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