fs/ocfs2/dlm/dlmmaster.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* -*- mode: c; c-basic-offset: 8; -*-
   3  * vim: noexpandtab sw=8 ts=8 sts=0:
   4  *
   5  * dlmmod.c
   6  *
   7  * standalone DLM module
   8  *
   9  * Copyright (C) 2004 Oracle.  All rights reserved.
  10  */
  11
  12
  13 #include <linux/module.h>
  14 #include <linux/fs.h>
  15 #include <linux/types.h>
  16 #include <linux/slab.h>
  17 #include <linux/highmem.h>
  18 #include <linux/init.h>
  19 #include <linux/sysctl.h>
  20 #include <linux/random.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/socket.h>
  23 #include <linux/inet.h>
  24 #include <linux/spinlock.h>
  25 #include <linux/delay.h>
  26
  27
  28 #include "../cluster/heartbeat.h"
  29 #include "../cluster/nodemanager.h"
  30 #include "../cluster/tcp.h"
  31
  32 #include "dlmapi.h"
  33 #include "dlmcommon.h"
  34 #include "dlmdomain.h"
  35 #include "dlmdebug.h"
  36
  37 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
  38 #include "../cluster/masklog.h"
  39
  40 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
  41                               struct dlm_master_list_entry *mle,
  42                               struct o2nm_node *node,
  43                               int idx);
  44 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
  45                             struct dlm_master_list_entry *mle,
  46                             struct o2nm_node *node,
  47                             int idx);
  48
  49 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
  50 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
  51                                 struct dlm_lock_resource *res,
  52                                 void *nodemap, u32 flags);
  53 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
  54
  55 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
  56                                 struct dlm_master_list_entry *mle,
  57                                 const char *name,
  58                                 unsigned int namelen)
  59 {
  60         if (dlm != mle->dlm)
  61                 return 0;
  62
  63         if (namelen != mle->mnamelen ||
  64             memcmp(name, mle->mname, namelen) != 0)
  65                 return 0;
  66
  67         return 1;
  68 }
  69
  70 static struct kmem_cache *dlm_lockres_cache;
  71 static struct kmem_cache *dlm_lockname_cache;
  72 static struct kmem_cache *dlm_mle_cache;
  73
  74 static void dlm_mle_release(struct kref *kref);
  75 static void dlm_init_mle(struct dlm_master_list_entry *mle,
  76                         enum dlm_mle_type type,
  77                         struct dlm_ctxt *dlm,
  78                         struct dlm_lock_resource *res,
  79                         const char *name,
  80                         unsigned int namelen);
  81 static void dlm_put_mle(struct dlm_master_list_entry *mle);
  82 static void __dlm_put_mle(struct dlm_master_list_entry *mle);
  83 static int dlm_find_mle(struct dlm_ctxt *dlm,
  84                         struct dlm_master_list_entry **mle,
  85                         char *name, unsigned int namelen);
  86
  87 static int dlm_do_master_request(struct dlm_lock_resource *res,
  88                                  struct dlm_master_list_entry *mle, int to);
  89
  90
  91 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
  92                                      struct dlm_lock_resource *res,
  93                                      struct dlm_master_list_entry *mle,
  94                                      int *blocked);
  95 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
  96                                     struct dlm_lock_resource *res,
  97                                     struct dlm_master_list_entry *mle,
  98                                     int blocked);
  99 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
 100                                  struct dlm_lock_resource *res,
 101                                  struct dlm_master_list_entry *mle,
 102                                  struct dlm_master_list_entry **oldmle,
 103                                  const char *name, unsigned int namelen,
 104                                  u8 new_master, u8 master);
 105
 106 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
 107                                     struct dlm_lock_resource *res);
 108 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
 109                                       struct dlm_lock_resource *res);
 110 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
 111                                        struct dlm_lock_resource *res,
 112                                        u8 target);
 113 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
 114                                        struct dlm_lock_resource *res);
 115
 116
 117 int dlm_is_host_down(int errno)
 118 {
 119         switch (errno) {
 120                 case -EBADF:
 121                 case -ECONNREFUSED:
 122                 case -ENOTCONN:
 123                 case -ECONNRESET:
 124                 case -EPIPE:
 125                 case -EHOSTDOWN:
 126                 case -EHOSTUNREACH:
 127                 case -ETIMEDOUT:
 128                 case -ECONNABORTED:
 129                 case -ENETDOWN:
 130                 case -ENETUNREACH:
 131                 case -ENETRESET:
 132                 case -ESHUTDOWN:
 133                 case -ENOPROTOOPT:
 134                 case -EINVAL:   /* if returned from our tcp code,
 135                                    this means there is no socket */
 136                         return 1;
 137         }
 138         return 0;
 139 }
 140
 141
 142 /*
 143  * MASTER LIST FUNCTIONS
 144  */
 145
 146
 147 /*
 148  * regarding master list entries and heartbeat callbacks:
 149  *
 150  * in order to avoid sleeping and allocation that occurs in
 151  * heartbeat, master list entries are simply attached to the
 152  * dlm's established heartbeat callbacks.  the mle is attached
 153  * when it is created, and since the dlm->spinlock is held at
 154  * that time, any heartbeat event will be properly discovered
 155  * by the mle.  the mle needs to be detached from the
 156  * dlm->mle_hb_events list as soon as heartbeat events are no
 157  * longer useful to the mle, and before the mle is freed.
 158  *
 159  * as a general rule, heartbeat events are no longer needed by
 160  * the mle once an "answer" regarding the lock master has been
 161  * received.
 162  */
 163 static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
 164                                               struct dlm_master_list_entry *mle)
 165 {
 166         assert_spin_locked(&dlm->spinlock);
 167
 168         list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
 169 }
 170
 171
 172 static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 173                                               struct dlm_master_list_entry *mle)
 174 {
 175         if (!list_empty(&mle->hb_events))
 176                 list_del_init(&mle->hb_events);
 177 }
 178
 179
 180 static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 181                                             struct dlm_master_list_entry *mle)
 182 {
 183         spin_lock(&dlm->spinlock);
 184         __dlm_mle_detach_hb_events(dlm, mle);
 185         spin_unlock(&dlm->spinlock);
 186 }
 187
 188 static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
 189 {
 190         struct dlm_ctxt *dlm;
 191         dlm = mle->dlm;
 192
 193         assert_spin_locked(&dlm->spinlock);
 194         assert_spin_locked(&dlm->master_lock);
 195         mle->inuse++;
 196         kref_get(&mle->mle_refs);
 197 }
 198
 199 static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
 200 {
 201         struct dlm_ctxt *dlm;
 202         dlm = mle->dlm;
 203
 204         spin_lock(&dlm->spinlock);
 205         spin_lock(&dlm->master_lock);
 206         mle->inuse--;
 207         __dlm_put_mle(mle);
 208         spin_unlock(&dlm->master_lock);
 209         spin_unlock(&dlm->spinlock);
 210
 211 }
 212
 213 /* remove from list and free */
 214 static void __dlm_put_mle(struct dlm_master_list_entry *mle)
 215 {
 216         struct dlm_ctxt *dlm;
 217         dlm = mle->dlm;
 218
 219         assert_spin_locked(&dlm->spinlock);
 220         assert_spin_locked(&dlm->master_lock);
 221         if (!kref_read(&mle->mle_refs)) {
 222                 /* this may or may not crash, but who cares.
 223                  * it's a BUG. */
 224                 mlog(ML_ERROR, "bad mle: %p\n", mle);
 225                 dlm_print_one_mle(mle);
 226                 BUG();
 227         } else
 228                 kref_put(&mle->mle_refs, dlm_mle_release);
 229 }
 230
 231
 232 /* must not have any spinlocks coming in */
 233 static void dlm_put_mle(struct dlm_master_list_entry *mle)
 234 {
 235         struct dlm_ctxt *dlm;
 236         dlm = mle->dlm;
 237
 238         spin_lock(&dlm->spinlock);
 239         spin_lock(&dlm->master_lock);
 240         __dlm_put_mle(mle);
 241         spin_unlock(&dlm->master_lock);
 242         spin_unlock(&dlm->spinlock);
 243 }
 244
 245 static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
 246 {
 247         kref_get(&mle->mle_refs);
 248 }
 249
 250 static void dlm_init_mle(struct dlm_master_list_entry *mle,
 251                         enum dlm_mle_type type,
 252                         struct dlm_ctxt *dlm,
 253                         struct dlm_lock_resource *res,
 254                         const char *name,
 255                         unsigned int namelen)
 256 {
 257         assert_spin_locked(&dlm->spinlock);
 258
 259         mle->dlm = dlm;
 260         mle->type = type;
 261         INIT_HLIST_NODE(&mle->master_hash_node);
 262         INIT_LIST_HEAD(&mle->hb_events);
 263         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
 264         spin_lock_init(&mle->spinlock);
 265         init_waitqueue_head(&mle->wq);
 266         atomic_set(&mle->woken, 0);
 267         kref_init(&mle->mle_refs);
 268         memset(mle->response_map, 0, sizeof(mle->response_map));
 269         mle->master = O2NM_MAX_NODES;
 270         mle->new_master = O2NM_MAX_NODES;
 271         mle->inuse = 0;
 272
 273         BUG_ON(mle->type != DLM_MLE_BLOCK &&
 274                mle->type != DLM_MLE_MASTER &&
 275                mle->type != DLM_MLE_MIGRATION);
 276
 277         if (mle->type == DLM_MLE_MASTER) {
 278                 BUG_ON(!res);
 279                 mle->mleres = res;
 280                 memcpy(mle->mname, res->lockname.name, res->lockname.len);
 281                 mle->mnamelen = res->lockname.len;
 282                 mle->mnamehash = res->lockname.hash;
 283         } else {
 284                 BUG_ON(!name);
 285                 mle->mleres = NULL;
 286                 memcpy(mle->mname, name, namelen);
 287                 mle->mnamelen = namelen;
 288                 mle->mnamehash = dlm_lockid_hash(name, namelen);
 289         }
 290
 291         atomic_inc(&dlm->mle_tot_count[mle->type]);
 292         atomic_inc(&dlm->mle_cur_count[mle->type]);
 293
 294         /* copy off the node_map and register hb callbacks on our copy */
 295         memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
 296         memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
 297         clear_bit(dlm->node_num, mle->vote_map);
 298         clear_bit(dlm->node_num, mle->node_map);
 299
 300         /* attach the mle to the domain node up/down events */
 301         __dlm_mle_attach_hb_events(dlm, mle);
 302 }
 303
 304 void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
 305 {
 306         assert_spin_locked(&dlm->spinlock);
 307         assert_spin_locked(&dlm->master_lock);
 308
 309         if (!hlist_unhashed(&mle->master_hash_node))
 310                 hlist_del_init(&mle->master_hash_node);
 311 }
 312
 313 void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
 314 {
 315         struct hlist_head *bucket;
 316
 317         assert_spin_locked(&dlm->master_lock);
 318
 319         bucket = dlm_master_hash(dlm, mle->mnamehash);
 320         hlist_add_head(&mle->master_hash_node, bucket);
 321 }
 322
 323 /* returns 1 if found, 0 if not */
 324 static int dlm_find_mle(struct dlm_ctxt *dlm,
 325                         struct dlm_master_list_entry **mle,
 326                         char *name, unsigned int namelen)
 327 {
 328         struct dlm_master_list_entry *tmpmle;
 329         struct hlist_head *bucket;
 330         unsigned int hash;
 331
 332         assert_spin_locked(&dlm->master_lock);
 333
 334         hash = dlm_lockid_hash(name, namelen);
 335         bucket = dlm_master_hash(dlm, hash);
 336         hlist_for_each_entry(tmpmle, bucket, master_hash_node) {
 337                 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
 338                         continue;
 339                 dlm_get_mle(tmpmle);
 340                 *mle = tmpmle;
 341                 return 1;
 342         }
 343         return 0;
 344 }
 345
 346 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
 347 {
 348         struct dlm_master_list_entry *mle;
 349
 350         assert_spin_locked(&dlm->spinlock);
 351
 352         list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
 353                 if (node_up)
 354                         dlm_mle_node_up(dlm, mle, NULL, idx);
 355                 else
 356                         dlm_mle_node_down(dlm, mle, NULL, idx);
 357         }
 358 }
 359
 360 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
 361                               struct dlm_master_list_entry *mle,
 362                               struct o2nm_node *node, int idx)
 363 {
 364         spin_lock(&mle->spinlock);
 365
 366         if (!test_bit(idx, mle->node_map))
 367                 mlog(0, "node %u already removed from nodemap!\n", idx);
 368         else
 369                 clear_bit(idx, mle->node_map);
 370
 371         spin_unlock(&mle->spinlock);
 372 }
 373
 374 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
 375                             struct dlm_master_list_entry *mle,
 376                             struct o2nm_node *node, int idx)
 377 {
 378         spin_lock(&mle->spinlock);
 379
 380         if (test_bit(idx, mle->node_map))
 381                 mlog(0, "node %u already in node map!\n", idx);
 382         else
 383                 set_bit(idx, mle->node_map);
 384
 385         spin_unlock(&mle->spinlock);
 386 }
 387
 388
 389 int dlm_init_mle_cache(void)
 390 {
 391         dlm_mle_cache = kmem_cache_create("o2dlm_mle",
 392                                           sizeof(struct dlm_master_list_entry),
 393                                           0, SLAB_HWCACHE_ALIGN,
 394                                           NULL);
 395         if (dlm_mle_cache == NULL)
 396                 return -ENOMEM;
 397         return 0;
 398 }
 399
 400 void dlm_destroy_mle_cache(void)
 401 {
 402         kmem_cache_destroy(dlm_mle_cache);
 403 }
 404
 405 static void dlm_mle_release(struct kref *kref)
 406 {
 407         struct dlm_master_list_entry *mle;
 408         struct dlm_ctxt *dlm;
 409
 410         mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
 411         dlm = mle->dlm;
 412
 413         assert_spin_locked(&dlm->spinlock);
 414         assert_spin_locked(&dlm->master_lock);
 415
 416         mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
 417              mle->type);
 418
 419         /* remove from list if not already */
 420         __dlm_unlink_mle(dlm, mle);
 421
 422         /* detach the mle from the domain node up/down events */
 423         __dlm_mle_detach_hb_events(dlm, mle);
 424
 425         atomic_dec(&dlm->mle_cur_count[mle->type]);
 426
 427         /* NOTE: kfree under spinlock here.
 428          * if this is bad, we can move this to a freelist. */
 429         kmem_cache_free(dlm_mle_cache, mle);
 430 }
 431
 432
 433 /*
 434  * LOCK RESOURCE FUNCTIONS
 435  */
 436
 437 int dlm_init_master_caches(void)
 438 {
 439         dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
 440                                               sizeof(struct dlm_lock_resource),
 441                                               0, SLAB_HWCACHE_ALIGN, NULL);
 442         if (!dlm_lockres_cache)
 443                 goto bail;
 444
 445         dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
 446                                                DLM_LOCKID_NAME_MAX, 0,
 447                                                SLAB_HWCACHE_ALIGN, NULL);
 448         if (!dlm_lockname_cache)
 449                 goto bail;
 450
 451         return 0;
 452 bail:
 453         dlm_destroy_master_caches();
 454         return -ENOMEM;
 455 }
 456
 457 void dlm_destroy_master_caches(void)
 458 {
 459         kmem_cache_destroy(dlm_lockname_cache);
 460         dlm_lockname_cache = NULL;
 461
 462         kmem_cache_destroy(dlm_lockres_cache);
 463         dlm_lockres_cache = NULL;
 464 }
 465
 466 static void dlm_lockres_release(struct kref *kref)
 467 {
 468         struct dlm_lock_resource *res;
 469         struct dlm_ctxt *dlm;
 470
 471         res = container_of(kref, struct dlm_lock_resource, refs);
 472         dlm = res->dlm;
 473
 474         /* This should not happen -- all lockres' have a name
 475          * associated with them at init time. */
 476         BUG_ON(!res->lockname.name);
 477
 478         mlog(0, "destroying lockres %.*s\n", res->lockname.len,
 479              res->lockname.name);
 480
 481         atomic_dec(&dlm->res_cur_count);
 482
 483         if (!hlist_unhashed(&res->hash_node) ||
 484             !list_empty(&res->granted) ||
 485             !list_empty(&res->converting) ||
 486             !list_empty(&res->blocked) ||
 487             !list_empty(&res->dirty) ||
 488             !list_empty(&res->recovering) ||
 489             !list_empty(&res->purge)) {
 490                 mlog(ML_ERROR,
 491                      "Going to BUG for resource %.*s."
 492                      "  We're on a list! [%c%c%c%c%c%c%c]\n",
 493                      res->lockname.len, res->lockname.name,
 494                      !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
 495                      !list_empty(&res->granted) ? 'G' : ' ',
 496                      !list_empty(&res->converting) ? 'C' : ' ',
 497                      !list_empty(&res->blocked) ? 'B' : ' ',
 498                      !list_empty(&res->dirty) ? 'D' : ' ',
 499                      !list_empty(&res->recovering) ? 'R' : ' ',
 500                      !list_empty(&res->purge) ? 'P' : ' ');
 501
 502                 dlm_print_one_lock_resource(res);
 503         }
 504
 505         /* By the time we're ready to blow this guy away, we shouldn't
 506          * be on any lists. */
 507         BUG_ON(!hlist_unhashed(&res->hash_node));
 508         BUG_ON(!list_empty(&res->granted));
 509         BUG_ON(!list_empty(&res->converting));
 510         BUG_ON(!list_empty(&res->blocked));
 511         BUG_ON(!list_empty(&res->dirty));
 512         BUG_ON(!list_empty(&res->recovering));
 513         BUG_ON(!list_empty(&res->purge));
 514
 515         kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
 516
 517         kmem_cache_free(dlm_lockres_cache, res);
 518 }
 519
 520 void dlm_lockres_put(struct dlm_lock_resource *res)
 521 {
 522         kref_put(&res->refs, dlm_lockres_release);
 523 }
 524
 525 static void dlm_init_lockres(struct dlm_ctxt *dlm,
 526                              struct dlm_lock_resource *res,
 527                              const char *name, unsigned int namelen)
 528 {
 529         char *qname;
 530
 531         /* If we memset here, we lose our reference to the kmalloc'd
 532          * res->lockname.name, so be sure to init every field
 533          * correctly! */
 534
 535         qname = (char *) res->lockname.name;
 536         memcpy(qname, name, namelen);
 537
 538         res->lockname.len = namelen;
 539         res->lockname.hash = dlm_lockid_hash(name, namelen);
 540
 541         init_waitqueue_head(&res->wq);
 542         spin_lock_init(&res->spinlock);
 543         INIT_HLIST_NODE(&res->hash_node);
 544         INIT_LIST_HEAD(&res->granted);
 545         INIT_LIST_HEAD(&res->converting);
 546         INIT_LIST_HEAD(&res->blocked);
 547         INIT_LIST_HEAD(&res->dirty);
 548         INIT_LIST_HEAD(&res->recovering);
 549         INIT_LIST_HEAD(&res->purge);
 550         INIT_LIST_HEAD(&res->tracking);
 551         atomic_set(&res->asts_reserved, 0);
 552         res->migration_pending = 0;
 553         res->inflight_locks = 0;
 554         res->inflight_assert_workers = 0;
 555
 556         res->dlm = dlm;
 557
 558         kref_init(&res->refs);
 559
 560         atomic_inc(&dlm->res_tot_count);
 561         atomic_inc(&dlm->res_cur_count);
 562
 563         /* just for consistency */
 564         spin_lock(&res->spinlock);
 565         dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
 566         spin_unlock(&res->spinlock);
 567
 568         res->state = DLM_LOCK_RES_IN_PROGRESS;
 569
 570         res->last_used = 0;
 571
 572         spin_lock(&dlm->track_lock);
 573         list_add_tail(&res->tracking, &dlm->tracking_list);
 574         spin_unlock(&dlm->track_lock);
 575
 576         memset(res->lvb, 0, DLM_LVB_LEN);
 577         memset(res->refmap, 0, sizeof(res->refmap));
 578 }
 579
 580 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
 581                                    const char *name,
 582                                    unsigned int namelen)
 583 {
 584         struct dlm_lock_resource *res = NULL;
 585
 586         res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
 587         if (!res)
 588                 goto error;
 589
 590         res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
 591         if (!res->lockname.name)
 592                 goto error;
 593
 594         dlm_init_lockres(dlm, res, name, namelen);
 595         return res;
 596
 597 error:
 598         if (res)
 599                 kmem_cache_free(dlm_lockres_cache, res);
 600         return NULL;
 601 }
 602
 603 void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm,
 604                                 struct dlm_lock_resource *res, int bit)
 605 {
 606         assert_spin_locked(&res->spinlock);
 607
 608         mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len,
 609              res->lockname.name, bit, __builtin_return_address(0));
 610
 611         set_bit(bit, res->refmap);
 612 }
 613
 614 void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm,
 615                                   struct dlm_lock_resource *res, int bit)
 616 {
 617         assert_spin_locked(&res->spinlock);
 618
 619         mlog(0, "res %.*s, clr node %u, %ps()\n", res->lockname.len,
 620              res->lockname.name, bit, __builtin_return_address(0));
 621
 622         clear_bit(bit, res->refmap);
 623 }
 624
 625 static void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
 626                                    struct dlm_lock_resource *res)
 627 {
 628         res->inflight_locks++;
 629
 630         mlog(0, "%s: res %.*s, inflight++: now %u, %ps()\n", dlm->name,
 631              res->lockname.len, res->lockname.name, res->inflight_locks,
 632              __builtin_return_address(0));
 633 }
 634
 635 void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
 636                                    struct dlm_lock_resource *res)
 637 {
 638         assert_spin_locked(&res->spinlock);
 639         __dlm_lockres_grab_inflight_ref(dlm, res);
 640 }
 641
 642 void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
 643                                    struct dlm_lock_resource *res)
 644 {
 645         assert_spin_locked(&res->spinlock);
 646
 647         BUG_ON(res->inflight_locks == 0);
 648
 649         res->inflight_locks--;
 650
 651         mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm->name,
 652              res->lockname.len, res->lockname.name, res->inflight_locks,
 653              __builtin_return_address(0));
 654
 655         wake_up(&res->wq);
 656 }
 657
 658 void __dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm,
 659                 struct dlm_lock_resource *res)
 660 {
 661         assert_spin_locked(&res->spinlock);
 662         res->inflight_assert_workers++;
 663         mlog(0, "%s:%.*s: inflight assert worker++: now %u\n",
 664                         dlm->name, res->lockname.len, res->lockname.name,
 665                         res->inflight_assert_workers);
 666 }
 667
 668 static void __dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
 669                 struct dlm_lock_resource *res)
 670 {
 671         assert_spin_locked(&res->spinlock);
 672         BUG_ON(res->inflight_assert_workers == 0);
 673         res->inflight_assert_workers--;
 674         mlog(0, "%s:%.*s: inflight assert worker--: now %u\n",
 675                         dlm->name, res->lockname.len, res->lockname.name,
 676                         res->inflight_assert_workers);
 677 }
 678
 679 static void dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
 680                 struct dlm_lock_resource *res)
 681 {
 682         spin_lock(&res->spinlock);
 683         __dlm_lockres_drop_inflight_worker(dlm, res);
 684         spin_unlock(&res->spinlock);
 685 }
 686
 687 /*
 688  * lookup a lock resource by name.
 689  * may already exist in the hashtable.
 690  * lockid is null terminated
 691  *
 692  * if not, allocate enough for the lockres and for
 693  * the temporary structure used in doing the mastering.
 694  *
 695  * also, do a lookup in the dlm->master_list to see
 696  * if another node has begun mastering the same lock.
 697  * if so, there should be a block entry in there
 698  * for this name, and we should *not* attempt to master
 699  * the lock here.   need to wait around for that node
 700  * to assert_master (or die).
 701  *
 702  */
 703 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
 704                                           const char *lockid,
 705                                           int namelen,
 706                                           int flags)
 707 {
 708         struct dlm_lock_resource *tmpres=NULL, *res=NULL;
 709         struct dlm_master_list_entry *mle = NULL;
 710         struct dlm_master_list_entry *alloc_mle = NULL;
 711         int blocked = 0;
 712         int ret, nodenum;
 713         struct dlm_node_iter iter;
 714         unsigned int hash;
 715         int tries = 0;
 716         int bit, wait_on_recovery = 0;
 717
 718         BUG_ON(!lockid);
 719
 720         hash = dlm_lockid_hash(lockid, namelen);
 721
 722         mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
 723
 724 lookup:
 725         spin_lock(&dlm->spinlock);
 726         tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
 727         if (tmpres) {
 728                 spin_unlock(&dlm->spinlock);
 729                 spin_lock(&tmpres->spinlock);
 730
 731                 /*
 732                  * Right after dlm spinlock was released, dlm_thread could have
 733                  * purged the lockres. Check if lockres got unhashed. If so
 734                  * start over.
 735                  */
 736                 if (hlist_unhashed(&tmpres->hash_node)) {
 737                         spin_unlock(&tmpres->spinlock);
 738                         dlm_lockres_put(tmpres);
 739                         tmpres = NULL;
 740                         goto lookup;
 741                 }
 742
 743                 /* Wait on the thread that is mastering the resource */
 744                 if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
 745                         __dlm_wait_on_lockres(tmpres);
 746                         BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
 747                         spin_unlock(&tmpres->spinlock);
 748                         dlm_lockres_put(tmpres);
 749                         tmpres = NULL;
 750                         goto lookup;
 751                 }
 752
 753                 /* Wait on the resource purge to complete before continuing */
 754                 if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) {
 755                         BUG_ON(tmpres->owner == dlm->node_num);
 756                         __dlm_wait_on_lockres_flags(tmpres,
 757                                                     DLM_LOCK_RES_DROPPING_REF);
 758                         spin_unlock(&tmpres->spinlock);
 759                         dlm_lockres_put(tmpres);
 760                         tmpres = NULL;
 761                         goto lookup;
 762                 }
 763
 764                 /* Grab inflight ref to pin the resource */
 765                 dlm_lockres_grab_inflight_ref(dlm, tmpres);
 766
 767                 spin_unlock(&tmpres->spinlock);
 768                 if (res) {
 769                         spin_lock(&dlm->track_lock);
 770                         if (!list_empty(&res->tracking))
 771                                 list_del_init(&res->tracking);
 772                         else
 773                                 mlog(ML_ERROR, "Resource %.*s not "
 774                                                 "on the Tracking list\n",
 775                                                 res->lockname.len,
 776                                                 res->lockname.name);
 777                         spin_unlock(&dlm->track_lock);
 778                         dlm_lockres_put(res);
 779                 }
 780                 res = tmpres;
 781                 goto leave;
 782         }
 783
 784         if (!res) {
 785                 spin_unlock(&dlm->spinlock);
 786                 mlog(0, "allocating a new resource\n");
 787                 /* nothing found and we need to allocate one. */
 788                 alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
 789                 if (!alloc_mle)
 790                         goto leave;
 791                 res = dlm_new_lockres(dlm, lockid, namelen);
 792                 if (!res)
 793                         goto leave;
 794                 goto lookup;
 795         }
 796
 797         mlog(0, "no lockres found, allocated our own: %p\n", res);
 798
 799         if (flags & LKM_LOCAL) {
 800                 /* caller knows it's safe to assume it's not mastered elsewhere
 801                  * DONE!  return right away */
 802                 spin_lock(&res->spinlock);
 803                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
 804                 __dlm_insert_lockres(dlm, res);
 805                 dlm_lockres_grab_inflight_ref(dlm, res);
 806                 spin_unlock(&res->spinlock);
 807                 spin_unlock(&dlm->spinlock);
 808                 /* lockres still marked IN_PROGRESS */
 809                 goto wake_waiters;
 810         }
 811
 812         /* check master list to see if another node has started mastering it */
 813         spin_lock(&dlm->master_lock);
 814
 815         /* if we found a block, wait for lock to be mastered by another node */
 816         blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
 817         if (blocked) {
 818                 int mig;
 819                 if (mle->type == DLM_MLE_MASTER) {
 820                         mlog(ML_ERROR, "master entry for nonexistent lock!\n");
 821                         BUG();
 822                 }
 823                 mig = (mle->type == DLM_MLE_MIGRATION);
 824                 /* if there is a migration in progress, let the migration
 825                  * finish before continuing.  we can wait for the absence
 826                  * of the MIGRATION mle: either the migrate finished or
 827                  * one of the nodes died and the mle was cleaned up.
 828                  * if there is a BLOCK here, but it already has a master
 829                  * set, we are too late.  the master does not have a ref
 830                  * for us in the refmap.  detach the mle and drop it.
 831                  * either way, go back to the top and start over. */
 832                 if (mig || mle->master != O2NM_MAX_NODES) {
 833                         BUG_ON(mig && mle->master == dlm->node_num);
 834                         /* we arrived too late.  the master does not
 835                          * have a ref for us. retry. */
 836                         mlog(0, "%s:%.*s: late on %s\n",
 837                              dlm->name, namelen, lockid,
 838                              mig ?  "MIGRATION" : "BLOCK");
 839                         spin_unlock(&dlm->master_lock);
 840                         spin_unlock(&dlm->spinlock);
 841
 842                         /* master is known, detach */
 843                         if (!mig)
 844                                 dlm_mle_detach_hb_events(dlm, mle);
 845                         dlm_put_mle(mle);
 846                         mle = NULL;
 847                         /* this is lame, but we can't wait on either
 848                          * the mle or lockres waitqueue here */
 849                         if (mig)
 850                                 msleep(100);
 851                         goto lookup;
 852                 }
 853         } else {
 854                 /* go ahead and try to master lock on this node */
 855                 mle = alloc_mle;
 856                 /* make sure this does not get freed below */
 857                 alloc_mle = NULL;
 858                 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
 859                 set_bit(dlm->node_num, mle->maybe_map);
 860                 __dlm_insert_mle(dlm, mle);
 861
 862                 /* still holding the dlm spinlock, check the recovery map
 863                  * to see if there are any nodes that still need to be
 864                  * considered.  these will not appear in the mle nodemap
 865                  * but they might own this lockres.  wait on them. */
 866                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 867                 if (bit < O2NM_MAX_NODES) {
 868                         mlog(0, "%s: res %.*s, At least one node (%d) "
 869                              "to recover before lock mastery can begin\n",
 870                              dlm->name, namelen, (char *)lockid, bit);
 871                         wait_on_recovery = 1;
 872                 }
 873         }
 874
 875         /* at this point there is either a DLM_MLE_BLOCK or a
 876          * DLM_MLE_MASTER on the master list, so it's safe to add the
 877          * lockres to the hashtable.  anyone who finds the lock will
 878          * still have to wait on the IN_PROGRESS. */
 879
 880         /* finally add the lockres to its hash bucket */
 881         __dlm_insert_lockres(dlm, res);
 882
 883         /* since this lockres is new it doesn't not require the spinlock */
 884         __dlm_lockres_grab_inflight_ref(dlm, res);
 885
 886         /* get an extra ref on the mle in case this is a BLOCK
 887          * if so, the creator of the BLOCK may try to put the last
 888          * ref at this time in the assert master handler, so we
 889          * need an extra one to keep from a bad ptr deref. */
 890         dlm_get_mle_inuse(mle);
 891         spin_unlock(&dlm->master_lock);
 892         spin_unlock(&dlm->spinlock);
 893
 894 redo_request:
 895         while (wait_on_recovery) {
 896                 /* any cluster changes that occurred after dropping the
 897                  * dlm spinlock would be detectable be a change on the mle,
 898                  * so we only need to clear out the recovery map once. */
 899                 if (dlm_is_recovery_lock(lockid, namelen)) {
 900                         mlog(0, "%s: Recovery map is not empty, but must "
 901                              "master $RECOVERY lock now\n", dlm->name);
 902                         if (!dlm_pre_master_reco_lockres(dlm, res))
 903                                 wait_on_recovery = 0;
 904                         else {
 905                                 mlog(0, "%s: waiting 500ms for heartbeat state "
 906                                     "change\n", dlm->name);
 907                                 msleep(500);
 908                         }
 909                         continue;
 910                 }
 911
 912                 dlm_kick_recovery_thread(dlm);
 913                 msleep(1000);
 914                 dlm_wait_for_recovery(dlm);
 915
 916                 spin_lock(&dlm->spinlock);
 917                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 918                 if (bit < O2NM_MAX_NODES) {
 919                         mlog(0, "%s: res %.*s, At least one node (%d) "
 920                              "to recover before lock mastery can begin\n",
 921                              dlm->name, namelen, (char *)lockid, bit);
 922                         wait_on_recovery = 1;
 923                 } else
 924                         wait_on_recovery = 0;
 925                 spin_unlock(&dlm->spinlock);
 926
 927                 if (wait_on_recovery)
 928                         dlm_wait_for_node_recovery(dlm, bit, 10000);
 929         }
 930
 931         /* must wait for lock to be mastered elsewhere */
 932         if (blocked)
 933                 goto wait;
 934
 935         ret = -EINVAL;
 936         dlm_node_iter_init(mle->vote_map, &iter);
 937         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
 938                 ret = dlm_do_master_request(res, mle, nodenum);
 939                 if (ret < 0)
 940                         mlog_errno(ret);
 941                 if (mle->master != O2NM_MAX_NODES) {
 942                         /* found a master ! */
 943                         if (mle->master <= nodenum)
 944                                 break;
 945                         /* if our master request has not reached the master
 946                          * yet, keep going until it does.  this is how the
 947                          * master will know that asserts are needed back to
 948                          * the lower nodes. */
 949                         mlog(0, "%s: res %.*s, Requests only up to %u but "
 950                              "master is %u, keep going\n", dlm->name, namelen,
 951                              lockid, nodenum, mle->master);
 952                 }
 953         }
 954
 955 wait:
 956         /* keep going until the response map includes all nodes */
 957         ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
 958         if (ret < 0) {
 959                 wait_on_recovery = 1;
 960                 mlog(0, "%s: res %.*s, Node map changed, redo the master "
 961                      "request now, blocked=%d\n", dlm->name, res->lockname.len,
 962                      res->lockname.name, blocked);
 963                 if (++tries > 20) {
 964                         mlog(ML_ERROR, "%s: res %.*s, Spinning on "
 965                              "dlm_wait_for_lock_mastery, blocked = %d\n",
 966                              dlm->name, res->lockname.len,
 967                              res->lockname.name, blocked);
 968                         dlm_print_one_lock_resource(res);
 969                         dlm_print_one_mle(mle);
 970                         tries = 0;
 971                 }
 972                 goto redo_request;
 973         }
 974
 975         mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len,
 976              res->lockname.name, res->owner);
 977         /* make sure we never continue without this */
 978         BUG_ON(res->owner == O2NM_MAX_NODES);
 979
 980         /* master is known, detach if not already detached */
 981         dlm_mle_detach_hb_events(dlm, mle);
 982         dlm_put_mle(mle);
 983         /* put the extra ref */
 984         dlm_put_mle_inuse(mle);
 985
 986 wake_waiters:
 987         spin_lock(&res->spinlock);
 988         res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
 989         spin_unlock(&res->spinlock);
 990         wake_up(&res->wq);
 991
 992 leave:
 993         /* need to free the unused mle */
 994         if (alloc_mle)
 995                 kmem_cache_free(dlm_mle_cache, alloc_mle);
 996
 997         return res;
 998 }
 999
1000
1001 #define DLM_MASTERY_TIMEOUT_MS   5000
1002
1003 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
1004                                      struct dlm_lock_resource *res,
1005                                      struct dlm_master_list_entry *mle,
1006                                      int *blocked)
1007 {
1008         u8 m;
1009         int ret, bit;
1010         int map_changed, voting_done;
1011         int assert, sleep;
1012
1013 recheck:
1014         ret = 0;
1015         assert = 0;
1016
1017         /* check if another node has already become the owner */
1018         spin_lock(&res->spinlock);
1019         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1020                 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1021                      res->lockname.len, res->lockname.name, res->owner);
1022                 spin_unlock(&res->spinlock);
1023                 /* this will cause the master to re-assert across
1024                  * the whole cluster, freeing up mles */
1025                 if (res->owner != dlm->node_num) {
1026                         ret = dlm_do_master_request(res, mle, res->owner);
1027                         if (ret < 0) {
1028                                 /* give recovery a chance to run */
1029                                 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1030                                 msleep(500);
1031                                 goto recheck;
1032                         }
1033                 }
1034                 ret = 0;
1035                 goto leave;
1036         }
1037         spin_unlock(&res->spinlock);
1038
1039         spin_lock(&mle->spinlock);
1040         m = mle->master;
1041         map_changed = (memcmp(mle->vote_map, mle->node_map,
1042                               sizeof(mle->vote_map)) != 0);
1043         voting_done = (memcmp(mle->vote_map, mle->response_map,
1044                              sizeof(mle->vote_map)) == 0);
1045
1046         /* restart if we hit any errors */
1047         if (map_changed) {
1048                 int b;
1049                 mlog(0, "%s: %.*s: node map changed, restarting\n",
1050                      dlm->name, res->lockname.len, res->lockname.name);
1051                 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1052                 b = (mle->type == DLM_MLE_BLOCK);
1053                 if ((*blocked && !b) || (!*blocked && b)) {
1054                         mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1055                              dlm->name, res->lockname.len, res->lockname.name,
1056                              *blocked, b);
1057                         *blocked = b;
1058                 }
1059                 spin_unlock(&mle->spinlock);
1060                 if (ret < 0) {
1061                         mlog_errno(ret);
1062                         goto leave;
1063                 }
1064                 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1065                      "rechecking now\n", dlm->name, res->lockname.len,
1066                      res->lockname.name);
1067                 goto recheck;
1068         } else {
1069                 if (!voting_done) {
1070                         mlog(0, "map not changed and voting not done "
1071                              "for %s:%.*s\n", dlm->name, res->lockname.len,
1072                              res->lockname.name);
1073                 }
1074         }
1075
1076         if (m != O2NM_MAX_NODES) {
1077                 /* another node has done an assert!
1078                  * all done! */
1079                 sleep = 0;
1080         } else {
1081                 sleep = 1;
1082                 /* have all nodes responded? */
1083                 if (voting_done && !*blocked) {
1084                         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1085                         if (dlm->node_num <= bit) {
1086                                 /* my node number is lowest.
1087                                  * now tell other nodes that I am
1088                                  * mastering this. */
1089                                 mle->master = dlm->node_num;
1090                                 /* ref was grabbed in get_lock_resource
1091                                  * will be dropped in dlmlock_master */
1092                                 assert = 1;
1093                                 sleep = 0;
1094                         }
1095                         /* if voting is done, but we have not received
1096                          * an assert master yet, we must sleep */
1097                 }
1098         }
1099
1100         spin_unlock(&mle->spinlock);
1101
1102         /* sleep if we haven't finished voting yet */
1103         if (sleep) {
1104                 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1105                 atomic_set(&mle->woken, 0);
1106                 (void)wait_event_timeout(mle->wq,
1107                                          (atomic_read(&mle->woken) == 1),
1108                                          timeo);
1109                 if (res->owner == O2NM_MAX_NODES) {
1110                         mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1111                              res->lockname.len, res->lockname.name);
1112                         goto recheck;
1113                 }
1114                 mlog(0, "done waiting, master is %u\n", res->owner);
1115                 ret = 0;
1116                 goto leave;
1117         }
1118
1119         ret = 0;   /* done */
1120         if (assert) {
1121                 m = dlm->node_num;
1122                 mlog(0, "about to master %.*s here, this=%u\n",
1123                      res->lockname.len, res->lockname.name, m);
1124                 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1125                 if (ret) {
1126                         /* This is a failure in the network path,
1127                          * not in the response to the assert_master
1128                          * (any nonzero response is a BUG on this node).
1129                          * Most likely a socket just got disconnected
1130                          * due to node death. */
1131                         mlog_errno(ret);
1132                 }
1133                 /* no longer need to restart lock mastery.
1134                  * all living nodes have been contacted. */
1135                 ret = 0;
1136         }
1137
1138         /* set the lockres owner */
1139         spin_lock(&res->spinlock);
1140         /* mastery reference obtained either during
1141          * assert_master_handler or in get_lock_resource */
1142         dlm_change_lockres_owner(dlm, res, m);
1143         spin_unlock(&res->spinlock);
1144
1145 leave:
1146         return ret;
1147 }
1148
1149 struct dlm_bitmap_diff_iter
1150 {
1151         int curnode;
1152         unsigned long *orig_bm;
1153         unsigned long *cur_bm;
1154         unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1155 };
1156
1157 enum dlm_node_state_change
1158 {
1159         NODE_DOWN = -1,
1160         NODE_NO_CHANGE = 0,
1161         NODE_UP
1162 };
1163
1164 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1165                                       unsigned long *orig_bm,
1166                                       unsigned long *cur_bm)
1167 {
1168         unsigned long p1, p2;
1169         int i;
1170
1171         iter->curnode = -1;
1172         iter->orig_bm = orig_bm;
1173         iter->cur_bm = cur_bm;
1174
1175         for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1176                 p1 = *(iter->orig_bm + i);
1177                 p2 = *(iter->cur_bm + i);
1178                 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1179         }
1180 }
1181
1182 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1183                                      enum dlm_node_state_change *state)
1184 {
1185         int bit;
1186
1187         if (iter->curnode >= O2NM_MAX_NODES)
1188                 return -ENOENT;
1189
1190         bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1191                             iter->curnode+1);
1192         if (bit >= O2NM_MAX_NODES) {
1193                 iter->curnode = O2NM_MAX_NODES;
1194                 return -ENOENT;
1195         }
1196
1197         /* if it was there in the original then this node died */
1198         if (test_bit(bit, iter->orig_bm))
1199                 *state = NODE_DOWN;
1200         else
1201                 *state = NODE_UP;
1202
1203         iter->curnode = bit;
1204         return bit;
1205 }
1206
1207
1208 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1209                                     struct dlm_lock_resource *res,
1210                                     struct dlm_master_list_entry *mle,
1211                                     int blocked)
1212 {
1213         struct dlm_bitmap_diff_iter bdi;
1214         enum dlm_node_state_change sc;
1215         int node;
1216         int ret = 0;
1217
1218         mlog(0, "something happened such that the "
1219              "master process may need to be restarted!\n");
1220
1221         assert_spin_locked(&mle->spinlock);
1222
1223         dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1224         node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1225         while (node >= 0) {
1226                 if (sc == NODE_UP) {
1227                         /* a node came up.  clear any old vote from
1228                          * the response map and set it in the vote map
1229                          * then restart the mastery. */
1230                         mlog(ML_NOTICE, "node %d up while restarting\n", node);
1231
1232                         /* redo the master request, but only for the new node */
1233                         mlog(0, "sending request to new node\n");
1234                         clear_bit(node, mle->response_map);
1235                         set_bit(node, mle->vote_map);
1236                 } else {
1237                         mlog(ML_ERROR, "node down! %d\n", node);
1238                         if (blocked) {
1239                                 int lowest = find_next_bit(mle->maybe_map,
1240                                                        O2NM_MAX_NODES, 0);
1241
1242                                 /* act like it was never there */
1243                                 clear_bit(node, mle->maybe_map);
1244
1245                                 if (node == lowest) {
1246                                         mlog(0, "expected master %u died"
1247                                             " while this node was blocked "
1248                                             "waiting on it!\n", node);
1249                                         lowest = find_next_bit(mle->maybe_map,
1250                                                         O2NM_MAX_NODES,
1251                                                         lowest+1);
1252                                         if (lowest < O2NM_MAX_NODES) {
1253                                                 mlog(0, "%s:%.*s:still "
1254                                                      "blocked. waiting on %u "
1255                                                      "now\n", dlm->name,
1256                                                      res->lockname.len,
1257                                                      res->lockname.name,
1258                                                      lowest);
1259                                         } else {
1260                                                 /* mle is an MLE_BLOCK, but
1261                                                  * there is now nothing left to
1262                                                  * block on.  we need to return
1263                                                  * all the way back out and try
1264                                                  * again with an MLE_MASTER.
1265                                                  * dlm_do_local_recovery_cleanup
1266                                                  * has already run, so the mle
1267                                                  * refcount is ok */
1268                                                 mlog(0, "%s:%.*s: no "
1269                                                      "longer blocking. try to "
1270                                                      "master this here\n",
1271                                                      dlm->name,
1272                                                      res->lockname.len,
1273                                                      res->lockname.name);
1274                                                 mle->type = DLM_MLE_MASTER;
1275                                                 mle->mleres = res;
1276                                         }
1277                                 }
1278                         }
1279
1280                         /* now blank out everything, as if we had never
1281                          * contacted anyone */
1282                         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1283                         memset(mle->response_map, 0, sizeof(mle->response_map));
1284                         /* reset the vote_map to the current node_map */
1285                         memcpy(mle->vote_map, mle->node_map,
1286                                sizeof(mle->node_map));
1287                         /* put myself into the maybe map */
1288                         if (mle->type != DLM_MLE_BLOCK)
1289                                 set_bit(dlm->node_num, mle->maybe_map);
1290                 }
1291                 ret = -EAGAIN;
1292                 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1293         }
1294         return ret;
1295 }
1296
1297
1298 /*
1299  * DLM_MASTER_REQUEST_MSG
1300  *
1301  * returns: 0 on success,
1302  *          -errno on a network error
1303  *
1304  * on error, the caller should assume the target node is "dead"
1305  *
1306  */
1307
1308 static int dlm_do_master_request(struct dlm_lock_resource *res,
1309                                  struct dlm_master_list_entry *mle, int to)
1310 {
1311         struct dlm_ctxt *dlm = mle->dlm;
1312         struct dlm_master_request request;
1313         int ret, response=0, resend;
1314
1315         memset(&request, 0, sizeof(request));
1316         request.node_idx = dlm->node_num;
1317
1318         BUG_ON(mle->type == DLM_MLE_MIGRATION);
1319
1320         request.namelen = (u8)mle->mnamelen;
1321         memcpy(request.name, mle->mname, request.namelen);
1322
1323 again:
1324         ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1325                                  sizeof(request), to, &response);
1326         if (ret < 0)  {
1327                 if (ret == -ESRCH) {
1328                         /* should never happen */
1329                         mlog(ML_ERROR, "TCP stack not ready!\n");
1330                         BUG();
1331                 } else if (ret == -EINVAL) {
1332                         mlog(ML_ERROR, "bad args passed to o2net!\n");
1333                         BUG();
1334                 } else if (ret == -ENOMEM) {
1335                         mlog(ML_ERROR, "out of memory while trying to send "
1336                              "network message!  retrying\n");
1337                         /* this is totally crude */
1338                         msleep(50);
1339                         goto again;
1340                 } else if (!dlm_is_host_down(ret)) {
1341                         /* not a network error. bad. */
1342                         mlog_errno(ret);
1343                         mlog(ML_ERROR, "unhandled error!");
1344                         BUG();
1345                 }
1346                 /* all other errors should be network errors,
1347                  * and likely indicate node death */
1348                 mlog(ML_ERROR, "link to %d went down!\n", to);
1349                 goto out;
1350         }
1351
1352         ret = 0;
1353         resend = 0;
1354         spin_lock(&mle->spinlock);
1355         switch (response) {
1356                 case DLM_MASTER_RESP_YES:
1357                         set_bit(to, mle->response_map);
1358                         mlog(0, "node %u is the master, response=YES\n", to);
1359                         mlog(0, "%s:%.*s: master node %u now knows I have a "
1360                              "reference\n", dlm->name, res->lockname.len,
1361                              res->lockname.name, to);
1362                         mle->master = to;
1363                         break;
1364                 case DLM_MASTER_RESP_NO:
1365                         mlog(0, "node %u not master, response=NO\n", to);
1366                         set_bit(to, mle->response_map);
1367                         break;
1368                 case DLM_MASTER_RESP_MAYBE:
1369                         mlog(0, "node %u not master, response=MAYBE\n", to);
1370                         set_bit(to, mle->response_map);
1371                         set_bit(to, mle->maybe_map);
1372                         break;
1373                 case DLM_MASTER_RESP_ERROR:
1374                         mlog(0, "node %u hit an error, resending\n", to);
1375                         resend = 1;
1376                         response = 0;
1377                         break;
1378                 default:
1379                         mlog(ML_ERROR, "bad response! %u\n", response);
1380                         BUG();
1381         }
1382         spin_unlock(&mle->spinlock);
1383         if (resend) {
1384                 /* this is also totally crude */
1385                 msleep(50);
1386                 goto again;
1387         }
1388
1389 out:
1390         return ret;
1391 }
1392
1393 /*
1394  * locks that can be taken here:
1395  * dlm->spinlock
1396  * res->spinlock
1397  * mle->spinlock
1398  * dlm->master_list
1399  *
1400  * if possible, TRIM THIS DOWN!!!
1401  */
1402 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1403                                void **ret_data)
1404 {
1405         u8 response = DLM_MASTER_RESP_MAYBE;
1406         struct dlm_ctxt *dlm = data;
1407         struct dlm_lock_resource *res = NULL;
1408         struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1409         struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1410         char *name;
1411         unsigned int namelen, hash;
1412         int found, ret;
1413         int set_maybe;
1414         int dispatch_assert = 0;
1415         int dispatched = 0;
1416
1417         if (!dlm_grab(dlm))
1418                 return DLM_MASTER_RESP_NO;
1419
1420         if (!dlm_domain_fully_joined(dlm)) {
1421                 response = DLM_MASTER_RESP_NO;
1422                 goto send_response;
1423         }
1424
1425         name = request->name;
1426         namelen = request->namelen;
1427         hash = dlm_lockid_hash(name, namelen);
1428
1429         if (namelen > DLM_LOCKID_NAME_MAX) {
1430                 response = DLM_IVBUFLEN;
1431                 goto send_response;
1432         }
1433
1434 way_up_top:
1435         spin_lock(&dlm->spinlock);
1436         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1437         if (res) {
1438                 spin_unlock(&dlm->spinlock);
1439
1440                 /* take care of the easy cases up front */
1441                 spin_lock(&res->spinlock);
1442
1443                 /*
1444                  * Right after dlm spinlock was released, dlm_thread could have
1445                  * purged the lockres. Check if lockres got unhashed. If so
1446                  * start over.
1447                  */
1448                 if (hlist_unhashed(&res->hash_node)) {
1449                         spin_unlock(&res->spinlock);
1450                         dlm_lockres_put(res);
1451                         goto way_up_top;
1452                 }
1453
1454                 if (res->state & (DLM_LOCK_RES_RECOVERING|
1455                                   DLM_LOCK_RES_MIGRATING)) {
1456                         spin_unlock(&res->spinlock);
1457                         mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1458                              "being recovered/migrated\n");
1459                         response = DLM_MASTER_RESP_ERROR;
1460                         if (mle)
1461                                 kmem_cache_free(dlm_mle_cache, mle);
1462                         goto send_response;
1463                 }
1464
1465                 if (res->owner == dlm->node_num) {
1466                         dlm_lockres_set_refmap_bit(dlm, res, request->node_idx);
1467                         spin_unlock(&res->spinlock);
1468                         response = DLM_MASTER_RESP_YES;
1469                         if (mle)
1470                                 kmem_cache_free(dlm_mle_cache, mle);
1471
1472                         /* this node is the owner.
1473                          * there is some extra work that needs to
1474                          * happen now.  the requesting node has
1475                          * caused all nodes up to this one to
1476                          * create mles.  this node now needs to
1477                          * go back and clean those up. */
1478                         dispatch_assert = 1;
1479                         goto send_response;
1480                 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1481                         spin_unlock(&res->spinlock);
1482                         // mlog(0, "node %u is the master\n", res->owner);
1483                         response = DLM_MASTER_RESP_NO;
1484                         if (mle)
1485                                 kmem_cache_free(dlm_mle_cache, mle);
1486                         goto send_response;
1487                 }
1488
1489                 /* ok, there is no owner.  either this node is
1490                  * being blocked, or it is actively trying to
1491                  * master this lock. */
1492                 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1493                         mlog(ML_ERROR, "lock with no owner should be "
1494                              "in-progress!\n");
1495                         BUG();
1496                 }
1497
1498                 // mlog(0, "lockres is in progress...\n");
1499                 spin_lock(&dlm->master_lock);
1500                 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1501                 if (!found) {
1502                         mlog(ML_ERROR, "no mle found for this lock!\n");
1503                         BUG();
1504                 }
1505                 set_maybe = 1;
1506                 spin_lock(&tmpmle->spinlock);
1507                 if (tmpmle->type == DLM_MLE_BLOCK) {
1508                         // mlog(0, "this node is waiting for "
1509                         // "lockres to be mastered\n");
1510                         response = DLM_MASTER_RESP_NO;
1511                 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1512                         mlog(0, "node %u is master, but trying to migrate to "
1513                              "node %u.\n", tmpmle->master, tmpmle->new_master);
1514                         if (tmpmle->master == dlm->node_num) {
1515                                 mlog(ML_ERROR, "no owner on lockres, but this "
1516                                      "node is trying to migrate it to %u?!\n",
1517                                      tmpmle->new_master);
1518                                 BUG();
1519                         } else {
1520                                 /* the real master can respond on its own */
1521                                 response = DLM_MASTER_RESP_NO;
1522                         }
1523                 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1524                         set_maybe = 0;
1525                         if (tmpmle->master == dlm->node_num) {
1526                                 response = DLM_MASTER_RESP_YES;
1527                                 /* this node will be the owner.
1528                                  * go back and clean the mles on any
1529                                  * other nodes */
1530                                 dispatch_assert = 1;
1531                                 dlm_lockres_set_refmap_bit(dlm, res,
1532                                                            request->node_idx);
1533                         } else
1534                                 response = DLM_MASTER_RESP_NO;
1535                 } else {
1536                         // mlog(0, "this node is attempting to "
1537                         // "master lockres\n");
1538                         response = DLM_MASTER_RESP_MAYBE;
1539                 }
1540                 if (set_maybe)
1541                         set_bit(request->node_idx, tmpmle->maybe_map);
1542                 spin_unlock(&tmpmle->spinlock);
1543
1544                 spin_unlock(&dlm->master_lock);
1545                 spin_unlock(&res->spinlock);
1546
1547                 /* keep the mle attached to heartbeat events */
1548                 dlm_put_mle(tmpmle);
1549                 if (mle)
1550                         kmem_cache_free(dlm_mle_cache, mle);
1551                 goto send_response;
1552         }
1553
1554         /*
1555          * lockres doesn't exist on this node
1556          * if there is an MLE_BLOCK, return NO
1557          * if there is an MLE_MASTER, return MAYBE
1558          * otherwise, add an MLE_BLOCK, return NO
1559          */
1560         spin_lock(&dlm->master_lock);
1561         found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1562         if (!found) {
1563                 /* this lockid has never been seen on this node yet */
1564                 // mlog(0, "no mle found\n");
1565                 if (!mle) {
1566                         spin_unlock(&dlm->master_lock);
1567                         spin_unlock(&dlm->spinlock);
1568
1569                         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1570                         if (!mle) {
1571                                 response = DLM_MASTER_RESP_ERROR;
1572                                 mlog_errno(-ENOMEM);
1573                                 goto send_response;
1574                         }
1575                         goto way_up_top;
1576                 }
1577
1578                 // mlog(0, "this is second time thru, already allocated, "
1579                 // "add the block.\n");
1580                 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1581                 set_bit(request->node_idx, mle->maybe_map);
1582                 __dlm_insert_mle(dlm, mle);
1583                 response = DLM_MASTER_RESP_NO;
1584         } else {
1585                 spin_lock(&tmpmle->spinlock);
1586                 if (tmpmle->master == dlm->node_num) {
1587                         mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1588                         BUG();
1589                 }
1590                 if (tmpmle->type == DLM_MLE_BLOCK)
1591                         response = DLM_MASTER_RESP_NO;
1592                 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1593                         mlog(0, "migration mle was found (%u->%u)\n",
1594                              tmpmle->master, tmpmle->new_master);
1595                         /* real master can respond on its own */
1596                         response = DLM_MASTER_RESP_NO;
1597                 } else
1598                         response = DLM_MASTER_RESP_MAYBE;
1599                 set_bit(request->node_idx, tmpmle->maybe_map);
1600                 spin_unlock(&tmpmle->spinlock);
1601         }
1602         spin_unlock(&dlm->master_lock);
1603         spin_unlock(&dlm->spinlock);
1604
1605         if (found) {
1606                 /* keep the mle attached to heartbeat events */
1607                 dlm_put_mle(tmpmle);
1608         }
1609 send_response:
1610         /*
1611          * __dlm_lookup_lockres() grabbed a reference to this lockres.
1612          * The reference is released by dlm_assert_master_worker() under
1613          * the call to dlm_dispatch_assert_master().  If
1614          * dlm_assert_master_worker() isn't called, we drop it here.
1615          */
1616         if (dispatch_assert) {
1617                 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1618                              dlm->node_num, res->lockname.len, res->lockname.name);
1619                 spin_lock(&res->spinlock);
1620                 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1621                                                  DLM_ASSERT_MASTER_MLE_CLEANUP);
1622                 if (ret < 0) {
1623                         mlog(ML_ERROR, "failed to dispatch assert master work\n");
1624                         response = DLM_MASTER_RESP_ERROR;
1625                         spin_unlock(&res->spinlock);
1626                         dlm_lockres_put(res);
1627                 } else {
1628                         dispatched = 1;
1629                         __dlm_lockres_grab_inflight_worker(dlm, res);
1630                         spin_unlock(&res->spinlock);
1631                 }
1632         } else {
1633                 if (res)
1634                         dlm_lockres_put(res);
1635         }
1636
1637         if (!dispatched)
1638                 dlm_put(dlm);
1639         return response;
1640 }
1641
1642 /*
1643  * DLM_ASSERT_MASTER_MSG
1644  */
1645
1646
1647 /*
1648  * NOTE: this can be used for debugging
1649  * can periodically run all locks owned by this node
1650  * and re-assert across the cluster...
1651  */
1652 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1653                                 struct dlm_lock_resource *res,
1654                                 void *nodemap, u32 flags)
1655 {
1656         struct dlm_assert_master assert;
1657         int to, tmpret;
1658         struct dlm_node_iter iter;
1659         int ret = 0;
1660         int reassert;
1661         const char *lockname = res->lockname.name;
1662         unsigned int namelen = res->lockname.len;
1663
1664         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1665
1666         spin_lock(&res->spinlock);
1667         res->state |= DLM_LOCK_RES_SETREF_INPROG;
1668         spin_unlock(&res->spinlock);
1669
1670 again:
1671         reassert = 0;
1672
1673         /* note that if this nodemap is empty, it returns 0 */
1674         dlm_node_iter_init(nodemap, &iter);
1675         while ((to = dlm_node_iter_next(&iter)) >= 0) {
1676                 int r = 0;
1677                 struct dlm_master_list_entry *mle = NULL;
1678
1679                 mlog(0, "sending assert master to %d (%.*s)\n", to,
1680                      namelen, lockname);
1681                 memset(&assert, 0, sizeof(assert));
1682                 assert.node_idx = dlm->node_num;
1683                 assert.namelen = namelen;
1684                 memcpy(assert.name, lockname, namelen);
1685                 assert.flags = cpu_to_be32(flags);
1686
1687                 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1688                                             &assert, sizeof(assert), to, &r);
1689                 if (tmpret < 0) {
1690                         mlog(ML_ERROR, "Error %d when sending message %u (key "
1691                              "0x%x) to node %u\n", tmpret,
1692                              DLM_ASSERT_MASTER_MSG, dlm->key, to);
1693                         if (!dlm_is_host_down(tmpret)) {
1694                                 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1695                                 BUG();
1696                         }
1697                         /* a node died.  finish out the rest of the nodes. */
1698                         mlog(0, "link to %d went down!\n", to);
1699                         /* any nonzero status return will do */
1700                         ret = tmpret;
1701                         r = 0;
1702                 } else if (r < 0) {
1703                         /* ok, something horribly messed.  kill thyself. */
1704                         mlog(ML_ERROR,"during assert master of %.*s to %u, "
1705                              "got %d.\n", namelen, lockname, to, r);
1706                         spin_lock(&dlm->spinlock);
1707                         spin_lock(&dlm->master_lock);
1708                         if (dlm_find_mle(dlm, &mle, (char *)lockname,
1709                                          namelen)) {
1710                                 dlm_print_one_mle(mle);
1711                                 __dlm_put_mle(mle);
1712                         }
1713                         spin_unlock(&dlm->master_lock);
1714                         spin_unlock(&dlm->spinlock);
1715                         BUG();
1716                 }
1717
1718                 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1719                     !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1720                                 mlog(ML_ERROR, "%.*s: very strange, "
1721                                      "master MLE but no lockres on %u\n",
1722                                      namelen, lockname, to);
1723                 }
1724
1725                 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1726                         mlog(0, "%.*s: node %u create mles on other "
1727                              "nodes and requests a re-assert\n",
1728                              namelen, lockname, to);
1729                         reassert = 1;
1730                 }
1731                 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1732                         mlog(0, "%.*s: node %u has a reference to this "
1733                              "lockres, set the bit in the refmap\n",
1734                              namelen, lockname, to);
1735                         spin_lock(&res->spinlock);
1736                         dlm_lockres_set_refmap_bit(dlm, res, to);
1737                         spin_unlock(&res->spinlock);
1738                 }
1739         }
1740
1741         if (reassert)
1742                 goto again;
1743
1744         spin_lock(&res->spinlock);
1745         res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1746         spin_unlock(&res->spinlock);
1747         wake_up(&res->wq);
1748
1749         return ret;
1750 }
1751
1752 /*
1753  * locks that can be taken here:
1754  * dlm->spinlock
1755  * res->spinlock
1756  * mle->spinlock
1757  * dlm->master_list
1758  *
1759  * if possible, TRIM THIS DOWN!!!
1760  */
1761 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1762                               void **ret_data)
1763 {
1764         struct dlm_ctxt *dlm = data;
1765         struct dlm_master_list_entry *mle = NULL;
1766         struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1767         struct dlm_lock_resource *res = NULL;
1768         char *name;
1769         unsigned int namelen, hash;
1770         u32 flags;
1771         int master_request = 0, have_lockres_ref = 0;
1772         int ret = 0;
1773
1774         if (!dlm_grab(dlm))
1775                 return 0;
1776
1777         name = assert->name;
1778         namelen = assert->namelen;
1779         hash = dlm_lockid_hash(name, namelen);
1780         flags = be32_to_cpu(assert->flags);
1781
1782         if (namelen > DLM_LOCKID_NAME_MAX) {
1783                 mlog(ML_ERROR, "Invalid name length!");
1784                 goto done;
1785         }
1786
1787         spin_lock(&dlm->spinlock);
1788
1789         if (flags)
1790                 mlog(0, "assert_master with flags: %u\n", flags);
1791
1792         /* find the MLE */
1793         spin_lock(&dlm->master_lock);
1794         if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1795                 /* not an error, could be master just re-asserting */
1796                 mlog(0, "just got an assert_master from %u, but no "
1797                      "MLE for it! (%.*s)\n", assert->node_idx,
1798                      namelen, name);
1799         } else {
1800                 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1801                 if (bit >= O2NM_MAX_NODES) {
1802                         /* not necessarily an error, though less likely.
1803                          * could be master just re-asserting. */
1804                         mlog(0, "no bits set in the maybe_map, but %u "
1805                              "is asserting! (%.*s)\n", assert->node_idx,
1806                              namelen, name);
1807                 } else if (bit != assert->node_idx) {
1808                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1809                                 mlog(0, "master %u was found, %u should "
1810                                      "back off\n", assert->node_idx, bit);
1811                         } else {
1812                                 /* with the fix for bug 569, a higher node
1813                                  * number winning the mastery will respond
1814                                  * YES to mastery requests, but this node
1815                                  * had no way of knowing.  let it pass. */
1816                                 mlog(0, "%u is the lowest node, "
1817                                      "%u is asserting. (%.*s)  %u must "
1818                                      "have begun after %u won.\n", bit,
1819                                      assert->node_idx, namelen, name, bit,
1820                                      assert->node_idx);
1821                         }
1822                 }
1823                 if (mle->type == DLM_MLE_MIGRATION) {
1824                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1825                                 mlog(0, "%s:%.*s: got cleanup assert"
1826                                      " from %u for migration\n",
1827                                      dlm->name, namelen, name,
1828                                      assert->node_idx);
1829                         } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1830                                 mlog(0, "%s:%.*s: got unrelated assert"
1831                                      " from %u for migration, ignoring\n",
1832                                      dlm->name, namelen, name,
1833                                      assert->node_idx);
1834                                 __dlm_put_mle(mle);
1835                                 spin_unlock(&dlm->master_lock);
1836                                 spin_unlock(&dlm->spinlock);
1837                                 goto done;
1838                         }
1839                 }
1840         }
1841         spin_unlock(&dlm->master_lock);
1842
1843         /* ok everything checks out with the MLE
1844          * now check to see if there is a lockres */
1845         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1846         if (res) {
1847                 spin_lock(&res->spinlock);
1848                 if (res->state & DLM_LOCK_RES_RECOVERING)  {
1849                         mlog(ML_ERROR, "%u asserting but %.*s is "
1850                              "RECOVERING!\n", assert->node_idx, namelen, name);
1851                         goto kill;
1852                 }
1853                 if (!mle) {
1854                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1855                             res->owner != assert->node_idx) {
1856                                 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1857                                      "but current owner is %u! (%.*s)\n",
1858                                      assert->node_idx, res->owner, namelen,
1859                                      name);
1860                                 __dlm_print_one_lock_resource(res);
1861                                 BUG();
1862                         }
1863                 } else if (mle->type != DLM_MLE_MIGRATION) {
1864                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1865                                 /* owner is just re-asserting */
1866                                 if (res->owner == assert->node_idx) {
1867                                         mlog(0, "owner %u re-asserting on "
1868                                              "lock %.*s\n", assert->node_idx,
1869                                              namelen, name);
1870                                         goto ok;
1871                                 }
1872                                 mlog(ML_ERROR, "got assert_master from "
1873                                      "node %u, but %u is the owner! "
1874                                      "(%.*s)\n", assert->node_idx,
1875                                      res->owner, namelen, name);
1876                                 goto kill;
1877                         }
1878                         if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1879                                 mlog(ML_ERROR, "got assert from %u, but lock "
1880                                      "with no owner should be "
1881                                      "in-progress! (%.*s)\n",
1882                                      assert->node_idx,
1883                                      namelen, name);
1884                                 goto kill;
1885                         }
1886                 } else /* mle->type == DLM_MLE_MIGRATION */ {
1887                         /* should only be getting an assert from new master */
1888                         if (assert->node_idx != mle->new_master) {
1889                                 mlog(ML_ERROR, "got assert from %u, but "
1890                                      "new master is %u, and old master "
1891                                      "was %u (%.*s)\n",
1892                                      assert->node_idx, mle->new_master,
1893                                      mle->master, namelen, name);
1894                                 goto kill;
1895                         }
1896
1897                 }
1898 ok:
1899                 spin_unlock(&res->spinlock);
1900         }
1901
1902         // mlog(0, "woo!  got an assert_master from node %u!\n",
1903         //           assert->node_idx);
1904         if (mle) {
1905                 int extra_ref = 0;
1906                 int nn = -1;
1907                 int rr, err = 0;
1908
1909                 spin_lock(&mle->spinlock);
1910                 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1911                         extra_ref = 1;
1912                 else {
1913                         /* MASTER mle: if any bits set in the response map
1914                          * then the calling node needs to re-assert to clear
1915                          * up nodes that this node contacted */
1916                         while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1917                                                     nn+1)) < O2NM_MAX_NODES) {
1918                                 if (nn != dlm->node_num && nn != assert->node_idx) {
1919                                         master_request = 1;
1920                                         break;
1921                                 }
1922                         }
1923                 }
1924                 mle->master = assert->node_idx;
1925                 atomic_set(&mle->woken, 1);
1926                 wake_up(&mle->wq);
1927                 spin_unlock(&mle->spinlock);
1928
1929                 if (res) {
1930                         int wake = 0;
1931                         spin_lock(&res->spinlock);
1932                         if (mle->type == DLM_MLE_MIGRATION) {
1933                                 mlog(0, "finishing off migration of lockres %.*s, "
1934                                         "from %u to %u\n",
1935                                         res->lockname.len, res->lockname.name,
1936                                         dlm->node_num, mle->new_master);
1937                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
1938                                 wake = 1;
1939                                 dlm_change_lockres_owner(dlm, res, mle->new_master);
1940                                 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1941                         } else {
1942                                 dlm_change_lockres_owner(dlm, res, mle->master);
1943                         }
1944                         spin_unlock(&res->spinlock);
1945                         have_lockres_ref = 1;
1946                         if (wake)
1947                                 wake_up(&res->wq);
1948                 }
1949
1950                 /* master is known, detach if not already detached.
1951                  * ensures that only one assert_master call will happen
1952                  * on this mle. */
1953                 spin_lock(&dlm->master_lock);
1954
1955                 rr = kref_read(&mle->mle_refs);
1956                 if (mle->inuse > 0) {
1957                         if (extra_ref && rr < 3)
1958                                 err = 1;
1959                         else if (!extra_ref && rr < 2)
1960                                 err = 1;
1961                 } else {
1962                         if (extra_ref && rr < 2)
1963                                 err = 1;
1964                         else if (!extra_ref && rr < 1)
1965                                 err = 1;
1966                 }
1967                 if (err) {
1968                         mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
1969                              "that will mess up this node, refs=%d, extra=%d, "
1970                              "inuse=%d\n", dlm->name, namelen, name,
1971                              assert->node_idx, rr, extra_ref, mle->inuse);
1972                         dlm_print_one_mle(mle);
1973                 }
1974                 __dlm_unlink_mle(dlm, mle);
1975                 __dlm_mle_detach_hb_events(dlm, mle);
1976                 __dlm_put_mle(mle);
1977                 if (extra_ref) {
1978                         /* the assert master message now balances the extra
1979                          * ref given by the master / migration request message.
1980                          * if this is the last put, it will be removed
1981                          * from the list. */
1982                         __dlm_put_mle(mle);
1983                 }
1984                 spin_unlock(&dlm->master_lock);
1985         } else if (res) {
1986                 if (res->owner != assert->node_idx) {
1987                         mlog(0, "assert_master from %u, but current "
1988                              "owner is %u (%.*s), no mle\n", assert->node_idx,
1989                              res->owner, namelen, name);
1990                 }
1991         }
1992         spin_unlock(&dlm->spinlock);
1993
1994 done:
1995         ret = 0;
1996         if (res) {
1997                 spin_lock(&res->spinlock);
1998                 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1999                 spin_unlock(&res->spinlock);
2000                 *ret_data = (void *)res;
2001         }
2002         dlm_put(dlm);
2003         if (master_request) {
2004                 mlog(0, "need to tell master to reassert\n");
2005                 /* positive. negative would shoot down the node. */
2006                 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2007                 if (!have_lockres_ref) {
2008                         mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2009                              "mle present here for %s:%.*s, but no lockres!\n",
2010                              assert->node_idx, dlm->name, namelen, name);
2011                 }
2012         }
2013         if (have_lockres_ref) {
2014                 /* let the master know we have a reference to the lockres */
2015                 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2016                 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2017                      dlm->name, namelen, name, assert->node_idx);
2018         }
2019         return ret;
2020
2021 kill:
2022         /* kill the caller! */
2023         mlog(ML_ERROR, "Bad message received from another node.  Dumping state "
2024              "and killing the other node now!  This node is OK and can continue.\n");
2025         __dlm_print_one_lock_resource(res);
2026         spin_unlock(&res->spinlock);
2027         spin_lock(&dlm->master_lock);
2028         if (mle)
2029                 __dlm_put_mle(mle);
2030         spin_unlock(&dlm->master_lock);
2031         spin_unlock(&dlm->spinlock);
2032         *ret_data = (void *)res;
2033         dlm_put(dlm);
2034         return -EINVAL;
2035 }
2036
2037 void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2038 {
2039         struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2040
2041         if (ret_data) {
2042                 spin_lock(&res->spinlock);
2043                 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2044                 spin_unlock(&res->spinlock);
2045                 wake_up(&res->wq);
2046                 dlm_lockres_put(res);
2047         }
2048         return;
2049 }
2050
2051 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2052                                struct dlm_lock_resource *res,
2053                                int ignore_higher, u8 request_from, u32 flags)
2054 {
2055         struct dlm_work_item *item;
2056         item = kzalloc(sizeof(*item), GFP_ATOMIC);
2057         if (!item)
2058                 return -ENOMEM;
2059
2060
2061         /* queue up work for dlm_assert_master_worker */
2062         dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2063         item->u.am.lockres = res; /* already have a ref */
2064         /* can optionally ignore node numbers higher than this node */
2065         item->u.am.ignore_higher = ignore_higher;
2066         item->u.am.request_from = request_from;
2067         item->u.am.flags = flags;
2068
2069         if (ignore_higher)
2070                 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2071                      res->lockname.name);
2072
2073         spin_lock(&dlm->work_lock);
2074         list_add_tail(&item->list, &dlm->work_list);
2075         spin_unlock(&dlm->work_lock);
2076
2077         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2078         return 0;
2079 }
2080
2081 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2082 {
2083         struct dlm_ctxt *dlm = data;
2084         int ret = 0;
2085         struct dlm_lock_resource *res;
2086         unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2087         int ignore_higher;
2088         int bit;
2089         u8 request_from;
2090         u32 flags;
2091
2092         dlm = item->dlm;
2093         res = item->u.am.lockres;
2094         ignore_higher = item->u.am.ignore_higher;
2095         request_from = item->u.am.request_from;
2096         flags = item->u.am.flags;
2097
2098         spin_lock(&dlm->spinlock);
2099         memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2100         spin_unlock(&dlm->spinlock);
2101
2102         clear_bit(dlm->node_num, nodemap);
2103         if (ignore_higher) {
2104                 /* if is this just to clear up mles for nodes below
2105                  * this node, do not send the message to the original
2106                  * caller or any node number higher than this */
2107                 clear_bit(request_from, nodemap);
2108                 bit = dlm->node_num;
2109                 while (1) {
2110                         bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2111                                             bit+1);
2112                         if (bit >= O2NM_MAX_NODES)
2113                                 break;
2114                         clear_bit(bit, nodemap);
2115                 }
2116         }
2117
2118         /*
2119          * If we're migrating this lock to someone else, we are no
2120          * longer allowed to assert out own mastery.  OTOH, we need to
2121          * prevent migration from starting while we're still asserting
2122          * our dominance.  The reserved ast delays migration.
2123          */
2124         spin_lock(&res->spinlock);
2125         if (res->state & DLM_LOCK_RES_MIGRATING) {
2126                 mlog(0, "Someone asked us to assert mastery, but we're "
2127                      "in the middle of migration.  Skipping assert, "
2128                      "the new master will handle that.\n");
2129                 spin_unlock(&res->spinlock);
2130                 goto put;
2131         } else
2132                 __dlm_lockres_reserve_ast(res);
2133         spin_unlock(&res->spinlock);
2134
2135         /* this call now finishes out the nodemap
2136          * even if one or more nodes die */
2137         mlog(0, "worker about to master %.*s here, this=%u\n",
2138                      res->lockname.len, res->lockname.name, dlm->node_num);
2139         ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2140         if (ret < 0) {
2141                 /* no need to restart, we are done */
2142                 if (!dlm_is_host_down(ret))
2143                         mlog_errno(ret);
2144         }
2145
2146         /* Ok, we've asserted ourselves.  Let's let migration start. */
2147         dlm_lockres_release_ast(dlm, res);
2148
2149 put:
2150         dlm_lockres_drop_inflight_worker(dlm, res);
2151
2152         dlm_lockres_put(res);
2153
2154         mlog(0, "finished with dlm_assert_master_worker\n");
2155 }
2156
2157 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2158  * We cannot wait for node recovery to complete to begin mastering this
2159  * lockres because this lockres is used to kick off recovery! ;-)
2160  * So, do a pre-check on all living nodes to see if any of those nodes
2161  * think that $RECOVERY is currently mastered by a dead node.  If so,
2162  * we wait a short time to allow that node to get notified by its own
2163  * heartbeat stack, then check again.  All $RECOVERY lock resources
2164  * mastered by dead nodes are purged when the heartbeat callback is
2165  * fired, so we can know for sure that it is safe to continue once
2166  * the node returns a live node or no node.  */
2167 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2168                                        struct dlm_lock_resource *res)
2169 {
2170         struct dlm_node_iter iter;
2171         int nodenum;
2172         int ret = 0;
2173         u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2174
2175         spin_lock(&dlm->spinlock);
2176         dlm_node_iter_init(dlm->domain_map, &iter);
2177         spin_unlock(&dlm->spinlock);
2178
2179         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2180                 /* do not send to self */
2181                 if (nodenum == dlm->node_num)
2182                         continue;
2183                 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2184                 if (ret < 0) {
2185                         mlog_errno(ret);
2186                         if (!dlm_is_host_down(ret))
2187                                 BUG();
2188                         /* host is down, so answer for that node would be
2189                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
2190                         ret = 0;
2191                 }
2192
2193                 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2194                         /* check to see if this master is in the recovery map */
2195                         spin_lock(&dlm->spinlock);
2196                         if (test_bit(master, dlm->recovery_map)) {
2197                                 mlog(ML_NOTICE, "%s: node %u has not seen "
2198                                      "node %u go down yet, and thinks the "
2199                                      "dead node is mastering the recovery "
2200                                      "lock.  must wait.\n", dlm->name,
2201                                      nodenum, master);
2202                                 ret = -EAGAIN;
2203                         }
2204                         spin_unlock(&dlm->spinlock);
2205                         mlog(0, "%s: reco lock master is %u\n", dlm->name,
2206                              master);
2207                         break;
2208                 }
2209         }
2210         return ret;
2211 }
2212
2213 /*
2214  * DLM_DEREF_LOCKRES_MSG
2215  */
2216
2217 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2218 {
2219         struct dlm_deref_lockres deref;
2220         int ret = 0, r;
2221         const char *lockname;
2222         unsigned int namelen;
2223
2224         lockname = res->lockname.name;
2225         namelen = res->lockname.len;
2226         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2227
2228         memset(&deref, 0, sizeof(deref));
2229         deref.node_idx = dlm->node_num;
2230         deref.namelen = namelen;
2231         memcpy(deref.name, lockname, namelen);
2232
2233         ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2234                                  &deref, sizeof(deref), res->owner, &r);
2235         if (ret < 0)
2236                 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n",
2237                      dlm->name, namelen, lockname, ret, res->owner);
2238         else if (r < 0) {
2239                 /* BAD.  other node says I did not have a ref. */
2240                 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2241                      dlm->name, namelen, lockname, res->owner, r);
2242                 dlm_print_one_lock_resource(res);
2243                 if (r == -ENOMEM)
2244                         BUG();
2245         } else
2246                 ret = r;
2247
2248         return ret;
2249 }
2250
2251 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2252                               void **ret_data)
2253 {
2254         struct dlm_ctxt *dlm = data;
2255         struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2256         struct dlm_lock_resource *res = NULL;
2257         char *name;
2258         unsigned int namelen;
2259         int ret = -EINVAL;
2260         u8 node;
2261         unsigned int hash;
2262         struct dlm_work_item *item;
2263         int cleared = 0;
2264         int dispatch = 0;
2265
2266         if (!dlm_grab(dlm))
2267                 return 0;
2268
2269         name = deref->name;
2270         namelen = deref->namelen;
2271         node = deref->node_idx;
2272
2273         if (namelen > DLM_LOCKID_NAME_MAX) {
2274                 mlog(ML_ERROR, "Invalid name length!");
2275                 goto done;
2276         }
2277         if (deref->node_idx >= O2NM_MAX_NODES) {
2278                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2279                 goto done;
2280         }
2281
2282         hash = dlm_lockid_hash(name, namelen);
2283
2284         spin_lock(&dlm->spinlock);
2285         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2286         if (!res) {
2287                 spin_unlock(&dlm->spinlock);
2288                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2289                      dlm->name, namelen, name);
2290                 goto done;
2291         }
2292         spin_unlock(&dlm->spinlock);
2293
2294         spin_lock(&res->spinlock);
2295         if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2296                 dispatch = 1;
2297         else {
2298                 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2299                 if (test_bit(node, res->refmap)) {
2300                         dlm_lockres_clear_refmap_bit(dlm, res, node);
2301                         cleared = 1;
2302                 }
2303         }
2304         spin_unlock(&res->spinlock);
2305
2306         if (!dispatch) {
2307                 if (cleared)
2308                         dlm_lockres_calc_usage(dlm, res);
2309                 else {
2310                         mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2311                         "but it is already dropped!\n", dlm->name,
2312                         res->lockname.len, res->lockname.name, node);
2313                         dlm_print_one_lock_resource(res);
2314                 }
2315                 ret = DLM_DEREF_RESPONSE_DONE;
2316                 goto done;
2317         }
2318
2319         item = kzalloc(sizeof(*item), GFP_NOFS);
2320         if (!item) {
2321                 ret = -ENOMEM;
2322                 mlog_errno(ret);
2323                 goto done;
2324         }
2325
2326         dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2327         item->u.dl.deref_res = res;
2328         item->u.dl.deref_node = node;
2329
2330         spin_lock(&dlm->work_lock);
2331         list_add_tail(&item->list, &dlm->work_list);
2332         spin_unlock(&dlm->work_lock);
2333
2334         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2335         return DLM_DEREF_RESPONSE_INPROG;
2336
2337 done:
2338         if (res)
2339                 dlm_lockres_put(res);
2340         dlm_put(dlm);
2341
2342         return ret;
2343 }
2344
2345 int dlm_deref_lockres_done_handler(struct o2net_msg *msg, u32 len, void *data,
2346                               void **ret_data)
2347 {
2348         struct dlm_ctxt *dlm = data;
2349         struct dlm_deref_lockres_done *deref
2350                         = (struct dlm_deref_lockres_done *)msg->buf;
2351         struct dlm_lock_resource *res = NULL;
2352         char *name;
2353         unsigned int namelen;
2354         int ret = -EINVAL;
2355         u8 node;
2356         unsigned int hash;
2357
2358         if (!dlm_grab(dlm))
2359                 return 0;
2360
2361         name = deref->name;
2362         namelen = deref->namelen;
2363         node = deref->node_idx;
2364
2365         if (namelen > DLM_LOCKID_NAME_MAX) {
2366                 mlog(ML_ERROR, "Invalid name length!");
2367                 goto done;
2368         }
2369         if (deref->node_idx >= O2NM_MAX_NODES) {
2370                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2371                 goto done;
2372         }
2373
2374         hash = dlm_lockid_hash(name, namelen);
2375
2376         spin_lock(&dlm->spinlock);
2377         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2378         if (!res) {
2379                 spin_unlock(&dlm->spinlock);
2380                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2381                      dlm->name, namelen, name);
2382                 goto done;
2383         }
2384
2385         spin_lock(&res->spinlock);
2386         if (!(res->state & DLM_LOCK_RES_DROPPING_REF)) {
2387                 spin_unlock(&res->spinlock);
2388                 spin_unlock(&dlm->spinlock);
2389                 mlog(ML_NOTICE, "%s:%.*s: node %u sends deref done "
2390                         "but it is already derefed!\n", dlm->name,
2391                         res->lockname.len, res->lockname.name, node);
2392                 ret = 0;
2393                 goto done;
2394         }
2395
2396         __dlm_do_purge_lockres(dlm, res);
2397         spin_unlock(&res->spinlock);
2398         wake_up(&res->wq);
2399
2400         spin_unlock(&dlm->spinlock);
2401
2402         ret = 0;
2403 done:
2404         if (res)
2405                 dlm_lockres_put(res);
2406         dlm_put(dlm);
2407         return ret;
2408 }
2409
2410 static void dlm_drop_lockres_ref_done(struct dlm_ctxt *dlm,
2411                 struct dlm_lock_resource *res, u8 node)
2412 {
2413         struct dlm_deref_lockres_done deref;
2414         int ret = 0, r;
2415         const char *lockname;
2416         unsigned int namelen;
2417
2418         lockname = res->lockname.name;
2419         namelen = res->lockname.len;
2420         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2421
2422         memset(&deref, 0, sizeof(deref));
2423         deref.node_idx = dlm->node_num;
2424         deref.namelen = namelen;
2425         memcpy(deref.name, lockname, namelen);
2426
2427         ret = o2net_send_message(DLM_DEREF_LOCKRES_DONE, dlm->key,
2428                                  &deref, sizeof(deref), node, &r);
2429         if (ret < 0) {
2430                 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF DONE "
2431                                 " to node %u\n", dlm->name, namelen,
2432                                 lockname, ret, node);
2433         } else if (r < 0) {
2434                 /* ignore the error */
2435                 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2436                      dlm->name, namelen, lockname, node, r);
2437                 dlm_print_one_lock_resource(res);
2438         }
2439 }
2440
2441 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2442 {
2443         struct dlm_ctxt *dlm;
2444         struct dlm_lock_resource *res;
2445         u8 node;
2446         u8 cleared = 0;
2447
2448         dlm = item->dlm;
2449         res = item->u.dl.deref_res;
2450         node = item->u.dl.deref_node;
2451
2452         spin_lock(&res->spinlock);
2453         BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2454         __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2455         if (test_bit(node, res->refmap)) {
2456                 dlm_lockres_clear_refmap_bit(dlm, res, node);
2457                 cleared = 1;
2458         }
2459         spin_unlock(&res->spinlock);
2460
2461         dlm_drop_lockres_ref_done(dlm, res, node);
2462
2463         if (cleared) {
2464                 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2465                      dlm->name, res->lockname.len, res->lockname.name, node);
2466                 dlm_lockres_calc_usage(dlm, res);
2467         } else {
2468                 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2469                      "but it is already dropped!\n", dlm->name,
2470                      res->lockname.len, res->lockname.name, node);
2471                 dlm_print_one_lock_resource(res);
2472         }
2473
2474         dlm_lockres_put(res);
2475 }
2476
2477 /*
2478  * A migratable resource is one that is :
2479  * 1. locally mastered, and,
2480  * 2. zero local locks, and,
2481  * 3. one or more non-local locks, or, one or more references
2482  * Returns 1 if yes, 0 if not.
2483  */
2484 static int dlm_is_lockres_migratable(struct dlm_ctxt *dlm,
2485                                       struct dlm_lock_resource *res)
2486 {
2487         enum dlm_lockres_list idx;
2488         int nonlocal = 0, node_ref;
2489         struct list_head *queue;
2490         struct dlm_lock *lock;
2491         u64 cookie;
2492
2493         assert_spin_locked(&res->spinlock);
2494
2495         /* delay migration when the lockres is in MIGRATING state */
2496         if (res->state & DLM_LOCK_RES_MIGRATING)
2497                 return 0;
2498
2499         /* delay migration when the lockres is in RECOCERING state */
2500         if (res->state & (DLM_LOCK_RES_RECOVERING|
2501                         DLM_LOCK_RES_RECOVERY_WAITING))
2502                 return 0;
2503
2504         if (res->owner != dlm->node_num)
2505                 return 0;
2506
2507         for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2508                 queue = dlm_list_idx_to_ptr(res, idx);
2509                 list_for_each_entry(lock, queue, list) {
2510                         if (lock->ml.node != dlm->node_num) {
2511                                 nonlocal++;
2512                                 continue;
2513                         }
2514                         cookie = be64_to_cpu(lock->ml.cookie);
2515                         mlog(0, "%s: Not migratable res %.*s, lock %u:%llu on "
2516                              "%s list\n", dlm->name, res->lockname.len,
2517                              res->lockname.name,
2518                              dlm_get_lock_cookie_node(cookie),
2519                              dlm_get_lock_cookie_seq(cookie),
2520                              dlm_list_in_text(idx));
2521                         return 0;
2522                 }
2523         }
2524
2525         if (!nonlocal) {
2526                 node_ref = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
2527                 if (node_ref >= O2NM_MAX_NODES)
2528                         return 0;
2529         }
2530
2531         mlog(0, "%s: res %.*s, Migratable\n", dlm->name, res->lockname.len,
2532              res->lockname.name);
2533
2534         return 1;
2535 }
2536
2537 /*
2538  * DLM_MIGRATE_LOCKRES
2539  */
2540
2541
2542 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2543                                struct dlm_lock_resource *res, u8 target)
2544 {
2545         struct dlm_master_list_entry *mle = NULL;
2546         struct dlm_master_list_entry *oldmle = NULL;
2547         struct dlm_migratable_lockres *mres = NULL;
2548         int ret = 0;
2549         const char *name;
2550         unsigned int namelen;
2551         int mle_added = 0;
2552         int wake = 0;
2553
2554         if (!dlm_grab(dlm))
2555                 return -EINVAL;
2556
2557         name = res->lockname.name;
2558         namelen = res->lockname.len;
2559
2560         mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name,
2561              target);
2562
2563         /* preallocate up front. if this fails, abort */
2564         ret = -ENOMEM;
2565         mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2566         if (!mres) {
2567                 mlog_errno(ret);
2568                 goto leave;
2569         }
2570
2571         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2572         if (!mle) {
2573                 mlog_errno(ret);
2574                 goto leave;
2575         }
2576         ret = 0;
2577
2578         /*
2579          * clear any existing master requests and
2580          * add the migration mle to the list
2581          */
2582         spin_lock(&dlm->spinlock);
2583         spin_lock(&dlm->master_lock);
2584         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2585                                     namelen, target, dlm->node_num);
2586         /* get an extra reference on the mle.
2587          * otherwise the assert_master from the new
2588          * master will destroy this.
2589          */
2590         if (ret != -EEXIST)
2591                 dlm_get_mle_inuse(mle);
2592
2593         spin_unlock(&dlm->master_lock);
2594         spin_unlock(&dlm->spinlock);
2595
2596         if (ret == -EEXIST) {
2597                 mlog(0, "another process is already migrating it\n");
2598                 goto fail;
2599         }
2600         mle_added = 1;
2601
2602         /*
2603          * set the MIGRATING flag and flush asts
2604          * if we fail after this we need to re-dirty the lockres
2605          */
2606         if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2607                 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2608                      "the target went down.\n", res->lockname.len,
2609                      res->lockname.name, target);
2610                 spin_lock(&res->spinlock);
2611                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2612                 wake = 1;
2613                 spin_unlock(&res->spinlock);
2614                 ret = -EINVAL;
2615         }
2616
2617 fail:
2618         if (ret != -EEXIST && oldmle) {
2619                 /* master is known, detach if not already detached */
2620                 dlm_mle_detach_hb_events(dlm, oldmle);
2621                 dlm_put_mle(oldmle);
2622         }
2623
2624         if (ret < 0) {
2625                 if (mle_added) {
2626                         dlm_mle_detach_hb_events(dlm, mle);
2627                         dlm_put_mle(mle);
2628                         dlm_put_mle_inuse(mle);
2629                 } else if (mle) {
2630                         kmem_cache_free(dlm_mle_cache, mle);
2631                         mle = NULL;
2632                 }
2633                 goto leave;
2634         }
2635
2636         /*
2637          * at this point, we have a migration target, an mle
2638          * in the master list, and the MIGRATING flag set on
2639          * the lockres
2640          */
2641
2642         /* now that remote nodes are spinning on the MIGRATING flag,
2643          * ensure that all assert_master work is flushed. */
2644         flush_workqueue(dlm->dlm_worker);
2645
2646         /* notify new node and send all lock state */
2647         /* call send_one_lockres with migration flag.
2648          * this serves as notice to the target node that a
2649          * migration is starting. */
2650         ret = dlm_send_one_lockres(dlm, res, mres, target,
2651                                    DLM_MRES_MIGRATION);
2652
2653         if (ret < 0) {
2654                 mlog(0, "migration to node %u failed with %d\n",
2655                      target, ret);
2656                 /* migration failed, detach and clean up mle */
2657                 dlm_mle_detach_hb_events(dlm, mle);
2658                 dlm_put_mle(mle);
2659                 dlm_put_mle_inuse(mle);
2660                 spin_lock(&res->spinlock);
2661                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2662                 wake = 1;
2663                 spin_unlock(&res->spinlock);
2664                 if (dlm_is_host_down(ret))
2665                         dlm_wait_for_node_death(dlm, target,
2666                                                 DLM_NODE_DEATH_WAIT_MAX);
2667                 goto leave;
2668         }
2669
2670         /* at this point, the target sends a message to all nodes,
2671          * (using dlm_do_migrate_request).  this node is skipped since
2672          * we had to put an mle in the list to begin the process.  this
2673          * node now waits for target to do an assert master.  this node
2674          * will be the last one notified, ensuring that the migration
2675          * is complete everywhere.  if the target dies while this is
2676          * going on, some nodes could potentially see the target as the
2677          * master, so it is important that my recovery finds the migration
2678          * mle and sets the master to UNKNOWN. */
2679
2680
2681         /* wait for new node to assert master */
2682         while (1) {
2683                 ret = wait_event_interruptible_timeout(mle->wq,
2684                                         (atomic_read(&mle->woken) == 1),
2685                                         msecs_to_jiffies(5000));
2686
2687                 if (ret >= 0) {
2688                         if (atomic_read(&mle->woken) == 1 ||
2689                             res->owner == target)
2690                                 break;
2691
2692                         mlog(0, "%s:%.*s: timed out during migration\n",
2693                              dlm->name, res->lockname.len, res->lockname.name);
2694                         /* avoid hang during shutdown when migrating lockres
2695                          * to a node which also goes down */
2696                         if (dlm_is_node_dead(dlm, target)) {
2697                                 mlog(0, "%s:%.*s: expected migration "
2698                                      "target %u is no longer up, restarting\n",
2699                                      dlm->name, res->lockname.len,
2700                                      res->lockname.name, target);
2701                                 ret = -EINVAL;
2702                                 /* migration failed, detach and clean up mle */
2703                                 dlm_mle_detach_hb_events(dlm, mle);
2704                                 dlm_put_mle(mle);
2705                                 dlm_put_mle_inuse(mle);
2706                                 spin_lock(&res->spinlock);
2707                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2708                                 wake = 1;
2709                                 spin_unlock(&res->spinlock);
2710                                 goto leave;
2711                         }
2712                 } else
2713                         mlog(0, "%s:%.*s: caught signal during migration\n",
2714                              dlm->name, res->lockname.len, res->lockname.name);
2715         }
2716
2717         /* all done, set the owner, clear the flag */
2718         spin_lock(&res->spinlock);
2719         dlm_set_lockres_owner(dlm, res, target);
2720         res->state &= ~DLM_LOCK_RES_MIGRATING;
2721         dlm_remove_nonlocal_locks(dlm, res);
2722         spin_unlock(&res->spinlock);
2723         wake_up(&res->wq);
2724
2725         /* master is known, detach if not already detached */
2726         dlm_mle_detach_hb_events(dlm, mle);
2727         dlm_put_mle_inuse(mle);
2728         ret = 0;
2729
2730         dlm_lockres_calc_usage(dlm, res);
2731
2732 leave:
2733         /* re-dirty the lockres if we failed */
2734         if (ret < 0)
2735                 dlm_kick_thread(dlm, res);
2736
2737         /* wake up waiters if the MIGRATING flag got set
2738          * but migration failed */
2739         if (wake)
2740                 wake_up(&res->wq);
2741
2742         if (mres)
2743                 free_page((unsigned long)mres);
2744
2745         dlm_put(dlm);
2746
2747         mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen,
2748              name, target, ret);
2749         return ret;
2750 }
2751
2752 /*
2753  * Should be called only after beginning the domain leave process.
2754  * There should not be any remaining locks on nonlocal lock resources,
2755  * and there should be no local locks left on locally mastered resources.
2756  *
2757  * Called with the dlm spinlock held, may drop it to do migration, but
2758  * will re-acquire before exit.
2759  *
2760  * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2761  */
2762 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2763         __must_hold(&dlm->spinlock)
2764 {
2765         int ret;
2766         int lock_dropped = 0;
2767         u8 target = O2NM_MAX_NODES;
2768
2769         assert_spin_locked(&dlm->spinlock);
2770
2771         spin_lock(&res->spinlock);
2772         if (dlm_is_lockres_migratable(dlm, res))
2773                 target = dlm_pick_migration_target(dlm, res);
2774         spin_unlock(&res->spinlock);
2775
2776         if (target == O2NM_MAX_NODES)
2777                 goto leave;
2778
2779         /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2780         spin_unlock(&dlm->spinlock);
2781         lock_dropped = 1;
2782         ret = dlm_migrate_lockres(dlm, res, target);
2783         if (ret)
2784                 mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2785                      dlm->name, res->lockname.len, res->lockname.name,
2786                      target, ret);
2787         spin_lock(&dlm->spinlock);
2788 leave:
2789         return lock_dropped;
2790 }
2791
2792 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2793 {
2794         int ret;
2795         spin_lock(&dlm->ast_lock);
2796         spin_lock(&lock->spinlock);
2797         ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2798         spin_unlock(&lock->spinlock);
2799         spin_unlock(&dlm->ast_lock);
2800         return ret;
2801 }
2802
2803 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2804                                      struct dlm_lock_resource *res,
2805                                      u8 mig_target)
2806 {
2807         int can_proceed;
2808         spin_lock(&res->spinlock);
2809         can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2810         spin_unlock(&res->spinlock);
2811
2812         /* target has died, so make the caller break out of the
2813          * wait_event, but caller must recheck the domain_map */
2814         spin_lock(&dlm->spinlock);
2815         if (!test_bit(mig_target, dlm->domain_map))
2816                 can_proceed = 1;
2817         spin_unlock(&dlm->spinlock);
2818         return can_proceed;
2819 }
2820
2821 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2822                                 struct dlm_lock_resource *res)
2823 {
2824         int ret;
2825         spin_lock(&res->spinlock);
2826         ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2827         spin_unlock(&res->spinlock);
2828         return ret;
2829 }
2830
2831
2832 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2833                                        struct dlm_lock_resource *res,
2834                                        u8 target)
2835 {
2836         int ret = 0;
2837
2838         mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2839                res->lockname.len, res->lockname.name, dlm->node_num,
2840                target);
2841         /* need to set MIGRATING flag on lockres.  this is done by
2842          * ensuring that all asts have been flushed for this lockres. */
2843         spin_lock(&res->spinlock);
2844         BUG_ON(res->migration_pending);
2845         res->migration_pending = 1;
2846         /* strategy is to reserve an extra ast then release
2847          * it below, letting the release do all of the work */
2848         __dlm_lockres_reserve_ast(res);
2849         spin_unlock(&res->spinlock);
2850
2851         /* now flush all the pending asts */
2852         dlm_kick_thread(dlm, res);
2853         /* before waiting on DIRTY, block processes which may
2854          * try to dirty the lockres before MIGRATING is set */
2855         spin_lock(&res->spinlock);
2856         BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2857         res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2858         spin_unlock(&res->spinlock);
2859         /* now wait on any pending asts and the DIRTY state */
2860         wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2861         dlm_lockres_release_ast(dlm, res);
2862
2863         mlog(0, "about to wait on migration_wq, dirty=%s\n",
2864                res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2865         /* if the extra ref we just put was the final one, this
2866          * will pass thru immediately.  otherwise, we need to wait
2867          * for the last ast to finish. */
2868 again:
2869         ret = wait_event_interruptible_timeout(dlm->migration_wq,
2870                    dlm_migration_can_proceed(dlm, res, target),
2871                    msecs_to_jiffies(1000));
2872         if (ret < 0) {
2873                 mlog(0, "woken again: migrating? %s, dead? %s\n",
2874                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2875                        test_bit(target, dlm->domain_map) ? "no":"yes");
2876         } else {
2877                 mlog(0, "all is well: migrating? %s, dead? %s\n",
2878                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2879                        test_bit(target, dlm->domain_map) ? "no":"yes");
2880         }
2881         if (!dlm_migration_can_proceed(dlm, res, target)) {
2882                 mlog(0, "trying again...\n");
2883                 goto again;
2884         }
2885
2886         ret = 0;
2887         /* did the target go down or die? */
2888         spin_lock(&dlm->spinlock);
2889         if (!test_bit(target, dlm->domain_map)) {
2890                 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2891                      target);
2892                 ret = -EHOSTDOWN;
2893         }
2894         spin_unlock(&dlm->spinlock);
2895
2896         /*
2897          * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2898          * another try; otherwise, we are sure the MIGRATING state is there,
2899          * drop the unneeded state which blocked threads trying to DIRTY
2900          */
2901         spin_lock(&res->spinlock);
2902         BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2903         res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2904         if (!ret)
2905                 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2906         else
2907                 res->migration_pending = 0;
2908         spin_unlock(&res->spinlock);
2909
2910         /*
2911          * at this point:
2912          *
2913          *   o the DLM_LOCK_RES_MIGRATING flag is set if target not down
2914          *   o there are no pending asts on this lockres
2915          *   o all processes trying to reserve an ast on this
2916          *     lockres must wait for the MIGRATING flag to clear
2917          */
2918         return ret;
2919 }
2920
2921 /* last step in the migration process.
2922  * original master calls this to free all of the dlm_lock
2923  * structures that used to be for other nodes. */
2924 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2925                                       struct dlm_lock_resource *res)
2926 {
2927         struct list_head *queue = &res->granted;
2928         int i, bit;
2929         struct dlm_lock *lock, *next;
2930
2931         assert_spin_locked(&res->spinlock);
2932
2933         BUG_ON(res->owner == dlm->node_num);
2934
2935         for (i=0; i<3; i++) {
2936                 list_for_each_entry_safe(lock, next, queue, list) {
2937                         if (lock->ml.node != dlm->node_num) {
2938                                 mlog(0, "putting lock for node %u\n",
2939                                      lock->ml.node);
2940                                 /* be extra careful */
2941                                 BUG_ON(!list_empty(&lock->ast_list));
2942                                 BUG_ON(!list_empty(&lock->bast_list));
2943                                 BUG_ON(lock->ast_pending);
2944                                 BUG_ON(lock->bast_pending);
2945                                 dlm_lockres_clear_refmap_bit(dlm, res,
2946                                                              lock->ml.node);
2947                                 list_del_init(&lock->list);
2948                                 dlm_lock_put(lock);
2949                                 /* In a normal unlock, we would have added a
2950                                  * DLM_UNLOCK_FREE_LOCK action. Force it. */
2951                                 dlm_lock_put(lock);
2952                         }
2953                 }
2954                 queue++;
2955         }
2956         bit = 0;
2957         while (1) {
2958                 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2959                 if (bit >= O2NM_MAX_NODES)
2960                         break;
2961                 /* do not clear the local node reference, if there is a
2962                  * process holding this, let it drop the ref itself */
2963                 if (bit != dlm->node_num) {
2964                         mlog(0, "%s:%.*s: node %u had a ref to this "
2965                              "migrating lockres, clearing\n", dlm->name,
2966                              res->lockname.len, res->lockname.name, bit);
2967                         dlm_lockres_clear_refmap_bit(dlm, res, bit);
2968                 }
2969                 bit++;
2970         }
2971 }
2972
2973 /*
2974  * Pick a node to migrate the lock resource to. This function selects a
2975  * potential target based first on the locks and then on refmap. It skips
2976  * nodes that are in the process of exiting the domain.
2977  */
2978 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2979                                     struct dlm_lock_resource *res)
2980 {
2981         enum dlm_lockres_list idx;
2982         struct list_head *queue = &res->granted;
2983         struct dlm_lock *lock;
2984         int noderef;
2985         u8 nodenum = O2NM_MAX_NODES;
2986
2987         assert_spin_locked(&dlm->spinlock);
2988         assert_spin_locked(&res->spinlock);
2989
2990         /* Go through all the locks */
2991         for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2992                 queue = dlm_list_idx_to_ptr(res, idx);
2993                 list_for_each_entry(lock, queue, list) {
2994                         if (lock->ml.node == dlm->node_num)
2995                                 continue;
2996                         if (test_bit(lock->ml.node, dlm->exit_domain_map))
2997                                 continue;
2998                         nodenum = lock->ml.node;
2999                         goto bail;
3000                 }
3001         }
3002
3003         /* Go thru the refmap */
3004         noderef = -1;
3005         while (1) {
3006                 noderef = find_next_bit(res->refmap, O2NM_MAX_NODES,
3007                                         noderef + 1);
3008                 if (noderef >= O2NM_MAX_NODES)
3009                         break;
3010                 if (noderef == dlm->node_num)
3011                         continue;
3012                 if (test_bit(noderef, dlm->exit_domain_map))
3013                         continue;
3014                 nodenum = noderef;
3015                 goto bail;
3016         }
3017
3018 bail:
3019         return nodenum;
3020 }
3021
3022 /* this is called by the new master once all lockres
3023  * data has been received */
3024 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
3025                                   struct dlm_lock_resource *res,
3026                                   u8 master, u8 new_master,
3027                                   struct dlm_node_iter *iter)
3028 {
3029         struct dlm_migrate_request migrate;
3030         int ret, skip, status = 0;
3031         int nodenum;
3032
3033         memset(&migrate, 0, sizeof(migrate));
3034         migrate.namelen = res->lockname.len;
3035         memcpy(migrate.name, res->lockname.name, migrate.namelen);
3036         migrate.new_master = new_master;
3037         migrate.master = master;
3038
3039         ret = 0;
3040
3041         /* send message to all nodes, except the master and myself */
3042         while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
3043                 if (nodenum == master ||
3044                     nodenum == new_master)
3045                         continue;
3046
3047                 /* We could race exit domain. If exited, skip. */
3048                 spin_lock(&dlm->spinlock);
3049                 skip = (!test_bit(nodenum, dlm->domain_map));
3050                 spin_unlock(&dlm->spinlock);
3051                 if (skip) {
3052                         clear_bit(nodenum, iter->node_map);
3053                         continue;
3054                 }
3055
3056                 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
3057                                          &migrate, sizeof(migrate), nodenum,
3058                                          &status);
3059                 if (ret < 0) {
3060                         mlog(ML_ERROR, "%s: res %.*s, Error %d send "
3061                              "MIGRATE_REQUEST to node %u\n", dlm->name,
3062                              migrate.namelen, migrate.name, ret, nodenum);
3063                         if (!dlm_is_host_down(ret)) {
3064                                 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
3065                                 BUG();
3066                         }
3067                         clear_bit(nodenum, iter->node_map);
3068                         ret = 0;
3069                 } else if (status < 0) {
3070                         mlog(0, "migrate request (node %u) returned %d!\n",
3071                              nodenum, status);
3072                         ret = status;
3073                 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
3074                         /* during the migration request we short-circuited
3075                          * the mastery of the lockres.  make sure we have
3076                          * a mastery ref for nodenum */
3077                         mlog(0, "%s:%.*s: need ref for node %u\n",
3078                              dlm->name, res->lockname.len, res->lockname.name,
3079                              nodenum);
3080                         spin_lock(&res->spinlock);
3081                         dlm_lockres_set_refmap_bit(dlm, res, nodenum);
3082                         spin_unlock(&res->spinlock);
3083                 }
3084         }
3085
3086         if (ret < 0)
3087                 mlog_errno(ret);
3088
3089         mlog(0, "returning ret=%d\n", ret);
3090         return ret;
3091 }
3092
3093
3094 /* if there is an existing mle for this lockres, we now know who the master is.
3095  * (the one who sent us *this* message) we can clear it up right away.
3096  * since the process that put the mle on the list still has a reference to it,
3097  * we can unhash it now, set the master and wake the process.  as a result,
3098  * we will have no mle in the list to start with.  now we can add an mle for
3099  * the migration and this should be the only one found for those scanning the
3100  * list.  */
3101 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
3102                                 void **ret_data)
3103 {
3104         struct dlm_ctxt *dlm = data;
3105         struct dlm_lock_resource *res = NULL;
3106         struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
3107         struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
3108         const char *name;
3109         unsigned int namelen, hash;
3110         int ret = 0;
3111
3112         if (!dlm_grab(dlm))
3113                 return 0;
3114
3115         name = migrate->name;
3116         namelen = migrate->namelen;
3117         hash = dlm_lockid_hash(name, namelen);
3118
3119         /* preallocate.. if this fails, abort */
3120         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
3121
3122         if (!mle) {
3123                 ret = -ENOMEM;
3124                 goto leave;
3125         }
3126
3127         /* check for pre-existing lock */
3128         spin_lock(&dlm->spinlock);
3129         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
3130         if (res) {
3131                 spin_lock(&res->spinlock);
3132                 if (res->state & DLM_LOCK_RES_RECOVERING) {
3133                         /* if all is working ok, this can only mean that we got
3134                         * a migrate request from a node that we now see as
3135                         * dead.  what can we do here?  drop it to the floor? */
3136                         spin_unlock(&res->spinlock);
3137                         mlog(ML_ERROR, "Got a migrate request, but the "
3138                              "lockres is marked as recovering!");
3139                         kmem_cache_free(dlm_mle_cache, mle);
3140                         ret = -EINVAL; /* need a better solution */
3141                         goto unlock;
3142                 }
3143                 res->state |= DLM_LOCK_RES_MIGRATING;
3144                 spin_unlock(&res->spinlock);
3145         }
3146
3147         spin_lock(&dlm->master_lock);
3148         /* ignore status.  only nonzero status would BUG. */
3149         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3150                                     name, namelen,
3151                                     migrate->new_master,
3152                                     migrate->master);
3153
3154         if (ret < 0)
3155                 kmem_cache_free(dlm_mle_cache, mle);
3156
3157         spin_unlock(&dlm->master_lock);
3158 unlock:
3159         spin_unlock(&dlm->spinlock);
3160
3161         if (oldmle) {
3162                 /* master is known, detach if not already detached */
3163                 dlm_mle_detach_hb_events(dlm, oldmle);
3164                 dlm_put_mle(oldmle);
3165         }
3166
3167         if (res)
3168                 dlm_lockres_put(res);
3169 leave:
3170         dlm_put(dlm);
3171         return ret;
3172 }
3173
3174 /* must be holding dlm->spinlock and dlm->master_lock
3175  * when adding a migration mle, we can clear any other mles
3176  * in the master list because we know with certainty that
3177  * the master is "master".  so we remove any old mle from
3178  * the list after setting it's master field, and then add
3179  * the new migration mle.  this way we can hold with the rule
3180  * of having only one mle for a given lock name at all times. */
3181 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3182                                  struct dlm_lock_resource *res,
3183                                  struct dlm_master_list_entry *mle,
3184                                  struct dlm_master_list_entry **oldmle,
3185                                  const char *name, unsigned int namelen,
3186                                  u8 new_master, u8 master)
3187 {
3188         int found;
3189         int ret = 0;
3190
3191         *oldmle = NULL;
3192
3193         assert_spin_locked(&dlm->spinlock);
3194         assert_spin_locked(&dlm->master_lock);
3195
3196         /* caller is responsible for any ref taken here on oldmle */
3197         found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3198         if (found) {
3199                 struct dlm_master_list_entry *tmp = *oldmle;
3200                 spin_lock(&tmp->spinlock);
3201                 if (tmp->type == DLM_MLE_MIGRATION) {
3202                         if (master == dlm->node_num) {
3203                                 /* ah another process raced me to it */
3204                                 mlog(0, "tried to migrate %.*s, but some "
3205                                      "process beat me to it\n",
3206                                      namelen, name);
3207                                 spin_unlock(&tmp->spinlock);
3208                                 return -EEXIST;
3209                         } else {
3210                                 /* bad.  2 NODES are trying to migrate! */
3211                                 mlog(ML_ERROR, "migration error  mle: "
3212                                      "master=%u new_master=%u // request: "
3213                                      "master=%u new_master=%u // "
3214                                      "lockres=%.*s\n",
3215                                      tmp->master, tmp->new_master,
3216                                      master, new_master,
3217                                      namelen, name);
3218                                 BUG();
3219                         }
3220                 } else {
3221                         /* this is essentially what assert_master does */
3222                         tmp->master = master;
3223                         atomic_set(&tmp->woken, 1);
3224                         wake_up(&tmp->wq);
3225                         /* remove it so that only one mle will be found */
3226                         __dlm_unlink_mle(dlm, tmp);
3227                         __dlm_mle_detach_hb_events(dlm, tmp);
3228                         if (tmp->type == DLM_MLE_MASTER) {
3229                                 ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3230                                 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3231                                                 "telling master to get ref "
3232                                                 "for cleared out mle during "
3233                                                 "migration\n", dlm->name,
3234                                                 namelen, name, master,
3235                                                 new_master);
3236                         }
3237                 }
3238                 spin_unlock(&tmp->spinlock);
3239         }
3240
3241         /* now add a migration mle to the tail of the list */
3242         dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3243         mle->new_master = new_master;
3244         /* the new master will be sending an assert master for this.
3245          * at that point we will get the refmap reference */
3246         mle->master = master;
3247         /* do this for consistency with other mle types */
3248         set_bit(new_master, mle->maybe_map);
3249         __dlm_insert_mle(dlm, mle);
3250
3251         return ret;
3252 }
3253
3254 /*
3255  * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3256  */
3257 static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3258                                         struct dlm_master_list_entry *mle)
3259 {
3260         struct dlm_lock_resource *res;
3261
3262         /* Find the lockres associated to the mle and set its owner to UNK */
3263         res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3264                                    mle->mnamehash);
3265         if (res) {
3266                 spin_unlock(&dlm->master_lock);
3267
3268                 /* move lockres onto recovery list */
3269                 spin_lock(&res->spinlock);
3270                 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3271                 dlm_move_lockres_to_recovery_list(dlm, res);
3272                 spin_unlock(&res->spinlock);
3273                 dlm_lockres_put(res);
3274
3275                 /* about to get rid of mle, detach from heartbeat */
3276                 __dlm_mle_detach_hb_events(dlm, mle);
3277
3278                 /* dump the mle */
3279                 spin_lock(&dlm->master_lock);
3280                 __dlm_put_mle(mle);
3281                 spin_unlock(&dlm->master_lock);
3282         }
3283
3284         return res;
3285 }
3286
3287 static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3288                                     struct dlm_master_list_entry *mle)
3289 {
3290         __dlm_mle_detach_hb_events(dlm, mle);
3291
3292         spin_lock(&mle->spinlock);
3293         __dlm_unlink_mle(dlm, mle);
3294         atomic_set(&mle->woken, 1);
3295         spin_unlock(&mle->spinlock);
3296
3297         wake_up(&mle->wq);
3298 }
3299
3300 static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3301                                 struct dlm_master_list_entry *mle, u8 dead_node)
3302 {
3303         int bit;
3304
3305         BUG_ON(mle->type != DLM_MLE_BLOCK);
3306
3307         spin_lock(&mle->spinlock);
3308         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3309         if (bit != dead_node) {
3310                 mlog(0, "mle found, but dead node %u would not have been "
3311                      "master\n", dead_node);
3312                 spin_unlock(&mle->spinlock);
3313         } else {
3314                 /* Must drop the refcount by one since the assert_master will
3315                  * never arrive. This may result in the mle being unlinked and
3316                  * freed, but there may still be a process waiting in the
3317                  * dlmlock path which is fine. */
3318                 mlog(0, "node %u was expected master\n", dead_node);
3319                 atomic_set(&mle->woken, 1);
3320                 spin_unlock(&mle->spinlock);
3321                 wake_up(&mle->wq);
3322
3323                 /* Do not need events any longer, so detach from heartbeat */
3324                 __dlm_mle_detach_hb_events(dlm, mle);
3325                 __dlm_put_mle(mle);
3326         }
3327 }
3328
3329 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3330 {
3331         struct dlm_master_list_entry *mle;
3332         struct dlm_lock_resource *res;
3333         struct hlist_head *bucket;
3334         struct hlist_node *tmp;
3335         unsigned int i;
3336
3337         mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
3338 top:
3339         assert_spin_locked(&dlm->spinlock);
3340
3341         /* clean the master list */
3342         spin_lock(&dlm->master_lock);
3343         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3344                 bucket = dlm_master_hash(dlm, i);
3345                 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3346                         BUG_ON(mle->type != DLM_MLE_BLOCK &&
3347                                mle->type != DLM_MLE_MASTER &&
3348                                mle->type != DLM_MLE_MIGRATION);
3349
3350                         /* MASTER mles are initiated locally. The waiting
3351                          * process will notice the node map change shortly.
3352                          * Let that happen as normal. */
3353                         if (mle->type == DLM_MLE_MASTER)
3354                                 continue;
3355
3356                         /* BLOCK mles are initiated by other nodes. Need to
3357                          * clean up if the dead node would have been the
3358                          * master. */
3359                         if (mle->type == DLM_MLE_BLOCK) {
3360                                 dlm_clean_block_mle(dlm, mle, dead_node);
3361                                 continue;
3362                         }
3363
3364                         /* Everything else is a MIGRATION mle */
3365
3366                         /* The rule for MIGRATION mles is that the master
3367                          * becomes UNKNOWN if *either* the original or the new
3368                          * master dies. All UNKNOWN lockres' are sent to
3369                          * whichever node becomes the recovery master. The new
3370                          * master is responsible for determining if there is
3371                          * still a master for this lockres, or if he needs to
3372                          * take over mastery. Either way, this node should
3373                          * expect another message to resolve this. */
3374
3375                         if (mle->master != dead_node &&
3376                             mle->new_master != dead_node)
3377                                 continue;
3378
3379                         if (mle->new_master == dead_node && mle->inuse) {
3380                                 mlog(ML_NOTICE, "%s: target %u died during "
3381                                                 "migration from %u, the MLE is "
3382                                                 "still keep used, ignore it!\n",
3383                                                 dlm->name, dead_node,
3384                                                 mle->master);
3385                                 continue;
3386                         }
3387
3388                         /* If we have reached this point, this mle needs to be
3389                          * removed from the list and freed. */
3390                         dlm_clean_migration_mle(dlm, mle);
3391
3392                         mlog(0, "%s: node %u died during migration from "
3393                              "%u to %u!\n", dlm->name, dead_node, mle->master,
3394                              mle->new_master);
3395
3396                         /* If we find a lockres associated with the mle, we've
3397                          * hit this rare case that messes up our lock ordering.
3398                          * If so, we need to drop the master lock so that we can
3399                          * take the lockres lock, meaning that we will have to
3400                          * restart from the head of list. */
3401                         res = dlm_reset_mleres_owner(dlm, mle);
3402                         if (res)
3403                                 /* restart */
3404                                 goto top;
3405
3406                         /* This may be the last reference */
3407                         __dlm_put_mle(mle);
3408                 }
3409         }
3410         spin_unlock(&dlm->master_lock);
3411 }
3412
3413 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3414                          u8 old_master)
3415 {
3416         struct dlm_node_iter iter;
3417         int ret = 0;
3418
3419         spin_lock(&dlm->spinlock);
3420         dlm_node_iter_init(dlm->domain_map, &iter);
3421         clear_bit(old_master, iter.node_map);
3422         clear_bit(dlm->node_num, iter.node_map);
3423         spin_unlock(&dlm->spinlock);
3424
3425         /* ownership of the lockres is changing.  account for the
3426          * mastery reference here since old_master will briefly have
3427          * a reference after the migration completes */
3428         spin_lock(&res->spinlock);
3429         dlm_lockres_set_refmap_bit(dlm, res, old_master);
3430         spin_unlock(&res->spinlock);
3431
3432         mlog(0, "now time to do a migrate request to other nodes\n");
3433         ret = dlm_do_migrate_request(dlm, res, old_master,
3434                                      dlm->node_num, &iter);
3435         if (ret < 0) {
3436                 mlog_errno(ret);
3437                 goto leave;
3438         }
3439
3440         mlog(0, "doing assert master of %.*s to all except the original node\n",
3441              res->lockname.len, res->lockname.name);
3442         /* this call now finishes out the nodemap
3443          * even if one or more nodes die */
3444         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3445                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3446         if (ret < 0) {
3447                 /* no longer need to retry.  all living nodes contacted. */
3448                 mlog_errno(ret);
3449                 ret = 0;
3450         }
3451
3452         memset(iter.node_map, 0, sizeof(iter.node_map));
3453         set_bit(old_master, iter.node_map);
3454         mlog(0, "doing assert master of %.*s back to %u\n",
3455              res->lockname.len, res->lockname.name, old_master);
3456         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3457                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3458         if (ret < 0) {
3459                 mlog(0, "assert master to original master failed "
3460                      "with %d.\n", ret);
3461                 /* the only nonzero status here would be because of
3462                  * a dead original node.  we're done. */
3463                 ret = 0;
3464         }
3465
3466         /* all done, set the owner, clear the flag */
3467         spin_lock(&res->spinlock);
3468         dlm_set_lockres_owner(dlm, res, dlm->node_num);
3469         res->state &= ~DLM_LOCK_RES_MIGRATING;
3470         spin_unlock(&res->spinlock);
3471         /* re-dirty it on the new master */
3472         dlm_kick_thread(dlm, res);
3473         wake_up(&res->wq);
3474 leave:
3475         return ret;
3476 }
3477
3478 /*
3479  * LOCKRES AST REFCOUNT
3480  * this is integral to migration
3481  */
3482
3483 /* for future intent to call an ast, reserve one ahead of time.
3484  * this should be called only after waiting on the lockres
3485  * with dlm_wait_on_lockres, and while still holding the
3486  * spinlock after the call. */
3487 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3488 {
3489         assert_spin_locked(&res->spinlock);
3490         if (res->state & DLM_LOCK_RES_MIGRATING) {
3491                 __dlm_print_one_lock_resource(res);
3492         }
3493         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3494
3495         atomic_inc(&res->asts_reserved);
3496 }
3497
3498 /*
3499  * used to drop the reserved ast, either because it went unused,
3500  * or because the ast/bast was actually called.
3501  *
3502  * also, if there is a pending migration on this lockres,
3503  * and this was the last pending ast on the lockres,
3504  * atomically set the MIGRATING flag before we drop the lock.
3505  * this is how we ensure that migration can proceed with no
3506  * asts in progress.  note that it is ok if the state of the
3507  * queues is such that a lock should be granted in the future
3508  * or that a bast should be fired, because the new master will
3509  * shuffle the lists on this lockres as soon as it is migrated.
3510  */
3511 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3512                              struct dlm_lock_resource *res)
3513 {
3514         if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3515                 return;
3516
3517         if (!res->migration_pending) {
3518                 spin_unlock(&res->spinlock);
3519                 return;
3520         }
3521
3522         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3523         res->migration_pending = 0;
3524         res->state |= DLM_LOCK_RES_MIGRATING;
3525         spin_unlock(&res->spinlock);
3526         wake_up(&res->wq);
3527         wake_up(&dlm->migration_wq);
3528 }
3529
3530 void dlm_force_free_mles(struct dlm_ctxt *dlm)
3531 {
3532         int i;
3533         struct hlist_head *bucket;
3534         struct dlm_master_list_entry *mle;
3535         struct hlist_node *tmp;
3536
3537         /*
3538          * We notified all other nodes that we are exiting the domain and
3539          * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3540          * around we force free them and wake any processes that are waiting
3541          * on the mles
3542          */
3543         spin_lock(&dlm->spinlock);
3544         spin_lock(&dlm->master_lock);
3545
3546         BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3547         BUG_ON((find_next_bit(dlm->domain_map, O2NM_MAX_NODES, 0) < O2NM_MAX_NODES));
3548
3549         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3550                 bucket = dlm_master_hash(dlm, i);
3551                 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3552                         if (mle->type != DLM_MLE_BLOCK) {
3553                                 mlog(ML_ERROR, "bad mle: %p\n", mle);
3554                                 dlm_print_one_mle(mle);
3555                         }
3556                         atomic_set(&mle->woken, 1);
3557                         wake_up(&mle->wq);
3558
3559                         __dlm_unlink_mle(dlm, mle);
3560                         __dlm_mle_detach_hb_events(dlm, mle);
3561                         __dlm_put_mle(mle);
3562                 }
3563         }
3564         spin_unlock(&dlm->master_lock);
3565         spin_unlock(&dlm->spinlock);
3566 }