fs/ocfs2/dlm/dlmmaster.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmmod.c
   5  *
   6  * standalone DLM module
   7  *
   8  * Copyright (C) 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  *
  25  */
  26
  27
  28 #include <linux/module.h>
  29 #include <linux/fs.h>
  30 #include <linux/types.h>
  31 #include <linux/slab.h>
  32 #include <linux/highmem.h>
  33 #include <linux/init.h>
  34 #include <linux/sysctl.h>
  35 #include <linux/random.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/socket.h>
  38 #include <linux/inet.h>
  39 #include <linux/spinlock.h>
  40 #include <linux/delay.h>
  41
  42
  43 #include "cluster/heartbeat.h"
  44 #include "cluster/nodemanager.h"
  45 #include "cluster/tcp.h"
  46
  47 #include "dlmapi.h"
  48 #include "dlmcommon.h"
  49 #include "dlmdomain.h"
  50 #include "dlmdebug.h"
  51
  52 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
  53 #include "cluster/masklog.h"
  54
  55 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
  56                               struct dlm_master_list_entry *mle,
  57                               struct o2nm_node *node,
  58                               int idx);
  59 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
  60                             struct dlm_master_list_entry *mle,
  61                             struct o2nm_node *node,
  62                             int idx);
  63
  64 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
  65 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
  66                                 struct dlm_lock_resource *res,
  67                                 void *nodemap, u32 flags);
  68 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
  69
  70 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
  71                                 struct dlm_master_list_entry *mle,
  72                                 const char *name,
  73                                 unsigned int namelen)
  74 {
  75         if (dlm != mle->dlm)
  76                 return 0;
  77
  78         if (namelen != mle->mnamelen ||
  79             memcmp(name, mle->mname, namelen) != 0)
  80                 return 0;
  81
  82         return 1;
  83 }
  84
  85 static struct kmem_cache *dlm_lockres_cache;
  86 static struct kmem_cache *dlm_lockname_cache;
  87 static struct kmem_cache *dlm_mle_cache;
  88
  89 static void dlm_mle_release(struct kref *kref);
  90 static void dlm_init_mle(struct dlm_master_list_entry *mle,
  91                         enum dlm_mle_type type,
  92                         struct dlm_ctxt *dlm,
  93                         struct dlm_lock_resource *res,
  94                         const char *name,
  95                         unsigned int namelen);
  96 static void dlm_put_mle(struct dlm_master_list_entry *mle);
  97 static void __dlm_put_mle(struct dlm_master_list_entry *mle);
  98 static int dlm_find_mle(struct dlm_ctxt *dlm,
  99                         struct dlm_master_list_entry **mle,
 100                         char *name, unsigned int namelen);
 101
 102 static int dlm_do_master_request(struct dlm_lock_resource *res,
 103                                  struct dlm_master_list_entry *mle, int to);
 104
 105
 106 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
 107                                      struct dlm_lock_resource *res,
 108                                      struct dlm_master_list_entry *mle,
 109                                      int *blocked);
 110 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
 111                                     struct dlm_lock_resource *res,
 112                                     struct dlm_master_list_entry *mle,
 113                                     int blocked);
 114 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
 115                                  struct dlm_lock_resource *res,
 116                                  struct dlm_master_list_entry *mle,
 117                                  struct dlm_master_list_entry **oldmle,
 118                                  const char *name, unsigned int namelen,
 119                                  u8 new_master, u8 master);
 120
 121 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
 122                                     struct dlm_lock_resource *res);
 123 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
 124                                       struct dlm_lock_resource *res);
 125 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
 126                                        struct dlm_lock_resource *res,
 127                                        u8 target);
 128 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
 129                                        struct dlm_lock_resource *res);
 130
 131
 132 int dlm_is_host_down(int errno)
 133 {
 134         switch (errno) {
 135                 case -EBADF:
 136                 case -ECONNREFUSED:
 137                 case -ENOTCONN:
 138                 case -ECONNRESET:
 139                 case -EPIPE:
 140                 case -EHOSTDOWN:
 141                 case -EHOSTUNREACH:
 142                 case -ETIMEDOUT:
 143                 case -ECONNABORTED:
 144                 case -ENETDOWN:
 145                 case -ENETUNREACH:
 146                 case -ENETRESET:
 147                 case -ESHUTDOWN:
 148                 case -ENOPROTOOPT:
 149                 case -EINVAL:   /* if returned from our tcp code,
 150                                    this means there is no socket */
 151                         return 1;
 152         }
 153         return 0;
 154 }
 155
 156
 157 /*
 158  * MASTER LIST FUNCTIONS
 159  */
 160
 161
 162 /*
 163  * regarding master list entries and heartbeat callbacks:
 164  *
 165  * in order to avoid sleeping and allocation that occurs in
 166  * heartbeat, master list entries are simply attached to the
 167  * dlm's established heartbeat callbacks.  the mle is attached
 168  * when it is created, and since the dlm->spinlock is held at
 169  * that time, any heartbeat event will be properly discovered
 170  * by the mle.  the mle needs to be detached from the
 171  * dlm->mle_hb_events list as soon as heartbeat events are no
 172  * longer useful to the mle, and before the mle is freed.
 173  *
 174  * as a general rule, heartbeat events are no longer needed by
 175  * the mle once an "answer" regarding the lock master has been
 176  * received.
 177  */
 178 static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
 179                                               struct dlm_master_list_entry *mle)
 180 {
 181         assert_spin_locked(&dlm->spinlock);
 182
 183         list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
 184 }
 185
 186
 187 static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 188                                               struct dlm_master_list_entry *mle)
 189 {
 190         if (!list_empty(&mle->hb_events))
 191                 list_del_init(&mle->hb_events);
 192 }
 193
 194
 195 static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 196                                             struct dlm_master_list_entry *mle)
 197 {
 198         spin_lock(&dlm->spinlock);
 199         __dlm_mle_detach_hb_events(dlm, mle);
 200         spin_unlock(&dlm->spinlock);
 201 }
 202
 203 static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
 204 {
 205         struct dlm_ctxt *dlm;
 206         dlm = mle->dlm;
 207
 208         assert_spin_locked(&dlm->spinlock);
 209         assert_spin_locked(&dlm->master_lock);
 210         mle->inuse++;
 211         kref_get(&mle->mle_refs);
 212 }
 213
 214 static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
 215 {
 216         struct dlm_ctxt *dlm;
 217         dlm = mle->dlm;
 218
 219         spin_lock(&dlm->spinlock);
 220         spin_lock(&dlm->master_lock);
 221         mle->inuse--;
 222         __dlm_put_mle(mle);
 223         spin_unlock(&dlm->master_lock);
 224         spin_unlock(&dlm->spinlock);
 225
 226 }
 227
 228 /* remove from list and free */
 229 static void __dlm_put_mle(struct dlm_master_list_entry *mle)
 230 {
 231         struct dlm_ctxt *dlm;
 232         dlm = mle->dlm;
 233
 234         assert_spin_locked(&dlm->spinlock);
 235         assert_spin_locked(&dlm->master_lock);
 236         if (!atomic_read(&mle->mle_refs.refcount)) {
 237                 /* this may or may not crash, but who cares.
 238                  * it's a BUG. */
 239                 mlog(ML_ERROR, "bad mle: %p\n", mle);
 240                 dlm_print_one_mle(mle);
 241                 BUG();
 242         } else
 243                 kref_put(&mle->mle_refs, dlm_mle_release);
 244 }
 245
 246
 247 /* must not have any spinlocks coming in */
 248 static void dlm_put_mle(struct dlm_master_list_entry *mle)
 249 {
 250         struct dlm_ctxt *dlm;
 251         dlm = mle->dlm;
 252
 253         spin_lock(&dlm->spinlock);
 254         spin_lock(&dlm->master_lock);
 255         __dlm_put_mle(mle);
 256         spin_unlock(&dlm->master_lock);
 257         spin_unlock(&dlm->spinlock);
 258 }
 259
 260 static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
 261 {
 262         kref_get(&mle->mle_refs);
 263 }
 264
 265 static void dlm_init_mle(struct dlm_master_list_entry *mle,
 266                         enum dlm_mle_type type,
 267                         struct dlm_ctxt *dlm,
 268                         struct dlm_lock_resource *res,
 269                         const char *name,
 270                         unsigned int namelen)
 271 {
 272         assert_spin_locked(&dlm->spinlock);
 273
 274         mle->dlm = dlm;
 275         mle->type = type;
 276         INIT_HLIST_NODE(&mle->master_hash_node);
 277         INIT_LIST_HEAD(&mle->hb_events);
 278         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
 279         spin_lock_init(&mle->spinlock);
 280         init_waitqueue_head(&mle->wq);
 281         atomic_set(&mle->woken, 0);
 282         kref_init(&mle->mle_refs);
 283         memset(mle->response_map, 0, sizeof(mle->response_map));
 284         mle->master = O2NM_MAX_NODES;
 285         mle->new_master = O2NM_MAX_NODES;
 286         mle->inuse = 0;
 287
 288         BUG_ON(mle->type != DLM_MLE_BLOCK &&
 289                mle->type != DLM_MLE_MASTER &&
 290                mle->type != DLM_MLE_MIGRATION);
 291
 292         if (mle->type == DLM_MLE_MASTER) {
 293                 BUG_ON(!res);
 294                 mle->mleres = res;
 295                 memcpy(mle->mname, res->lockname.name, res->lockname.len);
 296                 mle->mnamelen = res->lockname.len;
 297                 mle->mnamehash = res->lockname.hash;
 298         } else {
 299                 BUG_ON(!name);
 300                 mle->mleres = NULL;
 301                 memcpy(mle->mname, name, namelen);
 302                 mle->mnamelen = namelen;
 303                 mle->mnamehash = dlm_lockid_hash(name, namelen);
 304         }
 305
 306         atomic_inc(&dlm->mle_tot_count[mle->type]);
 307         atomic_inc(&dlm->mle_cur_count[mle->type]);
 308
 309         /* copy off the node_map and register hb callbacks on our copy */
 310         memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
 311         memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
 312         clear_bit(dlm->node_num, mle->vote_map);
 313         clear_bit(dlm->node_num, mle->node_map);
 314
 315         /* attach the mle to the domain node up/down events */
 316         __dlm_mle_attach_hb_events(dlm, mle);
 317 }
 318
 319 void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
 320 {
 321         assert_spin_locked(&dlm->spinlock);
 322         assert_spin_locked(&dlm->master_lock);
 323
 324         if (!hlist_unhashed(&mle->master_hash_node))
 325                 hlist_del_init(&mle->master_hash_node);
 326 }
 327
 328 void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
 329 {
 330         struct hlist_head *bucket;
 331
 332         assert_spin_locked(&dlm->master_lock);
 333
 334         bucket = dlm_master_hash(dlm, mle->mnamehash);
 335         hlist_add_head(&mle->master_hash_node, bucket);
 336 }
 337
 338 /* returns 1 if found, 0 if not */
 339 static int dlm_find_mle(struct dlm_ctxt *dlm,
 340                         struct dlm_master_list_entry **mle,
 341                         char *name, unsigned int namelen)
 342 {
 343         struct dlm_master_list_entry *tmpmle;
 344         struct hlist_head *bucket;
 345         unsigned int hash;
 346
 347         assert_spin_locked(&dlm->master_lock);
 348
 349         hash = dlm_lockid_hash(name, namelen);
 350         bucket = dlm_master_hash(dlm, hash);
 351         hlist_for_each_entry(tmpmle, bucket, master_hash_node) {
 352                 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
 353                         continue;
 354                 dlm_get_mle(tmpmle);
 355                 *mle = tmpmle;
 356                 return 1;
 357         }
 358         return 0;
 359 }
 360
 361 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
 362 {
 363         struct dlm_master_list_entry *mle;
 364
 365         assert_spin_locked(&dlm->spinlock);
 366
 367         list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
 368                 if (node_up)
 369                         dlm_mle_node_up(dlm, mle, NULL, idx);
 370                 else
 371                         dlm_mle_node_down(dlm, mle, NULL, idx);
 372         }
 373 }
 374
 375 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
 376                               struct dlm_master_list_entry *mle,
 377                               struct o2nm_node *node, int idx)
 378 {
 379         spin_lock(&mle->spinlock);
 380
 381         if (!test_bit(idx, mle->node_map))
 382                 mlog(0, "node %u already removed from nodemap!\n", idx);
 383         else
 384                 clear_bit(idx, mle->node_map);
 385
 386         spin_unlock(&mle->spinlock);
 387 }
 388
 389 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
 390                             struct dlm_master_list_entry *mle,
 391                             struct o2nm_node *node, int idx)
 392 {
 393         spin_lock(&mle->spinlock);
 394
 395         if (test_bit(idx, mle->node_map))
 396                 mlog(0, "node %u already in node map!\n", idx);
 397         else
 398                 set_bit(idx, mle->node_map);
 399
 400         spin_unlock(&mle->spinlock);
 401 }
 402
 403
 404 int dlm_init_mle_cache(void)
 405 {
 406         dlm_mle_cache = kmem_cache_create("o2dlm_mle",
 407                                           sizeof(struct dlm_master_list_entry),
 408                                           0, SLAB_HWCACHE_ALIGN,
 409                                           NULL);
 410         if (dlm_mle_cache == NULL)
 411                 return -ENOMEM;
 412         return 0;
 413 }
 414
 415 void dlm_destroy_mle_cache(void)
 416 {
 417         if (dlm_mle_cache)
 418                 kmem_cache_destroy(dlm_mle_cache);
 419 }
 420
 421 static void dlm_mle_release(struct kref *kref)
 422 {
 423         struct dlm_master_list_entry *mle;
 424         struct dlm_ctxt *dlm;
 425
 426         mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
 427         dlm = mle->dlm;
 428
 429         assert_spin_locked(&dlm->spinlock);
 430         assert_spin_locked(&dlm->master_lock);
 431
 432         mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
 433              mle->type);
 434
 435         /* remove from list if not already */
 436         __dlm_unlink_mle(dlm, mle);
 437
 438         /* detach the mle from the domain node up/down events */
 439         __dlm_mle_detach_hb_events(dlm, mle);
 440
 441         atomic_dec(&dlm->mle_cur_count[mle->type]);
 442
 443         /* NOTE: kfree under spinlock here.
 444          * if this is bad, we can move this to a freelist. */
 445         kmem_cache_free(dlm_mle_cache, mle);
 446 }
 447
 448
 449 /*
 450  * LOCK RESOURCE FUNCTIONS
 451  */
 452
 453 int dlm_init_master_caches(void)
 454 {
 455         dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
 456                                               sizeof(struct dlm_lock_resource),
 457                                               0, SLAB_HWCACHE_ALIGN, NULL);
 458         if (!dlm_lockres_cache)
 459                 goto bail;
 460
 461         dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
 462                                                DLM_LOCKID_NAME_MAX, 0,
 463                                                SLAB_HWCACHE_ALIGN, NULL);
 464         if (!dlm_lockname_cache)
 465                 goto bail;
 466
 467         return 0;
 468 bail:
 469         dlm_destroy_master_caches();
 470         return -ENOMEM;
 471 }
 472
 473 void dlm_destroy_master_caches(void)
 474 {
 475         if (dlm_lockname_cache) {
 476                 kmem_cache_destroy(dlm_lockname_cache);
 477                 dlm_lockname_cache = NULL;
 478         }
 479
 480         if (dlm_lockres_cache) {
 481                 kmem_cache_destroy(dlm_lockres_cache);
 482                 dlm_lockres_cache = NULL;
 483         }
 484 }
 485
 486 static void dlm_lockres_release(struct kref *kref)
 487 {
 488         struct dlm_lock_resource *res;
 489         struct dlm_ctxt *dlm;
 490
 491         res = container_of(kref, struct dlm_lock_resource, refs);
 492         dlm = res->dlm;
 493
 494         /* This should not happen -- all lockres' have a name
 495          * associated with them at init time. */
 496         BUG_ON(!res->lockname.name);
 497
 498         mlog(0, "destroying lockres %.*s\n", res->lockname.len,
 499              res->lockname.name);
 500
 501         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->spinlock);
 593         list_add_tail(&res->tracking, &dlm->tracking_list);
 594         spin_unlock(&dlm->spinlock);
 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 (atomic_read(&mle->mle_refs.refcount) < 2)
1128                         mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1129                         atomic_read(&mle->mle_refs.refcount),
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                 // mlog(0, "mle was found\n");
1613                 set_maybe = 1;
1614                 spin_lock(&tmpmle->spinlock);
1615                 if (tmpmle->master == dlm->node_num) {
1616                         mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1617                         BUG();
1618                 }
1619                 if (tmpmle->type == DLM_MLE_BLOCK)
1620                         response = DLM_MASTER_RESP_NO;
1621                 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1622                         mlog(0, "migration mle was found (%u->%u)\n",
1623                              tmpmle->master, tmpmle->new_master);
1624                         /* real master can respond on its own */
1625                         response = DLM_MASTER_RESP_NO;
1626                 } else
1627                         response = DLM_MASTER_RESP_MAYBE;
1628                 if (set_maybe)
1629                         set_bit(request->node_idx, tmpmle->maybe_map);
1630                 spin_unlock(&tmpmle->spinlock);
1631         }
1632         spin_unlock(&dlm->master_lock);
1633         spin_unlock(&dlm->spinlock);
1634
1635         if (found) {
1636                 /* keep the mle attached to heartbeat events */
1637                 dlm_put_mle(tmpmle);
1638         }
1639 send_response:
1640         /*
1641          * __dlm_lookup_lockres() grabbed a reference to this lockres.
1642          * The reference is released by dlm_assert_master_worker() under
1643          * the call to dlm_dispatch_assert_master().  If
1644          * dlm_assert_master_worker() isn't called, we drop it here.
1645          */
1646         if (dispatch_assert) {
1647                 if (response != DLM_MASTER_RESP_YES)
1648                         mlog(ML_ERROR, "invalid response %d\n", response);
1649                 if (!res) {
1650                         mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1651                         BUG();
1652                 }
1653                 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1654                              dlm->node_num, res->lockname.len, res->lockname.name);
1655                 spin_lock(&res->spinlock);
1656                 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1657                                                  DLM_ASSERT_MASTER_MLE_CLEANUP);
1658                 if (ret < 0) {
1659                         mlog(ML_ERROR, "failed to dispatch assert master work\n");
1660                         response = DLM_MASTER_RESP_ERROR;
1661                         spin_unlock(&res->spinlock);
1662                         dlm_lockres_put(res);
1663                 } else {
1664                         dispatched = 1;
1665                         __dlm_lockres_grab_inflight_worker(dlm, res);
1666                         spin_unlock(&res->spinlock);
1667                 }
1668         } else {
1669                 if (res)
1670                         dlm_lockres_put(res);
1671         }
1672
1673         if (!dispatched)
1674                 dlm_put(dlm);
1675         return response;
1676 }
1677
1678 /*
1679  * DLM_ASSERT_MASTER_MSG
1680  */
1681
1682
1683 /*
1684  * NOTE: this can be used for debugging
1685  * can periodically run all locks owned by this node
1686  * and re-assert across the cluster...
1687  */
1688 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1689                                 struct dlm_lock_resource *res,
1690                                 void *nodemap, u32 flags)
1691 {
1692         struct dlm_assert_master assert;
1693         int to, tmpret;
1694         struct dlm_node_iter iter;
1695         int ret = 0;
1696         int reassert;
1697         const char *lockname = res->lockname.name;
1698         unsigned int namelen = res->lockname.len;
1699
1700         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1701
1702         spin_lock(&res->spinlock);
1703         res->state |= DLM_LOCK_RES_SETREF_INPROG;
1704         spin_unlock(&res->spinlock);
1705
1706 again:
1707         reassert = 0;
1708
1709         /* note that if this nodemap is empty, it returns 0 */
1710         dlm_node_iter_init(nodemap, &iter);
1711         while ((to = dlm_node_iter_next(&iter)) >= 0) {
1712                 int r = 0;
1713                 struct dlm_master_list_entry *mle = NULL;
1714
1715                 mlog(0, "sending assert master to %d (%.*s)\n", to,
1716                      namelen, lockname);
1717                 memset(&assert, 0, sizeof(assert));
1718                 assert.node_idx = dlm->node_num;
1719                 assert.namelen = namelen;
1720                 memcpy(assert.name, lockname, namelen);
1721                 assert.flags = cpu_to_be32(flags);
1722
1723                 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1724                                             &assert, sizeof(assert), to, &r);
1725                 if (tmpret < 0) {
1726                         mlog(ML_ERROR, "Error %d when sending message %u (key "
1727                              "0x%x) to node %u\n", tmpret,
1728                              DLM_ASSERT_MASTER_MSG, dlm->key, to);
1729                         if (!dlm_is_host_down(tmpret)) {
1730                                 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1731                                 BUG();
1732                         }
1733                         /* a node died.  finish out the rest of the nodes. */
1734                         mlog(0, "link to %d went down!\n", to);
1735                         /* any nonzero status return will do */
1736                         ret = tmpret;
1737                         r = 0;
1738                 } else if (r < 0) {
1739                         /* ok, something horribly messed.  kill thyself. */
1740                         mlog(ML_ERROR,"during assert master of %.*s to %u, "
1741                              "got %d.\n", namelen, lockname, to, r);
1742                         spin_lock(&dlm->spinlock);
1743                         spin_lock(&dlm->master_lock);
1744                         if (dlm_find_mle(dlm, &mle, (char *)lockname,
1745                                          namelen)) {
1746                                 dlm_print_one_mle(mle);
1747                                 __dlm_put_mle(mle);
1748                         }
1749                         spin_unlock(&dlm->master_lock);
1750                         spin_unlock(&dlm->spinlock);
1751                         BUG();
1752                 }
1753
1754                 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1755                     !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1756                                 mlog(ML_ERROR, "%.*s: very strange, "
1757                                      "master MLE but no lockres on %u\n",
1758                                      namelen, lockname, to);
1759                 }
1760
1761                 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1762                         mlog(0, "%.*s: node %u create mles on other "
1763                              "nodes and requests a re-assert\n",
1764                              namelen, lockname, to);
1765                         reassert = 1;
1766                 }
1767                 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1768                         mlog(0, "%.*s: node %u has a reference to this "
1769                              "lockres, set the bit in the refmap\n",
1770                              namelen, lockname, to);
1771                         spin_lock(&res->spinlock);
1772                         dlm_lockres_set_refmap_bit(dlm, res, to);
1773                         spin_unlock(&res->spinlock);
1774                 }
1775         }
1776
1777         if (reassert)
1778                 goto again;
1779
1780         spin_lock(&res->spinlock);
1781         res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1782         spin_unlock(&res->spinlock);
1783         wake_up(&res->wq);
1784
1785         return ret;
1786 }
1787
1788 /*
1789  * locks that can be taken here:
1790  * dlm->spinlock
1791  * res->spinlock
1792  * mle->spinlock
1793  * dlm->master_list
1794  *
1795  * if possible, TRIM THIS DOWN!!!
1796  */
1797 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1798                               void **ret_data)
1799 {
1800         struct dlm_ctxt *dlm = data;
1801         struct dlm_master_list_entry *mle = NULL;
1802         struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1803         struct dlm_lock_resource *res = NULL;
1804         char *name;
1805         unsigned int namelen, hash;
1806         u32 flags;
1807         int master_request = 0, have_lockres_ref = 0;
1808         int ret = 0;
1809
1810         if (!dlm_grab(dlm))
1811                 return 0;
1812
1813         name = assert->name;
1814         namelen = assert->namelen;
1815         hash = dlm_lockid_hash(name, namelen);
1816         flags = be32_to_cpu(assert->flags);
1817
1818         if (namelen > DLM_LOCKID_NAME_MAX) {
1819                 mlog(ML_ERROR, "Invalid name length!");
1820                 goto done;
1821         }
1822
1823         spin_lock(&dlm->spinlock);
1824
1825         if (flags)
1826                 mlog(0, "assert_master with flags: %u\n", flags);
1827
1828         /* find the MLE */
1829         spin_lock(&dlm->master_lock);
1830         if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1831                 /* not an error, could be master just re-asserting */
1832                 mlog(0, "just got an assert_master from %u, but no "
1833                      "MLE for it! (%.*s)\n", assert->node_idx,
1834                      namelen, name);
1835         } else {
1836                 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1837                 if (bit >= O2NM_MAX_NODES) {
1838                         /* not necessarily an error, though less likely.
1839                          * could be master just re-asserting. */
1840                         mlog(0, "no bits set in the maybe_map, but %u "
1841                              "is asserting! (%.*s)\n", assert->node_idx,
1842                              namelen, name);
1843                 } else if (bit != assert->node_idx) {
1844                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1845                                 mlog(0, "master %u was found, %u should "
1846                                      "back off\n", assert->node_idx, bit);
1847                         } else {
1848                                 /* with the fix for bug 569, a higher node
1849                                  * number winning the mastery will respond
1850                                  * YES to mastery requests, but this node
1851                                  * had no way of knowing.  let it pass. */
1852                                 mlog(0, "%u is the lowest node, "
1853                                      "%u is asserting. (%.*s)  %u must "
1854                                      "have begun after %u won.\n", bit,
1855                                      assert->node_idx, namelen, name, bit,
1856                                      assert->node_idx);
1857                         }
1858                 }
1859                 if (mle->type == DLM_MLE_MIGRATION) {
1860                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1861                                 mlog(0, "%s:%.*s: got cleanup assert"
1862                                      " from %u for migration\n",
1863                                      dlm->name, namelen, name,
1864                                      assert->node_idx);
1865                         } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1866                                 mlog(0, "%s:%.*s: got unrelated assert"
1867                                      " from %u for migration, ignoring\n",
1868                                      dlm->name, namelen, name,
1869                                      assert->node_idx);
1870                                 __dlm_put_mle(mle);
1871                                 spin_unlock(&dlm->master_lock);
1872                                 spin_unlock(&dlm->spinlock);
1873                                 goto done;
1874                         }
1875                 }
1876         }
1877         spin_unlock(&dlm->master_lock);
1878
1879         /* ok everything checks out with the MLE
1880          * now check to see if there is a lockres */
1881         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1882         if (res) {
1883                 spin_lock(&res->spinlock);
1884                 if (res->state & DLM_LOCK_RES_RECOVERING)  {
1885                         mlog(ML_ERROR, "%u asserting but %.*s is "
1886                              "RECOVERING!\n", assert->node_idx, namelen, name);
1887                         goto kill;
1888                 }
1889                 if (!mle) {
1890                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1891                             res->owner != assert->node_idx) {
1892                                 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1893                                      "but current owner is %u! (%.*s)\n",
1894                                      assert->node_idx, res->owner, namelen,
1895                                      name);
1896                                 __dlm_print_one_lock_resource(res);
1897                                 BUG();
1898                         }
1899                 } else if (mle->type != DLM_MLE_MIGRATION) {
1900                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1901                                 /* owner is just re-asserting */
1902                                 if (res->owner == assert->node_idx) {
1903                                         mlog(0, "owner %u re-asserting on "
1904                                              "lock %.*s\n", assert->node_idx,
1905                                              namelen, name);
1906                                         goto ok;
1907                                 }
1908                                 mlog(ML_ERROR, "got assert_master from "
1909                                      "node %u, but %u is the owner! "
1910                                      "(%.*s)\n", assert->node_idx,
1911                                      res->owner, namelen, name);
1912                                 goto kill;
1913                         }
1914                         if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1915                                 mlog(ML_ERROR, "got assert from %u, but lock "
1916                                      "with no owner should be "
1917                                      "in-progress! (%.*s)\n",
1918                                      assert->node_idx,
1919                                      namelen, name);
1920                                 goto kill;
1921                         }
1922                 } else /* mle->type == DLM_MLE_MIGRATION */ {
1923                         /* should only be getting an assert from new master */
1924                         if (assert->node_idx != mle->new_master) {
1925                                 mlog(ML_ERROR, "got assert from %u, but "
1926                                      "new master is %u, and old master "
1927                                      "was %u (%.*s)\n",
1928                                      assert->node_idx, mle->new_master,
1929                                      mle->master, namelen, name);
1930                                 goto kill;
1931                         }
1932
1933                 }
1934 ok:
1935                 spin_unlock(&res->spinlock);
1936         }
1937
1938         // mlog(0, "woo!  got an assert_master from node %u!\n",
1939         //           assert->node_idx);
1940         if (mle) {
1941                 int extra_ref = 0;
1942                 int nn = -1;
1943                 int rr, err = 0;
1944
1945                 spin_lock(&mle->spinlock);
1946                 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1947                         extra_ref = 1;
1948                 else {
1949                         /* MASTER mle: if any bits set in the response map
1950                          * then the calling node needs to re-assert to clear
1951                          * up nodes that this node contacted */
1952                         while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1953                                                     nn+1)) < O2NM_MAX_NODES) {
1954                                 if (nn != dlm->node_num && nn != assert->node_idx) {
1955                                         master_request = 1;
1956                                         break;
1957                                 }
1958                         }
1959                 }
1960                 mle->master = assert->node_idx;
1961                 atomic_set(&mle->woken, 1);
1962                 wake_up(&mle->wq);
1963                 spin_unlock(&mle->spinlock);
1964
1965                 if (res) {
1966                         int wake = 0;
1967                         spin_lock(&res->spinlock);
1968                         if (mle->type == DLM_MLE_MIGRATION) {
1969                                 mlog(0, "finishing off migration of lockres %.*s, "
1970                                         "from %u to %u\n",
1971                                         res->lockname.len, res->lockname.name,
1972                                         dlm->node_num, mle->new_master);
1973                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
1974                                 wake = 1;
1975                                 dlm_change_lockres_owner(dlm, res, mle->new_master);
1976                                 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1977                         } else {
1978                                 dlm_change_lockres_owner(dlm, res, mle->master);
1979                         }
1980                         spin_unlock(&res->spinlock);
1981                         have_lockres_ref = 1;
1982                         if (wake)
1983                                 wake_up(&res->wq);
1984                 }
1985
1986                 /* master is known, detach if not already detached.
1987                  * ensures that only one assert_master call will happen
1988                  * on this mle. */
1989                 spin_lock(&dlm->master_lock);
1990
1991                 rr = atomic_read(&mle->mle_refs.refcount);
1992                 if (mle->inuse > 0) {
1993                         if (extra_ref && rr < 3)
1994                                 err = 1;
1995                         else if (!extra_ref && rr < 2)
1996                                 err = 1;
1997                 } else {
1998                         if (extra_ref && rr < 2)
1999                                 err = 1;
2000                         else if (!extra_ref && rr < 1)
2001                                 err = 1;
2002                 }
2003                 if (err) {
2004                         mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
2005                              "that will mess up this node, refs=%d, extra=%d, "
2006                              "inuse=%d\n", dlm->name, namelen, name,
2007                              assert->node_idx, rr, extra_ref, mle->inuse);
2008                         dlm_print_one_mle(mle);
2009                 }
2010                 __dlm_unlink_mle(dlm, mle);
2011                 __dlm_mle_detach_hb_events(dlm, mle);
2012                 __dlm_put_mle(mle);
2013                 if (extra_ref) {
2014                         /* the assert master message now balances the extra
2015                          * ref given by the master / migration request message.
2016                          * if this is the last put, it will be removed
2017                          * from the list. */
2018                         __dlm_put_mle(mle);
2019                 }
2020                 spin_unlock(&dlm->master_lock);
2021         } else if (res) {
2022                 if (res->owner != assert->node_idx) {
2023                         mlog(0, "assert_master from %u, but current "
2024                              "owner is %u (%.*s), no mle\n", assert->node_idx,
2025                              res->owner, namelen, name);
2026                 }
2027         }
2028         spin_unlock(&dlm->spinlock);
2029
2030 done:
2031         ret = 0;
2032         if (res) {
2033                 spin_lock(&res->spinlock);
2034                 res->state |= DLM_LOCK_RES_SETREF_INPROG;
2035                 spin_unlock(&res->spinlock);
2036                 *ret_data = (void *)res;
2037         }
2038         dlm_put(dlm);
2039         if (master_request) {
2040                 mlog(0, "need to tell master to reassert\n");
2041                 /* positive. negative would shoot down the node. */
2042                 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2043                 if (!have_lockres_ref) {
2044                         mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2045                              "mle present here for %s:%.*s, but no lockres!\n",
2046                              assert->node_idx, dlm->name, namelen, name);
2047                 }
2048         }
2049         if (have_lockres_ref) {
2050                 /* let the master know we have a reference to the lockres */
2051                 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2052                 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2053                      dlm->name, namelen, name, assert->node_idx);
2054         }
2055         return ret;
2056
2057 kill:
2058         /* kill the caller! */
2059         mlog(ML_ERROR, "Bad message received from another node.  Dumping state "
2060              "and killing the other node now!  This node is OK and can continue.\n");
2061         __dlm_print_one_lock_resource(res);
2062         spin_unlock(&res->spinlock);
2063         spin_lock(&dlm->master_lock);
2064         if (mle)
2065                 __dlm_put_mle(mle);
2066         spin_unlock(&dlm->master_lock);
2067         spin_unlock(&dlm->spinlock);
2068         *ret_data = (void *)res;
2069         dlm_put(dlm);
2070         return -EINVAL;
2071 }
2072
2073 void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2074 {
2075         struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2076
2077         if (ret_data) {
2078                 spin_lock(&res->spinlock);
2079                 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2080                 spin_unlock(&res->spinlock);
2081                 wake_up(&res->wq);
2082                 dlm_lockres_put(res);
2083         }
2084         return;
2085 }
2086
2087 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2088                                struct dlm_lock_resource *res,
2089                                int ignore_higher, u8 request_from, u32 flags)
2090 {
2091         struct dlm_work_item *item;
2092         item = kzalloc(sizeof(*item), GFP_ATOMIC);
2093         if (!item)
2094                 return -ENOMEM;
2095
2096
2097         /* queue up work for dlm_assert_master_worker */
2098         dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2099         item->u.am.lockres = res; /* already have a ref */
2100         /* can optionally ignore node numbers higher than this node */
2101         item->u.am.ignore_higher = ignore_higher;
2102         item->u.am.request_from = request_from;
2103         item->u.am.flags = flags;
2104
2105         if (ignore_higher)
2106                 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2107                      res->lockname.name);
2108
2109         spin_lock(&dlm->work_lock);
2110         list_add_tail(&item->list, &dlm->work_list);
2111         spin_unlock(&dlm->work_lock);
2112
2113         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2114         return 0;
2115 }
2116
2117 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2118 {
2119         struct dlm_ctxt *dlm = data;
2120         int ret = 0;
2121         struct dlm_lock_resource *res;
2122         unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2123         int ignore_higher;
2124         int bit;
2125         u8 request_from;
2126         u32 flags;
2127
2128         dlm = item->dlm;
2129         res = item->u.am.lockres;
2130         ignore_higher = item->u.am.ignore_higher;
2131         request_from = item->u.am.request_from;
2132         flags = item->u.am.flags;
2133
2134         spin_lock(&dlm->spinlock);
2135         memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2136         spin_unlock(&dlm->spinlock);
2137
2138         clear_bit(dlm->node_num, nodemap);
2139         if (ignore_higher) {
2140                 /* if is this just to clear up mles for nodes below
2141                  * this node, do not send the message to the original
2142                  * caller or any node number higher than this */
2143                 clear_bit(request_from, nodemap);
2144                 bit = dlm->node_num;
2145                 while (1) {
2146                         bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2147                                             bit+1);
2148                         if (bit >= O2NM_MAX_NODES)
2149                                 break;
2150                         clear_bit(bit, nodemap);
2151                 }
2152         }
2153
2154         /*
2155          * If we're migrating this lock to someone else, we are no
2156          * longer allowed to assert out own mastery.  OTOH, we need to
2157          * prevent migration from starting while we're still asserting
2158          * our dominance.  The reserved ast delays migration.
2159          */
2160         spin_lock(&res->spinlock);
2161         if (res->state & DLM_LOCK_RES_MIGRATING) {
2162                 mlog(0, "Someone asked us to assert mastery, but we're "
2163                      "in the middle of migration.  Skipping assert, "
2164                      "the new master will handle that.\n");
2165                 spin_unlock(&res->spinlock);
2166                 goto put;
2167         } else
2168                 __dlm_lockres_reserve_ast(res);
2169         spin_unlock(&res->spinlock);
2170
2171         /* this call now finishes out the nodemap
2172          * even if one or more nodes die */
2173         mlog(0, "worker about to master %.*s here, this=%u\n",
2174                      res->lockname.len, res->lockname.name, dlm->node_num);
2175         ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2176         if (ret < 0) {
2177                 /* no need to restart, we are done */
2178                 if (!dlm_is_host_down(ret))
2179                         mlog_errno(ret);
2180         }
2181
2182         /* Ok, we've asserted ourselves.  Let's let migration start. */
2183         dlm_lockres_release_ast(dlm, res);
2184
2185 put:
2186         dlm_lockres_drop_inflight_worker(dlm, res);
2187
2188         dlm_lockres_put(res);
2189
2190         mlog(0, "finished with dlm_assert_master_worker\n");
2191 }
2192
2193 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2194  * We cannot wait for node recovery to complete to begin mastering this
2195  * lockres because this lockres is used to kick off recovery! ;-)
2196  * So, do a pre-check on all living nodes to see if any of those nodes
2197  * think that $RECOVERY is currently mastered by a dead node.  If so,
2198  * we wait a short time to allow that node to get notified by its own
2199  * heartbeat stack, then check again.  All $RECOVERY lock resources
2200  * mastered by dead nodes are purged when the hearbeat callback is
2201  * fired, so we can know for sure that it is safe to continue once
2202  * the node returns a live node or no node.  */
2203 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2204                                        struct dlm_lock_resource *res)
2205 {
2206         struct dlm_node_iter iter;
2207         int nodenum;
2208         int ret = 0;
2209         u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2210
2211         spin_lock(&dlm->spinlock);
2212         dlm_node_iter_init(dlm->domain_map, &iter);
2213         spin_unlock(&dlm->spinlock);
2214
2215         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2216                 /* do not send to self */
2217                 if (nodenum == dlm->node_num)
2218                         continue;
2219                 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2220                 if (ret < 0) {
2221                         mlog_errno(ret);
2222                         if (!dlm_is_host_down(ret))
2223                                 BUG();
2224                         /* host is down, so answer for that node would be
2225                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
2226                         ret = 0;
2227                 }
2228
2229                 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2230                         /* check to see if this master is in the recovery map */
2231                         spin_lock(&dlm->spinlock);
2232                         if (test_bit(master, dlm->recovery_map)) {
2233                                 mlog(ML_NOTICE, "%s: node %u has not seen "
2234                                      "node %u go down yet, and thinks the "
2235                                      "dead node is mastering the recovery "
2236                                      "lock.  must wait.\n", dlm->name,
2237                                      nodenum, master);
2238                                 ret = -EAGAIN;
2239                         }
2240                         spin_unlock(&dlm->spinlock);
2241                         mlog(0, "%s: reco lock master is %u\n", dlm->name,
2242                              master);
2243                         break;
2244                 }
2245         }
2246         return ret;
2247 }
2248
2249 /*
2250  * DLM_DEREF_LOCKRES_MSG
2251  */
2252
2253 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2254 {
2255         struct dlm_deref_lockres deref;
2256         int ret = 0, r;
2257         const char *lockname;
2258         unsigned int namelen;
2259
2260         lockname = res->lockname.name;
2261         namelen = res->lockname.len;
2262         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2263
2264         memset(&deref, 0, sizeof(deref));
2265         deref.node_idx = dlm->node_num;
2266         deref.namelen = namelen;
2267         memcpy(deref.name, lockname, namelen);
2268
2269         ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2270                                  &deref, sizeof(deref), res->owner, &r);
2271         if (ret < 0)
2272                 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n",
2273                      dlm->name, namelen, lockname, ret, res->owner);
2274         else if (r < 0) {
2275                 /* BAD.  other node says I did not have a ref. */
2276                 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2277                      dlm->name, namelen, lockname, res->owner, r);
2278                 dlm_print_one_lock_resource(res);
2279                 BUG();
2280         }
2281         return ret;
2282 }
2283
2284 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2285                               void **ret_data)
2286 {
2287         struct dlm_ctxt *dlm = data;
2288         struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2289         struct dlm_lock_resource *res = NULL;
2290         char *name;
2291         unsigned int namelen;
2292         int ret = -EINVAL;
2293         u8 node;
2294         unsigned int hash;
2295         struct dlm_work_item *item;
2296         int cleared = 0;
2297         int dispatch = 0;
2298
2299         if (!dlm_grab(dlm))
2300                 return 0;
2301
2302         name = deref->name;
2303         namelen = deref->namelen;
2304         node = deref->node_idx;
2305
2306         if (namelen > DLM_LOCKID_NAME_MAX) {
2307                 mlog(ML_ERROR, "Invalid name length!");
2308                 goto done;
2309         }
2310         if (deref->node_idx >= O2NM_MAX_NODES) {
2311                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2312                 goto done;
2313         }
2314
2315         hash = dlm_lockid_hash(name, namelen);
2316
2317         spin_lock(&dlm->spinlock);
2318         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2319         if (!res) {
2320                 spin_unlock(&dlm->spinlock);
2321                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2322                      dlm->name, namelen, name);
2323                 goto done;
2324         }
2325         spin_unlock(&dlm->spinlock);
2326
2327         spin_lock(&res->spinlock);
2328         if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2329                 dispatch = 1;
2330         else {
2331                 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2332                 if (test_bit(node, res->refmap)) {
2333                         dlm_lockres_clear_refmap_bit(dlm, res, node);
2334                         cleared = 1;
2335                 }
2336         }
2337         spin_unlock(&res->spinlock);
2338
2339         if (!dispatch) {
2340                 if (cleared)
2341                         dlm_lockres_calc_usage(dlm, res);
2342                 else {
2343                         mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2344                         "but it is already dropped!\n", dlm->name,
2345                         res->lockname.len, res->lockname.name, node);
2346                         dlm_print_one_lock_resource(res);
2347                 }
2348                 ret = 0;
2349                 goto done;
2350         }
2351
2352         item = kzalloc(sizeof(*item), GFP_NOFS);
2353         if (!item) {
2354                 ret = -ENOMEM;
2355                 mlog_errno(ret);
2356                 goto done;
2357         }
2358
2359         dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2360         item->u.dl.deref_res = res;
2361         item->u.dl.deref_node = node;
2362
2363         spin_lock(&dlm->work_lock);
2364         list_add_tail(&item->list, &dlm->work_list);
2365         spin_unlock(&dlm->work_lock);
2366
2367         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2368         return 0;
2369
2370 done:
2371         if (res)
2372                 dlm_lockres_put(res);
2373         dlm_put(dlm);
2374
2375         return ret;
2376 }
2377
2378 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2379 {
2380         struct dlm_ctxt *dlm;
2381         struct dlm_lock_resource *res;
2382         u8 node;
2383         u8 cleared = 0;
2384
2385         dlm = item->dlm;
2386         res = item->u.dl.deref_res;
2387         node = item->u.dl.deref_node;
2388
2389         spin_lock(&res->spinlock);
2390         BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2391         if (test_bit(node, res->refmap)) {
2392                 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2393                 dlm_lockres_clear_refmap_bit(dlm, res, node);
2394                 cleared = 1;
2395         }
2396         spin_unlock(&res->spinlock);
2397
2398         if (cleared) {
2399                 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2400                      dlm->name, res->lockname.len, res->lockname.name, node);
2401                 dlm_lockres_calc_usage(dlm, res);
2402         } else {
2403                 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2404                      "but it is already dropped!\n", dlm->name,
2405                      res->lockname.len, res->lockname.name, node);
2406                 dlm_print_one_lock_resource(res);
2407         }
2408
2409         dlm_lockres_put(res);
2410 }
2411
2412 /*
2413  * A migrateable resource is one that is :
2414  * 1. locally mastered, and,
2415  * 2. zero local locks, and,
2416  * 3. one or more non-local locks, or, one or more references
2417  * Returns 1 if yes, 0 if not.
2418  */
2419 static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
2420                                       struct dlm_lock_resource *res)
2421 {
2422         enum dlm_lockres_list idx;
2423         int nonlocal = 0, node_ref;
2424         struct list_head *queue;
2425         struct dlm_lock *lock;
2426         u64 cookie;
2427
2428         assert_spin_locked(&res->spinlock);
2429
2430         /* delay migration when the lockres is in MIGRATING state */
2431         if (res->state & DLM_LOCK_RES_MIGRATING)
2432                 return 0;
2433
2434         /* delay migration when the lockres is in RECOCERING state */
2435         if (res->state & DLM_LOCK_RES_RECOVERING)
2436                 return 0;
2437
2438         if (res->owner != dlm->node_num)
2439                 return 0;
2440
2441         for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2442                 queue = dlm_list_idx_to_ptr(res, idx);
2443                 list_for_each_entry(lock, queue, list) {
2444                         if (lock->ml.node != dlm->node_num) {
2445                                 nonlocal++;
2446                                 continue;
2447                         }
2448                         cookie = be64_to_cpu(lock->ml.cookie);
2449                         mlog(0, "%s: Not migrateable res %.*s, lock %u:%llu on "
2450                              "%s list\n", dlm->name, res->lockname.len,
2451                              res->lockname.name,
2452                              dlm_get_lock_cookie_node(cookie),
2453                              dlm_get_lock_cookie_seq(cookie),
2454                              dlm_list_in_text(idx));
2455                         return 0;
2456                 }
2457         }
2458
2459         if (!nonlocal) {
2460                 node_ref = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
2461                 if (node_ref >= O2NM_MAX_NODES)
2462                         return 0;
2463         }
2464
2465         mlog(0, "%s: res %.*s, Migrateable\n", dlm->name, res->lockname.len,
2466              res->lockname.name);
2467
2468         return 1;
2469 }
2470
2471 /*
2472  * DLM_MIGRATE_LOCKRES
2473  */
2474
2475
2476 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2477                                struct dlm_lock_resource *res, u8 target)
2478 {
2479         struct dlm_master_list_entry *mle = NULL;
2480         struct dlm_master_list_entry *oldmle = NULL;
2481         struct dlm_migratable_lockres *mres = NULL;
2482         int ret = 0;
2483         const char *name;
2484         unsigned int namelen;
2485         int mle_added = 0;
2486         int wake = 0;
2487
2488         if (!dlm_grab(dlm))
2489                 return -EINVAL;
2490
2491         BUG_ON(target == O2NM_MAX_NODES);
2492
2493         name = res->lockname.name;
2494         namelen = res->lockname.len;
2495
2496         mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name,
2497              target);
2498
2499         /* preallocate up front. if this fails, abort */
2500         ret = -ENOMEM;
2501         mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2502         if (!mres) {
2503                 mlog_errno(ret);
2504                 goto leave;
2505         }
2506
2507         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2508         if (!mle) {
2509                 mlog_errno(ret);
2510                 goto leave;
2511         }
2512         ret = 0;
2513
2514         /*
2515          * clear any existing master requests and
2516          * add the migration mle to the list
2517          */
2518         spin_lock(&dlm->spinlock);
2519         spin_lock(&dlm->master_lock);
2520         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2521                                     namelen, target, dlm->node_num);
2522         spin_unlock(&dlm->master_lock);
2523         spin_unlock(&dlm->spinlock);
2524
2525         if (ret == -EEXIST) {
2526                 mlog(0, "another process is already migrating it\n");
2527                 goto fail;
2528         }
2529         mle_added = 1;
2530
2531         /*
2532          * set the MIGRATING flag and flush asts
2533          * if we fail after this we need to re-dirty the lockres
2534          */
2535         if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2536                 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2537                      "the target went down.\n", res->lockname.len,
2538                      res->lockname.name, target);
2539                 spin_lock(&res->spinlock);
2540                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2541                 wake = 1;
2542                 spin_unlock(&res->spinlock);
2543                 ret = -EINVAL;
2544         }
2545
2546 fail:
2547         if (oldmle) {
2548                 /* master is known, detach if not already detached */
2549                 dlm_mle_detach_hb_events(dlm, oldmle);
2550                 dlm_put_mle(oldmle);
2551         }
2552
2553         if (ret < 0) {
2554                 if (mle_added) {
2555                         dlm_mle_detach_hb_events(dlm, mle);
2556                         dlm_put_mle(mle);
2557                 } else if (mle) {
2558                         kmem_cache_free(dlm_mle_cache, mle);
2559                         mle = NULL;
2560                 }
2561                 goto leave;
2562         }
2563
2564         /*
2565          * at this point, we have a migration target, an mle
2566          * in the master list, and the MIGRATING flag set on
2567          * the lockres
2568          */
2569
2570         /* now that remote nodes are spinning on the MIGRATING flag,
2571          * ensure that all assert_master work is flushed. */
2572         flush_workqueue(dlm->dlm_worker);
2573
2574         /* get an extra reference on the mle.
2575          * otherwise the assert_master from the new
2576          * master will destroy this.
2577          * also, make sure that all callers of dlm_get_mle
2578          * take both dlm->spinlock and dlm->master_lock */
2579         spin_lock(&dlm->spinlock);
2580         spin_lock(&dlm->master_lock);
2581         dlm_get_mle_inuse(mle);
2582         spin_unlock(&dlm->master_lock);
2583         spin_unlock(&dlm->spinlock);
2584
2585         /* notify new node and send all lock state */
2586         /* call send_one_lockres with migration flag.
2587          * this serves as notice to the target node that a
2588          * migration is starting. */
2589         ret = dlm_send_one_lockres(dlm, res, mres, target,
2590                                    DLM_MRES_MIGRATION);
2591
2592         if (ret < 0) {
2593                 mlog(0, "migration to node %u failed with %d\n",
2594                      target, ret);
2595                 /* migration failed, detach and clean up mle */
2596                 dlm_mle_detach_hb_events(dlm, mle);
2597                 dlm_put_mle(mle);
2598                 dlm_put_mle_inuse(mle);
2599                 spin_lock(&res->spinlock);
2600                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2601                 wake = 1;
2602                 spin_unlock(&res->spinlock);
2603                 if (dlm_is_host_down(ret))
2604                         dlm_wait_for_node_death(dlm, target,
2605                                                 DLM_NODE_DEATH_WAIT_MAX);
2606                 goto leave;
2607         }
2608
2609         /* at this point, the target sends a message to all nodes,
2610          * (using dlm_do_migrate_request).  this node is skipped since
2611          * we had to put an mle in the list to begin the process.  this
2612          * node now waits for target to do an assert master.  this node
2613          * will be the last one notified, ensuring that the migration
2614          * is complete everywhere.  if the target dies while this is
2615          * going on, some nodes could potentially see the target as the
2616          * master, so it is important that my recovery finds the migration
2617          * mle and sets the master to UNKNOWN. */
2618
2619
2620         /* wait for new node to assert master */
2621         while (1) {
2622                 ret = wait_event_interruptible_timeout(mle->wq,
2623                                         (atomic_read(&mle->woken) == 1),
2624                                         msecs_to_jiffies(5000));
2625
2626                 if (ret >= 0) {
2627                         if (atomic_read(&mle->woken) == 1 ||
2628                             res->owner == target)
2629                                 break;
2630
2631                         mlog(0, "%s:%.*s: timed out during migration\n",
2632                              dlm->name, res->lockname.len, res->lockname.name);
2633                         /* avoid hang during shutdown when migrating lockres
2634                          * to a node which also goes down */
2635                         if (dlm_is_node_dead(dlm, target)) {
2636                                 mlog(0, "%s:%.*s: expected migration "
2637                                      "target %u is no longer up, restarting\n",
2638                                      dlm->name, res->lockname.len,
2639                                      res->lockname.name, target);
2640                                 ret = -EINVAL;
2641                                 /* migration failed, detach and clean up mle */
2642                                 dlm_mle_detach_hb_events(dlm, mle);
2643                                 dlm_put_mle(mle);
2644                                 dlm_put_mle_inuse(mle);
2645                                 spin_lock(&res->spinlock);
2646                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2647                                 wake = 1;
2648                                 spin_unlock(&res->spinlock);
2649                                 goto leave;
2650                         }
2651                 } else
2652                         mlog(0, "%s:%.*s: caught signal during migration\n",
2653                              dlm->name, res->lockname.len, res->lockname.name);
2654         }
2655
2656         /* all done, set the owner, clear the flag */
2657         spin_lock(&res->spinlock);
2658         dlm_set_lockres_owner(dlm, res, target);
2659         res->state &= ~DLM_LOCK_RES_MIGRATING;
2660         dlm_remove_nonlocal_locks(dlm, res);
2661         spin_unlock(&res->spinlock);
2662         wake_up(&res->wq);
2663
2664         /* master is known, detach if not already detached */
2665         dlm_mle_detach_hb_events(dlm, mle);
2666         dlm_put_mle_inuse(mle);
2667         ret = 0;
2668
2669         dlm_lockres_calc_usage(dlm, res);
2670
2671 leave:
2672         /* re-dirty the lockres if we failed */
2673         if (ret < 0)
2674                 dlm_kick_thread(dlm, res);
2675
2676         /* wake up waiters if the MIGRATING flag got set
2677          * but migration failed */
2678         if (wake)
2679                 wake_up(&res->wq);
2680
2681         if (mres)
2682                 free_page((unsigned long)mres);
2683
2684         dlm_put(dlm);
2685
2686         mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen,
2687              name, target, ret);
2688         return ret;
2689 }
2690
2691 #define DLM_MIGRATION_RETRY_MS  100
2692
2693 /*
2694  * Should be called only after beginning the domain leave process.
2695  * There should not be any remaining locks on nonlocal lock resources,
2696  * and there should be no local locks left on locally mastered resources.
2697  *
2698  * Called with the dlm spinlock held, may drop it to do migration, but
2699  * will re-acquire before exit.
2700  *
2701  * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2702  */
2703 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2704 {
2705         int ret;
2706         int lock_dropped = 0;
2707         u8 target = O2NM_MAX_NODES;
2708
2709         assert_spin_locked(&dlm->spinlock);
2710
2711         spin_lock(&res->spinlock);
2712         if (dlm_is_lockres_migrateable(dlm, res))
2713                 target = dlm_pick_migration_target(dlm, res);
2714         spin_unlock(&res->spinlock);
2715
2716         if (target == O2NM_MAX_NODES)
2717                 goto leave;
2718
2719         /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2720         spin_unlock(&dlm->spinlock);
2721         lock_dropped = 1;
2722         ret = dlm_migrate_lockres(dlm, res, target);
2723         if (ret)
2724                 mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2725                      dlm->name, res->lockname.len, res->lockname.name,
2726                      target, ret);
2727         spin_lock(&dlm->spinlock);
2728 leave:
2729         return lock_dropped;
2730 }
2731
2732 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2733 {
2734         int ret;
2735         spin_lock(&dlm->ast_lock);
2736         spin_lock(&lock->spinlock);
2737         ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2738         spin_unlock(&lock->spinlock);
2739         spin_unlock(&dlm->ast_lock);
2740         return ret;
2741 }
2742
2743 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2744                                      struct dlm_lock_resource *res,
2745                                      u8 mig_target)
2746 {
2747         int can_proceed;
2748         spin_lock(&res->spinlock);
2749         can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2750         spin_unlock(&res->spinlock);
2751
2752         /* target has died, so make the caller break out of the
2753          * wait_event, but caller must recheck the domain_map */
2754         spin_lock(&dlm->spinlock);
2755         if (!test_bit(mig_target, dlm->domain_map))
2756                 can_proceed = 1;
2757         spin_unlock(&dlm->spinlock);
2758         return can_proceed;
2759 }
2760
2761 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2762                                 struct dlm_lock_resource *res)
2763 {
2764         int ret;
2765         spin_lock(&res->spinlock);
2766         ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2767         spin_unlock(&res->spinlock);
2768         return ret;
2769 }
2770
2771
2772 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2773                                        struct dlm_lock_resource *res,
2774                                        u8 target)
2775 {
2776         int ret = 0;
2777
2778         mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2779                res->lockname.len, res->lockname.name, dlm->node_num,
2780                target);
2781         /* need to set MIGRATING flag on lockres.  this is done by
2782          * ensuring that all asts have been flushed for this lockres. */
2783         spin_lock(&res->spinlock);
2784         BUG_ON(res->migration_pending);
2785         res->migration_pending = 1;
2786         /* strategy is to reserve an extra ast then release
2787          * it below, letting the release do all of the work */
2788         __dlm_lockres_reserve_ast(res);
2789         spin_unlock(&res->spinlock);
2790
2791         /* now flush all the pending asts */
2792         dlm_kick_thread(dlm, res);
2793         /* before waiting on DIRTY, block processes which may
2794          * try to dirty the lockres before MIGRATING is set */
2795         spin_lock(&res->spinlock);
2796         BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2797         res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2798         spin_unlock(&res->spinlock);
2799         /* now wait on any pending asts and the DIRTY state */
2800         wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2801         dlm_lockres_release_ast(dlm, res);
2802
2803         mlog(0, "about to wait on migration_wq, dirty=%s\n",
2804                res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2805         /* if the extra ref we just put was the final one, this
2806          * will pass thru immediately.  otherwise, we need to wait
2807          * for the last ast to finish. */
2808 again:
2809         ret = wait_event_interruptible_timeout(dlm->migration_wq,
2810                    dlm_migration_can_proceed(dlm, res, target),
2811                    msecs_to_jiffies(1000));
2812         if (ret < 0) {
2813                 mlog(0, "woken again: migrating? %s, dead? %s\n",
2814                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2815                        test_bit(target, dlm->domain_map) ? "no":"yes");
2816         } else {
2817                 mlog(0, "all is well: migrating? %s, dead? %s\n",
2818                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2819                        test_bit(target, dlm->domain_map) ? "no":"yes");
2820         }
2821         if (!dlm_migration_can_proceed(dlm, res, target)) {
2822                 mlog(0, "trying again...\n");
2823                 goto again;
2824         }
2825
2826         ret = 0;
2827         /* did the target go down or die? */
2828         spin_lock(&dlm->spinlock);
2829         if (!test_bit(target, dlm->domain_map)) {
2830                 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2831                      target);
2832                 ret = -EHOSTDOWN;
2833         }
2834         spin_unlock(&dlm->spinlock);
2835
2836         /*
2837          * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2838          * another try; otherwise, we are sure the MIGRATING state is there,
2839          * drop the unneded state which blocked threads trying to DIRTY
2840          */
2841         spin_lock(&res->spinlock);
2842         BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2843         res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2844         if (!ret)
2845                 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2846         spin_unlock(&res->spinlock);
2847
2848         /*
2849          * at this point:
2850          *
2851          *   o the DLM_LOCK_RES_MIGRATING flag is set if target not down
2852          *   o there are no pending asts on this lockres
2853          *   o all processes trying to reserve an ast on this
2854          *     lockres must wait for the MIGRATING flag to clear
2855          */
2856         return ret;
2857 }
2858
2859 /* last step in the migration process.
2860  * original master calls this to free all of the dlm_lock
2861  * structures that used to be for other nodes. */
2862 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2863                                       struct dlm_lock_resource *res)
2864 {
2865         struct list_head *queue = &res->granted;
2866         int i, bit;
2867         struct dlm_lock *lock, *next;
2868
2869         assert_spin_locked(&res->spinlock);
2870
2871         BUG_ON(res->owner == dlm->node_num);
2872
2873         for (i=0; i<3; i++) {
2874                 list_for_each_entry_safe(lock, next, queue, list) {
2875                         if (lock->ml.node != dlm->node_num) {
2876                                 mlog(0, "putting lock for node %u\n",
2877                                      lock->ml.node);
2878                                 /* be extra careful */
2879                                 BUG_ON(!list_empty(&lock->ast_list));
2880                                 BUG_ON(!list_empty(&lock->bast_list));
2881                                 BUG_ON(lock->ast_pending);
2882                                 BUG_ON(lock->bast_pending);
2883                                 dlm_lockres_clear_refmap_bit(dlm, res,
2884                                                              lock->ml.node);
2885                                 list_del_init(&lock->list);
2886                                 dlm_lock_put(lock);
2887                                 /* In a normal unlock, we would have added a
2888                                  * DLM_UNLOCK_FREE_LOCK action. Force it. */
2889                                 dlm_lock_put(lock);
2890                         }
2891                 }
2892                 queue++;
2893         }
2894         bit = 0;
2895         while (1) {
2896                 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2897                 if (bit >= O2NM_MAX_NODES)
2898                         break;
2899                 /* do not clear the local node reference, if there is a
2900                  * process holding this, let it drop the ref itself */
2901                 if (bit != dlm->node_num) {
2902                         mlog(0, "%s:%.*s: node %u had a ref to this "
2903                              "migrating lockres, clearing\n", dlm->name,
2904                              res->lockname.len, res->lockname.name, bit);
2905                         dlm_lockres_clear_refmap_bit(dlm, res, bit);
2906                 }
2907                 bit++;
2908         }
2909 }
2910
2911 /*
2912  * Pick a node to migrate the lock resource to. This function selects a
2913  * potential target based first on the locks and then on refmap. It skips
2914  * nodes that are in the process of exiting the domain.
2915  */
2916 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2917                                     struct dlm_lock_resource *res)
2918 {
2919         enum dlm_lockres_list idx;
2920         struct list_head *queue = &res->granted;
2921         struct dlm_lock *lock;
2922         int noderef;
2923         u8 nodenum = O2NM_MAX_NODES;
2924
2925         assert_spin_locked(&dlm->spinlock);
2926         assert_spin_locked(&res->spinlock);
2927
2928         /* Go through all the locks */
2929         for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2930                 queue = dlm_list_idx_to_ptr(res, idx);
2931                 list_for_each_entry(lock, queue, list) {
2932                         if (lock->ml.node == dlm->node_num)
2933                                 continue;
2934                         if (test_bit(lock->ml.node, dlm->exit_domain_map))
2935                                 continue;
2936                         nodenum = lock->ml.node;
2937                         goto bail;
2938                 }
2939         }
2940
2941         /* Go thru the refmap */
2942         noderef = -1;
2943         while (1) {
2944                 noderef = find_next_bit(res->refmap, O2NM_MAX_NODES,
2945                                         noderef + 1);
2946                 if (noderef >= O2NM_MAX_NODES)
2947                         break;
2948                 if (noderef == dlm->node_num)
2949                         continue;
2950                 if (test_bit(noderef, dlm->exit_domain_map))
2951                         continue;
2952                 nodenum = noderef;
2953                 goto bail;
2954         }
2955
2956 bail:
2957         return nodenum;
2958 }
2959
2960 /* this is called by the new master once all lockres
2961  * data has been received */
2962 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2963                                   struct dlm_lock_resource *res,
2964                                   u8 master, u8 new_master,
2965                                   struct dlm_node_iter *iter)
2966 {
2967         struct dlm_migrate_request migrate;
2968         int ret, skip, status = 0;
2969         int nodenum;
2970
2971         memset(&migrate, 0, sizeof(migrate));
2972         migrate.namelen = res->lockname.len;
2973         memcpy(migrate.name, res->lockname.name, migrate.namelen);
2974         migrate.new_master = new_master;
2975         migrate.master = master;
2976
2977         ret = 0;
2978
2979         /* send message to all nodes, except the master and myself */
2980         while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2981                 if (nodenum == master ||
2982                     nodenum == new_master)
2983                         continue;
2984
2985                 /* We could race exit domain. If exited, skip. */
2986                 spin_lock(&dlm->spinlock);
2987                 skip = (!test_bit(nodenum, dlm->domain_map));
2988                 spin_unlock(&dlm->spinlock);
2989                 if (skip) {
2990                         clear_bit(nodenum, iter->node_map);
2991                         continue;
2992                 }
2993
2994                 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2995                                          &migrate, sizeof(migrate), nodenum,
2996                                          &status);
2997                 if (ret < 0) {
2998                         mlog(ML_ERROR, "%s: res %.*s, Error %d send "
2999                              "MIGRATE_REQUEST to node %u\n", dlm->name,
3000                              migrate.namelen, migrate.name, ret, nodenum);
3001                         if (!dlm_is_host_down(ret)) {
3002                                 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
3003                                 BUG();
3004                         }
3005                         clear_bit(nodenum, iter->node_map);
3006                         ret = 0;
3007                 } else if (status < 0) {
3008                         mlog(0, "migrate request (node %u) returned %d!\n",
3009                              nodenum, status);
3010                         ret = status;
3011                 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
3012                         /* during the migration request we short-circuited
3013                          * the mastery of the lockres.  make sure we have
3014                          * a mastery ref for nodenum */
3015                         mlog(0, "%s:%.*s: need ref for node %u\n",
3016                              dlm->name, res->lockname.len, res->lockname.name,
3017                              nodenum);
3018                         spin_lock(&res->spinlock);
3019                         dlm_lockres_set_refmap_bit(dlm, res, nodenum);
3020                         spin_unlock(&res->spinlock);
3021                 }
3022         }
3023
3024         if (ret < 0)
3025                 mlog_errno(ret);
3026
3027         mlog(0, "returning ret=%d\n", ret);
3028         return ret;
3029 }
3030
3031
3032 /* if there is an existing mle for this lockres, we now know who the master is.
3033  * (the one who sent us *this* message) we can clear it up right away.
3034  * since the process that put the mle on the list still has a reference to it,
3035  * we can unhash it now, set the master and wake the process.  as a result,
3036  * we will have no mle in the list to start with.  now we can add an mle for
3037  * the migration and this should be the only one found for those scanning the
3038  * list.  */
3039 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
3040                                 void **ret_data)
3041 {
3042         struct dlm_ctxt *dlm = data;
3043         struct dlm_lock_resource *res = NULL;
3044         struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
3045         struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
3046         const char *name;
3047         unsigned int namelen, hash;
3048         int ret = 0;
3049
3050         if (!dlm_grab(dlm))
3051                 return -EINVAL;
3052
3053         name = migrate->name;
3054         namelen = migrate->namelen;
3055         hash = dlm_lockid_hash(name, namelen);
3056
3057         /* preallocate.. if this fails, abort */
3058         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
3059
3060         if (!mle) {
3061                 ret = -ENOMEM;
3062                 goto leave;
3063         }
3064
3065         /* check for pre-existing lock */
3066         spin_lock(&dlm->spinlock);
3067         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
3068         if (res) {
3069                 spin_lock(&res->spinlock);
3070                 if (res->state & DLM_LOCK_RES_RECOVERING) {
3071                         /* if all is working ok, this can only mean that we got
3072                         * a migrate request from a node that we now see as
3073                         * dead.  what can we do here?  drop it to the floor? */
3074                         spin_unlock(&res->spinlock);
3075                         mlog(ML_ERROR, "Got a migrate request, but the "
3076                              "lockres is marked as recovering!");
3077                         kmem_cache_free(dlm_mle_cache, mle);
3078                         ret = -EINVAL; /* need a better solution */
3079                         goto unlock;
3080                 }
3081                 res->state |= DLM_LOCK_RES_MIGRATING;
3082                 spin_unlock(&res->spinlock);
3083         }
3084
3085         spin_lock(&dlm->master_lock);
3086         /* ignore status.  only nonzero status would BUG. */
3087         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3088                                     name, namelen,
3089                                     migrate->new_master,
3090                                     migrate->master);
3091
3092         spin_unlock(&dlm->master_lock);
3093 unlock:
3094         spin_unlock(&dlm->spinlock);
3095
3096         if (oldmle) {
3097                 /* master is known, detach if not already detached */
3098                 dlm_mle_detach_hb_events(dlm, oldmle);
3099                 dlm_put_mle(oldmle);
3100         }
3101
3102         if (res)
3103                 dlm_lockres_put(res);
3104 leave:
3105         dlm_put(dlm);
3106         return ret;
3107 }
3108
3109 /* must be holding dlm->spinlock and dlm->master_lock
3110  * when adding a migration mle, we can clear any other mles
3111  * in the master list because we know with certainty that
3112  * the master is "master".  so we remove any old mle from
3113  * the list after setting it's master field, and then add
3114  * the new migration mle.  this way we can hold with the rule
3115  * of having only one mle for a given lock name at all times. */
3116 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3117                                  struct dlm_lock_resource *res,
3118                                  struct dlm_master_list_entry *mle,
3119                                  struct dlm_master_list_entry **oldmle,
3120                                  const char *name, unsigned int namelen,
3121                                  u8 new_master, u8 master)
3122 {
3123         int found;
3124         int ret = 0;
3125
3126         *oldmle = NULL;
3127
3128         assert_spin_locked(&dlm->spinlock);
3129         assert_spin_locked(&dlm->master_lock);
3130
3131         /* caller is responsible for any ref taken here on oldmle */
3132         found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3133         if (found) {
3134                 struct dlm_master_list_entry *tmp = *oldmle;
3135                 spin_lock(&tmp->spinlock);
3136                 if (tmp->type == DLM_MLE_MIGRATION) {
3137                         if (master == dlm->node_num) {
3138                                 /* ah another process raced me to it */
3139                                 mlog(0, "tried to migrate %.*s, but some "
3140                                      "process beat me to it\n",
3141                                      namelen, name);
3142                                 ret = -EEXIST;
3143                         } else {
3144                                 /* bad.  2 NODES are trying to migrate! */
3145                                 mlog(ML_ERROR, "migration error  mle: "
3146                                      "master=%u new_master=%u // request: "
3147                                      "master=%u new_master=%u // "
3148                                      "lockres=%.*s\n",
3149                                      tmp->master, tmp->new_master,
3150                                      master, new_master,
3151                                      namelen, name);
3152                                 BUG();
3153                         }
3154                 } else {
3155                         /* this is essentially what assert_master does */
3156                         tmp->master = master;
3157                         atomic_set(&tmp->woken, 1);
3158                         wake_up(&tmp->wq);
3159                         /* remove it so that only one mle will be found */
3160                         __dlm_unlink_mle(dlm, tmp);
3161                         __dlm_mle_detach_hb_events(dlm, tmp);
3162                         if (tmp->type == DLM_MLE_MASTER) {
3163                                 ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3164                                 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3165                                                 "telling master to get ref "
3166                                                 "for cleared out mle during "
3167                                                 "migration\n", dlm->name,
3168                                                 namelen, name, master,
3169                                                 new_master);
3170                         }
3171                 }
3172                 spin_unlock(&tmp->spinlock);
3173         }
3174
3175         /* now add a migration mle to the tail of the list */
3176         dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3177         mle->new_master = new_master;
3178         /* the new master will be sending an assert master for this.
3179          * at that point we will get the refmap reference */
3180         mle->master = master;
3181         /* do this for consistency with other mle types */
3182         set_bit(new_master, mle->maybe_map);
3183         __dlm_insert_mle(dlm, mle);
3184
3185         return ret;
3186 }
3187
3188 /*
3189  * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3190  */
3191 static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3192                                         struct dlm_master_list_entry *mle)
3193 {
3194         struct dlm_lock_resource *res;
3195
3196         /* Find the lockres associated to the mle and set its owner to UNK */
3197         res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3198                                    mle->mnamehash);
3199         if (res) {
3200                 spin_unlock(&dlm->master_lock);
3201
3202                 /* move lockres onto recovery list */
3203                 spin_lock(&res->spinlock);
3204                 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3205                 dlm_move_lockres_to_recovery_list(dlm, res);
3206                 spin_unlock(&res->spinlock);
3207                 dlm_lockres_put(res);
3208
3209                 /* about to get rid of mle, detach from heartbeat */
3210                 __dlm_mle_detach_hb_events(dlm, mle);
3211
3212                 /* dump the mle */
3213                 spin_lock(&dlm->master_lock);
3214                 __dlm_put_mle(mle);
3215                 spin_unlock(&dlm->master_lock);
3216         }
3217
3218         return res;
3219 }
3220
3221 static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3222                                     struct dlm_master_list_entry *mle)
3223 {
3224         __dlm_mle_detach_hb_events(dlm, mle);
3225
3226         spin_lock(&mle->spinlock);
3227         __dlm_unlink_mle(dlm, mle);
3228         atomic_set(&mle->woken, 1);
3229         spin_unlock(&mle->spinlock);
3230
3231         wake_up(&mle->wq);
3232 }
3233
3234 static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3235                                 struct dlm_master_list_entry *mle, u8 dead_node)
3236 {
3237         int bit;
3238
3239         BUG_ON(mle->type != DLM_MLE_BLOCK);
3240
3241         spin_lock(&mle->spinlock);
3242         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3243         if (bit != dead_node) {
3244                 mlog(0, "mle found, but dead node %u would not have been "
3245                      "master\n", dead_node);
3246                 spin_unlock(&mle->spinlock);
3247         } else {
3248                 /* Must drop the refcount by one since the assert_master will
3249                  * never arrive. This may result in the mle being unlinked and
3250                  * freed, but there may still be a process waiting in the
3251                  * dlmlock path which is fine. */
3252                 mlog(0, "node %u was expected master\n", dead_node);
3253                 atomic_set(&mle->woken, 1);
3254                 spin_unlock(&mle->spinlock);
3255                 wake_up(&mle->wq);
3256
3257                 /* Do not need events any longer, so detach from heartbeat */
3258                 __dlm_mle_detach_hb_events(dlm, mle);
3259                 __dlm_put_mle(mle);
3260         }
3261 }
3262
3263 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3264 {
3265         struct dlm_master_list_entry *mle;
3266         struct dlm_lock_resource *res;
3267         struct hlist_head *bucket;
3268         struct hlist_node *tmp;
3269         unsigned int i;
3270
3271         mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
3272 top:
3273         assert_spin_locked(&dlm->spinlock);
3274
3275         /* clean the master list */
3276         spin_lock(&dlm->master_lock);
3277         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3278                 bucket = dlm_master_hash(dlm, i);
3279                 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3280                         BUG_ON(mle->type != DLM_MLE_BLOCK &&
3281                                mle->type != DLM_MLE_MASTER &&
3282                                mle->type != DLM_MLE_MIGRATION);
3283
3284                         /* MASTER mles are initiated locally. The waiting
3285                          * process will notice the node map change shortly.
3286                          * Let that happen as normal. */
3287                         if (mle->type == DLM_MLE_MASTER)
3288                                 continue;
3289
3290                         /* BLOCK mles are initiated by other nodes. Need to
3291                          * clean up if the dead node would have been the
3292                          * master. */
3293                         if (mle->type == DLM_MLE_BLOCK) {
3294                                 dlm_clean_block_mle(dlm, mle, dead_node);
3295                                 continue;
3296                         }
3297
3298                         /* Everything else is a MIGRATION mle */
3299
3300                         /* The rule for MIGRATION mles is that the master
3301                          * becomes UNKNOWN if *either* the original or the new
3302                          * master dies. All UNKNOWN lockres' are sent to
3303                          * whichever node becomes the recovery master. The new
3304                          * master is responsible for determining if there is
3305                          * still a master for this lockres, or if he needs to
3306                          * take over mastery. Either way, this node should
3307                          * expect another message to resolve this. */
3308
3309                         if (mle->master != dead_node &&
3310                             mle->new_master != dead_node)
3311                                 continue;
3312
3313                         /* If we have reached this point, this mle needs to be
3314                          * removed from the list and freed. */
3315                         dlm_clean_migration_mle(dlm, mle);
3316
3317                         mlog(0, "%s: node %u died during migration from "
3318                              "%u to %u!\n", dlm->name, dead_node, mle->master,
3319                              mle->new_master);
3320
3321                         /* If we find a lockres associated with the mle, we've
3322                          * hit this rare case that messes up our lock ordering.
3323                          * If so, we need to drop the master lock so that we can
3324                          * take the lockres lock, meaning that we will have to
3325                          * restart from the head of list. */
3326                         res = dlm_reset_mleres_owner(dlm, mle);
3327                         if (res)
3328                                 /* restart */
3329                                 goto top;
3330
3331                         /* This may be the last reference */
3332                         __dlm_put_mle(mle);
3333                 }
3334         }
3335         spin_unlock(&dlm->master_lock);
3336 }
3337
3338 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3339                          u8 old_master)
3340 {
3341         struct dlm_node_iter iter;
3342         int ret = 0;
3343
3344         spin_lock(&dlm->spinlock);
3345         dlm_node_iter_init(dlm->domain_map, &iter);
3346         clear_bit(old_master, iter.node_map);
3347         clear_bit(dlm->node_num, iter.node_map);
3348         spin_unlock(&dlm->spinlock);
3349
3350         /* ownership of the lockres is changing.  account for the
3351          * mastery reference here since old_master will briefly have
3352          * a reference after the migration completes */
3353         spin_lock(&res->spinlock);
3354         dlm_lockres_set_refmap_bit(dlm, res, old_master);
3355         spin_unlock(&res->spinlock);
3356
3357         mlog(0, "now time to do a migrate request to other nodes\n");
3358         ret = dlm_do_migrate_request(dlm, res, old_master,
3359                                      dlm->node_num, &iter);
3360         if (ret < 0) {
3361                 mlog_errno(ret);
3362                 goto leave;
3363         }
3364
3365         mlog(0, "doing assert master of %.*s to all except the original node\n",
3366              res->lockname.len, res->lockname.name);
3367         /* this call now finishes out the nodemap
3368          * even if one or more nodes die */
3369         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3370                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3371         if (ret < 0) {
3372                 /* no longer need to retry.  all living nodes contacted. */
3373                 mlog_errno(ret);
3374                 ret = 0;
3375         }
3376
3377         memset(iter.node_map, 0, sizeof(iter.node_map));
3378         set_bit(old_master, iter.node_map);
3379         mlog(0, "doing assert master of %.*s back to %u\n",
3380              res->lockname.len, res->lockname.name, old_master);
3381         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3382                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3383         if (ret < 0) {
3384                 mlog(0, "assert master to original master failed "
3385                      "with %d.\n", ret);
3386                 /* the only nonzero status here would be because of
3387                  * a dead original node.  we're done. */
3388                 ret = 0;
3389         }
3390
3391         /* all done, set the owner, clear the flag */
3392         spin_lock(&res->spinlock);
3393         dlm_set_lockres_owner(dlm, res, dlm->node_num);
3394         res->state &= ~DLM_LOCK_RES_MIGRATING;
3395         spin_unlock(&res->spinlock);
3396         /* re-dirty it on the new master */
3397         dlm_kick_thread(dlm, res);
3398         wake_up(&res->wq);
3399 leave:
3400         return ret;
3401 }
3402
3403 /*
3404  * LOCKRES AST REFCOUNT
3405  * this is integral to migration
3406  */
3407
3408 /* for future intent to call an ast, reserve one ahead of time.
3409  * this should be called only after waiting on the lockres
3410  * with dlm_wait_on_lockres, and while still holding the
3411  * spinlock after the call. */
3412 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3413 {
3414         assert_spin_locked(&res->spinlock);
3415         if (res->state & DLM_LOCK_RES_MIGRATING) {
3416                 __dlm_print_one_lock_resource(res);
3417         }
3418         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3419
3420         atomic_inc(&res->asts_reserved);
3421 }
3422
3423 /*
3424  * used to drop the reserved ast, either because it went unused,
3425  * or because the ast/bast was actually called.
3426  *
3427  * also, if there is a pending migration on this lockres,
3428  * and this was the last pending ast on the lockres,
3429  * atomically set the MIGRATING flag before we drop the lock.
3430  * this is how we ensure that migration can proceed with no
3431  * asts in progress.  note that it is ok if the state of the
3432  * queues is such that a lock should be granted in the future
3433  * or that a bast should be fired, because the new master will
3434  * shuffle the lists on this lockres as soon as it is migrated.
3435  */
3436 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3437                              struct dlm_lock_resource *res)
3438 {
3439         if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3440                 return;
3441
3442         if (!res->migration_pending) {
3443                 spin_unlock(&res->spinlock);
3444                 return;
3445         }
3446
3447         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3448         res->migration_pending = 0;
3449         res->state |= DLM_LOCK_RES_MIGRATING;
3450         spin_unlock(&res->spinlock);
3451         wake_up(&res->wq);
3452         wake_up(&dlm->migration_wq);
3453 }
3454
3455 void dlm_force_free_mles(struct dlm_ctxt *dlm)
3456 {
3457         int i;
3458         struct hlist_head *bucket;
3459         struct dlm_master_list_entry *mle;
3460         struct hlist_node *tmp;
3461
3462         /*
3463          * We notified all other nodes that we are exiting the domain and
3464          * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3465          * around we force free them and wake any processes that are waiting
3466          * on the mles
3467          */
3468         spin_lock(&dlm->spinlock);
3469         spin_lock(&dlm->master_lock);
3470
3471         BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3472         BUG_ON((find_next_bit(dlm->domain_map, O2NM_MAX_NODES, 0) < O2NM_MAX_NODES));
3473
3474         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3475                 bucket = dlm_master_hash(dlm, i);
3476                 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3477                         if (mle->type != DLM_MLE_BLOCK) {
3478                                 mlog(ML_ERROR, "bad mle: %p\n", mle);
3479                                 dlm_print_one_mle(mle);
3480                         }
3481                         atomic_set(&mle->woken, 1);
3482                         wake_up(&mle->wq);
3483
3484                         __dlm_unlink_mle(dlm, mle);
3485                         __dlm_mle_detach_hb_events(dlm, mle);
3486                         __dlm_put_mle(mle);
3487                 }
3488         }
3489         spin_unlock(&dlm->master_lock);
3490         spin_unlock(&dlm->spinlock);
3491 }