                 dlm_lockres_grab_inflight_ref(dlm, tmpres);
 786
 787                 spin_unlock(&tmpres->spinlock);
 788                 if (res) {
 789                         spin_lock(&dlm->track_lock);
 790                         if (!list_empty(&res->tracking))
 791                                 list_del_init(&res->tracking);
 792                         else
 793                                 mlog(ML_ERROR, "Resource %.*s not "
 794                                                 "on the Tracking list\n",
 795                                                 res->lockname.len,
 796                                                 res->lockname.name);
 797                         spin_unlock(&dlm->track_lock);
 798                         dlm_lockres_put(res);
 799                 }
 800                 res = tmpres;
 801                 goto leave;
 802         }
 803
 804         if (!res) {
 805                 spin_unlock(&dlm->spinlock);
 806                 mlog(0, "allocating a new resource\n");
 807                 /* nothing found and we need to allocate one. */
 808                 alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
 809                 if (!alloc_mle)
 810                         goto leave;
 811                 res = dlm_new_lockres(dlm, lockid, namelen);
 812                 if (!res)
 813                         goto leave;
 814                 goto lookup;
 815         }
 816
 817         mlog(0, "no lockres found, allocated our own: %p\n", res);
 818
 819         if (flags & LKM_LOCAL) {
 820                 /* caller knows it's safe to assume it's not mastered elsewhere
 821                  * DONE!  return right away */
 822                 spin_lock(&res->spinlock);
 823                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
 824                 __dlm_insert_lockres(dlm, res);
 825                 dlm_lockres_grab_inflight_ref(dlm, res);
 826                 spin_unlock(&res->spinlock);
 827                 spin_unlock(&dlm->spinlock);
 828                 /* lockres still marked IN_PROGRESS */
 829                 goto wake_waiters;
 830         }
 831
 832         /* check master list to see if another node has started mastering it */
 833         spin_lock(&dlm->master_lock);
 834
 835         /* if we found a block, wait for lock to be mastered by another node */
 836         blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
 837         if (blocked) {
 838                 int mig;
 839                 if (mle->type == DLM_MLE_MASTER) {
 840                         mlog(ML_ERROR, "master entry for nonexistent lock!\n");
 841                         BUG();
 842                 }
 843                 mig = (mle->type == DLM_MLE_MIGRATION);
 844                 /* if there is a migration in progress, let the migration
 845                  * finish before continuing.  we can wait for the absence
 846                  * of the MIGRATION mle: either the migrate finished or
 847                  * one of the nodes died and the mle was cleaned up.
 848                  * if there is a BLOCK here, but it already has a master
 849                  * set, we are too late.  the master does not have a ref
 850                  * for us in the refmap.  detach the mle and drop it.
 851                  * either way, go back to the top and start over. */
 852                 if (mig || mle->master != O2NM_MAX_NODES) {
 853                         BUG_ON(mig && mle->master == dlm->node_num);
 854                         /* we arrived too late.  the master does not
 855                          * have a ref for us. retry. */
 856                         mlog(0, "%s:%.*s: late on %s\n",
 857                              dlm->name, namelen, lockid,
 858                              mig ?  "MIGRATION" : "BLOCK");
 859                         spin_unlock(&dlm->master_lock);
 860                         spin_unlock(&dlm->spinlock);
 861
 862                         /* master is known, detach */
 863                         if (!mig)
 864                                 dlm_mle_detach_hb_events(dlm, mle);
 865                         dlm_put_mle(mle);
 866                         mle = NULL;
 867                         /* this is lame, but we can't wait on either
 868                          * the mle or lockres waitqueue here */
 869                         if (mig)
 870                                 msleep(100);
 871                         goto lookup;
 872                 }
 873         } else {
 874                 /* go ahead and try to master lock on this node */
 875                 mle = alloc_mle;
 876                 /* make sure this does not get freed below */
 877                 alloc_mle = NULL;
 878                 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
 879                 set_bit(dlm->node_num, mle->maybe_map);
 880                 __dlm_insert_mle(dlm, mle);
 881
 882                 /* still holding the dlm spinlock, check the recovery map
 883                  * to see if there are any nodes that still need to be
 884                  * considered.  these will not appear in the mle nodemap
 885                  * but they might own this lockres.  wait on them. */
 886                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 887                 if (bit < O2NM_MAX_NODES) {
 888                         mlog(0, "%s: res %.*s, At least one node (%d) "
 889                              "to recover before lock mastery can begin\n",
 890                              dlm->name, namelen, (char *)lockid, bit);
 891                         wait_on_recovery = 1;
 892                 }
 893         }
 894
 895         /* at this point there is either a DLM_MLE_BLOCK or a
 896          * DLM_MLE_MASTER on the master list, so it's safe to add the
 897          * lockres to the hashtable.  anyone who finds the lock will
 898          * still have to wait on the IN_PROGRESS. */
 899
 900         /* finally add the lockres to its hash bucket */
 901         __dlm_insert_lockres(dlm, res);
 902
 903         /* since this lockres is new it doesn't not require the spinlock */
 904         __dlm_lockres_grab_inflight_ref(dlm, res);
 905
 906         /* get an extra ref on the mle in case this is a BLOCK
 907          * if so, the creator of the BLOCK may try to put the last
 908          * ref at this time in the assert master handler, so we
 909          * need an extra one to keep from a bad ptr deref. */
 910         dlm_get_mle_inuse(mle);
 911         spin_unlock(&dlm->master_lock);
 912         spin_unlock(&dlm->spinlock);
 913
 914 redo_request:
 915         while (wait_on_recovery) {
 916                 /* any cluster changes that occurred after dropping the
 917                  * dlm spinlock would be detectable be a change on the mle,
 918                  * so we only need to clear out the recovery map once. */
 919                 if (dlm_is_recovery_lock(lockid, namelen)) {
 920                         mlog(0, "%s: Recovery map is not empty, but must "
 921                              "master $RECOVERY lock now\n", dlm->name);
 922                         if (!dlm_pre_master_reco_lockres(dlm, res))
 923                                 wait_on_recovery = 0;
 924                         else {
 925                                 mlog(0, "%s: waiting 500ms for heartbeat state "
 926                                     "change\n", dlm->name);
 927                                 msleep(500);
 928                         }
 929                         continue;
 930                 }
 931
 932                 dlm_kick_recovery_thread(dlm);
 933                 msleep(1000);
 934                 dlm_wait_for_recovery(dlm);
 935
 936                 spin_lock(&dlm->spinlock);
 937                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 938                 if (bit < O2NM_MAX_NODES) {
 939                         mlog(0, "%s: res %.*s, At least one node (%d) "
 940                              "to recover before lock mastery can begin\n",
 941                              dlm->name, namelen, (char *)lockid, bit);
 942                         wait_on_recovery = 1;
 943                 } else
 944                         wait_on_recovery = 0;
 945                 spin_unlock(&dlm->spinlock);
 946
 947                 if (wait_on_recovery)
 948                         dlm_wait_for_node_recovery(dlm, bit, 10000);
 949         }
 950
 951         /* must wait for lock to be mastered elsewhere */
 952         if (blocked)
 953                 goto wait;
 954
 955         ret = -EINVAL;
 956         dlm_node_iter_init(mle->vote_map, &iter);
 957         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
 958                 ret = dlm_do_master_request(res, mle, nodenum);
 959                 if (ret < 0)
 960                         mlog_errno(ret);
 961                 if (mle->master != O2NM_MAX_NODES) {
 962                         /* found a master ! */
 963                         if (mle->master <= nodenum)
 964                                 break;
 965                         /* if our master request has not reached the master
 966                          * yet, keep going until it does.  this is how the
 967                          * master will know that asserts are needed back to
 968                          * the lower nodes. */
 969                         mlog(0, "%s: res %.*s, Requests only up to %u but "
 970                              "master is %u, keep going\n", dlm->name, namelen,
 971                              lockid, nodenum, mle->master);
 972                 }
 973         }
 974
 975 wait:
 976         /* keep going until the response map includes all nodes */
 977         ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
 978         if (ret < 0) {
 979                 wait_on_recovery = 1;
 980                 mlog(0, "%s: res %.*s, Node map changed, redo the master "
 981                      "request now, blocked=%d\n", dlm->name, res->lockname.len,
 982                      res->lockname.name, blocked);
 983                 if (++tries > 20) {
 984                         mlog(ML_ERROR, "%s: res %.*s, Spinning on "
 985                              "dlm_wait_for_lock_mastery, blocked = %d\n",
 986                              dlm->name, res->lockname.len,
 987                              res->lockname.name, blocked);
 988                         dlm_print_one_lock_resource(res);
 989                         dlm_print_one_mle(mle);
 990                         tries = 0;
 991                 }
 992                 goto redo_request;
 993         }
 994
 995         mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len,
 996              res->lockname.name, res->owner);
 997         /* make sure we never continue without this */
 998         BUG_ON(res->owner == O2NM_MAX_NODES);
 999
1000         /* master is known, detach if not already detached */
1001         dlm_mle_detach_hb_events(dlm, mle);
1002         dlm_put_mle(mle);
1003         /* put the extra ref */
1004         dlm_put_mle_inuse(mle);
1005
1006 wake_waiters:
1007         spin_lock(&res->spinlock);
1008         res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1009         spin_unlock(&res->spinlock);
1010         wake_up(&res->wq);
1011
1012 leave:
1013         /* need to free the unused mle */
1014         if (alloc_mle)
1015                 kmem_cache_free(dlm_mle_cache, alloc_mle);
1016
1017         return res;
1018 }
1019
1020
1021 #define DLM_MASTERY_TIMEOUT_MS   5000
1022
1023 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
1024                                      struct dlm_lock_resource *res,
1025                                      struct dlm_master_list_entry *mle,
1026                                      int *blocked)
1027 {
1028         u8 m;
1029         int ret, bit;
1030         int map_changed, voting_done;
1031         int assert, sleep;
1032
1033 recheck:
1034         ret = 0;
1035         assert = 0;
1036
1037         /* check if another node has already become the owner */
1038         spin_lock(&res->spinlock);
1039         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1040                 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1041                      res->lockname.len, res->lockname.name, res->owner);
1042                 spin_unlock(&res->spinlock);
1043                 /* this will cause the master to re-assert across
1044                  * the whole cluster, freeing up mles */
1045                 if (res->owner != dlm->node_num) {
1046                         ret = dlm_do_master_request(res, mle, res->owner);
1047                         if (ret < 0) {
1048                                 /* give recovery a chance to run */
1049                                 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1050                                 msleep(500);
1051                                 goto recheck;
1052                         }
1053                 }
1054                 ret = 0;
1055                 goto leave;
1056         }
1057         spin_unlock(&res->spinlock);
1058
1059         spin_lock(&mle->spinlock);
1060         m = mle->master;
1061         map_changed = (memcmp(mle->vote_map, mle->node_map,
1062                               sizeof(mle->vote_map)) != 0);
1063         voting_done = (memcmp(mle->vote_map, mle->response_map,
1064                              sizeof(mle->vote_map)) == 0);
1065
1066         /* restart if we hit any errors */
1067         if (map_changed) {
1068                 int b;
1069                 mlog(0, "%s: %.*s: node map changed, restarting\n",
1070                      dlm->name, res->lockname.len, res->lockname.name);
1071                 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1072                 b = (mle->type == DLM_MLE_BLOCK);
1073                 if ((*blocked && !b) || (!*blocked && b)) {
1074                         mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1075                              dlm->name, res->lockname.len, res->lockname.name,
1076                              *blocked, b);
1077                         *blocked = b;
1078                 }
1079                 spin_unlock(&mle->spinlock);
1080                 if (ret < 0) {
1081                         mlog_errno(ret);
1082                         goto leave;
1083                 }
1084                 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1085                      "rechecking now\n", dlm->name, res->lockname.len,
1086                      res->lockname.name);
1087                 goto recheck;
1088         } else {
1089                 if (!voting_done) {
1090                         mlog(0, "map not changed and voting not done "
1091                              "for %s:%.*s\n", dlm->name, res->lockname.len,
1092                              res->lockname.name);
1093                 }
1094         }
1095
1096         if (m != O2NM_MAX_NODES) {
1097                 /* another node has done an assert!
1098                  * all done! */
1099                 sleep = 0;
1100         } else {
1101                 sleep = 1;
1102                 /* have all nodes responded? */
1103                 if (voting_done && !*blocked) {
1104                         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1105                         if (dlm->node_num <= bit) {
1106                                 /* my node number is lowest.
1107                                  * now tell other nodes that I am
1108                                  * mastering this. */
1109                                 mle->master = dlm->node_num;
1110                                 /* ref was grabbed in get_lock_resource
1111                                  * will be dropped in dlmlock_master */
1112                                 assert = 1;
1113                                 sleep = 0;
1114                         }
1115                         /* if voting is done, but we have not received
1116                          * an assert master yet, we must sleep */
1117                 }
1118         }
1119
1120         spin_unlock(&mle->spinlock);
1121
1122         /* sleep if we haven't finished voting yet */
1123         if (sleep) {
1124                 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1125
1126                 /*
1127                 if (kref_read(&mle->mle_refs) < 2)
1128                         mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1129                         kref_read(&mle->mle_refs),
1130                         res->lockname.len, res->lockname.name);
1131                 */
1132                 atomic_set(&mle->woken, 0);
1133                 (void)wait_event_timeout(mle->wq,
1134                                          (atomic_read(&mle->woken) == 1),
1135                                          timeo);
1136                 if (res->owner == O2NM_MAX_NODES) {
1137                         mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1138                              res->lockname.len, res->lockname.name);
1139                         goto recheck;
1140                 }
1141                 mlog(0, "done waiting, master is %u\n", res->owner);
1142                 ret = 0;
1143                 goto leave;
1144         }
1145
1146         ret = 0;   /* done */
1147         if (assert) {
1148                 m = dlm->node_num;
1149                 mlog(0, "about to master %.*s here, this=%u\n",
1150                      res->lockname.len, res->lockname.name, m);
1151                 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1152                 if (ret) {
1153                         /* This is a failure in the network path,
1154                          * not in the response to the assert_master
1155                          * (any nonzero response is a BUG on this node).
1156                          * Most likely a socket just got disconnected
1157                          * due to node death. */
1158                         mlog_errno(ret);
1159                 }
1160                 /* no longer need to restart lock mastery.
1161                  * all living nodes have been contacted. */
1162                 ret = 0;
1163         }
1164
1165         /* set the lockres owner */
1166         spin_lock(&res->spinlock);
1167         /* mastery reference obtained either during
1168          * assert_master_handler or in get_lock_resource */
1169         dlm_change_lockres_owner(dlm, res, m);
1170         spin_unlock(&res->spinlock);
1171
1172 leave:
1173         return ret;
1174 }
1175
1176 struct dlm_bitmap_diff_iter
1177 {
1178         int curnode;
1179         unsigned long *orig_bm;
1180         unsigned long *cur_bm;
1181         unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1182 };
1183
1184 enum dlm_node_state_change
1185 {
1186         NODE_DOWN = -1,
1187         NODE_NO_CHANGE = 0,
1188         NODE_UP
1189 };
1190
1191 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1192                                       unsigned long *orig_bm,
1193                                       unsigned long *cur_bm)
1194 {
1195         unsigned long p1, p2;
1196         int i;
1197
1198         iter->curnode = -1;
1199         iter->orig_bm = orig_bm;
1200         iter->cur_bm = cur_bm;
1201
1202         for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1203                 p1 = *(iter->orig_bm + i);
1204                 p2 = *(iter->cur_bm + i);
1205                 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1206         }
1207 }
1208
1209 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1210                                      enum dlm_node_state_change *state)
1211 {
1212         int bit;
1213
1214         if (iter->curnode >= O2NM_MAX_NODES)
1215                 return -ENOENT;
1216
1217         bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1218                             iter->curnode+1);
1219         if (bit >= O2NM_MAX_NODES) {
1220                 iter->curnode = O2NM_MAX_NODES;
1221                 return -ENOENT;
1222         }
1223
1224         /* if it was there in the original then this node died */
1225         if (test_bit(bit, iter->orig_bm))
1226                 *state = NODE_DOWN;
1227         else
1228                 *state = NODE_UP;
1229
1230         iter->curnode = bit;
1231         return bit;
1232 }
1233
1234
1235 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1236                                     struct dlm_lock_resource *res,
1237                                     struct dlm_master_list_entry *mle,
1238                                     int blocked)
1239 {
1240         struct dlm_bitmap_diff_iter bdi;
1241         enum dlm_node_state_change sc;
1242         int node;
1243         int ret = 0;
1244
1245         mlog(0, "something happened such that the "
1246              "master process may need to be restarted!\n");
1247
1248         assert_spin_locked(&mle->spinlock);
1249
1250         dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1251         node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1252         while (node >= 0) {
1253                 if (sc == NODE_UP) {
1254                         /* a node came up.  clear any old vote from
1255                          * the response map and set it in the vote map
1256                          * then restart the mastery. */
1257                         mlog(ML_NOTICE, "node %d up while restarting\n", node);
1258
1259                         /* redo the master request, but only for the new node */
1260                         mlog(0, "sending request to new node\n");
1261                         clear_bit(node, mle->response_map);
1262                         set_bit(node, mle->vote_map);
1263                 } else {
1264                         mlog(ML_ERROR, "node down! %d\n", node);
1265                         if (blocked) {
1266                                 int lowest = find_next_bit(mle->maybe_map,
1267                                                        O2NM_MAX_NODES, 0);
1268
1269                                 /* act like it was never there */
1270                                 clear_bit(node, mle->maybe_map);
1271
1272                                 if (node == lowest) {
1273                                         mlog(0, "expected master %u died"
1274                                             " while this node was blocked "
1275                                             "waiting on it!\n", node);
1276                                         lowest = find_next_bit(mle->maybe_map,
1277                                                         O2NM_MAX_NODES,
1278                                                         lowest+1);
1279                                         if (lowest < O2NM_MAX_NODES) {
1280                                                 mlog(0, "%s:%.*s:still "
1281                                                      "blocked. waiting on %u "
1282                                                      "now\n", dlm->name,
1283                                                      res->lockname.len,
1284                                                      res->lockname.name,
1285                                                      lowest);
1286                                         } else {
1287                                                 /* mle is an MLE_BLOCK, but
1288                                                  * there is now nothing left to
1289                                                  * block on.  we need to return
1290                                                  * all the way back out and try
1291                                                  * again with an MLE_MASTER.
1292                                                  * dlm_do_local_recovery_cleanup
1293                                                  * has already run, so the mle
1294                                                  * refcount is ok */
1295                                                 mlog(0, "%s:%.*s: no "
1296                                                      "longer blocking. try to "
1297                                                      "master this here\n",
1298                                                      dlm->name,
1299                                                      res->lockname.len,
1300                                                      res->lockname.name);
1301                                                 mle->type = DLM_MLE_MASTER;
1302                                                 mle->mleres = res;
1303                                         }
1304                                 }
1305                         }
1306
1307                         /* now blank out everything, as if we had never
1308                          * contacted anyone */
1309                         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1310                         memset(mle->response_map, 0, sizeof(mle->response_map));
1311                         /* reset the vote_map to the current node_map */
1312                         memcpy(mle->vote_map, mle->node_map,
1313                                sizeof(mle->node_map));
1314                         /* put myself into the maybe map */
1315                         if (mle->type != DLM_MLE_BLOCK)
1316                                 set_bit(dlm->node_num, mle->maybe_map);
1317                 }
1318                 ret = -EAGAIN;
1319                 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1320         }
1321         return ret;
1322 }
1323
1324
1325 /*
1326  * DLM_MASTER_REQUEST_MSG
1327  *
1328  * returns: 0 on success,
1329  *          -errno on a network error
1330  *
1331  * on error, the caller should assume the target node is "dead"
1332  *
1333  */
1334
1335 static int dlm_do_master_request(struct dlm_lock_resource *res,
1336                                  struct dlm_master_list_entry *mle, int to)
1337 {
1338         struct dlm_ctxt *dlm = mle->dlm;
1339         struct dlm_master_request request;
1340         int ret, response=0, resend;
1341
1342         memset(&request, 0, sizeof(request));
1343         request.node_idx = dlm->node_num;
1344
1345         BUG_ON(mle->type == DLM_MLE_MIGRATION);
1346
1347         request.namelen = (u8)mle->mnamelen;
1348         memcpy(request.name, mle->mname, request.namelen);
1349
1350 again:
1351         ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1352                                  sizeof(request), to, &response);
1353         if (ret < 0)  {
1354                 if (ret == -ESRCH) {
1355                         /* should never happen */
1356                         mlog(ML_ERROR, "TCP stack not ready!\n");
1357                         BUG();
1358                 } else if (ret == -EINVAL) {
1359                         mlog(ML_ERROR, "bad args passed to o2net!\n");
1360                         BUG();
1361                 } else if (ret == -ENOMEM) {
1362                         mlog(ML_ERROR, "out of memory while trying to send "
1363                              "network message!  retrying\n");
1364                         /* this is totally crude */
1365                         msleep(50);
1366                         goto again;
1367                 } else if (!dlm_is_host_down(ret)) {
1368                         /* not a network error. bad. */
1369                         mlog_errno(ret);
1370                         mlog(ML_ERROR, "unhandled error!");
1371                         BUG();
1372                 }
1373                 /* all other errors should be network errors,
1374                  * and likely indicate node death */
1375                 mlog(ML_ERROR, "link to %d went down!\n", to);
1376                 goto out;
1377         }
1378
1379         ret = 0;
1380         resend = 0;
1381         spin_lock(&mle->spinlock);
1382         switch (response) {
1383                 case DLM_MASTER_RESP_YES:
1384                         set_bit(to, mle->response_map);
1385                         mlog(0, "node %u is the master, response=YES\n", to);
1386                         mlog(0, "%s:%.*s: master node %u now knows I have a "
1387                              "reference\n", dlm->name, res->lockname.len,
1388                              res->lockname.name, to);
1389                         mle->master = to;
1390                         break;
1391                 case DLM_MASTER_RESP_NO:
1392                         mlog(0, "node %u not master, response=NO\n", to);
1393                         set_bit(to, mle->response_map);
1394                         break;
1395                 case DLM_MASTER_RESP_MAYBE:
1396                         mlog(0, "node %u not master, response=MAYBE\n", to);
1397                         set_bit(to, mle->response_map);
1398                         set_bit(to, mle->maybe_map);
1399                         break;
1400                 case DLM_MASTER_RESP_ERROR:
1401                         mlog(0, "node %u hit an error, resending\n", to);
1402                         resend = 1;
1403                         response = 0;
1404                         break;
1405                 default:
1406                         mlog(ML_ERROR, "bad response! %u\n", response);
1407                         BUG();
1408         }
1409         spin_unlock(&mle->spinlock);
1410         if (resend) {
1411                 /* this is also totally crude */
1412                 msleep(50);
1413                 goto again;
1414         }
1415
1416 out:
1417         return ret;
1418 }
1419
1420 /*
1421  * locks that can be taken here:
1422  * dlm->spinlock
1423  * res->spinlock
1424  * mle->spinlock
1425  * dlm->master_list
1426  *
1427  * if possible, TRIM THIS DOWN!!!
1428  */
1429 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1430                                void **ret_data)
1431 {
1432         u8 response = DLM_MASTER_RESP_MAYBE;
1433         struct dlm_ctxt *dlm = data;
1434         struct dlm_lock_resource *res = NULL;
1435         struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1436         struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1437         char *name;
1438         unsigned int namelen, hash;
1439         int found, ret;
1440         int set_maybe;
1441         int dispatch_assert = 0;
1442         int dispatched = 0;
1443
1444         if (!dlm_grab(dlm))
1445                 return DLM_MASTER_RESP_NO;
1446
1447         if (!dlm_domain_fully_joined(dlm)) {
1448                 response = DLM_MASTER_RESP_NO;
1449                 goto send_response;
1450         }
1451
1452         name = request->name;
1453         namelen = request->namelen;
1454         hash = dlm_lockid_hash(name, namelen);
1455
1456         if (namelen > DLM_LOCKID_NAME_MAX) {
1457                 response = DLM_IVBUFLEN;
1458                 goto send_response;
1459         }
1460
1461 way_up_top:
1462         spin_lock(&dlm->spinlock);
1463         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1464         if (res) {
1465                 spin_unlock(&dlm->spinlock);
1466
1467                 /* take care of the easy cases up front */
1468                 spin_lock(&res->spinlock);
1469
1470                 /*
1471                  * Right after dlm spinlock was released, dlm_thread could have
1472                  * purged the lockres. Check if lockres got unhashed. If so
1473                  * start over.
1474                  */
1475                 if (hlist_unhashed(&res->hash_node)) {
1476                         spin_unlock(&res->spinlock);
1477                         dlm_lockres_put(res);
1478                         goto way_up_top;
1479                 }
1480
1481                 if (res->state & (DLM_LOCK_RES_RECOVERING|
1482                                   DLM_LOCK_RES_MIGRATING)) {
1483                         spin_unlock(&res->spinlock);
1484                         mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1485                              "being recovered/migrated\n");
1486                         response = DLM_MASTER_RESP_ERROR;
1487                         if (mle)
1488                                 kmem_cache_free(dlm_mle_cache, mle);
1489                         goto send_response;
1490                 }
1491
1492                 if (res->owner == dlm->node_num) {
1493                         dlm_lockres_set_refmap_bit(dlm, res, request->node_idx);
1494                         spin_unlock(&res->spinlock);
1495                         response = DLM_MASTER_RESP_YES;
1496                         if (mle)
1497                                 kmem_cache_free(dlm_mle_cache, mle);
1498
1499                         /* this node is the owner.
1500                          * there is some extra work that needs to
1501                          * happen now.  the requesting node has
1502                          * caused all nodes up to this one to
1503                          * create mles.  this node now needs to
1504                          * go back and clean those up. */
1505                         dispatch_assert = 1;
1506                         goto send_response;
1507                 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1508                         spin_unlock(&res->spinlock);
1509                         // mlog(0, "node %u is the master\n", res->owner);
1510                         response = DLM_MASTER_RESP_NO;
1511                         if (mle)
1512                                 kmem_cache_free(dlm_mle_cache, mle);
1513                         goto send_response;
1514                 }
1515
1516                 /* ok, there is no owner.  either this node is
1517                  * being blocked, or it is actively trying to
1518                  * master this lock. */
1519                 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1520                         mlog(ML_ERROR, "lock with no owner should be "
1521                              "in-progress!\n");
1522                         BUG();
1523                 }
1524
1525                 // mlog(0, "lockres is in progress...\n");
1526                 spin_lock(&dlm->master_lock);
1527                 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1528                 if (!found) {
1529                         mlog(ML_ERROR, "no mle found for this lock!\n");
1530                         BUG();
1531                 }
1532                 set_maybe = 1;
1533                 spin_lock(&tmpmle->spinlock);
1534                 if (tmpmle->type == DLM_MLE_BLOCK) {
1535                         // mlog(0, "this node is waiting for "
1536                         // "lockres to be mastered\n");
1537                         response = DLM_MASTER_RESP_NO;
1538                 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1539                         mlog(0, "node %u is master, but trying to migrate to "
1540                              "node %u.\n", tmpmle->master, tmpmle->new_master);
1541                         if (tmpmle->master == dlm->node_num) {
1542                                 mlog(ML_ERROR, "no owner on lockres, but this "
1543                                      "node is trying to migrate it to %u?!\n",
1544                                      tmpmle->new_master);
1545                                 BUG();
1546                         } else {
1547                                 /* the real master can respond on its own */
1548                                 response = DLM_MASTER_RESP_NO;
1549                         }
1550                 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1551                         set_maybe = 0;
1552                         if (tmpmle->master == dlm->node_num) {
1553                                 response = DLM_MASTER_RESP_YES;
1554                                 /* this node will be the owner.
1555                                  * go back and clean the mles on any
1556                                  * other nodes */
1557                                 dispatch_assert = 1;
1558                                 dlm_lockres_set_refmap_bit(dlm, res,
1559                                                            request->node_idx);
1560                         } else
1561                                 response = DLM_MASTER_RESP_NO;
1562                 } else {
1563                         // mlog(0, "this node is attempting to "
1564                         // "master lockres\n");
1565                         response = DLM_MASTER_RESP_MAYBE;
1566                 }
1567                 if (set_maybe)
1568                         set_bit(request->node_idx, tmpmle->maybe_map);
1569                 spin_unlock(&tmpmle->spinlock);
1570
1571                 spin_unlock(&dlm->master_lock);
1572                 spin_unlock(&res->spinlock);
1573
1574                 /* keep the mle attached to heartbeat events */
1575                 dlm_put_mle(tmpmle);
1576                 if (mle)
1577                         kmem_cache_free(dlm_mle_cache, mle);
1578                 goto send_response;
1579         }
1580
1581         /*
1582          * lockres doesn't exist on this node
1583          * if there is an MLE_BLOCK, return NO
1584          * if there is an MLE_MASTER, return MAYBE
1585          * otherwise, add an MLE_BLOCK, return NO
1586          */
1587         spin_lock(&dlm->master_lock);
1588         found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1589         if (!found) {
1590                 /* this lockid has never been seen on this node yet */
1591                 // mlog(0, "no mle found\n");
1592                 if (!mle) {
1593                         spin_unlock(&dlm->master_lock);
1594                         spin_unlock(&dlm->spinlock);
1595
1596                         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1597                         if (!mle) {
1598                                 response = DLM_MASTER_RESP_ERROR;
1599                                 mlog_errno(-ENOMEM);
1600                                 goto send_response;
1601                         }
1602                         goto way_up_top;
1603                 }
1604
1605                 // mlog(0, "this is second time thru, already allocated, "
1606                 // "add the block.\n");
1607                 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1608                 set_bit(request->node_idx, mle->maybe_map);
1609                 __dlm_insert_mle(dlm, mle);
1610                 response = DLM_MASTER_RESP_NO;
1611         } else {
1612                 spin_lock(&tmpmle->spinlock);
1613                 if (tmpmle->master == dlm->node_num) {
1614                         mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1615                         BUG();
1616                 }
1617                 if (tmpmle->type == DLM_MLE_BLOCK)
1618                         response = DLM_MASTER_RESP_NO;
1619                 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1620                         mlog(0, "migration mle was found (%u->%u)\n",
1621                              tmpmle->master, tmpmle->new_master);
1622                         /* real master can respond on its own */
1623                         response = DLM_MASTER_RESP_NO;
1624                 } else
1625                         response = DLM_MASTER_RESP_MAYBE;
1626                 set_bit(request->node_idx, tmpmle->maybe_map);
1627                 spin_unlock(&tmpmle->spinlock);
1628         }
1629         spin_unlock(&dlm->master_lock);
1630         spin_unlock(&dlm->spinlock);
1631
1632         if (found) {
1633                 /* keep the mle attached to heartbeat events */
1634                 dlm_put_mle(tmpmle);
1635         }
1636 send_response:
1637         /*
1638          * __dlm_lookup_lockres() grabbed a reference to this lockres.
1639          * The reference is released by dlm_assert_master_worker() under
1640          * the call to dlm_dispatch_assert_master().  If
1641          * dlm_assert_master_worker() isn't called, we drop it here.
1642          */
1643         if (dispatch_assert) {
1644                 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1645                              dlm->node_num, res->lockname.len, res->lockname.name);
1646                 spin_lock(&res->spinlock);
1647                 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1648                                                  DLM_ASSERT_MASTER_MLE_CLEANUP);
1649                 if (ret < 0) {
1650                         mlog(ML_ERROR, "failed to dispatch assert master work\n");
1651                         response = DLM_MASTER_RESP_ERROR;
1652                         spin_unlock(&res->spinlock);
1653                         dlm_lockres_put(res);
1654                 } else {
1655                         dispatched = 1;
1656                         __dlm_lockres_grab_inflight_worker(dlm, res);
1657                         spin_unlock(&res->spinlock);
1658                 }
1659         } else {
1660                 if (res)
1661                         dlm_lockres_put(res);
1662         }
1663
1664         if (!dispatched)
1665                 dlm_put(dlm);
1666         return response;
1667 }
1668
1669 /*
1670  * DLM_ASSERT_MASTER_MSG
1671  */
1672
1673
1674 /*
1675  * NOTE: this can be used for debugging
1676  * can periodically run all locks owned by this node
1677  * and re-assert across the cluster...
1678  */
1679 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1680                                 struct dlm_lock_resource *res,
1681                                 void *nodemap, u32 flags)
1682 {
1683         struct dlm_assert_master assert;
1684         int to, tmpret;
1685         struct dlm_node_iter iter;
1686         int ret = 0;
1687         int reassert;
1688         const char *lockname = res->lockname.name;
1689         unsigned int namelen = res->lockname.len;
1690
1691         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1692
1693         spin_lock(&res->spinlock);
1694         res->state |= DLM_LOCK_RES_SETREF_INPROG;
1695         spin_unlock(&res->spinlock);
1696
1697 again:
1698         reassert = 0;
1699
1700         /* note that if this nodemap is empty, it returns 0 */
1701         dlm_node_iter_init(nodemap, &iter);
1702         while ((to = dlm_node_iter_next(&iter)) >= 0) {
1703                 int r = 0;
1704                 struct dlm_master_list_entry *mle = NULL;
1705
1706                 mlog(0, "sending assert master to %d (%.*s)\n", to,
1707                      namelen, lockname);
1708                 memset(&assert, 0, sizeof(assert));
1709                 assert.node_idx = dlm->node_num;
1710                 assert.namelen = namelen;
1711                 memcpy(assert.name, lockname, namelen);
1712                 assert.flags = cpu_to_be32(flags);
1713
1714                 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1715                                             &assert, sizeof(assert), to, &r);
1716                 if (tmpret < 0) {
1717                         mlog(ML_ERROR, "Error %d when sending message %u (key "
1718                              "0x%x) to node %u\n", tmpret,
1719                              DLM_ASSERT_MASTER_MSG, dlm->key, to);
1720                         if (!dlm_is_host_down(tmpret)) {
1721                                 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1722                                 BUG();
1723                         }
1724                         /* a node died.  finish out the rest of the nodes. */
1725                         mlog(0, "link to %d went down!\n", to);
1726                         /* any nonzero status return will do */
1727                         ret = tmpret;
1728                         r = 0;
1729                 } else if (r < 0) {
1730                         /* ok, something horribly messed.  kill thyself. */
1731                         mlog(ML_ERROR,"during assert master of %.*s to %u, "
1732                              "got %d.\n", namelen, lockname, to, r);
1733                         spin_lock(&dlm->spinlock);
1734                         spin_lock(&dlm->master_lock);
1735                         if (dlm_find_mle(dlm, &mle, (char *)lockname,
1736                                          namelen)) {
1737                                 dlm_print_one_mle(mle);
1738                                 __dlm_put_mle(mle);
1739                         }
1740                         spin_unlock(&dlm->master_lock);
1741                         spin_unlock(&dlm->spinlock);
1742                         BUG();
1743                 }
1744
1745                 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1746                     !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1747                                 mlog(ML_ERROR, "%.*s: very strange, "
1748                                      "master MLE but no lockres on %u\n",
1749                                      namelen, lockname, to);
1750                 }
1751
1752                 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1753                         mlog(0, "%.*s: node %u create mles on other "
1754                              "nodes and requests a re-assert\n",
1755                              namelen, lockname, to);
1756                         reassert = 1;
1757                 }
1758                 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1759                         mlog(0, "%.*s: node %u has a reference to this "
1760                              "lockres, set the bit in the refmap\n",
1761                              namelen, lockname, to);
1762                         spin_lock(&res->spinlock);
1763                         dlm_lockres_set_refmap_bit(dlm, res, to);
1764                         spin_unlock(&res->spinlock);
1765                 }
1766         }
1767
1768         if (reassert)
1769                 goto again;
1770
1771         spin_lock(&res->spinlock);
1772         res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1773         spin_unlock(&res->spinlock);
1774         wake_up(&res->wq);
1775
1776         return ret;
1777 }
1778
1779 /*
1780  * locks that can be taken here:
1781  * dlm->spinlock
1782  * res->spinlock
1783  * mle->spinlock
1784  * dlm->master_list
1785  *
1786  * if possible, TRIM THIS DOWN!!!
1787  */
1788 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1789                               void **ret_data)
1790 {
1791         struct dlm_ctxt *dlm = data;
1792         struct dlm_master_list_entry *mle = NULL;
1793         struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1794         struct dlm_lock_resource *res = NULL;
1795         char *name;
1796         unsigned int namelen, hash;
1797         u32 flags;
1798         int master_request = 0, have_lockres_ref = 0;
1799         int ret = 0;
1800
1801         if (!dlm_grab(dlm))
1802                 return 0;
1803
1804         name = assert->name;
1805         namelen = assert->namelen;
1806         hash = dlm_lockid_hash(name, namelen);
1807         flags = be32_to_cpu(assert->flags);
1808
1809         if (namelen > DLM_LOCKID_NAME_MAX) {
1810                 mlog(ML_ERROR, "Invalid name length!");
1811                 goto done;
1812         }
1813
1814         spin_lock(&dlm->spinlock);
1815
1816         if (flags)
1817                 mlog(0, "assert_master with flags: %u\n", flags);
1818
1819         /* find the MLE */
1820         spin_lock(&dlm->master_lock);
1821         if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1822                 /* not an error, could be master just re-asserting */
1823                 mlog(0, "just got an assert_master from %u, but no "
1824                      "MLE for it! (%.*s)\n", assert->node_idx,
1825                      namelen, name);
1826         } else {
1827                 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1828                 if (bit >= O2NM_MAX_NODES) {
1829                         /* not necessarily an error, though less likely.
1830                          * could be master just re-asserting. */
1831                         mlog(0, "no bits set in the maybe_map, but %u "
1832                              "is asserting! (%.*s)\n", assert->node_idx,
1833                              namelen, name);
1834                 } else if (bit != assert->node_idx) {
1835                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1836                                 mlog(0, "master %u was found, %u should "
1837                                      "back off\n", assert->node_idx, bit);
1838                         } else {
1839                                 /* with the fix for bug 569, a higher node
1840                                  * number winning the mastery will respond
1841                                  * YES to mastery requests, but this node
1842                                  * had no way of knowing.  let it pass. */
1843                                 mlog(0, "%u is the lowest node, "
1844                                      "%u is asserting. (%.*s)  %u must "
1845                                      "have begun after %u won.\n", bit,
1846                                      assert->node_idx, namelen, name, bit,
1847                                      assert->node_idx);
1848                         }
1849                 }
1850                 if (mle->type == DLM_MLE_MIGRATION) {
1851                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1852                                 mlog(0, "%s:%.*s: got cleanup assert"
1853                                      " from %u for migration\n",
1854                                      dlm->name, namelen, name,
1855                                      assert->node_idx);
1856                         } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1857                                 mlog(0, "%s:%.*s: got unrelated assert"
1858                                      " from %u for migration, ignoring\n",
1859                                      dlm->name, namelen, name,
1860                                      assert->node_idx);
1861                                 __dlm_put_mle(mle);
1862                                 spin_unlock(&dlm->master_lock);
1863                                 spin_unlock(&dlm->spinlock);
1864                                 goto done;
1865                         }
1866                 }
1867         }
1868         spin_unlock(&dlm->master_lock);
1869
1870         /* ok everything checks out with the MLE
1871          * now check to see if there is a lockres */
1872         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1873         if (res) {
1874                 spin_lock(&res->spinlock);
1875                 if (res->state & DLM_LOCK_RES_RECOVERING)  {
1876                         mlog(ML_ERROR, "%u asserting but %.*s is "
1877                              "RECOVERING!\n", assert->node_idx, namelen, name);
1878                         goto kill;
1879                 }
1880                 if (!mle) {
1881                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1882                             res->owner != assert->node_idx) {
1883                                 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1884                                      "but current owner is %u! (%.*s)\n",
1885                                      assert->node_idx, res->owner, namelen,
1886                                      name);
1887                                 __dlm_print_one_lock_resource(res);
1888                                 BUG();
1889                         }
1890                 } else if (mle->type != DLM_MLE_MIGRATION) {
1891                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1892                                 /* owner is just re-asserting */
1893                                 if (res->owner == assert->node_idx) {
1894                                         mlog(0, "owner %u re-asserting on "
1895                                              "lock %.*s\n", assert->node_idx,
1896                                              namelen, name);
1897                                         goto ok;
1898                                 }
1899                                 mlog(ML_ERROR, "got assert_master from "
1900                                      "node %u, but %u is the owner! "
1901                                      "(%.*s)\n", assert->node_idx,
1902                                      res->owner, namelen, name);
1903                                 goto kill;
1904                         }
1905                         if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1906                                 mlog(ML_ERROR, "got assert from %u, but lock "
1907                                      "with no owner should be "
1908                                      "in-progress! (%.*s)\n",
1909                                      assert->node_idx,
1910                                      namelen, name);
1911                                 goto kill;
1912                         }
1913                 } else /* mle->type == DLM_MLE_MIGRATION */ {
1914                         /* should only be getting an assert from new master */
1915                         if (assert->node_idx != mle->new_master) {
1916                                 mlog(ML_ERROR, "got assert from %u, but "
1917                                      "new master is %u, and old master "
1918                                      "was %u (%.*s)\n",
1919                                      assert->node_idx, mle->new_master,
1920                                      mle->master, namelen, name);
1921                                 goto kill;
1922                         }
1923
1924                 }
1925 ok:
1926                 spin_unlock(&res->spinlock);
1927         }
1928
1929         // mlog(0, "woo!  got an assert_master from node %u!\n",
1930         //           assert->node_idx);
1931         if (mle) {
1932                 int extra_ref = 0;
1933                 int nn = -1;
1934                 int rr, err = 0;
1935
1936                 spin_lock(&mle->spinlock);
1937                 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1938                         extra_ref = 1;
1939                 else {
1940                         /* MASTER mle: if any bits set in the response map
1941                          * then the calling node needs to re-assert to clear
1942                          * up nodes that this node contacted */
1943                         while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1944                                                     nn+1)) < O2NM_MAX_NODES) {
1945                                 if (nn != dlm->node_num && nn != assert->node_idx) {
1946                                         master_request = 1;
1947                                         break;
1948                                 }
1949                         }
1950                 }
1951                 mle->master = assert->node_idx;
1952                 atomic_set(&mle->woken, 1);
1953                 wake_up(&mle->wq);
1954                 spin_unlock(&mle->spinlock);
1955
1956                 if (res) {
1957                         int wake = 0;
1958                         spin_lock(&res->spinlock);
1959                         if (mle->type == DLM_MLE_MIGRATION) {
1960                                 mlog(0, "finishing off migration of lockres %.*s, "
1961                                         "from %u to %u\n",
1962                                         res->lockname.len, res->lockname.name,
1963                                         dlm->node_num, mle->new_master);
1964                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
1965                                 wake = 1;
1966                                 dlm_change_lockres_owner(dlm, res, mle->new_master);
1967                                 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1968                         } else {
1969                                 dlm_change_lockres_owner(dlm, res, mle->master);
1970                         }
1971                         spin_unlock(&res->spinlock);
1972                         have_lockres_ref = 1;
1973                         if (wake)
1974                                 wake_up(&res->wq);
1975                 }
1976
1977                 /* master is known, detach if not already detached.
1978                  * ensures that only one assert_master call will happen
1979                  * on this mle. */
1980                 spin_lock(&dlm->master_lock);
1981
1982                 rr = kref_read(&mle->mle_refs);
1983                 if (mle->inuse > 0) {
1984                         if (extra_ref && rr < 3)
1985                                 err = 1;
1986                         else if (!extra_ref && rr < 2)
1987                                 err = 1;
1988                 } else {
1989                         if (extra_ref && rr < 2)
1990                                 err = 1;
1991                         else if (!extra_ref && rr < 1)
1992                                 err = 1;
1993                 }
1994                 if (err) {
1995                         mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
1996                              "that will mess up this node, refs=%d, extra=%d, "
1997                              "inuse=%d\n", dlm->name, namelen, name,
1998                              assert->node_idx, rr, extra_ref, mle->inuse);
1999                         dlm_print_one_mle(mle);
2000                 }
2001                 __dlm_unlink_mle(dlm, mle);
2002                 __dlm_mle_detach_hb_events(dlm, mle);
2003                 __dlm_put_mle(mle);
2004                 if (extra_ref) {
2005                         /* the assert master message now balances the extra
2006                          * ref given by the master / migration request message.
2007                          * if this is the last put, it will be removed
2008                          * from the list. */
2009                         __dlm_put_mle(mle);
2010                 }
2011                 spin_unlock(&dlm->master_lock);
2012         } else if (res) {
2013                 if (res->owner != assert->node_idx) {
2014                         mlog(0, "assert_master from %u, but current "
2015                              "owner is %u (%.*s), no mle\n", assert->node_idx,
2016                              res->owner, namelen, name);
2017                 }
2018         }
2019         spin_unlock(&dlm->spinlock);
2020
2021 done:
2022         ret = 0;
2023         if (res) {
2024                 spin_lock(&res->spinlock);
2025                 res->state |= DLM_LOCK_RES_SETREF_INPROG;
2026                 spin_unlock(&res->spinlock);
2027                 *ret_data = (void *)res;
2028         }
2029         dlm_put(dlm);
2030         if (master_request) {
2031                 mlog(0, "need to tell master to reassert\n");
2032                 /* positive. negative would shoot down the node. */
2033                 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2034                 if (!have_lockres_ref) {
2035                         mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2036                              "mle present here for %s:%.*s, but no lockres!\n",
2037                              assert->node_idx, dlm->name, namelen, name);
2038                 }
2039         }
2040         if (have_lockres_ref) {
2041                 /* let the master know we have a reference to the lockres */
2042                 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2043                 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2044                      dlm->name, namelen, name, assert->node_idx);
2045         }
2046         return ret;
2047
2048 kill:
2049         /* kill the caller! */
2050         mlog(ML_ERROR, "Bad message received from another node.  Dumping state "
2051              "and killing the other node now!  This node is OK and can continue.\n");
2052         __dlm_print_one_lock_resource(res);
2053         spin_unlock(&res->spinlock);
2054         spin_lock(&dlm->master_lock);
2055         if (mle)
2056                 __dlm_put_mle(mle);
2057         spin_unlock(&dlm->master_lock);
2058         spin_unlock(&dlm->spinlock);
2059         *ret_data = (void *)res;
2060         dlm_put(dlm);
2061         return -EINVAL;
2062 }
2063
2064 void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2065 {
2066         struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2067
2068         if (ret_data) {
2069                 spin_lock(&res->spinlock);
2070                 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2071                 spin_unlock(&res->spinlock);
2072                 wake_up(&res->wq);
2073                 dlm_lockres_put(res);
2074         }
2075         return;
2076 }
2077
2078 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2079                                struct dlm_lock_resource *res,
2080                                int ignore_higher, u8 request_from, u32 flags)
2081 {
2082         struct dlm_work_item *item;
2083         item = kzalloc(sizeof(*item), GFP_ATOMIC);
2084         if (!item)
2085                 return -ENOMEM;
2086
2087
2088         /* queue up work for dlm_assert_master_worker */
2089         dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2090         item->u.am.lockres = res; /* already have a ref */
2091         /* can optionally ignore node numbers higher than this node */
2092         item->u.am.ignore_higher = ignore_higher;
2093         item->u.am.request_from = request_from;
2094         item->u.am.flags = flags;
2095
2096         if (ignore_higher)
2097                 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2098                      res->lockname.name);
2099
2100         spin_lock(&dlm->work_lock);
2101         list_add_tail(&item->list, &dlm->work_list);
2102         spin_unlock(&dlm->work_lock);
2103
2104         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2105         return 0;
2106 }
2107
2108 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2109 {
2110         struct dlm_ctxt *dlm = data;
2111         int ret = 0;
2112         struct dlm_lock_resource *res;
2113         unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2114         int ignore_higher;
2115         int bit;
2116         u8 request_from;
2117         u32 flags;
2118
2119         dlm = item->dlm;
2120         res = item->u.am.lockres;
2121         ignore_higher = item->u.am.ignore_higher;
2122         request_from = item->u.am.request_from;
2123         flags = item->u.am.flags;
2124
2125         spin_lock(&dlm->spinlock);
2126         memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2127         spin_unlock(&dlm->spinlock);
2128
2129         clear_bit(dlm->node_num, nodemap);
2130         if (ignore_higher) {
2131                 /* if is this just to clear up mles for nodes below
2132                  * this node, do not send the message to the original
2133                  * caller or any node number higher than this */
2134                 clear_bit(request_from, nodemap);
2135                 bit = dlm->node_num;
2136                 while (1) {
2137                         bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2138                                             bit+1);
2139                         if (bit >= O2NM_MAX_NODES)
2140                                 break;
2141                         clear_bit(bit, nodemap);
2142                 }
2143         }
2144
2145         /*
2146          * If we're migrating this lock to someone else, we are no
2147          * longer allowed to assert out own mastery.  OTOH, we need to
2148          * prevent migration from starting while we're still asserting
2149          * our dominance.  The reserved ast delays migration.
2150          */
2151         spin_lock(&res->spinlock);
2152         if (res->state & DLM_LOCK_RES_MIGRATING) {
2153                 mlog(0, "Someone asked us to assert mastery, but we're "
2154                      "in the middle of migration.  Skipping assert, "
2155                      "the new master will handle that.\n");
2156                 spin_unlock(&res->spinlock);
2157                 goto put;
2158         } else
2159                 __dlm_lockres_reserve_ast(res);
2160         spin_unlock(&res->spinlock);
2161
2162         /* this call now finishes out the nodemap
2163          * even if one or more nodes die */
2164         mlog(0, "worker about to master %.*s here, this=%u\n",
2165                      res->lockname.len, res->lockname.name, dlm->node_num);
2166         ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2167         if (ret < 0) {
2168                 /* no need to restart, we are done */
2169                 if (!dlm_is_host_down(ret))
2170                         mlog_errno(ret);
2171         }
2172
2173         /* Ok, we've asserted ourselves.  Let's let migration start. */
2174         dlm_lockres_release_ast(dlm, res);
2175
2176 put:
2177         dlm_lockres_drop_inflight_worker(dlm, res);
2178
2179         dlm_lockres_put(res);
2180
2181         mlog(0, "finished with dlm_assert_master_worker\n");
2182 }
2183
2184 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2185  * We cannot wait for node recovery to complete to begin mastering this
2186  * lockres because this lockres is used to kick off recovery! ;-)
2187  * So, do a pre-check on all living nodes to see if any of those nodes
2188  * think that $RECOVERY is currently mastered by a dead node.  If so,
2189  * we wait a short time to allow that node to get notified by its own
2190  * heartbeat stack, then check again.  All $RECOVERY lock resources
2191  * mastered by dead nodes are purged when the hearbeat callback is
2192  * fired, so we can know for sure that it is safe to continue once
2193  * the node returns a live node or no node.  */
2194 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2195                                        struct dlm_lock_resource *res)
2196 {
2197         struct dlm_node_iter iter;
2198         int nodenum;
2199         int ret = 0;
2200         u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2201
2202         spin_lock(&dlm->spinlock);
2203         dlm_node_iter_init(dlm->domain_map, &iter);
2204         spin_unlock(&dlm->spinlock);
2205
2206         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2207                 /* do not send to self */
2208                 if (nodenum == dlm->node_num)
2209                         continue;
2210                 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2211                 if (ret < 0) {
2212                         mlog_errno(ret);
2213                         if (!dlm_is_host_down(ret))
2214                                 BUG();
2215                         /* host is down, so answer for that node would be
2216                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
2217                         ret = 0;
2218                 }
2219
2220                 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2221                         /* check to see if this master is in the recovery map */
2222                         spin_lock(&dlm->spinlock);
2223                         if (test_bit(master, dlm->recovery_map)) {
2224                                 mlog(ML_NOTICE, "%s: node %u has not seen "
2225                                      "node %u go down yet, and thinks the "
2226                                      "dead node is mastering the recovery "
2227                                      "lock.  must wait.\n", dlm->name,
2228                                      nodenum, master);
2229                                 ret = -EAGAIN;
2230                         }
2231                         spin_unlock(&dlm->spinlock);
2232                         mlog(0, "%s: reco lock master is %u\n", dlm->name,
2233                              master);
2234                         break;
2235                 }
2236         }
2237         return ret;
2238 }
2239
2240 /*
2241  * DLM_DEREF_LOCKRES_MSG
2242  */
2243
2244 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2245 {
2246         struct dlm_deref_lockres deref;
2247         int ret = 0, r;
2248         const char *lockname;
2249         unsigned int namelen;
2250
2251         lockname = res->lockname.name;
2252         namelen = res->lockname.len;
2253         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2254
2255         memset(&deref, 0, sizeof(deref));
2256         deref.node_idx = dlm->node_num;
2257         deref.namelen = namelen;
2258         memcpy(deref.name, lockname, namelen);
2259
2260         ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2261                                  &deref, sizeof(deref), res->owner, &r);
2262         if (ret < 0)
2263                 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n",
2264                      dlm->name, namelen, lockname, ret, res->owner);
2265         else if (r < 0) {
2266                 /* BAD.  other node says I did not have a ref. */
2267                 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2268                      dlm->name, namelen, lockname, res->owner, r);
2269                 dlm_print_one_lock_resource(res);
2270                 if (r == -ENOMEM)
2271                         BUG();
2272         } else
2273                 ret = r;
2274
2275         return ret;
2276 }
2277
2278 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2279                               void **ret_data)
2280 {
2281         struct dlm_ctxt *dlm = data;
2282         struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2283         struct dlm_lock_resource *res = NULL;
2284         char *name;
2285         unsigned int namelen;
2286         int ret = -EINVAL;
2287         u8 node;
2288         unsigned int hash;
2289         struct dlm_work_item *item;
2290         int cleared = 0;
2291         int dispatch = 0;
2292
2293         if (!dlm_grab(dlm))
2294                 return 0;
2295
2296         name = deref->name;
2297         namelen = deref->namelen;
2298         node = deref->node_idx;
2299
2300         if (namelen > DLM_LOCKID_NAME_MAX) {
2301                 mlog(ML_ERROR, "Invalid name length!");
2302                 goto done;
2303         }
2304         if (deref->node_idx >= O2NM_MAX_NODES) {
2305                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2306                 goto done;
2307         }
2308
2309         hash = dlm_lockid_hash(name, namelen);
2310
2311         spin_lock(&dlm->spinlock);
2312         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2313         if (!res) {
2314                 spin_unlock(&dlm->spinlock);
2315                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2316                      dlm->name, namelen, name);
2317                 goto done;
2318         }
2319         spin_unlock(&dlm->spinlock);
2320
2321         spin_lock(&res->spinlock);
2322         if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2323                 dispatch = 1;
2324         else {
2325                 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2326                 if (test_bit(node, res->refmap)) {
2327                         dlm_lockres_clear_refmap_bit(dlm, res, node);
2328                         cleared = 1;
2329                 }
2330         }
2331         spin_unlock(&res->spinlock);
2332
2333         if (!dispatch) {
2334                 if (cleared)
2335                         dlm_lockres_calc_usage(dlm, res);
2336                 else {
2337                         mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2338                         "but it is already dropped!\n", dlm->name,
2339                         res->lockname.len, res->lockname.name, node);
2340                         dlm_print_one_lock_resource(res);
2341                 }
2342                 ret = DLM_DEREF_RESPONSE_DONE;
2343                 goto done;
2344         }
2345
2346         item = kzalloc(sizeof(*item), GFP_NOFS);
2347         if (!item) {
2348                 ret = -ENOMEM;
2349                 mlog_errno(ret);
2350                 goto done;
2351         }
2352
2353         dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2354         item->u.dl.deref_res = res;
2355         item->u.dl.deref_node = node;
2356
2357         spin_lock(&dlm->work_lock);
2358         list_add_tail(&item->list, &dlm->work_list);
2359         spin_unlock(&dlm->work_lock);
2360
2361         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2362         return DLM_DEREF_RESPONSE_INPROG;
2363
2364 done:
2365         if (res)
2366                 dlm_lockres_put(res);
2367         dlm_put(dlm);
2368
2369         return ret;
2370 }
2371
2372 int dlm_deref_lockres_done_handler(struct o2net_msg *msg, u32 len, void *data,
2373                               void **ret_data)
2374 {
2375         struct dlm_ctxt *dlm = data;
2376         struct dlm_deref_lockres_done *deref
2377                         = (struct dlm_deref_lockres_done *)msg->buf;
2378         struct dlm_lock_resource *res = NULL;
2379         char *name;
2380         unsigned int namelen;
2381         int ret = -EINVAL;
2382         u8 node;
2383         unsigned int hash;
2384
2385         if (!dlm_grab(dlm))
2386                 return 0;
2387
2388         name = deref->name;
2389         namelen = deref->namelen;
2390         node = deref->node_idx;
2391
2392         if (namelen > DLM_LOCKID_NAME_MAX) {
2393                 mlog(ML_ERROR, "Invalid name length!");
2394                 goto done;
2395         }
2396         if (deref->node_idx >= O2NM_MAX_NODES) {
2397                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2398                 goto done;
2399         }
2400
2401         hash = dlm_lockid_hash(name, namelen);
2402
2403         spin_lock(&dlm->spinlock);
2404         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2405         if (!res) {
2406                 spin_unlock(&dlm->spinlock);
2407                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2408                      dlm->name, namelen, name);
2409                 goto done;
2410         }
2411
2412         spin_lock(&res->spinlock);
2413         if (!(res->state & DLM_LOCK_RES_DROPPING_REF)) {
2414                 spin_unlock(&res->spinlock);
2415                 spin_unlock(&dlm->spinlock);
2416                 mlog(ML_NOTICE, "%s:%.*s: node %u sends deref done "
2417                         "but it is already derefed!\n", dlm->name,
2418                         res->lockname.len, res->lockname.name, node);
2419                 ret = 0;
2420                 goto done;
2421         }
2422
2423         __dlm_do_purge_lockres(dlm, res);
2424         spin_unlock(&res->spinlock);
2425         wake_up(&res->wq);
2426
2427         spin_unlock(&dlm->spinlock);
2428
2429         ret = 0;
2430 done:
2431         if (res)
2432                 dlm_lockres_put(res);
2433         dlm_put(dlm);
2434         return ret;
2435 }
2436
2437 static void dlm_drop_lockres_ref_done(struct dlm_ctxt *dlm,
2438                 struct dlm_lock_resource *res, u8 node)
2439 {
2440         struct dlm_deref_lockres_done deref;
2441         int ret = 0, r;
2442         const char *lockname;
2443         unsigned int namelen;
2444
2445         lockname = res->lockname.name;
2446         namelen = res->lockname.len;
2447         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2448
2449         memset(&deref, 0, sizeof(deref));
2450         deref.node_idx = dlm->node_num;
2451         deref.namelen = namelen;
2452         memcpy(deref.name, lockname, namelen);
2453
2454         ret = o2net_send_message(DLM_DEREF_LOCKRES_DONE, dlm->key,
2455                                  &deref, sizeof(deref), node, &r);
2456         if (ret < 0) {
2457                 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF DONE "
2458                                 " to node %u\n", dlm->name, namelen,
2459                                 lockname, ret, node);
2460         } else if (r < 0) {
2461                 /* ignore the error */
2462                 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2463                      dlm->name, namelen, lockname, node, r);
2464                 dlm_print_one_lock_resource(res);
2465         }
2466 }
2467
2468 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2469 {
2470         struct dlm_ctxt *dlm;
2471         struct dlm_lock_resource *res;
2472         u8 node;
2473         u8 cleared = 0;
2474
2475         dlm = item->dlm;
2476         res = item->u.dl.deref_res;
2477         node = item->u.dl.deref_node;
2478
2479         spin_lock(&res->spinlock);
2480         BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2481         __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2482         if (test_bit(node, res->refmap)) {
2483                 dlm_lockres_clear_refmap_bit(dlm, res, node);
2484                 cleared = 1;
2485         }
2486         spin_unlock(&res->spinlock);
2487
2488         dlm_drop_lockres_ref_done(dlm, res, node);
2489
2490         if (cleared) {
2491                 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2492                      dlm->name, res->lockname.len, res->lockname.name, node);
2493                 dlm_lockres_calc_usage(dlm, res);
2494         } else {
2495                 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2496                      "but it is already dropped!\n", dlm->name,
2497                      res->lockname.len, res->lockname.name, node);
2498                 dlm_print_one_lock_resource(res);
2499         }
2500
2501         dlm_lockres_put(res);
2502 }
2503
2504 /*
2505  * A migrateable resource is one that is :
2506  * 1. locally mastered, and,
2507  * 2. zero local locks, and,
2508  * 3. one or more non-local locks, or, one or more references
2509  * Returns 1 if yes, 0 if not.
2510  */
2511 static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
2512                                       struct dlm_lock_resource *res)
2513 {
2514         enum dlm_lockres_list idx;
2515         int nonlocal = 0, node_ref;
2516         struct list_head *queue;
2517         struct dlm_lock *lock;
2518         u64 cookie;
2519
2520         assert_spin_locked(&res->spinlock);
2521
2522         /* delay migration when the lockres is in MIGRATING state */
2523         if (res->state & DLM_LOCK_RES_MIGRATING)
2524                 return 0;
2525
2526         /* delay migration when the lockres is in RECOCERING state */
2527         if (res->state & (DLM_LOCK_RES_RECOVERING|
2528                         DLM_LOCK_RES_RECOVERY_WAITING))
2529                 return 0;
2530
2531         if (res->owner != dlm->node_num)
2532                 return 0;
2533
2534         for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2535                 queue = dlm_list_idx_to_ptr(res, idx);
2536                 list_for_each_entry(lock, queue, list) {
2537                         if (lock->ml.node != dlm->node_num) {
2538                                 nonlocal++;
2539                                 continue;
2540                         }
2541                         cookie = be64_to_cpu(lock->ml.cookie);
2542                         mlog(0, "%s: Not migrateable res %.*s, lock %u:%llu on "
2543                              "%s list\n", dlm->name, res->lockname.len,
2544                              res->lockname.name,
2545                              dlm_get_lock_cookie_node(cookie),
2546                              dlm_get_lock_cookie_seq(cookie),
2547                              dlm_list_in_text(idx));
2548                         return 0;
2549                 }
2550         }
2551
2552         if (!nonlocal) {
2553                 node_ref = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
2554                 if (node_ref >= O2NM_MAX_NODES)
2555                         return 0;
2556         }
2557
2558         mlog(0, "%s: res %.*s, Migrateable\n", dlm->name, res->lockname.len,
2559              res->lockname.name);
2560
2561         return 1;
2562 }
2563
2564 /*
2565  * DLM_MIGRATE_LOCKRES
2566  */
2567
2568
2569 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2570                                struct dlm_lock_resource *res, u8 target)
2571 {
2572         struct dlm_master_list_entry *mle = NULL;
2573         struct dlm_master_list_entry *oldmle = NULL;
2574         struct dlm_migratable_lockres *mres = NULL;
2575         int ret = 0;
2576         const char *name;
2577         unsigned int namelen;
2578         int mle_added = 0;
2579         int wake = 0;
2580
2581         if (!dlm_grab(dlm))
2582                 return -EINVAL;
2583
2584         BUG_ON(target == O2NM_MAX_NODES);
2585
2586         name = res->lockname.name;
2587         namelen = res->lockname.len;
2588
2589         mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name,
2590              target);
2591
2592         /* preallocate up front. if this fails, abort */
2593         ret = -ENOMEM;
2594         mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2595         if (!mres) {
2596                 mlog_errno(ret);
2597                 goto leave;
2598         }
2599
2600         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2601         if (!mle) {
2602                 mlog_errno(ret);
2603                 goto leave;
2604         }
2605         ret = 0;
2606
2607         /*
2608          * clear any existing master requests and
2609          * add the migration mle to the list
2610          */
2611         spin_lock(&dlm->spinlock);
2612         spin_lock(&dlm->master_lock);
2613         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2614                                     namelen, target, dlm->node_num);
2615         /* get an extra reference on the mle.
2616          * otherwise the assert_master from the new
2617          * master will destroy this.
2618          */
2619         dlm_get_mle_inuse(mle);
2620         spin_unlock(&dlm->master_lock);
2621         spin_unlock(&dlm->spinlock);
2622
2623         if (ret == -EEXIST) {
2624                 mlog(0, "another process is already migrating it\n");
2625                 goto fail;
2626         }
2627         mle_added = 1;
2628
2629         /*
2630          * set the MIGRATING flag and flush asts
2631          * if we fail after this we need to re-dirty the lockres
2632          */
2633         if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2634                 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2635                      "the target went down.\n", res->lockname.len,
2636                      res->lockname.name, target);
2637                 spin_lock(&res->spinlock);
2638                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2639                 wake = 1;
2640                 spin_unlock(&res->spinlock);
2641                 ret = -EINVAL;
2642         }
2643
2644 fail:
2645         if (ret != -EEXIST && oldmle) {
2646                 /* master is known, detach if not already detached */
2647                 dlm_mle_detach_hb_events(dlm, oldmle);
2648                 dlm_put_mle(oldmle);
2649         }
2650
2651         if (ret < 0) {
2652                 if (mle_added) {
2653                         dlm_mle_detach_hb_events(dlm, mle);
2654                         dlm_put_mle(mle);
2655                         dlm_put_mle_inuse(mle);
2656                 } else if (mle) {
2657                         kmem_cache_free(dlm_mle_cache, mle);
2658                         mle = NULL;
2659                 }
2660                 goto leave;
2661         }
2662
2663         /*
2664          * at this point, we have a migration target, an mle
2665          * in the master list, and the MIGRATING flag set on
2666          * the lockres
2667          */
2668
2669         /* now that remote nodes are spinning on the MIGRATING flag,
2670          * ensure that all assert_master work is flushed. */
2671         flush_workqueue(dlm->dlm_worker);
2672
2673         /* notify new node and send all lock state */
2674         /* call send_one_lockres with migration flag.
2675          * this serves as notice to the target node that a
2676          * migration is starting. */
2677         ret = dlm_send_one_lockres(dlm, res, mres, target,
2678                                    DLM_MRES_MIGRATION);
2679
2680         if (ret < 0) {
2681                 mlog(0, "migration to node %u failed with %d\n",
2682                      target, ret);
2683                 /* migration failed, detach and clean up mle */
2684                 dlm_mle_detach_hb_events(dlm, mle);
2685                 dlm_put_mle(mle);
2686                 dlm_put_mle_inuse(mle);
2687                 spin_lock(&res->spinlock);
2688                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2689                 wake = 1;
2690                 spin_unlock(&res->spinlock);
2691                 if (dlm_is_host_down(ret))
2692                         dlm_wait_for_node_death(dlm, target,
2693                                                 DLM_NODE_DEATH_WAIT_MAX);
2694                 goto leave;
2695         }
2696
2697         /* at this point, the target sends a message to all nodes,
2698          * (using dlm_do_migrate_request).  this node is skipped since
2699          * we had to put an mle in the list to begin the process.  this
2700          * node now waits for target to do an assert master.  this node
2701          * will be the last one notified, ensuring that the migration
2702          * is complete everywhere.  if the target dies while this is
2703          * going on, some nodes could potentially see the target as the
2704          * master, so it is important that my recovery finds the migration
2705          * mle and sets the master to UNKNOWN. */
2706
2707
2708         /* wait for new node to assert master */
2709         while (1) {
2710                 ret = wait_event_interruptible_timeout(mle->wq,
2711                                         (atomic_read(&mle->woken) == 1),
2712                                         msecs_to_jiffies(5000));
2713
2714                 if (ret >= 0) {
2715                         if (atomic_read(&mle->woken) == 1 ||
2716                             res->owner == target)
2717                                 break;
2718
2719                         mlog(0, "%s:%.*s: timed out during migration\n",
2720                              dlm->name, res->lockname.len, res->lockname.name);
2721                         /* avoid hang during shutdown when migrating lockres
2722                          * to a node which also goes down */
2723                         if (dlm_is_node_dead(dlm, target)) {
2724                                 mlog(0, "%s:%.*s: expected migration "
2725                                      "target %u is no longer up, restarting\n",
2726                                      dlm->name, res->lockname.len,
2727                                      res->lockname.name, target);
2728                                 ret = -EINVAL;
2729                                 /* migration failed, detach and clean up mle */
2730                                 dlm_mle_detach_hb_events(dlm, mle);
2731                                 dlm_put_mle(mle);
2732                                 dlm_put_mle_inuse(mle);
2733                                 spin_lock(&res->spinlock);
2734                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2735                                 wake = 1;
2736                                 spin_unlock(&res->spinlock);
2737                                 goto leave;
2738                         }
2739                 } else
2740                         mlog(0, "%s:%.*s: caught signal during migration\n",
2741                              dlm->name, res->lockname.len, res->lockname.name);
2742         }
2743
2744         /* all done, set the owner, clear the flag */
2745         spin_lock(&res->spinlock);
2746         dlm_set_lockres_owner(dlm, res, target);
2747         res->state &= ~DLM_LOCK_RES_MIGRATING;
2748         dlm_remove_nonlocal_locks(dlm, res);
2749         spin_unlock(&res->spinlock);
2750         wake_up(&res->wq);
2751
2752         /* master is known, detach if not already detached */
2753         dlm_mle_detach_hb_events(dlm, mle);
2754         dlm_put_mle_inuse(mle);
2755         ret = 0;
2756
2757         dlm_lockres_calc_usage(dlm, res);
2758
2759 leave:
2760         /* re-dirty the lockres if we failed */
2761         if (ret < 0)
2762                 dlm_kick_thread(dlm, res);
2763
2764         /* wake up waiters if the MIGRATING flag got set
2765          * but migration failed */
2766         if (wake)
2767                 wake_up(&res->wq);
2768
2769         if (mres)
2770                 free_page((unsigned long)mres);
2771
2772         dlm_put(dlm);
2773
2774         mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen,
2775              name, target, ret);
2776         return ret;
2777 }
2778
2779 #define DLM_MIGRATION_RETRY_MS  100
2780
2781 /*
2782  * Should be called only after beginning the domain leave process.
2783  * There should not be any remaining locks on nonlocal lock resources,
2784  * and there should be no local locks left on locally mastered resources.
2785  *
2786  * Called with the dlm spinlock held, may drop it to do migration, but
2787  * will re-acquire before exit.
2788  *
2789  * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2790  */
2791 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2792 {
2793         int ret;
2794         int lock_dropped = 0;
2795         u8 target = O2NM_MAX_NODES;
2796
2797         assert_spin_locked(&dlm->spinlock);
2798
2799         spin_lock(&res->spinlock);
2800         if (dlm_is_lockres_migrateable(dlm, res))
2801                 target = dlm_pick_migration_target(dlm, res);
2802         spin_unlock(&res->spinlock);
2803
2804         if (target == O2NM_MAX_NODES)
2805                 goto leave;
2806
2807         /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2808         spin_unlock(&dlm->spinlock);
2809         lock_dropped = 1;
2810         ret = dlm_migrate_lockres(dlm, res, target);
2811         if (ret)
2812                 mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2813                      dlm->name, res->lockname.len, res->lockname.name,
2814                      target, ret);
2815         spin_lock(&dlm->spinlock);
2816 leave:
2817         return lock_dropped;
2818 }
2819
2820 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2821 {
2822         int ret;
2823         spin_lock(&dlm->ast_lock);
2824         spin_lock(&lock->spinlock);
2825         ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2826         spin_unlock(&lock->spinlock);
2827         spin_unlock(&dlm->ast_lock);
2828         return ret;
2829 }
2830
2831 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2832                                      struct dlm_lock_resource *res,
2833                                      u8 mig_target)
2834 {
2835         int can_proceed;
2836         spin_lock(&res->spinlock);
2837         can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2838         spin_unlock(&res->spinlock);
2839
2840         /* target has died, so make the caller break out of the
2841          * wait_event, but caller must recheck the domain_map */
2842         spin_lock(&dlm->spinlock);
2843         if (!test_bit(mig_target, dlm->domain_map))
2844                 can_proceed = 1;
2845         spin_unlock(&dlm->spinlock);
2846         return can_proceed;
2847 }
2848
2849 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2850                                 struct dlm_lock_resource *res)
2851 {
2852         int ret;
2853         spin_lock(&res->spinlock);
2854         ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2855         spin_unlock(&res->spinlock);
2856         return ret;
2857 }
2858
2859
2860 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2861                                        struct dlm_lock_resource *res,
2862                                        u8 target)
2863 {
2864         int ret = 0;
2865
2866         mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2867                res->lockname.len, res->lockname.name, dlm->node_num,
2868                target);
2869         /* need to set MIGRATING flag on lockres.  this is done by
2870          * ensuring that all asts have been flushed for this lockres. */
2871         spin_lock(&res->spinlock);
2872         BUG_ON(res->migration_pending);
2873         res->migration_pending = 1;
2874         /* strategy is to reserve an extra ast then release
2875          * it below, letting the release do all of the work */
2876         __dlm_lockres_reserve_ast(res);
2877         spin_unlock(&res->spinlock);
2878
2879         /* now flush all the pending asts */
2880         dlm_kick_thread(dlm, res);
2881         /* before waiting on DIRTY, block processes which may
2882          * try to dirty the lockres before MIGRATING is set */
2883         spin_lock(&res->spinlock);
2884         BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2885         res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2886         spin_unlock(&res->spinlock);
2887         /* now wait on any pending asts and the DIRTY state */
2888         wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2889         dlm_lockres_release_ast(dlm, res);
2890
2891         mlog(0, "about to wait on migration_wq, dirty=%s\n",
2892                res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2893         /* if the extra ref we just put was the final one, this
2894          * will pass thru immediately.  otherwise, we need to wait
2895          * for the last ast to finish. */
2896 again:
2897         ret = wait_event_interruptible_timeout(dlm->migration_wq,
2898                    dlm_migration_can_proceed(dlm, res, target),
2899                    msecs_to_jiffies(1000));
2900         if (ret < 0) {
2901                 mlog(0, "woken again: migrating? %s, dead? %s\n",
2902                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2903                        test_bit(target, dlm->domain_map) ? "no":"yes");
2904         } else {
2905                 mlog(0, "all is well: migrating? %s, dead? %s\n",
2906                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2907                        test_bit(target, dlm->domain_map) ? "no":"yes");
2908         }
2909         if (!dlm_migration_can_proceed(dlm, res, target)) {
2910                 mlog(0, "trying again...\n");
2911                 goto again;
2912         }
2913
2914         ret = 0;
2915         /* did the target go down or die? */
2916         spin_lock(&dlm->spinlock);
2917         if (!test_bit(target, dlm->domain_map)) {
2918                 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2919                      target);
2920                 ret = -EHOSTDOWN;
2921         }
2922         spin_unlock(&dlm->spinlock);
2923
2924         /*
2925          * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2926          * another try; otherwise, we are sure the MIGRATING state is there,
2927          * drop the unneeded state which blocked threads trying to DIRTY
2928          */
2929         spin_lock(&res->spinlock);
2930         BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2931         res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2932         if (!ret)
2933                 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2934         else
2935                 res->migration_pending = 0;
2936         spin_unlock(&res->spinlock);
2937
2938         /*
2939          * at this point:
2940          *
2941          *   o the DLM_LOCK_RES_MIGRATING flag is set if target not down
2942          *   o there are no pending asts on this lockres
2943          *   o all processes trying to reserve an ast on this
2944          *     lockres must wait for the MIGRATING flag to clear
2945          */
2946         return ret;
2947 }
2948
2949 /* last step in the migration process.
2950  * original master calls this to free all of the dlm_lock
2951  * structures that used to be for other nodes. */
2952 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2953                                       struct dlm_lock_resource *res)
2954 {
2955         struct list_head *queue = &res->granted;
2956         int i, bit;
2957         struct dlm_lock *lock, *next;
2958
2959         assert_spin_locked(&res->spinlock);
2960
2961         BUG_ON(res->owner == dlm->node_num);
2962
2963         for (i=0; i<3; i++) {
2964                 list_for_each_entry_safe(lock, next, queue, list) {
2965                         if (lock->ml.node != dlm->node_num) {
2966                                 mlog(0, "putting lock for node %u\n",
2967                                      lock->ml.node);
2968                                 /* be extra careful */
2969                                 BUG_ON(!list_empty(&lock->ast_list));
2970                                 BUG_ON(!list_empty(&lock->bast_list));
2971                                 BUG_ON(lock->ast_pending);
2972                                 BUG_ON(lock->bast_pending);
2973                                 dlm_lockres_clear_refmap_bit(dlm, res,
2974                                                              lock->ml.node);
2975                                 list_del_init(&lock->list);
2976                                 dlm_lock_put(lock);
2977                                 /* In a normal unlock, we would have added a
2978                                  * DLM_UNLOCK_FREE_LOCK action. Force it. */
2979                                 dlm_lock_put(lock);
2980                         }
2981                 }
2982                 queue++;
2983         }
2984         bit = 0;
2985         while (1) {
2986                 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2987                 if (bit >= O2NM_MAX_NODES)
2988                         break;
2989                 /* do not clear the local node reference, if there is a
2990                  * process holding this, let it drop the ref itself */
2991                 if (bit != dlm->node_num) {
2992                         mlog(0, "%s:%.*s: node %u had a ref to this "
2993                              "migrating lockres, clearing\n", dlm->name,
2994                              res->lockname.len, res->lockname.name, bit);
2995                         dlm_lockres_clear_refmap_bit(dlm, res, bit);
2996                 }
2997                 bit++;
2998         }
2999 }
3000
3001 /*
3002  * Pick a node to migrate the lock resource to. This function selects a
3003  * potential target based first on the locks and then on refmap. It skips
3004  * nodes that are in the process of exiting the domain.
3005  */
3006 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
3007                                     struct dlm_lock_resource *res)
3008 {
3009         enum dlm_lockres_list idx;
3010         struct list_head *queue = &res->granted;
3011         struct dlm_lock *lock;
3012         int noderef;
3013         u8 nodenum = O2NM_MAX_NODES;
3014
3015         assert_spin_locked(&dlm->spinlock);
3016         assert_spin_locked(&res->spinlock);
3017
3018         /* Go through all the locks */
3019         for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
3020                 queue = dlm_list_idx_to_ptr(res, idx);
3021                 list_for_each_entry(lock, queue, list) {
3022                         if (lock->ml.node == dlm->node_num)
3023                                 continue;
3024                         if (test_bit(lock->ml.node, dlm->exit_domain_map))
3025                                 continue;
3026                         nodenum = lock->ml.node;
3027                         goto bail;
3028                 }
3029         }
3030
3031         /* Go thru the refmap */
3032         noderef = -1;
3033         while (1) {
3034                 noderef = find_next_bit(res->refmap, O2NM_MAX_NODES,
3035                                         noderef + 1);
3036                 if (noderef >= O2NM_MAX_NODES)
3037                         break;
3038                 if (noderef == dlm->node_num)
3039                         continue;
3040                 if (test_bit(noderef, dlm->exit_domain_map))
3041                         continue;
3042                 nodenum = noderef;
3043                 goto bail;
3044         }
3045
3046 bail:
3047         return nodenum;
3048 }
3049
3050 /* this is called by the new master once all lockres
3051  * data has been received */
3052 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
3053                                   struct dlm_lock_resource *res,
3054                                   u8 master, u8 new_master,
3055                                   struct dlm_node_iter *iter)
3056 {
3057         struct dlm_migrate_request migrate;
3058         int ret, skip, status = 0;
3059         int nodenum;
3060
3061         memset(&migrate, 0, sizeof(migrate));
3062         migrate.namelen = res->lockname.len;
3063         memcpy(migrate.name, res->lockname.name, migrate.namelen);
3064         migrate.new_master = new_master;
3065         migrate.master = master;
3066
3067         ret = 0;
3068
3069         /* send message to all nodes, except the master and myself */
3070         while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
3071                 if (nodenum == master ||
3072                     nodenum == new_master)
3073                         continue;
3074
3075                 /* We could race exit domain. If exited, skip. */
3076                 spin_lock(&dlm->spinlock);
3077                 skip = (!test_bit(nodenum, dlm->domain_map));
3078                 spin_unlock(&dlm->spinlock);
3079                 if (skip) {
3080                         clear_bit(nodenum, iter->node_map);
3081                         continue;
3082                 }
3083
3084                 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
3085                                          &migrate, sizeof(migrate), nodenum,
3086                                          &status);
3087                 if (ret < 0) {
3088                         mlog(ML_ERROR, "%s: res %.*s, Error %d send "
3089                              "MIGRATE_REQUEST to node %u\n", dlm->name,
3090                              migrate.namelen, migrate.name, ret, nodenum);
3091                         if (!dlm_is_host_down(ret)) {
3092                                 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
3093                                 BUG();
3094                         }
3095                         clear_bit(nodenum, iter->node_map);
3096                         ret = 0;
3097                 } else if (status < 0) {
3098                         mlog(0, "migrate request (node %u) returned %d!\n",
3099                              nodenum, status);
3100                         ret = status;
3101                 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
3102                         /* during the migration request we short-circuited
3103                          * the mastery of the lockres.  make sure we have
3104                          * a mastery ref for nodenum */
3105                         mlog(0, "%s:%.*s: need ref for node %u\n",
3106                              dlm->name, res->lockname.len, res->lockname.name,
3107                              nodenum);
3108                         spin_lock(&res->spinlock);
3109                         dlm_lockres_set_refmap_bit(dlm, res, nodenum);
3110                         spin_unlock(&res->spinlock);
3111                 }
3112         }
3113
3114         if (ret < 0)
3115                 mlog_errno(ret);
3116
3117         mlog(0, "returning ret=%d\n", ret);
3118         return ret;
3119 }
3120
3121
3122 /* if there is an existing mle for this lockres, we now know who the master is.
3123  * (the one who sent us *this* message) we can clear it up right away.
3124  * since the process that put the mle on the list still has a reference to it,
3125  * we can unhash it now, set the master and wake the process.  as a result,
3126  * we will have no mle in the list to start with.  now we can add an mle for
3127  * the migration and this should be the only one found for those scanning the
3128  * list.  */
3129 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
3130                                 void **ret_data)
3131 {
3132         struct dlm_ctxt *dlm = data;
3133         struct dlm_lock_resource *res = NULL;
3134         struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
3135         struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
3136         const char *name;
3137         unsigned int namelen, hash;
3138         int ret = 0;
3139
3140         if (!dlm_grab(dlm))
3141                 return 0;
3142
3143         name = migrate->name;
3144         namelen = migrate->namelen;
3145         hash = dlm_lockid_hash(name, namelen);
3146
3147         /* preallocate.. if this fails, abort */
3148         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
3149
3150         if (!mle) {
3151                 ret = -ENOMEM;
3152                 goto leave;
3153         }
3154
3155         /* check for pre-existing lock */
3156         spin_lock(&dlm->spinlock);
3157         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
3158         if (res) {
3159                 spin_lock(&res->spinlock);
3160                 if (res->state & DLM_LOCK_RES_RECOVERING) {
3161                         /* if all is working ok, this can only mean that we got
3162                         * a migrate request from a node that we now see as
3163                         * dead.  what can we do here?  drop it to the floor? */
3164                         spin_unlock(&res->spinlock);
3165                         mlog(ML_ERROR, "Got a migrate request, but the "
3166                              "lockres is marked as recovering!");
3167                         kmem_cache_free(dlm_mle_cache, mle);
3168                         ret = -EINVAL; /* need a better solution */
3169                         goto unlock;
3170                 }
3171                 res->state |= DLM_LOCK_RES_MIGRATING;
3172                 spin_unlock(&res->spinlock);
3173         }
3174
3175         spin_lock(&dlm->master_lock);
3176         /* ignore status.  only nonzero status would BUG. */
3177         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3178                                     name, namelen,
3179                                     migrate->new_master,
3180                                     migrate->master);
3181
3182         if (ret < 0)
3183                 kmem_cache_free(dlm_mle_cache, mle);
3184
3185         spin_unlock(&dlm->master_lock);
3186 unlock:
3187         spin_unlock(&dlm->spinlock);
3188
3189         if (oldmle) {
3190                 /* master is known, detach if not already detached */
3191                 dlm_mle_detach_hb_events(dlm, oldmle);
3192                 dlm_put_mle(oldmle);
3193         }
3194
3195         if (res)
3196                 dlm_lockres_put(res);
3197 leave:
3198         dlm_put(dlm);
3199         return ret;
3200 }
3201
3202 /* must be holding dlm->spinlock and dlm->master_lock
3203  * when adding a migration mle, we can clear any other mles
3204  * in the master list because we know with certainty that
3205  * the master is "master".  so we remove any old mle from
3206  * the list after setting it's master field, and then add
3207  * the new migration mle.  this way we can hold with the rule
3208  * of having only one mle for a given lock name at all times. */
3209 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3210                                  struct dlm_lock_resource *res,
3211                                  struct dlm_master_list_entry *mle,
3212                                  struct dlm_master_list_entry **oldmle,
3213                                  const char *name, unsigned int namelen,
3214                                  u8 new_master, u8 master)
3215 {
3216         int found;
3217         int ret = 0;
3218
3219         *oldmle = NULL;
3220
3221         assert_spin_locked(&dlm->spinlock);
3222         assert_spin_locked(&dlm->master_lock);
3223
3224         /* caller is responsible for any ref taken here on oldmle */
3225         found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3226         if (found) {
3227                 struct dlm_master_list_entry *tmp = *oldmle;
3228                 spin_lock(&tmp->spinlock);
3229                 if (tmp->type == DLM_MLE_MIGRATION) {
3230                         if (master == dlm->node_num) {
3231                                 /* ah another process raced me to it */
3232                                 mlog(0, "tried to migrate %.*s, but some "
3233                                      "process beat me to it\n",
3234                                      namelen, name);
3235                                 spin_unlock(&tmp->spinlock);
3236                                 return -EEXIST;
3237                         } else {
3238                                 /* bad.  2 NODES are trying to migrate! */
3239                                 mlog(ML_ERROR, "migration error  mle: "
3240                                      "master=%u new_master=%u // request: "
3241                                      "master=%u new_master=%u // "
3242                                      "lockres=%.*s\n",
3243                                      tmp->master, tmp->new_master,
3244                                      master, new_master,
3245                                      namelen, name);
3246                                 BUG();
3247                         }
3248                 } else {
3249                         /* this is essentially what assert_master does */
3250                         tmp->master = master;
3251                         atomic_set(&tmp->woken, 1);
3252                         wake_up(&tmp->wq);
3253                         /* remove it so that only one mle will be found */
3254                         __dlm_unlink_mle(dlm, tmp);
3255                         __dlm_mle_detach_hb_events(dlm, tmp);
3256                         if (tmp->type == DLM_MLE_MASTER) {
3257                                 ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3258                                 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3259                                                 "telling master to get ref "
3260                                                 "for cleared out mle during "
3261                                                 "migration\n", dlm->name,
3262                                                 namelen, name, master,
3263                                                 new_master);
3264                         }
3265                 }
3266                 spin_unlock(&tmp->spinlock);
3267         }
3268
3269         /* now add a migration mle to the tail of the list */
3270         dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3271         mle->new_master = new_master;
3272         /* the new master will be sending an assert master for this.
3273          * at that point we will get the refmap reference */
3274         mle->master = master;
3275         /* do this for consistency with other mle types */
3276         set_bit(new_master, mle->maybe_map);
3277         __dlm_insert_mle(dlm, mle);
3278
3279         return ret;
3280 }
3281
3282 /*
3283  * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3284  */
3285 static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3286                                         struct dlm_master_list_entry *mle)
3287 {
3288         struct dlm_lock_resource *res;
3289
3290         /* Find the lockres associated to the mle and set its owner to UNK */
3291         res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3292                                    mle->mnamehash);
3293         if (res) {
3294                 spin_unlock(&dlm->master_lock);
3295
3296                 /* move lockres onto recovery list */
3297                 spin_lock(&res->spinlock);
3298                 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3299                 dlm_move_lockres_to_recovery_list(dlm, res);
3300                 spin_unlock(&res->spinlock);
3301                 dlm_lockres_put(res);
3302
3303                 /* about to get rid of mle, detach from heartbeat */
3304                 __dlm_mle_detach_hb_events(dlm, mle);
3305
3306                 /* dump the mle */
3307                 spin_lock(&dlm->master_lock);
3308                 __dlm_put_mle(mle);
3309                 spin_unlock(&dlm->master_lock);
3310         }
3311
3312         return res;
3313 }
3314
3315 static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3316                                     struct dlm_master_list_entry *mle)
3317 {
3318         __dlm_mle_detach_hb_events(dlm, mle);
3319
3320         spin_lock(&mle->spinlock);
3321         __dlm_unlink_mle(dlm, mle);
3322         atomic_set(&mle->woken, 1);
3323         spin_unlock(&mle->spinlock);
3324
3325         wake_up(&mle->wq);
3326 }
3327
3328 static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3329                                 struct dlm_master_list_entry *mle, u8 dead_node)
3330 {
3331         int bit;
3332
3333         BUG_ON(mle->type != DLM_MLE_BLOCK);
3334
3335         spin_lock(&mle->spinlock);
3336         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3337         if (bit != dead_node) {
3338                 mlog(0, "mle found, but dead node %u would not have been "
3339                      "master\n", dead_node);
3340                 spin_unlock(&mle->spinlock);
3341         } else {
3342                 /* Must drop the refcount by one since the assert_master will
3343                  * never arrive. This may result in the mle being unlinked and
3344                  * freed, but there may still be a process waiting in the
3345                  * dlmlock path which is fine. */
3346                 mlog(0, "node %u was expected master\n", dead_node);
3347                 atomic_set(&mle->woken, 1);
3348                 spin_unlock(&mle->spinlock);
3349                 wake_up(&mle->wq);
3350
3351                 /* Do not need events any longer, so detach from heartbeat */
3352                 __dlm_mle_detach_hb_events(dlm, mle);
3353                 __dlm_put_mle(mle);
3354         }
3355 }
3356
3357 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3358 {
3359         struct dlm_master_list_entry *mle;
3360         struct dlm_lock_resource *res;
3361         struct hlist_head *bucket;
3362         struct hlist_node *tmp;
3363         unsigned int i;
3364
3365         mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
3366 top:
3367         assert_spin_locked(&dlm->spinlock);
3368
3369         /* clean the master list */
3370         spin_lock(&dlm->master_lock);
3371         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3372                 bucket = dlm_master_hash(dlm, i);
3373                 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3374                         BUG_ON(mle->type != DLM_MLE_BLOCK &&
3375                                mle->type != DLM_MLE_MASTER &&
3376                                mle->type != DLM_MLE_MIGRATION);
3377
3378                         /* MASTER mles are initiated locally. The waiting
3379                          * process will notice the node map change shortly.
3380                          * Let that happen as normal. */
3381                         if (mle->type == DLM_MLE_MASTER)
3382                                 continue;
3383
3384                         /* BLOCK mles are initiated by other nodes. Need to
3385                          * clean up if the dead node would have been the
3386                          * master. */
3387                         if (mle->type == DLM_MLE_BLOCK) {
3388                                 dlm_clean_block_mle(dlm, mle, dead_node);
3389                                 continue;
3390                         }
3391
3392                         /* Everything else is a MIGRATION mle */
3393
3394                         /* The rule for MIGRATION mles is that the master
3395                          * becomes UNKNOWN if *either* the original or the new
3396                          * master dies. All UNKNOWN lockres' are sent to
3397                          * whichever node becomes the recovery master. The new
3398                          * master is responsible for determining if there is
3399                          * still a master for this lockres, or if he needs to
3400                          * take over mastery. Either way, this node should
3401                          * expect another message to resolve this. */
3402
3403                         if (mle->master != dead_node &&
3404                             mle->new_master != dead_node)
3405                                 continue;
3406
3407                         if (mle->new_master == dead_node && mle->inuse) {
3408                                 mlog(ML_NOTICE, "%s: target %u died during "
3409                                                 "migration from %u, the MLE is "
3410                                                 "still keep used, ignore it!\n",
3411                                                 dlm->name, dead_node,
3412                                                 mle->master);
3413                                 continue;
3414                         }
3415
3416                         /* If we have reached this point, this mle needs to be
3417                          * removed from the list and freed. */
3418                         dlm_clean_migration_mle(dlm, mle);
3419
3420                         mlog(0, "%s: node %u died during migration from "
3421                              "%u to %u!\n", dlm->name, dead_node, mle->master,
3422                              mle->new_master);
3423
3424                         /* If we find a lockres associated with the mle, we've
3425                          * hit this rare case that messes up our lock ordering.
3426                          * If so, we need to drop the master lock so that we can
3427                          * take the lockres lock, meaning that we will have to
3428                          * restart from the head of list. */
3429                         res = dlm_reset_mleres_owner(dlm, mle);
3430                         if (res)
3431                                 /* restart */
3432                                 goto top;
3433
3434                         /* This may be the last reference */
3435                         __dlm_put_mle(mle);
3436                 }
3437         }
3438         spin_unlock(&dlm->master_lock);
3439 }
3440
3441 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3442                          u8 old_master)
3443 {
3444         struct dlm_node_iter iter;
3445         int ret = 0;
3446
3447         spin_lock(&dlm->spinlock);
3448         dlm_node_iter_init(dlm->domain_map, &iter);
3449         clear_bit(old_master, iter.node_map);
3450         clear_bit(dlm->node_num, iter.node_map);
3451         spin_unlock(&dlm->spinlock);
3452
3453         /* ownership of the lockres is changing.  account for the
3454          * mastery reference here since old_master will briefly have
3455          * a reference after the migration completes */
3456         spin_lock(&res->spinlock);
3457         dlm_lockres_set_refmap_bit(dlm, res, old_master);
3458         spin_unlock(&res->spinlock);
3459
3460         mlog(0, "now time to do a migrate request to other nodes\n");
3461         ret = dlm_do_migrate_request(dlm, res, old_master,
3462                                      dlm->node_num, &iter);
3463         if (ret < 0) {
3464                 mlog_errno(ret);
3465                 goto leave;
3466         }
3467
3468         mlog(0, "doing assert master of %.*s to all except the original node\n",
3469              res->lockname.len, res->lockname.name);
3470         /* this call now finishes out the nodemap
3471          * even if one or more nodes die */
3472         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3473                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3474         if (ret < 0) {
3475                 /* no longer need to retry.  all living nodes contacted. */
3476                 mlog_errno(ret);
3477                 ret = 0;
3478         }
3479
3480         memset(iter.node_map, 0, sizeof(iter.node_map));
3481         set_bit(old_master, iter.node_map);
3482         mlog(0, "doing assert master of %.*s back to %u\n",
3483              res->lockname.len, res->lockname.name, old_master);
3484         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3485                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3486         if (ret < 0) {
3487                 mlog(0, "assert master to original master failed "
3488                      "with %d.\n", ret);
3489                 /* the only nonzero status here would be because of
3490                  * a dead original node.  we're done. */
3491                 ret = 0;
3492         }
3493
3494         /* all done, set the owner, clear the flag */
3495         spin_lock(&res->spinlock);
3496         dlm_set_lockres_owner(dlm, res, dlm->node_num);
3497         res->state &= ~DLM_LOCK_RES_MIGRATING;
3498         spin_unlock(&res->spinlock);
3499         /* re-dirty it on the new master */
3500         dlm_kick_thread(dlm, res);
3501         wake_up(&res->wq);
3502 leave:
3503         return ret;
3504 }
3505
3506 /*
3507  * LOCKRES AST REFCOUNT
3508  * this is integral to migration
3509  */
3510
3511 /* for future intent to call an ast, reserve one ahead of time.
3512  * this should be called only after waiting on the lockres
3513  * with dlm_wait_on_lockres, and while still holding the
3514  * spinlock after the call. */
3515 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3516 {
3517         assert_spin_locked(&res->spinlock);
3518         if (res->state & DLM_LOCK_RES_MIGRATING) {
3519                 __dlm_print_one_lock_resource(res);
3520         }
3521         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3522
3523         atomic_inc(&res->asts_reserved);
3524 }
3525
3526 /*
3527  * used to drop the reserved ast, either because it went unused,
3528  * or because the ast/bast was actually called.
3529  *
3530  * also, if there is a pending migration on this lockres,
3531  * and this was the last pending ast on the lockres,
3532  * atomically set the MIGRATING flag before we drop the lock.
3533  * this is how we ensure that migration can proceed with no
3534  * asts in progress.  note that it is ok if the state of the
3535  * queues is such that a lock should be granted in the future
3536  * or that a bast should be fired, because the new master will
3537  * shuffle the lists on this lockres as soon as it is migrated.
3538  */
3539 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3540                              struct dlm_lock_resource *res)
3541 {
3542         if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3543                 return;
3544
3545         if (!res->migration_pending) {
3546                 spin_unlock(&res->spinlock);
3547                 return;
3548         }
3549
3550         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3551         res->migration_pending = 0;
3552         res->state |= DLM_LOCK_RES_MIGRATING;
3553         spin_unlock(&res->spinlock);
3554         wake_up(&res->wq);
3555         wake_up(&dlm->migration_wq);
3556 }
3557
3558 void dlm_force_free_mles(struct dlm_ctxt *dlm)
3559 {
3560         int i;
3561         struct hlist_head *bucket;
3562         struct dlm_master_list_entry *mle;
3563         struct hlist_node *tmp;
3564
3565         /*
3566          * We notified all other nodes that we are exiting the domain and
3567          * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3568          * around we force free them and wake any processes that are waiting
3569          * on the mles
3570          */
3571         spin_lock(&dlm->spinlock);
3572         spin_lock(&dlm->master_lock);
3573
3574         BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3575         BUG_ON((find_next_bit(dlm->domain_map, O2NM_MAX_NODES, 0) < O2NM_MAX_NODES));
3576
3577         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3578                 bucket = dlm_master_hash(dlm, i);
3579                 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3580                         if (mle->type != DLM_MLE_BLOCK) {
3581                                 mlog(ML_ERROR, "bad mle: %p\n", mle);
3582                                 dlm_print_one_mle(mle);
3583                         }
3584                         atomic_set(&mle->woken, 1);
3585                         wake_up(&mle->wq);
3586
3587                         __dlm_unlink_mle(dlm, mle);
3588                         __dlm_mle_detach_hb_events(dlm, mle);
3589                         __dlm_put_mle(mle);
3590                 }
3591         }
3592         spin_unlock(&dlm->master_lock);
3593         spin_unlock(&dlm->spinlock);
3594 }