fs/ocfs2/dlm/dlmmaster.c

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