fs/ocfs2/dlm/dlmmaster.c

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