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