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
 733                 /*
 734                  * Right after dlm spinlock was released, dlm_thread could have
 735                  * purged the lockres. Check if lockres got unhashed. If so
 736                  * start over.
 737                  */
 738                 if (hlist_unhashed(&tmpres->hash_node)) {
 739                         spin_unlock(&tmpres->spinlock);
 740                         dlm_lockres_put(tmpres);
 741                         tmpres = NULL;
 742                         goto lookup;
 743                 }
 744
 745                 /* Wait on the thread that is mastering the resource */
 746                 if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
 747                         __dlm_wait_on_lockres(tmpres);
 748                         BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
 749                         spin_unlock(&tmpres->spinlock);
 750                         dlm_lockres_put(tmpres);
 751                         tmpres = NULL;
 752                         goto lookup;
 753                 }
 754
 755                 /* Wait on the resource purge to complete before continuing */
 756                 if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) {
 757                         BUG_ON(tmpres->owner == dlm->node_num);
 758                         __dlm_wait_on_lockres_flags(tmpres,
 759                                                     DLM_LOCK_RES_DROPPING_REF);
 760                         spin_unlock(&tmpres->spinlock);
 761                         dlm_lockres_put(tmpres);
 762                         tmpres = NULL;
 763                         goto lookup;
 764                 }
 765
 766                 /* Grab inflight ref to pin the resource */
 767                 dlm_lockres_grab_inflight_ref(dlm, tmpres);
 768
 769                 spin_unlock(&tmpres->spinlock);
 770                 if (res)
 771                         dlm_lockres_put(res);
 772                 res = tmpres;
 773                 goto leave;
 774         }
 775
 776         if (!res) {
 777                 spin_unlock(&dlm->spinlock);
 778                 mlog(0, "allocating a new resource\n");
 779                 /* nothing found and we need to allocate one. */
 780                 alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
 781                 if (!alloc_mle)
 782                         goto leave;
 783                 res = dlm_new_lockres(dlm, lockid, namelen);
 784                 if (!res)
 785                         goto leave;
 786                 goto lookup;
 787         }
 788
 789         mlog(0, "no lockres found, allocated our own: %p\n", res);
 790
 791         if (flags & LKM_LOCAL) {
 792                 /* caller knows it's safe to assume it's not mastered elsewhere
 793                  * DONE!  return right away */
 794                 spin_lock(&res->spinlock);
 795                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
 796                 __dlm_insert_lockres(dlm, res);
 797                 dlm_lockres_grab_inflight_ref(dlm, res);
 798                 spin_unlock(&res->spinlock);
 799                 spin_unlock(&dlm->spinlock);
 800                 /* lockres still marked IN_PROGRESS */
 801                 goto wake_waiters;
 802         }
 803
 804         /* check master list to see if another node has started mastering it */
 805         spin_lock(&dlm->master_lock);
 806
 807         /* if we found a block, wait for lock to be mastered by another node */
 808         blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
 809         if (blocked) {
 810                 int mig;
 811                 if (mle->type == DLM_MLE_MASTER) {
 812                         mlog(ML_ERROR, "master entry for nonexistent lock!\n");
 813                         BUG();
 814                 }
 815                 mig = (mle->type == DLM_MLE_MIGRATION);
 816                 /* if there is a migration in progress, let the migration
 817                  * finish before continuing.  we can wait for the absence
 818                  * of the MIGRATION mle: either the migrate finished or
 819                  * one of the nodes died and the mle was cleaned up.
 820                  * if there is a BLOCK here, but it already has a master
 821                  * set, we are too late.  the master does not have a ref
 822                  * for us in the refmap.  detach the mle and drop it.
 823                  * either way, go back to the top and start over. */
 824                 if (mig || mle->master != O2NM_MAX_NODES) {
 825                         BUG_ON(mig && mle->master == dlm->node_num);
 826                         /* we arrived too late.  the master does not
 827                          * have a ref for us. retry. */
 828                         mlog(0, "%s:%.*s: late on %s\n",
 829                              dlm->name, namelen, lockid,
 830                              mig ?  "MIGRATION" : "BLOCK");
 831                         spin_unlock(&dlm->master_lock);
 832                         spin_unlock(&dlm->spinlock);
 833
 834                         /* master is known, detach */
 835                         if (!mig)
 836                                 dlm_mle_detach_hb_events(dlm, mle);
 837                         dlm_put_mle(mle);
 838                         mle = NULL;
 839                         /* this is lame, but we can't wait on either
 840                          * the mle or lockres waitqueue here */
 841                         if (mig)
 842                                 msleep(100);
 843                         goto lookup;
 844                 }
 845         } else {
 846                 /* go ahead and try to master lock on this node */
 847                 mle = alloc_mle;
 848                 /* make sure this does not get freed below */
 849                 alloc_mle = NULL;
 850                 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
 851                 set_bit(dlm->node_num, mle->maybe_map);
 852                 __dlm_insert_mle(dlm, mle);
 853
 854                 /* still holding the dlm spinlock, check the recovery map
 855                  * to see if there are any nodes that still need to be
 856                  * considered.  these will not appear in the mle nodemap
 857                  * but they might own this lockres.  wait on them. */
 858                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 859                 if (bit < O2NM_MAX_NODES) {
 860                         mlog(0, "%s: res %.*s, At least one node (%d) "
 861                              "to recover before lock mastery can begin\n",
 862                              dlm->name, namelen, (char *)lockid, bit);
 863                         wait_on_recovery = 1;
 864                 }
 865         }
 866
 867         /* at this point there is either a DLM_MLE_BLOCK or a
 868          * DLM_MLE_MASTER on the master list, so it's safe to add the
 869          * lockres to the hashtable.  anyone who finds the lock will
 870          * still have to wait on the IN_PROGRESS. */
 871
 872         /* finally add the lockres to its hash bucket */
 873         __dlm_insert_lockres(dlm, res);
 874
 875         /* since this lockres is new it doesn't not require the spinlock */
 876         __dlm_lockres_grab_inflight_ref(dlm, res);
 877
 878         /* get an extra ref on the mle in case this is a BLOCK
 879          * if so, the creator of the BLOCK may try to put the last
 880          * ref at this time in the assert master handler, so we
 881          * need an extra one to keep from a bad ptr deref. */
 882         dlm_get_mle_inuse(mle);
 883         spin_unlock(&dlm->master_lock);
 884         spin_unlock(&dlm->spinlock);
 885
 886 redo_request:
 887         while (wait_on_recovery) {
 888                 /* any cluster changes that occurred after dropping the
 889                  * dlm spinlock would be detectable be a change on the mle,
 890                  * so we only need to clear out the recovery map once. */
 891                 if (dlm_is_recovery_lock(lockid, namelen)) {
 892                         mlog(0, "%s: Recovery map is not empty, but must "
 893                              "master $RECOVERY lock now\n", dlm->name);
 894                         if (!dlm_pre_master_reco_lockres(dlm, res))
 895                                 wait_on_recovery = 0;
 896                         else {
 897                                 mlog(0, "%s: waiting 500ms for heartbeat state "
 898                                     "change\n", dlm->name);
 899                                 msleep(500);
 900                         }
 901                         continue;
 902                 }
 903
 904                 dlm_kick_recovery_thread(dlm);
 905                 msleep(1000);
 906                 dlm_wait_for_recovery(dlm);
 907
 908                 spin_lock(&dlm->spinlock);
 909                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 910                 if (bit < O2NM_MAX_NODES) {
 911                         mlog(0, "%s: res %.*s, At least one node (%d) "
 912                              "to recover before lock mastery can begin\n",
 913                              dlm->name, namelen, (char *)lockid, bit);
 914                         wait_on_recovery = 1;
 915                 } else
 916                         wait_on_recovery = 0;
 917                 spin_unlock(&dlm->spinlock);
 918
 919                 if (wait_on_recovery)
 920                         dlm_wait_for_node_recovery(dlm, bit, 10000);
 921         }
 922
 923         /* must wait for lock to be mastered elsewhere */
 924         if (blocked)
 925                 goto wait;
 926
 927         ret = -EINVAL;
 928         dlm_node_iter_init(mle->vote_map, &iter);
 929         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
 930                 ret = dlm_do_master_request(res, mle, nodenum);
 931                 if (ret < 0)
 932                         mlog_errno(ret);
 933                 if (mle->master != O2NM_MAX_NODES) {
 934                         /* found a master ! */
 935                         if (mle->master <= nodenum)
 936                                 break;
 937                         /* if our master request has not reached the master
 938                          * yet, keep going until it does.  this is how the
 939                          * master will know that asserts are needed back to
 940                          * the lower nodes. */
 941                         mlog(0, "%s: res %.*s, Requests only up to %u but "
 942                              "master is %u, keep going\n", dlm->name, namelen,
 943                              lockid, nodenum, mle->master);
 944                 }
 945         }
 946
 947 wait:
 948         /* keep going until the response map includes all nodes */
 949         ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
 950         if (ret < 0) {
 951                 wait_on_recovery = 1;
 952                 mlog(0, "%s: res %.*s, Node map changed, redo the master "
 953                      "request now, blocked=%d\n", dlm->name, res->lockname.len,
 954                      res->lockname.name, blocked);
 955                 if (++tries > 20) {
 956                         mlog(ML_ERROR, "%s: res %.*s, Spinning on "
 957                              "dlm_wait_for_lock_mastery, blocked = %d\n",
 958                              dlm->name, res->lockname.len,
 959                              res->lockname.name, blocked);
 960                         dlm_print_one_lock_resource(res);
 961                         dlm_print_one_mle(mle);
 962                         tries = 0;
 963                 }
 964                 goto redo_request;
 965         }
 966
 967         mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len,
 968              res->lockname.name, res->owner);
 969         /* make sure we never continue without this */
 970         BUG_ON(res->owner == O2NM_MAX_NODES);
 971
 972         /* master is known, detach if not already detached */
 973         dlm_mle_detach_hb_events(dlm, mle);
 974         dlm_put_mle(mle);
 975         /* put the extra ref */
 976         dlm_put_mle_inuse(mle);
 977
 978 wake_waiters:
 979         spin_lock(&res->spinlock);
 980         res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
 981         spin_unlock(&res->spinlock);
 982         wake_up(&res->wq);
 983
 984 leave:
 985         /* need to free the unused mle */
 986         if (alloc_mle)
 987                 kmem_cache_free(dlm_mle_cache, alloc_mle);
 988
 989         return res;
 990 }
 991
 992
 993 #define DLM_MASTERY_TIMEOUT_MS   5000
 994
 995 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
 996                                      struct dlm_lock_resource *res,
 997                                      struct dlm_master_list_entry *mle,
 998                                      int *blocked)
 999 {
1000         u8 m;
1001         int ret, bit;
1002         int map_changed, voting_done;
1003         int assert, sleep;
1004
1005 recheck:
1006         ret = 0;
1007         assert = 0;
1008
1009         /* check if another node has already become the owner */
1010         spin_lock(&res->spinlock);
1011         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1012                 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1013                      res->lockname.len, res->lockname.name, res->owner);
1014                 spin_unlock(&res->spinlock);
1015                 /* this will cause the master to re-assert across
1016                  * the whole cluster, freeing up mles */
1017                 if (res->owner != dlm->node_num) {
1018                         ret = dlm_do_master_request(res, mle, res->owner);
1019                         if (ret < 0) {
1020                                 /* give recovery a chance to run */
1021                                 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1022                                 msleep(500);
1023                                 goto recheck;
1024                         }
1025                 }
1026                 ret = 0;
1027                 goto leave;
1028         }
1029         spin_unlock(&res->spinlock);
1030
1031         spin_lock(&mle->spinlock);
1032         m = mle->master;
1033         map_changed = (memcmp(mle->vote_map, mle->node_map,
1034                               sizeof(mle->vote_map)) != 0);
1035         voting_done = (memcmp(mle->vote_map, mle->response_map,
1036                              sizeof(mle->vote_map)) == 0);
1037
1038         /* restart if we hit any errors */
1039         if (map_changed) {
1040                 int b;
1041                 mlog(0, "%s: %.*s: node map changed, restarting\n",
1042                      dlm->name, res->lockname.len, res->lockname.name);
1043                 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1044                 b = (mle->type == DLM_MLE_BLOCK);
1045                 if ((*blocked && !b) || (!*blocked && b)) {
1046                         mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1047                              dlm->name, res->lockname.len, res->lockname.name,
1048                              *blocked, b);
1049                         *blocked = b;
1050                 }
1051                 spin_unlock(&mle->spinlock);
1052                 if (ret < 0) {
1053                         mlog_errno(ret);
1054                         goto leave;
1055                 }
1056                 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1057                      "rechecking now\n", dlm->name, res->lockname.len,
1058                      res->lockname.name);
1059                 goto recheck;
1060         } else {
1061                 if (!voting_done) {
1062                         mlog(0, "map not changed and voting not done "
1063                              "for %s:%.*s\n", dlm->name, res->lockname.len,
1064                              res->lockname.name);
1065                 }
1066         }
1067
1068         if (m != O2NM_MAX_NODES) {
1069                 /* another node has done an assert!
1070                  * all done! */
1071                 sleep = 0;
1072         } else {
1073                 sleep = 1;
1074                 /* have all nodes responded? */
1075                 if (voting_done && !*blocked) {
1076                         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1077                         if (dlm->node_num <= bit) {
1078                                 /* my node number is lowest.
1079                                  * now tell other nodes that I am
1080                                  * mastering this. */
1081                                 mle->master = dlm->node_num;
1082                                 /* ref was grabbed in get_lock_resource
1083                                  * will be dropped in dlmlock_master */
1084                                 assert = 1;
1085                                 sleep = 0;
1086                         }
1087                         /* if voting is done, but we have not received
1088                          * an assert master yet, we must sleep */
1089                 }
1090         }
1091
1092         spin_unlock(&mle->spinlock);
1093
1094         /* sleep if we haven't finished voting yet */
1095         if (sleep) {
1096                 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1097
1098                 /*
1099                 if (atomic_read(&mle->mle_refs.refcount) < 2)
1100                         mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1101                         atomic_read(&mle->mle_refs.refcount),
1102                         res->lockname.len, res->lockname.name);
1103                 */
1104                 atomic_set(&mle->woken, 0);
1105                 (void)wait_event_timeout(mle->wq,
1106                                          (atomic_read(&mle->woken) == 1),
1107                                          timeo);
1108                 if (res->owner == O2NM_MAX_NODES) {
1109                         mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1110                              res->lockname.len, res->lockname.name);
1111                         goto recheck;
1112                 }
1113                 mlog(0, "done waiting, master is %u\n", res->owner);
1114                 ret = 0;
1115                 goto leave;
1116         }
1117
1118         ret = 0;   /* done */
1119         if (assert) {
1120                 m = dlm->node_num;
1121                 mlog(0, "about to master %.*s here, this=%u\n",
1122                      res->lockname.len, res->lockname.name, m);
1123                 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1124                 if (ret) {
1125                         /* This is a failure in the network path,
1126                          * not in the response to the assert_master
1127                          * (any nonzero response is a BUG on this node).
1128                          * Most likely a socket just got disconnected
1129                          * due to node death. */
1130                         mlog_errno(ret);
1131                 }
1132                 /* no longer need to restart lock mastery.
1133                  * all living nodes have been contacted. */
1134                 ret = 0;
1135         }
1136
1137         /* set the lockres owner */
1138         spin_lock(&res->spinlock);
1139         /* mastery reference obtained either during
1140          * assert_master_handler or in get_lock_resource */
1141         dlm_change_lockres_owner(dlm, res, m);
1142         spin_unlock(&res->spinlock);
1143
1144 leave:
1145         return ret;
1146 }
1147
1148 struct dlm_bitmap_diff_iter
1149 {
1150         int curnode;
1151         unsigned long *orig_bm;
1152         unsigned long *cur_bm;
1153         unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1154 };
1155
1156 enum dlm_node_state_change
1157 {
1158         NODE_DOWN = -1,
1159         NODE_NO_CHANGE = 0,
1160         NODE_UP
1161 };
1162
1163 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1164                                       unsigned long *orig_bm,
1165                                       unsigned long *cur_bm)
1166 {
1167         unsigned long p1, p2;
1168         int i;
1169
1170         iter->curnode = -1;
1171         iter->orig_bm = orig_bm;
1172         iter->cur_bm = cur_bm;
1173
1174         for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1175                 p1 = *(iter->orig_bm + i);
1176                 p2 = *(iter->cur_bm + i);
1177                 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1178         }
1179 }
1180
1181 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1182                                      enum dlm_node_state_change *state)
1183 {
1184         int bit;
1185
1186         if (iter->curnode >= O2NM_MAX_NODES)
1187                 return -ENOENT;
1188
1189         bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1190                             iter->curnode+1);
1191         if (bit >= O2NM_MAX_NODES) {
1192                 iter->curnode = O2NM_MAX_NODES;
1193                 return -ENOENT;
1194         }
1195
1196         /* if it was there in the original then this node died */
1197         if (test_bit(bit, iter->orig_bm))
1198                 *state = NODE_DOWN;
1199         else
1200                 *state = NODE_UP;
1201
1202         iter->curnode = bit;
1203         return bit;
1204 }
1205
1206
1207 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1208                                     struct dlm_lock_resource *res,
1209                                     struct dlm_master_list_entry *mle,
1210                                     int blocked)
1211 {
1212         struct dlm_bitmap_diff_iter bdi;
1213         enum dlm_node_state_change sc;
1214         int node;
1215         int ret = 0;
1216
1217         mlog(0, "something happened such that the "
1218              "master process may need to be restarted!\n");
1219
1220         assert_spin_locked(&mle->spinlock);
1221
1222         dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1223         node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1224         while (node >= 0) {
1225                 if (sc == NODE_UP) {
1226                         /* a node came up.  clear any old vote from
1227                          * the response map and set it in the vote map
1228                          * then restart the mastery. */
1229                         mlog(ML_NOTICE, "node %d up while restarting\n", node);
1230
1231                         /* redo the master request, but only for the new node */
1232                         mlog(0, "sending request to new node\n");
1233                         clear_bit(node, mle->response_map);
1234                         set_bit(node, mle->vote_map);
1235                 } else {
1236                         mlog(ML_ERROR, "node down! %d\n", node);
1237                         if (blocked) {
1238                                 int lowest = find_next_bit(mle->maybe_map,
1239                                                        O2NM_MAX_NODES, 0);
1240
1241                                 /* act like it was never there */
1242                                 clear_bit(node, mle->maybe_map);
1243
1244                                 if (node == lowest) {
1245                                         mlog(0, "expected master %u died"
1246                                             " while this node was blocked "
1247                                             "waiting on it!\n", node);
1248                                         lowest = find_next_bit(mle->maybe_map,
1249                                                         O2NM_MAX_NODES,
1250                                                         lowest+1);
1251                                         if (lowest < O2NM_MAX_NODES) {
1252                                                 mlog(0, "%s:%.*s:still "
1253                                                      "blocked. waiting on %u "
1254                                                      "now\n", dlm->name,
1255                                                      res->lockname.len,
1256                                                      res->lockname.name,
1257                                                      lowest);
1258                                         } else {
1259                                                 /* mle is an MLE_BLOCK, but
1260                                                  * there is now nothing left to
1261                                                  * block on.  we need to return
1262                                                  * all the way back out and try
1263                                                  * again with an MLE_MASTER.
1264                                                  * dlm_do_local_recovery_cleanup
1265                                                  * has already run, so the mle
1266                                                  * refcount is ok */
1267                                                 mlog(0, "%s:%.*s: no "
1268                                                      "longer blocking. try to "
1269                                                      "master this here\n",
1270                                                      dlm->name,
1271                                                      res->lockname.len,
1272                                                      res->lockname.name);
1273                                                 mle->type = DLM_MLE_MASTER;
1274                                                 mle->mleres = res;
1275                                         }
1276                                 }
1277                         }
1278
1279                         /* now blank out everything, as if we had never
1280                          * contacted anyone */
1281                         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1282                         memset(mle->response_map, 0, sizeof(mle->response_map));
1283                         /* reset the vote_map to the current node_map */
1284                         memcpy(mle->vote_map, mle->node_map,
1285                                sizeof(mle->node_map));
1286                         /* put myself into the maybe map */
1287                         if (mle->type != DLM_MLE_BLOCK)
1288                                 set_bit(dlm->node_num, mle->maybe_map);
1289                 }
1290                 ret = -EAGAIN;
1291                 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1292         }
1293         return ret;
1294 }
1295
1296
1297 /*
1298  * DLM_MASTER_REQUEST_MSG
1299  *
1300  * returns: 0 on success,
1301  *          -errno on a network error
1302  *
1303  * on error, the caller should assume the target node is "dead"
1304  *
1305  */
1306
1307 static int dlm_do_master_request(struct dlm_lock_resource *res,
1308                                  struct dlm_master_list_entry *mle, int to)
1309 {
1310         struct dlm_ctxt *dlm = mle->dlm;
1311         struct dlm_master_request request;
1312         int ret, response=0, resend;
1313
1314         memset(&request, 0, sizeof(request));
1315         request.node_idx = dlm->node_num;
1316
1317         BUG_ON(mle->type == DLM_MLE_MIGRATION);
1318
1319         request.namelen = (u8)mle->mnamelen;
1320         memcpy(request.name, mle->mname, request.namelen);
1321
1322 again:
1323         ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1324                                  sizeof(request), to, &response);
1325         if (ret < 0)  {
1326                 if (ret == -ESRCH) {
1327                         /* should never happen */
1328                         mlog(ML_ERROR, "TCP stack not ready!\n");
1329                         BUG();
1330                 } else if (ret == -EINVAL) {
1331                         mlog(ML_ERROR, "bad args passed to o2net!\n");
1332                         BUG();
1333                 } else if (ret == -ENOMEM) {
1334                         mlog(ML_ERROR, "out of memory while trying to send "
1335                              "network message!  retrying\n");
1336                         /* this is totally crude */
1337                         msleep(50);
1338                         goto again;
1339                 } else if (!dlm_is_host_down(ret)) {
1340                         /* not a network error. bad. */
1341                         mlog_errno(ret);
1342                         mlog(ML_ERROR, "unhandled error!");
1343                         BUG();
1344                 }
1345                 /* all other errors should be network errors,
1346                  * and likely indicate node death */
1347                 mlog(ML_ERROR, "link to %d went down!\n", to);
1348                 goto out;
1349         }
1350
1351         ret = 0;
1352         resend = 0;
1353         spin_lock(&mle->spinlock);
1354         switch (response) {
1355                 case DLM_MASTER_RESP_YES:
1356                         set_bit(to, mle->response_map);
1357                         mlog(0, "node %u is the master, response=YES\n", to);
1358                         mlog(0, "%s:%.*s: master node %u now knows I have a "
1359                              "reference\n", dlm->name, res->lockname.len,
1360                              res->lockname.name, to);
1361                         mle->master = to;
1362                         break;
1363                 case DLM_MASTER_RESP_NO:
1364                         mlog(0, "node %u not master, response=NO\n", to);
1365                         set_bit(to, mle->response_map);
1366                         break;
1367                 case DLM_MASTER_RESP_MAYBE:
1368                         mlog(0, "node %u not master, response=MAYBE\n", to);
1369                         set_bit(to, mle->response_map);
1370                         set_bit(to, mle->maybe_map);
1371                         break;
1372                 case DLM_MASTER_RESP_ERROR:
1373                         mlog(0, "node %u hit an error, resending\n", to);
1374                         resend = 1;
1375                         response = 0;
1376                         break;
1377                 default:
1378                         mlog(ML_ERROR, "bad response! %u\n", response);
1379                         BUG();
1380         }
1381         spin_unlock(&mle->spinlock);
1382         if (resend) {
1383                 /* this is also totally crude */
1384                 msleep(50);
1385                 goto again;
1386         }
1387
1388 out:
1389         return ret;
1390 }
1391
1392 /*
1393  * locks that can be taken here:
1394  * dlm->spinlock
1395  * res->spinlock
1396  * mle->spinlock
1397  * dlm->master_list
1398  *
1399  * if possible, TRIM THIS DOWN!!!
1400  */
1401 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1402                                void **ret_data)
1403 {
1404         u8 response = DLM_MASTER_RESP_MAYBE;
1405         struct dlm_ctxt *dlm = data;
1406         struct dlm_lock_resource *res = NULL;
1407         struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1408         struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1409         char *name;
1410         unsigned int namelen, hash;
1411         int found, ret;
1412         int set_maybe;
1413         int dispatch_assert = 0;
1414
1415         if (!dlm_grab(dlm))
1416                 return DLM_MASTER_RESP_NO;
1417
1418         if (!dlm_domain_fully_joined(dlm)) {
1419                 response = DLM_MASTER_RESP_NO;
1420                 goto send_response;
1421         }
1422
1423         name = request->name;
1424         namelen = request->namelen;
1425         hash = dlm_lockid_hash(name, namelen);
1426
1427         if (namelen > DLM_LOCKID_NAME_MAX) {
1428                 response = DLM_IVBUFLEN;
1429                 goto send_response;
1430         }
1431
1432 way_up_top:
1433         spin_lock(&dlm->spinlock);
1434         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1435         if (res) {
1436                 spin_unlock(&dlm->spinlock);
1437
1438                 /* take care of the easy cases up front */
1439                 spin_lock(&res->spinlock);
1440                 if (res->state & (DLM_LOCK_RES_RECOVERING|
1441                                   DLM_LOCK_RES_MIGRATING)) {
1442                         spin_unlock(&res->spinlock);
1443                         mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1444                              "being recovered/migrated\n");
1445                         response = DLM_MASTER_RESP_ERROR;
1446                         if (mle)
1447                                 kmem_cache_free(dlm_mle_cache, mle);
1448                         goto send_response;
1449                 }
1450
1451                 if (res->owner == dlm->node_num) {
1452                         dlm_lockres_set_refmap_bit(dlm, res, request->node_idx);
1453                         spin_unlock(&res->spinlock);
1454                         response = DLM_MASTER_RESP_YES;
1455                         if (mle)
1456                                 kmem_cache_free(dlm_mle_cache, mle);
1457
1458                         /* this node is the owner.
1459                          * there is some extra work that needs to
1460                          * happen now.  the requesting node has
1461                          * caused all nodes up to this one to
1462                          * create mles.  this node now needs to
1463                          * go back and clean those up. */
1464                         dispatch_assert = 1;
1465                         goto send_response;
1466                 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1467                         spin_unlock(&res->spinlock);
1468                         // mlog(0, "node %u is the master\n", res->owner);
1469                         response = DLM_MASTER_RESP_NO;
1470                         if (mle)
1471                                 kmem_cache_free(dlm_mle_cache, mle);
1472                         goto send_response;
1473                 }
1474
1475                 /* ok, there is no owner.  either this node is
1476                  * being blocked, or it is actively trying to
1477                  * master this lock. */
1478                 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1479                         mlog(ML_ERROR, "lock with no owner should be "
1480                              "in-progress!\n");
1481                         BUG();
1482                 }
1483
1484                 // mlog(0, "lockres is in progress...\n");
1485                 spin_lock(&dlm->master_lock);
1486                 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1487                 if (!found) {
1488                         mlog(ML_ERROR, "no mle found for this lock!\n");
1489                         BUG();
1490                 }
1491                 set_maybe = 1;
1492                 spin_lock(&tmpmle->spinlock);
1493                 if (tmpmle->type == DLM_MLE_BLOCK) {
1494                         // mlog(0, "this node is waiting for "
1495                         // "lockres to be mastered\n");
1496                         response = DLM_MASTER_RESP_NO;
1497                 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1498                         mlog(0, "node %u is master, but trying to migrate to "
1499                              "node %u.\n", tmpmle->master, tmpmle->new_master);
1500                         if (tmpmle->master == dlm->node_num) {
1501                                 mlog(ML_ERROR, "no owner on lockres, but this "
1502                                      "node is trying to migrate it to %u?!\n",
1503                                      tmpmle->new_master);
1504                                 BUG();
1505                         } else {
1506                                 /* the real master can respond on its own */
1507                                 response = DLM_MASTER_RESP_NO;
1508                         }
1509                 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1510                         set_maybe = 0;
1511                         if (tmpmle->master == dlm->node_num) {
1512                                 response = DLM_MASTER_RESP_YES;
1513                                 /* this node will be the owner.
1514                                  * go back and clean the mles on any
1515                                  * other nodes */
1516                                 dispatch_assert = 1;
1517                                 dlm_lockres_set_refmap_bit(dlm, res,
1518                                                            request->node_idx);
1519                         } else
1520                                 response = DLM_MASTER_RESP_NO;
1521                 } else {
1522                         // mlog(0, "this node is attempting to "
1523                         // "master lockres\n");
1524                         response = DLM_MASTER_RESP_MAYBE;
1525                 }
1526                 if (set_maybe)
1527                         set_bit(request->node_idx, tmpmle->maybe_map);
1528                 spin_unlock(&tmpmle->spinlock);
1529
1530                 spin_unlock(&dlm->master_lock);
1531                 spin_unlock(&res->spinlock);
1532
1533                 /* keep the mle attached to heartbeat events */
1534                 dlm_put_mle(tmpmle);
1535                 if (mle)
1536                         kmem_cache_free(dlm_mle_cache, mle);
1537                 goto send_response;
1538         }
1539
1540         /*
1541          * lockres doesn't exist on this node
1542          * if there is an MLE_BLOCK, return NO
1543          * if there is an MLE_MASTER, return MAYBE
1544          * otherwise, add an MLE_BLOCK, return NO
1545          */
1546         spin_lock(&dlm->master_lock);
1547         found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1548         if (!found) {
1549                 /* this lockid has never been seen on this node yet */
1550                 // mlog(0, "no mle found\n");
1551                 if (!mle) {
1552                         spin_unlock(&dlm->master_lock);
1553                         spin_unlock(&dlm->spinlock);
1554
1555                         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1556                         if (!mle) {
1557                                 response = DLM_MASTER_RESP_ERROR;
1558                                 mlog_errno(-ENOMEM);
1559                                 goto send_response;
1560                         }
1561                         goto way_up_top;
1562                 }
1563
1564                 // mlog(0, "this is second time thru, already allocated, "
1565                 // "add the block.\n");
1566                 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1567                 set_bit(request->node_idx, mle->maybe_map);
1568                 __dlm_insert_mle(dlm, mle);
1569                 response = DLM_MASTER_RESP_NO;
1570         } else {
1571                 // mlog(0, "mle was found\n");
1572                 set_maybe = 1;
1573                 spin_lock(&tmpmle->spinlock);
1574                 if (tmpmle->master == dlm->node_num) {
1575                         mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1576                         BUG();
1577                 }
1578                 if (tmpmle->type == DLM_MLE_BLOCK)
1579                         response = DLM_MASTER_RESP_NO;
1580                 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1581                         mlog(0, "migration mle was found (%u->%u)\n",
1582                              tmpmle->master, tmpmle->new_master);
1583                         /* real master can respond on its own */
1584                         response = DLM_MASTER_RESP_NO;
1585                 } else
1586                         response = DLM_MASTER_RESP_MAYBE;
1587                 if (set_maybe)
1588                         set_bit(request->node_idx, tmpmle->maybe_map);
1589                 spin_unlock(&tmpmle->spinlock);
1590         }
1591         spin_unlock(&dlm->master_lock);
1592         spin_unlock(&dlm->spinlock);
1593
1594         if (found) {
1595                 /* keep the mle attached to heartbeat events */
1596                 dlm_put_mle(tmpmle);
1597         }
1598 send_response:
1599         /*
1600          * __dlm_lookup_lockres() grabbed a reference to this lockres.
1601          * The reference is released by dlm_assert_master_worker() under
1602          * the call to dlm_dispatch_assert_master().  If
1603          * dlm_assert_master_worker() isn't called, we drop it here.
1604          */
1605         if (dispatch_assert) {
1606                 if (response != DLM_MASTER_RESP_YES)
1607                         mlog(ML_ERROR, "invalid response %d\n", response);
1608                 if (!res) {
1609                         mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1610                         BUG();
1611                 }
1612                 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1613                              dlm->node_num, res->lockname.len, res->lockname.name);
1614                 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1615                                                  DLM_ASSERT_MASTER_MLE_CLEANUP);
1616                 if (ret < 0) {
1617                         mlog(ML_ERROR, "failed to dispatch assert master work\n");
1618                         response = DLM_MASTER_RESP_ERROR;
1619                         dlm_lockres_put(res);
1620                 }
1621         } else {
1622                 if (res)
1623                         dlm_lockres_put(res);
1624         }
1625
1626         dlm_put(dlm);
1627         return response;
1628 }
1629
1630 /*
1631  * DLM_ASSERT_MASTER_MSG
1632  */
1633
1634
1635 /*
1636  * NOTE: this can be used for debugging
1637  * can periodically run all locks owned by this node
1638  * and re-assert across the cluster...
1639  */
1640 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1641                                 struct dlm_lock_resource *res,
1642                                 void *nodemap, u32 flags)
1643 {
1644         struct dlm_assert_master assert;
1645         int to, tmpret;
1646         struct dlm_node_iter iter;
1647         int ret = 0;
1648         int reassert;
1649         const char *lockname = res->lockname.name;
1650         unsigned int namelen = res->lockname.len;
1651
1652         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1653
1654         spin_lock(&res->spinlock);
1655         res->state |= DLM_LOCK_RES_SETREF_INPROG;
1656         spin_unlock(&res->spinlock);
1657
1658 again:
1659         reassert = 0;
1660
1661         /* note that if this nodemap is empty, it returns 0 */
1662         dlm_node_iter_init(nodemap, &iter);
1663         while ((to = dlm_node_iter_next(&iter)) >= 0) {
1664                 int r = 0;
1665                 struct dlm_master_list_entry *mle = NULL;
1666
1667                 mlog(0, "sending assert master to %d (%.*s)\n", to,
1668                      namelen, lockname);
1669                 memset(&assert, 0, sizeof(assert));
1670                 assert.node_idx = dlm->node_num;
1671                 assert.namelen = namelen;
1672                 memcpy(assert.name, lockname, namelen);
1673                 assert.flags = cpu_to_be32(flags);
1674
1675                 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1676                                             &assert, sizeof(assert), to, &r);
1677                 if (tmpret < 0) {
1678                         mlog(ML_ERROR, "Error %d when sending message %u (key "
1679                              "0x%x) to node %u\n", tmpret,
1680                              DLM_ASSERT_MASTER_MSG, dlm->key, to);
1681                         if (!dlm_is_host_down(tmpret)) {
1682                                 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1683                                 BUG();
1684                         }
1685                         /* a node died.  finish out the rest of the nodes. */
1686                         mlog(0, "link to %d went down!\n", to);
1687                         /* any nonzero status return will do */
1688                         ret = tmpret;
1689                         r = 0;
1690                 } else if (r < 0) {
1691                         /* ok, something horribly messed.  kill thyself. */
1692                         mlog(ML_ERROR,"during assert master of %.*s to %u, "
1693                              "got %d.\n", namelen, lockname, to, r);
1694                         spin_lock(&dlm->spinlock);
1695                         spin_lock(&dlm->master_lock);
1696                         if (dlm_find_mle(dlm, &mle, (char *)lockname,
1697                                          namelen)) {
1698                                 dlm_print_one_mle(mle);
1699                                 __dlm_put_mle(mle);
1700                         }
1701                         spin_unlock(&dlm->master_lock);
1702                         spin_unlock(&dlm->spinlock);
1703                         BUG();
1704                 }
1705
1706                 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1707                     !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1708                                 mlog(ML_ERROR, "%.*s: very strange, "
1709                                      "master MLE but no lockres on %u\n",
1710                                      namelen, lockname, to);
1711                 }
1712
1713                 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1714                         mlog(0, "%.*s: node %u create mles on other "
1715                              "nodes and requests a re-assert\n",
1716                              namelen, lockname, to);
1717                         reassert = 1;
1718                 }
1719                 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1720                         mlog(0, "%.*s: node %u has a reference to this "
1721                              "lockres, set the bit in the refmap\n",
1722                              namelen, lockname, to);
1723                         spin_lock(&res->spinlock);
1724                         dlm_lockres_set_refmap_bit(dlm, res, to);
1725                         spin_unlock(&res->spinlock);
1726                 }
1727         }
1728
1729         if (reassert)
1730                 goto again;
1731
1732         spin_lock(&res->spinlock);
1733         res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1734         spin_unlock(&res->spinlock);
1735         wake_up(&res->wq);
1736
1737         return ret;
1738 }
1739
1740 /*
1741  * locks that can be taken here:
1742  * dlm->spinlock
1743  * res->spinlock
1744  * mle->spinlock
1745  * dlm->master_list
1746  *
1747  * if possible, TRIM THIS DOWN!!!
1748  */
1749 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1750                               void **ret_data)
1751 {
1752         struct dlm_ctxt *dlm = data;
1753         struct dlm_master_list_entry *mle = NULL;
1754         struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1755         struct dlm_lock_resource *res = NULL;
1756         char *name;
1757         unsigned int namelen, hash;
1758         u32 flags;
1759         int master_request = 0, have_lockres_ref = 0;
1760         int ret = 0;
1761
1762         if (!dlm_grab(dlm))
1763                 return 0;
1764
1765         name = assert->name;
1766         namelen = assert->namelen;
1767         hash = dlm_lockid_hash(name, namelen);
1768         flags = be32_to_cpu(assert->flags);
1769
1770         if (namelen > DLM_LOCKID_NAME_MAX) {
1771                 mlog(ML_ERROR, "Invalid name length!");
1772                 goto done;
1773         }
1774
1775         spin_lock(&dlm->spinlock);
1776
1777         if (flags)
1778                 mlog(0, "assert_master with flags: %u\n", flags);
1779
1780         /* find the MLE */
1781         spin_lock(&dlm->master_lock);
1782         if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1783                 /* not an error, could be master just re-asserting */
1784                 mlog(0, "just got an assert_master from %u, but no "
1785                      "MLE for it! (%.*s)\n", assert->node_idx,
1786                      namelen, name);
1787         } else {
1788                 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1789                 if (bit >= O2NM_MAX_NODES) {
1790                         /* not necessarily an error, though less likely.
1791                          * could be master just re-asserting. */
1792                         mlog(0, "no bits set in the maybe_map, but %u "
1793                              "is asserting! (%.*s)\n", assert->node_idx,
1794                              namelen, name);
1795                 } else if (bit != assert->node_idx) {
1796                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1797                                 mlog(0, "master %u was found, %u should "
1798                                      "back off\n", assert->node_idx, bit);
1799                         } else {
1800                                 /* with the fix for bug 569, a higher node
1801                                  * number winning the mastery will respond
1802                                  * YES to mastery requests, but this node
1803                                  * had no way of knowing.  let it pass. */
1804                                 mlog(0, "%u is the lowest node, "
1805                                      "%u is asserting. (%.*s)  %u must "
1806                                      "have begun after %u won.\n", bit,
1807                                      assert->node_idx, namelen, name, bit,
1808                                      assert->node_idx);
1809                         }
1810                 }
1811                 if (mle->type == DLM_MLE_MIGRATION) {
1812                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1813                                 mlog(0, "%s:%.*s: got cleanup assert"
1814                                      " from %u for migration\n",
1815                                      dlm->name, namelen, name,
1816                                      assert->node_idx);
1817                         } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1818                                 mlog(0, "%s:%.*s: got unrelated assert"
1819                                      " from %u for migration, ignoring\n",
1820                                      dlm->name, namelen, name,
1821                                      assert->node_idx);
1822                                 __dlm_put_mle(mle);
1823                                 spin_unlock(&dlm->master_lock);
1824                                 spin_unlock(&dlm->spinlock);
1825                                 goto done;
1826                         }
1827                 }
1828         }
1829         spin_unlock(&dlm->master_lock);
1830
1831         /* ok everything checks out with the MLE
1832          * now check to see if there is a lockres */
1833         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1834         if (res) {
1835                 spin_lock(&res->spinlock);
1836                 if (res->state & DLM_LOCK_RES_RECOVERING)  {
1837                         mlog(ML_ERROR, "%u asserting but %.*s is "
1838                              "RECOVERING!\n", assert->node_idx, namelen, name);
1839                         goto kill;
1840                 }
1841                 if (!mle) {
1842                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1843                             res->owner != assert->node_idx) {
1844                                 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1845                                      "but current owner is %u! (%.*s)\n",
1846                                      assert->node_idx, res->owner, namelen,
1847                                      name);
1848                                 __dlm_print_one_lock_resource(res);
1849                                 BUG();
1850                         }
1851                 } else if (mle->type != DLM_MLE_MIGRATION) {
1852                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1853                                 /* owner is just re-asserting */
1854                                 if (res->owner == assert->node_idx) {
1855                                         mlog(0, "owner %u re-asserting on "
1856                                              "lock %.*s\n", assert->node_idx,
1857                                              namelen, name);
1858                                         goto ok;
1859                                 }
1860                                 mlog(ML_ERROR, "got assert_master from "
1861                                      "node %u, but %u is the owner! "
1862                                      "(%.*s)\n", assert->node_idx,
1863                                      res->owner, namelen, name);
1864                                 goto kill;
1865                         }
1866                         if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1867                                 mlog(ML_ERROR, "got assert from %u, but lock "
1868                                      "with no owner should be "
1869                                      "in-progress! (%.*s)\n",
1870                                      assert->node_idx,
1871                                      namelen, name);
1872                                 goto kill;
1873                         }
1874                 } else /* mle->type == DLM_MLE_MIGRATION */ {
1875                         /* should only be getting an assert from new master */
1876                         if (assert->node_idx != mle->new_master) {
1877                                 mlog(ML_ERROR, "got assert from %u, but "
1878                                      "new master is %u, and old master "
1879                                      "was %u (%.*s)\n",
1880                                      assert->node_idx, mle->new_master,
1881                                      mle->master, namelen, name);
1882                                 goto kill;
1883                         }
1884
1885                 }
1886 ok:
1887                 spin_unlock(&res->spinlock);
1888         }
1889
1890         // mlog(0, "woo!  got an assert_master from node %u!\n",
1891         //           assert->node_idx);
1892         if (mle) {
1893                 int extra_ref = 0;
1894                 int nn = -1;
1895                 int rr, err = 0;
1896
1897                 spin_lock(&mle->spinlock);
1898                 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1899                         extra_ref = 1;
1900                 else {
1901                         /* MASTER mle: if any bits set in the response map
1902                          * then the calling node needs to re-assert to clear
1903                          * up nodes that this node contacted */
1904                         while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1905                                                     nn+1)) < O2NM_MAX_NODES) {
1906                                 if (nn != dlm->node_num && nn != assert->node_idx)
1907                                         master_request = 1;
1908                         }
1909                 }
1910                 mle->master = assert->node_idx;
1911                 atomic_set(&mle->woken, 1);
1912                 wake_up(&mle->wq);
1913                 spin_unlock(&mle->spinlock);
1914
1915                 if (res) {
1916                         int wake = 0;
1917                         spin_lock(&res->spinlock);
1918                         if (mle->type == DLM_MLE_MIGRATION) {
1919                                 mlog(0, "finishing off migration of lockres %.*s, "
1920                                         "from %u to %u\n",
1921                                         res->lockname.len, res->lockname.name,
1922                                         dlm->node_num, mle->new_master);
1923                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
1924                                 wake = 1;
1925                                 dlm_change_lockres_owner(dlm, res, mle->new_master);
1926                                 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1927                         } else {
1928                                 dlm_change_lockres_owner(dlm, res, mle->master);
1929                         }
1930                         spin_unlock(&res->spinlock);
1931                         have_lockres_ref = 1;
1932                         if (wake)
1933                                 wake_up(&res->wq);
1934                 }
1935
1936                 /* master is known, detach if not already detached.
1937                  * ensures that only one assert_master call will happen
1938                  * on this mle. */
1939                 spin_lock(&dlm->master_lock);
1940
1941                 rr = atomic_read(&mle->mle_refs.refcount);
1942                 if (mle->inuse > 0) {
1943                         if (extra_ref && rr < 3)
1944                                 err = 1;
1945                         else if (!extra_ref && rr < 2)
1946                                 err = 1;
1947                 } else {
1948                         if (extra_ref && rr < 2)
1949                                 err = 1;
1950                         else if (!extra_ref && rr < 1)
1951                                 err = 1;
1952                 }
1953                 if (err) {
1954                         mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
1955                              "that will mess up this node, refs=%d, extra=%d, "
1956                              "inuse=%d\n", dlm->name, namelen, name,
1957                              assert->node_idx, rr, extra_ref, mle->inuse);
1958                         dlm_print_one_mle(mle);
1959                 }
1960                 __dlm_unlink_mle(dlm, mle);
1961                 __dlm_mle_detach_hb_events(dlm, mle);
1962                 __dlm_put_mle(mle);
1963                 if (extra_ref) {
1964                         /* the assert master message now balances the extra
1965                          * ref given by the master / migration request message.
1966                          * if this is the last put, it will be removed
1967                          * from the list. */
1968                         __dlm_put_mle(mle);
1969                 }
1970                 spin_unlock(&dlm->master_lock);
1971         } else if (res) {
1972                 if (res->owner != assert->node_idx) {
1973                         mlog(0, "assert_master from %u, but current "
1974                              "owner is %u (%.*s), no mle\n", assert->node_idx,
1975                              res->owner, namelen, name);
1976                 }
1977         }
1978         spin_unlock(&dlm->spinlock);
1979
1980 done:
1981         ret = 0;
1982         if (res) {
1983                 spin_lock(&res->spinlock);
1984                 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1985                 spin_unlock(&res->spinlock);
1986                 *ret_data = (void *)res;
1987         }
1988         dlm_put(dlm);
1989         if (master_request) {
1990                 mlog(0, "need to tell master to reassert\n");
1991                 /* positive. negative would shoot down the node. */
1992                 ret |= DLM_ASSERT_RESPONSE_REASSERT;
1993                 if (!have_lockres_ref) {
1994                         mlog(ML_ERROR, "strange, got assert from %u, MASTER "
1995                              "mle present here for %s:%.*s, but no lockres!\n",
1996                              assert->node_idx, dlm->name, namelen, name);
1997                 }
1998         }
1999         if (have_lockres_ref) {
2000                 /* let the master know we have a reference to the lockres */
2001                 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2002                 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2003                      dlm->name, namelen, name, assert->node_idx);
2004         }
2005         return ret;
2006
2007 kill:
2008         /* kill the caller! */
2009         mlog(ML_ERROR, "Bad message received from another node.  Dumping state "
2010              "and killing the other node now!  This node is OK and can continue.\n");
2011         __dlm_print_one_lock_resource(res);
2012         spin_unlock(&res->spinlock);
2013         spin_unlock(&dlm->spinlock);
2014         *ret_data = (void *)res;
2015         dlm_put(dlm);
2016         return -EINVAL;
2017 }
2018
2019 void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2020 {
2021         struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2022
2023         if (ret_data) {
2024                 spin_lock(&res->spinlock);
2025                 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2026                 spin_unlock(&res->spinlock);
2027                 wake_up(&res->wq);
2028                 dlm_lockres_put(res);
2029         }
2030         return;
2031 }
2032
2033 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2034                                struct dlm_lock_resource *res,
2035                                int ignore_higher, u8 request_from, u32 flags)
2036 {
2037         struct dlm_work_item *item;
2038         item = kzalloc(sizeof(*item), GFP_NOFS);
2039         if (!item)
2040                 return -ENOMEM;
2041
2042
2043         /* queue up work for dlm_assert_master_worker */
2044         dlm_grab(dlm);  /* get an extra ref for the work item */
2045         dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2046         item->u.am.lockres = res; /* already have a ref */
2047         /* can optionally ignore node numbers higher than this node */
2048         item->u.am.ignore_higher = ignore_higher;
2049         item->u.am.request_from = request_from;
2050         item->u.am.flags = flags;
2051
2052         if (ignore_higher)
2053                 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2054                      res->lockname.name);
2055
2056         spin_lock(&dlm->work_lock);
2057         list_add_tail(&item->list, &dlm->work_list);
2058         spin_unlock(&dlm->work_lock);
2059
2060         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2061         return 0;
2062 }
2063
2064 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2065 {
2066         struct dlm_ctxt *dlm = data;
2067         int ret = 0;
2068         struct dlm_lock_resource *res;
2069         unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2070         int ignore_higher;
2071         int bit;
2072         u8 request_from;
2073         u32 flags;
2074
2075         dlm = item->dlm;
2076         res = item->u.am.lockres;
2077         ignore_higher = item->u.am.ignore_higher;
2078         request_from = item->u.am.request_from;
2079         flags = item->u.am.flags;
2080
2081         spin_lock(&dlm->spinlock);
2082         memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2083         spin_unlock(&dlm->spinlock);
2084
2085         clear_bit(dlm->node_num, nodemap);
2086         if (ignore_higher) {
2087                 /* if is this just to clear up mles for nodes below
2088                  * this node, do not send the message to the original
2089                  * caller or any node number higher than this */
2090                 clear_bit(request_from, nodemap);
2091                 bit = dlm->node_num;
2092                 while (1) {
2093                         bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2094                                             bit+1);
2095                         if (bit >= O2NM_MAX_NODES)
2096                                 break;
2097                         clear_bit(bit, nodemap);
2098                 }
2099         }
2100
2101         /*
2102          * If we're migrating this lock to someone else, we are no
2103          * longer allowed to assert out own mastery.  OTOH, we need to
2104          * prevent migration from starting while we're still asserting
2105          * our dominance.  The reserved ast delays migration.
2106          */
2107         spin_lock(&res->spinlock);
2108         if (res->state & DLM_LOCK_RES_MIGRATING) {
2109                 mlog(0, "Someone asked us to assert mastery, but we're "
2110                      "in the middle of migration.  Skipping assert, "
2111                      "the new master will handle that.\n");
2112                 spin_unlock(&res->spinlock);
2113                 goto put;
2114         } else
2115                 __dlm_lockres_reserve_ast(res);
2116         spin_unlock(&res->spinlock);
2117
2118         /* this call now finishes out the nodemap
2119          * even if one or more nodes die */
2120         mlog(0, "worker about to master %.*s here, this=%u\n",
2121                      res->lockname.len, res->lockname.name, dlm->node_num);
2122         ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2123         if (ret < 0) {
2124                 /* no need to restart, we are done */
2125                 if (!dlm_is_host_down(ret))
2126                         mlog_errno(ret);
2127         }
2128
2129         /* Ok, we've asserted ourselves.  Let's let migration start. */
2130         dlm_lockres_release_ast(dlm, res);
2131
2132 put:
2133         dlm_lockres_put(res);
2134
2135         mlog(0, "finished with dlm_assert_master_worker\n");
2136 }
2137
2138 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2139  * We cannot wait for node recovery to complete to begin mastering this
2140  * lockres because this lockres is used to kick off recovery! ;-)
2141  * So, do a pre-check on all living nodes to see if any of those nodes
2142  * think that $RECOVERY is currently mastered by a dead node.  If so,
2143  * we wait a short time to allow that node to get notified by its own
2144  * heartbeat stack, then check again.  All $RECOVERY lock resources
2145  * mastered by dead nodes are purged when the hearbeat callback is
2146  * fired, so we can know for sure that it is safe to continue once
2147  * the node returns a live node or no node.  */
2148 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2149                                        struct dlm_lock_resource *res)
2150 {
2151         struct dlm_node_iter iter;
2152         int nodenum;
2153         int ret = 0;
2154         u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2155
2156         spin_lock(&dlm->spinlock);
2157         dlm_node_iter_init(dlm->domain_map, &iter);
2158         spin_unlock(&dlm->spinlock);
2159
2160         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2161                 /* do not send to self */
2162                 if (nodenum == dlm->node_num)
2163                         continue;
2164                 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2165                 if (ret < 0) {
2166                         mlog_errno(ret);
2167                         if (!dlm_is_host_down(ret))
2168                                 BUG();
2169                         /* host is down, so answer for that node would be
2170                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
2171                         ret = 0;
2172                 }
2173
2174                 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2175                         /* check to see if this master is in the recovery map */
2176                         spin_lock(&dlm->spinlock);
2177                         if (test_bit(master, dlm->recovery_map)) {
2178                                 mlog(ML_NOTICE, "%s: node %u has not seen "
2179                                      "node %u go down yet, and thinks the "
2180                                      "dead node is mastering the recovery "
2181                                      "lock.  must wait.\n", dlm->name,
2182                                      nodenum, master);
2183                                 ret = -EAGAIN;
2184                         }
2185                         spin_unlock(&dlm->spinlock);
2186                         mlog(0, "%s: reco lock master is %u\n", dlm->name,
2187                              master);
2188                         break;
2189                 }
2190         }
2191         return ret;
2192 }
2193
2194 /*
2195  * DLM_DEREF_LOCKRES_MSG
2196  */
2197
2198 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2199 {
2200         struct dlm_deref_lockres deref;
2201         int ret = 0, r;
2202         const char *lockname;
2203         unsigned int namelen;
2204
2205         lockname = res->lockname.name;
2206         namelen = res->lockname.len;
2207         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2208
2209         memset(&deref, 0, sizeof(deref));
2210         deref.node_idx = dlm->node_num;
2211         deref.namelen = namelen;
2212         memcpy(deref.name, lockname, namelen);
2213
2214         ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2215                                  &deref, sizeof(deref), res->owner, &r);
2216         if (ret < 0)
2217                 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n",
2218                      dlm->name, namelen, lockname, ret, res->owner);
2219         else if (r < 0) {
2220                 /* BAD.  other node says I did not have a ref. */
2221                 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2222                      dlm->name, namelen, lockname, res->owner, r);
2223                 dlm_print_one_lock_resource(res);
2224                 BUG();
2225         }
2226         return ret;
2227 }
2228
2229 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2230                               void **ret_data)
2231 {
2232         struct dlm_ctxt *dlm = data;
2233         struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2234         struct dlm_lock_resource *res = NULL;
2235         char *name;
2236         unsigned int namelen;
2237         int ret = -EINVAL;
2238         u8 node;
2239         unsigned int hash;
2240         struct dlm_work_item *item;
2241         int cleared = 0;
2242         int dispatch = 0;
2243
2244         if (!dlm_grab(dlm))
2245                 return 0;
2246
2247         name = deref->name;
2248         namelen = deref->namelen;
2249         node = deref->node_idx;
2250
2251         if (namelen > DLM_LOCKID_NAME_MAX) {
2252                 mlog(ML_ERROR, "Invalid name length!");
2253                 goto done;
2254         }
2255         if (deref->node_idx >= O2NM_MAX_NODES) {
2256                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2257                 goto done;
2258         }
2259
2260         hash = dlm_lockid_hash(name, namelen);
2261
2262         spin_lock(&dlm->spinlock);
2263         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2264         if (!res) {
2265                 spin_unlock(&dlm->spinlock);
2266                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2267                      dlm->name, namelen, name);
2268                 goto done;
2269         }
2270         spin_unlock(&dlm->spinlock);
2271
2272         spin_lock(&res->spinlock);
2273         if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2274                 dispatch = 1;
2275         else {
2276                 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2277                 if (test_bit(node, res->refmap)) {
2278                         dlm_lockres_clear_refmap_bit(dlm, res, node);
2279                         cleared = 1;
2280                 }
2281         }
2282         spin_unlock(&res->spinlock);
2283
2284         if (!dispatch) {
2285                 if (cleared)
2286                         dlm_lockres_calc_usage(dlm, res);
2287                 else {
2288                         mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2289                         "but it is already dropped!\n", dlm->name,
2290                         res->lockname.len, res->lockname.name, node);
2291                         dlm_print_one_lock_resource(res);
2292                 }
2293                 ret = 0;
2294                 goto done;
2295         }
2296
2297         item = kzalloc(sizeof(*item), GFP_NOFS);
2298         if (!item) {
2299                 ret = -ENOMEM;
2300                 mlog_errno(ret);
2301                 goto done;
2302         }
2303
2304         dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2305         item->u.dl.deref_res = res;
2306         item->u.dl.deref_node = node;
2307
2308         spin_lock(&dlm->work_lock);
2309         list_add_tail(&item->list, &dlm->work_list);
2310         spin_unlock(&dlm->work_lock);
2311
2312         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2313         return 0;
2314
2315 done:
2316         if (res)
2317                 dlm_lockres_put(res);
2318         dlm_put(dlm);
2319
2320         return ret;
2321 }
2322
2323 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2324 {
2325         struct dlm_ctxt *dlm;
2326         struct dlm_lock_resource *res;
2327         u8 node;
2328         u8 cleared = 0;
2329
2330         dlm = item->dlm;
2331         res = item->u.dl.deref_res;
2332         node = item->u.dl.deref_node;
2333
2334         spin_lock(&res->spinlock);
2335         BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2336         if (test_bit(node, res->refmap)) {
2337                 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2338                 dlm_lockres_clear_refmap_bit(dlm, res, node);
2339                 cleared = 1;
2340         }
2341         spin_unlock(&res->spinlock);
2342
2343         if (cleared) {
2344                 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2345                      dlm->name, res->lockname.len, res->lockname.name, node);
2346                 dlm_lockres_calc_usage(dlm, res);
2347         } else {
2348                 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2349                      "but it is already dropped!\n", dlm->name,
2350                      res->lockname.len, res->lockname.name, node);
2351                 dlm_print_one_lock_resource(res);
2352         }
2353
2354         dlm_lockres_put(res);
2355 }
2356
2357 /*
2358  * A migrateable resource is one that is :
2359  * 1. locally mastered, and,
2360  * 2. zero local locks, and,
2361  * 3. one or more non-local locks, or, one or more references
2362  * Returns 1 if yes, 0 if not.
2363  */
2364 static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
2365                                       struct dlm_lock_resource *res)
2366 {
2367         enum dlm_lockres_list idx;
2368         int nonlocal = 0, node_ref;
2369         struct list_head *queue;
2370         struct dlm_lock *lock;
2371         u64 cookie;
2372
2373         assert_spin_locked(&res->spinlock);
2374
2375         if (res->owner != dlm->node_num)
2376                 return 0;
2377
2378         for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2379                 queue = dlm_list_idx_to_ptr(res, idx);
2380                 list_for_each_entry(lock, queue, list) {
2381                         if (lock->ml.node != dlm->node_num) {
2382                                 nonlocal++;
2383                                 continue;
2384                         }
2385                         cookie = be64_to_cpu(lock->ml.cookie);
2386                         mlog(0, "%s: Not migrateable res %.*s, lock %u:%llu on "
2387                              "%s list\n", dlm->name, res->lockname.len,
2388                              res->lockname.name,
2389                              dlm_get_lock_cookie_node(cookie),
2390                              dlm_get_lock_cookie_seq(cookie),
2391                              dlm_list_in_text(idx));
2392                         return 0;
2393                 }
2394         }
2395
2396         if (!nonlocal) {
2397                 node_ref = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
2398                 if (node_ref >= O2NM_MAX_NODES)
2399                         return 0;
2400         }
2401
2402         mlog(0, "%s: res %.*s, Migrateable\n", dlm->name, res->lockname.len,
2403              res->lockname.name);
2404
2405         return 1;
2406 }
2407
2408 /*
2409  * DLM_MIGRATE_LOCKRES
2410  */
2411
2412
2413 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2414                                struct dlm_lock_resource *res, u8 target)
2415 {
2416         struct dlm_master_list_entry *mle = NULL;
2417         struct dlm_master_list_entry *oldmle = NULL;
2418         struct dlm_migratable_lockres *mres = NULL;
2419         int ret = 0;
2420         const char *name;
2421         unsigned int namelen;
2422         int mle_added = 0;
2423         int wake = 0;
2424
2425         if (!dlm_grab(dlm))
2426                 return -EINVAL;
2427
2428         BUG_ON(target == O2NM_MAX_NODES);
2429
2430         name = res->lockname.name;
2431         namelen = res->lockname.len;
2432
2433         mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name,
2434              target);
2435
2436         /* preallocate up front. if this fails, abort */
2437         ret = -ENOMEM;
2438         mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2439         if (!mres) {
2440                 mlog_errno(ret);
2441                 goto leave;
2442         }
2443
2444         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2445         if (!mle) {
2446                 mlog_errno(ret);
2447                 goto leave;
2448         }
2449         ret = 0;
2450
2451         /*
2452          * clear any existing master requests and
2453          * add the migration mle to the list
2454          */
2455         spin_lock(&dlm->spinlock);
2456         spin_lock(&dlm->master_lock);
2457         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2458                                     namelen, target, dlm->node_num);
2459         spin_unlock(&dlm->master_lock);
2460         spin_unlock(&dlm->spinlock);
2461
2462         if (ret == -EEXIST) {
2463                 mlog(0, "another process is already migrating it\n");
2464                 goto fail;
2465         }
2466         mle_added = 1;
2467
2468         /*
2469          * set the MIGRATING flag and flush asts
2470          * if we fail after this we need to re-dirty the lockres
2471          */
2472         if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2473                 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2474                      "the target went down.\n", res->lockname.len,
2475                      res->lockname.name, target);
2476                 spin_lock(&res->spinlock);
2477                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2478                 wake = 1;
2479                 spin_unlock(&res->spinlock);
2480                 ret = -EINVAL;
2481         }
2482
2483 fail:
2484         if (oldmle) {
2485                 /* master is known, detach if not already detached */
2486                 dlm_mle_detach_hb_events(dlm, oldmle);
2487                 dlm_put_mle(oldmle);
2488         }
2489
2490         if (ret < 0) {
2491                 if (mle_added) {
2492                         dlm_mle_detach_hb_events(dlm, mle);
2493                         dlm_put_mle(mle);
2494                 } else if (mle) {
2495                         kmem_cache_free(dlm_mle_cache, mle);
2496                         mle = NULL;
2497                 }
2498                 goto leave;
2499         }
2500
2501         /*
2502          * at this point, we have a migration target, an mle
2503          * in the master list, and the MIGRATING flag set on
2504          * the lockres
2505          */
2506
2507         /* now that remote nodes are spinning on the MIGRATING flag,
2508          * ensure that all assert_master work is flushed. */
2509         flush_workqueue(dlm->dlm_worker);
2510
2511         /* get an extra reference on the mle.
2512          * otherwise the assert_master from the new
2513          * master will destroy this.
2514          * also, make sure that all callers of dlm_get_mle
2515          * take both dlm->spinlock and dlm->master_lock */
2516         spin_lock(&dlm->spinlock);
2517         spin_lock(&dlm->master_lock);
2518         dlm_get_mle_inuse(mle);
2519         spin_unlock(&dlm->master_lock);
2520         spin_unlock(&dlm->spinlock);
2521
2522         /* notify new node and send all lock state */
2523         /* call send_one_lockres with migration flag.
2524          * this serves as notice to the target node that a
2525          * migration is starting. */
2526         ret = dlm_send_one_lockres(dlm, res, mres, target,
2527                                    DLM_MRES_MIGRATION);
2528
2529         if (ret < 0) {
2530                 mlog(0, "migration to node %u failed with %d\n",
2531                      target, ret);
2532                 /* migration failed, detach and clean up mle */
2533                 dlm_mle_detach_hb_events(dlm, mle);
2534                 dlm_put_mle(mle);
2535                 dlm_put_mle_inuse(mle);
2536                 spin_lock(&res->spinlock);
2537                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2538                 wake = 1;
2539                 spin_unlock(&res->spinlock);
2540                 if (dlm_is_host_down(ret))
2541                         dlm_wait_for_node_death(dlm, target,
2542                                                 DLM_NODE_DEATH_WAIT_MAX);
2543                 goto leave;
2544         }
2545
2546         /* at this point, the target sends a message to all nodes,
2547          * (using dlm_do_migrate_request).  this node is skipped since
2548          * we had to put an mle in the list to begin the process.  this
2549          * node now waits for target to do an assert master.  this node
2550          * will be the last one notified, ensuring that the migration
2551          * is complete everywhere.  if the target dies while this is
2552          * going on, some nodes could potentially see the target as the
2553          * master, so it is important that my recovery finds the migration
2554          * mle and sets the master to UNKNOWN. */
2555
2556
2557         /* wait for new node to assert master */
2558         while (1) {
2559                 ret = wait_event_interruptible_timeout(mle->wq,
2560                                         (atomic_read(&mle->woken) == 1),
2561                                         msecs_to_jiffies(5000));
2562
2563                 if (ret >= 0) {
2564                         if (atomic_read(&mle->woken) == 1 ||
2565                             res->owner == target)
2566                                 break;
2567
2568                         mlog(0, "%s:%.*s: timed out during migration\n",
2569                              dlm->name, res->lockname.len, res->lockname.name);
2570                         /* avoid hang during shutdown when migrating lockres
2571                          * to a node which also goes down */
2572                         if (dlm_is_node_dead(dlm, target)) {
2573                                 mlog(0, "%s:%.*s: expected migration "
2574                                      "target %u is no longer up, restarting\n",
2575                                      dlm->name, res->lockname.len,
2576                                      res->lockname.name, target);
2577                                 ret = -EINVAL;
2578                                 /* migration failed, detach and clean up mle */
2579                                 dlm_mle_detach_hb_events(dlm, mle);
2580                                 dlm_put_mle(mle);
2581                                 dlm_put_mle_inuse(mle);
2582                                 spin_lock(&res->spinlock);
2583                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2584                                 wake = 1;
2585                                 spin_unlock(&res->spinlock);
2586                                 goto leave;
2587                         }
2588                 } else
2589                         mlog(0, "%s:%.*s: caught signal during migration\n",
2590                              dlm->name, res->lockname.len, res->lockname.name);
2591         }
2592
2593         /* all done, set the owner, clear the flag */
2594         spin_lock(&res->spinlock);
2595         dlm_set_lockres_owner(dlm, res, target);
2596         res->state &= ~DLM_LOCK_RES_MIGRATING;
2597         dlm_remove_nonlocal_locks(dlm, res);
2598         spin_unlock(&res->spinlock);
2599         wake_up(&res->wq);
2600
2601         /* master is known, detach if not already detached */
2602         dlm_mle_detach_hb_events(dlm, mle);
2603         dlm_put_mle_inuse(mle);
2604         ret = 0;
2605
2606         dlm_lockres_calc_usage(dlm, res);
2607
2608 leave:
2609         /* re-dirty the lockres if we failed */
2610         if (ret < 0)
2611                 dlm_kick_thread(dlm, res);
2612
2613         /* wake up waiters if the MIGRATING flag got set
2614          * but migration failed */
2615         if (wake)
2616                 wake_up(&res->wq);
2617
2618         if (mres)
2619                 free_page((unsigned long)mres);
2620
2621         dlm_put(dlm);
2622
2623         mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen,
2624              name, target, ret);
2625         return ret;
2626 }
2627
2628 #define DLM_MIGRATION_RETRY_MS  100
2629
2630 /*
2631  * Should be called only after beginning the domain leave process.
2632  * There should not be any remaining locks on nonlocal lock resources,
2633  * and there should be no local locks left on locally mastered resources.
2634  *
2635  * Called with the dlm spinlock held, may drop it to do migration, but
2636  * will re-acquire before exit.
2637  *
2638  * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2639  */
2640 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2641 {
2642         int ret;
2643         int lock_dropped = 0;
2644         u8 target = O2NM_MAX_NODES;
2645
2646         assert_spin_locked(&dlm->spinlock);
2647
2648         spin_lock(&res->spinlock);
2649         if (dlm_is_lockres_migrateable(dlm, res))
2650                 target = dlm_pick_migration_target(dlm, res);
2651         spin_unlock(&res->spinlock);
2652
2653         if (target == O2NM_MAX_NODES)
2654                 goto leave;
2655
2656         /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2657         spin_unlock(&dlm->spinlock);
2658         lock_dropped = 1;
2659         ret = dlm_migrate_lockres(dlm, res, target);
2660         if (ret)
2661                 mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2662                      dlm->name, res->lockname.len, res->lockname.name,
2663                      target, ret);
2664         spin_lock(&dlm->spinlock);
2665 leave:
2666         return lock_dropped;
2667 }
2668
2669 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2670 {
2671         int ret;
2672         spin_lock(&dlm->ast_lock);
2673         spin_lock(&lock->spinlock);
2674         ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2675         spin_unlock(&lock->spinlock);
2676         spin_unlock(&dlm->ast_lock);
2677         return ret;
2678 }
2679
2680 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2681                                      struct dlm_lock_resource *res,
2682                                      u8 mig_target)
2683 {
2684         int can_proceed;
2685         spin_lock(&res->spinlock);
2686         can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2687         spin_unlock(&res->spinlock);
2688
2689         /* target has died, so make the caller break out of the
2690          * wait_event, but caller must recheck the domain_map */
2691         spin_lock(&dlm->spinlock);
2692         if (!test_bit(mig_target, dlm->domain_map))
2693                 can_proceed = 1;
2694         spin_unlock(&dlm->spinlock);
2695         return can_proceed;
2696 }
2697
2698 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2699                                 struct dlm_lock_resource *res)
2700 {
2701         int ret;
2702         spin_lock(&res->spinlock);
2703         ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2704         spin_unlock(&res->spinlock);
2705         return ret;
2706 }
2707
2708
2709 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2710                                        struct dlm_lock_resource *res,
2711                                        u8 target)
2712 {
2713         int ret = 0;
2714
2715         mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2716                res->lockname.len, res->lockname.name, dlm->node_num,
2717                target);
2718         /* need to set MIGRATING flag on lockres.  this is done by
2719          * ensuring that all asts have been flushed for this lockres. */
2720         spin_lock(&res->spinlock);
2721         BUG_ON(res->migration_pending);
2722         res->migration_pending = 1;
2723         /* strategy is to reserve an extra ast then release
2724          * it below, letting the release do all of the work */
2725         __dlm_lockres_reserve_ast(res);
2726         spin_unlock(&res->spinlock);
2727
2728         /* now flush all the pending asts */
2729         dlm_kick_thread(dlm, res);
2730         /* before waiting on DIRTY, block processes which may
2731          * try to dirty the lockres before MIGRATING is set */
2732         spin_lock(&res->spinlock);
2733         BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2734         res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2735         spin_unlock(&res->spinlock);
2736         /* now wait on any pending asts and the DIRTY state */
2737         wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2738         dlm_lockres_release_ast(dlm, res);
2739
2740         mlog(0, "about to wait on migration_wq, dirty=%s\n",
2741                res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2742         /* if the extra ref we just put was the final one, this
2743          * will pass thru immediately.  otherwise, we need to wait
2744          * for the last ast to finish. */
2745 again:
2746         ret = wait_event_interruptible_timeout(dlm->migration_wq,
2747                    dlm_migration_can_proceed(dlm, res, target),
2748                    msecs_to_jiffies(1000));
2749         if (ret < 0) {
2750                 mlog(0, "woken again: migrating? %s, dead? %s\n",
2751                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2752                        test_bit(target, dlm->domain_map) ? "no":"yes");
2753         } else {
2754                 mlog(0, "all is well: migrating? %s, dead? %s\n",
2755                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2756                        test_bit(target, dlm->domain_map) ? "no":"yes");
2757         }
2758         if (!dlm_migration_can_proceed(dlm, res, target)) {
2759                 mlog(0, "trying again...\n");
2760                 goto again;
2761         }
2762
2763         ret = 0;
2764         /* did the target go down or die? */
2765         spin_lock(&dlm->spinlock);
2766         if (!test_bit(target, dlm->domain_map)) {
2767                 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2768                      target);
2769                 ret = -EHOSTDOWN;
2770         }
2771         spin_unlock(&dlm->spinlock);
2772
2773         /*
2774          * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2775          * another try; otherwise, we are sure the MIGRATING state is there,
2776          * drop the unneded state which blocked threads trying to DIRTY
2777          */
2778         spin_lock(&res->spinlock);
2779         BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2780         res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2781         if (!ret)
2782                 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2783         spin_unlock(&res->spinlock);
2784
2785         /*
2786          * at this point:
2787          *
2788          *   o the DLM_LOCK_RES_MIGRATING flag is set if target not down
2789          *   o there are no pending asts on this lockres
2790          *   o all processes trying to reserve an ast on this
2791          *     lockres must wait for the MIGRATING flag to clear
2792          */
2793         return ret;
2794 }
2795
2796 /* last step in the migration process.
2797  * original master calls this to free all of the dlm_lock
2798  * structures that used to be for other nodes. */
2799 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2800                                       struct dlm_lock_resource *res)
2801 {
2802         struct list_head *queue = &res->granted;
2803         int i, bit;
2804         struct dlm_lock *lock, *next;
2805
2806         assert_spin_locked(&res->spinlock);
2807
2808         BUG_ON(res->owner == dlm->node_num);
2809
2810         for (i=0; i<3; i++) {
2811                 list_for_each_entry_safe(lock, next, queue, list) {
2812                         if (lock->ml.node != dlm->node_num) {
2813                                 mlog(0, "putting lock for node %u\n",
2814                                      lock->ml.node);
2815                                 /* be extra careful */
2816                                 BUG_ON(!list_empty(&lock->ast_list));
2817                                 BUG_ON(!list_empty(&lock->bast_list));
2818                                 BUG_ON(lock->ast_pending);
2819                                 BUG_ON(lock->bast_pending);
2820                                 dlm_lockres_clear_refmap_bit(dlm, res,
2821                                                              lock->ml.node);
2822                                 list_del_init(&lock->list);
2823                                 dlm_lock_put(lock);
2824                                 /* In a normal unlock, we would have added a
2825                                  * DLM_UNLOCK_FREE_LOCK action. Force it. */
2826                                 dlm_lock_put(lock);
2827                         }
2828                 }
2829                 queue++;
2830         }
2831         bit = 0;
2832         while (1) {
2833                 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2834                 if (bit >= O2NM_MAX_NODES)
2835                         break;
2836                 /* do not clear the local node reference, if there is a
2837                  * process holding this, let it drop the ref itself */
2838                 if (bit != dlm->node_num) {
2839                         mlog(0, "%s:%.*s: node %u had a ref to this "
2840                              "migrating lockres, clearing\n", dlm->name,
2841                              res->lockname.len, res->lockname.name, bit);
2842                         dlm_lockres_clear_refmap_bit(dlm, res, bit);
2843                 }
2844                 bit++;
2845         }
2846 }
2847
2848 /*
2849  * Pick a node to migrate the lock resource to. This function selects a
2850  * potential target based first on the locks and then on refmap. It skips
2851  * nodes that are in the process of exiting the domain.
2852  */
2853 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2854                                     struct dlm_lock_resource *res)
2855 {
2856         enum dlm_lockres_list idx;
2857         struct list_head *queue = &res->granted;
2858         struct dlm_lock *lock;
2859         int noderef;
2860         u8 nodenum = O2NM_MAX_NODES;
2861
2862         assert_spin_locked(&dlm->spinlock);
2863         assert_spin_locked(&res->spinlock);
2864
2865         /* Go through all the locks */
2866         for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2867                 queue = dlm_list_idx_to_ptr(res, idx);
2868                 list_for_each_entry(lock, queue, list) {
2869                         if (lock->ml.node == dlm->node_num)
2870                                 continue;
2871                         if (test_bit(lock->ml.node, dlm->exit_domain_map))
2872                                 continue;
2873                         nodenum = lock->ml.node;
2874                         goto bail;
2875                 }
2876         }
2877
2878         /* Go thru the refmap */
2879         noderef = -1;
2880         while (1) {
2881                 noderef = find_next_bit(res->refmap, O2NM_MAX_NODES,
2882                                         noderef + 1);
2883                 if (noderef >= O2NM_MAX_NODES)
2884                         break;
2885                 if (noderef == dlm->node_num)
2886                         continue;
2887                 if (test_bit(noderef, dlm->exit_domain_map))
2888                         continue;
2889                 nodenum = noderef;
2890                 goto bail;
2891         }
2892
2893 bail:
2894         return nodenum;
2895 }
2896
2897 /* this is called by the new master once all lockres
2898  * data has been received */
2899 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2900                                   struct dlm_lock_resource *res,
2901                                   u8 master, u8 new_master,
2902                                   struct dlm_node_iter *iter)
2903 {
2904         struct dlm_migrate_request migrate;
2905         int ret, skip, status = 0;
2906         int nodenum;
2907
2908         memset(&migrate, 0, sizeof(migrate));
2909         migrate.namelen = res->lockname.len;
2910         memcpy(migrate.name, res->lockname.name, migrate.namelen);
2911         migrate.new_master = new_master;
2912         migrate.master = master;
2913
2914         ret = 0;
2915
2916         /* send message to all nodes, except the master and myself */
2917         while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2918                 if (nodenum == master ||
2919                     nodenum == new_master)
2920                         continue;
2921
2922                 /* We could race exit domain. If exited, skip. */
2923                 spin_lock(&dlm->spinlock);
2924                 skip = (!test_bit(nodenum, dlm->domain_map));
2925                 spin_unlock(&dlm->spinlock);
2926                 if (skip) {
2927                         clear_bit(nodenum, iter->node_map);
2928                         continue;
2929                 }
2930
2931                 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2932                                          &migrate, sizeof(migrate), nodenum,
2933                                          &status);
2934                 if (ret < 0) {
2935                         mlog(ML_ERROR, "%s: res %.*s, Error %d send "
2936                              "MIGRATE_REQUEST to node %u\n", dlm->name,
2937                              migrate.namelen, migrate.name, ret, nodenum);
2938                         if (!dlm_is_host_down(ret)) {
2939                                 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
2940                                 BUG();
2941                         }
2942                         clear_bit(nodenum, iter->node_map);
2943                         ret = 0;
2944                 } else if (status < 0) {
2945                         mlog(0, "migrate request (node %u) returned %d!\n",
2946                              nodenum, status);
2947                         ret = status;
2948                 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
2949                         /* during the migration request we short-circuited
2950                          * the mastery of the lockres.  make sure we have
2951                          * a mastery ref for nodenum */
2952                         mlog(0, "%s:%.*s: need ref for node %u\n",
2953                              dlm->name, res->lockname.len, res->lockname.name,
2954                              nodenum);
2955                         spin_lock(&res->spinlock);
2956                         dlm_lockres_set_refmap_bit(dlm, res, nodenum);
2957                         spin_unlock(&res->spinlock);
2958                 }
2959         }
2960
2961         if (ret < 0)
2962                 mlog_errno(ret);
2963
2964         mlog(0, "returning ret=%d\n", ret);
2965         return ret;
2966 }
2967
2968
2969 /* if there is an existing mle for this lockres, we now know who the master is.
2970  * (the one who sent us *this* message) we can clear it up right away.
2971  * since the process that put the mle on the list still has a reference to it,
2972  * we can unhash it now, set the master and wake the process.  as a result,
2973  * we will have no mle in the list to start with.  now we can add an mle for
2974  * the migration and this should be the only one found for those scanning the
2975  * list.  */
2976 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
2977                                 void **ret_data)
2978 {
2979         struct dlm_ctxt *dlm = data;
2980         struct dlm_lock_resource *res = NULL;
2981         struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
2982         struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
2983         const char *name;
2984         unsigned int namelen, hash;
2985         int ret = 0;
2986
2987         if (!dlm_grab(dlm))
2988                 return -EINVAL;
2989
2990         name = migrate->name;
2991         namelen = migrate->namelen;
2992         hash = dlm_lockid_hash(name, namelen);
2993
2994         /* preallocate.. if this fails, abort */
2995         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2996
2997         if (!mle) {
2998                 ret = -ENOMEM;
2999                 goto leave;
3000         }
3001
3002         /* check for pre-existing lock */
3003         spin_lock(&dlm->spinlock);
3004         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
3005         if (res) {
3006                 spin_lock(&res->spinlock);
3007                 if (res->state & DLM_LOCK_RES_RECOVERING) {
3008                         /* if all is working ok, this can only mean that we got
3009                         * a migrate request from a node that we now see as
3010                         * dead.  what can we do here?  drop it to the floor? */
3011                         spin_unlock(&res->spinlock);
3012                         mlog(ML_ERROR, "Got a migrate request, but the "
3013                              "lockres is marked as recovering!");
3014                         kmem_cache_free(dlm_mle_cache, mle);
3015                         ret = -EINVAL; /* need a better solution */
3016                         goto unlock;
3017                 }
3018                 res->state |= DLM_LOCK_RES_MIGRATING;
3019                 spin_unlock(&res->spinlock);
3020         }
3021
3022         spin_lock(&dlm->master_lock);
3023         /* ignore status.  only nonzero status would BUG. */
3024         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3025                                     name, namelen,
3026                                     migrate->new_master,
3027                                     migrate->master);
3028
3029         spin_unlock(&dlm->master_lock);
3030 unlock:
3031         spin_unlock(&dlm->spinlock);
3032
3033         if (oldmle) {
3034                 /* master is known, detach if not already detached */
3035                 dlm_mle_detach_hb_events(dlm, oldmle);
3036                 dlm_put_mle(oldmle);
3037         }
3038
3039         if (res)
3040                 dlm_lockres_put(res);
3041 leave:
3042         dlm_put(dlm);
3043         return ret;
3044 }
3045
3046 /* must be holding dlm->spinlock and dlm->master_lock
3047  * when adding a migration mle, we can clear any other mles
3048  * in the master list because we know with certainty that
3049  * the master is "master".  so we remove any old mle from
3050  * the list after setting it's master field, and then add
3051  * the new migration mle.  this way we can hold with the rule
3052  * of having only one mle for a given lock name at all times. */
3053 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3054                                  struct dlm_lock_resource *res,
3055                                  struct dlm_master_list_entry *mle,
3056                                  struct dlm_master_list_entry **oldmle,
3057                                  const char *name, unsigned int namelen,
3058                                  u8 new_master, u8 master)
3059 {
3060         int found;
3061         int ret = 0;
3062
3063         *oldmle = NULL;
3064
3065         assert_spin_locked(&dlm->spinlock);
3066         assert_spin_locked(&dlm->master_lock);
3067
3068         /* caller is responsible for any ref taken here on oldmle */
3069         found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3070         if (found) {
3071                 struct dlm_master_list_entry *tmp = *oldmle;
3072                 spin_lock(&tmp->spinlock);
3073                 if (tmp->type == DLM_MLE_MIGRATION) {
3074                         if (master == dlm->node_num) {
3075                                 /* ah another process raced me to it */
3076                                 mlog(0, "tried to migrate %.*s, but some "
3077                                      "process beat me to it\n",
3078                                      namelen, name);
3079                                 ret = -EEXIST;
3080                         } else {
3081                                 /* bad.  2 NODES are trying to migrate! */
3082                                 mlog(ML_ERROR, "migration error  mle: "
3083                                      "master=%u new_master=%u // request: "
3084                                      "master=%u new_master=%u // "
3085                                      "lockres=%.*s\n",
3086                                      tmp->master, tmp->new_master,
3087                                      master, new_master,
3088                                      namelen, name);
3089                                 BUG();
3090                         }
3091                 } else {
3092                         /* this is essentially what assert_master does */
3093                         tmp->master = master;
3094                         atomic_set(&tmp->woken, 1);
3095                         wake_up(&tmp->wq);
3096                         /* remove it so that only one mle will be found */
3097                         __dlm_unlink_mle(dlm, tmp);
3098                         __dlm_mle_detach_hb_events(dlm, tmp);
3099                         ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3100                         mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3101                             "telling master to get ref for cleared out mle "
3102                             "during migration\n", dlm->name, namelen, name,
3103                             master, new_master);
3104                 }
3105                 spin_unlock(&tmp->spinlock);
3106         }
3107
3108         /* now add a migration mle to the tail of the list */
3109         dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3110         mle->new_master = new_master;
3111         /* the new master will be sending an assert master for this.
3112          * at that point we will get the refmap reference */
3113         mle->master = master;
3114         /* do this for consistency with other mle types */
3115         set_bit(new_master, mle->maybe_map);
3116         __dlm_insert_mle(dlm, mle);
3117
3118         return ret;
3119 }
3120
3121 /*
3122  * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3123  */
3124 static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3125                                         struct dlm_master_list_entry *mle)
3126 {
3127         struct dlm_lock_resource *res;
3128
3129         /* Find the lockres associated to the mle and set its owner to UNK */
3130         res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3131                                    mle->mnamehash);
3132         if (res) {
3133                 spin_unlock(&dlm->master_lock);
3134
3135                 /* move lockres onto recovery list */
3136                 spin_lock(&res->spinlock);
3137                 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3138                 dlm_move_lockres_to_recovery_list(dlm, res);
3139                 spin_unlock(&res->spinlock);
3140                 dlm_lockres_put(res);
3141
3142                 /* about to get rid of mle, detach from heartbeat */
3143                 __dlm_mle_detach_hb_events(dlm, mle);
3144
3145                 /* dump the mle */
3146                 spin_lock(&dlm->master_lock);
3147                 __dlm_put_mle(mle);
3148                 spin_unlock(&dlm->master_lock);
3149         }
3150
3151         return res;
3152 }
3153
3154 static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3155                                     struct dlm_master_list_entry *mle)
3156 {
3157         __dlm_mle_detach_hb_events(dlm, mle);
3158
3159         spin_lock(&mle->spinlock);
3160         __dlm_unlink_mle(dlm, mle);
3161         atomic_set(&mle->woken, 1);
3162         spin_unlock(&mle->spinlock);
3163
3164         wake_up(&mle->wq);
3165 }
3166
3167 static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3168                                 struct dlm_master_list_entry *mle, u8 dead_node)
3169 {
3170         int bit;
3171
3172         BUG_ON(mle->type != DLM_MLE_BLOCK);
3173
3174         spin_lock(&mle->spinlock);
3175         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3176         if (bit != dead_node) {
3177                 mlog(0, "mle found, but dead node %u would not have been "
3178                      "master\n", dead_node);
3179                 spin_unlock(&mle->spinlock);
3180         } else {
3181                 /* Must drop the refcount by one since the assert_master will
3182                  * never arrive. This may result in the mle being unlinked and
3183                  * freed, but there may still be a process waiting in the
3184                  * dlmlock path which is fine. */
3185                 mlog(0, "node %u was expected master\n", dead_node);
3186                 atomic_set(&mle->woken, 1);
3187                 spin_unlock(&mle->spinlock);
3188                 wake_up(&mle->wq);
3189
3190                 /* Do not need events any longer, so detach from heartbeat */
3191                 __dlm_mle_detach_hb_events(dlm, mle);
3192                 __dlm_put_mle(mle);
3193         }
3194 }
3195
3196 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3197 {
3198         struct dlm_master_list_entry *mle;
3199         struct dlm_lock_resource *res;
3200         struct hlist_head *bucket;
3201         struct hlist_node *list;
3202         unsigned int i;
3203
3204         mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
3205 top:
3206         assert_spin_locked(&dlm->spinlock);
3207
3208         /* clean the master list */
3209         spin_lock(&dlm->master_lock);
3210         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3211                 bucket = dlm_master_hash(dlm, i);
3212                 hlist_for_each(list, bucket) {
3213                         mle = hlist_entry(list, struct dlm_master_list_entry,
3214                                           master_hash_node);
3215
3216                         BUG_ON(mle->type != DLM_MLE_BLOCK &&
3217                                mle->type != DLM_MLE_MASTER &&
3218                                mle->type != DLM_MLE_MIGRATION);
3219
3220                         /* MASTER mles are initiated locally. The waiting
3221                          * process will notice the node map change shortly.
3222                          * Let that happen as normal. */
3223                         if (mle->type == DLM_MLE_MASTER)
3224                                 continue;
3225
3226                         /* BLOCK mles are initiated by other nodes. Need to
3227                          * clean up if the dead node would have been the
3228                          * master. */
3229                         if (mle->type == DLM_MLE_BLOCK) {
3230                                 dlm_clean_block_mle(dlm, mle, dead_node);
3231                                 continue;
3232                         }
3233
3234                         /* Everything else is a MIGRATION mle */
3235
3236                         /* The rule for MIGRATION mles is that the master
3237                          * becomes UNKNOWN if *either* the original or the new
3238                          * master dies. All UNKNOWN lockres' are sent to
3239                          * whichever node becomes the recovery master. The new
3240                          * master is responsible for determining if there is
3241                          * still a master for this lockres, or if he needs to
3242                          * take over mastery. Either way, this node should
3243                          * expect another message to resolve this. */
3244
3245                         if (mle->master != dead_node &&
3246                             mle->new_master != dead_node)
3247                                 continue;
3248
3249                         /* If we have reached this point, this mle needs to be
3250                          * removed from the list and freed. */
3251                         dlm_clean_migration_mle(dlm, mle);
3252
3253                         mlog(0, "%s: node %u died during migration from "
3254                              "%u to %u!\n", dlm->name, dead_node, mle->master,
3255                              mle->new_master);
3256
3257                         /* If we find a lockres associated with the mle, we've
3258                          * hit this rare case that messes up our lock ordering.
3259                          * If so, we need to drop the master lock so that we can
3260                          * take the lockres lock, meaning that we will have to
3261                          * restart from the head of list. */
3262                         res = dlm_reset_mleres_owner(dlm, mle);
3263                         if (res)
3264                                 /* restart */
3265                                 goto top;
3266
3267                         /* This may be the last reference */
3268                         __dlm_put_mle(mle);
3269                 }
3270         }
3271         spin_unlock(&dlm->master_lock);
3272 }
3273
3274 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3275                          u8 old_master)
3276 {
3277         struct dlm_node_iter iter;
3278         int ret = 0;
3279
3280         spin_lock(&dlm->spinlock);
3281         dlm_node_iter_init(dlm->domain_map, &iter);
3282         clear_bit(old_master, iter.node_map);
3283         clear_bit(dlm->node_num, iter.node_map);
3284         spin_unlock(&dlm->spinlock);
3285
3286         /* ownership of the lockres is changing.  account for the
3287          * mastery reference here since old_master will briefly have
3288          * a reference after the migration completes */
3289         spin_lock(&res->spinlock);
3290         dlm_lockres_set_refmap_bit(dlm, res, old_master);
3291         spin_unlock(&res->spinlock);
3292
3293         mlog(0, "now time to do a migrate request to other nodes\n");
3294         ret = dlm_do_migrate_request(dlm, res, old_master,
3295                                      dlm->node_num, &iter);
3296         if (ret < 0) {
3297                 mlog_errno(ret);
3298                 goto leave;
3299         }
3300
3301         mlog(0, "doing assert master of %.*s to all except the original node\n",
3302              res->lockname.len, res->lockname.name);
3303         /* this call now finishes out the nodemap
3304          * even if one or more nodes die */
3305         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3306                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3307         if (ret < 0) {
3308                 /* no longer need to retry.  all living nodes contacted. */
3309                 mlog_errno(ret);
3310                 ret = 0;
3311         }
3312
3313         memset(iter.node_map, 0, sizeof(iter.node_map));
3314         set_bit(old_master, iter.node_map);
3315         mlog(0, "doing assert master of %.*s back to %u\n",
3316              res->lockname.len, res->lockname.name, old_master);
3317         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3318                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3319         if (ret < 0) {
3320                 mlog(0, "assert master to original master failed "
3321                      "with %d.\n", ret);
3322                 /* the only nonzero status here would be because of
3323                  * a dead original node.  we're done. */
3324                 ret = 0;
3325         }
3326
3327         /* all done, set the owner, clear the flag */
3328         spin_lock(&res->spinlock);
3329         dlm_set_lockres_owner(dlm, res, dlm->node_num);
3330         res->state &= ~DLM_LOCK_RES_MIGRATING;
3331         spin_unlock(&res->spinlock);
3332         /* re-dirty it on the new master */
3333         dlm_kick_thread(dlm, res);
3334         wake_up(&res->wq);
3335 leave:
3336         return ret;
3337 }
3338
3339 /*
3340  * LOCKRES AST REFCOUNT
3341  * this is integral to migration
3342  */
3343
3344 /* for future intent to call an ast, reserve one ahead of time.
3345  * this should be called only after waiting on the lockres
3346  * with dlm_wait_on_lockres, and while still holding the
3347  * spinlock after the call. */
3348 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3349 {
3350         assert_spin_locked(&res->spinlock);
3351         if (res->state & DLM_LOCK_RES_MIGRATING) {
3352                 __dlm_print_one_lock_resource(res);
3353         }
3354         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3355
3356         atomic_inc(&res->asts_reserved);
3357 }
3358
3359 /*
3360  * used to drop the reserved ast, either because it went unused,
3361  * or because the ast/bast was actually called.
3362  *
3363  * also, if there is a pending migration on this lockres,
3364  * and this was the last pending ast on the lockres,
3365  * atomically set the MIGRATING flag before we drop the lock.
3366  * this is how we ensure that migration can proceed with no
3367  * asts in progress.  note that it is ok if the state of the
3368  * queues is such that a lock should be granted in the future
3369  * or that a bast should be fired, because the new master will
3370  * shuffle the lists on this lockres as soon as it is migrated.
3371  */
3372 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3373                              struct dlm_lock_resource *res)
3374 {
3375         if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3376                 return;
3377
3378         if (!res->migration_pending) {
3379                 spin_unlock(&res->spinlock);
3380                 return;
3381         }
3382
3383         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3384         res->migration_pending = 0;
3385         res->state |= DLM_LOCK_RES_MIGRATING;
3386         spin_unlock(&res->spinlock);
3387         wake_up(&res->wq);
3388         wake_up(&dlm->migration_wq);
3389 }
3390
3391 void dlm_force_free_mles(struct dlm_ctxt *dlm)
3392 {
3393         int i;
3394         struct hlist_head *bucket;
3395         struct dlm_master_list_entry *mle;
3396         struct hlist_node *tmp, *list;
3397
3398         /*
3399          * We notified all other nodes that we are exiting the domain and
3400          * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3401          * around we force free them and wake any processes that are waiting
3402          * on the mles
3403          */
3404         spin_lock(&dlm->spinlock);
3405         spin_lock(&dlm->master_lock);
3406
3407         BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3408         BUG_ON((find_next_bit(dlm->domain_map, O2NM_MAX_NODES, 0) < O2NM_MAX_NODES));
3409
3410         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3411                 bucket = dlm_master_hash(dlm, i);
3412                 hlist_for_each_safe(list, tmp, bucket) {
3413                         mle = hlist_entry(list, struct dlm_master_list_entry,
3414                                           master_hash_node);
3415                         if (mle->type != DLM_MLE_BLOCK) {
3416                                 mlog(ML_ERROR, "bad mle: %p\n", mle);
3417                                 dlm_print_one_mle(mle);
3418                         }
3419                         atomic_set(&mle->woken, 1);
3420                         wake_up(&mle->wq);
3421
3422                         __dlm_unlink_mle(dlm, mle);
3423                         __dlm_mle_detach_hb_events(dlm, mle);
3424                         __dlm_put_mle(mle);
3425                 }
3426         }
3427         spin_unlock(&dlm->master_lock);
3428         spin_unlock(&dlm->spinlock);
3429 }