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 {
1474         u8 response = DLM_MASTER_RESP_MAYBE;
1475         struct dlm_ctxt *dlm = data;
1476         struct dlm_lock_resource *res = NULL;
1477         struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1478         struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1479         char *name;
1480         unsigned int namelen, hash;
1481         int found, ret;
1482         int set_maybe;
1483         int dispatch_assert = 0;
1484
1485         if (!dlm_grab(dlm))
1486                 return DLM_MASTER_RESP_NO;
1487
1488         if (!dlm_domain_fully_joined(dlm)) {
1489                 response = DLM_MASTER_RESP_NO;
1490                 goto send_response;
1491         }
1492
1493         name = request->name;
1494         namelen = request->namelen;
1495         hash = dlm_lockid_hash(name, namelen);
1496
1497         if (namelen > DLM_LOCKID_NAME_MAX) {
1498                 response = DLM_IVBUFLEN;
1499                 goto send_response;
1500         }
1501
1502 way_up_top:
1503         spin_lock(&dlm->spinlock);
1504         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1505         if (res) {
1506                 spin_unlock(&dlm->spinlock);
1507
1508                 /* take care of the easy cases up front */
1509                 spin_lock(&res->spinlock);
1510                 if (res->state & DLM_LOCK_RES_RECOVERING) {
1511                         spin_unlock(&res->spinlock);
1512                         mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1513                              "being recovered\n");
1514                         response = DLM_MASTER_RESP_ERROR;
1515                         if (mle)
1516                                 kmem_cache_free(dlm_mle_cache, mle);
1517                         goto send_response;
1518                 }
1519
1520                 if (res->owner == dlm->node_num) {
1521                         mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1522                              dlm->name, namelen, name, request->node_idx);
1523                         dlm_lockres_set_refmap_bit(request->node_idx, res);
1524                         spin_unlock(&res->spinlock);
1525                         response = DLM_MASTER_RESP_YES;
1526                         if (mle)
1527                                 kmem_cache_free(dlm_mle_cache, mle);
1528
1529                         /* this node is the owner.
1530                          * there is some extra work that needs to
1531                          * happen now.  the requesting node has
1532                          * caused all nodes up to this one to
1533                          * create mles.  this node now needs to
1534                          * go back and clean those up. */
1535                         dispatch_assert = 1;
1536                         goto send_response;
1537                 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1538                         spin_unlock(&res->spinlock);
1539                         // mlog(0, "node %u is the master\n", res->owner);
1540                         response = DLM_MASTER_RESP_NO;
1541                         if (mle)
1542                                 kmem_cache_free(dlm_mle_cache, mle);
1543                         goto send_response;
1544                 }
1545
1546                 /* ok, there is no owner.  either this node is
1547                  * being blocked, or it is actively trying to
1548                  * master this lock. */
1549                 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1550                         mlog(ML_ERROR, "lock with no owner should be "
1551                              "in-progress!\n");
1552                         BUG();
1553                 }
1554
1555                 // mlog(0, "lockres is in progress...\n");
1556                 spin_lock(&dlm->master_lock);
1557                 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1558                 if (!found) {
1559                         mlog(ML_ERROR, "no mle found for this lock!\n");
1560                         BUG();
1561                 }
1562                 set_maybe = 1;
1563                 spin_lock(&tmpmle->spinlock);
1564                 if (tmpmle->type == DLM_MLE_BLOCK) {
1565                         // mlog(0, "this node is waiting for "
1566                         // "lockres to be mastered\n");
1567                         response = DLM_MASTER_RESP_NO;
1568                 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1569                         mlog(0, "node %u is master, but trying to migrate to "
1570                              "node %u.\n", tmpmle->master, tmpmle->new_master);
1571                         if (tmpmle->master == dlm->node_num) {
1572                                 mlog(ML_ERROR, "no owner on lockres, but this "
1573                                      "node is trying to migrate it to %u?!\n",
1574                                      tmpmle->new_master);
1575                                 BUG();
1576                         } else {
1577                                 /* the real master can respond on its own */
1578                                 response = DLM_MASTER_RESP_NO;
1579                         }
1580                 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1581                         set_maybe = 0;
1582                         if (tmpmle->master == dlm->node_num) {
1583                                 response = DLM_MASTER_RESP_YES;
1584                                 /* this node will be the owner.
1585                                  * go back and clean the mles on any
1586                                  * other nodes */
1587                                 dispatch_assert = 1;
1588                                 dlm_lockres_set_refmap_bit(request->node_idx, res);
1589                                 mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1590                                      dlm->name, namelen, name,
1591                                      request->node_idx);
1592                         } else
1593                                 response = DLM_MASTER_RESP_NO;
1594                 } else {
1595                         // mlog(0, "this node is attempting to "
1596                         // "master lockres\n");
1597                         response = DLM_MASTER_RESP_MAYBE;
1598                 }
1599                 if (set_maybe)
1600                         set_bit(request->node_idx, tmpmle->maybe_map);
1601                 spin_unlock(&tmpmle->spinlock);
1602
1603                 spin_unlock(&dlm->master_lock);
1604                 spin_unlock(&res->spinlock);
1605
1606                 /* keep the mle attached to heartbeat events */
1607                 dlm_put_mle(tmpmle);
1608                 if (mle)
1609                         kmem_cache_free(dlm_mle_cache, mle);
1610                 goto send_response;
1611         }
1612
1613         /*
1614          * lockres doesn't exist on this node
1615          * if there is an MLE_BLOCK, return NO
1616          * if there is an MLE_MASTER, return MAYBE
1617          * otherwise, add an MLE_BLOCK, return NO
1618          */
1619         spin_lock(&dlm->master_lock);
1620         found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1621         if (!found) {
1622                 /* this lockid has never been seen on this node yet */
1623                 // mlog(0, "no mle found\n");
1624                 if (!mle) {
1625                         spin_unlock(&dlm->master_lock);
1626                         spin_unlock(&dlm->spinlock);
1627
1628                         mle = (struct dlm_master_list_entry *)
1629                                 kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1630                         if (!mle) {
1631                                 response = DLM_MASTER_RESP_ERROR;
1632                                 mlog_errno(-ENOMEM);
1633                                 goto send_response;
1634                         }
1635                         goto way_up_top;
1636                 }
1637
1638                 // mlog(0, "this is second time thru, already allocated, "
1639                 // "add the block.\n");
1640                 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1641                 set_bit(request->node_idx, mle->maybe_map);
1642                 list_add(&mle->list, &dlm->master_list);
1643                 response = DLM_MASTER_RESP_NO;
1644         } else {
1645                 // mlog(0, "mle was found\n");
1646                 set_maybe = 1;
1647                 spin_lock(&tmpmle->spinlock);
1648                 if (tmpmle->master == dlm->node_num) {
1649                         mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1650                         BUG();
1651                 }
1652                 if (tmpmle->type == DLM_MLE_BLOCK)
1653                         response = DLM_MASTER_RESP_NO;
1654                 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1655                         mlog(0, "migration mle was found (%u->%u)\n",
1656                              tmpmle->master, tmpmle->new_master);
1657                         /* real master can respond on its own */
1658                         response = DLM_MASTER_RESP_NO;
1659                 } else
1660                         response = DLM_MASTER_RESP_MAYBE;
1661                 if (set_maybe)
1662                         set_bit(request->node_idx, tmpmle->maybe_map);
1663                 spin_unlock(&tmpmle->spinlock);
1664         }
1665         spin_unlock(&dlm->master_lock);
1666         spin_unlock(&dlm->spinlock);
1667
1668         if (found) {
1669                 /* keep the mle attached to heartbeat events */
1670                 dlm_put_mle(tmpmle);
1671         }
1672 send_response:
1673
1674         if (dispatch_assert) {
1675                 if (response != DLM_MASTER_RESP_YES)
1676                         mlog(ML_ERROR, "invalid response %d\n", response);
1677                 if (!res) {
1678                         mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1679                         BUG();
1680                 }
1681                 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1682                              dlm->node_num, res->lockname.len, res->lockname.name);
1683                 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1684                                                  DLM_ASSERT_MASTER_MLE_CLEANUP);
1685                 if (ret < 0) {
1686                         mlog(ML_ERROR, "failed to dispatch assert master work\n");
1687                         response = DLM_MASTER_RESP_ERROR;
1688                 }
1689         }
1690
1691         dlm_put(dlm);
1692         return response;
1693 }
1694
1695 /*
1696  * DLM_ASSERT_MASTER_MSG
1697  */
1698
1699
1700 /*
1701  * NOTE: this can be used for debugging
1702  * can periodically run all locks owned by this node
1703  * and re-assert across the cluster...
1704  */
1705 int dlm_do_assert_master(struct dlm_ctxt *dlm,
1706                          struct dlm_lock_resource *res,
1707                          void *nodemap, u32 flags)
1708 {
1709         struct dlm_assert_master assert;
1710         int to, tmpret;
1711         struct dlm_node_iter iter;
1712         int ret = 0;
1713         int reassert;
1714         const char *lockname = res->lockname.name;
1715         unsigned int namelen = res->lockname.len;
1716
1717         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1718 again:
1719         reassert = 0;
1720
1721         /* note that if this nodemap is empty, it returns 0 */
1722         dlm_node_iter_init(nodemap, &iter);
1723         while ((to = dlm_node_iter_next(&iter)) >= 0) {
1724                 int r = 0;
1725                 struct dlm_master_list_entry *mle = NULL;
1726
1727                 mlog(0, "sending assert master to %d (%.*s)\n", to,
1728                      namelen, lockname);
1729                 memset(&assert, 0, sizeof(assert));
1730                 assert.node_idx = dlm->node_num;
1731                 assert.namelen = namelen;
1732                 memcpy(assert.name, lockname, namelen);
1733                 assert.flags = cpu_to_be32(flags);
1734
1735                 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1736                                             &assert, sizeof(assert), to, &r);
1737                 if (tmpret < 0) {
1738                         mlog(0, "assert_master returned %d!\n", tmpret);
1739                         if (!dlm_is_host_down(tmpret)) {
1740                                 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1741                                 BUG();
1742                         }
1743                         /* a node died.  finish out the rest of the nodes. */
1744                         mlog(0, "link to %d went down!\n", to);
1745                         /* any nonzero status return will do */
1746                         ret = tmpret;
1747                         r = 0;
1748                 } else if (r < 0) {
1749                         /* ok, something horribly messed.  kill thyself. */
1750                         mlog(ML_ERROR,"during assert master of %.*s to %u, "
1751                              "got %d.\n", namelen, lockname, to, r);
1752                         spin_lock(&dlm->spinlock);
1753                         spin_lock(&dlm->master_lock);
1754                         if (dlm_find_mle(dlm, &mle, (char *)lockname,
1755                                          namelen)) {
1756                                 dlm_print_one_mle(mle);
1757                                 __dlm_put_mle(mle);
1758                         }
1759                         spin_unlock(&dlm->master_lock);
1760                         spin_unlock(&dlm->spinlock);
1761                         BUG();
1762                 }
1763
1764                 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1765                     !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1766                                 mlog(ML_ERROR, "%.*s: very strange, "
1767                                      "master MLE but no lockres on %u\n",
1768                                      namelen, lockname, to);
1769                 }
1770
1771                 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1772                         mlog(0, "%.*s: node %u create mles on other "
1773                              "nodes and requests a re-assert\n",
1774                              namelen, lockname, to);
1775                         reassert = 1;
1776                 }
1777                 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1778                         mlog(0, "%.*s: node %u has a reference to this "
1779                              "lockres, set the bit in the refmap\n",
1780                              namelen, lockname, to);
1781                         spin_lock(&res->spinlock);
1782                         dlm_lockres_set_refmap_bit(to, res);
1783                         spin_unlock(&res->spinlock);
1784                 }
1785         }
1786
1787         if (reassert)
1788                 goto again;
1789
1790         return ret;
1791 }
1792
1793 /*
1794  * locks that can be taken here:
1795  * dlm->spinlock
1796  * res->spinlock
1797  * mle->spinlock
1798  * dlm->master_list
1799  *
1800  * if possible, TRIM THIS DOWN!!!
1801  */
1802 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data)
1803 {
1804         struct dlm_ctxt *dlm = data;
1805         struct dlm_master_list_entry *mle = NULL;
1806         struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1807         struct dlm_lock_resource *res = NULL;
1808         char *name;
1809         unsigned int namelen, hash;
1810         u32 flags;
1811         int master_request = 0, have_lockres_ref = 0;
1812         int ret = 0;
1813
1814         if (!dlm_grab(dlm))
1815                 return 0;
1816
1817         name = assert->name;
1818         namelen = assert->namelen;
1819         hash = dlm_lockid_hash(name, namelen);
1820         flags = be32_to_cpu(assert->flags);
1821
1822         if (namelen > DLM_LOCKID_NAME_MAX) {
1823                 mlog(ML_ERROR, "Invalid name length!");
1824                 goto done;
1825         }
1826
1827         spin_lock(&dlm->spinlock);
1828
1829         if (flags)
1830                 mlog(0, "assert_master with flags: %u\n", flags);
1831
1832         /* find the MLE */
1833         spin_lock(&dlm->master_lock);
1834         if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1835                 /* not an error, could be master just re-asserting */
1836                 mlog(0, "just got an assert_master from %u, but no "
1837                      "MLE for it! (%.*s)\n", assert->node_idx,
1838                      namelen, name);
1839         } else {
1840                 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1841                 if (bit >= O2NM_MAX_NODES) {
1842                         /* not necessarily an error, though less likely.
1843                          * could be master just re-asserting. */
1844                         mlog(0, "no bits set in the maybe_map, but %u "
1845                              "is asserting! (%.*s)\n", assert->node_idx,
1846                              namelen, name);
1847                 } else if (bit != assert->node_idx) {
1848                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1849                                 mlog(0, "master %u was found, %u should "
1850                                      "back off\n", assert->node_idx, bit);
1851                         } else {
1852                                 /* with the fix for bug 569, a higher node
1853                                  * number winning the mastery will respond
1854                                  * YES to mastery requests, but this node
1855                                  * had no way of knowing.  let it pass. */
1856                                 mlog(0, "%u is the lowest node, "
1857                                      "%u is asserting. (%.*s)  %u must "
1858                                      "have begun after %u won.\n", bit,
1859                                      assert->node_idx, namelen, name, bit,
1860                                      assert->node_idx);
1861                         }
1862                 }
1863                 if (mle->type == DLM_MLE_MIGRATION) {
1864                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1865                                 mlog(0, "%s:%.*s: got cleanup assert"
1866                                      " from %u for migration\n",
1867                                      dlm->name, namelen, name,
1868                                      assert->node_idx);
1869                         } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1870                                 mlog(0, "%s:%.*s: got unrelated assert"
1871                                      " from %u for migration, ignoring\n",
1872                                      dlm->name, namelen, name,
1873                                      assert->node_idx);
1874                                 __dlm_put_mle(mle);
1875                                 spin_unlock(&dlm->master_lock);
1876                                 spin_unlock(&dlm->spinlock);
1877                                 goto done;
1878                         }
1879                 }
1880         }
1881         spin_unlock(&dlm->master_lock);
1882
1883         /* ok everything checks out with the MLE
1884          * now check to see if there is a lockres */
1885         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1886         if (res) {
1887                 spin_lock(&res->spinlock);
1888                 if (res->state & DLM_LOCK_RES_RECOVERING)  {
1889                         mlog(ML_ERROR, "%u asserting but %.*s is "
1890                              "RECOVERING!\n", assert->node_idx, namelen, name);
1891                         goto kill;
1892                 }
1893                 if (!mle) {
1894                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1895                             res->owner != assert->node_idx) {
1896                                 mlog(ML_ERROR, "assert_master from "
1897                                           "%u, but current owner is "
1898                                           "%u! (%.*s)\n",
1899                                        assert->node_idx, res->owner,
1900                                        namelen, name);
1901                                 goto kill;
1902                         }
1903                 } else if (mle->type != DLM_MLE_MIGRATION) {
1904                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1905                                 /* owner is just re-asserting */
1906                                 if (res->owner == assert->node_idx) {
1907                                         mlog(0, "owner %u re-asserting on "
1908                                              "lock %.*s\n", assert->node_idx,
1909                                              namelen, name);
1910                                         goto ok;
1911                                 }
1912                                 mlog(ML_ERROR, "got assert_master from "
1913                                      "node %u, but %u is the owner! "
1914                                      "(%.*s)\n", assert->node_idx,
1915                                      res->owner, namelen, name);
1916                                 goto kill;
1917                         }
1918                         if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1919                                 mlog(ML_ERROR, "got assert from %u, but lock "
1920                                      "with no owner should be "
1921                                      "in-progress! (%.*s)\n",
1922                                      assert->node_idx,
1923                                      namelen, name);
1924                                 goto kill;
1925                         }
1926                 } else /* mle->type == DLM_MLE_MIGRATION */ {
1927                         /* should only be getting an assert from new master */
1928                         if (assert->node_idx != mle->new_master) {
1929                                 mlog(ML_ERROR, "got assert from %u, but "
1930                                      "new master is %u, and old master "
1931                                      "was %u (%.*s)\n",
1932                                      assert->node_idx, mle->new_master,
1933                                      mle->master, namelen, name);
1934                                 goto kill;
1935                         }
1936
1937                 }
1938 ok:
1939                 spin_unlock(&res->spinlock);
1940         }
1941         spin_unlock(&dlm->spinlock);
1942
1943         // mlog(0, "woo!  got an assert_master from node %u!\n",
1944         //           assert->node_idx);
1945         if (mle) {
1946                 int extra_ref = 0;
1947                 int nn = -1;
1948                 int rr, err = 0;
1949
1950                 spin_lock(&mle->spinlock);
1951                 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1952                         extra_ref = 1;
1953                 else {
1954                         /* MASTER mle: if any bits set in the response map
1955                          * then the calling node needs to re-assert to clear
1956                          * up nodes that this node contacted */
1957                         while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1958                                                     nn+1)) < O2NM_MAX_NODES) {
1959                                 if (nn != dlm->node_num && nn != assert->node_idx)
1960                                         master_request = 1;
1961                         }
1962                 }
1963                 mle->master = assert->node_idx;
1964                 atomic_set(&mle->woken, 1);
1965                 wake_up(&mle->wq);
1966                 spin_unlock(&mle->spinlock);
1967
1968                 if (res) {
1969                         spin_lock(&res->spinlock);
1970                         if (mle->type == DLM_MLE_MIGRATION) {
1971                                 mlog(0, "finishing off migration of lockres %.*s, "
1972                                         "from %u to %u\n",
1973                                         res->lockname.len, res->lockname.name,
1974                                         dlm->node_num, mle->new_master);
1975                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
1976                                 dlm_change_lockres_owner(dlm, res, mle->new_master);
1977                                 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1978                         } else {
1979                                 dlm_change_lockres_owner(dlm, res, mle->master);
1980                         }
1981                         spin_unlock(&res->spinlock);
1982                         have_lockres_ref = 1;
1983                 }
1984
1985                 /* master is known, detach if not already detached.
1986                  * ensures that only one assert_master call will happen
1987                  * on this mle. */
1988                 spin_lock(&dlm->spinlock);
1989                 spin_lock(&dlm->master_lock);
1990
1991                 rr = atomic_read(&mle->mle_refs.refcount);
1992                 if (mle->inuse > 0) {
1993                         if (extra_ref && rr < 3)
1994                                 err = 1;
1995                         else if (!extra_ref && rr < 2)
1996                                 err = 1;
1997                 } else {
1998                         if (extra_ref && rr < 2)
1999                                 err = 1;
2000                         else if (!extra_ref && rr < 1)
2001                                 err = 1;
2002                 }
2003                 if (err) {
2004                         mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
2005                              "that will mess up this node, refs=%d, extra=%d, "
2006                              "inuse=%d\n", dlm->name, namelen, name,
2007                              assert->node_idx, rr, extra_ref, mle->inuse);
2008                         dlm_print_one_mle(mle);
2009                 }
2010                 list_del_init(&mle->list);
2011                 __dlm_mle_detach_hb_events(dlm, mle);
2012                 __dlm_put_mle(mle);
2013                 if (extra_ref) {
2014                         /* the assert master message now balances the extra
2015                          * ref given by the master / migration request message.
2016                          * if this is the last put, it will be removed
2017                          * from the list. */
2018                         __dlm_put_mle(mle);
2019                 }
2020                 spin_unlock(&dlm->master_lock);
2021                 spin_unlock(&dlm->spinlock);
2022         } else if (res) {
2023                 if (res->owner != assert->node_idx) {
2024                         mlog(0, "assert_master from %u, but current "
2025                              "owner is %u (%.*s), no mle\n", assert->node_idx,
2026                              res->owner, namelen, name);
2027                 }
2028         }
2029
2030 done:
2031         ret = 0;
2032         if (res)
2033                 dlm_lockres_put(res);
2034         dlm_put(dlm);
2035         if (master_request) {
2036                 mlog(0, "need to tell master to reassert\n");
2037                 /* positive. negative would shoot down the node. */
2038                 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2039                 if (!have_lockres_ref) {
2040                         mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2041                              "mle present here for %s:%.*s, but no lockres!\n",
2042                              assert->node_idx, dlm->name, namelen, name);
2043                 }
2044         }
2045         if (have_lockres_ref) {
2046                 /* let the master know we have a reference to the lockres */
2047                 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2048                 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2049                      dlm->name, namelen, name, assert->node_idx);
2050         }
2051         return ret;
2052
2053 kill:
2054         /* kill the caller! */
2055         mlog(ML_ERROR, "Bad message received from another node.  Dumping state "
2056              "and killing the other node now!  This node is OK and can continue.\n");
2057         __dlm_print_one_lock_resource(res);
2058         spin_unlock(&res->spinlock);
2059         spin_unlock(&dlm->spinlock);
2060         dlm_lockres_put(res);
2061         dlm_put(dlm);
2062         return -EINVAL;
2063 }
2064
2065 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2066                                struct dlm_lock_resource *res,
2067                                int ignore_higher, u8 request_from, u32 flags)
2068 {
2069         struct dlm_work_item *item;
2070         item = kzalloc(sizeof(*item), GFP_NOFS);
2071         if (!item)
2072                 return -ENOMEM;
2073
2074
2075         /* queue up work for dlm_assert_master_worker */
2076         dlm_grab(dlm);  /* get an extra ref for the work item */
2077         dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2078         item->u.am.lockres = res; /* already have a ref */
2079         /* can optionally ignore node numbers higher than this node */
2080         item->u.am.ignore_higher = ignore_higher;
2081         item->u.am.request_from = request_from;
2082         item->u.am.flags = flags;
2083
2084         if (ignore_higher)
2085                 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2086                      res->lockname.name);
2087
2088         spin_lock(&dlm->work_lock);
2089         list_add_tail(&item->list, &dlm->work_list);
2090         spin_unlock(&dlm->work_lock);
2091
2092         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2093         return 0;
2094 }
2095
2096 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2097 {
2098         struct dlm_ctxt *dlm = data;
2099         int ret = 0;
2100         struct dlm_lock_resource *res;
2101         unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2102         int ignore_higher;
2103         int bit;
2104         u8 request_from;
2105         u32 flags;
2106
2107         dlm = item->dlm;
2108         res = item->u.am.lockres;
2109         ignore_higher = item->u.am.ignore_higher;
2110         request_from = item->u.am.request_from;
2111         flags = item->u.am.flags;
2112
2113         spin_lock(&dlm->spinlock);
2114         memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2115         spin_unlock(&dlm->spinlock);
2116
2117         clear_bit(dlm->node_num, nodemap);
2118         if (ignore_higher) {
2119                 /* if is this just to clear up mles for nodes below
2120                  * this node, do not send the message to the original
2121                  * caller or any node number higher than this */
2122                 clear_bit(request_from, nodemap);
2123                 bit = dlm->node_num;
2124                 while (1) {
2125                         bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2126                                             bit+1);
2127                         if (bit >= O2NM_MAX_NODES)
2128                                 break;
2129                         clear_bit(bit, nodemap);
2130                 }
2131         }
2132
2133         /*
2134          * If we're migrating this lock to someone else, we are no
2135          * longer allowed to assert out own mastery.  OTOH, we need to
2136          * prevent migration from starting while we're still asserting
2137          * our dominance.  The reserved ast delays migration.
2138          */
2139         spin_lock(&res->spinlock);
2140         if (res->state & DLM_LOCK_RES_MIGRATING) {
2141                 mlog(0, "Someone asked us to assert mastery, but we're "
2142                      "in the middle of migration.  Skipping assert, "
2143                      "the new master will handle that.\n");
2144                 spin_unlock(&res->spinlock);
2145                 goto put;
2146         } else
2147                 __dlm_lockres_reserve_ast(res);
2148         spin_unlock(&res->spinlock);
2149
2150         /* this call now finishes out the nodemap
2151          * even if one or more nodes die */
2152         mlog(0, "worker about to master %.*s here, this=%u\n",
2153                      res->lockname.len, res->lockname.name, dlm->node_num);
2154         ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2155         if (ret < 0) {
2156                 /* no need to restart, we are done */
2157                 if (!dlm_is_host_down(ret))
2158                         mlog_errno(ret);
2159         }
2160
2161         /* Ok, we've asserted ourselves.  Let's let migration start. */
2162         dlm_lockres_release_ast(dlm, res);
2163
2164 put:
2165         dlm_lockres_put(res);
2166
2167         mlog(0, "finished with dlm_assert_master_worker\n");
2168 }
2169
2170 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2171  * We cannot wait for node recovery to complete to begin mastering this
2172  * lockres because this lockres is used to kick off recovery! ;-)
2173  * So, do a pre-check on all living nodes to see if any of those nodes
2174  * think that $RECOVERY is currently mastered by a dead node.  If so,
2175  * we wait a short time to allow that node to get notified by its own
2176  * heartbeat stack, then check again.  All $RECOVERY lock resources
2177  * mastered by dead nodes are purged when the hearbeat callback is
2178  * fired, so we can know for sure that it is safe to continue once
2179  * the node returns a live node or no node.  */
2180 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2181                                        struct dlm_lock_resource *res)
2182 {
2183         struct dlm_node_iter iter;
2184         int nodenum;
2185         int ret = 0;
2186         u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2187
2188         spin_lock(&dlm->spinlock);
2189         dlm_node_iter_init(dlm->domain_map, &iter);
2190         spin_unlock(&dlm->spinlock);
2191
2192         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2193                 /* do not send to self */
2194                 if (nodenum == dlm->node_num)
2195                         continue;
2196                 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2197                 if (ret < 0) {
2198                         mlog_errno(ret);
2199                         if (!dlm_is_host_down(ret))
2200                                 BUG();
2201                         /* host is down, so answer for that node would be
2202                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
2203                         ret = 0;
2204                 }
2205
2206                 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2207                         /* check to see if this master is in the recovery map */
2208                         spin_lock(&dlm->spinlock);
2209                         if (test_bit(master, dlm->recovery_map)) {
2210                                 mlog(ML_NOTICE, "%s: node %u has not seen "
2211                                      "node %u go down yet, and thinks the "
2212                                      "dead node is mastering the recovery "
2213                                      "lock.  must wait.\n", dlm->name,
2214                                      nodenum, master);
2215                                 ret = -EAGAIN;
2216                         }
2217                         spin_unlock(&dlm->spinlock);
2218                         mlog(0, "%s: reco lock master is %u\n", dlm->name,
2219                              master);
2220                         break;
2221                 }
2222         }
2223         return ret;
2224 }
2225
2226 /*
2227  * DLM_DEREF_LOCKRES_MSG
2228  */
2229
2230 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2231 {
2232         struct dlm_deref_lockres deref;
2233         int ret = 0, r;
2234         const char *lockname;
2235         unsigned int namelen;
2236
2237         lockname = res->lockname.name;
2238         namelen = res->lockname.len;
2239         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2240
2241         mlog(0, "%s:%.*s: sending deref to %d\n",
2242              dlm->name, namelen, lockname, res->owner);
2243         memset(&deref, 0, sizeof(deref));
2244         deref.node_idx = dlm->node_num;
2245         deref.namelen = namelen;
2246         memcpy(deref.name, lockname, namelen);
2247
2248         ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2249                                  &deref, sizeof(deref), res->owner, &r);
2250         if (ret < 0)
2251                 mlog_errno(ret);
2252         else if (r < 0) {
2253                 /* BAD.  other node says I did not have a ref. */
2254                 mlog(ML_ERROR,"while dropping ref on %s:%.*s "
2255                     "(master=%u) got %d.\n", dlm->name, namelen,
2256                     lockname, res->owner, r);
2257                 dlm_print_one_lock_resource(res);
2258                 BUG();
2259         }
2260         return ret;
2261 }
2262
2263 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data)
2264 {
2265         struct dlm_ctxt *dlm = data;
2266         struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2267         struct dlm_lock_resource *res = NULL;
2268         char *name;
2269         unsigned int namelen;
2270         int ret = -EINVAL;
2271         u8 node;
2272         unsigned int hash;
2273
2274         if (!dlm_grab(dlm))
2275                 return 0;
2276
2277         name = deref->name;
2278         namelen = deref->namelen;
2279         node = deref->node_idx;
2280
2281         if (namelen > DLM_LOCKID_NAME_MAX) {
2282                 mlog(ML_ERROR, "Invalid name length!");
2283                 goto done;
2284         }
2285         if (deref->node_idx >= O2NM_MAX_NODES) {
2286                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2287                 goto done;
2288         }
2289
2290         hash = dlm_lockid_hash(name, namelen);
2291
2292         spin_lock(&dlm->spinlock);
2293         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2294         if (!res) {
2295                 spin_unlock(&dlm->spinlock);
2296                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2297                      dlm->name, namelen, name);
2298                 goto done;
2299         }
2300         spin_unlock(&dlm->spinlock);
2301
2302         spin_lock(&res->spinlock);
2303         BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2304         if (test_bit(node, res->refmap)) {
2305                 ret = 0;
2306                 dlm_lockres_clear_refmap_bit(node, res);
2307         } else {
2308                 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2309                      "but it is already dropped!\n", dlm->name, namelen,
2310                      name, node);
2311                 __dlm_print_one_lock_resource(res);
2312         }
2313         spin_unlock(&res->spinlock);
2314
2315         if (!ret)
2316                 dlm_lockres_calc_usage(dlm, res);
2317 done:
2318         if (res)
2319                 dlm_lockres_put(res);
2320         dlm_put(dlm);
2321         return ret;
2322 }
2323
2324
2325 /*
2326  * DLM_MIGRATE_LOCKRES
2327  */
2328
2329
2330 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2331                                struct dlm_lock_resource *res,
2332                                u8 target)
2333 {
2334         struct dlm_master_list_entry *mle = NULL;
2335         struct dlm_master_list_entry *oldmle = NULL;
2336         struct dlm_migratable_lockres *mres = NULL;
2337         int ret = -EINVAL;
2338         const char *name;
2339         unsigned int namelen;
2340         int mle_added = 0;
2341         struct list_head *queue, *iter;
2342         int i;
2343         struct dlm_lock *lock;
2344         int empty = 1;
2345
2346         if (!dlm_grab(dlm))
2347                 return -EINVAL;
2348
2349         name = res->lockname.name;
2350         namelen = res->lockname.len;
2351
2352         mlog(0, "migrating %.*s to %u\n", namelen, name, target);
2353
2354         /*
2355          * ensure this lockres is a proper candidate for migration
2356          */
2357         spin_lock(&res->spinlock);
2358         if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
2359                 mlog(0, "cannot migrate lockres with unknown owner!\n");
2360                 spin_unlock(&res->spinlock);
2361                 goto leave;
2362         }
2363         if (res->owner != dlm->node_num) {
2364                 mlog(0, "cannot migrate lockres this node doesn't own!\n");
2365                 spin_unlock(&res->spinlock);
2366                 goto leave;
2367         }
2368         mlog(0, "checking queues...\n");
2369         queue = &res->granted;
2370         for (i=0; i<3; i++) {
2371                 list_for_each(iter, queue) {
2372                         lock = list_entry (iter, struct dlm_lock, list);
2373                         empty = 0;
2374                         if (lock->ml.node == dlm->node_num) {
2375                                 mlog(0, "found a lock owned by this node "
2376                                      "still on the %s queue!  will not "
2377                                      "migrate this lockres\n",
2378                                      i==0 ? "granted" :
2379                                      (i==1 ? "converting" : "blocked"));
2380                                 spin_unlock(&res->spinlock);
2381                                 ret = -ENOTEMPTY;
2382                                 goto leave;
2383                         }
2384                 }
2385                 queue++;
2386         }
2387         mlog(0, "all locks on this lockres are nonlocal.  continuing\n");
2388         spin_unlock(&res->spinlock);
2389
2390         /* no work to do */
2391         if (empty) {
2392                 mlog(0, "no locks were found on this lockres! done!\n");
2393                 ret = 0;
2394                 goto leave;
2395         }
2396
2397         /*
2398          * preallocate up front
2399          * if this fails, abort
2400          */
2401
2402         ret = -ENOMEM;
2403         mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2404         if (!mres) {
2405                 mlog_errno(ret);
2406                 goto leave;
2407         }
2408
2409         mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
2410                                                                 GFP_NOFS);
2411         if (!mle) {
2412                 mlog_errno(ret);
2413                 goto leave;
2414         }
2415         ret = 0;
2416
2417         /*
2418          * find a node to migrate the lockres to
2419          */
2420
2421         mlog(0, "picking a migration node\n");
2422         spin_lock(&dlm->spinlock);
2423         /* pick a new node */
2424         if (!test_bit(target, dlm->domain_map) ||
2425             target >= O2NM_MAX_NODES) {
2426                 target = dlm_pick_migration_target(dlm, res);
2427         }
2428         mlog(0, "node %u chosen for migration\n", target);
2429
2430         if (target >= O2NM_MAX_NODES ||
2431             !test_bit(target, dlm->domain_map)) {
2432                 /* target chosen is not alive */
2433                 ret = -EINVAL;
2434         }
2435
2436         if (ret) {
2437                 spin_unlock(&dlm->spinlock);
2438                 goto fail;
2439         }
2440
2441         mlog(0, "continuing with target = %u\n", target);
2442
2443         /*
2444          * clear any existing master requests and
2445          * add the migration mle to the list
2446          */
2447         spin_lock(&dlm->master_lock);
2448         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2449                                     namelen, target, dlm->node_num);
2450         spin_unlock(&dlm->master_lock);
2451         spin_unlock(&dlm->spinlock);
2452
2453         if (ret == -EEXIST) {
2454                 mlog(0, "another process is already migrating it\n");
2455                 goto fail;
2456         }
2457         mle_added = 1;
2458
2459         /*
2460          * set the MIGRATING flag and flush asts
2461          * if we fail after this we need to re-dirty the lockres
2462          */
2463         if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2464                 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2465                      "the target went down.\n", res->lockname.len,
2466                      res->lockname.name, target);
2467                 spin_lock(&res->spinlock);
2468                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2469                 spin_unlock(&res->spinlock);
2470                 ret = -EINVAL;
2471         }
2472
2473 fail:
2474         if (oldmle) {
2475                 /* master is known, detach if not already detached */
2476                 dlm_mle_detach_hb_events(dlm, oldmle);
2477                 dlm_put_mle(oldmle);
2478         }
2479
2480         if (ret < 0) {
2481                 if (mle_added) {
2482                         dlm_mle_detach_hb_events(dlm, mle);
2483                         dlm_put_mle(mle);
2484                 } else if (mle) {
2485                         kmem_cache_free(dlm_mle_cache, mle);
2486                 }
2487                 goto leave;
2488         }
2489
2490         /*
2491          * at this point, we have a migration target, an mle
2492          * in the master list, and the MIGRATING flag set on
2493          * the lockres
2494          */
2495
2496
2497         /* get an extra reference on the mle.
2498          * otherwise the assert_master from the new
2499          * master will destroy this.
2500          * also, make sure that all callers of dlm_get_mle
2501          * take both dlm->spinlock and dlm->master_lock */
2502         spin_lock(&dlm->spinlock);
2503         spin_lock(&dlm->master_lock);
2504         dlm_get_mle_inuse(mle);
2505         spin_unlock(&dlm->master_lock);
2506         spin_unlock(&dlm->spinlock);
2507
2508         /* notify new node and send all lock state */
2509         /* call send_one_lockres with migration flag.
2510          * this serves as notice to the target node that a
2511          * migration is starting. */
2512         ret = dlm_send_one_lockres(dlm, res, mres, target,
2513                                    DLM_MRES_MIGRATION);
2514
2515         if (ret < 0) {
2516                 mlog(0, "migration to node %u failed with %d\n",
2517                      target, ret);
2518                 /* migration failed, detach and clean up mle */
2519                 dlm_mle_detach_hb_events(dlm, mle);
2520                 dlm_put_mle(mle);
2521                 dlm_put_mle_inuse(mle);
2522                 spin_lock(&res->spinlock);
2523                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2524                 spin_unlock(&res->spinlock);
2525                 goto leave;
2526         }
2527
2528         /* at this point, the target sends a message to all nodes,
2529          * (using dlm_do_migrate_request).  this node is skipped since
2530          * we had to put an mle in the list to begin the process.  this
2531          * node now waits for target to do an assert master.  this node
2532          * will be the last one notified, ensuring that the migration
2533          * is complete everywhere.  if the target dies while this is
2534          * going on, some nodes could potentially see the target as the
2535          * master, so it is important that my recovery finds the migration
2536          * mle and sets the master to UNKNONWN. */
2537
2538
2539         /* wait for new node to assert master */
2540         while (1) {
2541                 ret = wait_event_interruptible_timeout(mle->wq,
2542                                         (atomic_read(&mle->woken) == 1),
2543                                         msecs_to_jiffies(5000));
2544
2545                 if (ret >= 0) {
2546                         if (atomic_read(&mle->woken) == 1 ||
2547                             res->owner == target)
2548                                 break;
2549
2550                         mlog(0, "timed out during migration\n");
2551                         /* avoid hang during shutdown when migrating lockres
2552                          * to a node which also goes down */
2553                         if (dlm_is_node_dead(dlm, target)) {
2554                                 mlog(0, "%s:%.*s: expected migration "
2555                                      "target %u is no longer up, restarting\n",
2556                                      dlm->name, res->lockname.len,
2557                                      res->lockname.name, target);
2558                                 ret = -ERESTARTSYS;
2559                         }
2560                 }
2561                 if (ret == -ERESTARTSYS) {
2562                         /* migration failed, detach and clean up mle */
2563                         dlm_mle_detach_hb_events(dlm, mle);
2564                         dlm_put_mle(mle);
2565                         dlm_put_mle_inuse(mle);
2566                         spin_lock(&res->spinlock);
2567                         res->state &= ~DLM_LOCK_RES_MIGRATING;
2568                         spin_unlock(&res->spinlock);
2569                         goto leave;
2570                 }
2571                 /* TODO: if node died: stop, clean up, return error */
2572         }
2573
2574         /* all done, set the owner, clear the flag */
2575         spin_lock(&res->spinlock);
2576         dlm_set_lockres_owner(dlm, res, target);
2577         res->state &= ~DLM_LOCK_RES_MIGRATING;
2578         dlm_remove_nonlocal_locks(dlm, res);
2579         spin_unlock(&res->spinlock);
2580         wake_up(&res->wq);
2581
2582         /* master is known, detach if not already detached */
2583         dlm_mle_detach_hb_events(dlm, mle);
2584         dlm_put_mle_inuse(mle);
2585         ret = 0;
2586
2587         dlm_lockres_calc_usage(dlm, res);
2588
2589 leave:
2590         /* re-dirty the lockres if we failed */
2591         if (ret < 0)
2592                 dlm_kick_thread(dlm, res);
2593
2594         /* TODO: cleanup */
2595         if (mres)
2596                 free_page((unsigned long)mres);
2597
2598         dlm_put(dlm);
2599
2600         mlog(0, "returning %d\n", ret);
2601         return ret;
2602 }
2603
2604 #define DLM_MIGRATION_RETRY_MS  100
2605
2606 /* Should be called only after beginning the domain leave process.
2607  * There should not be any remaining locks on nonlocal lock resources,
2608  * and there should be no local locks left on locally mastered resources.
2609  *
2610  * Called with the dlm spinlock held, may drop it to do migration, but
2611  * will re-acquire before exit.
2612  *
2613  * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped */
2614 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2615 {
2616         int ret;
2617         int lock_dropped = 0;
2618
2619         if (res->owner != dlm->node_num) {
2620                 if (!__dlm_lockres_unused(res)) {
2621                         mlog(ML_ERROR, "%s:%.*s: this node is not master, "
2622                              "trying to free this but locks remain\n",
2623                              dlm->name, res->lockname.len, res->lockname.name);
2624                 }
2625                 goto leave;
2626         }
2627
2628         /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2629         spin_unlock(&dlm->spinlock);
2630         lock_dropped = 1;
2631         while (1) {
2632                 ret = dlm_migrate_lockres(dlm, res, O2NM_MAX_NODES);
2633                 if (ret >= 0)
2634                         break;
2635                 if (ret == -ENOTEMPTY) {
2636                         mlog(ML_ERROR, "lockres %.*s still has local locks!\n",
2637                                 res->lockname.len, res->lockname.name);
2638                         BUG();
2639                 }
2640
2641                 mlog(0, "lockres %.*s: migrate failed, "
2642                      "retrying\n", res->lockname.len,
2643                      res->lockname.name);
2644                 msleep(DLM_MIGRATION_RETRY_MS);
2645         }
2646         spin_lock(&dlm->spinlock);
2647 leave:
2648         return lock_dropped;
2649 }
2650
2651 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2652 {
2653         int ret;
2654         spin_lock(&dlm->ast_lock);
2655         spin_lock(&lock->spinlock);
2656         ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2657         spin_unlock(&lock->spinlock);
2658         spin_unlock(&dlm->ast_lock);
2659         return ret;
2660 }
2661
2662 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2663                                      struct dlm_lock_resource *res,
2664                                      u8 mig_target)
2665 {
2666         int can_proceed;
2667         spin_lock(&res->spinlock);
2668         can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2669         spin_unlock(&res->spinlock);
2670
2671         /* target has died, so make the caller break out of the
2672          * wait_event, but caller must recheck the domain_map */
2673         spin_lock(&dlm->spinlock);
2674         if (!test_bit(mig_target, dlm->domain_map))
2675                 can_proceed = 1;
2676         spin_unlock(&dlm->spinlock);
2677         return can_proceed;
2678 }
2679
2680 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2681                                 struct dlm_lock_resource *res)
2682 {
2683         int ret;
2684         spin_lock(&res->spinlock);
2685         ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2686         spin_unlock(&res->spinlock);
2687         return ret;
2688 }
2689
2690
2691 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2692                                        struct dlm_lock_resource *res,
2693                                        u8 target)
2694 {
2695         int ret = 0;
2696
2697         mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2698                res->lockname.len, res->lockname.name, dlm->node_num,
2699                target);
2700         /* need to set MIGRATING flag on lockres.  this is done by
2701          * ensuring that all asts have been flushed for this lockres. */
2702         spin_lock(&res->spinlock);
2703         BUG_ON(res->migration_pending);
2704         res->migration_pending = 1;
2705         /* strategy is to reserve an extra ast then release
2706          * it below, letting the release do all of the work */
2707         __dlm_lockres_reserve_ast(res);
2708         spin_unlock(&res->spinlock);
2709
2710         /* now flush all the pending asts.. hang out for a bit */
2711         dlm_kick_thread(dlm, res);
2712         wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2713         dlm_lockres_release_ast(dlm, res);
2714
2715         mlog(0, "about to wait on migration_wq, dirty=%s\n",
2716                res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2717         /* if the extra ref we just put was the final one, this
2718          * will pass thru immediately.  otherwise, we need to wait
2719          * for the last ast to finish. */
2720 again:
2721         ret = wait_event_interruptible_timeout(dlm->migration_wq,
2722                    dlm_migration_can_proceed(dlm, res, target),
2723                    msecs_to_jiffies(1000));
2724         if (ret < 0) {
2725                 mlog(0, "woken again: migrating? %s, dead? %s\n",
2726                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2727                        test_bit(target, dlm->domain_map) ? "no":"yes");
2728         } else {
2729                 mlog(0, "all is well: migrating? %s, dead? %s\n",
2730                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2731                        test_bit(target, dlm->domain_map) ? "no":"yes");
2732         }
2733         if (!dlm_migration_can_proceed(dlm, res, target)) {
2734                 mlog(0, "trying again...\n");
2735                 goto again;
2736         }
2737
2738         /* did the target go down or die? */
2739         spin_lock(&dlm->spinlock);
2740         if (!test_bit(target, dlm->domain_map)) {
2741                 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2742                      target);
2743                 ret = -EHOSTDOWN;
2744         }
2745         spin_unlock(&dlm->spinlock);
2746
2747         /*
2748          * at this point:
2749          *
2750          *   o the DLM_LOCK_RES_MIGRATING flag is set
2751          *   o there are no pending asts on this lockres
2752          *   o all processes trying to reserve an ast on this
2753          *     lockres must wait for the MIGRATING flag to clear
2754          */
2755         return ret;
2756 }
2757
2758 /* last step in the migration process.
2759  * original master calls this to free all of the dlm_lock
2760  * structures that used to be for other nodes. */
2761 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2762                                       struct dlm_lock_resource *res)
2763 {
2764         struct list_head *iter, *iter2;
2765         struct list_head *queue = &res->granted;
2766         int i, bit;
2767         struct dlm_lock *lock;
2768
2769         assert_spin_locked(&res->spinlock);
2770
2771         BUG_ON(res->owner == dlm->node_num);
2772
2773         for (i=0; i<3; i++) {
2774                 list_for_each_safe(iter, iter2, queue) {
2775                         lock = list_entry (iter, struct dlm_lock, list);
2776                         if (lock->ml.node != dlm->node_num) {
2777                                 mlog(0, "putting lock for node %u\n",
2778                                      lock->ml.node);
2779                                 /* be extra careful */
2780                                 BUG_ON(!list_empty(&lock->ast_list));
2781                                 BUG_ON(!list_empty(&lock->bast_list));
2782                                 BUG_ON(lock->ast_pending);
2783                                 BUG_ON(lock->bast_pending);
2784                                 dlm_lockres_clear_refmap_bit(lock->ml.node, res);
2785                                 list_del_init(&lock->list);
2786                                 dlm_lock_put(lock);
2787                         }
2788                 }
2789                 queue++;
2790         }
2791         bit = 0;
2792         while (1) {
2793                 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2794                 if (bit >= O2NM_MAX_NODES)
2795                         break;
2796                 /* do not clear the local node reference, if there is a
2797                  * process holding this, let it drop the ref itself */
2798                 if (bit != dlm->node_num) {
2799                         mlog(0, "%s:%.*s: node %u had a ref to this "
2800                              "migrating lockres, clearing\n", dlm->name,
2801                              res->lockname.len, res->lockname.name, bit);
2802                         dlm_lockres_clear_refmap_bit(bit, res);
2803                 }
2804                 bit++;
2805         }
2806 }
2807
2808 /* for now this is not too intelligent.  we will
2809  * need stats to make this do the right thing.
2810  * this just finds the first lock on one of the
2811  * queues and uses that node as the target. */
2812 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2813                                     struct dlm_lock_resource *res)
2814 {
2815         int i;
2816         struct list_head *queue = &res->granted;
2817         struct list_head *iter;
2818         struct dlm_lock *lock;
2819         int nodenum;
2820
2821         assert_spin_locked(&dlm->spinlock);
2822
2823         spin_lock(&res->spinlock);
2824         for (i=0; i<3; i++) {
2825                 list_for_each(iter, queue) {
2826                         /* up to the caller to make sure this node
2827                          * is alive */
2828                         lock = list_entry (iter, struct dlm_lock, list);
2829                         if (lock->ml.node != dlm->node_num) {
2830                                 spin_unlock(&res->spinlock);
2831                                 return lock->ml.node;
2832                         }
2833                 }
2834                 queue++;
2835         }
2836         spin_unlock(&res->spinlock);
2837         mlog(0, "have not found a suitable target yet! checking domain map\n");
2838
2839         /* ok now we're getting desperate.  pick anyone alive. */
2840         nodenum = -1;
2841         while (1) {
2842                 nodenum = find_next_bit(dlm->domain_map,
2843                                         O2NM_MAX_NODES, nodenum+1);
2844                 mlog(0, "found %d in domain map\n", nodenum);
2845                 if (nodenum >= O2NM_MAX_NODES)
2846                         break;
2847                 if (nodenum != dlm->node_num) {
2848                         mlog(0, "picking %d\n", nodenum);
2849                         return nodenum;
2850                 }
2851         }
2852
2853         mlog(0, "giving up.  no master to migrate to\n");
2854         return DLM_LOCK_RES_OWNER_UNKNOWN;
2855 }
2856
2857
2858
2859 /* this is called by the new master once all lockres
2860  * data has been received */
2861 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2862                                   struct dlm_lock_resource *res,
2863                                   u8 master, u8 new_master,
2864                                   struct dlm_node_iter *iter)
2865 {
2866         struct dlm_migrate_request migrate;
2867         int ret, status = 0;
2868         int nodenum;
2869
2870         memset(&migrate, 0, sizeof(migrate));
2871         migrate.namelen = res->lockname.len;
2872         memcpy(migrate.name, res->lockname.name, migrate.namelen);
2873         migrate.new_master = new_master;
2874         migrate.master = master;
2875
2876         ret = 0;
2877
2878         /* send message to all nodes, except the master and myself */
2879         while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2880                 if (nodenum == master ||
2881                     nodenum == new_master)
2882                         continue;
2883
2884                 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2885                                          &migrate, sizeof(migrate), nodenum,
2886                                          &status);
2887                 if (ret < 0)
2888                         mlog_errno(ret);
2889                 else if (status < 0) {
2890                         mlog(0, "migrate request (node %u) returned %d!\n",
2891                              nodenum, status);
2892                         ret = status;
2893                 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
2894                         /* during the migration request we short-circuited
2895                          * the mastery of the lockres.  make sure we have
2896                          * a mastery ref for nodenum */
2897                         mlog(0, "%s:%.*s: need ref for node %u\n",
2898                              dlm->name, res->lockname.len, res->lockname.name,
2899                              nodenum);
2900                         spin_lock(&res->spinlock);
2901                         dlm_lockres_set_refmap_bit(nodenum, res);
2902                         spin_unlock(&res->spinlock);
2903                 }
2904         }
2905
2906         if (ret < 0)
2907                 mlog_errno(ret);
2908
2909         mlog(0, "returning ret=%d\n", ret);
2910         return ret;
2911 }
2912
2913
2914 /* if there is an existing mle for this lockres, we now know who the master is.
2915  * (the one who sent us *this* message) we can clear it up right away.
2916  * since the process that put the mle on the list still has a reference to it,
2917  * we can unhash it now, set the master and wake the process.  as a result,
2918  * we will have no mle in the list to start with.  now we can add an mle for
2919  * the migration and this should be the only one found for those scanning the
2920  * list.  */
2921 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data)
2922 {
2923         struct dlm_ctxt *dlm = data;
2924         struct dlm_lock_resource *res = NULL;
2925         struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
2926         struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
2927         const char *name;
2928         unsigned int namelen, hash;
2929         int ret = 0;
2930
2931         if (!dlm_grab(dlm))
2932                 return -EINVAL;
2933
2934         name = migrate->name;
2935         namelen = migrate->namelen;
2936         hash = dlm_lockid_hash(name, namelen);
2937
2938         /* preallocate.. if this fails, abort */
2939         mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
2940                                                          GFP_NOFS);
2941
2942         if (!mle) {
2943                 ret = -ENOMEM;
2944                 goto leave;
2945         }
2946
2947         /* check for pre-existing lock */
2948         spin_lock(&dlm->spinlock);
2949         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
2950         spin_lock(&dlm->master_lock);
2951
2952         if (res) {
2953                 spin_lock(&res->spinlock);
2954                 if (res->state & DLM_LOCK_RES_RECOVERING) {
2955                         /* if all is working ok, this can only mean that we got
2956                         * a migrate request from a node that we now see as
2957                         * dead.  what can we do here?  drop it to the floor? */
2958                         spin_unlock(&res->spinlock);
2959                         mlog(ML_ERROR, "Got a migrate request, but the "
2960                              "lockres is marked as recovering!");
2961                         kmem_cache_free(dlm_mle_cache, mle);
2962                         ret = -EINVAL; /* need a better solution */
2963                         goto unlock;
2964                 }
2965                 res->state |= DLM_LOCK_RES_MIGRATING;
2966                 spin_unlock(&res->spinlock);
2967         }
2968
2969         /* ignore status.  only nonzero status would BUG. */
2970         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
2971                                     name, namelen,
2972                                     migrate->new_master,
2973                                     migrate->master);
2974
2975 unlock:
2976         spin_unlock(&dlm->master_lock);
2977         spin_unlock(&dlm->spinlock);
2978
2979         if (oldmle) {
2980                 /* master is known, detach if not already detached */
2981                 dlm_mle_detach_hb_events(dlm, oldmle);
2982                 dlm_put_mle(oldmle);
2983         }
2984
2985         if (res)
2986                 dlm_lockres_put(res);
2987 leave:
2988         dlm_put(dlm);
2989         return ret;
2990 }
2991
2992 /* must be holding dlm->spinlock and dlm->master_lock
2993  * when adding a migration mle, we can clear any other mles
2994  * in the master list because we know with certainty that
2995  * the master is "master".  so we remove any old mle from
2996  * the list after setting it's master field, and then add
2997  * the new migration mle.  this way we can hold with the rule
2998  * of having only one mle for a given lock name at all times. */
2999 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3000                                  struct dlm_lock_resource *res,
3001                                  struct dlm_master_list_entry *mle,
3002                                  struct dlm_master_list_entry **oldmle,
3003                                  const char *name, unsigned int namelen,
3004                                  u8 new_master, u8 master)
3005 {
3006         int found;
3007         int ret = 0;
3008
3009         *oldmle = NULL;
3010
3011         mlog_entry_void();
3012
3013         assert_spin_locked(&dlm->spinlock);
3014         assert_spin_locked(&dlm->master_lock);
3015
3016         /* caller is responsible for any ref taken here on oldmle */
3017         found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3018         if (found) {
3019                 struct dlm_master_list_entry *tmp = *oldmle;
3020                 spin_lock(&tmp->spinlock);
3021                 if (tmp->type == DLM_MLE_MIGRATION) {
3022                         if (master == dlm->node_num) {
3023                                 /* ah another process raced me to it */
3024                                 mlog(0, "tried to migrate %.*s, but some "
3025                                      "process beat me to it\n",
3026                                      namelen, name);
3027                                 ret = -EEXIST;
3028                         } else {
3029                                 /* bad.  2 NODES are trying to migrate! */
3030                                 mlog(ML_ERROR, "migration error  mle: "
3031                                      "master=%u new_master=%u // request: "
3032                                      "master=%u new_master=%u // "
3033                                      "lockres=%.*s\n",
3034                                      tmp->master, tmp->new_master,
3035                                      master, new_master,
3036                                      namelen, name);
3037                                 BUG();
3038                         }
3039                 } else {
3040                         /* this is essentially what assert_master does */
3041                         tmp->master = master;
3042                         atomic_set(&tmp->woken, 1);
3043                         wake_up(&tmp->wq);
3044                         /* remove it from the list so that only one
3045                          * mle will be found */
3046                         list_del_init(&tmp->list);
3047                         /* this was obviously WRONG.  mle is uninited here.  should be tmp. */
3048                         __dlm_mle_detach_hb_events(dlm, tmp);
3049                         ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3050                         mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3051                             "telling master to get ref for cleared out mle "
3052                             "during migration\n", dlm->name, namelen, name,
3053                             master, new_master);
3054                 }
3055                 spin_unlock(&tmp->spinlock);
3056         }
3057
3058         /* now add a migration mle to the tail of the list */
3059         dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3060         mle->new_master = new_master;
3061         /* the new master will be sending an assert master for this.
3062          * at that point we will get the refmap reference */
3063         mle->master = master;
3064         /* do this for consistency with other mle types */
3065         set_bit(new_master, mle->maybe_map);
3066         list_add(&mle->list, &dlm->master_list);
3067
3068         return ret;
3069 }
3070
3071
3072 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3073 {
3074         struct list_head *iter, *iter2;
3075         struct dlm_master_list_entry *mle;
3076         struct dlm_lock_resource *res;
3077         unsigned int hash;
3078
3079         mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node);
3080 top:
3081         assert_spin_locked(&dlm->spinlock);
3082
3083         /* clean the master list */
3084         spin_lock(&dlm->master_lock);
3085         list_for_each_safe(iter, iter2, &dlm->master_list) {
3086                 mle = list_entry(iter, struct dlm_master_list_entry, list);
3087
3088                 BUG_ON(mle->type != DLM_MLE_BLOCK &&
3089                        mle->type != DLM_MLE_MASTER &&
3090                        mle->type != DLM_MLE_MIGRATION);
3091
3092                 /* MASTER mles are initiated locally.  the waiting
3093                  * process will notice the node map change
3094                  * shortly.  let that happen as normal. */
3095                 if (mle->type == DLM_MLE_MASTER)
3096                         continue;
3097
3098
3099                 /* BLOCK mles are initiated by other nodes.
3100                  * need to clean up if the dead node would have
3101                  * been the master. */
3102                 if (mle->type == DLM_MLE_BLOCK) {
3103                         int bit;
3104
3105                         spin_lock(&mle->spinlock);
3106                         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3107                         if (bit != dead_node) {
3108                                 mlog(0, "mle found, but dead node %u would "
3109                                      "not have been master\n", dead_node);
3110                                 spin_unlock(&mle->spinlock);
3111                         } else {
3112                                 /* must drop the refcount by one since the
3113                                  * assert_master will never arrive.  this
3114                                  * may result in the mle being unlinked and
3115                                  * freed, but there may still be a process
3116                                  * waiting in the dlmlock path which is fine. */
3117                                 mlog(0, "node %u was expected master\n",
3118                                      dead_node);
3119                                 atomic_set(&mle->woken, 1);
3120                                 spin_unlock(&mle->spinlock);
3121                                 wake_up(&mle->wq);
3122                                 /* do not need events any longer, so detach
3123                                  * from heartbeat */
3124                                 __dlm_mle_detach_hb_events(dlm, mle);
3125                                 __dlm_put_mle(mle);
3126                         }
3127                         continue;
3128                 }
3129
3130                 /* everything else is a MIGRATION mle */
3131
3132                 /* the rule for MIGRATION mles is that the master
3133                  * becomes UNKNOWN if *either* the original or
3134                  * the new master dies.  all UNKNOWN lockreses
3135                  * are sent to whichever node becomes the recovery
3136                  * master.  the new master is responsible for
3137                  * determining if there is still a master for
3138                  * this lockres, or if he needs to take over
3139                  * mastery.  either way, this node should expect
3140                  * another message to resolve this. */
3141                 if (mle->master != dead_node &&
3142                     mle->new_master != dead_node)
3143                         continue;
3144
3145                 /* if we have reached this point, this mle needs to
3146                  * be removed from the list and freed. */
3147
3148                 /* remove from the list early.  NOTE: unlinking
3149                  * list_head while in list_for_each_safe */
3150                 __dlm_mle_detach_hb_events(dlm, mle);
3151                 spin_lock(&mle->spinlock);
3152                 list_del_init(&mle->list);
3153                 atomic_set(&mle->woken, 1);
3154                 spin_unlock(&mle->spinlock);
3155                 wake_up(&mle->wq);
3156
3157                 mlog(0, "%s: node %u died during migration from "
3158                      "%u to %u!\n", dlm->name, dead_node,
3159                      mle->master, mle->new_master);
3160                 /* if there is a lockres associated with this
3161                  * mle, find it and set its owner to UNKNOWN */
3162                 hash = dlm_lockid_hash(mle->u.name.name, mle->u.name.len);
3163                 res = __dlm_lookup_lockres(dlm, mle->u.name.name,
3164                                            mle->u.name.len, hash);
3165                 if (res) {
3166                         /* unfortunately if we hit this rare case, our
3167                          * lock ordering is messed.  we need to drop
3168                          * the master lock so that we can take the
3169                          * lockres lock, meaning that we will have to
3170                          * restart from the head of list. */
3171                         spin_unlock(&dlm->master_lock);
3172
3173                         /* move lockres onto recovery list */
3174                         spin_lock(&res->spinlock);
3175                         dlm_set_lockres_owner(dlm, res,
3176                                         DLM_LOCK_RES_OWNER_UNKNOWN);
3177                         dlm_move_lockres_to_recovery_list(dlm, res);
3178                         spin_unlock(&res->spinlock);
3179                         dlm_lockres_put(res);
3180
3181                         /* about to get rid of mle, detach from heartbeat */
3182                         __dlm_mle_detach_hb_events(dlm, mle);
3183
3184                         /* dump the mle */
3185                         spin_lock(&dlm->master_lock);
3186                         __dlm_put_mle(mle);
3187                         spin_unlock(&dlm->master_lock);
3188
3189                         /* restart */
3190                         goto top;
3191                 }
3192
3193                 /* this may be the last reference */
3194                 __dlm_put_mle(mle);
3195         }
3196         spin_unlock(&dlm->master_lock);
3197 }
3198
3199
3200 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3201                          u8 old_master)
3202 {
3203         struct dlm_node_iter iter;
3204         int ret = 0;
3205
3206         spin_lock(&dlm->spinlock);
3207         dlm_node_iter_init(dlm->domain_map, &iter);
3208         clear_bit(old_master, iter.node_map);
3209         clear_bit(dlm->node_num, iter.node_map);
3210         spin_unlock(&dlm->spinlock);
3211
3212         /* ownership of the lockres is changing.  account for the
3213          * mastery reference here since old_master will briefly have
3214          * a reference after the migration completes */
3215         spin_lock(&res->spinlock);
3216         dlm_lockres_set_refmap_bit(old_master, res);
3217         spin_unlock(&res->spinlock);
3218
3219         mlog(0, "now time to do a migrate request to other nodes\n");
3220         ret = dlm_do_migrate_request(dlm, res, old_master,
3221                                      dlm->node_num, &iter);
3222         if (ret < 0) {
3223                 mlog_errno(ret);
3224                 goto leave;
3225         }
3226
3227         mlog(0, "doing assert master of %.*s to all except the original node\n",
3228              res->lockname.len, res->lockname.name);
3229         /* this call now finishes out the nodemap
3230          * even if one or more nodes die */
3231         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3232                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3233         if (ret < 0) {
3234                 /* no longer need to retry.  all living nodes contacted. */
3235                 mlog_errno(ret);
3236                 ret = 0;
3237         }
3238
3239         memset(iter.node_map, 0, sizeof(iter.node_map));
3240         set_bit(old_master, iter.node_map);
3241         mlog(0, "doing assert master of %.*s back to %u\n",
3242              res->lockname.len, res->lockname.name, old_master);
3243         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3244                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3245         if (ret < 0) {
3246                 mlog(0, "assert master to original master failed "
3247                      "with %d.\n", ret);
3248                 /* the only nonzero status here would be because of
3249                  * a dead original node.  we're done. */
3250                 ret = 0;
3251         }
3252
3253         /* all done, set the owner, clear the flag */
3254         spin_lock(&res->spinlock);
3255         dlm_set_lockres_owner(dlm, res, dlm->node_num);
3256         res->state &= ~DLM_LOCK_RES_MIGRATING;
3257         spin_unlock(&res->spinlock);
3258         /* re-dirty it on the new master */
3259         dlm_kick_thread(dlm, res);
3260         wake_up(&res->wq);
3261 leave:
3262         return ret;
3263 }
3264
3265 /*
3266  * LOCKRES AST REFCOUNT
3267  * this is integral to migration
3268  */
3269
3270 /* for future intent to call an ast, reserve one ahead of time.
3271  * this should be called only after waiting on the lockres
3272  * with dlm_wait_on_lockres, and while still holding the
3273  * spinlock after the call. */
3274 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3275 {
3276         assert_spin_locked(&res->spinlock);
3277         if (res->state & DLM_LOCK_RES_MIGRATING) {
3278                 __dlm_print_one_lock_resource(res);
3279         }
3280         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3281
3282         atomic_inc(&res->asts_reserved);
3283 }
3284
3285 /*
3286  * used to drop the reserved ast, either because it went unused,
3287  * or because the ast/bast was actually called.
3288  *
3289  * also, if there is a pending migration on this lockres,
3290  * and this was the last pending ast on the lockres,
3291  * atomically set the MIGRATING flag before we drop the lock.
3292  * this is how we ensure that migration can proceed with no
3293  * asts in progress.  note that it is ok if the state of the
3294  * queues is such that a lock should be granted in the future
3295  * or that a bast should be fired, because the new master will
3296  * shuffle the lists on this lockres as soon as it is migrated.
3297  */
3298 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3299                              struct dlm_lock_resource *res)
3300 {
3301         if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3302                 return;
3303
3304         if (!res->migration_pending) {
3305                 spin_unlock(&res->spinlock);
3306                 return;
3307         }
3308
3309         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3310         res->migration_pending = 0;
3311         res->state |= DLM_LOCK_RES_MIGRATING;
3312         spin_unlock(&res->spinlock);
3313         wake_up(&res->wq);
3314         wake_up(&dlm->migration_wq);
3315 }