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