fs/ocfs2/dlm/dlmmaster.c

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