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