git.oblomov.eu Git - linux-2.6/blob - 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/utsname.h>
  34 #include <linux/init.h>
  35 #include <linux/sysctl.h>
  36 #include <linux/random.h>
  37 #include <linux/blkdev.h>
  38 #include <linux/socket.h>
  39 #include <linux/inet.h>
  40 #include <linux/spinlock.h>
  41 #include <linux/delay.h>
  42
  43
  44 #include "cluster/heartbeat.h"
  45 #include "cluster/nodemanager.h"
  46 #include "cluster/tcp.h"
  47
  48 #include "dlmapi.h"
  49 #include "dlmcommon.h"
  50 #include "dlmdomain.h"
  51 #include "dlmdebug.h"
  52
  53 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
  54 #include "cluster/masklog.h"
  55
  56 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
  57                               struct dlm_master_list_entry *mle,
  58                               struct o2nm_node *node,
  59                               int idx);
  60 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
  61                             struct dlm_master_list_entry *mle,
  62                             struct o2nm_node *node,
  63                             int idx);
  64
  65 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
  66 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
  67                                 struct dlm_lock_resource *res,
  68                                 void *nodemap, u32 flags);
  69 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
  70
  71 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
  72                                 struct dlm_master_list_entry *mle,
  73                                 const char *name,
  74                                 unsigned int namelen)
  75 {
  76         if (dlm != mle->dlm)
  77                 return 0;
  78
  79         if (namelen != mle->mnamelen ||
  80             memcmp(name, mle->mname, namelen) != 0)
  81                 return 0;
  82
  83         return 1;
  84 }
  85
  86 static struct kmem_cache *dlm_lockres_cache = NULL;
  87 static struct kmem_cache *dlm_lockname_cache = NULL;
  88 static struct kmem_cache *dlm_mle_cache = NULL;
  89
  90 static void dlm_mle_release(struct kref *kref);
  91 static void dlm_init_mle(struct dlm_master_list_entry *mle,
  92                         enum dlm_mle_type type,
  93                         struct dlm_ctxt *dlm,
  94                         struct dlm_lock_resource *res,
  95                         const char *name,
  96                         unsigned int namelen);
  97 static void dlm_put_mle(struct dlm_master_list_entry *mle);
  98 static void __dlm_put_mle(struct dlm_master_list_entry *mle);
  99 static int dlm_find_mle(struct dlm_ctxt *dlm,
 100                         struct dlm_master_list_entry **mle,
 101                         char *name, unsigned int namelen);
 102
 103 static int dlm_do_master_request(struct dlm_lock_resource *res,
 104                                  struct dlm_master_list_entry *mle, int to);
 105
 106
 107 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
 108                                      struct dlm_lock_resource *res,
 109                                      struct dlm_master_list_entry *mle,
 110                                      int *blocked);
 111 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
 112                                     struct dlm_lock_resource *res,
 113                                     struct dlm_master_list_entry *mle,
 114                                     int blocked);
 115 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
 116                                  struct dlm_lock_resource *res,
 117                                  struct dlm_master_list_entry *mle,
 118                                  struct dlm_master_list_entry **oldmle,
 119                                  const char *name, unsigned int namelen,
 120                                  u8 new_master, u8 master);
 121
 122 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
 123                                     struct dlm_lock_resource *res);
 124 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
 125                                       struct dlm_lock_resource *res);
 126 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
 127                                        struct dlm_lock_resource *res,
 128                                        u8 target);
 129 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
 130                                        struct dlm_lock_resource *res);
 131
 132
 133 int dlm_is_host_down(int errno)
 134 {
 135         switch (errno) {
 136                 case -EBADF:
 137                 case -ECONNREFUSED:
 138                 case -ENOTCONN:
 139                 case -ECONNRESET:
 140                 case -EPIPE:
 141                 case -EHOSTDOWN:
 142                 case -EHOSTUNREACH:
 143                 case -ETIMEDOUT:
 144                 case -ECONNABORTED:
 145                 case -ENETDOWN:
 146                 case -ENETUNREACH:
 147                 case -ENETRESET:
 148                 case -ESHUTDOWN:
 149                 case -ENOPROTOOPT:
 150                 case -EINVAL:   /* if returned from our tcp code,
 151                                    this means there is no socket */
 152                         return 1;
 153         }
 154         return 0;
 155 }
 156
 157
 158 /*
 159  * MASTER LIST FUNCTIONS
 160  */
 161
 162
 163 /*
 164  * regarding master list entries and heartbeat callbacks:
 165  *
 166  * in order to avoid sleeping and allocation that occurs in
 167  * heartbeat, master list entries are simply attached to the
 168  * dlm's established heartbeat callbacks.  the mle is attached
 169  * when it is created, and since the dlm->spinlock is held at
 170  * that time, any heartbeat event will be properly discovered
 171  * by the mle.  the mle needs to be detached from the
 172  * dlm->mle_hb_events list as soon as heartbeat events are no
 173  * longer useful to the mle, and before the mle is freed.
 174  *
 175  * as a general rule, heartbeat events are no longer needed by
 176  * the mle once an "answer" regarding the lock master has been
 177  * received.
 178  */
 179 static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
 180                                               struct dlm_master_list_entry *mle)
 181 {
 182         assert_spin_locked(&dlm->spinlock);
 183
 184         list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
 185 }
 186
 187
 188 static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 189                                               struct dlm_master_list_entry *mle)
 190 {
 191         if (!list_empty(&mle->hb_events))
 192                 list_del_init(&mle->hb_events);
 193 }
 194
 195
 196 static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 197                                             struct dlm_master_list_entry *mle)
 198 {
 199         spin_lock(&dlm->spinlock);
 200         __dlm_mle_detach_hb_events(dlm, mle);
 201         spin_unlock(&dlm->spinlock);
 202 }
 203
 204 static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
 205 {
 206         struct dlm_ctxt *dlm;
 207         dlm = mle->dlm;
 208
 209         assert_spin_locked(&dlm->spinlock);
 210         assert_spin_locked(&dlm->master_lock);
 211         mle->inuse++;
 212         kref_get(&mle->mle_refs);
 213 }
 214
 215 static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
 216 {
 217         struct dlm_ctxt *dlm;
 218         dlm = mle->dlm;
 219
 220         spin_lock(&dlm->spinlock);
 221         spin_lock(&dlm->master_lock);
 222         mle->inuse--;
 223         __dlm_put_mle(mle);
 224         spin_unlock(&dlm->master_lock);
 225         spin_unlock(&dlm->spinlock);
 226
 227 }
 228
 229 /* remove from list and free */
 230 static void __dlm_put_mle(struct dlm_master_list_entry *mle)
 231 {
 232         struct dlm_ctxt *dlm;
 233         dlm = mle->dlm;
 234
 235         assert_spin_locked(&dlm->spinlock);
 236         assert_spin_locked(&dlm->master_lock);
 237         if (!atomic_read(&mle->mle_refs.refcount)) {
 238                 /* this may or may not crash, but who cares.
 239                  * it's a BUG. */
 240                 mlog(ML_ERROR, "bad mle: %p\n", mle);
 241                 dlm_print_one_mle(mle);
 242                 BUG();
 243         } else
 244                 kref_put(&mle->mle_refs, dlm_mle_release);
 245 }
 246
 247
 248 /* must not have any spinlocks coming in */
 249 static void dlm_put_mle(struct dlm_master_list_entry *mle)
 250 {
 251         struct dlm_ctxt *dlm;
 252         dlm = mle->dlm;
 253
 254         spin_lock(&dlm->spinlock);
 255         spin_lock(&dlm->master_lock);
 256         __dlm_put_mle(mle);
 257         spin_unlock(&dlm->master_lock);
 258         spin_unlock(&dlm->spinlock);
 259 }
 260
 261 static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
 262 {
 263         kref_get(&mle->mle_refs);
 264 }
 265
 266 static void dlm_init_mle(struct dlm_master_list_entry *mle,
 267                         enum dlm_mle_type type,
 268                         struct dlm_ctxt *dlm,
 269                         struct dlm_lock_resource *res,
 270                         const char *name,
 271                         unsigned int namelen)
 272 {
 273         assert_spin_locked(&dlm->spinlock);
 274
 275         mle->dlm = dlm;
 276         mle->type = type;
 277         INIT_HLIST_NODE(&mle->master_hash_node);
 278         INIT_LIST_HEAD(&mle->hb_events);
 279         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
 280         spin_lock_init(&mle->spinlock);
 281         init_waitqueue_head(&mle->wq);
 282         atomic_set(&mle->woken, 0);
 283         kref_init(&mle->mle_refs);
 284         memset(mle->response_map, 0, sizeof(mle->response_map));
 285         mle->master = O2NM_MAX_NODES;
 286         mle->new_master = O2NM_MAX_NODES;
 287         mle->inuse = 0;
 288
 289         BUG_ON(mle->type != DLM_MLE_BLOCK &&
 290                mle->type != DLM_MLE_MASTER &&
 291                mle->type != DLM_MLE_MIGRATION);
 292
 293         if (mle->type == DLM_MLE_MASTER) {
 294                 BUG_ON(!res);
 295                 mle->mleres = res;
 296                 memcpy(mle->mname, res->lockname.name, res->lockname.len);
 297                 mle->mnamelen = res->lockname.len;
 298                 mle->mnamehash = res->lockname.hash;
 299         } else {
 300                 BUG_ON(!name);
 301                 mle->mleres = NULL;
 302                 memcpy(mle->mname, name, namelen);
 303                 mle->mnamelen = namelen;
 304                 mle->mnamehash = dlm_lockid_hash(name, namelen);
 305         }
 306
 307         atomic_inc(&dlm->mle_tot_count[mle->type]);
 308         atomic_inc(&dlm->mle_cur_count[mle->type]);
 309
 310         /* copy off the node_map and register hb callbacks on our copy */
 311         memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
 312         memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
 313         clear_bit(dlm->node_num, mle->vote_map);
 314         clear_bit(dlm->node_num, mle->node_map);
 315
 316         /* attach the mle to the domain node up/down events */
 317         __dlm_mle_attach_hb_events(dlm, mle);
 318 }
 319
 320 void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
 321 {
 322         assert_spin_locked(&dlm->spinlock);
 323         assert_spin_locked(&dlm->master_lock);
 324
 325         if (!hlist_unhashed(&mle->master_hash_node))
 326                 hlist_del_init(&mle->master_hash_node);
 327 }
 328
 329 void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
 330 {
 331         struct hlist_head *bucket;
 332
 333         assert_spin_locked(&dlm->master_lock);
 334
 335         bucket = dlm_master_hash(dlm, mle->mnamehash);
 336         hlist_add_head(&mle->master_hash_node, bucket);
 337 }
 338
 339 /* returns 1 if found, 0 if not */
 340 static int dlm_find_mle(struct dlm_ctxt *dlm,
 341                         struct dlm_master_list_entry **mle,
 342                         char *name, unsigned int namelen)
 343 {
 344         struct dlm_master_list_entry *tmpmle;
 345         struct hlist_head *bucket;
 346         struct hlist_node *list;
 347         unsigned int hash;
 348
 349         assert_spin_locked(&dlm->master_lock);
 350
 351         hash = dlm_lockid_hash(name, namelen);
 352         bucket = dlm_master_hash(dlm, hash);
 353         hlist_for_each(list, bucket) {
 354                 tmpmle = hlist_entry(list, struct dlm_master_list_entry,
 355                                      master_hash_node);
 356                 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
 357                         continue;
 358                 dlm_get_mle(tmpmle);
 359                 *mle = tmpmle;
 360                 return 1;
 361         }
 362         return 0;
 363 }
 364
 365 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
 366 {
 367         struct dlm_master_list_entry *mle;
 368
 369         assert_spin_locked(&dlm->spinlock);
 370
 371         list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
 372                 if (node_up)
 373                         dlm_mle_node_up(dlm, mle, NULL, idx);
 374                 else
 375                         dlm_mle_node_down(dlm, mle, NULL, idx);
 376         }
 377 }
 378
 379 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
 380                               struct dlm_master_list_entry *mle,
 381                               struct o2nm_node *node, int idx)
 382 {
 383         spin_lock(&mle->spinlock);
 384
 385         if (!test_bit(idx, mle->node_map))
 386                 mlog(0, "node %u already removed from nodemap!\n", idx);
 387         else
 388                 clear_bit(idx, mle->node_map);
 389
 390         spin_unlock(&mle->spinlock);
 391 }
 392
 393 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
 394                             struct dlm_master_list_entry *mle,
 395                             struct o2nm_node *node, int idx)
 396 {
 397         spin_lock(&mle->spinlock);
 398
 399         if (test_bit(idx, mle->node_map))
 400                 mlog(0, "node %u already in node map!\n", idx);
 401         else
 402                 set_bit(idx, mle->node_map);
 403
 404         spin_unlock(&mle->spinlock);
 405 }
 406
 407
 408 int dlm_init_mle_cache(void)
 409 {
 410         dlm_mle_cache = kmem_cache_create("o2dlm_mle",
 411                                           sizeof(struct dlm_master_list_entry),
 412                                           0, SLAB_HWCACHE_ALIGN,
 413                                           NULL);
 414         if (dlm_mle_cache == NULL)
 415                 return -ENOMEM;
 416         return 0;
 417 }
 418
 419 void dlm_destroy_mle_cache(void)
 420 {
 421         if (dlm_mle_cache)
 422                 kmem_cache_destroy(dlm_mle_cache);
 423 }
 424
 425 static void dlm_mle_release(struct kref *kref)
 426 {
 427         struct dlm_master_list_entry *mle;
 428         struct dlm_ctxt *dlm;
 429
 430         mlog_entry_void();
 431
 432         mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
 433         dlm = mle->dlm;
 434
 435         assert_spin_locked(&dlm->spinlock);
 436         assert_spin_locked(&dlm->master_lock);
 437
 438         mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
 439              mle->type);
 440
 441         /* remove from list if not already */
 442         __dlm_unlink_mle(dlm, mle);
 443
 444         /* detach the mle from the domain node up/down events */
 445         __dlm_mle_detach_hb_events(dlm, mle);
 446
 447         atomic_dec(&dlm->mle_cur_count[mle->type]);
 448
 449         /* NOTE: kfree under spinlock here.
 450          * if this is bad, we can move this to a freelist. */
 451         kmem_cache_free(dlm_mle_cache, mle);
 452 }
 453
 454
 455 /*
 456  * LOCK RESOURCE FUNCTIONS
 457  */
 458
 459 int dlm_init_master_caches(void)
 460 {
 461         dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
 462                                               sizeof(struct dlm_lock_resource),
 463                                               0, SLAB_HWCACHE_ALIGN, NULL);
 464         if (!dlm_lockres_cache)
 465                 goto bail;
 466
 467         dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
 468                                                DLM_LOCKID_NAME_MAX, 0,
 469                                                SLAB_HWCACHE_ALIGN, NULL);
 470         if (!dlm_lockname_cache)
 471                 goto bail;
 472
 473         return 0;
 474 bail:
 475         dlm_destroy_master_caches();
 476         return -ENOMEM;
 477 }
 478
 479 void dlm_destroy_master_caches(void)
 480 {
 481         if (dlm_lockname_cache)
 482                 kmem_cache_destroy(dlm_lockname_cache);
 483
 484         if (dlm_lockres_cache)
 485                 kmem_cache_destroy(dlm_lockres_cache);
 486 }
 487
 488 static void dlm_lockres_release(struct kref *kref)
 489 {
 490         struct dlm_lock_resource *res;
 491         struct dlm_ctxt *dlm;
 492
 493         res = container_of(kref, struct dlm_lock_resource, refs);
 494         dlm = res->dlm;
 495
 496         /* This should not happen -- all lockres' have a name
 497          * associated with them at init time. */
 498         BUG_ON(!res->lockname.name);
 499
 500         mlog(0, "destroying lockres %.*s\n", res->lockname.len,
 501              res->lockname.name);
 502
 503         spin_lock(&dlm->track_lock);
 504         if (!list_empty(&res->tracking))
 505                 list_del_init(&res->tracking);
 506         else {
 507                 mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n",
 508                      res->lockname.len, res->lockname.name);
 509                 dlm_print_one_lock_resource(res);
 510         }
 511         spin_unlock(&dlm->track_lock);
 512
 513         atomic_dec(&dlm->res_cur_count);
 514
 515         dlm_put(dlm);
 516
 517         if (!hlist_unhashed(&res->hash_node) ||
 518             !list_empty(&res->granted) ||
 519             !list_empty(&res->converting) ||
 520             !list_empty(&res->blocked) ||
 521             !list_empty(&res->dirty) ||
 522             !list_empty(&res->recovering) ||
 523             !list_empty(&res->purge)) {
 524                 mlog(ML_ERROR,
 525                      "Going to BUG for resource %.*s."
 526                      "  We're on a list! [%c%c%c%c%c%c%c]\n",
 527                      res->lockname.len, res->lockname.name,
 528                      !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
 529                      !list_empty(&res->granted) ? 'G' : ' ',
 530                      !list_empty(&res->converting) ? 'C' : ' ',
 531                      !list_empty(&res->blocked) ? 'B' : ' ',
 532                      !list_empty(&res->dirty) ? 'D' : ' ',
 533                      !list_empty(&res->recovering) ? 'R' : ' ',
 534                      !list_empty(&res->purge) ? 'P' : ' ');
 535
 536                 dlm_print_one_lock_resource(res);
 537         }
 538
 539         /* By the time we're ready to blow this guy away, we shouldn't
 540          * be on any lists. */
 541         BUG_ON(!hlist_unhashed(&res->hash_node));
 542         BUG_ON(!list_empty(&res->granted));
 543         BUG_ON(!list_empty(&res->converting));
 544         BUG_ON(!list_empty(&res->blocked));
 545         BUG_ON(!list_empty(&res->dirty));
 546         BUG_ON(!list_empty(&res->recovering));
 547         BUG_ON(!list_empty(&res->purge));
 548
 549         kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
 550
 551         kmem_cache_free(dlm_lockres_cache, res);
 552 }
 553
 554 void dlm_lockres_put(struct dlm_lock_resource *res)
 555 {
 556         kref_put(&res->refs, dlm_lockres_release);
 557 }
 558
 559 static void dlm_init_lockres(struct dlm_ctxt *dlm,
 560                              struct dlm_lock_resource *res,
 561                              const char *name, unsigned int namelen)
 562 {
 563         char *qname;
 564
 565         /* If we memset here, we lose our reference to the kmalloc'd
 566          * res->lockname.name, so be sure to init every field
 567          * correctly! */
 568
 569         qname = (char *) res->lockname.name;
 570         memcpy(qname, name, namelen);
 571
 572         res->lockname.len = namelen;
 573         res->lockname.hash = dlm_lockid_hash(name, namelen);
 574
 575         init_waitqueue_head(&res->wq);
 576         spin_lock_init(&res->spinlock);
 577         INIT_HLIST_NODE(&res->hash_node);
 578         INIT_LIST_HEAD(&res->granted);
 579         INIT_LIST_HEAD(&res->converting);
 580         INIT_LIST_HEAD(&res->blocked);
 581         INIT_LIST_HEAD(&res->dirty);
 582         INIT_LIST_HEAD(&res->recovering);
 583         INIT_LIST_HEAD(&res->purge);
 584         INIT_LIST_HEAD(&res->tracking);
 585         atomic_set(&res->asts_reserved, 0);
 586         res->migration_pending = 0;
 587         res->inflight_locks = 0;
 588
 589         /* put in dlm_lockres_release */
 590         dlm_grab(dlm);
 591         res->dlm = dlm;
 592
 593         kref_init(&res->refs);
 594
 595         atomic_inc(&dlm->res_tot_count);
 596         atomic_inc(&dlm->res_cur_count);
 597
 598         /* just for consistency */
 599         spin_lock(&res->spinlock);
 600         dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
 601         spin_unlock(&res->spinlock);
 602
 603         res->state = DLM_LOCK_RES_IN_PROGRESS;
 604
 605         res->last_used = 0;
 606
 607         spin_lock(&dlm->spinlock);
 608         list_add_tail(&res->tracking, &dlm->tracking_list);
 609         spin_unlock(&dlm->spinlock);
 610
 611         memset(res->lvb, 0, DLM_LVB_LEN);
 612         memset(res->refmap, 0, sizeof(res->refmap));
 613 }
 614
 615 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
 616                                    const char *name,
 617                                    unsigned int namelen)
 618 {
 619         struct dlm_lock_resource *res = NULL;
 620
 621         res = (struct dlm_lock_resource *)
 622                                 kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
 623         if (!res)
 624                 goto error;
 625
 626         res->lockname.name = (char *)
 627                                 kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
 628         if (!res->lockname.name)
 629                 goto error;
 630
 631         dlm_init_lockres(dlm, res, name, namelen);
 632         return res;
 633
 634 error:
 635         if (res && res->lockname.name)
 636                 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
 637
 638         if (res)
 639                 kmem_cache_free(dlm_lockres_cache, res);
 640         return NULL;
 641 }
 642
 643 void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
 644                                    struct dlm_lock_resource *res,
 645                                    int new_lockres,
 646                                    const char *file,
 647                                    int line)
 648 {
 649         if (!new_lockres)
 650                 assert_spin_locked(&res->spinlock);
 651
 652         if (!test_bit(dlm->node_num, res->refmap)) {
 653                 BUG_ON(res->inflight_locks != 0);
 654                 dlm_lockres_set_refmap_bit(dlm->node_num, res);
 655         }
 656         res->inflight_locks++;
 657         mlog(0, "%s:%.*s: inflight++: now %u\n",
 658              dlm->name, res->lockname.len, res->lockname.name,
 659              res->inflight_locks);
 660 }
 661
 662 void __dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
 663                                    struct dlm_lock_resource *res,
 664                                    const char *file,
 665                                    int line)
 666 {
 667         assert_spin_locked(&res->spinlock);
 668
 669         BUG_ON(res->inflight_locks == 0);
 670         res->inflight_locks--;
 671         mlog(0, "%s:%.*s: inflight--: now %u\n",
 672              dlm->name, res->lockname.len, res->lockname.name,
 673              res->inflight_locks);
 674         if (res->inflight_locks == 0)
 675                 dlm_lockres_clear_refmap_bit(dlm->node_num, res);
 676         wake_up(&res->wq);
 677 }
 678
 679 /*
 680  * lookup a lock resource by name.
 681  * may already exist in the hashtable.
 682  * lockid is null terminated
 683  *
 684  * if not, allocate enough for the lockres and for
 685  * the temporary structure used in doing the mastering.
 686  *
 687  * also, do a lookup in the dlm->master_list to see
 688  * if another node has begun mastering the same lock.
 689  * if so, there should be a block entry in there
 690  * for this name, and we should *not* attempt to master
 691  * the lock here.   need to wait around for that node
 692  * to assert_master (or die).
 693  *
 694  */
 695 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
 696                                           const char *lockid,
 697                                           int namelen,
 698                                           int flags)
 699 {
 700         struct dlm_lock_resource *tmpres=NULL, *res=NULL;
 701         struct dlm_master_list_entry *mle = NULL;
 702         struct dlm_master_list_entry *alloc_mle = NULL;
 703         int blocked = 0;
 704         int ret, nodenum;
 705         struct dlm_node_iter iter;
 706         unsigned int hash;
 707         int tries = 0;
 708         int bit, wait_on_recovery = 0;
 709         int drop_inflight_if_nonlocal = 0;
 710
 711         BUG_ON(!lockid);
 712
 713         hash = dlm_lockid_hash(lockid, namelen);
 714
 715         mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
 716
 717 lookup:
 718         spin_lock(&dlm->spinlock);
 719         tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
 720         if (tmpres) {
 721                 int dropping_ref = 0;
 722
 723                 spin_unlock(&dlm->spinlock);
 724
 725                 spin_lock(&tmpres->spinlock);
 726                 /* We wait for the other thread that is mastering the resource */
 727                 if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
 728                         __dlm_wait_on_lockres(tmpres);
 729                         BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
 730                 }
 731
 732                 if (tmpres->owner == dlm->node_num) {
 733                         BUG_ON(tmpres->state & DLM_LOCK_RES_DROPPING_REF);
 734                         dlm_lockres_grab_inflight_ref(dlm, tmpres);
 735                 } else if (tmpres->state & DLM_LOCK_RES_DROPPING_REF)
 736                         dropping_ref = 1;
 737                 spin_unlock(&tmpres->spinlock);
 738
 739                 /* wait until done messaging the master, drop our ref to allow
 740                  * the lockres to be purged, start over. */
 741                 if (dropping_ref) {
 742                         spin_lock(&tmpres->spinlock);
 743                         __dlm_wait_on_lockres_flags(tmpres, DLM_LOCK_RES_DROPPING_REF);
 744                         spin_unlock(&tmpres->spinlock);
 745                         dlm_lockres_put(tmpres);
 746                         tmpres = NULL;
 747                         goto lookup;
 748                 }
 749
 750                 mlog(0, "found in hash!\n");
 751                 if (res)
 752                         dlm_lockres_put(res);
 753                 res = tmpres;
 754                 goto leave;
 755         }
 756
 757         if (!res) {
 758                 spin_unlock(&dlm->spinlock);
 759                 mlog(0, "allocating a new resource\n");
 760                 /* nothing found and we need to allocate one. */
 761                 alloc_mle = (struct dlm_master_list_entry *)
 762                         kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
 763                 if (!alloc_mle)
 764                         goto leave;
 765                 res = dlm_new_lockres(dlm, lockid, namelen);
 766                 if (!res)
 767                         goto leave;
 768                 goto lookup;
 769         }
 770
 771         mlog(0, "no lockres found, allocated our own: %p\n", res);
 772
 773         if (flags & LKM_LOCAL) {
 774                 /* caller knows it's safe to assume it's not mastered elsewhere
 775                  * DONE!  return right away */
 776                 spin_lock(&res->spinlock);
 777                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
 778                 __dlm_insert_lockres(dlm, res);
 779                 dlm_lockres_grab_inflight_ref(dlm, res);
 780                 spin_unlock(&res->spinlock);
 781                 spin_unlock(&dlm->spinlock);
 782                 /* lockres still marked IN_PROGRESS */
 783                 goto wake_waiters;
 784         }
 785
 786         /* check master list to see if another node has started mastering it */
 787         spin_lock(&dlm->master_lock);
 788
 789         /* if we found a block, wait for lock to be mastered by another node */
 790         blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
 791         if (blocked) {
 792                 int mig;
 793                 if (mle->type == DLM_MLE_MASTER) {
 794                         mlog(ML_ERROR, "master entry for nonexistent lock!\n");
 795                         BUG();
 796                 }
 797                 mig = (mle->type == DLM_MLE_MIGRATION);
 798                 /* if there is a migration in progress, let the migration
 799                  * finish before continuing.  we can wait for the absence
 800                  * of the MIGRATION mle: either the migrate finished or
 801                  * one of the nodes died and the mle was cleaned up.
 802                  * if there is a BLOCK here, but it already has a master
 803                  * set, we are too late.  the master does not have a ref
 804                  * for us in the refmap.  detach the mle and drop it.
 805                  * either way, go back to the top and start over. */
 806                 if (mig || mle->master != O2NM_MAX_NODES) {
 807                         BUG_ON(mig && mle->master == dlm->node_num);
 808                         /* we arrived too late.  the master does not
 809                          * have a ref for us. retry. */
 810                         mlog(0, "%s:%.*s: late on %s\n",
 811                              dlm->name, namelen, lockid,
 812                              mig ?  "MIGRATION" : "BLOCK");
 813                         spin_unlock(&dlm->master_lock);
 814                         spin_unlock(&dlm->spinlock);
 815
 816                         /* master is known, detach */
 817                         if (!mig)
 818                                 dlm_mle_detach_hb_events(dlm, mle);
 819                         dlm_put_mle(mle);
 820                         mle = NULL;
 821                         /* this is lame, but we cant wait on either
 822                          * the mle or lockres waitqueue here */
 823                         if (mig)
 824                                 msleep(100);
 825                         goto lookup;
 826                 }
 827         } else {
 828                 /* go ahead and try to master lock on this node */
 829                 mle = alloc_mle;
 830                 /* make sure this does not get freed below */
 831                 alloc_mle = NULL;
 832                 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
 833                 set_bit(dlm->node_num, mle->maybe_map);
 834                 __dlm_insert_mle(dlm, mle);
 835
 836                 /* still holding the dlm spinlock, check the recovery map
 837                  * to see if there are any nodes that still need to be
 838                  * considered.  these will not appear in the mle nodemap
 839                  * but they might own this lockres.  wait on them. */
 840                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 841                 if (bit < O2NM_MAX_NODES) {
 842                         mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to "
 843                              "recover before lock mastery can begin\n",
 844                              dlm->name, namelen, (char *)lockid, bit);
 845                         wait_on_recovery = 1;
 846                 }
 847         }
 848
 849         /* at this point there is either a DLM_MLE_BLOCK or a
 850          * DLM_MLE_MASTER on the master list, so it's safe to add the
 851          * lockres to the hashtable.  anyone who finds the lock will
 852          * still have to wait on the IN_PROGRESS. */
 853
 854         /* finally add the lockres to its hash bucket */
 855         __dlm_insert_lockres(dlm, res);
 856         /* since this lockres is new it doesnt not require the spinlock */
 857         dlm_lockres_grab_inflight_ref_new(dlm, res);
 858
 859         /* if this node does not become the master make sure to drop
 860          * this inflight reference below */
 861         drop_inflight_if_nonlocal = 1;
 862
 863         /* get an extra ref on the mle in case this is a BLOCK
 864          * if so, the creator of the BLOCK may try to put the last
 865          * ref at this time in the assert master handler, so we
 866          * need an extra one to keep from a bad ptr deref. */
 867         dlm_get_mle_inuse(mle);
 868         spin_unlock(&dlm->master_lock);
 869         spin_unlock(&dlm->spinlock);
 870
 871 redo_request:
 872         while (wait_on_recovery) {
 873                 /* any cluster changes that occurred after dropping the
 874                  * dlm spinlock would be detectable be a change on the mle,
 875                  * so we only need to clear out the recovery map once. */
 876                 if (dlm_is_recovery_lock(lockid, namelen)) {
 877                         mlog(ML_NOTICE, "%s: recovery map is not empty, but "
 878                              "must master $RECOVERY lock now\n", dlm->name);
 879                         if (!dlm_pre_master_reco_lockres(dlm, res))
 880                                 wait_on_recovery = 0;
 881                         else {
 882                                 mlog(0, "%s: waiting 500ms for heartbeat state "
 883                                     "change\n", dlm->name);
 884                                 msleep(500);
 885                         }
 886                         continue;
 887                 }
 888
 889                 dlm_kick_recovery_thread(dlm);
 890                 msleep(1000);
 891                 dlm_wait_for_recovery(dlm);
 892
 893                 spin_lock(&dlm->spinlock);
 894                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 895                 if (bit < O2NM_MAX_NODES) {
 896                         mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to "
 897                              "recover before lock mastery can begin\n",
 898                              dlm->name, namelen, (char *)lockid, bit);
 899                         wait_on_recovery = 1;
 900                 } else
 901                         wait_on_recovery = 0;
 902                 spin_unlock(&dlm->spinlock);
 903
 904                 if (wait_on_recovery)
 905                         dlm_wait_for_node_recovery(dlm, bit, 10000);
 906         }
 907
 908         /* must wait for lock to be mastered elsewhere */
 909         if (blocked)
 910                 goto wait;
 911
 912         ret = -EINVAL;
 913         dlm_node_iter_init(mle->vote_map, &iter);
 914         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
 915                 ret = dlm_do_master_request(res, mle, nodenum);
 916                 if (ret < 0)
 917                         mlog_errno(ret);
 918                 if (mle->master != O2NM_MAX_NODES) {
 919                         /* found a master ! */
 920                         if (mle->master <= nodenum)
 921                                 break;
 922                         /* if our master request has not reached the master
 923                          * yet, keep going until it does.  this is how the
 924                          * master will know that asserts are needed back to
 925                          * the lower nodes. */
 926                         mlog(0, "%s:%.*s: requests only up to %u but master "
 927                              "is %u, keep going\n", dlm->name, namelen,
 928                              lockid, nodenum, mle->master);
 929                 }
 930         }
 931
 932 wait:
 933         /* keep going until the response map includes all nodes */
 934         ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
 935         if (ret < 0) {
 936                 wait_on_recovery = 1;
 937                 mlog(0, "%s:%.*s: node map changed, redo the "
 938                      "master request now, blocked=%d\n",
 939                      dlm->name, res->lockname.len,
 940                      res->lockname.name, blocked);
 941                 if (++tries > 20) {
 942                         mlog(ML_ERROR, "%s:%.*s: spinning on "
 943                              "dlm_wait_for_lock_mastery, blocked=%d\n",
 944                              dlm->name, res->lockname.len,
 945                              res->lockname.name, blocked);
 946                         dlm_print_one_lock_resource(res);
 947                         dlm_print_one_mle(mle);
 948                         tries = 0;
 949                 }
 950                 goto redo_request;
 951         }
 952
 953         mlog(0, "lockres mastered by %u\n", res->owner);
 954         /* make sure we never continue without this */
 955         BUG_ON(res->owner == O2NM_MAX_NODES);
 956
 957         /* master is known, detach if not already detached */
 958         dlm_mle_detach_hb_events(dlm, mle);
 959         dlm_put_mle(mle);
 960         /* put the extra ref */
 961         dlm_put_mle_inuse(mle);
 962
 963 wake_waiters:
 964         spin_lock(&res->spinlock);
 965         if (res->owner != dlm->node_num && drop_inflight_if_nonlocal)
 966                 dlm_lockres_drop_inflight_ref(dlm, res);
 967         res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
 968         spin_unlock(&res->spinlock);
 969         wake_up(&res->wq);
 970
 971 leave:
 972         /* need to free the unused mle */
 973         if (alloc_mle)
 974                 kmem_cache_free(dlm_mle_cache, alloc_mle);
 975
 976         return res;
 977 }
 978
 979
 980 #define DLM_MASTERY_TIMEOUT_MS   5000
 981
 982 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
 983                                      struct dlm_lock_resource *res,
 984                                      struct dlm_master_list_entry *mle,
 985                                      int *blocked)
 986 {
 987         u8 m;
 988         int ret, bit;
 989         int map_changed, voting_done;
 990         int assert, sleep;
 991
 992 recheck:
 993         ret = 0;
 994         assert = 0;
 995
 996         /* check if another node has already become the owner */
 997         spin_lock(&res->spinlock);
 998         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
 999                 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1000                      res->lockname.len, res->lockname.name, res->owner);
1001                 spin_unlock(&res->spinlock);
1002                 /* this will cause the master to re-assert across
1003                  * the whole cluster, freeing up mles */
1004                 if (res->owner != dlm->node_num) {
1005                         ret = dlm_do_master_request(res, mle, res->owner);
1006                         if (ret < 0) {
1007                                 /* give recovery a chance to run */
1008                                 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1009                                 msleep(500);
1010                                 goto recheck;
1011                         }
1012                 }
1013                 ret = 0;
1014                 goto leave;
1015         }
1016         spin_unlock(&res->spinlock);
1017
1018         spin_lock(&mle->spinlock);
1019         m = mle->master;
1020         map_changed = (memcmp(mle->vote_map, mle->node_map,
1021                               sizeof(mle->vote_map)) != 0);
1022         voting_done = (memcmp(mle->vote_map, mle->response_map,
1023                              sizeof(mle->vote_map)) == 0);
1024
1025         /* restart if we hit any errors */
1026         if (map_changed) {
1027                 int b;
1028                 mlog(0, "%s: %.*s: node map changed, restarting\n",
1029                      dlm->name, res->lockname.len, res->lockname.name);
1030                 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1031                 b = (mle->type == DLM_MLE_BLOCK);
1032                 if ((*blocked && !b) || (!*blocked && b)) {
1033                         mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1034                              dlm->name, res->lockname.len, res->lockname.name,
1035                              *blocked, b);
1036                         *blocked = b;
1037                 }
1038                 spin_unlock(&mle->spinlock);
1039                 if (ret < 0) {
1040                         mlog_errno(ret);
1041                         goto leave;
1042                 }
1043                 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1044                      "rechecking now\n", dlm->name, res->lockname.len,
1045                      res->lockname.name);
1046                 goto recheck;
1047         } else {
1048                 if (!voting_done) {
1049                         mlog(0, "map not changed and voting not done "
1050                              "for %s:%.*s\n", dlm->name, res->lockname.len,
1051                              res->lockname.name);
1052                 }
1053         }
1054
1055         if (m != O2NM_MAX_NODES) {
1056                 /* another node has done an assert!
1057                  * all done! */
1058                 sleep = 0;
1059         } else {
1060                 sleep = 1;
1061                 /* have all nodes responded? */
1062                 if (voting_done && !*blocked) {
1063                         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1064                         if (dlm->node_num <= bit) {
1065                                 /* my node number is lowest.
1066                                  * now tell other nodes that I am
1067                                  * mastering this. */
1068                                 mle->master = dlm->node_num;
1069                                 /* ref was grabbed in get_lock_resource
1070                                  * will be dropped in dlmlock_master */
1071                                 assert = 1;
1072                                 sleep = 0;
1073                         }
1074                         /* if voting is done, but we have not received
1075                          * an assert master yet, we must sleep */
1076                 }
1077         }
1078
1079         spin_unlock(&mle->spinlock);
1080
1081         /* sleep if we haven't finished voting yet */
1082         if (sleep) {
1083                 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1084
1085                 /*
1086                 if (atomic_read(&mle->mle_refs.refcount) < 2)
1087                         mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1088                         atomic_read(&mle->mle_refs.refcount),
1089                         res->lockname.len, res->lockname.name);
1090                 */
1091                 atomic_set(&mle->woken, 0);
1092                 (void)wait_event_timeout(mle->wq,
1093                                          (atomic_read(&mle->woken) == 1),
1094                                          timeo);
1095                 if (res->owner == O2NM_MAX_NODES) {
1096                         mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1097                              res->lockname.len, res->lockname.name);
1098                         goto recheck;
1099                 }
1100                 mlog(0, "done waiting, master is %u\n", res->owner);
1101                 ret = 0;
1102                 goto leave;
1103         }
1104
1105         ret = 0;   /* done */
1106         if (assert) {
1107                 m = dlm->node_num;
1108                 mlog(0, "about to master %.*s here, this=%u\n",
1109                      res->lockname.len, res->lockname.name, m);
1110                 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1111                 if (ret) {
1112                         /* This is a failure in the network path,
1113                          * not in the response to the assert_master
1114                          * (any nonzero response is a BUG on this node).
1115                          * Most likely a socket just got disconnected
1116                          * due to node death. */
1117                         mlog_errno(ret);
1118                 }
1119                 /* no longer need to restart lock mastery.
1120                  * all living nodes have been contacted. */
1121                 ret = 0;
1122         }
1123
1124         /* set the lockres owner */
1125         spin_lock(&res->spinlock);
1126         /* mastery reference obtained either during
1127          * assert_master_handler or in get_lock_resource */
1128         dlm_change_lockres_owner(dlm, res, m);
1129         spin_unlock(&res->spinlock);
1130
1131 leave:
1132         return ret;
1133 }
1134
1135 struct dlm_bitmap_diff_iter
1136 {
1137         int curnode;
1138         unsigned long *orig_bm;
1139         unsigned long *cur_bm;
1140         unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1141 };
1142
1143 enum dlm_node_state_change
1144 {
1145         NODE_DOWN = -1,
1146         NODE_NO_CHANGE = 0,
1147         NODE_UP
1148 };
1149
1150 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1151                                       unsigned long *orig_bm,
1152                                       unsigned long *cur_bm)
1153 {
1154         unsigned long p1, p2;
1155         int i;
1156
1157         iter->curnode = -1;
1158         iter->orig_bm = orig_bm;
1159         iter->cur_bm = cur_bm;
1160
1161         for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1162                 p1 = *(iter->orig_bm + i);
1163                 p2 = *(iter->cur_bm + i);
1164                 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1165         }
1166 }
1167
1168 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1169                                      enum dlm_node_state_change *state)
1170 {
1171         int bit;
1172
1173         if (iter->curnode >= O2NM_MAX_NODES)
1174                 return -ENOENT;
1175
1176         bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1177                             iter->curnode+1);
1178         if (bit >= O2NM_MAX_NODES) {
1179                 iter->curnode = O2NM_MAX_NODES;
1180                 return -ENOENT;
1181         }
1182
1183         /* if it was there in the original then this node died */
1184         if (test_bit(bit, iter->orig_bm))
1185                 *state = NODE_DOWN;
1186         else
1187                 *state = NODE_UP;
1188
1189         iter->curnode = bit;
1190         return bit;
1191 }
1192
1193
1194 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1195                                     struct dlm_lock_resource *res,
1196                                     struct dlm_master_list_entry *mle,
1197                                     int blocked)
1198 {
1199         struct dlm_bitmap_diff_iter bdi;
1200         enum dlm_node_state_change sc;
1201         int node;
1202         int ret = 0;
1203
1204         mlog(0, "something happened such that the "
1205              "master process may need to be restarted!\n");
1206
1207         assert_spin_locked(&mle->spinlock);
1208
1209         dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1210         node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1211         while (node >= 0) {
1212                 if (sc == NODE_UP) {
1213                         /* a node came up.  clear any old vote from
1214                          * the response map and set it in the vote map
1215                          * then restart the mastery. */
1216                         mlog(ML_NOTICE, "node %d up while restarting\n", node);
1217
1218                         /* redo the master request, but only for the new node */
1219                         mlog(0, "sending request to new node\n");
1220                         clear_bit(node, mle->response_map);
1221                         set_bit(node, mle->vote_map);
1222                 } else {
1223                         mlog(ML_ERROR, "node down! %d\n", node);
1224                         if (blocked) {
1225                                 int lowest = find_next_bit(mle->maybe_map,
1226                                                        O2NM_MAX_NODES, 0);
1227
1228                                 /* act like it was never there */
1229                                 clear_bit(node, mle->maybe_map);
1230
1231                                 if (node == lowest) {
1232                                         mlog(0, "expected master %u died"
1233                                             " while this node was blocked "
1234                                             "waiting on it!\n", node);
1235                                         lowest = find_next_bit(mle->maybe_map,
1236                                                         O2NM_MAX_NODES,
1237                                                         lowest+1);
1238                                         if (lowest < O2NM_MAX_NODES) {
1239                                                 mlog(0, "%s:%.*s:still "
1240                                                      "blocked. waiting on %u "
1241                                                      "now\n", dlm->name,
1242                                                      res->lockname.len,
1243                                                      res->lockname.name,
1244                                                      lowest);
1245                                         } else {
1246                                                 /* mle is an MLE_BLOCK, but
1247                                                  * there is now nothing left to
1248                                                  * block on.  we need to return
1249                                                  * all the way back out and try
1250                                                  * again with an MLE_MASTER.
1251                                                  * dlm_do_local_recovery_cleanup
1252                                                  * has already run, so the mle
1253                                                  * refcount is ok */
1254                                                 mlog(0, "%s:%.*s: no "
1255                                                      "longer blocking. try to "
1256                                                      "master this here\n",
1257                                                      dlm->name,
1258                                                      res->lockname.len,
1259                                                      res->lockname.name);
1260                                                 mle->type = DLM_MLE_MASTER;
1261                                                 mle->mleres = res;
1262                                         }
1263                                 }
1264                         }
1265
1266                         /* now blank out everything, as if we had never
1267                          * contacted anyone */
1268                         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1269                         memset(mle->response_map, 0, sizeof(mle->response_map));
1270                         /* reset the vote_map to the current node_map */
1271                         memcpy(mle->vote_map, mle->node_map,
1272                                sizeof(mle->node_map));
1273                         /* put myself into the maybe map */
1274                         if (mle->type != DLM_MLE_BLOCK)
1275                                 set_bit(dlm->node_num, mle->maybe_map);
1276                 }
1277                 ret = -EAGAIN;
1278                 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1279         }
1280         return ret;
1281 }
1282
1283
1284 /*
1285  * DLM_MASTER_REQUEST_MSG
1286  *
1287  * returns: 0 on success,
1288  *          -errno on a network error
1289  *
1290  * on error, the caller should assume the target node is "dead"
1291  *
1292  */
1293
1294 static int dlm_do_master_request(struct dlm_lock_resource *res,
1295                                  struct dlm_master_list_entry *mle, int to)
1296 {
1297         struct dlm_ctxt *dlm = mle->dlm;
1298         struct dlm_master_request request;
1299         int ret, response=0, resend;
1300
1301         memset(&request, 0, sizeof(request));
1302         request.node_idx = dlm->node_num;
1303
1304         BUG_ON(mle->type == DLM_MLE_MIGRATION);
1305
1306         request.namelen = (u8)mle->mnamelen;
1307         memcpy(request.name, mle->mname, request.namelen);
1308
1309 again:
1310         ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1311                                  sizeof(request), to, &response);
1312         if (ret < 0)  {
1313                 if (ret == -ESRCH) {
1314                         /* should never happen */
1315                         mlog(ML_ERROR, "TCP stack not ready!\n");
1316                         BUG();
1317                 } else if (ret == -EINVAL) {
1318                         mlog(ML_ERROR, "bad args passed to o2net!\n");
1319                         BUG();
1320                 } else if (ret == -ENOMEM) {
1321                         mlog(ML_ERROR, "out of memory while trying to send "
1322                              "network message!  retrying\n");
1323                         /* this is totally crude */
1324                         msleep(50);
1325                         goto again;
1326                 } else if (!dlm_is_host_down(ret)) {
1327                         /* not a network error. bad. */
1328                         mlog_errno(ret);
1329                         mlog(ML_ERROR, "unhandled error!");
1330                         BUG();
1331                 }
1332                 /* all other errors should be network errors,
1333                  * and likely indicate node death */
1334                 mlog(ML_ERROR, "link to %d went down!\n", to);
1335                 goto out;
1336         }
1337
1338         ret = 0;
1339         resend = 0;
1340         spin_lock(&mle->spinlock);
1341         switch (response) {
1342                 case DLM_MASTER_RESP_YES:
1343                         set_bit(to, mle->response_map);
1344                         mlog(0, "node %u is the master, response=YES\n", to);
1345                         mlog(0, "%s:%.*s: master node %u now knows I have a "
1346                              "reference\n", dlm->name, res->lockname.len,
1347                              res->lockname.name, to);
1348                         mle->master = to;
1349                         break;
1350                 case DLM_MASTER_RESP_NO:
1351                         mlog(0, "node %u not master, response=NO\n", to);
1352                         set_bit(to, mle->response_map);
1353                         break;
1354                 case DLM_MASTER_RESP_MAYBE:
1355                         mlog(0, "node %u not master, response=MAYBE\n", to);
1356                         set_bit(to, mle->response_map);
1357                         set_bit(to, mle->maybe_map);
1358                         break;
1359                 case DLM_MASTER_RESP_ERROR:
1360                         mlog(0, "node %u hit an error, resending\n", to);
1361                         resend = 1;
1362                         response = 0;
1363                         break;
1364                 default:
1365                         mlog(ML_ERROR, "bad response! %u\n", response);
1366                         BUG();
1367         }
1368         spin_unlock(&mle->spinlock);
1369         if (resend) {
1370                 /* this is also totally crude */
1371                 msleep(50);
1372                 goto again;
1373         }
1374
1375 out:
1376         return ret;
1377 }
1378
1379 /*
1380  * locks that can be taken here:
1381  * dlm->spinlock
1382  * res->spinlock
1383  * mle->spinlock
1384  * dlm->master_list
1385  *
1386  * if possible, TRIM THIS DOWN!!!
1387  */
1388 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1389                                void **ret_data)
1390 {
1391         u8 response = DLM_MASTER_RESP_MAYBE;
1392         struct dlm_ctxt *dlm = data;
1393         struct dlm_lock_resource *res = NULL;
1394         struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1395         struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1396         char *name;
1397         unsigned int namelen, hash;
1398         int found, ret;
1399         int set_maybe;
1400         int dispatch_assert = 0;
1401
1402         if (!dlm_grab(dlm))
1403                 return DLM_MASTER_RESP_NO;
1404
1405         if (!dlm_domain_fully_joined(dlm)) {
1406                 response = DLM_MASTER_RESP_NO;
1407                 goto send_response;
1408         }
1409
1410         name = request->name;
1411         namelen = request->namelen;
1412         hash = dlm_lockid_hash(name, namelen);
1413
1414         if (namelen > DLM_LOCKID_NAME_MAX) {
1415                 response = DLM_IVBUFLEN;
1416                 goto send_response;
1417         }
1418
1419 way_up_top:
1420         spin_lock(&dlm->spinlock);
1421         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1422         if (res) {
1423                 spin_unlock(&dlm->spinlock);
1424
1425                 /* take care of the easy cases up front */
1426                 spin_lock(&res->spinlock);
1427                 if (res->state & (DLM_LOCK_RES_RECOVERING|
1428                                   DLM_LOCK_RES_MIGRATING)) {
1429                         spin_unlock(&res->spinlock);
1430                         mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1431                              "being recovered/migrated\n");
1432                         response = DLM_MASTER_RESP_ERROR;
1433                         if (mle)
1434                                 kmem_cache_free(dlm_mle_cache, mle);
1435                         goto send_response;
1436                 }
1437
1438                 if (res->owner == dlm->node_num) {
1439                         mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1440                              dlm->name, namelen, name, request->node_idx);
1441                         dlm_lockres_set_refmap_bit(request->node_idx, res);
1442                         spin_unlock(&res->spinlock);
1443                         response = DLM_MASTER_RESP_YES;
1444                         if (mle)
1445                                 kmem_cache_free(dlm_mle_cache, mle);
1446
1447                         /* this node is the owner.
1448                          * there is some extra work that needs to
1449                          * happen now.  the requesting node has
1450                          * caused all nodes up to this one to
1451                          * create mles.  this node now needs to
1452                          * go back and clean those up. */
1453                         dispatch_assert = 1;
1454                         goto send_response;
1455                 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1456                         spin_unlock(&res->spinlock);
1457                         // mlog(0, "node %u is the master\n", res->owner);
1458                         response = DLM_MASTER_RESP_NO;
1459                         if (mle)
1460                                 kmem_cache_free(dlm_mle_cache, mle);
1461                         goto send_response;
1462                 }
1463
1464                 /* ok, there is no owner.  either this node is
1465                  * being blocked, or it is actively trying to
1466                  * master this lock. */
1467                 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1468                         mlog(ML_ERROR, "lock with no owner should be "
1469                              "in-progress!\n");
1470                         BUG();
1471                 }
1472
1473                 // mlog(0, "lockres is in progress...\n");
1474                 spin_lock(&dlm->master_lock);
1475                 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1476                 if (!found) {
1477                         mlog(ML_ERROR, "no mle found for this lock!\n");
1478                         BUG();
1479                 }
1480                 set_maybe = 1;
1481                 spin_lock(&tmpmle->spinlock);
1482                 if (tmpmle->type == DLM_MLE_BLOCK) {
1483                         // mlog(0, "this node is waiting for "
1484                         // "lockres to be mastered\n");
1485                         response = DLM_MASTER_RESP_NO;
1486                 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1487                         mlog(0, "node %u is master, but trying to migrate to "
1488                              "node %u.\n", tmpmle->master, tmpmle->new_master);
1489                         if (tmpmle->master == dlm->node_num) {
1490                                 mlog(ML_ERROR, "no owner on lockres, but this "
1491                                      "node is trying to migrate it to %u?!\n",
1492                                      tmpmle->new_master);
1493                                 BUG();
1494                         } else {
1495                                 /* the real master can respond on its own */
1496                                 response = DLM_MASTER_RESP_NO;
1497                         }
1498                 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1499                         set_maybe = 0;
1500                         if (tmpmle->master == dlm->node_num) {
1501                                 response = DLM_MASTER_RESP_YES;
1502                                 /* this node will be the owner.
1503                                  * go back and clean the mles on any
1504                                  * other nodes */
1505                                 dispatch_assert = 1;
1506                                 dlm_lockres_set_refmap_bit(request->node_idx, res);
1507                                 mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1508                                      dlm->name, namelen, name,
1509                                      request->node_idx);
1510                         } else
1511                                 response = DLM_MASTER_RESP_NO;
1512                 } else {
1513                         // mlog(0, "this node is attempting to "
1514                         // "master lockres\n");
1515                         response = DLM_MASTER_RESP_MAYBE;
1516                 }
1517                 if (set_maybe)
1518                         set_bit(request->node_idx, tmpmle->maybe_map);
1519                 spin_unlock(&tmpmle->spinlock);
1520
1521                 spin_unlock(&dlm->master_lock);
1522                 spin_unlock(&res->spinlock);
1523
1524                 /* keep the mle attached to heartbeat events */
1525                 dlm_put_mle(tmpmle);
1526                 if (mle)
1527                         kmem_cache_free(dlm_mle_cache, mle);
1528                 goto send_response;
1529         }
1530
1531         /*
1532          * lockres doesn't exist on this node
1533          * if there is an MLE_BLOCK, return NO
1534          * if there is an MLE_MASTER, return MAYBE
1535          * otherwise, add an MLE_BLOCK, return NO
1536          */
1537         spin_lock(&dlm->master_lock);
1538         found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1539         if (!found) {
1540                 /* this lockid has never been seen on this node yet */
1541                 // mlog(0, "no mle found\n");
1542                 if (!mle) {
1543                         spin_unlock(&dlm->master_lock);
1544                         spin_unlock(&dlm->spinlock);
1545
1546                         mle = (struct dlm_master_list_entry *)
1547                                 kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1548                         if (!mle) {
1549                                 response = DLM_MASTER_RESP_ERROR;
1550                                 mlog_errno(-ENOMEM);
1551                                 goto send_response;
1552                         }
1553                         goto way_up_top;
1554                 }
1555
1556                 // mlog(0, "this is second time thru, already allocated, "
1557                 // "add the block.\n");
1558                 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1559                 set_bit(request->node_idx, mle->maybe_map);
1560                 __dlm_insert_mle(dlm, mle);
1561                 response = DLM_MASTER_RESP_NO;
1562         } else {
1563                 // mlog(0, "mle was found\n");
1564                 set_maybe = 1;
1565                 spin_lock(&tmpmle->spinlock);
1566                 if (tmpmle->master == dlm->node_num) {
1567                         mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1568                         BUG();
1569                 }
1570                 if (tmpmle->type == DLM_MLE_BLOCK)
1571                         response = DLM_MASTER_RESP_NO;
1572                 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1573                         mlog(0, "migration mle was found (%u->%u)\n",
1574                              tmpmle->master, tmpmle->new_master);
1575                         /* real master can respond on its own */
1576                         response = DLM_MASTER_RESP_NO;
1577                 } else
1578                         response = DLM_MASTER_RESP_MAYBE;
1579                 if (set_maybe)
1580                         set_bit(request->node_idx, tmpmle->maybe_map);
1581                 spin_unlock(&tmpmle->spinlock);
1582         }
1583         spin_unlock(&dlm->master_lock);
1584         spin_unlock(&dlm->spinlock);
1585
1586         if (found) {
1587                 /* keep the mle attached to heartbeat events */
1588                 dlm_put_mle(tmpmle);
1589         }
1590 send_response:
1591         /*
1592          * __dlm_lookup_lockres() grabbed a reference to this lockres.
1593          * The reference is released by dlm_assert_master_worker() under
1594          * the call to dlm_dispatch_assert_master().  If
1595          * dlm_assert_master_worker() isn't called, we drop it here.
1596          */
1597         if (dispatch_assert) {
1598                 if (response != DLM_MASTER_RESP_YES)
1599                         mlog(ML_ERROR, "invalid response %d\n", response);
1600                 if (!res) {
1601                         mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1602                         BUG();
1603                 }
1604                 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1605                              dlm->node_num, res->lockname.len, res->lockname.name);
1606                 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1607                                                  DLM_ASSERT_MASTER_MLE_CLEANUP);
1608                 if (ret < 0) {
1609                         mlog(ML_ERROR, "failed to dispatch assert master work\n");
1610                         response = DLM_MASTER_RESP_ERROR;
1611                         dlm_lockres_put(res);
1612                 }
1613         } else {
1614                 if (res)
1615                         dlm_lockres_put(res);
1616         }
1617
1618         dlm_put(dlm);
1619         return response;
1620 }
1621
1622 /*
1623  * DLM_ASSERT_MASTER_MSG
1624  */
1625
1626
1627 /*
1628  * NOTE: this can be used for debugging
1629  * can periodically run all locks owned by this node
1630  * and re-assert across the cluster...
1631  */
1632 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1633                                 struct dlm_lock_resource *res,
1634                                 void *nodemap, u32 flags)
1635 {
1636         struct dlm_assert_master assert;
1637         int to, tmpret;
1638         struct dlm_node_iter iter;
1639         int ret = 0;
1640         int reassert;
1641         const char *lockname = res->lockname.name;
1642         unsigned int namelen = res->lockname.len;
1643
1644         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1645
1646         spin_lock(&res->spinlock);
1647         res->state |= DLM_LOCK_RES_SETREF_INPROG;
1648         spin_unlock(&res->spinlock);
1649
1650 again:
1651         reassert = 0;
1652
1653         /* note that if this nodemap is empty, it returns 0 */
1654         dlm_node_iter_init(nodemap, &iter);
1655         while ((to = dlm_node_iter_next(&iter)) >= 0) {
1656                 int r = 0;
1657                 struct dlm_master_list_entry *mle = NULL;
1658
1659                 mlog(0, "sending assert master to %d (%.*s)\n", to,
1660                      namelen, lockname);
1661                 memset(&assert, 0, sizeof(assert));
1662                 assert.node_idx = dlm->node_num;
1663                 assert.namelen = namelen;
1664                 memcpy(assert.name, lockname, namelen);
1665                 assert.flags = cpu_to_be32(flags);
1666
1667                 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1668                                             &assert, sizeof(assert), to, &r);
1669                 if (tmpret < 0) {
1670                         mlog(0, "assert_master returned %d!\n", tmpret);
1671                         if (!dlm_is_host_down(tmpret)) {
1672                                 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1673                                 BUG();
1674                         }
1675                         /* a node died.  finish out the rest of the nodes. */
1676                         mlog(0, "link to %d went down!\n", to);
1677                         /* any nonzero status return will do */
1678                         ret = tmpret;
1679                         r = 0;
1680                 } else if (r < 0) {
1681                         /* ok, something horribly messed.  kill thyself. */
1682                         mlog(ML_ERROR,"during assert master of %.*s to %u, "
1683                              "got %d.\n", namelen, lockname, to, r);
1684                         spin_lock(&dlm->spinlock);
1685                         spin_lock(&dlm->master_lock);
1686                         if (dlm_find_mle(dlm, &mle, (char *)lockname,
1687                                          namelen)) {
1688                                 dlm_print_one_mle(mle);
1689                                 __dlm_put_mle(mle);
1690                         }
1691                         spin_unlock(&dlm->master_lock);
1692                         spin_unlock(&dlm->spinlock);
1693                         BUG();
1694                 }
1695
1696                 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1697                     !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1698                                 mlog(ML_ERROR, "%.*s: very strange, "
1699                                      "master MLE but no lockres on %u\n",
1700                                      namelen, lockname, to);
1701                 }
1702
1703                 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1704                         mlog(0, "%.*s: node %u create mles on other "
1705                              "nodes and requests a re-assert\n",
1706                              namelen, lockname, to);
1707                         reassert = 1;
1708                 }
1709                 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1710                         mlog(0, "%.*s: node %u has a reference to this "
1711                              "lockres, set the bit in the refmap\n",
1712                              namelen, lockname, to);
1713                         spin_lock(&res->spinlock);
1714                         dlm_lockres_set_refmap_bit(to, res);
1715                         spin_unlock(&res->spinlock);
1716                 }
1717         }
1718
1719         if (reassert)
1720                 goto again;
1721
1722         spin_lock(&res->spinlock);
1723         res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1724         spin_unlock(&res->spinlock);
1725         wake_up(&res->wq);
1726
1727         return ret;
1728 }
1729
1730 /*
1731  * locks that can be taken here:
1732  * dlm->spinlock
1733  * res->spinlock
1734  * mle->spinlock
1735  * dlm->master_list
1736  *
1737  * if possible, TRIM THIS DOWN!!!
1738  */
1739 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1740                               void **ret_data)
1741 {
1742         struct dlm_ctxt *dlm = data;
1743         struct dlm_master_list_entry *mle = NULL;
1744         struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1745         struct dlm_lock_resource *res = NULL;
1746         char *name;
1747         unsigned int namelen, hash;
1748         u32 flags;
1749         int master_request = 0, have_lockres_ref = 0;
1750         int ret = 0;
1751
1752         if (!dlm_grab(dlm))
1753                 return 0;
1754
1755         name = assert->name;
1756         namelen = assert->namelen;
1757         hash = dlm_lockid_hash(name, namelen);
1758         flags = be32_to_cpu(assert->flags);
1759
1760         if (namelen > DLM_LOCKID_NAME_MAX) {
1761                 mlog(ML_ERROR, "Invalid name length!");
1762                 goto done;
1763         }
1764
1765         spin_lock(&dlm->spinlock);
1766
1767         if (flags)
1768                 mlog(0, "assert_master with flags: %u\n", flags);
1769
1770         /* find the MLE */
1771         spin_lock(&dlm->master_lock);
1772         if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1773                 /* not an error, could be master just re-asserting */
1774                 mlog(0, "just got an assert_master from %u, but no "
1775                      "MLE for it! (%.*s)\n", assert->node_idx,
1776                      namelen, name);
1777         } else {
1778                 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1779                 if (bit >= O2NM_MAX_NODES) {
1780                         /* not necessarily an error, though less likely.
1781                          * could be master just re-asserting. */
1782                         mlog(0, "no bits set in the maybe_map, but %u "
1783                              "is asserting! (%.*s)\n", assert->node_idx,
1784                              namelen, name);
1785                 } else if (bit != assert->node_idx) {
1786                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1787                                 mlog(0, "master %u was found, %u should "
1788                                      "back off\n", assert->node_idx, bit);
1789                         } else {
1790                                 /* with the fix for bug 569, a higher node
1791                                  * number winning the mastery will respond
1792                                  * YES to mastery requests, but this node
1793                                  * had no way of knowing.  let it pass. */
1794                                 mlog(0, "%u is the lowest node, "
1795                                      "%u is asserting. (%.*s)  %u must "
1796                                      "have begun after %u won.\n", bit,
1797                                      assert->node_idx, namelen, name, bit,
1798                                      assert->node_idx);
1799                         }
1800                 }
1801                 if (mle->type == DLM_MLE_MIGRATION) {
1802                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1803                                 mlog(0, "%s:%.*s: got cleanup assert"
1804                                      " from %u for migration\n",
1805                                      dlm->name, namelen, name,
1806                                      assert->node_idx);
1807                         } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1808                                 mlog(0, "%s:%.*s: got unrelated assert"
1809                                      " from %u for migration, ignoring\n",
1810                                      dlm->name, namelen, name,
1811                                      assert->node_idx);
1812                                 __dlm_put_mle(mle);
1813                                 spin_unlock(&dlm->master_lock);
1814                                 spin_unlock(&dlm->spinlock);
1815                                 goto done;
1816                         }
1817                 }
1818         }
1819         spin_unlock(&dlm->master_lock);
1820
1821         /* ok everything checks out with the MLE
1822          * now check to see if there is a lockres */
1823         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1824         if (res) {
1825                 spin_lock(&res->spinlock);
1826                 if (res->state & DLM_LOCK_RES_RECOVERING)  {
1827                         mlog(ML_ERROR, "%u asserting but %.*s is "
1828                              "RECOVERING!\n", assert->node_idx, namelen, name);
1829                         goto kill;
1830                 }
1831                 if (!mle) {
1832                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1833                             res->owner != assert->node_idx) {
1834                                 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1835                                      "but current owner is %u! (%.*s)\n",
1836                                      assert->node_idx, res->owner, namelen,
1837                                      name);
1838                                 __dlm_print_one_lock_resource(res);
1839                                 BUG();
1840                         }
1841                 } else if (mle->type != DLM_MLE_MIGRATION) {
1842                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1843                                 /* owner is just re-asserting */
1844                                 if (res->owner == assert->node_idx) {
1845                                         mlog(0, "owner %u re-asserting on "
1846                                              "lock %.*s\n", assert->node_idx,
1847                                              namelen, name);
1848                                         goto ok;
1849                                 }
1850                                 mlog(ML_ERROR, "got assert_master from "
1851                                      "node %u, but %u is the owner! "
1852                                      "(%.*s)\n", assert->node_idx,
1853                                      res->owner, namelen, name);
1854                                 goto kill;
1855                         }
1856                         if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1857                                 mlog(ML_ERROR, "got assert from %u, but lock "
1858                                      "with no owner should be "
1859                                      "in-progress! (%.*s)\n",
1860                                      assert->node_idx,
1861                                      namelen, name);
1862                                 goto kill;
1863                         }
1864                 } else /* mle->type == DLM_MLE_MIGRATION */ {
1865                         /* should only be getting an assert from new master */
1866                         if (assert->node_idx != mle->new_master) {
1867                                 mlog(ML_ERROR, "got assert from %u, but "
1868                                      "new master is %u, and old master "
1869                                      "was %u (%.*s)\n",
1870                                      assert->node_idx, mle->new_master,
1871                                      mle->master, namelen, name);
1872                                 goto kill;
1873                         }
1874
1875                 }
1876 ok:
1877                 spin_unlock(&res->spinlock);
1878         }
1879         spin_unlock(&dlm->spinlock);
1880
1881         // mlog(0, "woo!  got an assert_master from node %u!\n",
1882         //           assert->node_idx);
1883         if (mle) {
1884                 int extra_ref = 0;
1885                 int nn = -1;
1886                 int rr, err = 0;
1887
1888                 spin_lock(&mle->spinlock);
1889                 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1890                         extra_ref = 1;
1891                 else {
1892                         /* MASTER mle: if any bits set in the response map
1893                          * then the calling node needs to re-assert to clear
1894                          * up nodes that this node contacted */
1895                         while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1896                                                     nn+1)) < O2NM_MAX_NODES) {
1897                                 if (nn != dlm->node_num && nn != assert->node_idx)
1898                                         master_request = 1;
1899                         }
1900                 }
1901                 mle->master = assert->node_idx;
1902                 atomic_set(&mle->woken, 1);
1903                 wake_up(&mle->wq);
1904                 spin_unlock(&mle->spinlock);
1905
1906                 if (res) {
1907                         int wake = 0;
1908                         spin_lock(&res->spinlock);
1909                         if (mle->type == DLM_MLE_MIGRATION) {
1910                                 mlog(0, "finishing off migration of lockres %.*s, "
1911                                         "from %u to %u\n",
1912                                         res->lockname.len, res->lockname.name,
1913                                         dlm->node_num, mle->new_master);
1914                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
1915                                 wake = 1;
1916                                 dlm_change_lockres_owner(dlm, res, mle->new_master);
1917                                 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1918                         } else {
1919                                 dlm_change_lockres_owner(dlm, res, mle->master);
1920                         }
1921                         spin_unlock(&res->spinlock);
1922                         have_lockres_ref = 1;
1923                         if (wake)
1924                                 wake_up(&res->wq);
1925                 }
1926
1927                 /* master is known, detach if not already detached.
1928                  * ensures that only one assert_master call will happen
1929                  * on this mle. */
1930                 spin_lock(&dlm->spinlock);
1931                 spin_lock(&dlm->master_lock);
1932
1933                 rr = atomic_read(&mle->mle_refs.refcount);
1934                 if (mle->inuse > 0) {
1935                         if (extra_ref && rr < 3)
1936                                 err = 1;
1937                         else if (!extra_ref && rr < 2)
1938                                 err = 1;
1939                 } else {
1940                         if (extra_ref && rr < 2)
1941                                 err = 1;
1942                         else if (!extra_ref && rr < 1)
1943                                 err = 1;
1944                 }
1945                 if (err) {
1946                         mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
1947                              "that will mess up this node, refs=%d, extra=%d, "
1948                              "inuse=%d\n", dlm->name, namelen, name,
1949                              assert->node_idx, rr, extra_ref, mle->inuse);
1950                         dlm_print_one_mle(mle);
1951                 }
1952                 __dlm_unlink_mle(dlm, mle);
1953                 __dlm_mle_detach_hb_events(dlm, mle);
1954                 __dlm_put_mle(mle);
1955                 if (extra_ref) {
1956                         /* the assert master message now balances the extra
1957                          * ref given by the master / migration request message.
1958                          * if this is the last put, it will be removed
1959                          * from the list. */
1960                         __dlm_put_mle(mle);
1961                 }
1962                 spin_unlock(&dlm->master_lock);
1963                 spin_unlock(&dlm->spinlock);
1964         } else if (res) {
1965                 if (res->owner != assert->node_idx) {
1966                         mlog(0, "assert_master from %u, but current "
1967                              "owner is %u (%.*s), no mle\n", assert->node_idx,
1968                              res->owner, namelen, name);
1969                 }
1970         }
1971
1972 done:
1973         ret = 0;
1974         if (res) {
1975                 spin_lock(&res->spinlock);
1976                 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1977                 spin_unlock(&res->spinlock);
1978                 *ret_data = (void *)res;
1979         }
1980         dlm_put(dlm);
1981         if (master_request) {
1982                 mlog(0, "need to tell master to reassert\n");
1983                 /* positive. negative would shoot down the node. */
1984                 ret |= DLM_ASSERT_RESPONSE_REASSERT;
1985                 if (!have_lockres_ref) {
1986                         mlog(ML_ERROR, "strange, got assert from %u, MASTER "
1987                              "mle present here for %s:%.*s, but no lockres!\n",
1988                              assert->node_idx, dlm->name, namelen, name);
1989                 }
1990         }
1991         if (have_lockres_ref) {
1992                 /* let the master know we have a reference to the lockres */
1993                 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
1994                 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
1995                      dlm->name, namelen, name, assert->node_idx);
1996         }
1997         return ret;
1998
1999 kill:
2000         /* kill the caller! */
2001         mlog(ML_ERROR, "Bad message received from another node.  Dumping state "
2002              "and killing the other node now!  This node is OK and can continue.\n");
2003         __dlm_print_one_lock_resource(res);
2004         spin_unlock(&res->spinlock);
2005         spin_unlock(&dlm->spinlock);
2006         *ret_data = (void *)res;
2007         dlm_put(dlm);
2008         return -EINVAL;
2009 }
2010
2011 void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2012 {
2013         struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2014
2015         if (ret_data) {
2016                 spin_lock(&res->spinlock);
2017                 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2018                 spin_unlock(&res->spinlock);
2019                 wake_up(&res->wq);
2020                 dlm_lockres_put(res);
2021         }
2022         return;
2023 }
2024
2025 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2026                                struct dlm_lock_resource *res,
2027                                int ignore_higher, u8 request_from, u32 flags)
2028 {
2029         struct dlm_work_item *item;
2030         item = kzalloc(sizeof(*item), GFP_NOFS);
2031         if (!item)
2032                 return -ENOMEM;
2033
2034
2035         /* queue up work for dlm_assert_master_worker */
2036         dlm_grab(dlm);  /* get an extra ref for the work item */
2037         dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2038         item->u.am.lockres = res; /* already have a ref */
2039         /* can optionally ignore node numbers higher than this node */
2040         item->u.am.ignore_higher = ignore_higher;
2041         item->u.am.request_from = request_from;
2042         item->u.am.flags = flags;
2043
2044         if (ignore_higher)
2045                 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2046                      res->lockname.name);
2047
2048         spin_lock(&dlm->work_lock);
2049         list_add_tail(&item->list, &dlm->work_list);
2050         spin_unlock(&dlm->work_lock);
2051
2052         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2053         return 0;
2054 }
2055
2056 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2057 {
2058         struct dlm_ctxt *dlm = data;
2059         int ret = 0;
2060         struct dlm_lock_resource *res;
2061         unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2062         int ignore_higher;
2063         int bit;
2064         u8 request_from;
2065         u32 flags;
2066
2067         dlm = item->dlm;
2068         res = item->u.am.lockres;
2069         ignore_higher = item->u.am.ignore_higher;
2070         request_from = item->u.am.request_from;
2071         flags = item->u.am.flags;
2072
2073         spin_lock(&dlm->spinlock);
2074         memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2075         spin_unlock(&dlm->spinlock);
2076
2077         clear_bit(dlm->node_num, nodemap);
2078         if (ignore_higher) {
2079                 /* if is this just to clear up mles for nodes below
2080                  * this node, do not send the message to the original
2081                  * caller or any node number higher than this */
2082                 clear_bit(request_from, nodemap);
2083                 bit = dlm->node_num;
2084                 while (1) {
2085                         bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2086                                             bit+1);
2087                         if (bit >= O2NM_MAX_NODES)
2088                                 break;
2089                         clear_bit(bit, nodemap);
2090                 }
2091         }
2092
2093         /*
2094          * If we're migrating this lock to someone else, we are no
2095          * longer allowed to assert out own mastery.  OTOH, we need to
2096          * prevent migration from starting while we're still asserting
2097          * our dominance.  The reserved ast delays migration.
2098          */
2099         spin_lock(&res->spinlock);
2100         if (res->state & DLM_LOCK_RES_MIGRATING) {
2101                 mlog(0, "Someone asked us to assert mastery, but we're "
2102                      "in the middle of migration.  Skipping assert, "
2103                      "the new master will handle that.\n");
2104                 spin_unlock(&res->spinlock);
2105                 goto put;
2106         } else
2107                 __dlm_lockres_reserve_ast(res);
2108         spin_unlock(&res->spinlock);
2109
2110         /* this call now finishes out the nodemap
2111          * even if one or more nodes die */
2112         mlog(0, "worker about to master %.*s here, this=%u\n",
2113                      res->lockname.len, res->lockname.name, dlm->node_num);
2114         ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2115         if (ret < 0) {
2116                 /* no need to restart, we are done */
2117                 if (!dlm_is_host_down(ret))
2118                         mlog_errno(ret);
2119         }
2120
2121         /* Ok, we've asserted ourselves.  Let's let migration start. */
2122         dlm_lockres_release_ast(dlm, res);
2123
2124 put:
2125         dlm_lockres_put(res);
2126
2127         mlog(0, "finished with dlm_assert_master_worker\n");
2128 }
2129
2130 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2131  * We cannot wait for node recovery to complete to begin mastering this
2132  * lockres because this lockres is used to kick off recovery! ;-)
2133  * So, do a pre-check on all living nodes to see if any of those nodes
2134  * think that $RECOVERY is currently mastered by a dead node.  If so,
2135  * we wait a short time to allow that node to get notified by its own
2136  * heartbeat stack, then check again.  All $RECOVERY lock resources
2137  * mastered by dead nodes are purged when the hearbeat callback is
2138  * fired, so we can know for sure that it is safe to continue once
2139  * the node returns a live node or no node.  */
2140 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2141                                        struct dlm_lock_resource *res)
2142 {
2143         struct dlm_node_iter iter;
2144         int nodenum;
2145         int ret = 0;
2146         u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2147
2148         spin_lock(&dlm->spinlock);
2149         dlm_node_iter_init(dlm->domain_map, &iter);
2150         spin_unlock(&dlm->spinlock);
2151
2152         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2153                 /* do not send to self */
2154                 if (nodenum == dlm->node_num)
2155                         continue;
2156                 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2157                 if (ret < 0) {
2158                         mlog_errno(ret);
2159                         if (!dlm_is_host_down(ret))
2160                                 BUG();
2161                         /* host is down, so answer for that node would be
2162                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
2163                         ret = 0;
2164                 }
2165
2166                 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2167                         /* check to see if this master is in the recovery map */
2168                         spin_lock(&dlm->spinlock);
2169                         if (test_bit(master, dlm->recovery_map)) {
2170                                 mlog(ML_NOTICE, "%s: node %u has not seen "
2171                                      "node %u go down yet, and thinks the "
2172                                      "dead node is mastering the recovery "
2173                                      "lock.  must wait.\n", dlm->name,
2174                                      nodenum, master);
2175                                 ret = -EAGAIN;
2176                         }
2177                         spin_unlock(&dlm->spinlock);
2178                         mlog(0, "%s: reco lock master is %u\n", dlm->name,
2179                              master);
2180                         break;
2181                 }
2182         }
2183         return ret;
2184 }
2185
2186 /*
2187  * DLM_DEREF_LOCKRES_MSG
2188  */
2189
2190 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2191 {
2192         struct dlm_deref_lockres deref;
2193         int ret = 0, r;
2194         const char *lockname;
2195         unsigned int namelen;
2196
2197         lockname = res->lockname.name;
2198         namelen = res->lockname.len;
2199         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2200
2201         mlog(0, "%s:%.*s: sending deref to %d\n",
2202              dlm->name, namelen, lockname, res->owner);
2203         memset(&deref, 0, sizeof(deref));
2204         deref.node_idx = dlm->node_num;
2205         deref.namelen = namelen;
2206         memcpy(deref.name, lockname, namelen);
2207
2208         ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2209                                  &deref, sizeof(deref), res->owner, &r);
2210         if (ret < 0)
2211                 mlog_errno(ret);
2212         else if (r < 0) {
2213                 /* BAD.  other node says I did not have a ref. */
2214                 mlog(ML_ERROR,"while dropping ref on %s:%.*s "
2215                     "(master=%u) got %d.\n", dlm->name, namelen,
2216                     lockname, res->owner, r);
2217                 dlm_print_one_lock_resource(res);
2218                 BUG();
2219         }
2220         return ret;
2221 }
2222
2223 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2224                               void **ret_data)
2225 {
2226         struct dlm_ctxt *dlm = data;
2227         struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2228         struct dlm_lock_resource *res = NULL;
2229         char *name;
2230         unsigned int namelen;
2231         int ret = -EINVAL;
2232         u8 node;
2233         unsigned int hash;
2234         struct dlm_work_item *item;
2235         int cleared = 0;
2236         int dispatch = 0;
2237
2238         if (!dlm_grab(dlm))
2239                 return 0;
2240
2241         name = deref->name;
2242         namelen = deref->namelen;
2243         node = deref->node_idx;
2244
2245         if (namelen > DLM_LOCKID_NAME_MAX) {
2246                 mlog(ML_ERROR, "Invalid name length!");
2247                 goto done;
2248         }
2249         if (deref->node_idx >= O2NM_MAX_NODES) {
2250                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2251                 goto done;
2252         }
2253
2254         hash = dlm_lockid_hash(name, namelen);
2255
2256         spin_lock(&dlm->spinlock);
2257         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2258         if (!res) {
2259                 spin_unlock(&dlm->spinlock);
2260                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2261                      dlm->name, namelen, name);
2262                 goto done;
2263         }
2264         spin_unlock(&dlm->spinlock);
2265
2266         spin_lock(&res->spinlock);
2267         if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2268                 dispatch = 1;
2269         else {
2270                 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2271                 if (test_bit(node, res->refmap)) {
2272                         dlm_lockres_clear_refmap_bit(node, res);
2273                         cleared = 1;
2274                 }
2275         }
2276         spin_unlock(&res->spinlock);
2277
2278         if (!dispatch) {
2279                 if (cleared)
2280                         dlm_lockres_calc_usage(dlm, res);
2281                 else {
2282                         mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2283                         "but it is already dropped!\n", dlm->name,
2284                         res->lockname.len, res->lockname.name, node);
2285                         dlm_print_one_lock_resource(res);
2286                 }
2287                 ret = 0;
2288                 goto done;
2289         }
2290
2291         item = kzalloc(sizeof(*item), GFP_NOFS);
2292         if (!item) {
2293                 ret = -ENOMEM;
2294                 mlog_errno(ret);
2295                 goto done;
2296         }
2297
2298         dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2299         item->u.dl.deref_res = res;
2300         item->u.dl.deref_node = node;
2301
2302         spin_lock(&dlm->work_lock);
2303         list_add_tail(&item->list, &dlm->work_list);
2304         spin_unlock(&dlm->work_lock);
2305
2306         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2307         return 0;
2308
2309 done:
2310         if (res)
2311                 dlm_lockres_put(res);
2312         dlm_put(dlm);
2313
2314         return ret;
2315 }
2316
2317 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2318 {
2319         struct dlm_ctxt *dlm;
2320         struct dlm_lock_resource *res;
2321         u8 node;
2322         u8 cleared = 0;
2323
2324         dlm = item->dlm;
2325         res = item->u.dl.deref_res;
2326         node = item->u.dl.deref_node;
2327
2328         spin_lock(&res->spinlock);
2329         BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2330         if (test_bit(node, res->refmap)) {
2331                 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2332                 dlm_lockres_clear_refmap_bit(node, res);
2333                 cleared = 1;
2334         }
2335         spin_unlock(&res->spinlock);
2336
2337         if (cleared) {
2338                 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2339                      dlm->name, res->lockname.len, res->lockname.name, node);
2340                 dlm_lockres_calc_usage(dlm, res);
2341         } else {
2342                 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2343                      "but it is already dropped!\n", dlm->name,
2344                      res->lockname.len, res->lockname.name, node);
2345                 dlm_print_one_lock_resource(res);
2346         }
2347
2348         dlm_lockres_put(res);
2349 }
2350
2351 /* Checks whether the lockres can be migrated. Returns 0 if yes, < 0
2352  * if not. If 0, numlocks is set to the number of locks in the lockres.
2353  */
2354 static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
2355                                       struct dlm_lock_resource *res,
2356                                       int *numlocks)
2357 {
2358         int ret;
2359         int i;
2360         int count = 0;
2361         struct list_head *queue;
2362         struct dlm_lock *lock;
2363
2364         assert_spin_locked(&res->spinlock);
2365
2366         ret = -EINVAL;
2367         if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
2368                 mlog(0, "cannot migrate lockres with unknown owner!\n");
2369                 goto leave;
2370         }
2371
2372         if (res->owner != dlm->node_num) {
2373                 mlog(0, "cannot migrate lockres this node doesn't own!\n");
2374                 goto leave;
2375         }
2376
2377         ret = 0;
2378         queue = &res->granted;
2379         for (i = 0; i < 3; i++) {
2380                 list_for_each_entry(lock, queue, list) {
2381                         ++count;
2382                         if (lock->ml.node == dlm->node_num) {
2383                                 mlog(0, "found a lock owned by this node still "
2384                                      "on the %s queue!  will not migrate this "
2385                                      "lockres\n", (i == 0 ? "granted" :
2386                                                    (i == 1 ? "converting" :
2387                                                     "blocked")));
2388                                 ret = -ENOTEMPTY;
2389                                 goto leave;
2390                         }
2391                 }
2392                 queue++;
2393         }
2394
2395         *numlocks = count;
2396         mlog(0, "migrateable lockres having %d locks\n", *numlocks);
2397
2398 leave:
2399         return ret;
2400 }
2401
2402 /*
2403  * DLM_MIGRATE_LOCKRES
2404  */
2405
2406
2407 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2408                                struct dlm_lock_resource *res,
2409                                u8 target)
2410 {
2411         struct dlm_master_list_entry *mle = NULL;
2412         struct dlm_master_list_entry *oldmle = NULL;
2413         struct dlm_migratable_lockres *mres = NULL;
2414         int ret = 0;
2415         const char *name;
2416         unsigned int namelen;
2417         int mle_added = 0;
2418         int numlocks;
2419         int wake = 0;
2420
2421         if (!dlm_grab(dlm))
2422                 return -EINVAL;
2423
2424         name = res->lockname.name;
2425         namelen = res->lockname.len;
2426
2427         mlog(0, "migrating %.*s to %u\n", namelen, name, target);
2428
2429         /*
2430          * ensure this lockres is a proper candidate for migration
2431          */
2432         spin_lock(&res->spinlock);
2433         ret = dlm_is_lockres_migrateable(dlm, res, &numlocks);
2434         if (ret < 0) {
2435                 spin_unlock(&res->spinlock);
2436                 goto leave;
2437         }
2438         spin_unlock(&res->spinlock);
2439
2440         /* no work to do */
2441         if (numlocks == 0) {
2442                 mlog(0, "no locks were found on this lockres! done!\n");
2443                 goto leave;
2444         }
2445
2446         /*
2447          * preallocate up front
2448          * if this fails, abort
2449          */
2450
2451         ret = -ENOMEM;
2452         mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2453         if (!mres) {
2454                 mlog_errno(ret);
2455                 goto leave;
2456         }
2457
2458         mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
2459                                                                 GFP_NOFS);
2460         if (!mle) {
2461                 mlog_errno(ret);
2462                 goto leave;
2463         }
2464         ret = 0;
2465
2466         /*
2467          * find a node to migrate the lockres to
2468          */
2469
2470         mlog(0, "picking a migration node\n");
2471         spin_lock(&dlm->spinlock);
2472         /* pick a new node */
2473         if (!test_bit(target, dlm->domain_map) ||
2474             target >= O2NM_MAX_NODES) {
2475                 target = dlm_pick_migration_target(dlm, res);
2476         }
2477         mlog(0, "node %u chosen for migration\n", target);
2478
2479         if (target >= O2NM_MAX_NODES ||
2480             !test_bit(target, dlm->domain_map)) {
2481                 /* target chosen is not alive */
2482                 ret = -EINVAL;
2483         }
2484
2485         if (ret) {
2486                 spin_unlock(&dlm->spinlock);
2487                 goto fail;
2488         }
2489
2490         mlog(0, "continuing with target = %u\n", target);
2491
2492         /*
2493          * clear any existing master requests and
2494          * add the migration mle to the list
2495          */
2496         spin_lock(&dlm->master_lock);
2497         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2498                                     namelen, target, dlm->node_num);
2499         spin_unlock(&dlm->master_lock);
2500         spin_unlock(&dlm->spinlock);
2501
2502         if (ret == -EEXIST) {
2503                 mlog(0, "another process is already migrating it\n");
2504                 goto fail;
2505         }
2506         mle_added = 1;
2507
2508         /*
2509          * set the MIGRATING flag and flush asts
2510          * if we fail after this we need to re-dirty the lockres
2511          */
2512         if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2513                 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2514                      "the target went down.\n", res->lockname.len,
2515                      res->lockname.name, target);
2516                 spin_lock(&res->spinlock);
2517                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2518                 wake = 1;
2519                 spin_unlock(&res->spinlock);
2520                 ret = -EINVAL;
2521         }
2522
2523 fail:
2524         if (oldmle) {
2525                 /* master is known, detach if not already detached */
2526                 dlm_mle_detach_hb_events(dlm, oldmle);
2527                 dlm_put_mle(oldmle);
2528         }
2529
2530         if (ret < 0) {
2531                 if (mle_added) {
2532                         dlm_mle_detach_hb_events(dlm, mle);
2533                         dlm_put_mle(mle);
2534                 } else if (mle) {
2535                         kmem_cache_free(dlm_mle_cache, mle);
2536                 }
2537                 goto leave;
2538         }
2539
2540         /*
2541          * at this point, we have a migration target, an mle
2542          * in the master list, and the MIGRATING flag set on
2543          * the lockres
2544          */
2545
2546         /* now that remote nodes are spinning on the MIGRATING flag,
2547          * ensure that all assert_master work is flushed. */
2548         flush_workqueue(dlm->dlm_worker);
2549
2550         /* get an extra reference on the mle.
2551          * otherwise the assert_master from the new
2552          * master will destroy this.
2553          * also, make sure that all callers of dlm_get_mle
2554          * take both dlm->spinlock and dlm->master_lock */
2555         spin_lock(&dlm->spinlock);
2556         spin_lock(&dlm->master_lock);
2557         dlm_get_mle_inuse(mle);
2558         spin_unlock(&dlm->master_lock);
2559         spin_unlock(&dlm->spinlock);
2560
2561         /* notify new node and send all lock state */
2562         /* call send_one_lockres with migration flag.
2563          * this serves as notice to the target node that a
2564          * migration is starting. */
2565         ret = dlm_send_one_lockres(dlm, res, mres, target,
2566                                    DLM_MRES_MIGRATION);
2567
2568         if (ret < 0) {
2569                 mlog(0, "migration to node %u failed with %d\n",
2570                      target, ret);
2571                 /* migration failed, detach and clean up mle */
2572                 dlm_mle_detach_hb_events(dlm, mle);
2573                 dlm_put_mle(mle);
2574                 dlm_put_mle_inuse(mle);
2575                 spin_lock(&res->spinlock);
2576                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2577                 wake = 1;
2578                 spin_unlock(&res->spinlock);
2579                 goto leave;
2580         }
2581
2582         /* at this point, the target sends a message to all nodes,
2583          * (using dlm_do_migrate_request).  this node is skipped since
2584          * we had to put an mle in the list to begin the process.  this
2585          * node now waits for target to do an assert master.  this node
2586          * will be the last one notified, ensuring that the migration
2587          * is complete everywhere.  if the target dies while this is
2588          * going on, some nodes could potentially see the target as the
2589          * master, so it is important that my recovery finds the migration
2590          * mle and sets the master to UNKNONWN. */
2591
2592
2593         /* wait for new node to assert master */
2594         while (1) {
2595                 ret = wait_event_interruptible_timeout(mle->wq,
2596                                         (atomic_read(&mle->woken) == 1),
2597                                         msecs_to_jiffies(5000));
2598
2599                 if (ret >= 0) {
2600                         if (atomic_read(&mle->woken) == 1 ||
2601                             res->owner == target)
2602                                 break;
2603
2604                         mlog(0, "%s:%.*s: timed out during migration\n",
2605                              dlm->name, res->lockname.len, res->lockname.name);
2606                         /* avoid hang during shutdown when migrating lockres
2607                          * to a node which also goes down */
2608                         if (dlm_is_node_dead(dlm, target)) {
2609                                 mlog(0, "%s:%.*s: expected migration "
2610                                      "target %u is no longer up, restarting\n",
2611                                      dlm->name, res->lockname.len,
2612                                      res->lockname.name, target);
2613                                 ret = -EINVAL;
2614                                 /* migration failed, detach and clean up mle */
2615                                 dlm_mle_detach_hb_events(dlm, mle);
2616                                 dlm_put_mle(mle);
2617                                 dlm_put_mle_inuse(mle);
2618                                 spin_lock(&res->spinlock);
2619                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2620                                 wake = 1;
2621                                 spin_unlock(&res->spinlock);
2622                                 goto leave;
2623                         }
2624                 } else
2625                         mlog(0, "%s:%.*s: caught signal during migration\n",
2626                              dlm->name, res->lockname.len, res->lockname.name);
2627         }
2628
2629         /* all done, set the owner, clear the flag */
2630         spin_lock(&res->spinlock);
2631         dlm_set_lockres_owner(dlm, res, target);
2632         res->state &= ~DLM_LOCK_RES_MIGRATING;
2633         dlm_remove_nonlocal_locks(dlm, res);
2634         spin_unlock(&res->spinlock);
2635         wake_up(&res->wq);
2636
2637         /* master is known, detach if not already detached */
2638         dlm_mle_detach_hb_events(dlm, mle);
2639         dlm_put_mle_inuse(mle);
2640         ret = 0;
2641
2642         dlm_lockres_calc_usage(dlm, res);
2643
2644 leave:
2645         /* re-dirty the lockres if we failed */
2646         if (ret < 0)
2647                 dlm_kick_thread(dlm, res);
2648
2649         /* wake up waiters if the MIGRATING flag got set
2650          * but migration failed */
2651         if (wake)
2652                 wake_up(&res->wq);
2653
2654         /* TODO: cleanup */
2655         if (mres)
2656                 free_page((unsigned long)mres);
2657
2658         dlm_put(dlm);
2659
2660         mlog(0, "returning %d\n", ret);
2661         return ret;
2662 }
2663
2664 #define DLM_MIGRATION_RETRY_MS  100
2665
2666 /* Should be called only after beginning the domain leave process.
2667  * There should not be any remaining locks on nonlocal lock resources,
2668  * and there should be no local locks left on locally mastered resources.
2669  *
2670  * Called with the dlm spinlock held, may drop it to do migration, but
2671  * will re-acquire before exit.
2672  *
2673  * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped */
2674 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2675 {
2676         int ret;
2677         int lock_dropped = 0;
2678         int numlocks;
2679
2680         spin_lock(&res->spinlock);
2681         if (res->owner != dlm->node_num) {
2682                 if (!__dlm_lockres_unused(res)) {
2683                         mlog(ML_ERROR, "%s:%.*s: this node is not master, "
2684                              "trying to free this but locks remain\n",
2685                              dlm->name, res->lockname.len, res->lockname.name);
2686                 }
2687                 spin_unlock(&res->spinlock);
2688                 goto leave;
2689         }
2690
2691         /* No need to migrate a lockres having no locks */
2692         ret = dlm_is_lockres_migrateable(dlm, res, &numlocks);
2693         if (ret >= 0 && numlocks == 0) {
2694                 spin_unlock(&res->spinlock);
2695                 goto leave;
2696         }
2697         spin_unlock(&res->spinlock);
2698
2699         /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2700         spin_unlock(&dlm->spinlock);
2701         lock_dropped = 1;
2702         while (1) {
2703                 ret = dlm_migrate_lockres(dlm, res, O2NM_MAX_NODES);
2704                 if (ret >= 0)
2705                         break;
2706                 if (ret == -ENOTEMPTY) {
2707                         mlog(ML_ERROR, "lockres %.*s still has local locks!\n",
2708                                 res->lockname.len, res->lockname.name);
2709                         BUG();
2710                 }
2711
2712                 mlog(0, "lockres %.*s: migrate failed, "
2713                      "retrying\n", res->lockname.len,
2714                      res->lockname.name);
2715                 msleep(DLM_MIGRATION_RETRY_MS);
2716         }
2717         spin_lock(&dlm->spinlock);
2718 leave:
2719         return lock_dropped;
2720 }
2721
2722 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2723 {
2724         int ret;
2725         spin_lock(&dlm->ast_lock);
2726         spin_lock(&lock->spinlock);
2727         ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2728         spin_unlock(&lock->spinlock);
2729         spin_unlock(&dlm->ast_lock);
2730         return ret;
2731 }
2732
2733 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2734                                      struct dlm_lock_resource *res,
2735                                      u8 mig_target)
2736 {
2737         int can_proceed;
2738         spin_lock(&res->spinlock);
2739         can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2740         spin_unlock(&res->spinlock);
2741
2742         /* target has died, so make the caller break out of the
2743          * wait_event, but caller must recheck the domain_map */
2744         spin_lock(&dlm->spinlock);
2745         if (!test_bit(mig_target, dlm->domain_map))
2746                 can_proceed = 1;
2747         spin_unlock(&dlm->spinlock);
2748         return can_proceed;
2749 }
2750
2751 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2752                                 struct dlm_lock_resource *res)
2753 {
2754         int ret;
2755         spin_lock(&res->spinlock);
2756         ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2757         spin_unlock(&res->spinlock);
2758         return ret;
2759 }
2760
2761
2762 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2763                                        struct dlm_lock_resource *res,
2764                                        u8 target)
2765 {
2766         int ret = 0;
2767
2768         mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2769                res->lockname.len, res->lockname.name, dlm->node_num,
2770                target);
2771         /* need to set MIGRATING flag on lockres.  this is done by
2772          * ensuring that all asts have been flushed for this lockres. */
2773         spin_lock(&res->spinlock);
2774         BUG_ON(res->migration_pending);
2775         res->migration_pending = 1;
2776         /* strategy is to reserve an extra ast then release
2777          * it below, letting the release do all of the work */
2778         __dlm_lockres_reserve_ast(res);
2779         spin_unlock(&res->spinlock);
2780
2781         /* now flush all the pending asts */
2782         dlm_kick_thread(dlm, res);
2783         /* before waiting on DIRTY, block processes which may
2784          * try to dirty the lockres before MIGRATING is set */
2785         spin_lock(&res->spinlock);
2786         BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2787         res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2788         spin_unlock(&res->spinlock);
2789         /* now wait on any pending asts and the DIRTY state */
2790         wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2791         dlm_lockres_release_ast(dlm, res);
2792
2793         mlog(0, "about to wait on migration_wq, dirty=%s\n",
2794                res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2795         /* if the extra ref we just put was the final one, this
2796          * will pass thru immediately.  otherwise, we need to wait
2797          * for the last ast to finish. */
2798 again:
2799         ret = wait_event_interruptible_timeout(dlm->migration_wq,
2800                    dlm_migration_can_proceed(dlm, res, target),
2801                    msecs_to_jiffies(1000));
2802         if (ret < 0) {
2803                 mlog(0, "woken again: migrating? %s, dead? %s\n",
2804                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2805                        test_bit(target, dlm->domain_map) ? "no":"yes");
2806         } else {
2807                 mlog(0, "all is well: migrating? %s, dead? %s\n",
2808                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2809                        test_bit(target, dlm->domain_map) ? "no":"yes");
2810         }
2811         if (!dlm_migration_can_proceed(dlm, res, target)) {
2812                 mlog(0, "trying again...\n");
2813                 goto again;
2814         }
2815         /* now that we are sure the MIGRATING state is there, drop
2816          * the unneded state which blocked threads trying to DIRTY */
2817         spin_lock(&res->spinlock);
2818         BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2819         BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2820         res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2821         spin_unlock(&res->spinlock);
2822
2823         /* did the target go down or die? */
2824         spin_lock(&dlm->spinlock);
2825         if (!test_bit(target, dlm->domain_map)) {
2826                 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2827                      target);
2828                 ret = -EHOSTDOWN;
2829         }
2830         spin_unlock(&dlm->spinlock);
2831
2832         /*
2833          * at this point:
2834          *
2835          *   o the DLM_LOCK_RES_MIGRATING flag is set
2836          *   o there are no pending asts on this lockres
2837          *   o all processes trying to reserve an ast on this
2838          *     lockres must wait for the MIGRATING flag to clear
2839          */
2840         return ret;
2841 }
2842
2843 /* last step in the migration process.
2844  * original master calls this to free all of the dlm_lock
2845  * structures that used to be for other nodes. */
2846 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2847                                       struct dlm_lock_resource *res)
2848 {
2849         struct list_head *queue = &res->granted;
2850         int i, bit;
2851         struct dlm_lock *lock, *next;
2852
2853         assert_spin_locked(&res->spinlock);
2854
2855         BUG_ON(res->owner == dlm->node_num);
2856
2857         for (i=0; i<3; i++) {
2858                 list_for_each_entry_safe(lock, next, queue, list) {
2859                         if (lock->ml.node != dlm->node_num) {
2860                                 mlog(0, "putting lock for node %u\n",
2861                                      lock->ml.node);
2862                                 /* be extra careful */
2863                                 BUG_ON(!list_empty(&lock->ast_list));
2864                                 BUG_ON(!list_empty(&lock->bast_list));
2865                                 BUG_ON(lock->ast_pending);
2866                                 BUG_ON(lock->bast_pending);
2867                                 dlm_lockres_clear_refmap_bit(lock->ml.node, res);
2868                                 list_del_init(&lock->list);
2869                                 dlm_lock_put(lock);
2870                                 /* In a normal unlock, we would have added a
2871                                  * DLM_UNLOCK_FREE_LOCK action. Force it. */
2872                                 dlm_lock_put(lock);
2873                         }
2874                 }
2875                 queue++;
2876         }
2877         bit = 0;
2878         while (1) {
2879                 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2880                 if (bit >= O2NM_MAX_NODES)
2881                         break;
2882                 /* do not clear the local node reference, if there is a
2883                  * process holding this, let it drop the ref itself */
2884                 if (bit != dlm->node_num) {
2885                         mlog(0, "%s:%.*s: node %u had a ref to this "
2886                              "migrating lockres, clearing\n", dlm->name,
2887                              res->lockname.len, res->lockname.name, bit);
2888                         dlm_lockres_clear_refmap_bit(bit, res);
2889                 }
2890                 bit++;
2891         }
2892 }
2893
2894 /* for now this is not too intelligent.  we will
2895  * need stats to make this do the right thing.
2896  * this just finds the first lock on one of the
2897  * queues and uses that node as the target. */
2898 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2899                                     struct dlm_lock_resource *res)
2900 {
2901         int i;
2902         struct list_head *queue = &res->granted;
2903         struct dlm_lock *lock;
2904         int nodenum;
2905
2906         assert_spin_locked(&dlm->spinlock);
2907
2908         spin_lock(&res->spinlock);
2909         for (i=0; i<3; i++) {
2910                 list_for_each_entry(lock, queue, list) {
2911                         /* up to the caller to make sure this node
2912                          * is alive */
2913                         if (lock->ml.node != dlm->node_num) {
2914                                 spin_unlock(&res->spinlock);
2915                                 return lock->ml.node;
2916                         }
2917                 }
2918                 queue++;
2919         }
2920         spin_unlock(&res->spinlock);
2921         mlog(0, "have not found a suitable target yet! checking domain map\n");
2922
2923         /* ok now we're getting desperate.  pick anyone alive. */
2924         nodenum = -1;
2925         while (1) {
2926                 nodenum = find_next_bit(dlm->domain_map,
2927                                         O2NM_MAX_NODES, nodenum+1);
2928                 mlog(0, "found %d in domain map\n", nodenum);
2929                 if (nodenum >= O2NM_MAX_NODES)
2930                         break;
2931                 if (nodenum != dlm->node_num) {
2932                         mlog(0, "picking %d\n", nodenum);
2933                         return nodenum;
2934                 }
2935         }
2936
2937         mlog(0, "giving up.  no master to migrate to\n");
2938         return DLM_LOCK_RES_OWNER_UNKNOWN;
2939 }
2940
2941
2942
2943 /* this is called by the new master once all lockres
2944  * data has been received */
2945 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2946                                   struct dlm_lock_resource *res,
2947                                   u8 master, u8 new_master,
2948                                   struct dlm_node_iter *iter)
2949 {
2950         struct dlm_migrate_request migrate;
2951         int ret, skip, status = 0;
2952         int nodenum;
2953
2954         memset(&migrate, 0, sizeof(migrate));
2955         migrate.namelen = res->lockname.len;
2956         memcpy(migrate.name, res->lockname.name, migrate.namelen);
2957         migrate.new_master = new_master;
2958         migrate.master = master;
2959
2960         ret = 0;
2961
2962         /* send message to all nodes, except the master and myself */
2963         while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2964                 if (nodenum == master ||
2965                     nodenum == new_master)
2966                         continue;
2967
2968                 /* We could race exit domain. If exited, skip. */
2969                 spin_lock(&dlm->spinlock);
2970                 skip = (!test_bit(nodenum, dlm->domain_map));
2971                 spin_unlock(&dlm->spinlock);
2972                 if (skip) {
2973                         clear_bit(nodenum, iter->node_map);
2974                         continue;
2975                 }
2976
2977                 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2978                                          &migrate, sizeof(migrate), nodenum,
2979                                          &status);
2980                 if (ret < 0) {
2981                         mlog(0, "migrate_request returned %d!\n", ret);
2982                         if (!dlm_is_host_down(ret)) {
2983                                 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
2984                                 BUG();
2985                         }
2986                         clear_bit(nodenum, iter->node_map);
2987                         ret = 0;
2988                 } else if (status < 0) {
2989                         mlog(0, "migrate request (node %u) returned %d!\n",
2990                              nodenum, status);
2991                         ret = status;
2992                 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
2993                         /* during the migration request we short-circuited
2994                          * the mastery of the lockres.  make sure we have
2995                          * a mastery ref for nodenum */
2996                         mlog(0, "%s:%.*s: need ref for node %u\n",
2997                              dlm->name, res->lockname.len, res->lockname.name,
2998                              nodenum);
2999                         spin_lock(&res->spinlock);
3000                         dlm_lockres_set_refmap_bit(nodenum, res);
3001                         spin_unlock(&res->spinlock);
3002                 }
3003         }
3004
3005         if (ret < 0)
3006                 mlog_errno(ret);
3007
3008         mlog(0, "returning ret=%d\n", ret);
3009         return ret;
3010 }
3011
3012
3013 /* if there is an existing mle for this lockres, we now know who the master is.
3014  * (the one who sent us *this* message) we can clear it up right away.
3015  * since the process that put the mle on the list still has a reference to it,
3016  * we can unhash it now, set the master and wake the process.  as a result,
3017  * we will have no mle in the list to start with.  now we can add an mle for
3018  * the migration and this should be the only one found for those scanning the
3019  * list.  */
3020 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
3021                                 void **ret_data)
3022 {
3023         struct dlm_ctxt *dlm = data;
3024         struct dlm_lock_resource *res = NULL;
3025         struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
3026         struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
3027         const char *name;
3028         unsigned int namelen, hash;
3029         int ret = 0;
3030
3031         if (!dlm_grab(dlm))
3032                 return -EINVAL;
3033
3034         name = migrate->name;
3035         namelen = migrate->namelen;
3036         hash = dlm_lockid_hash(name, namelen);
3037
3038         /* preallocate.. if this fails, abort */
3039         mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
3040                                                          GFP_NOFS);
3041
3042         if (!mle) {
3043                 ret = -ENOMEM;
3044                 goto leave;
3045         }
3046
3047         /* check for pre-existing lock */
3048         spin_lock(&dlm->spinlock);
3049         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
3050         spin_lock(&dlm->master_lock);
3051
3052         if (res) {
3053                 spin_lock(&res->spinlock);
3054                 if (res->state & DLM_LOCK_RES_RECOVERING) {
3055                         /* if all is working ok, this can only mean that we got
3056                         * a migrate request from a node that we now see as
3057                         * dead.  what can we do here?  drop it to the floor? */
3058                         spin_unlock(&res->spinlock);
3059                         mlog(ML_ERROR, "Got a migrate request, but the "
3060                              "lockres is marked as recovering!");
3061                         kmem_cache_free(dlm_mle_cache, mle);
3062                         ret = -EINVAL; /* need a better solution */
3063                         goto unlock;
3064                 }
3065                 res->state |= DLM_LOCK_RES_MIGRATING;
3066                 spin_unlock(&res->spinlock);
3067         }
3068
3069         /* ignore status.  only nonzero status would BUG. */
3070         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3071                                     name, namelen,
3072                                     migrate->new_master,
3073                                     migrate->master);
3074
3075 unlock:
3076         spin_unlock(&dlm->master_lock);
3077         spin_unlock(&dlm->spinlock);
3078
3079         if (oldmle) {
3080                 /* master is known, detach if not already detached */
3081                 dlm_mle_detach_hb_events(dlm, oldmle);
3082                 dlm_put_mle(oldmle);
3083         }
3084
3085         if (res)
3086                 dlm_lockres_put(res);
3087 leave:
3088         dlm_put(dlm);
3089         return ret;
3090 }
3091
3092 /* must be holding dlm->spinlock and dlm->master_lock
3093  * when adding a migration mle, we can clear any other mles
3094  * in the master list because we know with certainty that
3095  * the master is "master".  so we remove any old mle from
3096  * the list after setting it's master field, and then add
3097  * the new migration mle.  this way we can hold with the rule
3098  * of having only one mle for a given lock name at all times. */
3099 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3100                                  struct dlm_lock_resource *res,
3101                                  struct dlm_master_list_entry *mle,
3102                                  struct dlm_master_list_entry **oldmle,
3103                                  const char *name, unsigned int namelen,
3104                                  u8 new_master, u8 master)
3105 {
3106         int found;
3107         int ret = 0;
3108
3109         *oldmle = NULL;
3110
3111         mlog_entry_void();
3112
3113         assert_spin_locked(&dlm->spinlock);
3114         assert_spin_locked(&dlm->master_lock);
3115
3116         /* caller is responsible for any ref taken here on oldmle */
3117         found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3118         if (found) {
3119                 struct dlm_master_list_entry *tmp = *oldmle;
3120                 spin_lock(&tmp->spinlock);
3121                 if (tmp->type == DLM_MLE_MIGRATION) {
3122                         if (master == dlm->node_num) {
3123                                 /* ah another process raced me to it */
3124                                 mlog(0, "tried to migrate %.*s, but some "
3125                                      "process beat me to it\n",
3126                                      namelen, name);
3127                                 ret = -EEXIST;
3128                         } else {
3129                                 /* bad.  2 NODES are trying to migrate! */
3130                                 mlog(ML_ERROR, "migration error  mle: "
3131                                      "master=%u new_master=%u // request: "
3132                                      "master=%u new_master=%u // "
3133                                      "lockres=%.*s\n",
3134                                      tmp->master, tmp->new_master,
3135                                      master, new_master,
3136                                      namelen, name);
3137                                 BUG();
3138                         }
3139                 } else {
3140                         /* this is essentially what assert_master does */
3141                         tmp->master = master;
3142                         atomic_set(&tmp->woken, 1);
3143                         wake_up(&tmp->wq);
3144                         /* remove it so that only one mle will be found */
3145                         __dlm_unlink_mle(dlm, tmp);
3146                         __dlm_mle_detach_hb_events(dlm, tmp);
3147                         ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3148                         mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3149                             "telling master to get ref for cleared out mle "
3150                             "during migration\n", dlm->name, namelen, name,
3151                             master, new_master);
3152                 }
3153                 spin_unlock(&tmp->spinlock);
3154         }
3155
3156         /* now add a migration mle to the tail of the list */
3157         dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3158         mle->new_master = new_master;
3159         /* the new master will be sending an assert master for this.
3160          * at that point we will get the refmap reference */
3161         mle->master = master;
3162         /* do this for consistency with other mle types */
3163         set_bit(new_master, mle->maybe_map);
3164         __dlm_insert_mle(dlm, mle);
3165
3166         return ret;
3167 }
3168
3169 /*
3170  * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3171  */
3172 static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3173                                         struct dlm_master_list_entry *mle)
3174 {
3175         struct dlm_lock_resource *res;
3176
3177         /* Find the lockres associated to the mle and set its owner to UNK */
3178         res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3179                                    mle->mnamehash);
3180         if (res) {
3181                 spin_unlock(&dlm->master_lock);
3182
3183                 /* move lockres onto recovery list */
3184                 spin_lock(&res->spinlock);
3185                 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3186                 dlm_move_lockres_to_recovery_list(dlm, res);
3187                 spin_unlock(&res->spinlock);
3188                 dlm_lockres_put(res);
3189
3190                 /* about to get rid of mle, detach from heartbeat */
3191                 __dlm_mle_detach_hb_events(dlm, mle);
3192
3193                 /* dump the mle */
3194                 spin_lock(&dlm->master_lock);
3195                 __dlm_put_mle(mle);
3196                 spin_unlock(&dlm->master_lock);
3197         }
3198
3199         return res;
3200 }
3201
3202 static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3203                                     struct dlm_master_list_entry *mle)
3204 {
3205         __dlm_mle_detach_hb_events(dlm, mle);
3206
3207         spin_lock(&mle->spinlock);
3208         __dlm_unlink_mle(dlm, mle);
3209         atomic_set(&mle->woken, 1);
3210         spin_unlock(&mle->spinlock);
3211
3212         wake_up(&mle->wq);
3213 }
3214
3215 static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3216                                 struct dlm_master_list_entry *mle, u8 dead_node)
3217 {
3218         int bit;
3219
3220         BUG_ON(mle->type != DLM_MLE_BLOCK);
3221
3222         spin_lock(&mle->spinlock);
3223         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3224         if (bit != dead_node) {
3225                 mlog(0, "mle found, but dead node %u would not have been "
3226                      "master\n", dead_node);
3227                 spin_unlock(&mle->spinlock);
3228         } else {
3229                 /* Must drop the refcount by one since the assert_master will
3230                  * never arrive. This may result in the mle being unlinked and
3231                  * freed, but there may still be a process waiting in the
3232                  * dlmlock path which is fine. */
3233                 mlog(0, "node %u was expected master\n", dead_node);
3234                 atomic_set(&mle->woken, 1);
3235                 spin_unlock(&mle->spinlock);
3236                 wake_up(&mle->wq);
3237
3238                 /* Do not need events any longer, so detach from heartbeat */
3239                 __dlm_mle_detach_hb_events(dlm, mle);
3240                 __dlm_put_mle(mle);
3241         }
3242 }
3243
3244 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3245 {
3246         struct dlm_master_list_entry *mle;
3247         struct dlm_lock_resource *res;
3248         struct hlist_head *bucket;
3249         struct hlist_node *list;
3250         unsigned int i;
3251
3252         mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node);
3253 top:
3254         assert_spin_locked(&dlm->spinlock);
3255
3256         /* clean the master list */
3257         spin_lock(&dlm->master_lock);
3258         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3259                 bucket = dlm_master_hash(dlm, i);
3260                 hlist_for_each(list, bucket) {
3261                         mle = hlist_entry(list, struct dlm_master_list_entry,
3262                                           master_hash_node);
3263
3264                         BUG_ON(mle->type != DLM_MLE_BLOCK &&
3265                                mle->type != DLM_MLE_MASTER &&
3266                                mle->type != DLM_MLE_MIGRATION);
3267
3268                         /* MASTER mles are initiated locally. The waiting
3269                          * process will notice the node map change shortly.
3270                          * Let that happen as normal. */
3271                         if (mle->type == DLM_MLE_MASTER)
3272                                 continue;
3273
3274                         /* BLOCK mles are initiated by other nodes. Need to
3275                          * clean up if the dead node would have been the
3276                          * master. */
3277                         if (mle->type == DLM_MLE_BLOCK) {
3278                                 dlm_clean_block_mle(dlm, mle, dead_node);
3279                                 continue;
3280                         }
3281
3282                         /* Everything else is a MIGRATION mle */
3283
3284                         /* The rule for MIGRATION mles is that the master
3285                          * becomes UNKNOWN if *either* the original or the new
3286                          * master dies. All UNKNOWN lockres' are sent to
3287                          * whichever node becomes the recovery master. The new
3288                          * master is responsible for determining if there is
3289                          * still a master for this lockres, or if he needs to
3290                          * take over mastery. Either way, this node should
3291                          * expect another message to resolve this. */
3292
3293                         if (mle->master != dead_node &&
3294                             mle->new_master != dead_node)
3295                                 continue;
3296
3297                         /* If we have reached this point, this mle needs to be
3298                          * removed from the list and freed. */
3299                         dlm_clean_migration_mle(dlm, mle);
3300
3301                         mlog(0, "%s: node %u died during migration from "
3302                              "%u to %u!\n", dlm->name, dead_node, mle->master,
3303                              mle->new_master);
3304
3305                         /* If we find a lockres associated with the mle, we've
3306                          * hit this rare case that messes up our lock ordering.
3307                          * If so, we need to drop the master lock so that we can
3308                          * take the lockres lock, meaning that we will have to
3309                          * restart from the head of list. */
3310                         res = dlm_reset_mleres_owner(dlm, mle);
3311                         if (res)
3312                                 /* restart */
3313                                 goto top;
3314
3315                         /* This may be the last reference */
3316                         __dlm_put_mle(mle);
3317                 }
3318         }
3319         spin_unlock(&dlm->master_lock);
3320 }
3321
3322 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3323                          u8 old_master)
3324 {
3325         struct dlm_node_iter iter;
3326         int ret = 0;
3327
3328         spin_lock(&dlm->spinlock);
3329         dlm_node_iter_init(dlm->domain_map, &iter);
3330         clear_bit(old_master, iter.node_map);
3331         clear_bit(dlm->node_num, iter.node_map);
3332         spin_unlock(&dlm->spinlock);
3333
3334         /* ownership of the lockres is changing.  account for the
3335          * mastery reference here since old_master will briefly have
3336          * a reference after the migration completes */
3337         spin_lock(&res->spinlock);
3338         dlm_lockres_set_refmap_bit(old_master, res);
3339         spin_unlock(&res->spinlock);
3340
3341         mlog(0, "now time to do a migrate request to other nodes\n");
3342         ret = dlm_do_migrate_request(dlm, res, old_master,
3343                                      dlm->node_num, &iter);
3344         if (ret < 0) {
3345                 mlog_errno(ret);
3346                 goto leave;
3347         }
3348
3349         mlog(0, "doing assert master of %.*s to all except the original node\n",
3350              res->lockname.len, res->lockname.name);
3351         /* this call now finishes out the nodemap
3352          * even if one or more nodes die */
3353         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3354                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3355         if (ret < 0) {
3356                 /* no longer need to retry.  all living nodes contacted. */
3357                 mlog_errno(ret);
3358                 ret = 0;
3359         }
3360
3361         memset(iter.node_map, 0, sizeof(iter.node_map));
3362         set_bit(old_master, iter.node_map);
3363         mlog(0, "doing assert master of %.*s back to %u\n",
3364              res->lockname.len, res->lockname.name, old_master);
3365         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3366                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3367         if (ret < 0) {
3368                 mlog(0, "assert master to original master failed "
3369                      "with %d.\n", ret);
3370                 /* the only nonzero status here would be because of
3371                  * a dead original node.  we're done. */
3372                 ret = 0;
3373         }
3374
3375         /* all done, set the owner, clear the flag */
3376         spin_lock(&res->spinlock);
3377         dlm_set_lockres_owner(dlm, res, dlm->node_num);
3378         res->state &= ~DLM_LOCK_RES_MIGRATING;
3379         spin_unlock(&res->spinlock);
3380         /* re-dirty it on the new master */
3381         dlm_kick_thread(dlm, res);
3382         wake_up(&res->wq);
3383 leave:
3384         return ret;
3385 }
3386
3387 /*
3388  * LOCKRES AST REFCOUNT
3389  * this is integral to migration
3390  */
3391
3392 /* for future intent to call an ast, reserve one ahead of time.
3393  * this should be called only after waiting on the lockres
3394  * with dlm_wait_on_lockres, and while still holding the
3395  * spinlock after the call. */
3396 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3397 {
3398         assert_spin_locked(&res->spinlock);
3399         if (res->state & DLM_LOCK_RES_MIGRATING) {
3400                 __dlm_print_one_lock_resource(res);
3401         }
3402         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3403
3404         atomic_inc(&res->asts_reserved);
3405 }
3406
3407 /*
3408  * used to drop the reserved ast, either because it went unused,
3409  * or because the ast/bast was actually called.
3410  *
3411  * also, if there is a pending migration on this lockres,
3412  * and this was the last pending ast on the lockres,
3413  * atomically set the MIGRATING flag before we drop the lock.
3414  * this is how we ensure that migration can proceed with no
3415  * asts in progress.  note that it is ok if the state of the
3416  * queues is such that a lock should be granted in the future
3417  * or that a bast should be fired, because the new master will
3418  * shuffle the lists on this lockres as soon as it is migrated.
3419  */
3420 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3421                              struct dlm_lock_resource *res)
3422 {
3423         if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3424                 return;
3425
3426         if (!res->migration_pending) {
3427                 spin_unlock(&res->spinlock);
3428                 return;
3429         }
3430
3431         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3432         res->migration_pending = 0;
3433         res->state |= DLM_LOCK_RES_MIGRATING;
3434         spin_unlock(&res->spinlock);
3435         wake_up(&res->wq);
3436         wake_up(&dlm->migration_wq);
3437 }