fs/ocfs2/dlm/dlmmaster.c

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