fs/ocfs2/dlm/dlmmaster.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmmod.c
   5  *
   6  * standalone DLM module
   7  *
   8  * Copyright (C) 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  *
  25  */
  26
  27
  28 #include <linux/module.h>
  29 #include <linux/fs.h>
  30 #include <linux/types.h>
  31 #include <linux/slab.h>
  32 #include <linux/highmem.h>
  33 #include <linux/init.h>
  34 #include <linux/sysctl.h>
  35 #include <linux/random.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/socket.h>
  38 #include <linux/inet.h>
  39 #include <linux/spinlock.h>
  40 #include <linux/delay.h>
  41
  42
  43 #include "cluster/heartbeat.h"
  44 #include "cluster/nodemanager.h"
  45 #include "cluster/tcp.h"
  46
  47 #include "dlmapi.h"
  48 #include "dlmcommon.h"
  49 #include "dlmdomain.h"
  50 #include "dlmdebug.h"
  51
  52 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
  53 #include "cluster/masklog.h"
  54
  55 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
  56                               struct dlm_master_list_entry *mle,
  57                               struct o2nm_node *node,
  58                               int idx);
  59 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
  60                             struct dlm_master_list_entry *mle,
  61                             struct o2nm_node *node,
  62                             int idx);
  63
  64 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
  65 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
  66                                 struct dlm_lock_resource *res,
  67                                 void *nodemap, u32 flags);
  68 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
  69
  70 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
  71                                 struct dlm_master_list_entry *mle,
  72                                 const char *name,
  73                                 unsigned int namelen)
  74 {
  75         if (dlm != mle->dlm)
  76                 return 0;
  77
  78         if (namelen != mle->mnamelen ||
  79             memcmp(name, mle->mname, namelen) != 0)
  80                 return 0;
  81
  82         return 1;
  83 }
  84
  85 static struct kmem_cache *dlm_lockres_cache = NULL;
  86 static struct kmem_cache *dlm_lockname_cache = NULL;
  87 static struct kmem_cache *dlm_mle_cache = NULL;
  88
  89 static void dlm_mle_release(struct kref *kref);
  90 static void dlm_init_mle(struct dlm_master_list_entry *mle,
  91                         enum dlm_mle_type type,
  92                         struct dlm_ctxt *dlm,
  93                         struct dlm_lock_resource *res,
  94                         const char *name,
  95                         unsigned int namelen);
  96 static void dlm_put_mle(struct dlm_master_list_entry *mle);
  97 static void __dlm_put_mle(struct dlm_master_list_entry *mle);
  98 static int dlm_find_mle(struct dlm_ctxt *dlm,
  99                         struct dlm_master_list_entry **mle,
 100                         char *name, unsigned int namelen);
 101
 102 static int dlm_do_master_request(struct dlm_lock_resource *res,
 103                                  struct dlm_master_list_entry *mle, int to);
 104
 105
 106 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
 107                                      struct dlm_lock_resource *res,
 108                                      struct dlm_master_list_entry *mle,
 109                                      int *blocked);
 110 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
 111                                     struct dlm_lock_resource *res,
 112                                     struct dlm_master_list_entry *mle,
 113                                     int blocked);
 114 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
 115                                  struct dlm_lock_resource *res,
 116                                  struct dlm_master_list_entry *mle,
 117                                  struct dlm_master_list_entry **oldmle,
 118                                  const char *name, unsigned int namelen,
 119                                  u8 new_master, u8 master);
 120
 121 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
 122                                     struct dlm_lock_resource *res);
 123 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
 124                                       struct dlm_lock_resource *res);
 125 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
 126                                        struct dlm_lock_resource *res,
 127                                        u8 target);
 128 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
 129                                        struct dlm_lock_resource *res);
 130
 131
 132 int dlm_is_host_down(int errno)
 133 {
 134         switch (errno) {
 135                 case -EBADF:
 136                 case -ECONNREFUSED:
 137                 case -ENOTCONN:
 138                 case -ECONNRESET:
 139                 case -EPIPE:
 140                 case -EHOSTDOWN:
 141                 case -EHOSTUNREACH:
 142                 case -ETIMEDOUT:
 143                 case -ECONNABORTED:
 144                 case -ENETDOWN:
 145                 case -ENETUNREACH:
 146                 case -ENETRESET:
 147                 case -ESHUTDOWN:
 148                 case -ENOPROTOOPT:
 149                 case -EINVAL:   /* if returned from our tcp code,
 150                                    this means there is no socket */
 151                         return 1;
 152         }
 153         return 0;
 154 }
 155
 156
 157 /*
 158  * MASTER LIST FUNCTIONS
 159  */
 160
 161
 162 /*
 163  * regarding master list entries and heartbeat callbacks:
 164  *
 165  * in order to avoid sleeping and allocation that occurs in
 166  * heartbeat, master list entries are simply attached to the
 167  * dlm's established heartbeat callbacks.  the mle is attached
 168  * when it is created, and since the dlm->spinlock is held at
 169  * that time, any heartbeat event will be properly discovered
 170  * by the mle.  the mle needs to be detached from the
 171  * dlm->mle_hb_events list as soon as heartbeat events are no
 172  * longer useful to the mle, and before the mle is freed.
 173  *
 174  * as a general rule, heartbeat events are no longer needed by
 175  * the mle once an "answer" regarding the lock master has been
 176  * received.
 177  */
 178 static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
 179                                               struct dlm_master_list_entry *mle)
 180 {
 181         assert_spin_locked(&dlm->spinlock);
 182
 183         list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
 184 }
 185
 186
 187 static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 188                                               struct dlm_master_list_entry *mle)
 189 {
 190         if (!list_empty(&mle->hb_events))
 191                 list_del_init(&mle->hb_events);
 192 }
 193
 194
 195 static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 196                                             struct dlm_master_list_entry *mle)
 197 {
 198         spin_lock(&dlm->spinlock);
 199         __dlm_mle_detach_hb_events(dlm, mle);
 200         spin_unlock(&dlm->spinlock);
 201 }
 202
 203 static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
 204 {
 205         struct dlm_ctxt *dlm;
 206         dlm = mle->dlm;
 207
 208         assert_spin_locked(&dlm->spinlock);
 209         assert_spin_locked(&dlm->master_lock);
 210         mle->inuse++;
 211         kref_get(&mle->mle_refs);
 212 }
 213
 214 static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
 215 {
 216         struct dlm_ctxt *dlm;
 217         dlm = mle->dlm;
 218
 219         spin_lock(&dlm->spinlock);
 220         spin_lock(&dlm->master_lock);
 221         mle->inuse--;
 222         __dlm_put_mle(mle);
 223         spin_unlock(&dlm->master_lock);
 224         spin_unlock(&dlm->spinlock);
 225
 226 }
 227
 228 /* remove from list and free */
 229 static void __dlm_put_mle(struct dlm_master_list_entry *mle)
 230 {
 231         struct dlm_ctxt *dlm;
 232         dlm = mle->dlm;
 233
 234         assert_spin_locked(&dlm->spinlock);
 235         assert_spin_locked(&dlm->master_lock);
 236         if (!atomic_read(&mle->mle_refs.refcount)) {
 237                 /* this may or may not crash, but who cares.
 238                  * it's a BUG. */
 239                 mlog(ML_ERROR, "bad mle: %p\n", mle);
 240                 dlm_print_one_mle(mle);
 241                 BUG();
 242         } else
 243                 kref_put(&mle->mle_refs, dlm_mle_release);
 244 }
 245
 246
 247 /* must not have any spinlocks coming in */
 248 static void dlm_put_mle(struct dlm_master_list_entry *mle)
 249 {
 250         struct dlm_ctxt *dlm;
 251         dlm = mle->dlm;
 252
 253         spin_lock(&dlm->spinlock);
 254         spin_lock(&dlm->master_lock);
 255         __dlm_put_mle(mle);
 256         spin_unlock(&dlm->master_lock);
 257         spin_unlock(&dlm->spinlock);
 258 }
 259
 260 static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
 261 {
 262         kref_get(&mle->mle_refs);
 263 }
 264
 265 static void dlm_init_mle(struct dlm_master_list_entry *mle,
 266                         enum dlm_mle_type type,
 267                         struct dlm_ctxt *dlm,
 268                         struct dlm_lock_resource *res,
 269                         const char *name,
 270                         unsigned int namelen)
 271 {
 272         assert_spin_locked(&dlm->spinlock);
 273
 274         mle->dlm = dlm;
 275         mle->type = type;
 276         INIT_HLIST_NODE(&mle->master_hash_node);
 277         INIT_LIST_HEAD(&mle->hb_events);
 278         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
 279         spin_lock_init(&mle->spinlock);
 280         init_waitqueue_head(&mle->wq);
 281         atomic_set(&mle->woken, 0);
 282         kref_init(&mle->mle_refs);
 283         memset(mle->response_map, 0, sizeof(mle->response_map));
 284         mle->master = O2NM_MAX_NODES;
 285         mle->new_master = O2NM_MAX_NODES;
 286         mle->inuse = 0;
 287
 288         BUG_ON(mle->type != DLM_MLE_BLOCK &&
 289                mle->type != DLM_MLE_MASTER &&
 290                mle->type != DLM_MLE_MIGRATION);
 291
 292         if (mle->type == DLM_MLE_MASTER) {
 293                 BUG_ON(!res);
 294                 mle->mleres = res;
 295                 memcpy(mle->mname, res->lockname.name, res->lockname.len);
 296                 mle->mnamelen = res->lockname.len;
 297                 mle->mnamehash = res->lockname.hash;
 298         } else {
 299                 BUG_ON(!name);
 300                 mle->mleres = NULL;
 301                 memcpy(mle->mname, name, namelen);
 302                 mle->mnamelen = namelen;
 303                 mle->mnamehash = dlm_lockid_hash(name, namelen);
 304         }
 305
 306         atomic_inc(&dlm->mle_tot_count[mle->type]);
 307         atomic_inc(&dlm->mle_cur_count[mle->type]);
 308
 309         /* copy off the node_map and register hb callbacks on our copy */
 310         memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
 311         memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
 312         clear_bit(dlm->node_num, mle->vote_map);
 313         clear_bit(dlm->node_num, mle->node_map);
 314
 315         /* attach the mle to the domain node up/down events */
 316         __dlm_mle_attach_hb_events(dlm, mle);
 317 }
 318
 319 void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
 320 {
 321         assert_spin_locked(&dlm->spinlock);
 322         assert_spin_locked(&dlm->master_lock);
 323
 324         if (!hlist_unhashed(&mle->master_hash_node))
 325                 hlist_del_init(&mle->master_hash_node);
 326 }
 327
 328 void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
 329 {
 330         struct hlist_head *bucket;
 331
 332         assert_spin_locked(&dlm->master_lock);
 333
 334         bucket = dlm_master_hash(dlm, mle->mnamehash);
 335         hlist_add_head(&mle->master_hash_node, bucket);
 336 }
 337
 338 /* returns 1 if found, 0 if not */
 339 static int dlm_find_mle(struct dlm_ctxt *dlm,
 340                         struct dlm_master_list_entry **mle,
 341                         char *name, unsigned int namelen)
 342 {
 343         struct dlm_master_list_entry *tmpmle;
 344         struct hlist_head *bucket;
 345         struct hlist_node *list;
 346         unsigned int hash;
 347
 348         assert_spin_locked(&dlm->master_lock);
 349
 350         hash = dlm_lockid_hash(name, namelen);
 351         bucket = dlm_master_hash(dlm, hash);
 352         hlist_for_each(list, bucket) {
 353                 tmpmle = hlist_entry(list, struct dlm_master_list_entry,
 354                                      master_hash_node);
 355                 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
 356                         continue;
 357                 dlm_get_mle(tmpmle);
 358                 *mle = tmpmle;
 359                 return 1;
 360         }
 361         return 0;
 362 }
 363
 364 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
 365 {
 366         struct dlm_master_list_entry *mle;
 367
 368         assert_spin_locked(&dlm->spinlock);
 369
 370         list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
 371                 if (node_up)
 372                         dlm_mle_node_up(dlm, mle, NULL, idx);
 373                 else
 374                         dlm_mle_node_down(dlm, mle, NULL, idx);
 375         }
 376 }
 377
 378 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
 379                               struct dlm_master_list_entry *mle,
 380                               struct o2nm_node *node, int idx)
 381 {
 382         spin_lock(&mle->spinlock);
 383
 384         if (!test_bit(idx, mle->node_map))
 385                 mlog(0, "node %u already removed from nodemap!\n", idx);
 386         else
 387                 clear_bit(idx, mle->node_map);
 388
 389         spin_unlock(&mle->spinlock);
 390 }
 391
 392 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
 393                             struct dlm_master_list_entry *mle,
 394                             struct o2nm_node *node, int idx)
 395 {
 396         spin_lock(&mle->spinlock);
 397
 398         if (test_bit(idx, mle->node_map))
 399                 mlog(0, "node %u already in node map!\n", idx);
 400         else
 401                 set_bit(idx, mle->node_map);
 402
 403         spin_unlock(&mle->spinlock);
 404 }
 405
 406
 407 int dlm_init_mle_cache(void)
 408 {
 409         dlm_mle_cache = kmem_cache_create("o2dlm_mle",
 410                                           sizeof(struct dlm_master_list_entry),
 411                                           0, SLAB_HWCACHE_ALIGN,
 412                                           NULL);
 413         if (dlm_mle_cache == NULL)
 414                 return -ENOMEM;
 415         return 0;
 416 }
 417
 418 void dlm_destroy_mle_cache(void)
 419 {
 420         if (dlm_mle_cache)
 421                 kmem_cache_destroy(dlm_mle_cache);
 422 }
 423
 424 static void dlm_mle_release(struct kref *kref)
 425 {
 426         struct dlm_master_list_entry *mle;
 427         struct dlm_ctxt *dlm;
 428
 429         mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
 430         dlm = mle->dlm;
 431
 432         assert_spin_locked(&dlm->spinlock);
 433         assert_spin_locked(&dlm->master_lock);
 434
 435         mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
 436              mle->type);
 437
 438         /* remove from list if not already */
 439         __dlm_unlink_mle(dlm, mle);
 440
 441         /* detach the mle from the domain node up/down events */
 442         __dlm_mle_detach_hb_events(dlm, mle);
 443
 444         atomic_dec(&dlm->mle_cur_count[mle->type]);
 445
 446         /* NOTE: kfree under spinlock here.
 447          * if this is bad, we can move this to a freelist. */
 448         kmem_cache_free(dlm_mle_cache, mle);
 449 }
 450
 451
 452 /*
 453  * LOCK RESOURCE FUNCTIONS
 454  */
 455
 456 int dlm_init_master_caches(void)
 457 {
 458         dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
 459                                               sizeof(struct dlm_lock_resource),
 460                                               0, SLAB_HWCACHE_ALIGN, NULL);
 461         if (!dlm_lockres_cache)
 462                 goto bail;
 463
 464         dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
 465                                                DLM_LOCKID_NAME_MAX, 0,
 466                                                SLAB_HWCACHE_ALIGN, NULL);
 467         if (!dlm_lockname_cache)
 468                 goto bail;
 469
 470         return 0;
 471 bail:
 472         dlm_destroy_master_caches();
 473         return -ENOMEM;
 474 }
 475
 476 void dlm_destroy_master_caches(void)
 477 {
 478         if (dlm_lockname_cache)
 479                 kmem_cache_destroy(dlm_lockname_cache);
 480
 481         if (dlm_lockres_cache)
 482                 kmem_cache_destroy(dlm_lockres_cache);
 483 }
 484
 485 static void dlm_lockres_release(struct kref *kref)
 486 {
 487         struct dlm_lock_resource *res;
 488         struct dlm_ctxt *dlm;
 489
 490         res = container_of(kref, struct dlm_lock_resource, refs);
 491         dlm = res->dlm;
 492
 493         /* This should not happen -- all lockres' have a name
 494          * associated with them at init time. */
 495         BUG_ON(!res->lockname.name);
 496
 497         mlog(0, "destroying lockres %.*s\n", res->lockname.len,
 498              res->lockname.name);
 499
 500         spin_lock(&dlm->track_lock);
 501         if (!list_empty(&res->tracking))
 502                 list_del_init(&res->tracking);
 503         else {
 504                 mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n",
 505                      res->lockname.len, res->lockname.name);
 506                 dlm_print_one_lock_resource(res);
 507         }
 508         spin_unlock(&dlm->track_lock);
 509
 510         atomic_dec(&dlm->res_cur_count);
 511
 512         if (!hlist_unhashed(&res->hash_node) ||
 513             !list_empty(&res->granted) ||
 514             !list_empty(&res->converting) ||
 515             !list_empty(&res->blocked) ||
 516             !list_empty(&res->dirty) ||
 517             !list_empty(&res->recovering) ||
 518             !list_empty(&res->purge)) {
 519                 mlog(ML_ERROR,
 520                      "Going to BUG for resource %.*s."
 521                      "  We're on a list! [%c%c%c%c%c%c%c]\n",
 522                      res->lockname.len, res->lockname.name,
 523                      !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
 524                      !list_empty(&res->granted) ? 'G' : ' ',
 525                      !list_empty(&res->converting) ? 'C' : ' ',
 526                      !list_empty(&res->blocked) ? 'B' : ' ',
 527                      !list_empty(&res->dirty) ? 'D' : ' ',
 528                      !list_empty(&res->recovering) ? 'R' : ' ',
 529                      !list_empty(&res->purge) ? 'P' : ' ');
 530
 531                 dlm_print_one_lock_resource(res);
 532         }
 533
 534         /* By the time we're ready to blow this guy away, we shouldn't
 535          * be on any lists. */
 536         BUG_ON(!hlist_unhashed(&res->hash_node));
 537         BUG_ON(!list_empty(&res->granted));
 538         BUG_ON(!list_empty(&res->converting));
 539         BUG_ON(!list_empty(&res->blocked));
 540         BUG_ON(!list_empty(&res->dirty));
 541         BUG_ON(!list_empty(&res->recovering));
 542         BUG_ON(!list_empty(&res->purge));
 543
 544         kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
 545
 546         kmem_cache_free(dlm_lockres_cache, res);
 547 }
 548
 549 void dlm_lockres_put(struct dlm_lock_resource *res)
 550 {
 551         kref_put(&res->refs, dlm_lockres_release);
 552 }
 553
 554 static void dlm_init_lockres(struct dlm_ctxt *dlm,
 555                              struct dlm_lock_resource *res,
 556                              const char *name, unsigned int namelen)
 557 {
 558         char *qname;
 559
 560         /* If we memset here, we lose our reference to the kmalloc'd
 561          * res->lockname.name, so be sure to init every field
 562          * correctly! */
 563
 564         qname = (char *) res->lockname.name;
 565         memcpy(qname, name, namelen);
 566
 567         res->lockname.len = namelen;
 568         res->lockname.hash = dlm_lockid_hash(name, namelen);
 569
 570         init_waitqueue_head(&res->wq);
 571         spin_lock_init(&res->spinlock);
 572         INIT_HLIST_NODE(&res->hash_node);
 573         INIT_LIST_HEAD(&res->granted);
 574         INIT_LIST_HEAD(&res->converting);
 575         INIT_LIST_HEAD(&res->blocked);
 576         INIT_LIST_HEAD(&res->dirty);
 577         INIT_LIST_HEAD(&res->recovering);
 578         INIT_LIST_HEAD(&res->purge);
 579         INIT_LIST_HEAD(&res->tracking);
 580         atomic_set(&res->asts_reserved, 0);
 581         res->migration_pending = 0;
 582         res->inflight_locks = 0;
 583
 584         res->dlm = dlm;
 585
 586         kref_init(&res->refs);
 587
 588         atomic_inc(&dlm->res_tot_count);
 589         atomic_inc(&dlm->res_cur_count);
 590
 591         /* just for consistency */
 592         spin_lock(&res->spinlock);
 593         dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
 594         spin_unlock(&res->spinlock);
 595
 596         res->state = DLM_LOCK_RES_IN_PROGRESS;
 597
 598         res->last_used = 0;
 599
 600         spin_lock(&dlm->spinlock);
 601         list_add_tail(&res->tracking, &dlm->tracking_list);
 602         spin_unlock(&dlm->spinlock);
 603
 604         memset(res->lvb, 0, DLM_LVB_LEN);
 605         memset(res->refmap, 0, sizeof(res->refmap));
 606 }
 607
 608 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
 609                                    const char *name,
 610                                    unsigned int namelen)
 611 {
 612         struct dlm_lock_resource *res = NULL;
 613
 614         res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
 615         if (!res)
 616                 goto error;
 617
 618         res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
 619         if (!res->lockname.name)
 620                 goto error;
 621
 622         dlm_init_lockres(dlm, res, name, namelen);
 623         return res;
 624
 625 error:
 626         if (res && res->lockname.name)
 627                 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
 628
 629         if (res)
 630                 kmem_cache_free(dlm_lockres_cache, res);
 631         return NULL;
 632 }
 633
 634 void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
 635                                    struct dlm_lock_resource *res,
 636                                    int new_lockres,
 637                                    const char *file,
 638                                    int line)
 639 {
 640         if (!new_lockres)
 641                 assert_spin_locked(&res->spinlock);
 642
 643         if (!test_bit(dlm->node_num, res->refmap)) {
 644                 BUG_ON(res->inflight_locks != 0);
 645                 dlm_lockres_set_refmap_bit(dlm->node_num, res);
 646         }
 647         res->inflight_locks++;
 648         mlog(0, "%s:%.*s: inflight++: now %u\n",
 649              dlm->name, res->lockname.len, res->lockname.name,
 650              res->inflight_locks);
 651 }
 652
 653 void __dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
 654                                    struct dlm_lock_resource *res,
 655                                    const char *file,
 656                                    int line)
 657 {
 658         assert_spin_locked(&res->spinlock);
 659
 660         BUG_ON(res->inflight_locks == 0);
 661         res->inflight_locks--;
 662         mlog(0, "%s:%.*s: inflight--: now %u\n",
 663              dlm->name, res->lockname.len, res->lockname.name,
 664              res->inflight_locks);
 665         if (res->inflight_locks == 0)
 666                 dlm_lockres_clear_refmap_bit(dlm->node_num, res);
 667         wake_up(&res->wq);
 668 }
 669
 670 /*
 671  * lookup a lock resource by name.
 672  * may already exist in the hashtable.
 673  * lockid is null terminated
 674  *
 675  * if not, allocate enough for the lockres and for
 676  * the temporary structure used in doing the mastering.
 677  *
 678  * also, do a lookup in the dlm->master_list to see
 679  * if another node has begun mastering the same lock.
 680  * if so, there should be a block entry in there
 681  * for this name, and we should *not* attempt to master
 682  * the lock here.   need to wait around for that node
 683  * to assert_master (or die).
 684  *
 685  */
 686 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
 687                                           const char *lockid,
 688                                           int namelen,
 689                                           int flags)
 690 {
 691         struct dlm_lock_resource *tmpres=NULL, *res=NULL;
 692         struct dlm_master_list_entry *mle = NULL;
 693         struct dlm_master_list_entry *alloc_mle = NULL;
 694         int blocked = 0;
 695         int ret, nodenum;
 696         struct dlm_node_iter iter;
 697         unsigned int hash;
 698         int tries = 0;
 699         int bit, wait_on_recovery = 0;
 700         int drop_inflight_if_nonlocal = 0;
 701
 702         BUG_ON(!lockid);
 703
 704         hash = dlm_lockid_hash(lockid, namelen);
 705
 706         mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
 707
 708 lookup:
 709         spin_lock(&dlm->spinlock);
 710         tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
 711         if (tmpres) {
 712                 int dropping_ref = 0;
 713
 714                 spin_unlock(&dlm->spinlock);
 715
 716                 spin_lock(&tmpres->spinlock);
 717                 /* We wait for the other thread that is mastering the resource */
 718                 if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
 719                         __dlm_wait_on_lockres(tmpres);
 720                         BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
 721                 }
 722
 723                 if (tmpres->owner == dlm->node_num) {
 724                         BUG_ON(tmpres->state & DLM_LOCK_RES_DROPPING_REF);
 725                         dlm_lockres_grab_inflight_ref(dlm, tmpres);
 726                 } else if (tmpres->state & DLM_LOCK_RES_DROPPING_REF)
 727                         dropping_ref = 1;
 728                 spin_unlock(&tmpres->spinlock);
 729
 730                 /* wait until done messaging the master, drop our ref to allow
 731                  * the lockres to be purged, start over. */
 732                 if (dropping_ref) {
 733                         spin_lock(&tmpres->spinlock);
 734                         __dlm_wait_on_lockres_flags(tmpres, DLM_LOCK_RES_DROPPING_REF);
 735                         spin_unlock(&tmpres->spinlock);
 736                         dlm_lockres_put(tmpres);
 737                         tmpres = NULL;
 738                         goto lookup;
 739                 }
 740
 741                 mlog(0, "found in hash!\n");
 742                 if (res)
 743                         dlm_lockres_put(res);
 744                 res = tmpres;
 745                 goto leave;
 746         }
 747
 748         if (!res) {
 749                 spin_unlock(&dlm->spinlock);
 750                 mlog(0, "allocating a new resource\n");
 751                 /* nothing found and we need to allocate one. */
 752                 alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
 753                 if (!alloc_mle)
 754                         goto leave;
 755                 res = dlm_new_lockres(dlm, lockid, namelen);
 756                 if (!res)
 757                         goto leave;
 758                 goto lookup;
 759         }
 760
 761         mlog(0, "no lockres found, allocated our own: %p\n", res);
 762
 763         if (flags & LKM_LOCAL) {
 764                 /* caller knows it's safe to assume it's not mastered elsewhere
 765                  * DONE!  return right away */
 766                 spin_lock(&res->spinlock);
 767                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
 768                 __dlm_insert_lockres(dlm, res);
 769                 dlm_lockres_grab_inflight_ref(dlm, res);
 770                 spin_unlock(&res->spinlock);
 771                 spin_unlock(&dlm->spinlock);
 772                 /* lockres still marked IN_PROGRESS */
 773                 goto wake_waiters;
 774         }
 775
 776         /* check master list to see if another node has started mastering it */
 777         spin_lock(&dlm->master_lock);
 778
 779         /* if we found a block, wait for lock to be mastered by another node */
 780         blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
 781         if (blocked) {
 782                 int mig;
 783                 if (mle->type == DLM_MLE_MASTER) {
 784                         mlog(ML_ERROR, "master entry for nonexistent lock!\n");
 785                         BUG();
 786                 }
 787                 mig = (mle->type == DLM_MLE_MIGRATION);
 788                 /* if there is a migration in progress, let the migration
 789                  * finish before continuing.  we can wait for the absence
 790                  * of the MIGRATION mle: either the migrate finished or
 791                  * one of the nodes died and the mle was cleaned up.
 792                  * if there is a BLOCK here, but it already has a master
 793                  * set, we are too late.  the master does not have a ref
 794                  * for us in the refmap.  detach the mle and drop it.
 795                  * either way, go back to the top and start over. */
 796                 if (mig || mle->master != O2NM_MAX_NODES) {
 797                         BUG_ON(mig && mle->master == dlm->node_num);
 798                         /* we arrived too late.  the master does not
 799                          * have a ref for us. retry. */
 800                         mlog(0, "%s:%.*s: late on %s\n",
 801                              dlm->name, namelen, lockid,
 802                              mig ?  "MIGRATION" : "BLOCK");
 803                         spin_unlock(&dlm->master_lock);
 804                         spin_unlock(&dlm->spinlock);
 805
 806                         /* master is known, detach */
 807                         if (!mig)
 808                                 dlm_mle_detach_hb_events(dlm, mle);
 809                         dlm_put_mle(mle);
 810                         mle = NULL;
 811                         /* this is lame, but we can't wait on either
 812                          * the mle or lockres waitqueue here */
 813                         if (mig)
 814                                 msleep(100);
 815                         goto lookup;
 816                 }
 817         } else {
 818                 /* go ahead and try to master lock on this node */
 819                 mle = alloc_mle;
 820                 /* make sure this does not get freed below */
 821                 alloc_mle = NULL;
 822                 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
 823                 set_bit(dlm->node_num, mle->maybe_map);
 824                 __dlm_insert_mle(dlm, mle);
 825
 826                 /* still holding the dlm spinlock, check the recovery map
 827                  * to see if there are any nodes that still need to be
 828                  * considered.  these will not appear in the mle nodemap
 829                  * but they might own this lockres.  wait on them. */
 830                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 831                 if (bit < O2NM_MAX_NODES) {
 832                         mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to "
 833                              "recover before lock mastery can begin\n",
 834                              dlm->name, namelen, (char *)lockid, bit);
 835                         wait_on_recovery = 1;
 836                 }
 837         }
 838
 839         /* at this point there is either a DLM_MLE_BLOCK or a
 840          * DLM_MLE_MASTER on the master list, so it's safe to add the
 841          * lockres to the hashtable.  anyone who finds the lock will
 842          * still have to wait on the IN_PROGRESS. */
 843
 844         /* finally add the lockres to its hash bucket */
 845         __dlm_insert_lockres(dlm, res);
 846         /* since this lockres is new it doesn't not require the spinlock */
 847         dlm_lockres_grab_inflight_ref_new(dlm, res);
 848
 849         /* if this node does not become the master make sure to drop
 850          * this inflight reference below */
 851         drop_inflight_if_nonlocal = 1;
 852
 853         /* get an extra ref on the mle in case this is a BLOCK
 854          * if so, the creator of the BLOCK may try to put the last
 855          * ref at this time in the assert master handler, so we
 856          * need an extra one to keep from a bad ptr deref. */
 857         dlm_get_mle_inuse(mle);
 858         spin_unlock(&dlm->master_lock);
 859         spin_unlock(&dlm->spinlock);
 860
 861 redo_request:
 862         while (wait_on_recovery) {
 863                 /* any cluster changes that occurred after dropping the
 864                  * dlm spinlock would be detectable be a change on the mle,
 865                  * so we only need to clear out the recovery map once. */
 866                 if (dlm_is_recovery_lock(lockid, namelen)) {
 867                         mlog(ML_NOTICE, "%s: recovery map is not empty, but "
 868                              "must master $RECOVERY lock now\n", dlm->name);
 869                         if (!dlm_pre_master_reco_lockres(dlm, res))
 870                                 wait_on_recovery = 0;
 871                         else {
 872                                 mlog(0, "%s: waiting 500ms for heartbeat state "
 873                                     "change\n", dlm->name);
 874                                 msleep(500);
 875                         }
 876                         continue;
 877                 }
 878
 879                 dlm_kick_recovery_thread(dlm);
 880                 msleep(1000);
 881                 dlm_wait_for_recovery(dlm);
 882
 883                 spin_lock(&dlm->spinlock);
 884                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 885                 if (bit < O2NM_MAX_NODES) {
 886                         mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to "
 887                              "recover before lock mastery can begin\n",
 888                              dlm->name, namelen, (char *)lockid, bit);
 889                         wait_on_recovery = 1;
 890                 } else
 891                         wait_on_recovery = 0;
 892                 spin_unlock(&dlm->spinlock);
 893
 894                 if (wait_on_recovery)
 895                         dlm_wait_for_node_recovery(dlm, bit, 10000);
 896         }
 897
 898         /* must wait for lock to be mastered elsewhere */
 899         if (blocked)
 900                 goto wait;
 901
 902         ret = -EINVAL;
 903         dlm_node_iter_init(mle->vote_map, &iter);
 904         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
 905                 ret = dlm_do_master_request(res, mle, nodenum);
 906                 if (ret < 0)
 907                         mlog_errno(ret);
 908                 if (mle->master != O2NM_MAX_NODES) {
 909                         /* found a master ! */
 910                         if (mle->master <= nodenum)
 911                                 break;
 912                         /* if our master request has not reached the master
 913                          * yet, keep going until it does.  this is how the
 914                          * master will know that asserts are needed back to
 915                          * the lower nodes. */
 916                         mlog(0, "%s:%.*s: requests only up to %u but master "
 917                              "is %u, keep going\n", dlm->name, namelen,
 918                              lockid, nodenum, mle->master);
 919                 }
 920         }
 921
 922 wait:
 923         /* keep going until the response map includes all nodes */
 924         ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
 925         if (ret < 0) {
 926                 wait_on_recovery = 1;
 927                 mlog(0, "%s:%.*s: node map changed, redo the "
 928                      "master request now, blocked=%d\n",
 929                      dlm->name, res->lockname.len,
 930                      res->lockname.name, blocked);
 931                 if (++tries > 20) {
 932                         mlog(ML_ERROR, "%s:%.*s: spinning on "
 933                              "dlm_wait_for_lock_mastery, blocked=%d\n",
 934                              dlm->name, res->lockname.len,
 935                              res->lockname.name, blocked);
 936                         dlm_print_one_lock_resource(res);
 937                         dlm_print_one_mle(mle);
 938                         tries = 0;
 939                 }
 940                 goto redo_request;
 941         }
 942
 943         mlog(0, "lockres mastered by %u\n", res->owner);
 944         /* make sure we never continue without this */
 945         BUG_ON(res->owner == O2NM_MAX_NODES);
 946
 947         /* master is known, detach if not already detached */
 948         dlm_mle_detach_hb_events(dlm, mle);
 949         dlm_put_mle(mle);
 950         /* put the extra ref */
 951         dlm_put_mle_inuse(mle);
 952
 953 wake_waiters:
 954         spin_lock(&res->spinlock);
 955         if (res->owner != dlm->node_num && drop_inflight_if_nonlocal)
 956                 dlm_lockres_drop_inflight_ref(dlm, res);
 957         res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
 958         spin_unlock(&res->spinlock);
 959         wake_up(&res->wq);
 960
 961 leave:
 962         /* need to free the unused mle */
 963         if (alloc_mle)
 964                 kmem_cache_free(dlm_mle_cache, alloc_mle);
 965
 966         return res;
 967 }
 968
 969
 970 #define DLM_MASTERY_TIMEOUT_MS   5000
 971
 972 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
 973                                      struct dlm_lock_resource *res,
 974                                      struct dlm_master_list_entry *mle,
 975                                      int *blocked)
 976 {
 977         u8 m;
 978         int ret, bit;
 979         int map_changed, voting_done;
 980         int assert, sleep;
 981
 982 recheck:
 983         ret = 0;
 984         assert = 0;
 985
 986         /* check if another node has already become the owner */
 987         spin_lock(&res->spinlock);
 988         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
 989                 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
 990                      res->lockname.len, res->lockname.name, res->owner);
 991                 spin_unlock(&res->spinlock);
 992                 /* this will cause the master to re-assert across
 993                  * the whole cluster, freeing up mles */
 994                 if (res->owner != dlm->node_num) {
 995                         ret = dlm_do_master_request(res, mle, res->owner);
 996                         if (ret < 0) {
 997                                 /* give recovery a chance to run */
 998                                 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
 999                                 msleep(500);
1000                                 goto recheck;
1001                         }
1002                 }
1003                 ret = 0;
1004                 goto leave;
1005         }
1006         spin_unlock(&res->spinlock);
1007
1008         spin_lock(&mle->spinlock);
1009         m = mle->master;
1010         map_changed = (memcmp(mle->vote_map, mle->node_map,
1011                               sizeof(mle->vote_map)) != 0);
1012         voting_done = (memcmp(mle->vote_map, mle->response_map,
1013                              sizeof(mle->vote_map)) == 0);
1014
1015         /* restart if we hit any errors */
1016         if (map_changed) {
1017                 int b;
1018                 mlog(0, "%s: %.*s: node map changed, restarting\n",
1019                      dlm->name, res->lockname.len, res->lockname.name);
1020                 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1021                 b = (mle->type == DLM_MLE_BLOCK);
1022                 if ((*blocked && !b) || (!*blocked && b)) {
1023                         mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1024                              dlm->name, res->lockname.len, res->lockname.name,
1025                              *blocked, b);
1026                         *blocked = b;
1027                 }
1028                 spin_unlock(&mle->spinlock);
1029                 if (ret < 0) {
1030                         mlog_errno(ret);
1031                         goto leave;
1032                 }
1033                 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1034                      "rechecking now\n", dlm->name, res->lockname.len,
1035                      res->lockname.name);
1036                 goto recheck;
1037         } else {
1038                 if (!voting_done) {
1039                         mlog(0, "map not changed and voting not done "
1040                              "for %s:%.*s\n", dlm->name, res->lockname.len,
1041                              res->lockname.name);
1042                 }
1043         }
1044
1045         if (m != O2NM_MAX_NODES) {
1046                 /* another node has done an assert!
1047                  * all done! */
1048                 sleep = 0;
1049         } else {
1050                 sleep = 1;
1051                 /* have all nodes responded? */
1052                 if (voting_done && !*blocked) {
1053                         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1054                         if (dlm->node_num <= bit) {
1055                                 /* my node number is lowest.
1056                                  * now tell other nodes that I am
1057                                  * mastering this. */
1058                                 mle->master = dlm->node_num;
1059                                 /* ref was grabbed in get_lock_resource
1060                                  * will be dropped in dlmlock_master */
1061                                 assert = 1;
1062                                 sleep = 0;
1063                         }
1064                         /* if voting is done, but we have not received
1065                          * an assert master yet, we must sleep */
1066                 }
1067         }
1068
1069         spin_unlock(&mle->spinlock);
1070
1071         /* sleep if we haven't finished voting yet */
1072         if (sleep) {
1073                 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1074
1075                 /*
1076                 if (atomic_read(&mle->mle_refs.refcount) < 2)
1077                         mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1078                         atomic_read(&mle->mle_refs.refcount),
1079                         res->lockname.len, res->lockname.name);
1080                 */
1081                 atomic_set(&mle->woken, 0);
1082                 (void)wait_event_timeout(mle->wq,
1083                                          (atomic_read(&mle->woken) == 1),
1084                                          timeo);
1085                 if (res->owner == O2NM_MAX_NODES) {
1086                         mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1087                              res->lockname.len, res->lockname.name);
1088                         goto recheck;
1089                 }
1090                 mlog(0, "done waiting, master is %u\n", res->owner);
1091                 ret = 0;
1092                 goto leave;
1093         }
1094
1095         ret = 0;   /* done */
1096         if (assert) {
1097                 m = dlm->node_num;
1098                 mlog(0, "about to master %.*s here, this=%u\n",
1099                      res->lockname.len, res->lockname.name, m);
1100                 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1101                 if (ret) {
1102                         /* This is a failure in the network path,
1103                          * not in the response to the assert_master
1104                          * (any nonzero response is a BUG on this node).
1105                          * Most likely a socket just got disconnected
1106                          * due to node death. */
1107                         mlog_errno(ret);
1108                 }
1109                 /* no longer need to restart lock mastery.
1110                  * all living nodes have been contacted. */
1111                 ret = 0;
1112         }
1113
1114         /* set the lockres owner */
1115         spin_lock(&res->spinlock);
1116         /* mastery reference obtained either during
1117          * assert_master_handler or in get_lock_resource */
1118         dlm_change_lockres_owner(dlm, res, m);
1119         spin_unlock(&res->spinlock);
1120
1121 leave:
1122         return ret;
1123 }
1124
1125 struct dlm_bitmap_diff_iter
1126 {
1127         int curnode;
1128         unsigned long *orig_bm;
1129         unsigned long *cur_bm;
1130         unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1131 };
1132
1133 enum dlm_node_state_change
1134 {
1135         NODE_DOWN = -1,
1136         NODE_NO_CHANGE = 0,
1137         NODE_UP
1138 };
1139
1140 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1141                                       unsigned long *orig_bm,
1142                                       unsigned long *cur_bm)
1143 {
1144         unsigned long p1, p2;
1145         int i;
1146
1147         iter->curnode = -1;
1148         iter->orig_bm = orig_bm;
1149         iter->cur_bm = cur_bm;
1150
1151         for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1152                 p1 = *(iter->orig_bm + i);
1153                 p2 = *(iter->cur_bm + i);
1154                 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1155         }
1156 }
1157
1158 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1159                                      enum dlm_node_state_change *state)
1160 {
1161         int bit;
1162
1163         if (iter->curnode >= O2NM_MAX_NODES)
1164                 return -ENOENT;
1165
1166         bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1167                             iter->curnode+1);
1168         if (bit >= O2NM_MAX_NODES) {
1169                 iter->curnode = O2NM_MAX_NODES;
1170                 return -ENOENT;
1171         }
1172
1173         /* if it was there in the original then this node died */
1174         if (test_bit(bit, iter->orig_bm))
1175                 *state = NODE_DOWN;
1176         else
1177                 *state = NODE_UP;
1178
1179         iter->curnode = bit;
1180         return bit;
1181 }
1182
1183
1184 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1185                                     struct dlm_lock_resource *res,
1186                                     struct dlm_master_list_entry *mle,
1187                                     int blocked)
1188 {
1189         struct dlm_bitmap_diff_iter bdi;
1190         enum dlm_node_state_change sc;
1191         int node;
1192         int ret = 0;
1193
1194         mlog(0, "something happened such that the "
1195              "master process may need to be restarted!\n");
1196
1197         assert_spin_locked(&mle->spinlock);
1198
1199         dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1200         node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1201         while (node >= 0) {
1202                 if (sc == NODE_UP) {
1203                         /* a node came up.  clear any old vote from
1204                          * the response map and set it in the vote map
1205                          * then restart the mastery. */
1206                         mlog(ML_NOTICE, "node %d up while restarting\n", node);
1207
1208                         /* redo the master request, but only for the new node */
1209                         mlog(0, "sending request to new node\n");
1210                         clear_bit(node, mle->response_map);
1211                         set_bit(node, mle->vote_map);
1212                 } else {
1213                         mlog(ML_ERROR, "node down! %d\n", node);
1214                         if (blocked) {
1215                                 int lowest = find_next_bit(mle->maybe_map,
1216                                                        O2NM_MAX_NODES, 0);
1217
1218                                 /* act like it was never there */
1219                                 clear_bit(node, mle->maybe_map);
1220
1221                                 if (node == lowest) {
1222                                         mlog(0, "expected master %u died"
1223                                             " while this node was blocked "
1224                                             "waiting on it!\n", node);
1225                                         lowest = find_next_bit(mle->maybe_map,
1226                                                         O2NM_MAX_NODES,
1227                                                         lowest+1);
1228                                         if (lowest < O2NM_MAX_NODES) {
1229                                                 mlog(0, "%s:%.*s:still "
1230                                                      "blocked. waiting on %u "
1231                                                      "now\n", dlm->name,
1232                                                      res->lockname.len,
1233                                                      res->lockname.name,
1234                                                      lowest);
1235                                         } else {
1236                                                 /* mle is an MLE_BLOCK, but
1237                                                  * there is now nothing left to
1238                                                  * block on.  we need to return
1239                                                  * all the way back out and try
1240                                                  * again with an MLE_MASTER.
1241                                                  * dlm_do_local_recovery_cleanup
1242                                                  * has already run, so the mle
1243                                                  * refcount is ok */
1244                                                 mlog(0, "%s:%.*s: no "
1245                                                      "longer blocking. try to "
1246                                                      "master this here\n",
1247                                                      dlm->name,
1248                                                      res->lockname.len,
1249                                                      res->lockname.name);
1250                                                 mle->type = DLM_MLE_MASTER;
1251                                                 mle->mleres = res;
1252                                         }
1253                                 }
1254                         }
1255
1256                         /* now blank out everything, as if we had never
1257                          * contacted anyone */
1258                         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1259                         memset(mle->response_map, 0, sizeof(mle->response_map));
1260                         /* reset the vote_map to the current node_map */
1261                         memcpy(mle->vote_map, mle->node_map,
1262                                sizeof(mle->node_map));
1263                         /* put myself into the maybe map */
1264                         if (mle->type != DLM_MLE_BLOCK)
1265                                 set_bit(dlm->node_num, mle->maybe_map);
1266                 }
1267                 ret = -EAGAIN;
1268                 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1269         }
1270         return ret;
1271 }
1272
1273
1274 /*
1275  * DLM_MASTER_REQUEST_MSG
1276  *
1277  * returns: 0 on success,
1278  *          -errno on a network error
1279  *
1280  * on error, the caller should assume the target node is "dead"
1281  *
1282  */
1283
1284 static int dlm_do_master_request(struct dlm_lock_resource *res,
1285                                  struct dlm_master_list_entry *mle, int to)
1286 {
1287         struct dlm_ctxt *dlm = mle->dlm;
1288         struct dlm_master_request request;
1289         int ret, response=0, resend;
1290
1291         memset(&request, 0, sizeof(request));
1292         request.node_idx = dlm->node_num;
1293
1294         BUG_ON(mle->type == DLM_MLE_MIGRATION);
1295
1296         request.namelen = (u8)mle->mnamelen;
1297         memcpy(request.name, mle->mname, request.namelen);
1298
1299 again:
1300         ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1301                                  sizeof(request), to, &response);
1302         if (ret < 0)  {
1303                 if (ret == -ESRCH) {
1304                         /* should never happen */
1305                         mlog(ML_ERROR, "TCP stack not ready!\n");
1306                         BUG();
1307                 } else if (ret == -EINVAL) {
1308                         mlog(ML_ERROR, "bad args passed to o2net!\n");
1309                         BUG();
1310                 } else if (ret == -ENOMEM) {
1311                         mlog(ML_ERROR, "out of memory while trying to send "
1312                              "network message!  retrying\n");
1313                         /* this is totally crude */
1314                         msleep(50);
1315                         goto again;
1316                 } else if (!dlm_is_host_down(ret)) {
1317                         /* not a network error. bad. */
1318                         mlog_errno(ret);
1319                         mlog(ML_ERROR, "unhandled error!");
1320                         BUG();
1321                 }
1322                 /* all other errors should be network errors,
1323                  * and likely indicate node death */
1324                 mlog(ML_ERROR, "link to %d went down!\n", to);
1325                 goto out;
1326         }
1327
1328         ret = 0;
1329         resend = 0;
1330         spin_lock(&mle->spinlock);
1331         switch (response) {
1332                 case DLM_MASTER_RESP_YES:
1333                         set_bit(to, mle->response_map);
1334                         mlog(0, "node %u is the master, response=YES\n", to);
1335                         mlog(0, "%s:%.*s: master node %u now knows I have a "
1336                              "reference\n", dlm->name, res->lockname.len,
1337                              res->lockname.name, to);
1338                         mle->master = to;
1339                         break;
1340                 case DLM_MASTER_RESP_NO:
1341                         mlog(0, "node %u not master, response=NO\n", to);
1342                         set_bit(to, mle->response_map);
1343                         break;
1344                 case DLM_MASTER_RESP_MAYBE:
1345                         mlog(0, "node %u not master, response=MAYBE\n", to);
1346                         set_bit(to, mle->response_map);
1347                         set_bit(to, mle->maybe_map);
1348                         break;
1349                 case DLM_MASTER_RESP_ERROR:
1350                         mlog(0, "node %u hit an error, resending\n", to);
1351                         resend = 1;
1352                         response = 0;
1353                         break;
1354                 default:
1355                         mlog(ML_ERROR, "bad response! %u\n", response);
1356                         BUG();
1357         }
1358         spin_unlock(&mle->spinlock);
1359         if (resend) {
1360                 /* this is also totally crude */
1361                 msleep(50);
1362                 goto again;
1363         }
1364
1365 out:
1366         return ret;
1367 }
1368
1369 /*
1370  * locks that can be taken here:
1371  * dlm->spinlock
1372  * res->spinlock
1373  * mle->spinlock
1374  * dlm->master_list
1375  *
1376  * if possible, TRIM THIS DOWN!!!
1377  */
1378 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1379                                void **ret_data)
1380 {
1381         u8 response = DLM_MASTER_RESP_MAYBE;
1382         struct dlm_ctxt *dlm = data;
1383         struct dlm_lock_resource *res = NULL;
1384         struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1385         struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1386         char *name;
1387         unsigned int namelen, hash;
1388         int found, ret;
1389         int set_maybe;
1390         int dispatch_assert = 0;
1391
1392         if (!dlm_grab(dlm))
1393                 return DLM_MASTER_RESP_NO;
1394
1395         if (!dlm_domain_fully_joined(dlm)) {
1396                 response = DLM_MASTER_RESP_NO;
1397                 goto send_response;
1398         }
1399
1400         name = request->name;
1401         namelen = request->namelen;
1402         hash = dlm_lockid_hash(name, namelen);
1403
1404         if (namelen > DLM_LOCKID_NAME_MAX) {
1405                 response = DLM_IVBUFLEN;
1406                 goto send_response;
1407         }
1408
1409 way_up_top:
1410         spin_lock(&dlm->spinlock);
1411         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1412         if (res) {
1413                 spin_unlock(&dlm->spinlock);
1414
1415                 /* take care of the easy cases up front */
1416                 spin_lock(&res->spinlock);
1417                 if (res->state & (DLM_LOCK_RES_RECOVERING|
1418                                   DLM_LOCK_RES_MIGRATING)) {
1419                         spin_unlock(&res->spinlock);
1420                         mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1421                              "being recovered/migrated\n");
1422                         response = DLM_MASTER_RESP_ERROR;
1423                         if (mle)
1424                                 kmem_cache_free(dlm_mle_cache, mle);
1425                         goto send_response;
1426                 }
1427
1428                 if (res->owner == dlm->node_num) {
1429                         mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1430                              dlm->name, namelen, name, request->node_idx);
1431                         dlm_lockres_set_refmap_bit(request->node_idx, res);
1432                         spin_unlock(&res->spinlock);
1433                         response = DLM_MASTER_RESP_YES;
1434                         if (mle)
1435                                 kmem_cache_free(dlm_mle_cache, mle);
1436
1437                         /* this node is the owner.
1438                          * there is some extra work that needs to
1439                          * happen now.  the requesting node has
1440                          * caused all nodes up to this one to
1441                          * create mles.  this node now needs to
1442                          * go back and clean those up. */
1443                         dispatch_assert = 1;
1444                         goto send_response;
1445                 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1446                         spin_unlock(&res->spinlock);
1447                         // mlog(0, "node %u is the master\n", res->owner);
1448                         response = DLM_MASTER_RESP_NO;
1449                         if (mle)
1450                                 kmem_cache_free(dlm_mle_cache, mle);
1451                         goto send_response;
1452                 }
1453
1454                 /* ok, there is no owner.  either this node is
1455                  * being blocked, or it is actively trying to
1456                  * master this lock. */
1457                 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1458                         mlog(ML_ERROR, "lock with no owner should be "
1459                              "in-progress!\n");
1460                         BUG();
1461                 }
1462
1463                 // mlog(0, "lockres is in progress...\n");
1464                 spin_lock(&dlm->master_lock);
1465                 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1466                 if (!found) {
1467                         mlog(ML_ERROR, "no mle found for this lock!\n");
1468                         BUG();
1469                 }
1470                 set_maybe = 1;
1471                 spin_lock(&tmpmle->spinlock);
1472                 if (tmpmle->type == DLM_MLE_BLOCK) {
1473                         // mlog(0, "this node is waiting for "
1474                         // "lockres to be mastered\n");
1475                         response = DLM_MASTER_RESP_NO;
1476                 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1477                         mlog(0, "node %u is master, but trying to migrate to "
1478                              "node %u.\n", tmpmle->master, tmpmle->new_master);
1479                         if (tmpmle->master == dlm->node_num) {
1480                                 mlog(ML_ERROR, "no owner on lockres, but this "
1481                                      "node is trying to migrate it to %u?!\n",
1482                                      tmpmle->new_master);
1483                                 BUG();
1484                         } else {
1485                                 /* the real master can respond on its own */
1486                                 response = DLM_MASTER_RESP_NO;
1487                         }
1488                 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1489                         set_maybe = 0;
1490                         if (tmpmle->master == dlm->node_num) {
1491                                 response = DLM_MASTER_RESP_YES;
1492                                 /* this node will be the owner.
1493                                  * go back and clean the mles on any
1494                                  * other nodes */
1495                                 dispatch_assert = 1;
1496                                 dlm_lockres_set_refmap_bit(request->node_idx, res);
1497                                 mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1498                                      dlm->name, namelen, name,
1499                                      request->node_idx);
1500                         } else
1501                                 response = DLM_MASTER_RESP_NO;
1502                 } else {
1503                         // mlog(0, "this node is attempting to "
1504                         // "master lockres\n");
1505                         response = DLM_MASTER_RESP_MAYBE;
1506                 }
1507                 if (set_maybe)
1508                         set_bit(request->node_idx, tmpmle->maybe_map);
1509                 spin_unlock(&tmpmle->spinlock);
1510
1511                 spin_unlock(&dlm->master_lock);
1512                 spin_unlock(&res->spinlock);
1513
1514                 /* keep the mle attached to heartbeat events */
1515                 dlm_put_mle(tmpmle);
1516                 if (mle)
1517                         kmem_cache_free(dlm_mle_cache, mle);
1518                 goto send_response;
1519         }
1520
1521         /*
1522          * lockres doesn't exist on this node
1523          * if there is an MLE_BLOCK, return NO
1524          * if there is an MLE_MASTER, return MAYBE
1525          * otherwise, add an MLE_BLOCK, return NO
1526          */
1527         spin_lock(&dlm->master_lock);
1528         found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1529         if (!found) {
1530                 /* this lockid has never been seen on this node yet */
1531                 // mlog(0, "no mle found\n");
1532                 if (!mle) {
1533                         spin_unlock(&dlm->master_lock);
1534                         spin_unlock(&dlm->spinlock);
1535
1536                         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1537                         if (!mle) {
1538                                 response = DLM_MASTER_RESP_ERROR;
1539                                 mlog_errno(-ENOMEM);
1540                                 goto send_response;
1541                         }
1542                         goto way_up_top;
1543                 }
1544
1545                 // mlog(0, "this is second time thru, already allocated, "
1546                 // "add the block.\n");
1547                 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1548                 set_bit(request->node_idx, mle->maybe_map);
1549                 __dlm_insert_mle(dlm, mle);
1550                 response = DLM_MASTER_RESP_NO;
1551         } else {
1552                 // mlog(0, "mle was found\n");
1553                 set_maybe = 1;
1554                 spin_lock(&tmpmle->spinlock);
1555                 if (tmpmle->master == dlm->node_num) {
1556                         mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1557                         BUG();
1558                 }
1559                 if (tmpmle->type == DLM_MLE_BLOCK)
1560                         response = DLM_MASTER_RESP_NO;
1561                 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1562                         mlog(0, "migration mle was found (%u->%u)\n",
1563                              tmpmle->master, tmpmle->new_master);
1564                         /* real master can respond on its own */
1565                         response = DLM_MASTER_RESP_NO;
1566                 } else
1567                         response = DLM_MASTER_RESP_MAYBE;
1568                 if (set_maybe)
1569                         set_bit(request->node_idx, tmpmle->maybe_map);
1570                 spin_unlock(&tmpmle->spinlock);
1571         }
1572         spin_unlock(&dlm->master_lock);
1573         spin_unlock(&dlm->spinlock);
1574
1575         if (found) {
1576                 /* keep the mle attached to heartbeat events */
1577                 dlm_put_mle(tmpmle);
1578         }
1579 send_response:
1580         /*
1581          * __dlm_lookup_lockres() grabbed a reference to this lockres.
1582          * The reference is released by dlm_assert_master_worker() under
1583          * the call to dlm_dispatch_assert_master().  If
1584          * dlm_assert_master_worker() isn't called, we drop it here.
1585          */
1586         if (dispatch_assert) {
1587                 if (response != DLM_MASTER_RESP_YES)
1588                         mlog(ML_ERROR, "invalid response %d\n", response);
1589                 if (!res) {
1590                         mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1591                         BUG();
1592                 }
1593                 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1594                              dlm->node_num, res->lockname.len, res->lockname.name);
1595                 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1596                                                  DLM_ASSERT_MASTER_MLE_CLEANUP);
1597                 if (ret < 0) {
1598                         mlog(ML_ERROR, "failed to dispatch assert master work\n");
1599                         response = DLM_MASTER_RESP_ERROR;
1600                         dlm_lockres_put(res);
1601                 }
1602         } else {
1603                 if (res)
1604                         dlm_lockres_put(res);
1605         }
1606
1607         dlm_put(dlm);
1608         return response;
1609 }
1610
1611 /*
1612  * DLM_ASSERT_MASTER_MSG
1613  */
1614
1615
1616 /*
1617  * NOTE: this can be used for debugging
1618  * can periodically run all locks owned by this node
1619  * and re-assert across the cluster...
1620  */
1621 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1622                                 struct dlm_lock_resource *res,
1623                                 void *nodemap, u32 flags)
1624 {
1625         struct dlm_assert_master assert;
1626         int to, tmpret;
1627         struct dlm_node_iter iter;
1628         int ret = 0;
1629         int reassert;
1630         const char *lockname = res->lockname.name;
1631         unsigned int namelen = res->lockname.len;
1632
1633         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1634
1635         spin_lock(&res->spinlock);
1636         res->state |= DLM_LOCK_RES_SETREF_INPROG;
1637         spin_unlock(&res->spinlock);
1638
1639 again:
1640         reassert = 0;
1641
1642         /* note that if this nodemap is empty, it returns 0 */
1643         dlm_node_iter_init(nodemap, &iter);
1644         while ((to = dlm_node_iter_next(&iter)) >= 0) {
1645                 int r = 0;
1646                 struct dlm_master_list_entry *mle = NULL;
1647
1648                 mlog(0, "sending assert master to %d (%.*s)\n", to,
1649                      namelen, lockname);
1650                 memset(&assert, 0, sizeof(assert));
1651                 assert.node_idx = dlm->node_num;
1652                 assert.namelen = namelen;
1653                 memcpy(assert.name, lockname, namelen);
1654                 assert.flags = cpu_to_be32(flags);
1655
1656                 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1657                                             &assert, sizeof(assert), to, &r);
1658                 if (tmpret < 0) {
1659                         mlog(ML_ERROR, "Error %d when sending message %u (key "
1660                              "0x%x) to node %u\n", tmpret,
1661                              DLM_ASSERT_MASTER_MSG, dlm->key, to);
1662                         if (!dlm_is_host_down(tmpret)) {
1663                                 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1664                                 BUG();
1665                         }
1666                         /* a node died.  finish out the rest of the nodes. */
1667                         mlog(0, "link to %d went down!\n", to);
1668                         /* any nonzero status return will do */
1669                         ret = tmpret;
1670                         r = 0;
1671                 } else if (r < 0) {
1672                         /* ok, something horribly messed.  kill thyself. */
1673                         mlog(ML_ERROR,"during assert master of %.*s to %u, "
1674                              "got %d.\n", namelen, lockname, to, r);
1675                         spin_lock(&dlm->spinlock);
1676                         spin_lock(&dlm->master_lock);
1677                         if (dlm_find_mle(dlm, &mle, (char *)lockname,
1678                                          namelen)) {
1679                                 dlm_print_one_mle(mle);
1680                                 __dlm_put_mle(mle);
1681                         }
1682                         spin_unlock(&dlm->master_lock);
1683                         spin_unlock(&dlm->spinlock);
1684                         BUG();
1685                 }
1686
1687                 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1688                     !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1689                                 mlog(ML_ERROR, "%.*s: very strange, "
1690                                      "master MLE but no lockres on %u\n",
1691                                      namelen, lockname, to);
1692                 }
1693
1694                 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1695                         mlog(0, "%.*s: node %u create mles on other "
1696                              "nodes and requests a re-assert\n",
1697                              namelen, lockname, to);
1698                         reassert = 1;
1699                 }
1700                 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1701                         mlog(0, "%.*s: node %u has a reference to this "
1702                              "lockres, set the bit in the refmap\n",
1703                              namelen, lockname, to);
1704                         spin_lock(&res->spinlock);
1705                         dlm_lockres_set_refmap_bit(to, res);
1706                         spin_unlock(&res->spinlock);
1707                 }
1708         }
1709
1710         if (reassert)
1711                 goto again;
1712
1713         spin_lock(&res->spinlock);
1714         res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1715         spin_unlock(&res->spinlock);
1716         wake_up(&res->wq);
1717
1718         return ret;
1719 }
1720
1721 /*
1722  * locks that can be taken here:
1723  * dlm->spinlock
1724  * res->spinlock
1725  * mle->spinlock
1726  * dlm->master_list
1727  *
1728  * if possible, TRIM THIS DOWN!!!
1729  */
1730 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1731                               void **ret_data)
1732 {
1733         struct dlm_ctxt *dlm = data;
1734         struct dlm_master_list_entry *mle = NULL;
1735         struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1736         struct dlm_lock_resource *res = NULL;
1737         char *name;
1738         unsigned int namelen, hash;
1739         u32 flags;
1740         int master_request = 0, have_lockres_ref = 0;
1741         int ret = 0;
1742
1743         if (!dlm_grab(dlm))
1744                 return 0;
1745
1746         name = assert->name;
1747         namelen = assert->namelen;
1748         hash = dlm_lockid_hash(name, namelen);
1749         flags = be32_to_cpu(assert->flags);
1750
1751         if (namelen > DLM_LOCKID_NAME_MAX) {
1752                 mlog(ML_ERROR, "Invalid name length!");
1753                 goto done;
1754         }
1755
1756         spin_lock(&dlm->spinlock);
1757
1758         if (flags)
1759                 mlog(0, "assert_master with flags: %u\n", flags);
1760
1761         /* find the MLE */
1762         spin_lock(&dlm->master_lock);
1763         if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1764                 /* not an error, could be master just re-asserting */
1765                 mlog(0, "just got an assert_master from %u, but no "
1766                      "MLE for it! (%.*s)\n", assert->node_idx,
1767                      namelen, name);
1768         } else {
1769                 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1770                 if (bit >= O2NM_MAX_NODES) {
1771                         /* not necessarily an error, though less likely.
1772                          * could be master just re-asserting. */
1773                         mlog(0, "no bits set in the maybe_map, but %u "
1774                              "is asserting! (%.*s)\n", assert->node_idx,
1775                              namelen, name);
1776                 } else if (bit != assert->node_idx) {
1777                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1778                                 mlog(0, "master %u was found, %u should "
1779                                      "back off\n", assert->node_idx, bit);
1780                         } else {
1781                                 /* with the fix for bug 569, a higher node
1782                                  * number winning the mastery will respond
1783                                  * YES to mastery requests, but this node
1784                                  * had no way of knowing.  let it pass. */
1785                                 mlog(0, "%u is the lowest node, "
1786                                      "%u is asserting. (%.*s)  %u must "
1787                                      "have begun after %u won.\n", bit,
1788                                      assert->node_idx, namelen, name, bit,
1789                                      assert->node_idx);
1790                         }
1791                 }
1792                 if (mle->type == DLM_MLE_MIGRATION) {
1793                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1794                                 mlog(0, "%s:%.*s: got cleanup assert"
1795                                      " from %u for migration\n",
1796                                      dlm->name, namelen, name,
1797                                      assert->node_idx);
1798                         } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1799                                 mlog(0, "%s:%.*s: got unrelated assert"
1800                                      " from %u for migration, ignoring\n",
1801                                      dlm->name, namelen, name,
1802                                      assert->node_idx);
1803                                 __dlm_put_mle(mle);
1804                                 spin_unlock(&dlm->master_lock);
1805                                 spin_unlock(&dlm->spinlock);
1806                                 goto done;
1807                         }
1808                 }
1809         }
1810         spin_unlock(&dlm->master_lock);
1811
1812         /* ok everything checks out with the MLE
1813          * now check to see if there is a lockres */
1814         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1815         if (res) {
1816                 spin_lock(&res->spinlock);
1817                 if (res->state & DLM_LOCK_RES_RECOVERING)  {
1818                         mlog(ML_ERROR, "%u asserting but %.*s is "
1819                              "RECOVERING!\n", assert->node_idx, namelen, name);
1820                         goto kill;
1821                 }
1822                 if (!mle) {
1823                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1824                             res->owner != assert->node_idx) {
1825                                 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1826                                      "but current owner is %u! (%.*s)\n",
1827                                      assert->node_idx, res->owner, namelen,
1828                                      name);
1829                                 __dlm_print_one_lock_resource(res);
1830                                 BUG();
1831                         }
1832                 } else if (mle->type != DLM_MLE_MIGRATION) {
1833                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1834                                 /* owner is just re-asserting */
1835                                 if (res->owner == assert->node_idx) {
1836                                         mlog(0, "owner %u re-asserting on "
1837                                              "lock %.*s\n", assert->node_idx,
1838                                              namelen, name);
1839                                         goto ok;
1840                                 }
1841                                 mlog(ML_ERROR, "got assert_master from "
1842                                      "node %u, but %u is the owner! "
1843                                      "(%.*s)\n", assert->node_idx,
1844                                      res->owner, namelen, name);
1845                                 goto kill;
1846                         }
1847                         if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1848                                 mlog(ML_ERROR, "got assert from %u, but lock "
1849                                      "with no owner should be "
1850                                      "in-progress! (%.*s)\n",
1851                                      assert->node_idx,
1852                                      namelen, name);
1853                                 goto kill;
1854                         }
1855                 } else /* mle->type == DLM_MLE_MIGRATION */ {
1856                         /* should only be getting an assert from new master */
1857                         if (assert->node_idx != mle->new_master) {
1858                                 mlog(ML_ERROR, "got assert from %u, but "
1859                                      "new master is %u, and old master "
1860                                      "was %u (%.*s)\n",
1861                                      assert->node_idx, mle->new_master,
1862                                      mle->master, namelen, name);
1863                                 goto kill;
1864                         }
1865
1866                 }
1867 ok:
1868                 spin_unlock(&res->spinlock);
1869         }
1870
1871         // mlog(0, "woo!  got an assert_master from node %u!\n",
1872         //           assert->node_idx);
1873         if (mle) {
1874                 int extra_ref = 0;
1875                 int nn = -1;
1876                 int rr, err = 0;
1877
1878                 spin_lock(&mle->spinlock);
1879                 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1880                         extra_ref = 1;
1881                 else {
1882                         /* MASTER mle: if any bits set in the response map
1883                          * then the calling node needs to re-assert to clear
1884                          * up nodes that this node contacted */
1885                         while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1886                                                     nn+1)) < O2NM_MAX_NODES) {
1887                                 if (nn != dlm->node_num && nn != assert->node_idx)
1888                                         master_request = 1;
1889                         }
1890                 }
1891                 mle->master = assert->node_idx;
1892                 atomic_set(&mle->woken, 1);
1893                 wake_up(&mle->wq);
1894                 spin_unlock(&mle->spinlock);
1895
1896                 if (res) {
1897                         int wake = 0;
1898                         spin_lock(&res->spinlock);
1899                         if (mle->type == DLM_MLE_MIGRATION) {
1900                                 mlog(0, "finishing off migration of lockres %.*s, "
1901                                         "from %u to %u\n",
1902                                         res->lockname.len, res->lockname.name,
1903                                         dlm->node_num, mle->new_master);
1904                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
1905                                 wake = 1;
1906                                 dlm_change_lockres_owner(dlm, res, mle->new_master);
1907                                 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1908                         } else {
1909                                 dlm_change_lockres_owner(dlm, res, mle->master);
1910                         }
1911                         spin_unlock(&res->spinlock);
1912                         have_lockres_ref = 1;
1913                         if (wake)
1914                                 wake_up(&res->wq);
1915                 }
1916
1917                 /* master is known, detach if not already detached.
1918                  * ensures that only one assert_master call will happen
1919                  * on this mle. */
1920                 spin_lock(&dlm->master_lock);
1921
1922                 rr = atomic_read(&mle->mle_refs.refcount);
1923                 if (mle->inuse > 0) {
1924                         if (extra_ref && rr < 3)
1925                                 err = 1;
1926                         else if (!extra_ref && rr < 2)
1927                                 err = 1;
1928                 } else {
1929                         if (extra_ref && rr < 2)
1930                                 err = 1;
1931                         else if (!extra_ref && rr < 1)
1932                                 err = 1;
1933                 }
1934                 if (err) {
1935                         mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
1936                              "that will mess up this node, refs=%d, extra=%d, "
1937                              "inuse=%d\n", dlm->name, namelen, name,
1938                              assert->node_idx, rr, extra_ref, mle->inuse);
1939                         dlm_print_one_mle(mle);
1940                 }
1941                 __dlm_unlink_mle(dlm, mle);
1942                 __dlm_mle_detach_hb_events(dlm, mle);
1943                 __dlm_put_mle(mle);
1944                 if (extra_ref) {
1945                         /* the assert master message now balances the extra
1946                          * ref given by the master / migration request message.
1947                          * if this is the last put, it will be removed
1948                          * from the list. */
1949                         __dlm_put_mle(mle);
1950                 }
1951                 spin_unlock(&dlm->master_lock);
1952         } else if (res) {
1953                 if (res->owner != assert->node_idx) {
1954                         mlog(0, "assert_master from %u, but current "
1955                              "owner is %u (%.*s), no mle\n", assert->node_idx,
1956                              res->owner, namelen, name);
1957                 }
1958         }
1959         spin_unlock(&dlm->spinlock);
1960
1961 done:
1962         ret = 0;
1963         if (res) {
1964                 spin_lock(&res->spinlock);
1965                 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1966                 spin_unlock(&res->spinlock);
1967                 *ret_data = (void *)res;
1968         }
1969         dlm_put(dlm);
1970         if (master_request) {
1971                 mlog(0, "need to tell master to reassert\n");
1972                 /* positive. negative would shoot down the node. */
1973                 ret |= DLM_ASSERT_RESPONSE_REASSERT;
1974                 if (!have_lockres_ref) {
1975                         mlog(ML_ERROR, "strange, got assert from %u, MASTER "
1976                              "mle present here for %s:%.*s, but no lockres!\n",
1977                              assert->node_idx, dlm->name, namelen, name);
1978                 }
1979         }
1980         if (have_lockres_ref) {
1981                 /* let the master know we have a reference to the lockres */
1982                 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
1983                 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
1984                      dlm->name, namelen, name, assert->node_idx);
1985         }
1986         return ret;
1987
1988 kill:
1989         /* kill the caller! */
1990         mlog(ML_ERROR, "Bad message received from another node.  Dumping state "
1991              "and killing the other node now!  This node is OK and can continue.\n");
1992         __dlm_print_one_lock_resource(res);
1993         spin_unlock(&res->spinlock);
1994         spin_unlock(&dlm->spinlock);
1995         *ret_data = (void *)res;
1996         dlm_put(dlm);
1997         return -EINVAL;
1998 }
1999
2000 void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2001 {
2002         struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2003
2004         if (ret_data) {
2005                 spin_lock(&res->spinlock);
2006                 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2007                 spin_unlock(&res->spinlock);
2008                 wake_up(&res->wq);
2009                 dlm_lockres_put(res);
2010         }
2011         return;
2012 }
2013
2014 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2015                                struct dlm_lock_resource *res,
2016                                int ignore_higher, u8 request_from, u32 flags)
2017 {
2018         struct dlm_work_item *item;
2019         item = kzalloc(sizeof(*item), GFP_NOFS);
2020         if (!item)
2021                 return -ENOMEM;
2022
2023
2024         /* queue up work for dlm_assert_master_worker */
2025         dlm_grab(dlm);  /* get an extra ref for the work item */
2026         dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2027         item->u.am.lockres = res; /* already have a ref */
2028         /* can optionally ignore node numbers higher than this node */
2029         item->u.am.ignore_higher = ignore_higher;
2030         item->u.am.request_from = request_from;
2031         item->u.am.flags = flags;
2032
2033         if (ignore_higher)
2034                 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2035                      res->lockname.name);
2036
2037         spin_lock(&dlm->work_lock);
2038         list_add_tail(&item->list, &dlm->work_list);
2039         spin_unlock(&dlm->work_lock);
2040
2041         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2042         return 0;
2043 }
2044
2045 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2046 {
2047         struct dlm_ctxt *dlm = data;
2048         int ret = 0;
2049         struct dlm_lock_resource *res;
2050         unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2051         int ignore_higher;
2052         int bit;
2053         u8 request_from;
2054         u32 flags;
2055
2056         dlm = item->dlm;
2057         res = item->u.am.lockres;
2058         ignore_higher = item->u.am.ignore_higher;
2059         request_from = item->u.am.request_from;
2060         flags = item->u.am.flags;
2061
2062         spin_lock(&dlm->spinlock);
2063         memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2064         spin_unlock(&dlm->spinlock);
2065
2066         clear_bit(dlm->node_num, nodemap);
2067         if (ignore_higher) {
2068                 /* if is this just to clear up mles for nodes below
2069                  * this node, do not send the message to the original
2070                  * caller or any node number higher than this */
2071                 clear_bit(request_from, nodemap);
2072                 bit = dlm->node_num;
2073                 while (1) {
2074                         bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2075                                             bit+1);
2076                         if (bit >= O2NM_MAX_NODES)
2077                                 break;
2078                         clear_bit(bit, nodemap);
2079                 }
2080         }
2081
2082         /*
2083          * If we're migrating this lock to someone else, we are no
2084          * longer allowed to assert out own mastery.  OTOH, we need to
2085          * prevent migration from starting while we're still asserting
2086          * our dominance.  The reserved ast delays migration.
2087          */
2088         spin_lock(&res->spinlock);
2089         if (res->state & DLM_LOCK_RES_MIGRATING) {
2090                 mlog(0, "Someone asked us to assert mastery, but we're "
2091                      "in the middle of migration.  Skipping assert, "
2092                      "the new master will handle that.\n");
2093                 spin_unlock(&res->spinlock);
2094                 goto put;
2095         } else
2096                 __dlm_lockres_reserve_ast(res);
2097         spin_unlock(&res->spinlock);
2098
2099         /* this call now finishes out the nodemap
2100          * even if one or more nodes die */
2101         mlog(0, "worker about to master %.*s here, this=%u\n",
2102                      res->lockname.len, res->lockname.name, dlm->node_num);
2103         ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2104         if (ret < 0) {
2105                 /* no need to restart, we are done */
2106                 if (!dlm_is_host_down(ret))
2107                         mlog_errno(ret);
2108         }
2109
2110         /* Ok, we've asserted ourselves.  Let's let migration start. */
2111         dlm_lockres_release_ast(dlm, res);
2112
2113 put:
2114         dlm_lockres_put(res);
2115
2116         mlog(0, "finished with dlm_assert_master_worker\n");
2117 }
2118
2119 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2120  * We cannot wait for node recovery to complete to begin mastering this
2121  * lockres because this lockres is used to kick off recovery! ;-)
2122  * So, do a pre-check on all living nodes to see if any of those nodes
2123  * think that $RECOVERY is currently mastered by a dead node.  If so,
2124  * we wait a short time to allow that node to get notified by its own
2125  * heartbeat stack, then check again.  All $RECOVERY lock resources
2126  * mastered by dead nodes are purged when the hearbeat callback is
2127  * fired, so we can know for sure that it is safe to continue once
2128  * the node returns a live node or no node.  */
2129 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2130                                        struct dlm_lock_resource *res)
2131 {
2132         struct dlm_node_iter iter;
2133         int nodenum;
2134         int ret = 0;
2135         u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2136
2137         spin_lock(&dlm->spinlock);
2138         dlm_node_iter_init(dlm->domain_map, &iter);
2139         spin_unlock(&dlm->spinlock);
2140
2141         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2142                 /* do not send to self */
2143                 if (nodenum == dlm->node_num)
2144                         continue;
2145                 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2146                 if (ret < 0) {
2147                         mlog_errno(ret);
2148                         if (!dlm_is_host_down(ret))
2149                                 BUG();
2150                         /* host is down, so answer for that node would be
2151                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
2152                         ret = 0;
2153                 }
2154
2155                 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2156                         /* check to see if this master is in the recovery map */
2157                         spin_lock(&dlm->spinlock);
2158                         if (test_bit(master, dlm->recovery_map)) {
2159                                 mlog(ML_NOTICE, "%s: node %u has not seen "
2160                                      "node %u go down yet, and thinks the "
2161                                      "dead node is mastering the recovery "
2162                                      "lock.  must wait.\n", dlm->name,
2163                                      nodenum, master);
2164                                 ret = -EAGAIN;
2165                         }
2166                         spin_unlock(&dlm->spinlock);
2167                         mlog(0, "%s: reco lock master is %u\n", dlm->name,
2168                              master);
2169                         break;
2170                 }
2171         }
2172         return ret;
2173 }
2174
2175 /*
2176  * DLM_DEREF_LOCKRES_MSG
2177  */
2178
2179 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2180 {
2181         struct dlm_deref_lockres deref;
2182         int ret = 0, r;
2183         const char *lockname;
2184         unsigned int namelen;
2185
2186         lockname = res->lockname.name;
2187         namelen = res->lockname.len;
2188         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2189
2190         mlog(0, "%s:%.*s: sending deref to %d\n",
2191              dlm->name, namelen, lockname, res->owner);
2192         memset(&deref, 0, sizeof(deref));
2193         deref.node_idx = dlm->node_num;
2194         deref.namelen = namelen;
2195         memcpy(deref.name, lockname, namelen);
2196
2197         ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2198                                  &deref, sizeof(deref), res->owner, &r);
2199         if (ret < 0)
2200                 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
2201                      "node %u\n", ret, DLM_DEREF_LOCKRES_MSG, dlm->key,
2202                      res->owner);
2203         else if (r < 0) {
2204                 /* BAD.  other node says I did not have a ref. */
2205                 mlog(ML_ERROR,"while dropping ref on %s:%.*s "
2206                     "(master=%u) got %d.\n", dlm->name, namelen,
2207                     lockname, res->owner, r);
2208                 dlm_print_one_lock_resource(res);
2209                 BUG();
2210         }
2211         return ret;
2212 }
2213
2214 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2215                               void **ret_data)
2216 {
2217         struct dlm_ctxt *dlm = data;
2218         struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2219         struct dlm_lock_resource *res = NULL;
2220         char *name;
2221         unsigned int namelen;
2222         int ret = -EINVAL;
2223         u8 node;
2224         unsigned int hash;
2225         struct dlm_work_item *item;
2226         int cleared = 0;
2227         int dispatch = 0;
2228
2229         if (!dlm_grab(dlm))
2230                 return 0;
2231
2232         name = deref->name;
2233         namelen = deref->namelen;
2234         node = deref->node_idx;
2235
2236         if (namelen > DLM_LOCKID_NAME_MAX) {
2237                 mlog(ML_ERROR, "Invalid name length!");
2238                 goto done;
2239         }
2240         if (deref->node_idx >= O2NM_MAX_NODES) {
2241                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2242                 goto done;
2243         }
2244
2245         hash = dlm_lockid_hash(name, namelen);
2246
2247         spin_lock(&dlm->spinlock);
2248         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2249         if (!res) {
2250                 spin_unlock(&dlm->spinlock);
2251                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2252                      dlm->name, namelen, name);
2253                 goto done;
2254         }
2255         spin_unlock(&dlm->spinlock);
2256
2257         spin_lock(&res->spinlock);
2258         if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2259                 dispatch = 1;
2260         else {
2261                 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2262                 if (test_bit(node, res->refmap)) {
2263                         dlm_lockres_clear_refmap_bit(node, res);
2264                         cleared = 1;
2265                 }
2266         }
2267         spin_unlock(&res->spinlock);
2268
2269         if (!dispatch) {
2270                 if (cleared)
2271                         dlm_lockres_calc_usage(dlm, res);
2272                 else {
2273                         mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2274                         "but it is already dropped!\n", dlm->name,
2275                         res->lockname.len, res->lockname.name, node);
2276                         dlm_print_one_lock_resource(res);
2277                 }
2278                 ret = 0;
2279                 goto done;
2280         }
2281
2282         item = kzalloc(sizeof(*item), GFP_NOFS);
2283         if (!item) {
2284                 ret = -ENOMEM;
2285                 mlog_errno(ret);
2286                 goto done;
2287         }
2288
2289         dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2290         item->u.dl.deref_res = res;
2291         item->u.dl.deref_node = node;
2292
2293         spin_lock(&dlm->work_lock);
2294         list_add_tail(&item->list, &dlm->work_list);
2295         spin_unlock(&dlm->work_lock);
2296
2297         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2298         return 0;
2299
2300 done:
2301         if (res)
2302                 dlm_lockres_put(res);
2303         dlm_put(dlm);
2304
2305         return ret;
2306 }
2307
2308 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2309 {
2310         struct dlm_ctxt *dlm;
2311         struct dlm_lock_resource *res;
2312         u8 node;
2313         u8 cleared = 0;
2314
2315         dlm = item->dlm;
2316         res = item->u.dl.deref_res;
2317         node = item->u.dl.deref_node;
2318
2319         spin_lock(&res->spinlock);
2320         BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2321         if (test_bit(node, res->refmap)) {
2322                 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2323                 dlm_lockres_clear_refmap_bit(node, res);
2324                 cleared = 1;
2325         }
2326         spin_unlock(&res->spinlock);
2327
2328         if (cleared) {
2329                 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2330                      dlm->name, res->lockname.len, res->lockname.name, node);
2331                 dlm_lockres_calc_usage(dlm, res);
2332         } else {
2333                 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2334                      "but it is already dropped!\n", dlm->name,
2335                      res->lockname.len, res->lockname.name, node);
2336                 dlm_print_one_lock_resource(res);
2337         }
2338
2339         dlm_lockres_put(res);
2340 }
2341
2342 /* Checks whether the lockres can be migrated. Returns 0 if yes, < 0
2343  * if not. If 0, numlocks is set to the number of locks in the lockres.
2344  */
2345 static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
2346                                       struct dlm_lock_resource *res,
2347                                       int *numlocks,
2348                                       int *hasrefs)
2349 {
2350         int ret;
2351         int i;
2352         int count = 0;
2353         struct list_head *queue;
2354         struct dlm_lock *lock;
2355
2356         assert_spin_locked(&res->spinlock);
2357
2358         *numlocks = 0;
2359         *hasrefs = 0;
2360
2361         ret = -EINVAL;
2362         if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
2363                 mlog(0, "cannot migrate lockres with unknown owner!\n");
2364                 goto leave;
2365         }
2366
2367         if (res->owner != dlm->node_num) {
2368                 mlog(0, "cannot migrate lockres this node doesn't own!\n");
2369                 goto leave;
2370         }
2371
2372         ret = 0;
2373         queue = &res->granted;
2374         for (i = 0; i < 3; i++) {
2375                 list_for_each_entry(lock, queue, list) {
2376                         ++count;
2377                         if (lock->ml.node == dlm->node_num) {
2378                                 mlog(0, "found a lock owned by this node still "
2379                                      "on the %s queue!  will not migrate this "
2380                                      "lockres\n", (i == 0 ? "granted" :
2381                                                    (i == 1 ? "converting" :
2382                                                     "blocked")));
2383                                 ret = -ENOTEMPTY;
2384                                 goto leave;
2385                         }
2386                 }
2387                 queue++;
2388         }
2389
2390         *numlocks = count;
2391
2392         count = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
2393         if (count < O2NM_MAX_NODES)
2394                 *hasrefs = 1;
2395
2396         mlog(0, "%s: res %.*s, Migrateable, locks %d, refs %d\n", dlm->name,
2397              res->lockname.len, res->lockname.name, *numlocks, *hasrefs);
2398
2399 leave:
2400         return ret;
2401 }
2402
2403 /*
2404  * DLM_MIGRATE_LOCKRES
2405  */
2406
2407
2408 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2409                                struct dlm_lock_resource *res,
2410                                u8 target)
2411 {
2412         struct dlm_master_list_entry *mle = NULL;
2413         struct dlm_master_list_entry *oldmle = NULL;
2414         struct dlm_migratable_lockres *mres = NULL;
2415         int ret = 0;
2416         const char *name;
2417         unsigned int namelen;
2418         int mle_added = 0;
2419         int numlocks, hasrefs;
2420         int wake = 0;
2421
2422         if (!dlm_grab(dlm))
2423                 return -EINVAL;
2424
2425         name = res->lockname.name;
2426         namelen = res->lockname.len;
2427
2428         mlog(0, "%s: Migrating %.*s to %u\n", dlm->name, namelen, name, target);
2429
2430         /*
2431          * ensure this lockres is a proper candidate for migration
2432          */
2433         spin_lock(&res->spinlock);
2434         ret = dlm_is_lockres_migrateable(dlm, res, &numlocks, &hasrefs);
2435         if (ret < 0) {
2436                 spin_unlock(&res->spinlock);
2437                 goto leave;
2438         }
2439         spin_unlock(&res->spinlock);
2440
2441         /* no work to do */
2442         if (numlocks == 0 && !hasrefs)
2443                 goto leave;
2444
2445         /*
2446          * preallocate up front
2447          * if this fails, abort
2448          */
2449
2450         ret = -ENOMEM;
2451         mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2452         if (!mres) {
2453                 mlog_errno(ret);
2454                 goto leave;
2455         }
2456
2457         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2458         if (!mle) {
2459                 mlog_errno(ret);
2460                 goto leave;
2461         }
2462         ret = 0;
2463
2464         /*
2465          * find a node to migrate the lockres to
2466          */
2467
2468         spin_lock(&dlm->spinlock);
2469         /* pick a new node */
2470         if (!test_bit(target, dlm->domain_map) ||
2471             target >= O2NM_MAX_NODES) {
2472                 target = dlm_pick_migration_target(dlm, res);
2473         }
2474         mlog(0, "%s: res %.*s, Node %u chosen for migration\n", dlm->name,
2475              namelen, name, target);
2476
2477         if (target >= O2NM_MAX_NODES ||
2478             !test_bit(target, dlm->domain_map)) {
2479                 /* target chosen is not alive */
2480                 ret = -EINVAL;
2481         }
2482
2483         if (ret) {
2484                 spin_unlock(&dlm->spinlock);
2485                 goto fail;
2486         }
2487
2488         mlog(0, "continuing with target = %u\n", target);
2489
2490         /*
2491          * clear any existing master requests and
2492          * add the migration mle to the list
2493          */
2494         spin_lock(&dlm->master_lock);
2495         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2496                                     namelen, target, dlm->node_num);
2497         spin_unlock(&dlm->master_lock);
2498         spin_unlock(&dlm->spinlock);
2499
2500         if (ret == -EEXIST) {
2501                 mlog(0, "another process is already migrating it\n");
2502                 goto fail;
2503         }
2504         mle_added = 1;
2505
2506         /*
2507          * set the MIGRATING flag and flush asts
2508          * if we fail after this we need to re-dirty the lockres
2509          */
2510         if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2511                 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2512                      "the target went down.\n", res->lockname.len,
2513                      res->lockname.name, target);
2514                 spin_lock(&res->spinlock);
2515                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2516                 wake = 1;
2517                 spin_unlock(&res->spinlock);
2518                 ret = -EINVAL;
2519         }
2520
2521 fail:
2522         if (oldmle) {
2523                 /* master is known, detach if not already detached */
2524                 dlm_mle_detach_hb_events(dlm, oldmle);
2525                 dlm_put_mle(oldmle);
2526         }
2527
2528         if (ret < 0) {
2529                 if (mle_added) {
2530                         dlm_mle_detach_hb_events(dlm, mle);
2531                         dlm_put_mle(mle);
2532                 } else if (mle) {
2533                         kmem_cache_free(dlm_mle_cache, mle);
2534                 }
2535                 goto leave;
2536         }
2537
2538         /*
2539          * at this point, we have a migration target, an mle
2540          * in the master list, and the MIGRATING flag set on
2541          * the lockres
2542          */
2543
2544         /* now that remote nodes are spinning on the MIGRATING flag,
2545          * ensure that all assert_master work is flushed. */
2546         flush_workqueue(dlm->dlm_worker);
2547
2548         /* get an extra reference on the mle.
2549          * otherwise the assert_master from the new
2550          * master will destroy this.
2551          * also, make sure that all callers of dlm_get_mle
2552          * take both dlm->spinlock and dlm->master_lock */
2553         spin_lock(&dlm->spinlock);
2554         spin_lock(&dlm->master_lock);
2555         dlm_get_mle_inuse(mle);
2556         spin_unlock(&dlm->master_lock);
2557         spin_unlock(&dlm->spinlock);
2558
2559         /* notify new node and send all lock state */
2560         /* call send_one_lockres with migration flag.
2561          * this serves as notice to the target node that a
2562          * migration is starting. */
2563         ret = dlm_send_one_lockres(dlm, res, mres, target,
2564                                    DLM_MRES_MIGRATION);
2565
2566         if (ret < 0) {
2567                 mlog(0, "migration to node %u failed with %d\n",
2568                      target, ret);
2569                 /* migration failed, detach and clean up mle */
2570                 dlm_mle_detach_hb_events(dlm, mle);
2571                 dlm_put_mle(mle);
2572                 dlm_put_mle_inuse(mle);
2573                 spin_lock(&res->spinlock);
2574                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2575                 wake = 1;
2576                 spin_unlock(&res->spinlock);
2577                 goto leave;
2578         }
2579
2580         /* at this point, the target sends a message to all nodes,
2581          * (using dlm_do_migrate_request).  this node is skipped since
2582          * we had to put an mle in the list to begin the process.  this
2583          * node now waits for target to do an assert master.  this node
2584          * will be the last one notified, ensuring that the migration
2585          * is complete everywhere.  if the target dies while this is
2586          * going on, some nodes could potentially see the target as the
2587          * master, so it is important that my recovery finds the migration
2588          * mle and sets the master to UNKNOWN. */
2589
2590
2591         /* wait for new node to assert master */
2592         while (1) {
2593                 ret = wait_event_interruptible_timeout(mle->wq,
2594                                         (atomic_read(&mle->woken) == 1),
2595                                         msecs_to_jiffies(5000));
2596
2597                 if (ret >= 0) {
2598                         if (atomic_read(&mle->woken) == 1 ||
2599                             res->owner == target)
2600                                 break;
2601
2602                         mlog(0, "%s:%.*s: timed out during migration\n",
2603                              dlm->name, res->lockname.len, res->lockname.name);
2604                         /* avoid hang during shutdown when migrating lockres
2605                          * to a node which also goes down */
2606                         if (dlm_is_node_dead(dlm, target)) {
2607                                 mlog(0, "%s:%.*s: expected migration "
2608                                      "target %u is no longer up, restarting\n",
2609                                      dlm->name, res->lockname.len,
2610                                      res->lockname.name, target);
2611                                 ret = -EINVAL;
2612                                 /* migration failed, detach and clean up mle */
2613                                 dlm_mle_detach_hb_events(dlm, mle);
2614                                 dlm_put_mle(mle);
2615                                 dlm_put_mle_inuse(mle);
2616                                 spin_lock(&res->spinlock);
2617                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2618                                 wake = 1;
2619                                 spin_unlock(&res->spinlock);
2620                                 goto leave;
2621                         }
2622                 } else
2623                         mlog(0, "%s:%.*s: caught signal during migration\n",
2624                              dlm->name, res->lockname.len, res->lockname.name);
2625         }
2626
2627         /* all done, set the owner, clear the flag */
2628         spin_lock(&res->spinlock);
2629         dlm_set_lockres_owner(dlm, res, target);
2630         res->state &= ~DLM_LOCK_RES_MIGRATING;
2631         dlm_remove_nonlocal_locks(dlm, res);
2632         spin_unlock(&res->spinlock);
2633         wake_up(&res->wq);
2634
2635         /* master is known, detach if not already detached */
2636         dlm_mle_detach_hb_events(dlm, mle);
2637         dlm_put_mle_inuse(mle);
2638         ret = 0;
2639
2640         dlm_lockres_calc_usage(dlm, res);
2641
2642 leave:
2643         /* re-dirty the lockres if we failed */
2644         if (ret < 0)
2645                 dlm_kick_thread(dlm, res);
2646
2647         /* wake up waiters if the MIGRATING flag got set
2648          * but migration failed */
2649         if (wake)
2650                 wake_up(&res->wq);
2651
2652         /* TODO: cleanup */
2653         if (mres)
2654                 free_page((unsigned long)mres);
2655
2656         dlm_put(dlm);
2657
2658         mlog(0, "returning %d\n", ret);
2659         return ret;
2660 }
2661
2662 #define DLM_MIGRATION_RETRY_MS  100
2663
2664 /* Should be called only after beginning the domain leave process.
2665  * There should not be any remaining locks on nonlocal lock resources,
2666  * and there should be no local locks left on locally mastered resources.
2667  *
2668  * Called with the dlm spinlock held, may drop it to do migration, but
2669  * will re-acquire before exit.
2670  *
2671  * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped */
2672 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2673 {
2674         int ret;
2675         int lock_dropped = 0;
2676         int numlocks, hasrefs;
2677
2678         spin_lock(&res->spinlock);
2679         if (res->owner != dlm->node_num) {
2680                 if (!__dlm_lockres_unused(res)) {
2681                         mlog(ML_ERROR, "%s:%.*s: this node is not master, "
2682                              "trying to free this but locks remain\n",
2683                              dlm->name, res->lockname.len, res->lockname.name);
2684                 }
2685                 spin_unlock(&res->spinlock);
2686                 goto leave;
2687         }
2688
2689         /* No need to migrate a lockres having no locks */
2690         ret = dlm_is_lockres_migrateable(dlm, res, &numlocks, &hasrefs);
2691         if (ret >= 0 && numlocks == 0 && !hasrefs) {
2692                 spin_unlock(&res->spinlock);
2693                 goto leave;
2694         }
2695         spin_unlock(&res->spinlock);
2696
2697         /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2698         spin_unlock(&dlm->spinlock);
2699         lock_dropped = 1;
2700         while (1) {
2701                 ret = dlm_migrate_lockres(dlm, res, O2NM_MAX_NODES);
2702                 if (ret >= 0)
2703                         break;
2704                 if (ret == -ENOTEMPTY) {
2705                         mlog(ML_ERROR, "lockres %.*s still has local locks!\n",
2706                                 res->lockname.len, res->lockname.name);
2707                         BUG();
2708                 }
2709
2710                 mlog(0, "lockres %.*s: migrate failed, "
2711                      "retrying\n", res->lockname.len,
2712                      res->lockname.name);
2713                 msleep(DLM_MIGRATION_RETRY_MS);
2714         }
2715         spin_lock(&dlm->spinlock);
2716 leave:
2717         return lock_dropped;
2718 }
2719
2720 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2721 {
2722         int ret;
2723         spin_lock(&dlm->ast_lock);
2724         spin_lock(&lock->spinlock);
2725         ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2726         spin_unlock(&lock->spinlock);
2727         spin_unlock(&dlm->ast_lock);
2728         return ret;
2729 }
2730
2731 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2732                                      struct dlm_lock_resource *res,
2733                                      u8 mig_target)
2734 {
2735         int can_proceed;
2736         spin_lock(&res->spinlock);
2737         can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2738         spin_unlock(&res->spinlock);
2739
2740         /* target has died, so make the caller break out of the
2741          * wait_event, but caller must recheck the domain_map */
2742         spin_lock(&dlm->spinlock);
2743         if (!test_bit(mig_target, dlm->domain_map))
2744                 can_proceed = 1;
2745         spin_unlock(&dlm->spinlock);
2746         return can_proceed;
2747 }
2748
2749 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2750                                 struct dlm_lock_resource *res)
2751 {
2752         int ret;
2753         spin_lock(&res->spinlock);
2754         ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2755         spin_unlock(&res->spinlock);
2756         return ret;
2757 }
2758
2759
2760 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2761                                        struct dlm_lock_resource *res,
2762                                        u8 target)
2763 {
2764         int ret = 0;
2765
2766         mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2767                res->lockname.len, res->lockname.name, dlm->node_num,
2768                target);
2769         /* need to set MIGRATING flag on lockres.  this is done by
2770          * ensuring that all asts have been flushed for this lockres. */
2771         spin_lock(&res->spinlock);
2772         BUG_ON(res->migration_pending);
2773         res->migration_pending = 1;
2774         /* strategy is to reserve an extra ast then release
2775          * it below, letting the release do all of the work */
2776         __dlm_lockres_reserve_ast(res);
2777         spin_unlock(&res->spinlock);
2778
2779         /* now flush all the pending asts */
2780         dlm_kick_thread(dlm, res);
2781         /* before waiting on DIRTY, block processes which may
2782          * try to dirty the lockres before MIGRATING is set */
2783         spin_lock(&res->spinlock);
2784         BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2785         res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2786         spin_unlock(&res->spinlock);
2787         /* now wait on any pending asts and the DIRTY state */
2788         wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2789         dlm_lockres_release_ast(dlm, res);
2790
2791         mlog(0, "about to wait on migration_wq, dirty=%s\n",
2792                res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2793         /* if the extra ref we just put was the final one, this
2794          * will pass thru immediately.  otherwise, we need to wait
2795          * for the last ast to finish. */
2796 again:
2797         ret = wait_event_interruptible_timeout(dlm->migration_wq,
2798                    dlm_migration_can_proceed(dlm, res, target),
2799                    msecs_to_jiffies(1000));
2800         if (ret < 0) {
2801                 mlog(0, "woken again: migrating? %s, dead? %s\n",
2802                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2803                        test_bit(target, dlm->domain_map) ? "no":"yes");
2804         } else {
2805                 mlog(0, "all is well: migrating? %s, dead? %s\n",
2806                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2807                        test_bit(target, dlm->domain_map) ? "no":"yes");
2808         }
2809         if (!dlm_migration_can_proceed(dlm, res, target)) {
2810                 mlog(0, "trying again...\n");
2811                 goto again;
2812         }
2813
2814         ret = 0;
2815         /* did the target go down or die? */
2816         spin_lock(&dlm->spinlock);
2817         if (!test_bit(target, dlm->domain_map)) {
2818                 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2819                      target);
2820                 ret = -EHOSTDOWN;
2821         }
2822         spin_unlock(&dlm->spinlock);
2823
2824         /*
2825          * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2826          * another try; otherwise, we are sure the MIGRATING state is there,
2827          * drop the unneded state which blocked threads trying to DIRTY
2828          */
2829         spin_lock(&res->spinlock);
2830         BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2831         res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2832         if (!ret)
2833                 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2834         spin_unlock(&res->spinlock);
2835
2836         /*
2837          * at this point:
2838          *
2839          *   o the DLM_LOCK_RES_MIGRATING flag is set if target not down
2840          *   o there are no pending asts on this lockres
2841          *   o all processes trying to reserve an ast on this
2842          *     lockres must wait for the MIGRATING flag to clear
2843          */
2844         return ret;
2845 }
2846
2847 /* last step in the migration process.
2848  * original master calls this to free all of the dlm_lock
2849  * structures that used to be for other nodes. */
2850 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2851                                       struct dlm_lock_resource *res)
2852 {
2853         struct list_head *queue = &res->granted;
2854         int i, bit;
2855         struct dlm_lock *lock, *next;
2856
2857         assert_spin_locked(&res->spinlock);
2858
2859         BUG_ON(res->owner == dlm->node_num);
2860
2861         for (i=0; i<3; i++) {
2862                 list_for_each_entry_safe(lock, next, queue, list) {
2863                         if (lock->ml.node != dlm->node_num) {
2864                                 mlog(0, "putting lock for node %u\n",
2865                                      lock->ml.node);
2866                                 /* be extra careful */
2867                                 BUG_ON(!list_empty(&lock->ast_list));
2868                                 BUG_ON(!list_empty(&lock->bast_list));
2869                                 BUG_ON(lock->ast_pending);
2870                                 BUG_ON(lock->bast_pending);
2871                                 dlm_lockres_clear_refmap_bit(lock->ml.node, res);
2872                                 list_del_init(&lock->list);
2873                                 dlm_lock_put(lock);
2874                                 /* In a normal unlock, we would have added a
2875                                  * DLM_UNLOCK_FREE_LOCK action. Force it. */
2876                                 dlm_lock_put(lock);
2877                         }
2878                 }
2879                 queue++;
2880         }
2881         bit = 0;
2882         while (1) {
2883                 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2884                 if (bit >= O2NM_MAX_NODES)
2885                         break;
2886                 /* do not clear the local node reference, if there is a
2887                  * process holding this, let it drop the ref itself */
2888                 if (bit != dlm->node_num) {
2889                         mlog(0, "%s:%.*s: node %u had a ref to this "
2890                              "migrating lockres, clearing\n", dlm->name,
2891                              res->lockname.len, res->lockname.name, bit);
2892                         dlm_lockres_clear_refmap_bit(bit, res);
2893                 }
2894                 bit++;
2895         }
2896 }
2897
2898 /* for now this is not too intelligent.  we will
2899  * need stats to make this do the right thing.
2900  * this just finds the first lock on one of the
2901  * queues and uses that node as the target. */
2902 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2903                                     struct dlm_lock_resource *res)
2904 {
2905         int i;
2906         struct list_head *queue = &res->granted;
2907         struct dlm_lock *lock;
2908         int nodenum;
2909
2910         assert_spin_locked(&dlm->spinlock);
2911
2912         spin_lock(&res->spinlock);
2913         for (i=0; i<3; i++) {
2914                 list_for_each_entry(lock, queue, list) {
2915                         /* up to the caller to make sure this node
2916                          * is alive */
2917                         if (lock->ml.node != dlm->node_num) {
2918                                 spin_unlock(&res->spinlock);
2919                                 return lock->ml.node;
2920                         }
2921                 }
2922                 queue++;
2923         }
2924
2925         nodenum = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
2926         if (nodenum < O2NM_MAX_NODES) {
2927                 spin_unlock(&res->spinlock);
2928                 return nodenum;
2929         }
2930         spin_unlock(&res->spinlock);
2931         mlog(0, "have not found a suitable target yet! checking domain map\n");
2932
2933         /* ok now we're getting desperate.  pick anyone alive. */
2934         nodenum = -1;
2935         while (1) {
2936                 nodenum = find_next_bit(dlm->domain_map,
2937                                         O2NM_MAX_NODES, nodenum+1);
2938                 mlog(0, "found %d in domain map\n", nodenum);
2939                 if (nodenum >= O2NM_MAX_NODES)
2940                         break;
2941                 if (nodenum != dlm->node_num) {
2942                         mlog(0, "picking %d\n", nodenum);
2943                         return nodenum;
2944                 }
2945         }
2946
2947         mlog(0, "giving up.  no master to migrate to\n");
2948         return DLM_LOCK_RES_OWNER_UNKNOWN;
2949 }
2950
2951
2952
2953 /* this is called by the new master once all lockres
2954  * data has been received */
2955 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2956                                   struct dlm_lock_resource *res,
2957                                   u8 master, u8 new_master,
2958                                   struct dlm_node_iter *iter)
2959 {
2960         struct dlm_migrate_request migrate;
2961         int ret, skip, status = 0;
2962         int nodenum;
2963
2964         memset(&migrate, 0, sizeof(migrate));
2965         migrate.namelen = res->lockname.len;
2966         memcpy(migrate.name, res->lockname.name, migrate.namelen);
2967         migrate.new_master = new_master;
2968         migrate.master = master;
2969
2970         ret = 0;
2971
2972         /* send message to all nodes, except the master and myself */
2973         while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2974                 if (nodenum == master ||
2975                     nodenum == new_master)
2976                         continue;
2977
2978                 /* We could race exit domain. If exited, skip. */
2979                 spin_lock(&dlm->spinlock);
2980                 skip = (!test_bit(nodenum, dlm->domain_map));
2981                 spin_unlock(&dlm->spinlock);
2982                 if (skip) {
2983                         clear_bit(nodenum, iter->node_map);
2984                         continue;
2985                 }
2986
2987                 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2988                                          &migrate, sizeof(migrate), nodenum,
2989                                          &status);
2990                 if (ret < 0) {
2991                         mlog(ML_ERROR, "Error %d when sending message %u (key "
2992                              "0x%x) to node %u\n", ret, DLM_MIGRATE_REQUEST_MSG,
2993                              dlm->key, nodenum);
2994                         if (!dlm_is_host_down(ret)) {
2995                                 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
2996                                 BUG();
2997                         }
2998                         clear_bit(nodenum, iter->node_map);
2999                         ret = 0;
3000                 } else if (status < 0) {
3001                         mlog(0, "migrate request (node %u) returned %d!\n",
3002                              nodenum, status);
3003                         ret = status;
3004                 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
3005                         /* during the migration request we short-circuited
3006                          * the mastery of the lockres.  make sure we have
3007                          * a mastery ref for nodenum */
3008                         mlog(0, "%s:%.*s: need ref for node %u\n",
3009                              dlm->name, res->lockname.len, res->lockname.name,
3010                              nodenum);
3011                         spin_lock(&res->spinlock);
3012                         dlm_lockres_set_refmap_bit(nodenum, res);
3013                         spin_unlock(&res->spinlock);
3014                 }
3015         }
3016
3017         if (ret < 0)
3018                 mlog_errno(ret);
3019
3020         mlog(0, "returning ret=%d\n", ret);
3021         return ret;
3022 }
3023
3024
3025 /* if there is an existing mle for this lockres, we now know who the master is.
3026  * (the one who sent us *this* message) we can clear it up right away.
3027  * since the process that put the mle on the list still has a reference to it,
3028  * we can unhash it now, set the master and wake the process.  as a result,
3029  * we will have no mle in the list to start with.  now we can add an mle for
3030  * the migration and this should be the only one found for those scanning the
3031  * list.  */
3032 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
3033                                 void **ret_data)
3034 {
3035         struct dlm_ctxt *dlm = data;
3036         struct dlm_lock_resource *res = NULL;
3037         struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
3038         struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
3039         const char *name;
3040         unsigned int namelen, hash;
3041         int ret = 0;
3042
3043         if (!dlm_grab(dlm))
3044                 return -EINVAL;
3045
3046         name = migrate->name;
3047         namelen = migrate->namelen;
3048         hash = dlm_lockid_hash(name, namelen);
3049
3050         /* preallocate.. if this fails, abort */
3051         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
3052
3053         if (!mle) {
3054                 ret = -ENOMEM;
3055                 goto leave;
3056         }
3057
3058         /* check for pre-existing lock */
3059         spin_lock(&dlm->spinlock);
3060         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
3061         if (res) {
3062                 spin_lock(&res->spinlock);
3063                 if (res->state & DLM_LOCK_RES_RECOVERING) {
3064                         /* if all is working ok, this can only mean that we got
3065                         * a migrate request from a node that we now see as
3066                         * dead.  what can we do here?  drop it to the floor? */
3067                         spin_unlock(&res->spinlock);
3068                         mlog(ML_ERROR, "Got a migrate request, but the "
3069                              "lockres is marked as recovering!");
3070                         kmem_cache_free(dlm_mle_cache, mle);
3071                         ret = -EINVAL; /* need a better solution */
3072                         goto unlock;
3073                 }
3074                 res->state |= DLM_LOCK_RES_MIGRATING;
3075                 spin_unlock(&res->spinlock);
3076         }
3077
3078         spin_lock(&dlm->master_lock);
3079         /* ignore status.  only nonzero status would BUG. */
3080         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3081                                     name, namelen,
3082                                     migrate->new_master,
3083                                     migrate->master);
3084
3085         spin_unlock(&dlm->master_lock);
3086 unlock:
3087         spin_unlock(&dlm->spinlock);
3088
3089         if (oldmle) {
3090                 /* master is known, detach if not already detached */
3091                 dlm_mle_detach_hb_events(dlm, oldmle);
3092                 dlm_put_mle(oldmle);
3093         }
3094
3095         if (res)
3096                 dlm_lockres_put(res);
3097 leave:
3098         dlm_put(dlm);
3099         return ret;
3100 }
3101
3102 /* must be holding dlm->spinlock and dlm->master_lock
3103  * when adding a migration mle, we can clear any other mles
3104  * in the master list because we know with certainty that
3105  * the master is "master".  so we remove any old mle from
3106  * the list after setting it's master field, and then add
3107  * the new migration mle.  this way we can hold with the rule
3108  * of having only one mle for a given lock name at all times. */
3109 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3110                                  struct dlm_lock_resource *res,
3111                                  struct dlm_master_list_entry *mle,
3112                                  struct dlm_master_list_entry **oldmle,
3113                                  const char *name, unsigned int namelen,
3114                                  u8 new_master, u8 master)
3115 {
3116         int found;
3117         int ret = 0;
3118
3119         *oldmle = NULL;
3120
3121         assert_spin_locked(&dlm->spinlock);
3122         assert_spin_locked(&dlm->master_lock);
3123
3124         /* caller is responsible for any ref taken here on oldmle */
3125         found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3126         if (found) {
3127                 struct dlm_master_list_entry *tmp = *oldmle;
3128                 spin_lock(&tmp->spinlock);
3129                 if (tmp->type == DLM_MLE_MIGRATION) {
3130                         if (master == dlm->node_num) {
3131                                 /* ah another process raced me to it */
3132                                 mlog(0, "tried to migrate %.*s, but some "
3133                                      "process beat me to it\n",
3134                                      namelen, name);
3135                                 ret = -EEXIST;
3136                         } else {
3137                                 /* bad.  2 NODES are trying to migrate! */
3138                                 mlog(ML_ERROR, "migration error  mle: "
3139                                      "master=%u new_master=%u // request: "
3140                                      "master=%u new_master=%u // "
3141                                      "lockres=%.*s\n",
3142                                      tmp->master, tmp->new_master,
3143                                      master, new_master,
3144                                      namelen, name);
3145                                 BUG();
3146                         }
3147                 } else {
3148                         /* this is essentially what assert_master does */
3149                         tmp->master = master;
3150                         atomic_set(&tmp->woken, 1);
3151                         wake_up(&tmp->wq);
3152                         /* remove it so that only one mle will be found */
3153                         __dlm_unlink_mle(dlm, tmp);
3154                         __dlm_mle_detach_hb_events(dlm, tmp);
3155                         ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3156                         mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3157                             "telling master to get ref for cleared out mle "
3158                             "during migration\n", dlm->name, namelen, name,
3159                             master, new_master);
3160                 }
3161                 spin_unlock(&tmp->spinlock);
3162         }
3163
3164         /* now add a migration mle to the tail of the list */
3165         dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3166         mle->new_master = new_master;
3167         /* the new master will be sending an assert master for this.
3168          * at that point we will get the refmap reference */
3169         mle->master = master;
3170         /* do this for consistency with other mle types */
3171         set_bit(new_master, mle->maybe_map);
3172         __dlm_insert_mle(dlm, mle);
3173
3174         return ret;
3175 }
3176
3177 /*
3178  * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3179  */
3180 static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3181                                         struct dlm_master_list_entry *mle)
3182 {
3183         struct dlm_lock_resource *res;
3184
3185         /* Find the lockres associated to the mle and set its owner to UNK */
3186         res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3187                                    mle->mnamehash);
3188         if (res) {
3189                 spin_unlock(&dlm->master_lock);
3190
3191                 /* move lockres onto recovery list */
3192                 spin_lock(&res->spinlock);
3193                 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3194                 dlm_move_lockres_to_recovery_list(dlm, res);
3195                 spin_unlock(&res->spinlock);
3196                 dlm_lockres_put(res);
3197
3198                 /* about to get rid of mle, detach from heartbeat */
3199                 __dlm_mle_detach_hb_events(dlm, mle);
3200
3201                 /* dump the mle */
3202                 spin_lock(&dlm->master_lock);
3203                 __dlm_put_mle(mle);
3204                 spin_unlock(&dlm->master_lock);
3205         }
3206
3207         return res;
3208 }
3209
3210 static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3211                                     struct dlm_master_list_entry *mle)
3212 {
3213         __dlm_mle_detach_hb_events(dlm, mle);
3214
3215         spin_lock(&mle->spinlock);
3216         __dlm_unlink_mle(dlm, mle);
3217         atomic_set(&mle->woken, 1);
3218         spin_unlock(&mle->spinlock);
3219
3220         wake_up(&mle->wq);
3221 }
3222
3223 static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3224                                 struct dlm_master_list_entry *mle, u8 dead_node)
3225 {
3226         int bit;
3227
3228         BUG_ON(mle->type != DLM_MLE_BLOCK);
3229
3230         spin_lock(&mle->spinlock);
3231         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3232         if (bit != dead_node) {
3233                 mlog(0, "mle found, but dead node %u would not have been "
3234                      "master\n", dead_node);
3235                 spin_unlock(&mle->spinlock);
3236         } else {
3237                 /* Must drop the refcount by one since the assert_master will
3238                  * never arrive. This may result in the mle being unlinked and
3239                  * freed, but there may still be a process waiting in the
3240                  * dlmlock path which is fine. */
3241                 mlog(0, "node %u was expected master\n", dead_node);
3242                 atomic_set(&mle->woken, 1);
3243                 spin_unlock(&mle->spinlock);
3244                 wake_up(&mle->wq);
3245
3246                 /* Do not need events any longer, so detach from heartbeat */
3247                 __dlm_mle_detach_hb_events(dlm, mle);
3248                 __dlm_put_mle(mle);
3249         }
3250 }
3251
3252 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3253 {
3254         struct dlm_master_list_entry *mle;
3255         struct dlm_lock_resource *res;
3256         struct hlist_head *bucket;
3257         struct hlist_node *list;
3258         unsigned int i;
3259
3260         mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
3261 top:
3262         assert_spin_locked(&dlm->spinlock);
3263
3264         /* clean the master list */
3265         spin_lock(&dlm->master_lock);
3266         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3267                 bucket = dlm_master_hash(dlm, i);
3268                 hlist_for_each(list, bucket) {
3269                         mle = hlist_entry(list, struct dlm_master_list_entry,
3270                                           master_hash_node);
3271
3272                         BUG_ON(mle->type != DLM_MLE_BLOCK &&
3273                                mle->type != DLM_MLE_MASTER &&
3274                                mle->type != DLM_MLE_MIGRATION);
3275
3276                         /* MASTER mles are initiated locally. The waiting
3277                          * process will notice the node map change shortly.
3278                          * Let that happen as normal. */
3279                         if (mle->type == DLM_MLE_MASTER)
3280                                 continue;
3281
3282                         /* BLOCK mles are initiated by other nodes. Need to
3283                          * clean up if the dead node would have been the
3284                          * master. */
3285                         if (mle->type == DLM_MLE_BLOCK) {
3286                                 dlm_clean_block_mle(dlm, mle, dead_node);
3287                                 continue;
3288                         }
3289
3290                         /* Everything else is a MIGRATION mle */
3291
3292                         /* The rule for MIGRATION mles is that the master
3293                          * becomes UNKNOWN if *either* the original or the new
3294                          * master dies. All UNKNOWN lockres' are sent to
3295                          * whichever node becomes the recovery master. The new
3296                          * master is responsible for determining if there is
3297                          * still a master for this lockres, or if he needs to
3298                          * take over mastery. Either way, this node should
3299                          * expect another message to resolve this. */
3300
3301                         if (mle->master != dead_node &&
3302                             mle->new_master != dead_node)
3303                                 continue;
3304
3305                         /* If we have reached this point, this mle needs to be
3306                          * removed from the list and freed. */
3307                         dlm_clean_migration_mle(dlm, mle);
3308
3309                         mlog(0, "%s: node %u died during migration from "
3310                              "%u to %u!\n", dlm->name, dead_node, mle->master,
3311                              mle->new_master);
3312
3313                         /* If we find a lockres associated with the mle, we've
3314                          * hit this rare case that messes up our lock ordering.
3315                          * If so, we need to drop the master lock so that we can
3316                          * take the lockres lock, meaning that we will have to
3317                          * restart from the head of list. */
3318                         res = dlm_reset_mleres_owner(dlm, mle);
3319                         if (res)
3320                                 /* restart */
3321                                 goto top;
3322
3323                         /* This may be the last reference */
3324                         __dlm_put_mle(mle);
3325                 }
3326         }
3327         spin_unlock(&dlm->master_lock);
3328 }
3329
3330 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3331                          u8 old_master)
3332 {
3333         struct dlm_node_iter iter;
3334         int ret = 0;
3335
3336         spin_lock(&dlm->spinlock);
3337         dlm_node_iter_init(dlm->domain_map, &iter);
3338         clear_bit(old_master, iter.node_map);
3339         clear_bit(dlm->node_num, iter.node_map);
3340         spin_unlock(&dlm->spinlock);
3341
3342         /* ownership of the lockres is changing.  account for the
3343          * mastery reference here since old_master will briefly have
3344          * a reference after the migration completes */
3345         spin_lock(&res->spinlock);
3346         dlm_lockres_set_refmap_bit(old_master, res);
3347         spin_unlock(&res->spinlock);
3348
3349         mlog(0, "now time to do a migrate request to other nodes\n");
3350         ret = dlm_do_migrate_request(dlm, res, old_master,
3351                                      dlm->node_num, &iter);
3352         if (ret < 0) {
3353                 mlog_errno(ret);
3354                 goto leave;
3355         }
3356
3357         mlog(0, "doing assert master of %.*s to all except the original node\n",
3358              res->lockname.len, res->lockname.name);
3359         /* this call now finishes out the nodemap
3360          * even if one or more nodes die */
3361         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3362                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3363         if (ret < 0) {
3364                 /* no longer need to retry.  all living nodes contacted. */
3365                 mlog_errno(ret);
3366                 ret = 0;
3367         }
3368
3369         memset(iter.node_map, 0, sizeof(iter.node_map));
3370         set_bit(old_master, iter.node_map);
3371         mlog(0, "doing assert master of %.*s back to %u\n",
3372              res->lockname.len, res->lockname.name, old_master);
3373         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3374                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3375         if (ret < 0) {
3376                 mlog(0, "assert master to original master failed "
3377                      "with %d.\n", ret);
3378                 /* the only nonzero status here would be because of
3379                  * a dead original node.  we're done. */
3380                 ret = 0;
3381         }
3382
3383         /* all done, set the owner, clear the flag */
3384         spin_lock(&res->spinlock);
3385         dlm_set_lockres_owner(dlm, res, dlm->node_num);
3386         res->state &= ~DLM_LOCK_RES_MIGRATING;
3387         spin_unlock(&res->spinlock);
3388         /* re-dirty it on the new master */
3389         dlm_kick_thread(dlm, res);
3390         wake_up(&res->wq);
3391 leave:
3392         return ret;
3393 }
3394
3395 /*
3396  * LOCKRES AST REFCOUNT
3397  * this is integral to migration
3398  */
3399
3400 /* for future intent to call an ast, reserve one ahead of time.
3401  * this should be called only after waiting on the lockres
3402  * with dlm_wait_on_lockres, and while still holding the
3403  * spinlock after the call. */
3404 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3405 {
3406         assert_spin_locked(&res->spinlock);
3407         if (res->state & DLM_LOCK_RES_MIGRATING) {
3408                 __dlm_print_one_lock_resource(res);
3409         }
3410         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3411
3412         atomic_inc(&res->asts_reserved);
3413 }
3414
3415 /*
3416  * used to drop the reserved ast, either because it went unused,
3417  * or because the ast/bast was actually called.
3418  *
3419  * also, if there is a pending migration on this lockres,
3420  * and this was the last pending ast on the lockres,
3421  * atomically set the MIGRATING flag before we drop the lock.
3422  * this is how we ensure that migration can proceed with no
3423  * asts in progress.  note that it is ok if the state of the
3424  * queues is such that a lock should be granted in the future
3425  * or that a bast should be fired, because the new master will
3426  * shuffle the lists on this lockres as soon as it is migrated.
3427  */
3428 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3429                              struct dlm_lock_resource *res)
3430 {
3431         if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3432                 return;
3433
3434         if (!res->migration_pending) {
3435                 spin_unlock(&res->spinlock);
3436                 return;
3437         }
3438
3439         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3440         res->migration_pending = 0;
3441         res->state |= DLM_LOCK_RES_MIGRATING;
3442         spin_unlock(&res->spinlock);
3443         wake_up(&res->wq);
3444         wake_up(&dlm->migration_wq);
3445 }
3446
3447 void dlm_force_free_mles(struct dlm_ctxt *dlm)
3448 {
3449         int i;
3450         struct hlist_head *bucket;
3451         struct dlm_master_list_entry *mle;
3452         struct hlist_node *tmp, *list;
3453
3454         /*
3455          * We notified all other nodes that we are exiting the domain and
3456          * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3457          * around we force free them and wake any processes that are waiting
3458          * on the mles
3459          */
3460         spin_lock(&dlm->spinlock);
3461         spin_lock(&dlm->master_lock);
3462
3463         BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3464         BUG_ON((find_next_bit(dlm->domain_map, O2NM_MAX_NODES, 0) < O2NM_MAX_NODES));
3465
3466         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3467                 bucket = dlm_master_hash(dlm, i);
3468                 hlist_for_each_safe(list, tmp, bucket) {
3469                         mle = hlist_entry(list, struct dlm_master_list_entry,
3470                                           master_hash_node);
3471                         if (mle->type != DLM_MLE_BLOCK) {
3472                                 mlog(ML_ERROR, "bad mle: %p\n", mle);
3473                                 dlm_print_one_mle(mle);
3474                         }
3475                         atomic_set(&mle->woken, 1);
3476                         wake_up(&mle->wq);
3477
3478                         __dlm_unlink_mle(dlm, mle);
3479                         __dlm_mle_detach_hb_events(dlm, mle);
3480                         __dlm_put_mle(mle);
3481                 }
3482         }
3483         spin_unlock(&dlm->master_lock);
3484         spin_unlock(&dlm->spinlock);
3485 }