fs/ocfs2/dlm/dlmmaster.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmmod.c
   5  *
   6  * standalone DLM module
   7  *
   8  * Copyright (C) 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  *
  25  */
  26
  27
  28 #include <linux/module.h>
  29 #include <linux/fs.h>
  30 #include <linux/types.h>
  31 #include <linux/slab.h>
  32 #include <linux/highmem.h>
  33 #include <linux/utsname.h>
  34 #include <linux/init.h>
  35 #include <linux/sysctl.h>
  36 #include <linux/random.h>
  37 #include <linux/blkdev.h>
  38 #include <linux/socket.h>
  39 #include <linux/inet.h>
  40 #include <linux/spinlock.h>
  41 #include <linux/delay.h>
  42
  43
  44 #include "cluster/heartbeat.h"
  45 #include "cluster/nodemanager.h"
  46 #include "cluster/tcp.h"
  47
  48 #include "dlmapi.h"
  49 #include "dlmcommon.h"
  50 #include "dlmdomain.h"
  51
  52 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
  53 #include "cluster/masklog.h"
  54
  55 enum dlm_mle_type {
  56         DLM_MLE_BLOCK,
  57         DLM_MLE_MASTER,
  58         DLM_MLE_MIGRATION
  59 };
  60
  61 struct dlm_lock_name
  62 {
  63         u8 len;
  64         u8 name[DLM_LOCKID_NAME_MAX];
  65 };
  66
  67 struct dlm_master_list_entry
  68 {
  69         struct list_head list;
  70         struct list_head hb_events;
  71         struct dlm_ctxt *dlm;
  72         spinlock_t spinlock;
  73         wait_queue_head_t wq;
  74         atomic_t woken;
  75         struct kref mle_refs;
  76         int inuse;
  77         unsigned long maybe_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  78         unsigned long vote_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  79         unsigned long response_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  80         unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  81         u8 master;
  82         u8 new_master;
  83         enum dlm_mle_type type;
  84         struct o2hb_callback_func mle_hb_up;
  85         struct o2hb_callback_func mle_hb_down;
  86         union {
  87                 struct dlm_lock_resource *res;
  88                 struct dlm_lock_name name;
  89         } u;
  90 };
  91
  92 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
  93                               struct dlm_master_list_entry *mle,
  94                               struct o2nm_node *node,
  95                               int idx);
  96 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
  97                             struct dlm_master_list_entry *mle,
  98                             struct o2nm_node *node,
  99                             int idx);
 100
 101 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
 102 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
 103                                 struct dlm_lock_resource *res,
 104                                 void *nodemap, u32 flags);
 105
 106 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
 107                                 struct dlm_master_list_entry *mle,
 108                                 const char *name,
 109                                 unsigned int namelen)
 110 {
 111         struct dlm_lock_resource *res;
 112
 113         if (dlm != mle->dlm)
 114                 return 0;
 115
 116         if (mle->type == DLM_MLE_BLOCK ||
 117             mle->type == DLM_MLE_MIGRATION) {
 118                 if (namelen != mle->u.name.len ||
 119                     memcmp(name, mle->u.name.name, namelen)!=0)
 120                         return 0;
 121         } else {
 122                 res = mle->u.res;
 123                 if (namelen != res->lockname.len ||
 124                     memcmp(res->lockname.name, name, namelen) != 0)
 125                         return 0;
 126         }
 127         return 1;
 128 }
 129
 130 #define dlm_print_nodemap(m)  _dlm_print_nodemap(m,#m)
 131 static void _dlm_print_nodemap(unsigned long *map, const char *mapname)
 132 {
 133         int i;
 134         printk("%s=[ ", mapname);
 135         for (i=0; i<O2NM_MAX_NODES; i++)
 136                 if (test_bit(i, map))
 137                         printk("%d ", i);
 138         printk("]");
 139 }
 140
 141 static void dlm_print_one_mle(struct dlm_master_list_entry *mle)
 142 {
 143         int refs;
 144         char *type;
 145         char attached;
 146         u8 master;
 147         unsigned int namelen;
 148         const char *name;
 149         struct kref *k;
 150         unsigned long *maybe = mle->maybe_map,
 151                       *vote = mle->vote_map,
 152                       *resp = mle->response_map,
 153                       *node = mle->node_map;
 154
 155         k = &mle->mle_refs;
 156         if (mle->type == DLM_MLE_BLOCK)
 157                 type = "BLK";
 158         else if (mle->type == DLM_MLE_MASTER)
 159                 type = "MAS";
 160         else
 161                 type = "MIG";
 162         refs = atomic_read(&k->refcount);
 163         master = mle->master;
 164         attached = (list_empty(&mle->hb_events) ? 'N' : 'Y');
 165
 166         if (mle->type != DLM_MLE_MASTER) {
 167                 namelen = mle->u.name.len;
 168                 name = mle->u.name.name;
 169         } else {
 170                 namelen = mle->u.res->lockname.len;
 171                 name = mle->u.res->lockname.name;
 172         }
 173
 174         mlog(ML_NOTICE, "%.*s: %3s refs=%3d mas=%3u new=%3u evt=%c inuse=%d ",
 175                   namelen, name, type, refs, master, mle->new_master, attached,
 176                   mle->inuse);
 177         dlm_print_nodemap(maybe);
 178         printk(", ");
 179         dlm_print_nodemap(vote);
 180         printk(", ");
 181         dlm_print_nodemap(resp);
 182         printk(", ");
 183         dlm_print_nodemap(node);
 184         printk(", ");
 185         printk("\n");
 186 }
 187
 188 #if 0
 189 /* Code here is included but defined out as it aids debugging */
 190
 191 static void dlm_dump_mles(struct dlm_ctxt *dlm)
 192 {
 193         struct dlm_master_list_entry *mle;
 194         struct list_head *iter;
 195
 196         mlog(ML_NOTICE, "dumping all mles for domain %s:\n", dlm->name);
 197         spin_lock(&dlm->master_lock);
 198         list_for_each(iter, &dlm->master_list) {
 199                 mle = list_entry(iter, struct dlm_master_list_entry, list);
 200                 dlm_print_one_mle(mle);
 201         }
 202         spin_unlock(&dlm->master_lock);
 203 }
 204
 205 int dlm_dump_all_mles(const char __user *data, unsigned int len)
 206 {
 207         struct list_head *iter;
 208         struct dlm_ctxt *dlm;
 209
 210         spin_lock(&dlm_domain_lock);
 211         list_for_each(iter, &dlm_domains) {
 212                 dlm = list_entry (iter, struct dlm_ctxt, list);
 213                 mlog(ML_NOTICE, "found dlm: %p, name=%s\n", dlm, dlm->name);
 214                 dlm_dump_mles(dlm);
 215         }
 216         spin_unlock(&dlm_domain_lock);
 217         return len;
 218 }
 219 EXPORT_SYMBOL_GPL(dlm_dump_all_mles);
 220
 221 #endif  /*  0  */
 222
 223
 224 static struct kmem_cache *dlm_mle_cache = NULL;
 225
 226
 227 static void dlm_mle_release(struct kref *kref);
 228 static void dlm_init_mle(struct dlm_master_list_entry *mle,
 229                         enum dlm_mle_type type,
 230                         struct dlm_ctxt *dlm,
 231                         struct dlm_lock_resource *res,
 232                         const char *name,
 233                         unsigned int namelen);
 234 static void dlm_put_mle(struct dlm_master_list_entry *mle);
 235 static void __dlm_put_mle(struct dlm_master_list_entry *mle);
 236 static int dlm_find_mle(struct dlm_ctxt *dlm,
 237                         struct dlm_master_list_entry **mle,
 238                         char *name, unsigned int namelen);
 239
 240 static int dlm_do_master_request(struct dlm_lock_resource *res,
 241                                  struct dlm_master_list_entry *mle, int to);
 242
 243
 244 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
 245                                      struct dlm_lock_resource *res,
 246                                      struct dlm_master_list_entry *mle,
 247                                      int *blocked);
 248 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
 249                                     struct dlm_lock_resource *res,
 250                                     struct dlm_master_list_entry *mle,
 251                                     int blocked);
 252 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
 253                                  struct dlm_lock_resource *res,
 254                                  struct dlm_master_list_entry *mle,
 255                                  struct dlm_master_list_entry **oldmle,
 256                                  const char *name, unsigned int namelen,
 257                                  u8 new_master, u8 master);
 258
 259 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
 260                                     struct dlm_lock_resource *res);
 261 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
 262                                       struct dlm_lock_resource *res);
 263 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
 264                                        struct dlm_lock_resource *res,
 265                                        u8 target);
 266 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
 267                                        struct dlm_lock_resource *res);
 268
 269
 270 int dlm_is_host_down(int errno)
 271 {
 272         switch (errno) {
 273                 case -EBADF:
 274                 case -ECONNREFUSED:
 275                 case -ENOTCONN:
 276                 case -ECONNRESET:
 277                 case -EPIPE:
 278                 case -EHOSTDOWN:
 279                 case -EHOSTUNREACH:
 280                 case -ETIMEDOUT:
 281                 case -ECONNABORTED:
 282                 case -ENETDOWN:
 283                 case -ENETUNREACH:
 284                 case -ENETRESET:
 285                 case -ESHUTDOWN:
 286                 case -ENOPROTOOPT:
 287                 case -EINVAL:   /* if returned from our tcp code,
 288                                    this means there is no socket */
 289                         return 1;
 290         }
 291         return 0;
 292 }
 293
 294
 295 /*
 296  * MASTER LIST FUNCTIONS
 297  */
 298
 299
 300 /*
 301  * regarding master list entries and heartbeat callbacks:
 302  *
 303  * in order to avoid sleeping and allocation that occurs in
 304  * heartbeat, master list entries are simply attached to the
 305  * dlm's established heartbeat callbacks.  the mle is attached
 306  * when it is created, and since the dlm->spinlock is held at
 307  * that time, any heartbeat event will be properly discovered
 308  * by the mle.  the mle needs to be detached from the
 309  * dlm->mle_hb_events list as soon as heartbeat events are no
 310  * longer useful to the mle, and before the mle is freed.
 311  *
 312  * as a general rule, heartbeat events are no longer needed by
 313  * the mle once an "answer" regarding the lock master has been
 314  * received.
 315  */
 316 static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
 317                                               struct dlm_master_list_entry *mle)
 318 {
 319         assert_spin_locked(&dlm->spinlock);
 320
 321         list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
 322 }
 323
 324
 325 static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 326                                               struct dlm_master_list_entry *mle)
 327 {
 328         if (!list_empty(&mle->hb_events))
 329                 list_del_init(&mle->hb_events);
 330 }
 331
 332
 333 static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 334                                             struct dlm_master_list_entry *mle)
 335 {
 336         spin_lock(&dlm->spinlock);
 337         __dlm_mle_detach_hb_events(dlm, mle);
 338         spin_unlock(&dlm->spinlock);
 339 }
 340
 341 static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
 342 {
 343         struct dlm_ctxt *dlm;
 344         dlm = mle->dlm;
 345
 346         assert_spin_locked(&dlm->spinlock);
 347         assert_spin_locked(&dlm->master_lock);
 348         mle->inuse++;
 349         kref_get(&mle->mle_refs);
 350 }
 351
 352 static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
 353 {
 354         struct dlm_ctxt *dlm;
 355         dlm = mle->dlm;
 356
 357         spin_lock(&dlm->spinlock);
 358         spin_lock(&dlm->master_lock);
 359         mle->inuse--;
 360         __dlm_put_mle(mle);
 361         spin_unlock(&dlm->master_lock);
 362         spin_unlock(&dlm->spinlock);
 363
 364 }
 365
 366 /* remove from list and free */
 367 static void __dlm_put_mle(struct dlm_master_list_entry *mle)
 368 {
 369         struct dlm_ctxt *dlm;
 370         dlm = mle->dlm;
 371
 372         assert_spin_locked(&dlm->spinlock);
 373         assert_spin_locked(&dlm->master_lock);
 374         if (!atomic_read(&mle->mle_refs.refcount)) {
 375                 /* this may or may not crash, but who cares.
 376                  * it's a BUG. */
 377                 mlog(ML_ERROR, "bad mle: %p\n", mle);
 378                 dlm_print_one_mle(mle);
 379                 BUG();
 380         } else
 381                 kref_put(&mle->mle_refs, dlm_mle_release);
 382 }
 383
 384
 385 /* must not have any spinlocks coming in */
 386 static void dlm_put_mle(struct dlm_master_list_entry *mle)
 387 {
 388         struct dlm_ctxt *dlm;
 389         dlm = mle->dlm;
 390
 391         spin_lock(&dlm->spinlock);
 392         spin_lock(&dlm->master_lock);
 393         __dlm_put_mle(mle);
 394         spin_unlock(&dlm->master_lock);
 395         spin_unlock(&dlm->spinlock);
 396 }
 397
 398 static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
 399 {
 400         kref_get(&mle->mle_refs);
 401 }
 402
 403 static void dlm_init_mle(struct dlm_master_list_entry *mle,
 404                         enum dlm_mle_type type,
 405                         struct dlm_ctxt *dlm,
 406                         struct dlm_lock_resource *res,
 407                         const char *name,
 408                         unsigned int namelen)
 409 {
 410         assert_spin_locked(&dlm->spinlock);
 411
 412         mle->dlm = dlm;
 413         mle->type = type;
 414         INIT_LIST_HEAD(&mle->list);
 415         INIT_LIST_HEAD(&mle->hb_events);
 416         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
 417         spin_lock_init(&mle->spinlock);
 418         init_waitqueue_head(&mle->wq);
 419         atomic_set(&mle->woken, 0);
 420         kref_init(&mle->mle_refs);
 421         memset(mle->response_map, 0, sizeof(mle->response_map));
 422         mle->master = O2NM_MAX_NODES;
 423         mle->new_master = O2NM_MAX_NODES;
 424         mle->inuse = 0;
 425
 426         if (mle->type == DLM_MLE_MASTER) {
 427                 BUG_ON(!res);
 428                 mle->u.res = res;
 429         } else if (mle->type == DLM_MLE_BLOCK) {
 430                 BUG_ON(!name);
 431                 memcpy(mle->u.name.name, name, namelen);
 432                 mle->u.name.len = namelen;
 433         } else /* DLM_MLE_MIGRATION */ {
 434                 BUG_ON(!name);
 435                 memcpy(mle->u.name.name, name, namelen);
 436                 mle->u.name.len = namelen;
 437         }
 438
 439         /* copy off the node_map and register hb callbacks on our copy */
 440         memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
 441         memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
 442         clear_bit(dlm->node_num, mle->vote_map);
 443         clear_bit(dlm->node_num, mle->node_map);
 444
 445         /* attach the mle to the domain node up/down events */
 446         __dlm_mle_attach_hb_events(dlm, mle);
 447 }
 448
 449
 450 /* returns 1 if found, 0 if not */
 451 static int dlm_find_mle(struct dlm_ctxt *dlm,
 452                         struct dlm_master_list_entry **mle,
 453                         char *name, unsigned int namelen)
 454 {
 455         struct dlm_master_list_entry *tmpmle;
 456         struct list_head *iter;
 457
 458         assert_spin_locked(&dlm->master_lock);
 459
 460         list_for_each(iter, &dlm->master_list) {
 461                 tmpmle = list_entry(iter, struct dlm_master_list_entry, list);
 462                 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
 463                         continue;
 464                 dlm_get_mle(tmpmle);
 465                 *mle = tmpmle;
 466                 return 1;
 467         }
 468         return 0;
 469 }
 470
 471 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
 472 {
 473         struct dlm_master_list_entry *mle;
 474         struct list_head *iter;
 475
 476         assert_spin_locked(&dlm->spinlock);
 477
 478         list_for_each(iter, &dlm->mle_hb_events) {
 479                 mle = list_entry(iter, struct dlm_master_list_entry,
 480                                  hb_events);
 481                 if (node_up)
 482                         dlm_mle_node_up(dlm, mle, NULL, idx);
 483                 else
 484                         dlm_mle_node_down(dlm, mle, NULL, idx);
 485         }
 486 }
 487
 488 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
 489                               struct dlm_master_list_entry *mle,
 490                               struct o2nm_node *node, int idx)
 491 {
 492         spin_lock(&mle->spinlock);
 493
 494         if (!test_bit(idx, mle->node_map))
 495                 mlog(0, "node %u already removed from nodemap!\n", idx);
 496         else
 497                 clear_bit(idx, mle->node_map);
 498
 499         spin_unlock(&mle->spinlock);
 500 }
 501
 502 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
 503                             struct dlm_master_list_entry *mle,
 504                             struct o2nm_node *node, int idx)
 505 {
 506         spin_lock(&mle->spinlock);
 507
 508         if (test_bit(idx, mle->node_map))
 509                 mlog(0, "node %u already in node map!\n", idx);
 510         else
 511                 set_bit(idx, mle->node_map);
 512
 513         spin_unlock(&mle->spinlock);
 514 }
 515
 516
 517 int dlm_init_mle_cache(void)
 518 {
 519         dlm_mle_cache = kmem_cache_create("dlm_mle_cache",
 520                                           sizeof(struct dlm_master_list_entry),
 521                                           0, SLAB_HWCACHE_ALIGN,
 522                                           NULL, NULL);
 523         if (dlm_mle_cache == NULL)
 524                 return -ENOMEM;
 525         return 0;
 526 }
 527
 528 void dlm_destroy_mle_cache(void)
 529 {
 530         if (dlm_mle_cache)
 531                 kmem_cache_destroy(dlm_mle_cache);
 532 }
 533
 534 static void dlm_mle_release(struct kref *kref)
 535 {
 536         struct dlm_master_list_entry *mle;
 537         struct dlm_ctxt *dlm;
 538
 539         mlog_entry_void();
 540
 541         mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
 542         dlm = mle->dlm;
 543
 544         if (mle->type != DLM_MLE_MASTER) {
 545                 mlog(0, "calling mle_release for %.*s, type %d\n",
 546                      mle->u.name.len, mle->u.name.name, mle->type);
 547         } else {
 548                 mlog(0, "calling mle_release for %.*s, type %d\n",
 549                      mle->u.res->lockname.len,
 550                      mle->u.res->lockname.name, mle->type);
 551         }
 552         assert_spin_locked(&dlm->spinlock);
 553         assert_spin_locked(&dlm->master_lock);
 554
 555         /* remove from list if not already */
 556         if (!list_empty(&mle->list))
 557                 list_del_init(&mle->list);
 558
 559         /* detach the mle from the domain node up/down events */
 560         __dlm_mle_detach_hb_events(dlm, mle);
 561
 562         /* NOTE: kfree under spinlock here.
 563          * if this is bad, we can move this to a freelist. */
 564         kmem_cache_free(dlm_mle_cache, mle);
 565 }
 566
 567
 568 /*
 569  * LOCK RESOURCE FUNCTIONS
 570  */
 571
 572 static void dlm_set_lockres_owner(struct dlm_ctxt *dlm,
 573                                   struct dlm_lock_resource *res,
 574                                   u8 owner)
 575 {
 576         assert_spin_locked(&res->spinlock);
 577
 578         mlog_entry("%.*s, %u\n", res->lockname.len, res->lockname.name, owner);
 579
 580         if (owner == dlm->node_num)
 581                 atomic_inc(&dlm->local_resources);
 582         else if (owner == DLM_LOCK_RES_OWNER_UNKNOWN)
 583                 atomic_inc(&dlm->unknown_resources);
 584         else
 585                 atomic_inc(&dlm->remote_resources);
 586
 587         res->owner = owner;
 588 }
 589
 590 void dlm_change_lockres_owner(struct dlm_ctxt *dlm,
 591                               struct dlm_lock_resource *res, u8 owner)
 592 {
 593         assert_spin_locked(&res->spinlock);
 594
 595         if (owner == res->owner)
 596                 return;
 597
 598         if (res->owner == dlm->node_num)
 599                 atomic_dec(&dlm->local_resources);
 600         else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN)
 601                 atomic_dec(&dlm->unknown_resources);
 602         else
 603                 atomic_dec(&dlm->remote_resources);
 604
 605         dlm_set_lockres_owner(dlm, res, owner);
 606 }
 607
 608
 609 static void dlm_lockres_release(struct kref *kref)
 610 {
 611         struct dlm_lock_resource *res;
 612
 613         res = container_of(kref, struct dlm_lock_resource, refs);
 614
 615         /* This should not happen -- all lockres' have a name
 616          * associated with them at init time. */
 617         BUG_ON(!res->lockname.name);
 618
 619         mlog(0, "destroying lockres %.*s\n", res->lockname.len,
 620              res->lockname.name);
 621
 622         if (!hlist_unhashed(&res->hash_node) ||
 623             !list_empty(&res->granted) ||
 624             !list_empty(&res->converting) ||
 625             !list_empty(&res->blocked) ||
 626             !list_empty(&res->dirty) ||
 627             !list_empty(&res->recovering) ||
 628             !list_empty(&res->purge)) {
 629                 mlog(ML_ERROR,
 630                      "Going to BUG for resource %.*s."
 631                      "  We're on a list! [%c%c%c%c%c%c%c]\n",
 632                      res->lockname.len, res->lockname.name,
 633                      !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
 634                      !list_empty(&res->granted) ? 'G' : ' ',
 635                      !list_empty(&res->converting) ? 'C' : ' ',
 636                      !list_empty(&res->blocked) ? 'B' : ' ',
 637                      !list_empty(&res->dirty) ? 'D' : ' ',
 638                      !list_empty(&res->recovering) ? 'R' : ' ',
 639                      !list_empty(&res->purge) ? 'P' : ' ');
 640
 641                 dlm_print_one_lock_resource(res);
 642         }
 643
 644         /* By the time we're ready to blow this guy away, we shouldn't
 645          * be on any lists. */
 646         BUG_ON(!hlist_unhashed(&res->hash_node));
 647         BUG_ON(!list_empty(&res->granted));
 648         BUG_ON(!list_empty(&res->converting));
 649         BUG_ON(!list_empty(&res->blocked));
 650         BUG_ON(!list_empty(&res->dirty));
 651         BUG_ON(!list_empty(&res->recovering));
 652         BUG_ON(!list_empty(&res->purge));
 653
 654         kfree(res->lockname.name);
 655
 656         kfree(res);
 657 }
 658
 659 void dlm_lockres_put(struct dlm_lock_resource *res)
 660 {
 661         kref_put(&res->refs, dlm_lockres_release);
 662 }
 663
 664 static void dlm_init_lockres(struct dlm_ctxt *dlm,
 665                              struct dlm_lock_resource *res,
 666                              const char *name, unsigned int namelen)
 667 {
 668         char *qname;
 669
 670         /* If we memset here, we lose our reference to the kmalloc'd
 671          * res->lockname.name, so be sure to init every field
 672          * correctly! */
 673
 674         qname = (char *) res->lockname.name;
 675         memcpy(qname, name, namelen);
 676
 677         res->lockname.len = namelen;
 678         res->lockname.hash = dlm_lockid_hash(name, namelen);
 679
 680         init_waitqueue_head(&res->wq);
 681         spin_lock_init(&res->spinlock);
 682         INIT_HLIST_NODE(&res->hash_node);
 683         INIT_LIST_HEAD(&res->granted);
 684         INIT_LIST_HEAD(&res->converting);
 685         INIT_LIST_HEAD(&res->blocked);
 686         INIT_LIST_HEAD(&res->dirty);
 687         INIT_LIST_HEAD(&res->recovering);
 688         INIT_LIST_HEAD(&res->purge);
 689         atomic_set(&res->asts_reserved, 0);
 690         res->migration_pending = 0;
 691         res->inflight_locks = 0;
 692
 693         kref_init(&res->refs);
 694
 695         /* just for consistency */
 696         spin_lock(&res->spinlock);
 697         dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
 698         spin_unlock(&res->spinlock);
 699
 700         res->state = DLM_LOCK_RES_IN_PROGRESS;
 701
 702         res->last_used = 0;
 703
 704         memset(res->lvb, 0, DLM_LVB_LEN);
 705         memset(res->refmap, 0, sizeof(res->refmap));
 706 }
 707
 708 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
 709                                    const char *name,
 710                                    unsigned int namelen)
 711 {
 712         struct dlm_lock_resource *res;
 713
 714         res = kmalloc(sizeof(struct dlm_lock_resource), GFP_NOFS);
 715         if (!res)
 716                 return NULL;
 717
 718         res->lockname.name = kmalloc(namelen, GFP_NOFS);
 719         if (!res->lockname.name) {
 720                 kfree(res);
 721                 return NULL;
 722         }
 723
 724         dlm_init_lockres(dlm, res, name, namelen);
 725         return res;
 726 }
 727
 728 void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
 729                                    struct dlm_lock_resource *res,
 730                                    int new_lockres,
 731                                    const char *file,
 732                                    int line)
 733 {
 734         if (!new_lockres)
 735                 assert_spin_locked(&res->spinlock);
 736
 737         if (!test_bit(dlm->node_num, res->refmap)) {
 738                 BUG_ON(res->inflight_locks != 0);
 739                 dlm_lockres_set_refmap_bit(dlm->node_num, res);
 740         }
 741         res->inflight_locks++;
 742         mlog(0, "%s:%.*s: inflight++: now %u\n",
 743              dlm->name, res->lockname.len, res->lockname.name,
 744              res->inflight_locks);
 745 }
 746
 747 void __dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
 748                                    struct dlm_lock_resource *res,
 749                                    const char *file,
 750                                    int line)
 751 {
 752         assert_spin_locked(&res->spinlock);
 753
 754         BUG_ON(res->inflight_locks == 0);
 755         res->inflight_locks--;
 756         mlog(0, "%s:%.*s: inflight--: now %u\n",
 757              dlm->name, res->lockname.len, res->lockname.name,
 758              res->inflight_locks);
 759         if (res->inflight_locks == 0)
 760                 dlm_lockres_clear_refmap_bit(dlm->node_num, res);
 761         wake_up(&res->wq);
 762 }
 763
 764 /*
 765  * lookup a lock resource by name.
 766  * may already exist in the hashtable.
 767  * lockid is null terminated
 768  *
 769  * if not, allocate enough for the lockres and for
 770  * the temporary structure used in doing the mastering.
 771  *
 772  * also, do a lookup in the dlm->master_list to see
 773  * if another node has begun mastering the same lock.
 774  * if so, there should be a block entry in there
 775  * for this name, and we should *not* attempt to master
 776  * the lock here.   need to wait around for that node
 777  * to assert_master (or die).
 778  *
 779  */
 780 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
 781                                           const char *lockid,
 782                                           int namelen,
 783                                           int flags)
 784 {
 785         struct dlm_lock_resource *tmpres=NULL, *res=NULL;
 786         struct dlm_master_list_entry *mle = NULL;
 787         struct dlm_master_list_entry *alloc_mle = NULL;
 788         int blocked = 0;
 789         int ret, nodenum;
 790         struct dlm_node_iter iter;
 791         unsigned int hash;
 792         int tries = 0;
 793         int bit, wait_on_recovery = 0;
 794         int drop_inflight_if_nonlocal = 0;
 795
 796         BUG_ON(!lockid);
 797
 798         hash = dlm_lockid_hash(lockid, namelen);
 799
 800         mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
 801
 802 lookup:
 803         spin_lock(&dlm->spinlock);
 804         tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
 805         if (tmpres) {
 806                 int dropping_ref = 0;
 807
 808                 spin_lock(&tmpres->spinlock);
 809                 if (tmpres->owner == dlm->node_num) {
 810                         BUG_ON(tmpres->state & DLM_LOCK_RES_DROPPING_REF);
 811                         dlm_lockres_grab_inflight_ref(dlm, tmpres);
 812                 } else if (tmpres->state & DLM_LOCK_RES_DROPPING_REF)
 813                         dropping_ref = 1;
 814                 spin_unlock(&tmpres->spinlock);
 815                 spin_unlock(&dlm->spinlock);
 816
 817                 /* wait until done messaging the master, drop our ref to allow
 818                  * the lockres to be purged, start over. */
 819                 if (dropping_ref) {
 820                         spin_lock(&tmpres->spinlock);
 821                         __dlm_wait_on_lockres_flags(tmpres, DLM_LOCK_RES_DROPPING_REF);
 822                         spin_unlock(&tmpres->spinlock);
 823                         dlm_lockres_put(tmpres);
 824                         tmpres = NULL;
 825                         goto lookup;
 826                 }
 827
 828                 mlog(0, "found in hash!\n");
 829                 if (res)
 830                         dlm_lockres_put(res);
 831                 res = tmpres;
 832                 goto leave;
 833         }
 834
 835         if (!res) {
 836                 spin_unlock(&dlm->spinlock);
 837                 mlog(0, "allocating a new resource\n");
 838                 /* nothing found and we need to allocate one. */
 839                 alloc_mle = (struct dlm_master_list_entry *)
 840                         kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
 841                 if (!alloc_mle)
 842                         goto leave;
 843                 res = dlm_new_lockres(dlm, lockid, namelen);
 844                 if (!res)
 845                         goto leave;
 846                 goto lookup;
 847         }
 848
 849         mlog(0, "no lockres found, allocated our own: %p\n", res);
 850
 851         if (flags & LKM_LOCAL) {
 852                 /* caller knows it's safe to assume it's not mastered elsewhere
 853                  * DONE!  return right away */
 854                 spin_lock(&res->spinlock);
 855                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
 856                 __dlm_insert_lockres(dlm, res);
 857                 dlm_lockres_grab_inflight_ref(dlm, res);
 858                 spin_unlock(&res->spinlock);
 859                 spin_unlock(&dlm->spinlock);
 860                 /* lockres still marked IN_PROGRESS */
 861                 goto wake_waiters;
 862         }
 863
 864         /* check master list to see if another node has started mastering it */
 865         spin_lock(&dlm->master_lock);
 866
 867         /* if we found a block, wait for lock to be mastered by another node */
 868         blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
 869         if (blocked) {
 870                 int mig;
 871                 if (mle->type == DLM_MLE_MASTER) {
 872                         mlog(ML_ERROR, "master entry for nonexistent lock!\n");
 873                         BUG();
 874                 }
 875                 mig = (mle->type == DLM_MLE_MIGRATION);
 876                 /* if there is a migration in progress, let the migration
 877                  * finish before continuing.  we can wait for the absence
 878                  * of the MIGRATION mle: either the migrate finished or
 879                  * one of the nodes died and the mle was cleaned up.
 880                  * if there is a BLOCK here, but it already has a master
 881                  * set, we are too late.  the master does not have a ref
 882                  * for us in the refmap.  detach the mle and drop it.
 883                  * either way, go back to the top and start over. */
 884                 if (mig || mle->master != O2NM_MAX_NODES) {
 885                         BUG_ON(mig && mle->master == dlm->node_num);
 886                         /* we arrived too late.  the master does not
 887                          * have a ref for us. retry. */
 888                         mlog(0, "%s:%.*s: late on %s\n",
 889                              dlm->name, namelen, lockid,
 890                              mig ?  "MIGRATION" : "BLOCK");
 891                         spin_unlock(&dlm->master_lock);
 892                         spin_unlock(&dlm->spinlock);
 893
 894                         /* master is known, detach */
 895                         if (!mig)
 896                                 dlm_mle_detach_hb_events(dlm, mle);
 897                         dlm_put_mle(mle);
 898                         mle = NULL;
 899                         /* this is lame, but we cant wait on either
 900                          * the mle or lockres waitqueue here */
 901                         if (mig)
 902                                 msleep(100);
 903                         goto lookup;
 904                 }
 905         } else {
 906                 /* go ahead and try to master lock on this node */
 907                 mle = alloc_mle;
 908                 /* make sure this does not get freed below */
 909                 alloc_mle = NULL;
 910                 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
 911                 set_bit(dlm->node_num, mle->maybe_map);
 912                 list_add(&mle->list, &dlm->master_list);
 913
 914                 /* still holding the dlm spinlock, check the recovery map
 915                  * to see if there are any nodes that still need to be
 916                  * considered.  these will not appear in the mle nodemap
 917                  * but they might own this lockres.  wait on them. */
 918                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 919                 if (bit < O2NM_MAX_NODES) {
 920                         mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to"
 921                              "recover before lock mastery can begin\n",
 922                              dlm->name, namelen, (char *)lockid, bit);
 923                         wait_on_recovery = 1;
 924                 }
 925         }
 926
 927         /* at this point there is either a DLM_MLE_BLOCK or a
 928          * DLM_MLE_MASTER on the master list, so it's safe to add the
 929          * lockres to the hashtable.  anyone who finds the lock will
 930          * still have to wait on the IN_PROGRESS. */
 931
 932         /* finally add the lockres to its hash bucket */
 933         __dlm_insert_lockres(dlm, res);
 934         /* since this lockres is new it doesnt not require the spinlock */
 935         dlm_lockres_grab_inflight_ref_new(dlm, res);
 936
 937         /* if this node does not become the master make sure to drop
 938          * this inflight reference below */
 939         drop_inflight_if_nonlocal = 1;
 940
 941         /* get an extra ref on the mle in case this is a BLOCK
 942          * if so, the creator of the BLOCK may try to put the last
 943          * ref at this time in the assert master handler, so we
 944          * need an extra one to keep from a bad ptr deref. */
 945         dlm_get_mle_inuse(mle);
 946         spin_unlock(&dlm->master_lock);
 947         spin_unlock(&dlm->spinlock);
 948
 949 redo_request:
 950         while (wait_on_recovery) {
 951                 /* any cluster changes that occurred after dropping the
 952                  * dlm spinlock would be detectable be a change on the mle,
 953                  * so we only need to clear out the recovery map once. */
 954                 if (dlm_is_recovery_lock(lockid, namelen)) {
 955                         mlog(ML_NOTICE, "%s: recovery map is not empty, but "
 956                              "must master $RECOVERY lock now\n", dlm->name);
 957                         if (!dlm_pre_master_reco_lockres(dlm, res))
 958                                 wait_on_recovery = 0;
 959                         else {
 960                                 mlog(0, "%s: waiting 500ms for heartbeat state "
 961                                     "change\n", dlm->name);
 962                                 msleep(500);
 963                         }
 964                         continue;
 965                 }
 966
 967                 dlm_kick_recovery_thread(dlm);
 968                 msleep(1000);
 969                 dlm_wait_for_recovery(dlm);
 970
 971                 spin_lock(&dlm->spinlock);
 972                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 973                 if (bit < O2NM_MAX_NODES) {
 974                         mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to"
 975                              "recover before lock mastery can begin\n",
 976                              dlm->name, namelen, (char *)lockid, bit);
 977                         wait_on_recovery = 1;
 978                 } else
 979                         wait_on_recovery = 0;
 980                 spin_unlock(&dlm->spinlock);
 981
 982                 if (wait_on_recovery)
 983                         dlm_wait_for_node_recovery(dlm, bit, 10000);
 984         }
 985
 986         /* must wait for lock to be mastered elsewhere */
 987         if (blocked)
 988                 goto wait;
 989
 990         ret = -EINVAL;
 991         dlm_node_iter_init(mle->vote_map, &iter);
 992         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
 993                 ret = dlm_do_master_request(res, mle, nodenum);
 994                 if (ret < 0)
 995                         mlog_errno(ret);
 996                 if (mle->master != O2NM_MAX_NODES) {
 997                         /* found a master ! */
 998                         if (mle->master <= nodenum)
 999                                 break;
1000                         /* if our master request has not reached the master
1001                          * yet, keep going until it does.  this is how the
1002                          * master will know that asserts are needed back to
1003                          * the lower nodes. */
1004                         mlog(0, "%s:%.*s: requests only up to %u but master "
1005                              "is %u, keep going\n", dlm->name, namelen,
1006                              lockid, nodenum, mle->master);
1007                 }
1008         }
1009
1010 wait:
1011         /* keep going until the response map includes all nodes */
1012         ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
1013         if (ret < 0) {
1014                 wait_on_recovery = 1;
1015                 mlog(0, "%s:%.*s: node map changed, redo the "
1016                      "master request now, blocked=%d\n",
1017                      dlm->name, res->lockname.len,
1018                      res->lockname.name, blocked);
1019                 if (++tries > 20) {
1020                         mlog(ML_ERROR, "%s:%.*s: spinning on "
1021                              "dlm_wait_for_lock_mastery, blocked=%d\n",
1022                              dlm->name, res->lockname.len,
1023                              res->lockname.name, blocked);
1024                         dlm_print_one_lock_resource(res);
1025                         dlm_print_one_mle(mle);
1026                         tries = 0;
1027                 }
1028                 goto redo_request;
1029         }
1030
1031         mlog(0, "lockres mastered by %u\n", res->owner);
1032         /* make sure we never continue without this */
1033         BUG_ON(res->owner == O2NM_MAX_NODES);
1034
1035         /* master is known, detach if not already detached */
1036         dlm_mle_detach_hb_events(dlm, mle);
1037         dlm_put_mle(mle);
1038         /* put the extra ref */
1039         dlm_put_mle_inuse(mle);
1040
1041 wake_waiters:
1042         spin_lock(&res->spinlock);
1043         if (res->owner != dlm->node_num && drop_inflight_if_nonlocal)
1044                 dlm_lockres_drop_inflight_ref(dlm, res);
1045         res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1046         spin_unlock(&res->spinlock);
1047         wake_up(&res->wq);
1048
1049 leave:
1050         /* need to free the unused mle */
1051         if (alloc_mle)
1052                 kmem_cache_free(dlm_mle_cache, alloc_mle);
1053
1054         return res;
1055 }
1056
1057
1058 #define DLM_MASTERY_TIMEOUT_MS   5000
1059
1060 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
1061                                      struct dlm_lock_resource *res,
1062                                      struct dlm_master_list_entry *mle,
1063                                      int *blocked)
1064 {
1065         u8 m;
1066         int ret, bit;
1067         int map_changed, voting_done;
1068         int assert, sleep;
1069
1070 recheck:
1071         ret = 0;
1072         assert = 0;
1073
1074         /* check if another node has already become the owner */
1075         spin_lock(&res->spinlock);
1076         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1077                 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1078                      res->lockname.len, res->lockname.name, res->owner);
1079                 spin_unlock(&res->spinlock);
1080                 /* this will cause the master to re-assert across
1081                  * the whole cluster, freeing up mles */
1082                 if (res->owner != dlm->node_num) {
1083                         ret = dlm_do_master_request(res, mle, res->owner);
1084                         if (ret < 0) {
1085                                 /* give recovery a chance to run */
1086                                 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1087                                 msleep(500);
1088                                 goto recheck;
1089                         }
1090                 }
1091                 ret = 0;
1092                 goto leave;
1093         }
1094         spin_unlock(&res->spinlock);
1095
1096         spin_lock(&mle->spinlock);
1097         m = mle->master;
1098         map_changed = (memcmp(mle->vote_map, mle->node_map,
1099                               sizeof(mle->vote_map)) != 0);
1100         voting_done = (memcmp(mle->vote_map, mle->response_map,
1101                              sizeof(mle->vote_map)) == 0);
1102
1103         /* restart if we hit any errors */
1104         if (map_changed) {
1105                 int b;
1106                 mlog(0, "%s: %.*s: node map changed, restarting\n",
1107                      dlm->name, res->lockname.len, res->lockname.name);
1108                 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1109                 b = (mle->type == DLM_MLE_BLOCK);
1110                 if ((*blocked && !b) || (!*blocked && b)) {
1111                         mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1112                              dlm->name, res->lockname.len, res->lockname.name,
1113                              *blocked, b);
1114                         *blocked = b;
1115                 }
1116                 spin_unlock(&mle->spinlock);
1117                 if (ret < 0) {
1118                         mlog_errno(ret);
1119                         goto leave;
1120                 }
1121                 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1122                      "rechecking now\n", dlm->name, res->lockname.len,
1123                      res->lockname.name);
1124                 goto recheck;
1125         } else {
1126                 if (!voting_done) {
1127                         mlog(0, "map not changed and voting not done "
1128                              "for %s:%.*s\n", dlm->name, res->lockname.len,
1129                              res->lockname.name);
1130                 }
1131         }
1132
1133         if (m != O2NM_MAX_NODES) {
1134                 /* another node has done an assert!
1135                  * all done! */
1136                 sleep = 0;
1137         } else {
1138                 sleep = 1;
1139                 /* have all nodes responded? */
1140                 if (voting_done && !*blocked) {
1141                         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1142                         if (dlm->node_num <= bit) {
1143                                 /* my node number is lowest.
1144                                  * now tell other nodes that I am
1145                                  * mastering this. */
1146                                 mle->master = dlm->node_num;
1147                                 /* ref was grabbed in get_lock_resource
1148                                  * will be dropped in dlmlock_master */
1149                                 assert = 1;
1150                                 sleep = 0;
1151                         }
1152                         /* if voting is done, but we have not received
1153                          * an assert master yet, we must sleep */
1154                 }
1155         }
1156
1157         spin_unlock(&mle->spinlock);
1158
1159         /* sleep if we haven't finished voting yet */
1160         if (sleep) {
1161                 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1162
1163                 /*
1164                 if (atomic_read(&mle->mle_refs.refcount) < 2)
1165                         mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1166                         atomic_read(&mle->mle_refs.refcount),
1167                         res->lockname.len, res->lockname.name);
1168                 */
1169                 atomic_set(&mle->woken, 0);
1170                 (void)wait_event_timeout(mle->wq,
1171                                          (atomic_read(&mle->woken) == 1),
1172                                          timeo);
1173                 if (res->owner == O2NM_MAX_NODES) {
1174                         mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1175                              res->lockname.len, res->lockname.name);
1176                         goto recheck;
1177                 }
1178                 mlog(0, "done waiting, master is %u\n", res->owner);
1179                 ret = 0;
1180                 goto leave;
1181         }
1182
1183         ret = 0;   /* done */
1184         if (assert) {
1185                 m = dlm->node_num;
1186                 mlog(0, "about to master %.*s here, this=%u\n",
1187                      res->lockname.len, res->lockname.name, m);
1188                 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1189                 if (ret) {
1190                         /* This is a failure in the network path,
1191                          * not in the response to the assert_master
1192                          * (any nonzero response is a BUG on this node).
1193                          * Most likely a socket just got disconnected
1194                          * due to node death. */
1195                         mlog_errno(ret);
1196                 }
1197                 /* no longer need to restart lock mastery.
1198                  * all living nodes have been contacted. */
1199                 ret = 0;
1200         }
1201
1202         /* set the lockres owner */
1203         spin_lock(&res->spinlock);
1204         /* mastery reference obtained either during
1205          * assert_master_handler or in get_lock_resource */
1206         dlm_change_lockres_owner(dlm, res, m);
1207         spin_unlock(&res->spinlock);
1208
1209 leave:
1210         return ret;
1211 }
1212
1213 struct dlm_bitmap_diff_iter
1214 {
1215         int curnode;
1216         unsigned long *orig_bm;
1217         unsigned long *cur_bm;
1218         unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1219 };
1220
1221 enum dlm_node_state_change
1222 {
1223         NODE_DOWN = -1,
1224         NODE_NO_CHANGE = 0,
1225         NODE_UP
1226 };
1227
1228 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1229                                       unsigned long *orig_bm,
1230                                       unsigned long *cur_bm)
1231 {
1232         unsigned long p1, p2;
1233         int i;
1234
1235         iter->curnode = -1;
1236         iter->orig_bm = orig_bm;
1237         iter->cur_bm = cur_bm;
1238
1239         for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1240                 p1 = *(iter->orig_bm + i);
1241                 p2 = *(iter->cur_bm + i);
1242                 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1243         }
1244 }
1245
1246 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1247                                      enum dlm_node_state_change *state)
1248 {
1249         int bit;
1250
1251         if (iter->curnode >= O2NM_MAX_NODES)
1252                 return -ENOENT;
1253
1254         bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1255                             iter->curnode+1);
1256         if (bit >= O2NM_MAX_NODES) {
1257                 iter->curnode = O2NM_MAX_NODES;
1258                 return -ENOENT;
1259         }
1260
1261         /* if it was there in the original then this node died */
1262         if (test_bit(bit, iter->orig_bm))
1263                 *state = NODE_DOWN;
1264         else
1265                 *state = NODE_UP;
1266
1267         iter->curnode = bit;
1268         return bit;
1269 }
1270
1271
1272 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1273                                     struct dlm_lock_resource *res,
1274                                     struct dlm_master_list_entry *mle,
1275                                     int blocked)
1276 {
1277         struct dlm_bitmap_diff_iter bdi;
1278         enum dlm_node_state_change sc;
1279         int node;
1280         int ret = 0;
1281
1282         mlog(0, "something happened such that the "
1283              "master process may need to be restarted!\n");
1284
1285         assert_spin_locked(&mle->spinlock);
1286
1287         dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1288         node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1289         while (node >= 0) {
1290                 if (sc == NODE_UP) {
1291                         /* a node came up.  clear any old vote from
1292                          * the response map and set it in the vote map
1293                          * then restart the mastery. */
1294                         mlog(ML_NOTICE, "node %d up while restarting\n", node);
1295
1296                         /* redo the master request, but only for the new node */
1297                         mlog(0, "sending request to new node\n");
1298                         clear_bit(node, mle->response_map);
1299                         set_bit(node, mle->vote_map);
1300                 } else {
1301                         mlog(ML_ERROR, "node down! %d\n", node);
1302                         if (blocked) {
1303                                 int lowest = find_next_bit(mle->maybe_map,
1304                                                        O2NM_MAX_NODES, 0);
1305
1306                                 /* act like it was never there */
1307                                 clear_bit(node, mle->maybe_map);
1308
1309                                 if (node == lowest) {
1310                                         mlog(0, "expected master %u died"
1311                                             " while this node was blocked "
1312                                             "waiting on it!\n", node);
1313                                         lowest = find_next_bit(mle->maybe_map,
1314                                                         O2NM_MAX_NODES,
1315                                                         lowest+1);
1316                                         if (lowest < O2NM_MAX_NODES) {
1317                                                 mlog(0, "%s:%.*s:still "
1318                                                      "blocked. waiting on %u "
1319                                                      "now\n", dlm->name,
1320                                                      res->lockname.len,
1321                                                      res->lockname.name,
1322                                                      lowest);
1323                                         } else {
1324                                                 /* mle is an MLE_BLOCK, but
1325                                                  * there is now nothing left to
1326                                                  * block on.  we need to return
1327                                                  * all the way back out and try
1328                                                  * again with an MLE_MASTER.
1329                                                  * dlm_do_local_recovery_cleanup
1330                                                  * has already run, so the mle
1331                                                  * refcount is ok */
1332                                                 mlog(0, "%s:%.*s: no "
1333                                                      "longer blocking. try to "
1334                                                      "master this here\n",
1335                                                      dlm->name,
1336                                                      res->lockname.len,
1337                                                      res->lockname.name);
1338                                                 mle->type = DLM_MLE_MASTER;
1339                                                 mle->u.res = res;
1340                                         }
1341                                 }
1342                         }
1343
1344                         /* now blank out everything, as if we had never
1345                          * contacted anyone */
1346                         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1347                         memset(mle->response_map, 0, sizeof(mle->response_map));
1348                         /* reset the vote_map to the current node_map */
1349                         memcpy(mle->vote_map, mle->node_map,
1350                                sizeof(mle->node_map));
1351                         /* put myself into the maybe map */
1352                         if (mle->type != DLM_MLE_BLOCK)
1353                                 set_bit(dlm->node_num, mle->maybe_map);
1354                 }
1355                 ret = -EAGAIN;
1356                 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1357         }
1358         return ret;
1359 }
1360
1361
1362 /*
1363  * DLM_MASTER_REQUEST_MSG
1364  *
1365  * returns: 0 on success,
1366  *          -errno on a network error
1367  *
1368  * on error, the caller should assume the target node is "dead"
1369  *
1370  */
1371
1372 static int dlm_do_master_request(struct dlm_lock_resource *res,
1373                                  struct dlm_master_list_entry *mle, int to)
1374 {
1375         struct dlm_ctxt *dlm = mle->dlm;
1376         struct dlm_master_request request;
1377         int ret, response=0, resend;
1378
1379         memset(&request, 0, sizeof(request));
1380         request.node_idx = dlm->node_num;
1381
1382         BUG_ON(mle->type == DLM_MLE_MIGRATION);
1383
1384         if (mle->type != DLM_MLE_MASTER) {
1385                 request.namelen = mle->u.name.len;
1386                 memcpy(request.name, mle->u.name.name, request.namelen);
1387         } else {
1388                 request.namelen = mle->u.res->lockname.len;
1389                 memcpy(request.name, mle->u.res->lockname.name,
1390                         request.namelen);
1391         }
1392
1393 again:
1394         ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1395                                  sizeof(request), to, &response);
1396         if (ret < 0)  {
1397                 if (ret == -ESRCH) {
1398                         /* should never happen */
1399                         mlog(ML_ERROR, "TCP stack not ready!\n");
1400                         BUG();
1401                 } else if (ret == -EINVAL) {
1402                         mlog(ML_ERROR, "bad args passed to o2net!\n");
1403                         BUG();
1404                 } else if (ret == -ENOMEM) {
1405                         mlog(ML_ERROR, "out of memory while trying to send "
1406                              "network message!  retrying\n");
1407                         /* this is totally crude */
1408                         msleep(50);
1409                         goto again;
1410                 } else if (!dlm_is_host_down(ret)) {
1411                         /* not a network error. bad. */
1412                         mlog_errno(ret);
1413                         mlog(ML_ERROR, "unhandled error!");
1414                         BUG();
1415                 }
1416                 /* all other errors should be network errors,
1417                  * and likely indicate node death */
1418                 mlog(ML_ERROR, "link to %d went down!\n", to);
1419                 goto out;
1420         }
1421
1422         ret = 0;
1423         resend = 0;
1424         spin_lock(&mle->spinlock);
1425         switch (response) {
1426                 case DLM_MASTER_RESP_YES:
1427                         set_bit(to, mle->response_map);
1428                         mlog(0, "node %u is the master, response=YES\n", to);
1429                         mlog(0, "%s:%.*s: master node %u now knows I have a "
1430                              "reference\n", dlm->name, res->lockname.len,
1431                              res->lockname.name, to);
1432                         mle->master = to;
1433                         break;
1434                 case DLM_MASTER_RESP_NO:
1435                         mlog(0, "node %u not master, response=NO\n", to);
1436                         set_bit(to, mle->response_map);
1437                         break;
1438                 case DLM_MASTER_RESP_MAYBE:
1439                         mlog(0, "node %u not master, response=MAYBE\n", to);
1440                         set_bit(to, mle->response_map);
1441                         set_bit(to, mle->maybe_map);
1442                         break;
1443                 case DLM_MASTER_RESP_ERROR:
1444                         mlog(0, "node %u hit an error, resending\n", to);
1445                         resend = 1;
1446                         response = 0;
1447                         break;
1448                 default:
1449                         mlog(ML_ERROR, "bad response! %u\n", response);
1450                         BUG();
1451         }
1452         spin_unlock(&mle->spinlock);
1453         if (resend) {
1454                 /* this is also totally crude */
1455                 msleep(50);
1456                 goto again;
1457         }
1458
1459 out:
1460         return ret;
1461 }
1462
1463 /*
1464  * locks that can be taken here:
1465  * dlm->spinlock
1466  * res->spinlock
1467  * mle->spinlock
1468  * dlm->master_list
1469  *
1470  * if possible, TRIM THIS DOWN!!!
1471  */
1472 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1473                                void **ret_data)
1474 {
1475         u8 response = DLM_MASTER_RESP_MAYBE;
1476         struct dlm_ctxt *dlm = data;
1477         struct dlm_lock_resource *res = NULL;
1478         struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1479         struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1480         char *name;
1481         unsigned int namelen, hash;
1482         int found, ret;
1483         int set_maybe;
1484         int dispatch_assert = 0;
1485
1486         if (!dlm_grab(dlm))
1487                 return DLM_MASTER_RESP_NO;
1488
1489         if (!dlm_domain_fully_joined(dlm)) {
1490                 response = DLM_MASTER_RESP_NO;
1491                 goto send_response;
1492         }
1493
1494         name = request->name;
1495         namelen = request->namelen;
1496         hash = dlm_lockid_hash(name, namelen);
1497
1498         if (namelen > DLM_LOCKID_NAME_MAX) {
1499                 response = DLM_IVBUFLEN;
1500                 goto send_response;
1501         }
1502
1503 way_up_top:
1504         spin_lock(&dlm->spinlock);
1505         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1506         if (res) {
1507                 spin_unlock(&dlm->spinlock);
1508
1509                 /* take care of the easy cases up front */
1510                 spin_lock(&res->spinlock);
1511                 if (res->state & (DLM_LOCK_RES_RECOVERING|
1512                                   DLM_LOCK_RES_MIGRATING)) {
1513                         spin_unlock(&res->spinlock);
1514                         mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1515                              "being recovered/migrated\n");
1516                         response = DLM_MASTER_RESP_ERROR;
1517                         if (mle)
1518                                 kmem_cache_free(dlm_mle_cache, mle);
1519                         goto send_response;
1520                 }
1521
1522                 if (res->owner == dlm->node_num) {
1523                         mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1524                              dlm->name, namelen, name, request->node_idx);
1525                         dlm_lockres_set_refmap_bit(request->node_idx, res);
1526                         spin_unlock(&res->spinlock);
1527                         response = DLM_MASTER_RESP_YES;
1528                         if (mle)
1529                                 kmem_cache_free(dlm_mle_cache, mle);
1530
1531                         /* this node is the owner.
1532                          * there is some extra work that needs to
1533                          * happen now.  the requesting node has
1534                          * caused all nodes up to this one to
1535                          * create mles.  this node now needs to
1536                          * go back and clean those up. */
1537                         dispatch_assert = 1;
1538                         goto send_response;
1539                 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1540                         spin_unlock(&res->spinlock);
1541                         // mlog(0, "node %u is the master\n", res->owner);
1542                         response = DLM_MASTER_RESP_NO;
1543                         if (mle)
1544                                 kmem_cache_free(dlm_mle_cache, mle);
1545                         goto send_response;
1546                 }
1547
1548                 /* ok, there is no owner.  either this node is
1549                  * being blocked, or it is actively trying to
1550                  * master this lock. */
1551                 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1552                         mlog(ML_ERROR, "lock with no owner should be "
1553                              "in-progress!\n");
1554                         BUG();
1555                 }
1556
1557                 // mlog(0, "lockres is in progress...\n");
1558                 spin_lock(&dlm->master_lock);
1559                 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1560                 if (!found) {
1561                         mlog(ML_ERROR, "no mle found for this lock!\n");
1562                         BUG();
1563                 }
1564                 set_maybe = 1;
1565                 spin_lock(&tmpmle->spinlock);
1566                 if (tmpmle->type == DLM_MLE_BLOCK) {
1567                         // mlog(0, "this node is waiting for "
1568                         // "lockres to be mastered\n");
1569                         response = DLM_MASTER_RESP_NO;
1570                 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1571                         mlog(0, "node %u is master, but trying to migrate to "
1572                              "node %u.\n", tmpmle->master, tmpmle->new_master);
1573                         if (tmpmle->master == dlm->node_num) {
1574                                 mlog(ML_ERROR, "no owner on lockres, but this "
1575                                      "node is trying to migrate it to %u?!\n",
1576                                      tmpmle->new_master);
1577                                 BUG();
1578                         } else {
1579                                 /* the real master can respond on its own */
1580                                 response = DLM_MASTER_RESP_NO;
1581                         }
1582                 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1583                         set_maybe = 0;
1584                         if (tmpmle->master == dlm->node_num) {
1585                                 response = DLM_MASTER_RESP_YES;
1586                                 /* this node will be the owner.
1587                                  * go back and clean the mles on any
1588                                  * other nodes */
1589                                 dispatch_assert = 1;
1590                                 dlm_lockres_set_refmap_bit(request->node_idx, res);
1591                                 mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1592                                      dlm->name, namelen, name,
1593                                      request->node_idx);
1594                         } else
1595                                 response = DLM_MASTER_RESP_NO;
1596                 } else {
1597                         // mlog(0, "this node is attempting to "
1598                         // "master lockres\n");
1599                         response = DLM_MASTER_RESP_MAYBE;
1600                 }
1601                 if (set_maybe)
1602                         set_bit(request->node_idx, tmpmle->maybe_map);
1603                 spin_unlock(&tmpmle->spinlock);
1604
1605                 spin_unlock(&dlm->master_lock);
1606                 spin_unlock(&res->spinlock);
1607
1608                 /* keep the mle attached to heartbeat events */
1609                 dlm_put_mle(tmpmle);
1610                 if (mle)
1611                         kmem_cache_free(dlm_mle_cache, mle);
1612                 goto send_response;
1613         }
1614
1615         /*
1616          * lockres doesn't exist on this node
1617          * if there is an MLE_BLOCK, return NO
1618          * if there is an MLE_MASTER, return MAYBE
1619          * otherwise, add an MLE_BLOCK, return NO
1620          */
1621         spin_lock(&dlm->master_lock);
1622         found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1623         if (!found) {
1624                 /* this lockid has never been seen on this node yet */
1625                 // mlog(0, "no mle found\n");
1626                 if (!mle) {
1627                         spin_unlock(&dlm->master_lock);
1628                         spin_unlock(&dlm->spinlock);
1629
1630                         mle = (struct dlm_master_list_entry *)
1631                                 kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1632                         if (!mle) {
1633                                 response = DLM_MASTER_RESP_ERROR;
1634                                 mlog_errno(-ENOMEM);
1635                                 goto send_response;
1636                         }
1637                         goto way_up_top;
1638                 }
1639
1640                 // mlog(0, "this is second time thru, already allocated, "
1641                 // "add the block.\n");
1642                 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1643                 set_bit(request->node_idx, mle->maybe_map);
1644                 list_add(&mle->list, &dlm->master_list);
1645                 response = DLM_MASTER_RESP_NO;
1646         } else {
1647                 // mlog(0, "mle was found\n");
1648                 set_maybe = 1;
1649                 spin_lock(&tmpmle->spinlock);
1650                 if (tmpmle->master == dlm->node_num) {
1651                         mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1652                         BUG();
1653                 }
1654                 if (tmpmle->type == DLM_MLE_BLOCK)
1655                         response = DLM_MASTER_RESP_NO;
1656                 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1657                         mlog(0, "migration mle was found (%u->%u)\n",
1658                              tmpmle->master, tmpmle->new_master);
1659                         /* real master can respond on its own */
1660                         response = DLM_MASTER_RESP_NO;
1661                 } else
1662                         response = DLM_MASTER_RESP_MAYBE;
1663                 if (set_maybe)
1664                         set_bit(request->node_idx, tmpmle->maybe_map);
1665                 spin_unlock(&tmpmle->spinlock);
1666         }
1667         spin_unlock(&dlm->master_lock);
1668         spin_unlock(&dlm->spinlock);
1669
1670         if (found) {
1671                 /* keep the mle attached to heartbeat events */
1672                 dlm_put_mle(tmpmle);
1673         }
1674 send_response:
1675
1676         if (dispatch_assert) {
1677                 if (response != DLM_MASTER_RESP_YES)
1678                         mlog(ML_ERROR, "invalid response %d\n", response);
1679                 if (!res) {
1680                         mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1681                         BUG();
1682                 }
1683                 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1684                              dlm->node_num, res->lockname.len, res->lockname.name);
1685                 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1686                                                  DLM_ASSERT_MASTER_MLE_CLEANUP);
1687                 if (ret < 0) {
1688                         mlog(ML_ERROR, "failed to dispatch assert master work\n");
1689                         response = DLM_MASTER_RESP_ERROR;
1690                 }
1691         }
1692
1693         dlm_put(dlm);
1694         return response;
1695 }
1696
1697 /*
1698  * DLM_ASSERT_MASTER_MSG
1699  */
1700
1701
1702 /*
1703  * NOTE: this can be used for debugging
1704  * can periodically run all locks owned by this node
1705  * and re-assert across the cluster...
1706  */
1707 int dlm_do_assert_master(struct dlm_ctxt *dlm,
1708                          struct dlm_lock_resource *res,
1709                          void *nodemap, u32 flags)
1710 {
1711         struct dlm_assert_master assert;
1712         int to, tmpret;
1713         struct dlm_node_iter iter;
1714         int ret = 0;
1715         int reassert;
1716         const char *lockname = res->lockname.name;
1717         unsigned int namelen = res->lockname.len;
1718
1719         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1720 again:
1721         reassert = 0;
1722
1723         /* note that if this nodemap is empty, it returns 0 */
1724         dlm_node_iter_init(nodemap, &iter);
1725         while ((to = dlm_node_iter_next(&iter)) >= 0) {
1726                 int r = 0;
1727                 struct dlm_master_list_entry *mle = NULL;
1728
1729                 mlog(0, "sending assert master to %d (%.*s)\n", to,
1730                      namelen, lockname);
1731                 memset(&assert, 0, sizeof(assert));
1732                 assert.node_idx = dlm->node_num;
1733                 assert.namelen = namelen;
1734                 memcpy(assert.name, lockname, namelen);
1735                 assert.flags = cpu_to_be32(flags);
1736
1737                 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1738                                             &assert, sizeof(assert), to, &r);
1739                 if (tmpret < 0) {
1740                         mlog(0, "assert_master returned %d!\n", tmpret);
1741                         if (!dlm_is_host_down(tmpret)) {
1742                                 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1743                                 BUG();
1744                         }
1745                         /* a node died.  finish out the rest of the nodes. */
1746                         mlog(0, "link to %d went down!\n", to);
1747                         /* any nonzero status return will do */
1748                         ret = tmpret;
1749                         r = 0;
1750                 } else if (r < 0) {
1751                         /* ok, something horribly messed.  kill thyself. */
1752                         mlog(ML_ERROR,"during assert master of %.*s to %u, "
1753                              "got %d.\n", namelen, lockname, to, r);
1754                         spin_lock(&dlm->spinlock);
1755                         spin_lock(&dlm->master_lock);
1756                         if (dlm_find_mle(dlm, &mle, (char *)lockname,
1757                                          namelen)) {
1758                                 dlm_print_one_mle(mle);
1759                                 __dlm_put_mle(mle);
1760                         }
1761                         spin_unlock(&dlm->master_lock);
1762                         spin_unlock(&dlm->spinlock);
1763                         BUG();
1764                 }
1765
1766                 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1767                     !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1768                                 mlog(ML_ERROR, "%.*s: very strange, "
1769                                      "master MLE but no lockres on %u\n",
1770                                      namelen, lockname, to);
1771                 }
1772
1773                 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1774                         mlog(0, "%.*s: node %u create mles on other "
1775                              "nodes and requests a re-assert\n",
1776                              namelen, lockname, to);
1777                         reassert = 1;
1778                 }
1779                 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1780                         mlog(0, "%.*s: node %u has a reference to this "
1781                              "lockres, set the bit in the refmap\n",
1782                              namelen, lockname, to);
1783                         spin_lock(&res->spinlock);
1784                         dlm_lockres_set_refmap_bit(to, res);
1785                         spin_unlock(&res->spinlock);
1786                 }
1787         }
1788
1789         if (reassert)
1790                 goto again;
1791
1792         return ret;
1793 }
1794
1795 /*
1796  * locks that can be taken here:
1797  * dlm->spinlock
1798  * res->spinlock
1799  * mle->spinlock
1800  * dlm->master_list
1801  *
1802  * if possible, TRIM THIS DOWN!!!
1803  */
1804 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1805                               void **ret_data)
1806 {
1807         struct dlm_ctxt *dlm = data;
1808         struct dlm_master_list_entry *mle = NULL;
1809         struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1810         struct dlm_lock_resource *res = NULL;
1811         char *name;
1812         unsigned int namelen, hash;
1813         u32 flags;
1814         int master_request = 0, have_lockres_ref = 0;
1815         int ret = 0;
1816
1817         if (!dlm_grab(dlm))
1818                 return 0;
1819
1820         name = assert->name;
1821         namelen = assert->namelen;
1822         hash = dlm_lockid_hash(name, namelen);
1823         flags = be32_to_cpu(assert->flags);
1824
1825         if (namelen > DLM_LOCKID_NAME_MAX) {
1826                 mlog(ML_ERROR, "Invalid name length!");
1827                 goto done;
1828         }
1829
1830         spin_lock(&dlm->spinlock);
1831
1832         if (flags)
1833                 mlog(0, "assert_master with flags: %u\n", flags);
1834
1835         /* find the MLE */
1836         spin_lock(&dlm->master_lock);
1837         if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1838                 /* not an error, could be master just re-asserting */
1839                 mlog(0, "just got an assert_master from %u, but no "
1840                      "MLE for it! (%.*s)\n", assert->node_idx,
1841                      namelen, name);
1842         } else {
1843                 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1844                 if (bit >= O2NM_MAX_NODES) {
1845                         /* not necessarily an error, though less likely.
1846                          * could be master just re-asserting. */
1847                         mlog(0, "no bits set in the maybe_map, but %u "
1848                              "is asserting! (%.*s)\n", assert->node_idx,
1849                              namelen, name);
1850                 } else if (bit != assert->node_idx) {
1851                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1852                                 mlog(0, "master %u was found, %u should "
1853                                      "back off\n", assert->node_idx, bit);
1854                         } else {
1855                                 /* with the fix for bug 569, a higher node
1856                                  * number winning the mastery will respond
1857                                  * YES to mastery requests, but this node
1858                                  * had no way of knowing.  let it pass. */
1859                                 mlog(0, "%u is the lowest node, "
1860                                      "%u is asserting. (%.*s)  %u must "
1861                                      "have begun after %u won.\n", bit,
1862                                      assert->node_idx, namelen, name, bit,
1863                                      assert->node_idx);
1864                         }
1865                 }
1866                 if (mle->type == DLM_MLE_MIGRATION) {
1867                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1868                                 mlog(0, "%s:%.*s: got cleanup assert"
1869                                      " from %u for migration\n",
1870                                      dlm->name, namelen, name,
1871                                      assert->node_idx);
1872                         } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1873                                 mlog(0, "%s:%.*s: got unrelated assert"
1874                                      " from %u for migration, ignoring\n",
1875                                      dlm->name, namelen, name,
1876                                      assert->node_idx);
1877                                 __dlm_put_mle(mle);
1878                                 spin_unlock(&dlm->master_lock);
1879                                 spin_unlock(&dlm->spinlock);
1880                                 goto done;
1881                         }
1882                 }
1883         }
1884         spin_unlock(&dlm->master_lock);
1885
1886         /* ok everything checks out with the MLE
1887          * now check to see if there is a lockres */
1888         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1889         if (res) {
1890                 spin_lock(&res->spinlock);
1891                 if (res->state & DLM_LOCK_RES_RECOVERING)  {
1892                         mlog(ML_ERROR, "%u asserting but %.*s is "
1893                              "RECOVERING!\n", assert->node_idx, namelen, name);
1894                         goto kill;
1895                 }
1896                 if (!mle) {
1897                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1898                             res->owner != assert->node_idx) {
1899                                 mlog(ML_ERROR, "assert_master from "
1900                                           "%u, but current owner is "
1901                                           "%u! (%.*s)\n",
1902                                        assert->node_idx, res->owner,
1903                                        namelen, name);
1904                                 goto kill;
1905                         }
1906                 } else if (mle->type != DLM_MLE_MIGRATION) {
1907                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1908                                 /* owner is just re-asserting */
1909                                 if (res->owner == assert->node_idx) {
1910                                         mlog(0, "owner %u re-asserting on "
1911                                              "lock %.*s\n", assert->node_idx,
1912                                              namelen, name);
1913                                         goto ok;
1914                                 }
1915                                 mlog(ML_ERROR, "got assert_master from "
1916                                      "node %u, but %u is the owner! "
1917                                      "(%.*s)\n", assert->node_idx,
1918                                      res->owner, namelen, name);
1919                                 goto kill;
1920                         }
1921                         if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1922                                 mlog(ML_ERROR, "got assert from %u, but lock "
1923                                      "with no owner should be "
1924                                      "in-progress! (%.*s)\n",
1925                                      assert->node_idx,
1926                                      namelen, name);
1927                                 goto kill;
1928                         }
1929                 } else /* mle->type == DLM_MLE_MIGRATION */ {
1930                         /* should only be getting an assert from new master */
1931                         if (assert->node_idx != mle->new_master) {
1932                                 mlog(ML_ERROR, "got assert from %u, but "
1933                                      "new master is %u, and old master "
1934                                      "was %u (%.*s)\n",
1935                                      assert->node_idx, mle->new_master,
1936                                      mle->master, namelen, name);
1937                                 goto kill;
1938                         }
1939
1940                 }
1941 ok:
1942                 spin_unlock(&res->spinlock);
1943         }
1944         spin_unlock(&dlm->spinlock);
1945
1946         // mlog(0, "woo!  got an assert_master from node %u!\n",
1947         //           assert->node_idx);
1948         if (mle) {
1949                 int extra_ref = 0;
1950                 int nn = -1;
1951                 int rr, err = 0;
1952
1953                 spin_lock(&mle->spinlock);
1954                 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1955                         extra_ref = 1;
1956                 else {
1957                         /* MASTER mle: if any bits set in the response map
1958                          * then the calling node needs to re-assert to clear
1959                          * up nodes that this node contacted */
1960                         while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1961                                                     nn+1)) < O2NM_MAX_NODES) {
1962                                 if (nn != dlm->node_num && nn != assert->node_idx)
1963                                         master_request = 1;
1964                         }
1965                 }
1966                 mle->master = assert->node_idx;
1967                 atomic_set(&mle->woken, 1);
1968                 wake_up(&mle->wq);
1969                 spin_unlock(&mle->spinlock);
1970
1971                 if (res) {
1972                         int wake = 0;
1973                         spin_lock(&res->spinlock);
1974                         if (mle->type == DLM_MLE_MIGRATION) {
1975                                 mlog(0, "finishing off migration of lockres %.*s, "
1976                                         "from %u to %u\n",
1977                                         res->lockname.len, res->lockname.name,
1978                                         dlm->node_num, mle->new_master);
1979                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
1980                                 wake = 1;
1981                                 dlm_change_lockres_owner(dlm, res, mle->new_master);
1982                                 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1983                         } else {
1984                                 dlm_change_lockres_owner(dlm, res, mle->master);
1985                         }
1986                         spin_unlock(&res->spinlock);
1987                         have_lockres_ref = 1;
1988                         if (wake)
1989                                 wake_up(&res->wq);
1990                 }
1991
1992                 /* master is known, detach if not already detached.
1993                  * ensures that only one assert_master call will happen
1994                  * on this mle. */
1995                 spin_lock(&dlm->spinlock);
1996                 spin_lock(&dlm->master_lock);
1997
1998                 rr = atomic_read(&mle->mle_refs.refcount);
1999                 if (mle->inuse > 0) {
2000                         if (extra_ref && rr < 3)
2001                                 err = 1;
2002                         else if (!extra_ref && rr < 2)
2003                                 err = 1;
2004                 } else {
2005                         if (extra_ref && rr < 2)
2006                                 err = 1;
2007                         else if (!extra_ref && rr < 1)
2008                                 err = 1;
2009                 }
2010                 if (err) {
2011                         mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
2012                              "that will mess up this node, refs=%d, extra=%d, "
2013                              "inuse=%d\n", dlm->name, namelen, name,
2014                              assert->node_idx, rr, extra_ref, mle->inuse);
2015                         dlm_print_one_mle(mle);
2016                 }
2017                 list_del_init(&mle->list);
2018                 __dlm_mle_detach_hb_events(dlm, mle);
2019                 __dlm_put_mle(mle);
2020                 if (extra_ref) {
2021                         /* the assert master message now balances the extra
2022                          * ref given by the master / migration request message.
2023                          * if this is the last put, it will be removed
2024                          * from the list. */
2025                         __dlm_put_mle(mle);
2026                 }
2027                 spin_unlock(&dlm->master_lock);
2028                 spin_unlock(&dlm->spinlock);
2029         } else if (res) {
2030                 if (res->owner != assert->node_idx) {
2031                         mlog(0, "assert_master from %u, but current "
2032                              "owner is %u (%.*s), no mle\n", assert->node_idx,
2033                              res->owner, namelen, name);
2034                 }
2035         }
2036
2037 done:
2038         ret = 0;
2039         if (res)
2040                 dlm_lockres_put(res);
2041         dlm_put(dlm);
2042         if (master_request) {
2043                 mlog(0, "need to tell master to reassert\n");
2044                 /* positive. negative would shoot down the node. */
2045                 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2046                 if (!have_lockres_ref) {
2047                         mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2048                              "mle present here for %s:%.*s, but no lockres!\n",
2049                              assert->node_idx, dlm->name, namelen, name);
2050                 }
2051         }
2052         if (have_lockres_ref) {
2053                 /* let the master know we have a reference to the lockres */
2054                 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2055                 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2056                      dlm->name, namelen, name, assert->node_idx);
2057         }
2058         return ret;
2059
2060 kill:
2061         /* kill the caller! */
2062         mlog(ML_ERROR, "Bad message received from another node.  Dumping state "
2063              "and killing the other node now!  This node is OK and can continue.\n");
2064         __dlm_print_one_lock_resource(res);
2065         spin_unlock(&res->spinlock);
2066         spin_unlock(&dlm->spinlock);
2067         dlm_lockres_put(res);
2068         dlm_put(dlm);
2069         return -EINVAL;
2070 }
2071
2072 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2073                                struct dlm_lock_resource *res,
2074                                int ignore_higher, u8 request_from, u32 flags)
2075 {
2076         struct dlm_work_item *item;
2077         item = kzalloc(sizeof(*item), GFP_NOFS);
2078         if (!item)
2079                 return -ENOMEM;
2080
2081
2082         /* queue up work for dlm_assert_master_worker */
2083         dlm_grab(dlm);  /* get an extra ref for the work item */
2084         dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2085         item->u.am.lockres = res; /* already have a ref */
2086         /* can optionally ignore node numbers higher than this node */
2087         item->u.am.ignore_higher = ignore_higher;
2088         item->u.am.request_from = request_from;
2089         item->u.am.flags = flags;
2090
2091         if (ignore_higher)
2092                 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2093                      res->lockname.name);
2094
2095         spin_lock(&dlm->work_lock);
2096         list_add_tail(&item->list, &dlm->work_list);
2097         spin_unlock(&dlm->work_lock);
2098
2099         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2100         return 0;
2101 }
2102
2103 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2104 {
2105         struct dlm_ctxt *dlm = data;
2106         int ret = 0;
2107         struct dlm_lock_resource *res;
2108         unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2109         int ignore_higher;
2110         int bit;
2111         u8 request_from;
2112         u32 flags;
2113
2114         dlm = item->dlm;
2115         res = item->u.am.lockres;
2116         ignore_higher = item->u.am.ignore_higher;
2117         request_from = item->u.am.request_from;
2118         flags = item->u.am.flags;
2119
2120         spin_lock(&dlm->spinlock);
2121         memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2122         spin_unlock(&dlm->spinlock);
2123
2124         clear_bit(dlm->node_num, nodemap);
2125         if (ignore_higher) {
2126                 /* if is this just to clear up mles for nodes below
2127                  * this node, do not send the message to the original
2128                  * caller or any node number higher than this */
2129                 clear_bit(request_from, nodemap);
2130                 bit = dlm->node_num;
2131                 while (1) {
2132                         bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2133                                             bit+1);
2134                         if (bit >= O2NM_MAX_NODES)
2135                                 break;
2136                         clear_bit(bit, nodemap);
2137                 }
2138         }
2139
2140         /*
2141          * If we're migrating this lock to someone else, we are no
2142          * longer allowed to assert out own mastery.  OTOH, we need to
2143          * prevent migration from starting while we're still asserting
2144          * our dominance.  The reserved ast delays migration.
2145          */
2146         spin_lock(&res->spinlock);
2147         if (res->state & DLM_LOCK_RES_MIGRATING) {
2148                 mlog(0, "Someone asked us to assert mastery, but we're "
2149                      "in the middle of migration.  Skipping assert, "
2150                      "the new master will handle that.\n");
2151                 spin_unlock(&res->spinlock);
2152                 goto put;
2153         } else
2154                 __dlm_lockres_reserve_ast(res);
2155         spin_unlock(&res->spinlock);
2156
2157         /* this call now finishes out the nodemap
2158          * even if one or more nodes die */
2159         mlog(0, "worker about to master %.*s here, this=%u\n",
2160                      res->lockname.len, res->lockname.name, dlm->node_num);
2161         ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2162         if (ret < 0) {
2163                 /* no need to restart, we are done */
2164                 if (!dlm_is_host_down(ret))
2165                         mlog_errno(ret);
2166         }
2167
2168         /* Ok, we've asserted ourselves.  Let's let migration start. */
2169         dlm_lockres_release_ast(dlm, res);
2170
2171 put:
2172         dlm_lockres_put(res);
2173
2174         mlog(0, "finished with dlm_assert_master_worker\n");
2175 }
2176
2177 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2178  * We cannot wait for node recovery to complete to begin mastering this
2179  * lockres because this lockres is used to kick off recovery! ;-)
2180  * So, do a pre-check on all living nodes to see if any of those nodes
2181  * think that $RECOVERY is currently mastered by a dead node.  If so,
2182  * we wait a short time to allow that node to get notified by its own
2183  * heartbeat stack, then check again.  All $RECOVERY lock resources
2184  * mastered by dead nodes are purged when the hearbeat callback is
2185  * fired, so we can know for sure that it is safe to continue once
2186  * the node returns a live node or no node.  */
2187 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2188                                        struct dlm_lock_resource *res)
2189 {
2190         struct dlm_node_iter iter;
2191         int nodenum;
2192         int ret = 0;
2193         u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2194
2195         spin_lock(&dlm->spinlock);
2196         dlm_node_iter_init(dlm->domain_map, &iter);
2197         spin_unlock(&dlm->spinlock);
2198
2199         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2200                 /* do not send to self */
2201                 if (nodenum == dlm->node_num)
2202                         continue;
2203                 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2204                 if (ret < 0) {
2205                         mlog_errno(ret);
2206                         if (!dlm_is_host_down(ret))
2207                                 BUG();
2208                         /* host is down, so answer for that node would be
2209                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
2210                         ret = 0;
2211                 }
2212
2213                 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2214                         /* check to see if this master is in the recovery map */
2215                         spin_lock(&dlm->spinlock);
2216                         if (test_bit(master, dlm->recovery_map)) {
2217                                 mlog(ML_NOTICE, "%s: node %u has not seen "
2218                                      "node %u go down yet, and thinks the "
2219                                      "dead node is mastering the recovery "
2220                                      "lock.  must wait.\n", dlm->name,
2221                                      nodenum, master);
2222                                 ret = -EAGAIN;
2223                         }
2224                         spin_unlock(&dlm->spinlock);
2225                         mlog(0, "%s: reco lock master is %u\n", dlm->name,
2226                              master);
2227                         break;
2228                 }
2229         }
2230         return ret;
2231 }
2232
2233 /*
2234  * DLM_DEREF_LOCKRES_MSG
2235  */
2236
2237 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2238 {
2239         struct dlm_deref_lockres deref;
2240         int ret = 0, r;
2241         const char *lockname;
2242         unsigned int namelen;
2243
2244         lockname = res->lockname.name;
2245         namelen = res->lockname.len;
2246         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2247
2248         mlog(0, "%s:%.*s: sending deref to %d\n",
2249              dlm->name, namelen, lockname, res->owner);
2250         memset(&deref, 0, sizeof(deref));
2251         deref.node_idx = dlm->node_num;
2252         deref.namelen = namelen;
2253         memcpy(deref.name, lockname, namelen);
2254
2255         ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2256                                  &deref, sizeof(deref), res->owner, &r);
2257         if (ret < 0)
2258                 mlog_errno(ret);
2259         else if (r < 0) {
2260                 /* BAD.  other node says I did not have a ref. */
2261                 mlog(ML_ERROR,"while dropping ref on %s:%.*s "
2262                     "(master=%u) got %d.\n", dlm->name, namelen,
2263                     lockname, res->owner, r);
2264                 dlm_print_one_lock_resource(res);
2265                 BUG();
2266         }
2267         return ret;
2268 }
2269
2270 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2271                               void **ret_data)
2272 {
2273         struct dlm_ctxt *dlm = data;
2274         struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2275         struct dlm_lock_resource *res = NULL;
2276         char *name;
2277         unsigned int namelen;
2278         int ret = -EINVAL;
2279         u8 node;
2280         unsigned int hash;
2281
2282         if (!dlm_grab(dlm))
2283                 return 0;
2284
2285         name = deref->name;
2286         namelen = deref->namelen;
2287         node = deref->node_idx;
2288
2289         if (namelen > DLM_LOCKID_NAME_MAX) {
2290                 mlog(ML_ERROR, "Invalid name length!");
2291                 goto done;
2292         }
2293         if (deref->node_idx >= O2NM_MAX_NODES) {
2294                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2295                 goto done;
2296         }
2297
2298         hash = dlm_lockid_hash(name, namelen);
2299
2300         spin_lock(&dlm->spinlock);
2301         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2302         if (!res) {
2303                 spin_unlock(&dlm->spinlock);
2304                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2305                      dlm->name, namelen, name);
2306                 goto done;
2307         }
2308         spin_unlock(&dlm->spinlock);
2309
2310         spin_lock(&res->spinlock);
2311         BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2312         if (test_bit(node, res->refmap)) {
2313                 ret = 0;
2314                 dlm_lockres_clear_refmap_bit(node, res);
2315         } else {
2316                 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2317                      "but it is already dropped!\n", dlm->name, namelen,
2318                      name, node);
2319                 __dlm_print_one_lock_resource(res);
2320         }
2321         spin_unlock(&res->spinlock);
2322
2323         if (!ret)
2324                 dlm_lockres_calc_usage(dlm, res);
2325 done:
2326         if (res)
2327                 dlm_lockres_put(res);
2328         dlm_put(dlm);
2329         return ret;
2330 }
2331
2332
2333 /*
2334  * DLM_MIGRATE_LOCKRES
2335  */
2336
2337
2338 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2339                                struct dlm_lock_resource *res,
2340                                u8 target)
2341 {
2342         struct dlm_master_list_entry *mle = NULL;
2343         struct dlm_master_list_entry *oldmle = NULL;
2344         struct dlm_migratable_lockres *mres = NULL;
2345         int ret = -EINVAL;
2346         const char *name;
2347         unsigned int namelen;
2348         int mle_added = 0;
2349         struct list_head *queue, *iter;
2350         int i;
2351         struct dlm_lock *lock;
2352         int empty = 1, wake = 0;
2353
2354         if (!dlm_grab(dlm))
2355                 return -EINVAL;
2356
2357         name = res->lockname.name;
2358         namelen = res->lockname.len;
2359
2360         mlog(0, "migrating %.*s to %u\n", namelen, name, target);
2361
2362         /*
2363          * ensure this lockres is a proper candidate for migration
2364          */
2365         spin_lock(&res->spinlock);
2366         if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
2367                 mlog(0, "cannot migrate lockres with unknown owner!\n");
2368                 spin_unlock(&res->spinlock);
2369                 goto leave;
2370         }
2371         if (res->owner != dlm->node_num) {
2372                 mlog(0, "cannot migrate lockres this node doesn't own!\n");
2373                 spin_unlock(&res->spinlock);
2374                 goto leave;
2375         }
2376         mlog(0, "checking queues...\n");
2377         queue = &res->granted;
2378         for (i=0; i<3; i++) {
2379                 list_for_each(iter, queue) {
2380                         lock = list_entry (iter, struct dlm_lock, list);
2381                         empty = 0;
2382                         if (lock->ml.node == dlm->node_num) {
2383                                 mlog(0, "found a lock owned by this node "
2384                                      "still on the %s queue!  will not "
2385                                      "migrate this lockres\n",
2386                                      i==0 ? "granted" :
2387                                      (i==1 ? "converting" : "blocked"));
2388                                 spin_unlock(&res->spinlock);
2389                                 ret = -ENOTEMPTY;
2390                                 goto leave;
2391                         }
2392                 }
2393                 queue++;
2394         }
2395         mlog(0, "all locks on this lockres are nonlocal.  continuing\n");
2396         spin_unlock(&res->spinlock);
2397
2398         /* no work to do */
2399         if (empty) {
2400                 mlog(0, "no locks were found on this lockres! done!\n");
2401                 ret = 0;
2402                 goto leave;
2403         }
2404
2405         /*
2406          * preallocate up front
2407          * if this fails, abort
2408          */
2409
2410         ret = -ENOMEM;
2411         mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2412         if (!mres) {
2413                 mlog_errno(ret);
2414                 goto leave;
2415         }
2416
2417         mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
2418                                                                 GFP_NOFS);
2419         if (!mle) {
2420                 mlog_errno(ret);
2421                 goto leave;
2422         }
2423         ret = 0;
2424
2425         /*
2426          * find a node to migrate the lockres to
2427          */
2428
2429         mlog(0, "picking a migration node\n");
2430         spin_lock(&dlm->spinlock);
2431         /* pick a new node */
2432         if (!test_bit(target, dlm->domain_map) ||
2433             target >= O2NM_MAX_NODES) {
2434                 target = dlm_pick_migration_target(dlm, res);
2435         }
2436         mlog(0, "node %u chosen for migration\n", target);
2437
2438         if (target >= O2NM_MAX_NODES ||
2439             !test_bit(target, dlm->domain_map)) {
2440                 /* target chosen is not alive */
2441                 ret = -EINVAL;
2442         }
2443
2444         if (ret) {
2445                 spin_unlock(&dlm->spinlock);
2446                 goto fail;
2447         }
2448
2449         mlog(0, "continuing with target = %u\n", target);
2450
2451         /*
2452          * clear any existing master requests and
2453          * add the migration mle to the list
2454          */
2455         spin_lock(&dlm->master_lock);
2456         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2457                                     namelen, target, dlm->node_num);
2458         spin_unlock(&dlm->master_lock);
2459         spin_unlock(&dlm->spinlock);
2460
2461         if (ret == -EEXIST) {
2462                 mlog(0, "another process is already migrating it\n");
2463                 goto fail;
2464         }
2465         mle_added = 1;
2466
2467         /*
2468          * set the MIGRATING flag and flush asts
2469          * if we fail after this we need to re-dirty the lockres
2470          */
2471         if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2472                 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2473                      "the target went down.\n", res->lockname.len,
2474                      res->lockname.name, target);
2475                 spin_lock(&res->spinlock);
2476                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2477                 wake = 1;
2478                 spin_unlock(&res->spinlock);
2479                 ret = -EINVAL;
2480         }
2481
2482 fail:
2483         if (oldmle) {
2484                 /* master is known, detach if not already detached */
2485                 dlm_mle_detach_hb_events(dlm, oldmle);
2486                 dlm_put_mle(oldmle);
2487         }
2488
2489         if (ret < 0) {
2490                 if (mle_added) {
2491                         dlm_mle_detach_hb_events(dlm, mle);
2492                         dlm_put_mle(mle);
2493                 } else if (mle) {
2494                         kmem_cache_free(dlm_mle_cache, mle);
2495                 }
2496                 goto leave;
2497         }
2498
2499         /*
2500          * at this point, we have a migration target, an mle
2501          * in the master list, and the MIGRATING flag set on
2502          * the lockres
2503          */
2504
2505         /* now that remote nodes are spinning on the MIGRATING flag,
2506          * ensure that all assert_master work is flushed. */
2507         flush_workqueue(dlm->dlm_worker);
2508
2509         /* get an extra reference on the mle.
2510          * otherwise the assert_master from the new
2511          * master will destroy this.
2512          * also, make sure that all callers of dlm_get_mle
2513          * take both dlm->spinlock and dlm->master_lock */
2514         spin_lock(&dlm->spinlock);
2515         spin_lock(&dlm->master_lock);
2516         dlm_get_mle_inuse(mle);
2517         spin_unlock(&dlm->master_lock);
2518         spin_unlock(&dlm->spinlock);
2519
2520         /* notify new node and send all lock state */
2521         /* call send_one_lockres with migration flag.
2522          * this serves as notice to the target node that a
2523          * migration is starting. */
2524         ret = dlm_send_one_lockres(dlm, res, mres, target,
2525                                    DLM_MRES_MIGRATION);
2526
2527         if (ret < 0) {
2528                 mlog(0, "migration to node %u failed with %d\n",
2529                      target, ret);
2530                 /* migration failed, detach and clean up mle */
2531                 dlm_mle_detach_hb_events(dlm, mle);
2532                 dlm_put_mle(mle);
2533                 dlm_put_mle_inuse(mle);
2534                 spin_lock(&res->spinlock);
2535                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2536                 wake = 1;
2537                 spin_unlock(&res->spinlock);
2538                 goto leave;
2539         }
2540
2541         /* at this point, the target sends a message to all nodes,
2542          * (using dlm_do_migrate_request).  this node is skipped since
2543          * we had to put an mle in the list to begin the process.  this
2544          * node now waits for target to do an assert master.  this node
2545          * will be the last one notified, ensuring that the migration
2546          * is complete everywhere.  if the target dies while this is
2547          * going on, some nodes could potentially see the target as the
2548          * master, so it is important that my recovery finds the migration
2549          * mle and sets the master to UNKNONWN. */
2550
2551
2552         /* wait for new node to assert master */
2553         while (1) {
2554                 ret = wait_event_interruptible_timeout(mle->wq,
2555                                         (atomic_read(&mle->woken) == 1),
2556                                         msecs_to_jiffies(5000));
2557
2558                 if (ret >= 0) {
2559                         if (atomic_read(&mle->woken) == 1 ||
2560                             res->owner == target)
2561                                 break;
2562
2563                         mlog(0, "%s:%.*s: timed out during migration\n",
2564                              dlm->name, res->lockname.len, res->lockname.name);
2565                         /* avoid hang during shutdown when migrating lockres
2566                          * to a node which also goes down */
2567                         if (dlm_is_node_dead(dlm, target)) {
2568                                 mlog(0, "%s:%.*s: expected migration "
2569                                      "target %u is no longer up, restarting\n",
2570                                      dlm->name, res->lockname.len,
2571                                      res->lockname.name, target);
2572                                 ret = -EINVAL;
2573                                 /* migration failed, detach and clean up mle */
2574                                 dlm_mle_detach_hb_events(dlm, mle);
2575                                 dlm_put_mle(mle);
2576                                 dlm_put_mle_inuse(mle);
2577                                 spin_lock(&res->spinlock);
2578                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2579                                 wake = 1;
2580                                 spin_unlock(&res->spinlock);
2581                                 goto leave;
2582                         }
2583                 } else
2584                         mlog(0, "%s:%.*s: caught signal during migration\n",
2585                              dlm->name, res->lockname.len, res->lockname.name);
2586         }
2587
2588         /* all done, set the owner, clear the flag */
2589         spin_lock(&res->spinlock);
2590         dlm_set_lockres_owner(dlm, res, target);
2591         res->state &= ~DLM_LOCK_RES_MIGRATING;
2592         dlm_remove_nonlocal_locks(dlm, res);
2593         spin_unlock(&res->spinlock);
2594         wake_up(&res->wq);
2595
2596         /* master is known, detach if not already detached */
2597         dlm_mle_detach_hb_events(dlm, mle);
2598         dlm_put_mle_inuse(mle);
2599         ret = 0;
2600
2601         dlm_lockres_calc_usage(dlm, res);
2602
2603 leave:
2604         /* re-dirty the lockres if we failed */
2605         if (ret < 0)
2606                 dlm_kick_thread(dlm, res);
2607
2608         /* wake up waiters if the MIGRATING flag got set
2609          * but migration failed */
2610         if (wake)
2611                 wake_up(&res->wq);
2612
2613         /* TODO: cleanup */
2614         if (mres)
2615                 free_page((unsigned long)mres);
2616
2617         dlm_put(dlm);
2618
2619         mlog(0, "returning %d\n", ret);
2620         return ret;
2621 }
2622
2623 #define DLM_MIGRATION_RETRY_MS  100
2624
2625 /* Should be called only after beginning the domain leave process.
2626  * There should not be any remaining locks on nonlocal lock resources,
2627  * and there should be no local locks left on locally mastered resources.
2628  *
2629  * Called with the dlm spinlock held, may drop it to do migration, but
2630  * will re-acquire before exit.
2631  *
2632  * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped */
2633 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2634 {
2635         int ret;
2636         int lock_dropped = 0;
2637
2638         if (res->owner != dlm->node_num) {
2639                 if (!__dlm_lockres_unused(res)) {
2640                         mlog(ML_ERROR, "%s:%.*s: this node is not master, "
2641                              "trying to free this but locks remain\n",
2642                              dlm->name, res->lockname.len, res->lockname.name);
2643                 }
2644                 goto leave;
2645         }
2646
2647         /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2648         spin_unlock(&dlm->spinlock);
2649         lock_dropped = 1;
2650         while (1) {
2651                 ret = dlm_migrate_lockres(dlm, res, O2NM_MAX_NODES);
2652                 if (ret >= 0)
2653                         break;
2654                 if (ret == -ENOTEMPTY) {
2655                         mlog(ML_ERROR, "lockres %.*s still has local locks!\n",
2656                                 res->lockname.len, res->lockname.name);
2657                         BUG();
2658                 }
2659
2660                 mlog(0, "lockres %.*s: migrate failed, "
2661                      "retrying\n", res->lockname.len,
2662                      res->lockname.name);
2663                 msleep(DLM_MIGRATION_RETRY_MS);
2664         }
2665         spin_lock(&dlm->spinlock);
2666 leave:
2667         return lock_dropped;
2668 }
2669
2670 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2671 {
2672         int ret;
2673         spin_lock(&dlm->ast_lock);
2674         spin_lock(&lock->spinlock);
2675         ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2676         spin_unlock(&lock->spinlock);
2677         spin_unlock(&dlm->ast_lock);
2678         return ret;
2679 }
2680
2681 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2682                                      struct dlm_lock_resource *res,
2683                                      u8 mig_target)
2684 {
2685         int can_proceed;
2686         spin_lock(&res->spinlock);
2687         can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2688         spin_unlock(&res->spinlock);
2689
2690         /* target has died, so make the caller break out of the
2691          * wait_event, but caller must recheck the domain_map */
2692         spin_lock(&dlm->spinlock);
2693         if (!test_bit(mig_target, dlm->domain_map))
2694                 can_proceed = 1;
2695         spin_unlock(&dlm->spinlock);
2696         return can_proceed;
2697 }
2698
2699 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2700                                 struct dlm_lock_resource *res)
2701 {
2702         int ret;
2703         spin_lock(&res->spinlock);
2704         ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2705         spin_unlock(&res->spinlock);
2706         return ret;
2707 }
2708
2709
2710 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2711                                        struct dlm_lock_resource *res,
2712                                        u8 target)
2713 {
2714         int ret = 0;
2715
2716         mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2717                res->lockname.len, res->lockname.name, dlm->node_num,
2718                target);
2719         /* need to set MIGRATING flag on lockres.  this is done by
2720          * ensuring that all asts have been flushed for this lockres. */
2721         spin_lock(&res->spinlock);
2722         BUG_ON(res->migration_pending);
2723         res->migration_pending = 1;
2724         /* strategy is to reserve an extra ast then release
2725          * it below, letting the release do all of the work */
2726         __dlm_lockres_reserve_ast(res);
2727         spin_unlock(&res->spinlock);
2728
2729         /* now flush all the pending asts */
2730         dlm_kick_thread(dlm, res);
2731         /* before waiting on DIRTY, block processes which may
2732          * try to dirty the lockres before MIGRATING is set */
2733         spin_lock(&res->spinlock);
2734         BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2735         res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2736         spin_unlock(&res->spinlock);
2737         /* now wait on any pending asts and the DIRTY state */
2738         wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2739         dlm_lockres_release_ast(dlm, res);
2740
2741         mlog(0, "about to wait on migration_wq, dirty=%s\n",
2742                res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2743         /* if the extra ref we just put was the final one, this
2744          * will pass thru immediately.  otherwise, we need to wait
2745          * for the last ast to finish. */
2746 again:
2747         ret = wait_event_interruptible_timeout(dlm->migration_wq,
2748                    dlm_migration_can_proceed(dlm, res, target),
2749                    msecs_to_jiffies(1000));
2750         if (ret < 0) {
2751                 mlog(0, "woken again: migrating? %s, dead? %s\n",
2752                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2753                        test_bit(target, dlm->domain_map) ? "no":"yes");
2754         } else {
2755                 mlog(0, "all is well: migrating? %s, dead? %s\n",
2756                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2757                        test_bit(target, dlm->domain_map) ? "no":"yes");
2758         }
2759         if (!dlm_migration_can_proceed(dlm, res, target)) {
2760                 mlog(0, "trying again...\n");
2761                 goto again;
2762         }
2763         /* now that we are sure the MIGRATING state is there, drop
2764          * the unneded state which blocked threads trying to DIRTY */
2765         spin_lock(&res->spinlock);
2766         BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2767         BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2768         res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2769         spin_unlock(&res->spinlock);
2770
2771         /* did the target go down or die? */
2772         spin_lock(&dlm->spinlock);
2773         if (!test_bit(target, dlm->domain_map)) {
2774                 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2775                      target);
2776                 ret = -EHOSTDOWN;
2777         }
2778         spin_unlock(&dlm->spinlock);
2779
2780         /*
2781          * at this point:
2782          *
2783          *   o the DLM_LOCK_RES_MIGRATING flag is set
2784          *   o there are no pending asts on this lockres
2785          *   o all processes trying to reserve an ast on this
2786          *     lockres must wait for the MIGRATING flag to clear
2787          */
2788         return ret;
2789 }
2790
2791 /* last step in the migration process.
2792  * original master calls this to free all of the dlm_lock
2793  * structures that used to be for other nodes. */
2794 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2795                                       struct dlm_lock_resource *res)
2796 {
2797         struct list_head *iter, *iter2;
2798         struct list_head *queue = &res->granted;
2799         int i, bit;
2800         struct dlm_lock *lock;
2801
2802         assert_spin_locked(&res->spinlock);
2803
2804         BUG_ON(res->owner == dlm->node_num);
2805
2806         for (i=0; i<3; i++) {
2807                 list_for_each_safe(iter, iter2, queue) {
2808                         lock = list_entry (iter, struct dlm_lock, list);
2809                         if (lock->ml.node != dlm->node_num) {
2810                                 mlog(0, "putting lock for node %u\n",
2811                                      lock->ml.node);
2812                                 /* be extra careful */
2813                                 BUG_ON(!list_empty(&lock->ast_list));
2814                                 BUG_ON(!list_empty(&lock->bast_list));
2815                                 BUG_ON(lock->ast_pending);
2816                                 BUG_ON(lock->bast_pending);
2817                                 dlm_lockres_clear_refmap_bit(lock->ml.node, res);
2818                                 list_del_init(&lock->list);
2819                                 dlm_lock_put(lock);
2820                         }
2821                 }
2822                 queue++;
2823         }
2824         bit = 0;
2825         while (1) {
2826                 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2827                 if (bit >= O2NM_MAX_NODES)
2828                         break;
2829                 /* do not clear the local node reference, if there is a
2830                  * process holding this, let it drop the ref itself */
2831                 if (bit != dlm->node_num) {
2832                         mlog(0, "%s:%.*s: node %u had a ref to this "
2833                              "migrating lockres, clearing\n", dlm->name,
2834                              res->lockname.len, res->lockname.name, bit);
2835                         dlm_lockres_clear_refmap_bit(bit, res);
2836                 }
2837                 bit++;
2838         }
2839 }
2840
2841 /* for now this is not too intelligent.  we will
2842  * need stats to make this do the right thing.
2843  * this just finds the first lock on one of the
2844  * queues and uses that node as the target. */
2845 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2846                                     struct dlm_lock_resource *res)
2847 {
2848         int i;
2849         struct list_head *queue = &res->granted;
2850         struct list_head *iter;
2851         struct dlm_lock *lock;
2852         int nodenum;
2853
2854         assert_spin_locked(&dlm->spinlock);
2855
2856         spin_lock(&res->spinlock);
2857         for (i=0; i<3; i++) {
2858                 list_for_each(iter, queue) {
2859                         /* up to the caller to make sure this node
2860                          * is alive */
2861                         lock = list_entry (iter, struct dlm_lock, list);
2862                         if (lock->ml.node != dlm->node_num) {
2863                                 spin_unlock(&res->spinlock);
2864                                 return lock->ml.node;
2865                         }
2866                 }
2867                 queue++;
2868         }
2869         spin_unlock(&res->spinlock);
2870         mlog(0, "have not found a suitable target yet! checking domain map\n");
2871
2872         /* ok now we're getting desperate.  pick anyone alive. */
2873         nodenum = -1;
2874         while (1) {
2875                 nodenum = find_next_bit(dlm->domain_map,
2876                                         O2NM_MAX_NODES, nodenum+1);
2877                 mlog(0, "found %d in domain map\n", nodenum);
2878                 if (nodenum >= O2NM_MAX_NODES)
2879                         break;
2880                 if (nodenum != dlm->node_num) {
2881                         mlog(0, "picking %d\n", nodenum);
2882                         return nodenum;
2883                 }
2884         }
2885
2886         mlog(0, "giving up.  no master to migrate to\n");
2887         return DLM_LOCK_RES_OWNER_UNKNOWN;
2888 }
2889
2890
2891
2892 /* this is called by the new master once all lockres
2893  * data has been received */
2894 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2895                                   struct dlm_lock_resource *res,
2896                                   u8 master, u8 new_master,
2897                                   struct dlm_node_iter *iter)
2898 {
2899         struct dlm_migrate_request migrate;
2900         int ret, status = 0;
2901         int nodenum;
2902
2903         memset(&migrate, 0, sizeof(migrate));
2904         migrate.namelen = res->lockname.len;
2905         memcpy(migrate.name, res->lockname.name, migrate.namelen);
2906         migrate.new_master = new_master;
2907         migrate.master = master;
2908
2909         ret = 0;
2910
2911         /* send message to all nodes, except the master and myself */
2912         while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2913                 if (nodenum == master ||
2914                     nodenum == new_master)
2915                         continue;
2916
2917                 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2918                                          &migrate, sizeof(migrate), nodenum,
2919                                          &status);
2920                 if (ret < 0)
2921                         mlog_errno(ret);
2922                 else if (status < 0) {
2923                         mlog(0, "migrate request (node %u) returned %d!\n",
2924                              nodenum, status);
2925                         ret = status;
2926                 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
2927                         /* during the migration request we short-circuited
2928                          * the mastery of the lockres.  make sure we have
2929                          * a mastery ref for nodenum */
2930                         mlog(0, "%s:%.*s: need ref for node %u\n",
2931                              dlm->name, res->lockname.len, res->lockname.name,
2932                              nodenum);
2933                         spin_lock(&res->spinlock);
2934                         dlm_lockres_set_refmap_bit(nodenum, res);
2935                         spin_unlock(&res->spinlock);
2936                 }
2937         }
2938
2939         if (ret < 0)
2940                 mlog_errno(ret);
2941
2942         mlog(0, "returning ret=%d\n", ret);
2943         return ret;
2944 }
2945
2946
2947 /* if there is an existing mle for this lockres, we now know who the master is.
2948  * (the one who sent us *this* message) we can clear it up right away.
2949  * since the process that put the mle on the list still has a reference to it,
2950  * we can unhash it now, set the master and wake the process.  as a result,
2951  * we will have no mle in the list to start with.  now we can add an mle for
2952  * the migration and this should be the only one found for those scanning the
2953  * list.  */
2954 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
2955                                 void **ret_data)
2956 {
2957         struct dlm_ctxt *dlm = data;
2958         struct dlm_lock_resource *res = NULL;
2959         struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
2960         struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
2961         const char *name;
2962         unsigned int namelen, hash;
2963         int ret = 0;
2964
2965         if (!dlm_grab(dlm))
2966                 return -EINVAL;
2967
2968         name = migrate->name;
2969         namelen = migrate->namelen;
2970         hash = dlm_lockid_hash(name, namelen);
2971
2972         /* preallocate.. if this fails, abort */
2973         mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
2974                                                          GFP_NOFS);
2975
2976         if (!mle) {
2977                 ret = -ENOMEM;
2978                 goto leave;
2979         }
2980
2981         /* check for pre-existing lock */
2982         spin_lock(&dlm->spinlock);
2983         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
2984         spin_lock(&dlm->master_lock);
2985
2986         if (res) {
2987                 spin_lock(&res->spinlock);
2988                 if (res->state & DLM_LOCK_RES_RECOVERING) {
2989                         /* if all is working ok, this can only mean that we got
2990                         * a migrate request from a node that we now see as
2991                         * dead.  what can we do here?  drop it to the floor? */
2992                         spin_unlock(&res->spinlock);
2993                         mlog(ML_ERROR, "Got a migrate request, but the "
2994                              "lockres is marked as recovering!");
2995                         kmem_cache_free(dlm_mle_cache, mle);
2996                         ret = -EINVAL; /* need a better solution */
2997                         goto unlock;
2998                 }
2999                 res->state |= DLM_LOCK_RES_MIGRATING;
3000                 spin_unlock(&res->spinlock);
3001         }
3002
3003         /* ignore status.  only nonzero status would BUG. */
3004         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3005                                     name, namelen,
3006                                     migrate->new_master,
3007                                     migrate->master);
3008
3009 unlock:
3010         spin_unlock(&dlm->master_lock);
3011         spin_unlock(&dlm->spinlock);
3012
3013         if (oldmle) {
3014                 /* master is known, detach if not already detached */
3015                 dlm_mle_detach_hb_events(dlm, oldmle);
3016                 dlm_put_mle(oldmle);
3017         }
3018
3019         if (res)
3020                 dlm_lockres_put(res);
3021 leave:
3022         dlm_put(dlm);
3023         return ret;
3024 }
3025
3026 /* must be holding dlm->spinlock and dlm->master_lock
3027  * when adding a migration mle, we can clear any other mles
3028  * in the master list because we know with certainty that
3029  * the master is "master".  so we remove any old mle from
3030  * the list after setting it's master field, and then add
3031  * the new migration mle.  this way we can hold with the rule
3032  * of having only one mle for a given lock name at all times. */
3033 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3034                                  struct dlm_lock_resource *res,
3035                                  struct dlm_master_list_entry *mle,
3036                                  struct dlm_master_list_entry **oldmle,
3037                                  const char *name, unsigned int namelen,
3038                                  u8 new_master, u8 master)
3039 {
3040         int found;
3041         int ret = 0;
3042
3043         *oldmle = NULL;
3044
3045         mlog_entry_void();
3046
3047         assert_spin_locked(&dlm->spinlock);
3048         assert_spin_locked(&dlm->master_lock);
3049
3050         /* caller is responsible for any ref taken here on oldmle */
3051         found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3052         if (found) {
3053                 struct dlm_master_list_entry *tmp = *oldmle;
3054                 spin_lock(&tmp->spinlock);
3055                 if (tmp->type == DLM_MLE_MIGRATION) {
3056                         if (master == dlm->node_num) {
3057                                 /* ah another process raced me to it */
3058                                 mlog(0, "tried to migrate %.*s, but some "
3059                                      "process beat me to it\n",
3060                                      namelen, name);
3061                                 ret = -EEXIST;
3062                         } else {
3063                                 /* bad.  2 NODES are trying to migrate! */
3064                                 mlog(ML_ERROR, "migration error  mle: "
3065                                      "master=%u new_master=%u // request: "
3066                                      "master=%u new_master=%u // "
3067                                      "lockres=%.*s\n",
3068                                      tmp->master, tmp->new_master,
3069                                      master, new_master,
3070                                      namelen, name);
3071                                 BUG();
3072                         }
3073                 } else {
3074                         /* this is essentially what assert_master does */
3075                         tmp->master = master;
3076                         atomic_set(&tmp->woken, 1);
3077                         wake_up(&tmp->wq);
3078                         /* remove it from the list so that only one
3079                          * mle will be found */
3080                         list_del_init(&tmp->list);
3081                         /* this was obviously WRONG.  mle is uninited here.  should be tmp. */
3082                         __dlm_mle_detach_hb_events(dlm, tmp);
3083                         ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3084                         mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3085                             "telling master to get ref for cleared out mle "
3086                             "during migration\n", dlm->name, namelen, name,
3087                             master, new_master);
3088                 }
3089                 spin_unlock(&tmp->spinlock);
3090         }
3091
3092         /* now add a migration mle to the tail of the list */
3093         dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3094         mle->new_master = new_master;
3095         /* the new master will be sending an assert master for this.
3096          * at that point we will get the refmap reference */
3097         mle->master = master;
3098         /* do this for consistency with other mle types */
3099         set_bit(new_master, mle->maybe_map);
3100         list_add(&mle->list, &dlm->master_list);
3101
3102         return ret;
3103 }
3104
3105
3106 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3107 {
3108         struct list_head *iter, *iter2;
3109         struct dlm_master_list_entry *mle;
3110         struct dlm_lock_resource *res;
3111         unsigned int hash;
3112
3113         mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node);
3114 top:
3115         assert_spin_locked(&dlm->spinlock);
3116
3117         /* clean the master list */
3118         spin_lock(&dlm->master_lock);
3119         list_for_each_safe(iter, iter2, &dlm->master_list) {
3120                 mle = list_entry(iter, struct dlm_master_list_entry, list);
3121
3122                 BUG_ON(mle->type != DLM_MLE_BLOCK &&
3123                        mle->type != DLM_MLE_MASTER &&
3124                        mle->type != DLM_MLE_MIGRATION);
3125
3126                 /* MASTER mles are initiated locally.  the waiting
3127                  * process will notice the node map change
3128                  * shortly.  let that happen as normal. */
3129                 if (mle->type == DLM_MLE_MASTER)
3130                         continue;
3131
3132
3133                 /* BLOCK mles are initiated by other nodes.
3134                  * need to clean up if the dead node would have
3135                  * been the master. */
3136                 if (mle->type == DLM_MLE_BLOCK) {
3137                         int bit;
3138
3139                         spin_lock(&mle->spinlock);
3140                         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3141                         if (bit != dead_node) {
3142                                 mlog(0, "mle found, but dead node %u would "
3143                                      "not have been master\n", dead_node);
3144                                 spin_unlock(&mle->spinlock);
3145                         } else {
3146                                 /* must drop the refcount by one since the
3147                                  * assert_master will never arrive.  this
3148                                  * may result in the mle being unlinked and
3149                                  * freed, but there may still be a process
3150                                  * waiting in the dlmlock path which is fine. */
3151                                 mlog(0, "node %u was expected master\n",
3152                                      dead_node);
3153                                 atomic_set(&mle->woken, 1);
3154                                 spin_unlock(&mle->spinlock);
3155                                 wake_up(&mle->wq);
3156                                 /* do not need events any longer, so detach
3157                                  * from heartbeat */
3158                                 __dlm_mle_detach_hb_events(dlm, mle);
3159                                 __dlm_put_mle(mle);
3160                         }
3161                         continue;
3162                 }
3163
3164                 /* everything else is a MIGRATION mle */
3165
3166                 /* the rule for MIGRATION mles is that the master
3167                  * becomes UNKNOWN if *either* the original or
3168                  * the new master dies.  all UNKNOWN lockreses
3169                  * are sent to whichever node becomes the recovery
3170                  * master.  the new master is responsible for
3171                  * determining if there is still a master for
3172                  * this lockres, or if he needs to take over
3173                  * mastery.  either way, this node should expect
3174                  * another message to resolve this. */
3175                 if (mle->master != dead_node &&
3176                     mle->new_master != dead_node)
3177                         continue;
3178
3179                 /* if we have reached this point, this mle needs to
3180                  * be removed from the list and freed. */
3181
3182                 /* remove from the list early.  NOTE: unlinking
3183                  * list_head while in list_for_each_safe */
3184                 __dlm_mle_detach_hb_events(dlm, mle);
3185                 spin_lock(&mle->spinlock);
3186                 list_del_init(&mle->list);
3187                 atomic_set(&mle->woken, 1);
3188                 spin_unlock(&mle->spinlock);
3189                 wake_up(&mle->wq);
3190
3191                 mlog(0, "%s: node %u died during migration from "
3192                      "%u to %u!\n", dlm->name, dead_node,
3193                      mle->master, mle->new_master);
3194                 /* if there is a lockres associated with this
3195                  * mle, find it and set its owner to UNKNOWN */
3196                 hash = dlm_lockid_hash(mle->u.name.name, mle->u.name.len);
3197                 res = __dlm_lookup_lockres(dlm, mle->u.name.name,
3198                                            mle->u.name.len, hash);
3199                 if (res) {
3200                         /* unfortunately if we hit this rare case, our
3201                          * lock ordering is messed.  we need to drop
3202                          * the master lock so that we can take the
3203                          * lockres lock, meaning that we will have to
3204                          * restart from the head of list. */
3205                         spin_unlock(&dlm->master_lock);
3206
3207                         /* move lockres onto recovery list */
3208                         spin_lock(&res->spinlock);
3209                         dlm_set_lockres_owner(dlm, res,
3210                                         DLM_LOCK_RES_OWNER_UNKNOWN);
3211                         dlm_move_lockres_to_recovery_list(dlm, res);
3212                         spin_unlock(&res->spinlock);
3213                         dlm_lockres_put(res);
3214
3215                         /* about to get rid of mle, detach from heartbeat */
3216                         __dlm_mle_detach_hb_events(dlm, mle);
3217
3218                         /* dump the mle */
3219                         spin_lock(&dlm->master_lock);
3220                         __dlm_put_mle(mle);
3221                         spin_unlock(&dlm->master_lock);
3222
3223                         /* restart */
3224                         goto top;
3225                 }
3226
3227                 /* this may be the last reference */
3228                 __dlm_put_mle(mle);
3229         }
3230         spin_unlock(&dlm->master_lock);
3231 }
3232
3233
3234 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3235                          u8 old_master)
3236 {
3237         struct dlm_node_iter iter;
3238         int ret = 0;
3239
3240         spin_lock(&dlm->spinlock);
3241         dlm_node_iter_init(dlm->domain_map, &iter);
3242         clear_bit(old_master, iter.node_map);
3243         clear_bit(dlm->node_num, iter.node_map);
3244         spin_unlock(&dlm->spinlock);
3245
3246         /* ownership of the lockres is changing.  account for the
3247          * mastery reference here since old_master will briefly have
3248          * a reference after the migration completes */
3249         spin_lock(&res->spinlock);
3250         dlm_lockres_set_refmap_bit(old_master, res);
3251         spin_unlock(&res->spinlock);
3252
3253         mlog(0, "now time to do a migrate request to other nodes\n");
3254         ret = dlm_do_migrate_request(dlm, res, old_master,
3255                                      dlm->node_num, &iter);
3256         if (ret < 0) {
3257                 mlog_errno(ret);
3258                 goto leave;
3259         }
3260
3261         mlog(0, "doing assert master of %.*s to all except the original node\n",
3262              res->lockname.len, res->lockname.name);
3263         /* this call now finishes out the nodemap
3264          * even if one or more nodes die */
3265         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3266                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3267         if (ret < 0) {
3268                 /* no longer need to retry.  all living nodes contacted. */
3269                 mlog_errno(ret);
3270                 ret = 0;
3271         }
3272
3273         memset(iter.node_map, 0, sizeof(iter.node_map));
3274         set_bit(old_master, iter.node_map);
3275         mlog(0, "doing assert master of %.*s back to %u\n",
3276              res->lockname.len, res->lockname.name, old_master);
3277         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3278                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3279         if (ret < 0) {
3280                 mlog(0, "assert master to original master failed "
3281                      "with %d.\n", ret);
3282                 /* the only nonzero status here would be because of
3283                  * a dead original node.  we're done. */
3284                 ret = 0;
3285         }
3286
3287         /* all done, set the owner, clear the flag */
3288         spin_lock(&res->spinlock);
3289         dlm_set_lockres_owner(dlm, res, dlm->node_num);
3290         res->state &= ~DLM_LOCK_RES_MIGRATING;
3291         spin_unlock(&res->spinlock);
3292         /* re-dirty it on the new master */
3293         dlm_kick_thread(dlm, res);
3294         wake_up(&res->wq);
3295 leave:
3296         return ret;
3297 }
3298
3299 /*
3300  * LOCKRES AST REFCOUNT
3301  * this is integral to migration
3302  */
3303
3304 /* for future intent to call an ast, reserve one ahead of time.
3305  * this should be called only after waiting on the lockres
3306  * with dlm_wait_on_lockres, and while still holding the
3307  * spinlock after the call. */
3308 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3309 {
3310         assert_spin_locked(&res->spinlock);
3311         if (res->state & DLM_LOCK_RES_MIGRATING) {
3312                 __dlm_print_one_lock_resource(res);
3313         }
3314         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3315
3316         atomic_inc(&res->asts_reserved);
3317 }
3318
3319 /*
3320  * used to drop the reserved ast, either because it went unused,
3321  * or because the ast/bast was actually called.
3322  *
3323  * also, if there is a pending migration on this lockres,
3324  * and this was the last pending ast on the lockres,
3325  * atomically set the MIGRATING flag before we drop the lock.
3326  * this is how we ensure that migration can proceed with no
3327  * asts in progress.  note that it is ok if the state of the
3328  * queues is such that a lock should be granted in the future
3329  * or that a bast should be fired, because the new master will
3330  * shuffle the lists on this lockres as soon as it is migrated.
3331  */
3332 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3333                              struct dlm_lock_resource *res)
3334 {
3335         if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3336                 return;
3337
3338         if (!res->migration_pending) {
3339                 spin_unlock(&res->spinlock);
3340                 return;
3341         }
3342
3343         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3344         res->migration_pending = 0;
3345         res->state |= DLM_LOCK_RES_MIGRATING;
3346         spin_unlock(&res->spinlock);
3347         wake_up(&res->wq);
3348         wake_up(&dlm->migration_wq);
3349 }