fs/ocfs2/dlm/dlmmaster.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmmod.c
   5  *
   6  * standalone DLM module
   7  *
   8  * Copyright (C) 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  *
  25  */
  26
  27
  28 #include <linux/module.h>
  29 #include <linux/fs.h>
  30 #include <linux/types.h>
  31 #include <linux/slab.h>
  32 #include <linux/highmem.h>
  33 #include <linux/utsname.h>
  34 #include <linux/init.h>
  35 #include <linux/sysctl.h>
  36 #include <linux/random.h>
  37 #include <linux/blkdev.h>
  38 #include <linux/socket.h>
  39 #include <linux/inet.h>
  40 #include <linux/spinlock.h>
  41 #include <linux/delay.h>
  42
  43
  44 #include "cluster/heartbeat.h"
  45 #include "cluster/nodemanager.h"
  46 #include "cluster/tcp.h"
  47
  48 #include "dlmapi.h"
  49 #include "dlmcommon.h"
  50 #include "dlmdomain.h"
  51
  52 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
  53 #include "cluster/masklog.h"
  54
  55 enum dlm_mle_type {
  56         DLM_MLE_BLOCK,
  57         DLM_MLE_MASTER,
  58         DLM_MLE_MIGRATION
  59 };
  60
  61 struct dlm_lock_name
  62 {
  63         u8 len;
  64         u8 name[DLM_LOCKID_NAME_MAX];
  65 };
  66
  67 struct dlm_master_list_entry
  68 {
  69         struct list_head list;
  70         struct list_head hb_events;
  71         struct dlm_ctxt *dlm;
  72         spinlock_t spinlock;
  73         wait_queue_head_t wq;
  74         atomic_t woken;
  75         struct kref mle_refs;
  76         int inuse;
  77         unsigned long maybe_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  78         unsigned long vote_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  79         unsigned long response_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  80         unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  81         u8 master;
  82         u8 new_master;
  83         enum dlm_mle_type type;
  84         struct o2hb_callback_func mle_hb_up;
  85         struct o2hb_callback_func mle_hb_down;
  86         union {
  87                 struct dlm_lock_resource *res;
  88                 struct dlm_lock_name name;
  89         } u;
  90 };
  91
  92 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
  93                               struct dlm_master_list_entry *mle,
  94                               struct o2nm_node *node,
  95                               int idx);
  96 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
  97                             struct dlm_master_list_entry *mle,
  98                             struct o2nm_node *node,
  99                             int idx);
 100
 101 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
 102 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
 103                                 struct dlm_lock_resource *res,
 104                                 void *nodemap, u32 flags);
 105 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
 106
 107 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
 108                                 struct dlm_master_list_entry *mle,
 109                                 const char *name,
 110                                 unsigned int namelen)
 111 {
 112         struct dlm_lock_resource *res;
 113
 114         if (dlm != mle->dlm)
 115                 return 0;
 116
 117         if (mle->type == DLM_MLE_BLOCK ||
 118             mle->type == DLM_MLE_MIGRATION) {
 119                 if (namelen != mle->u.name.len ||
 120                     memcmp(name, mle->u.name.name, namelen)!=0)
 121                         return 0;
 122         } else {
 123                 res = mle->u.res;
 124                 if (namelen != res->lockname.len ||
 125                     memcmp(res->lockname.name, name, namelen) != 0)
 126                         return 0;
 127         }
 128         return 1;
 129 }
 130
 131 #define dlm_print_nodemap(m)  _dlm_print_nodemap(m,#m)
 132 static void _dlm_print_nodemap(unsigned long *map, const char *mapname)
 133 {
 134         int i;
 135         printk("%s=[ ", mapname);
 136         for (i=0; i<O2NM_MAX_NODES; i++)
 137                 if (test_bit(i, map))
 138                         printk("%d ", i);
 139         printk("]");
 140 }
 141
 142 static void dlm_print_one_mle(struct dlm_master_list_entry *mle)
 143 {
 144         int refs;
 145         char *type;
 146         char attached;
 147         u8 master;
 148         unsigned int namelen;
 149         const char *name;
 150         struct kref *k;
 151         unsigned long *maybe = mle->maybe_map,
 152                       *vote = mle->vote_map,
 153                       *resp = mle->response_map,
 154                       *node = mle->node_map;
 155
 156         k = &mle->mle_refs;
 157         if (mle->type == DLM_MLE_BLOCK)
 158                 type = "BLK";
 159         else if (mle->type == DLM_MLE_MASTER)
 160                 type = "MAS";
 161         else
 162                 type = "MIG";
 163         refs = atomic_read(&k->refcount);
 164         master = mle->master;
 165         attached = (list_empty(&mle->hb_events) ? 'N' : 'Y');
 166
 167         if (mle->type != DLM_MLE_MASTER) {
 168                 namelen = mle->u.name.len;
 169                 name = mle->u.name.name;
 170         } else {
 171                 namelen = mle->u.res->lockname.len;
 172                 name = mle->u.res->lockname.name;
 173         }
 174
 175         mlog(ML_NOTICE, "%.*s: %3s refs=%3d mas=%3u new=%3u evt=%c inuse=%d ",
 176                   namelen, name, type, refs, master, mle->new_master, attached,
 177                   mle->inuse);
 178         dlm_print_nodemap(maybe);
 179         printk(", ");
 180         dlm_print_nodemap(vote);
 181         printk(", ");
 182         dlm_print_nodemap(resp);
 183         printk(", ");
 184         dlm_print_nodemap(node);
 185         printk(", ");
 186         printk("\n");
 187 }
 188
 189 #if 0
 190 /* Code here is included but defined out as it aids debugging */
 191
 192 static void dlm_dump_mles(struct dlm_ctxt *dlm)
 193 {
 194         struct dlm_master_list_entry *mle;
 195
 196         mlog(ML_NOTICE, "dumping all mles for domain %s:\n", dlm->name);
 197         spin_lock(&dlm->master_lock);
 198         list_for_each_entry(mle, &dlm->master_list, list)
 199                 dlm_print_one_mle(mle);
 200         spin_unlock(&dlm->master_lock);
 201 }
 202
 203 int dlm_dump_all_mles(const char __user *data, unsigned int len)
 204 {
 205         struct dlm_ctxt *dlm;
 206
 207         spin_lock(&dlm_domain_lock);
 208         list_for_each_entry(dlm, &dlm_domains, list) {
 209                 mlog(ML_NOTICE, "found dlm: %p, name=%s\n", dlm, dlm->name);
 210                 dlm_dump_mles(dlm);
 211         }
 212         spin_unlock(&dlm_domain_lock);
 213         return len;
 214 }
 215 EXPORT_SYMBOL_GPL(dlm_dump_all_mles);
 216
 217 #endif  /*  0  */
 218
 219 static struct kmem_cache *dlm_lockres_cache = NULL;
 220 static struct kmem_cache *dlm_lockname_cache = NULL;
 221 static struct kmem_cache *dlm_mle_cache = NULL;
 222
 223 static void dlm_mle_release(struct kref *kref);
 224 static void dlm_init_mle(struct dlm_master_list_entry *mle,
 225                         enum dlm_mle_type type,
 226                         struct dlm_ctxt *dlm,
 227                         struct dlm_lock_resource *res,
 228                         const char *name,
 229                         unsigned int namelen);
 230 static void dlm_put_mle(struct dlm_master_list_entry *mle);
 231 static void __dlm_put_mle(struct dlm_master_list_entry *mle);
 232 static int dlm_find_mle(struct dlm_ctxt *dlm,
 233                         struct dlm_master_list_entry **mle,
 234                         char *name, unsigned int namelen);
 235
 236 static int dlm_do_master_request(struct dlm_lock_resource *res,
 237                                  struct dlm_master_list_entry *mle, int to);
 238
 239
 240 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
 241                                      struct dlm_lock_resource *res,
 242                                      struct dlm_master_list_entry *mle,
 243                                      int *blocked);
 244 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
 245                                     struct dlm_lock_resource *res,
 246                                     struct dlm_master_list_entry *mle,
 247                                     int blocked);
 248 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
 249                                  struct dlm_lock_resource *res,
 250                                  struct dlm_master_list_entry *mle,
 251                                  struct dlm_master_list_entry **oldmle,
 252                                  const char *name, unsigned int namelen,
 253                                  u8 new_master, u8 master);
 254
 255 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
 256                                     struct dlm_lock_resource *res);
 257 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
 258                                       struct dlm_lock_resource *res);
 259 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
 260                                        struct dlm_lock_resource *res,
 261                                        u8 target);
 262 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
 263                                        struct dlm_lock_resource *res);
 264
 265
 266 int dlm_is_host_down(int errno)
 267 {
 268         switch (errno) {
 269                 case -EBADF:
 270                 case -ECONNREFUSED:
 271                 case -ENOTCONN:
 272                 case -ECONNRESET:
 273                 case -EPIPE:
 274                 case -EHOSTDOWN:
 275                 case -EHOSTUNREACH:
 276                 case -ETIMEDOUT:
 277                 case -ECONNABORTED:
 278                 case -ENETDOWN:
 279                 case -ENETUNREACH:
 280                 case -ENETRESET:
 281                 case -ESHUTDOWN:
 282                 case -ENOPROTOOPT:
 283                 case -EINVAL:   /* if returned from our tcp code,
 284                                    this means there is no socket */
 285                         return 1;
 286         }
 287         return 0;
 288 }
 289
 290
 291 /*
 292  * MASTER LIST FUNCTIONS
 293  */
 294
 295
 296 /*
 297  * regarding master list entries and heartbeat callbacks:
 298  *
 299  * in order to avoid sleeping and allocation that occurs in
 300  * heartbeat, master list entries are simply attached to the
 301  * dlm's established heartbeat callbacks.  the mle is attached
 302  * when it is created, and since the dlm->spinlock is held at
 303  * that time, any heartbeat event will be properly discovered
 304  * by the mle.  the mle needs to be detached from the
 305  * dlm->mle_hb_events list as soon as heartbeat events are no
 306  * longer useful to the mle, and before the mle is freed.
 307  *
 308  * as a general rule, heartbeat events are no longer needed by
 309  * the mle once an "answer" regarding the lock master has been
 310  * received.
 311  */
 312 static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
 313                                               struct dlm_master_list_entry *mle)
 314 {
 315         assert_spin_locked(&dlm->spinlock);
 316
 317         list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
 318 }
 319
 320
 321 static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 322                                               struct dlm_master_list_entry *mle)
 323 {
 324         if (!list_empty(&mle->hb_events))
 325                 list_del_init(&mle->hb_events);
 326 }
 327
 328
 329 static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
 330                                             struct dlm_master_list_entry *mle)
 331 {
 332         spin_lock(&dlm->spinlock);
 333         __dlm_mle_detach_hb_events(dlm, mle);
 334         spin_unlock(&dlm->spinlock);
 335 }
 336
 337 static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
 338 {
 339         struct dlm_ctxt *dlm;
 340         dlm = mle->dlm;
 341
 342         assert_spin_locked(&dlm->spinlock);
 343         assert_spin_locked(&dlm->master_lock);
 344         mle->inuse++;
 345         kref_get(&mle->mle_refs);
 346 }
 347
 348 static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
 349 {
 350         struct dlm_ctxt *dlm;
 351         dlm = mle->dlm;
 352
 353         spin_lock(&dlm->spinlock);
 354         spin_lock(&dlm->master_lock);
 355         mle->inuse--;
 356         __dlm_put_mle(mle);
 357         spin_unlock(&dlm->master_lock);
 358         spin_unlock(&dlm->spinlock);
 359
 360 }
 361
 362 /* remove from list and free */
 363 static void __dlm_put_mle(struct dlm_master_list_entry *mle)
 364 {
 365         struct dlm_ctxt *dlm;
 366         dlm = mle->dlm;
 367
 368         assert_spin_locked(&dlm->spinlock);
 369         assert_spin_locked(&dlm->master_lock);
 370         if (!atomic_read(&mle->mle_refs.refcount)) {
 371                 /* this may or may not crash, but who cares.
 372                  * it's a BUG. */
 373                 mlog(ML_ERROR, "bad mle: %p\n", mle);
 374                 dlm_print_one_mle(mle);
 375                 BUG();
 376         } else
 377                 kref_put(&mle->mle_refs, dlm_mle_release);
 378 }
 379
 380
 381 /* must not have any spinlocks coming in */
 382 static void dlm_put_mle(struct dlm_master_list_entry *mle)
 383 {
 384         struct dlm_ctxt *dlm;
 385         dlm = mle->dlm;
 386
 387         spin_lock(&dlm->spinlock);
 388         spin_lock(&dlm->master_lock);
 389         __dlm_put_mle(mle);
 390         spin_unlock(&dlm->master_lock);
 391         spin_unlock(&dlm->spinlock);
 392 }
 393
 394 static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
 395 {
 396         kref_get(&mle->mle_refs);
 397 }
 398
 399 static void dlm_init_mle(struct dlm_master_list_entry *mle,
 400                         enum dlm_mle_type type,
 401                         struct dlm_ctxt *dlm,
 402                         struct dlm_lock_resource *res,
 403                         const char *name,
 404                         unsigned int namelen)
 405 {
 406         assert_spin_locked(&dlm->spinlock);
 407
 408         mle->dlm = dlm;
 409         mle->type = type;
 410         INIT_LIST_HEAD(&mle->list);
 411         INIT_LIST_HEAD(&mle->hb_events);
 412         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
 413         spin_lock_init(&mle->spinlock);
 414         init_waitqueue_head(&mle->wq);
 415         atomic_set(&mle->woken, 0);
 416         kref_init(&mle->mle_refs);
 417         memset(mle->response_map, 0, sizeof(mle->response_map));
 418         mle->master = O2NM_MAX_NODES;
 419         mle->new_master = O2NM_MAX_NODES;
 420         mle->inuse = 0;
 421
 422         if (mle->type == DLM_MLE_MASTER) {
 423                 BUG_ON(!res);
 424                 mle->u.res = res;
 425         } else if (mle->type == DLM_MLE_BLOCK) {
 426                 BUG_ON(!name);
 427                 memcpy(mle->u.name.name, name, namelen);
 428                 mle->u.name.len = namelen;
 429         } else /* DLM_MLE_MIGRATION */ {
 430                 BUG_ON(!name);
 431                 memcpy(mle->u.name.name, name, namelen);
 432                 mle->u.name.len = namelen;
 433         }
 434
 435         /* copy off the node_map and register hb callbacks on our copy */
 436         memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
 437         memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
 438         clear_bit(dlm->node_num, mle->vote_map);
 439         clear_bit(dlm->node_num, mle->node_map);
 440
 441         /* attach the mle to the domain node up/down events */
 442         __dlm_mle_attach_hb_events(dlm, mle);
 443 }
 444
 445
 446 /* returns 1 if found, 0 if not */
 447 static int dlm_find_mle(struct dlm_ctxt *dlm,
 448                         struct dlm_master_list_entry **mle,
 449                         char *name, unsigned int namelen)
 450 {
 451         struct dlm_master_list_entry *tmpmle;
 452
 453         assert_spin_locked(&dlm->master_lock);
 454
 455         list_for_each_entry(tmpmle, &dlm->master_list, list) {
 456                 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
 457                         continue;
 458                 dlm_get_mle(tmpmle);
 459                 *mle = tmpmle;
 460                 return 1;
 461         }
 462         return 0;
 463 }
 464
 465 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
 466 {
 467         struct dlm_master_list_entry *mle;
 468
 469         assert_spin_locked(&dlm->spinlock);
 470
 471         list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
 472                 if (node_up)
 473                         dlm_mle_node_up(dlm, mle, NULL, idx);
 474                 else
 475                         dlm_mle_node_down(dlm, mle, NULL, idx);
 476         }
 477 }
 478
 479 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
 480                               struct dlm_master_list_entry *mle,
 481                               struct o2nm_node *node, int idx)
 482 {
 483         spin_lock(&mle->spinlock);
 484
 485         if (!test_bit(idx, mle->node_map))
 486                 mlog(0, "node %u already removed from nodemap!\n", idx);
 487         else
 488                 clear_bit(idx, mle->node_map);
 489
 490         spin_unlock(&mle->spinlock);
 491 }
 492
 493 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
 494                             struct dlm_master_list_entry *mle,
 495                             struct o2nm_node *node, int idx)
 496 {
 497         spin_lock(&mle->spinlock);
 498
 499         if (test_bit(idx, mle->node_map))
 500                 mlog(0, "node %u already in node map!\n", idx);
 501         else
 502                 set_bit(idx, mle->node_map);
 503
 504         spin_unlock(&mle->spinlock);
 505 }
 506
 507
 508 int dlm_init_mle_cache(void)
 509 {
 510         dlm_mle_cache = kmem_cache_create("o2dlm_mle",
 511                                           sizeof(struct dlm_master_list_entry),
 512                                           0, SLAB_HWCACHE_ALIGN,
 513                                           NULL);
 514         if (dlm_mle_cache == NULL)
 515                 return -ENOMEM;
 516         return 0;
 517 }
 518
 519 void dlm_destroy_mle_cache(void)
 520 {
 521         if (dlm_mle_cache)
 522                 kmem_cache_destroy(dlm_mle_cache);
 523 }
 524
 525 static void dlm_mle_release(struct kref *kref)
 526 {
 527         struct dlm_master_list_entry *mle;
 528         struct dlm_ctxt *dlm;
 529
 530         mlog_entry_void();
 531
 532         mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
 533         dlm = mle->dlm;
 534
 535         if (mle->type != DLM_MLE_MASTER) {
 536                 mlog(0, "calling mle_release for %.*s, type %d\n",
 537                      mle->u.name.len, mle->u.name.name, mle->type);
 538         } else {
 539                 mlog(0, "calling mle_release for %.*s, type %d\n",
 540                      mle->u.res->lockname.len,
 541                      mle->u.res->lockname.name, mle->type);
 542         }
 543         assert_spin_locked(&dlm->spinlock);
 544         assert_spin_locked(&dlm->master_lock);
 545
 546         /* remove from list if not already */
 547         if (!list_empty(&mle->list))
 548                 list_del_init(&mle->list);
 549
 550         /* detach the mle from the domain node up/down events */
 551         __dlm_mle_detach_hb_events(dlm, mle);
 552
 553         /* NOTE: kfree under spinlock here.
 554          * if this is bad, we can move this to a freelist. */
 555         kmem_cache_free(dlm_mle_cache, mle);
 556 }
 557
 558
 559 /*
 560  * LOCK RESOURCE FUNCTIONS
 561  */
 562
 563 int dlm_init_master_caches(void)
 564 {
 565         dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
 566                                               sizeof(struct dlm_lock_resource),
 567                                               0, SLAB_HWCACHE_ALIGN, NULL);
 568         if (!dlm_lockres_cache)
 569                 goto bail;
 570
 571         dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
 572                                                DLM_LOCKID_NAME_MAX, 0,
 573                                                SLAB_HWCACHE_ALIGN, NULL);
 574         if (!dlm_lockname_cache)
 575                 goto bail;
 576
 577         return 0;
 578 bail:
 579         dlm_destroy_master_caches();
 580         return -ENOMEM;
 581 }
 582
 583 void dlm_destroy_master_caches(void)
 584 {
 585         if (dlm_lockname_cache)
 586                 kmem_cache_destroy(dlm_lockname_cache);
 587
 588         if (dlm_lockres_cache)
 589                 kmem_cache_destroy(dlm_lockres_cache);
 590 }
 591
 592 static void dlm_set_lockres_owner(struct dlm_ctxt *dlm,
 593                                   struct dlm_lock_resource *res,
 594                                   u8 owner)
 595 {
 596         assert_spin_locked(&res->spinlock);
 597
 598         mlog_entry("%.*s, %u\n", res->lockname.len, res->lockname.name, owner);
 599
 600         if (owner == dlm->node_num)
 601                 atomic_inc(&dlm->local_resources);
 602         else if (owner == DLM_LOCK_RES_OWNER_UNKNOWN)
 603                 atomic_inc(&dlm->unknown_resources);
 604         else
 605                 atomic_inc(&dlm->remote_resources);
 606
 607         res->owner = owner;
 608 }
 609
 610 void dlm_change_lockres_owner(struct dlm_ctxt *dlm,
 611                               struct dlm_lock_resource *res, u8 owner)
 612 {
 613         assert_spin_locked(&res->spinlock);
 614
 615         if (owner == res->owner)
 616                 return;
 617
 618         if (res->owner == dlm->node_num)
 619                 atomic_dec(&dlm->local_resources);
 620         else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN)
 621                 atomic_dec(&dlm->unknown_resources);
 622         else
 623                 atomic_dec(&dlm->remote_resources);
 624
 625         dlm_set_lockres_owner(dlm, res, owner);
 626 }
 627
 628
 629 static void dlm_lockres_release(struct kref *kref)
 630 {
 631         struct dlm_lock_resource *res;
 632
 633         res = container_of(kref, struct dlm_lock_resource, refs);
 634
 635         /* This should not happen -- all lockres' have a name
 636          * associated with them at init time. */
 637         BUG_ON(!res->lockname.name);
 638
 639         mlog(0, "destroying lockres %.*s\n", res->lockname.len,
 640              res->lockname.name);
 641
 642         if (!hlist_unhashed(&res->hash_node) ||
 643             !list_empty(&res->granted) ||
 644             !list_empty(&res->converting) ||
 645             !list_empty(&res->blocked) ||
 646             !list_empty(&res->dirty) ||
 647             !list_empty(&res->recovering) ||
 648             !list_empty(&res->purge)) {
 649                 mlog(ML_ERROR,
 650                      "Going to BUG for resource %.*s."
 651                      "  We're on a list! [%c%c%c%c%c%c%c]\n",
 652                      res->lockname.len, res->lockname.name,
 653                      !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
 654                      !list_empty(&res->granted) ? 'G' : ' ',
 655                      !list_empty(&res->converting) ? 'C' : ' ',
 656                      !list_empty(&res->blocked) ? 'B' : ' ',
 657                      !list_empty(&res->dirty) ? 'D' : ' ',
 658                      !list_empty(&res->recovering) ? 'R' : ' ',
 659                      !list_empty(&res->purge) ? 'P' : ' ');
 660
 661                 dlm_print_one_lock_resource(res);
 662         }
 663
 664         /* By the time we're ready to blow this guy away, we shouldn't
 665          * be on any lists. */
 666         BUG_ON(!hlist_unhashed(&res->hash_node));
 667         BUG_ON(!list_empty(&res->granted));
 668         BUG_ON(!list_empty(&res->converting));
 669         BUG_ON(!list_empty(&res->blocked));
 670         BUG_ON(!list_empty(&res->dirty));
 671         BUG_ON(!list_empty(&res->recovering));
 672         BUG_ON(!list_empty(&res->purge));
 673
 674         kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
 675
 676         kmem_cache_free(dlm_lockres_cache, res);
 677 }
 678
 679 void dlm_lockres_put(struct dlm_lock_resource *res)
 680 {
 681         kref_put(&res->refs, dlm_lockres_release);
 682 }
 683
 684 static void dlm_init_lockres(struct dlm_ctxt *dlm,
 685                              struct dlm_lock_resource *res,
 686                              const char *name, unsigned int namelen)
 687 {
 688         char *qname;
 689
 690         /* If we memset here, we lose our reference to the kmalloc'd
 691          * res->lockname.name, so be sure to init every field
 692          * correctly! */
 693
 694         qname = (char *) res->lockname.name;
 695         memcpy(qname, name, namelen);
 696
 697         res->lockname.len = namelen;
 698         res->lockname.hash = dlm_lockid_hash(name, namelen);
 699
 700         init_waitqueue_head(&res->wq);
 701         spin_lock_init(&res->spinlock);
 702         INIT_HLIST_NODE(&res->hash_node);
 703         INIT_LIST_HEAD(&res->granted);
 704         INIT_LIST_HEAD(&res->converting);
 705         INIT_LIST_HEAD(&res->blocked);
 706         INIT_LIST_HEAD(&res->dirty);
 707         INIT_LIST_HEAD(&res->recovering);
 708         INIT_LIST_HEAD(&res->purge);
 709         atomic_set(&res->asts_reserved, 0);
 710         res->migration_pending = 0;
 711         res->inflight_locks = 0;
 712
 713         kref_init(&res->refs);
 714
 715         /* just for consistency */
 716         spin_lock(&res->spinlock);
 717         dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
 718         spin_unlock(&res->spinlock);
 719
 720         res->state = DLM_LOCK_RES_IN_PROGRESS;
 721
 722         res->last_used = 0;
 723
 724         memset(res->lvb, 0, DLM_LVB_LEN);
 725         memset(res->refmap, 0, sizeof(res->refmap));
 726 }
 727
 728 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
 729                                    const char *name,
 730                                    unsigned int namelen)
 731 {
 732         struct dlm_lock_resource *res = NULL;
 733
 734         res = (struct dlm_lock_resource *)
 735                                 kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
 736         if (!res)
 737                 goto error;
 738
 739         res->lockname.name = (char *)
 740                                 kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
 741         if (!res->lockname.name)
 742                 goto error;
 743
 744         dlm_init_lockres(dlm, res, name, namelen);
 745         return res;
 746
 747 error:
 748         if (res && res->lockname.name)
 749                 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
 750
 751         if (res)
 752                 kmem_cache_free(dlm_lockres_cache, res);
 753         return NULL;
 754 }
 755
 756 void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
 757                                    struct dlm_lock_resource *res,
 758                                    int new_lockres,
 759                                    const char *file,
 760                                    int line)
 761 {
 762         if (!new_lockres)
 763                 assert_spin_locked(&res->spinlock);
 764
 765         if (!test_bit(dlm->node_num, res->refmap)) {
 766                 BUG_ON(res->inflight_locks != 0);
 767                 dlm_lockres_set_refmap_bit(dlm->node_num, res);
 768         }
 769         res->inflight_locks++;
 770         mlog(0, "%s:%.*s: inflight++: now %u\n",
 771              dlm->name, res->lockname.len, res->lockname.name,
 772              res->inflight_locks);
 773 }
 774
 775 void __dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
 776                                    struct dlm_lock_resource *res,
 777                                    const char *file,
 778                                    int line)
 779 {
 780         assert_spin_locked(&res->spinlock);
 781
 782         BUG_ON(res->inflight_locks == 0);
 783         res->inflight_locks--;
 784         mlog(0, "%s:%.*s: inflight--: now %u\n",
 785              dlm->name, res->lockname.len, res->lockname.name,
 786              res->inflight_locks);
 787         if (res->inflight_locks == 0)
 788                 dlm_lockres_clear_refmap_bit(dlm->node_num, res);
 789         wake_up(&res->wq);
 790 }
 791
 792 /*
 793  * lookup a lock resource by name.
 794  * may already exist in the hashtable.
 795  * lockid is null terminated
 796  *
 797  * if not, allocate enough for the lockres and for
 798  * the temporary structure used in doing the mastering.
 799  *
 800  * also, do a lookup in the dlm->master_list to see
 801  * if another node has begun mastering the same lock.
 802  * if so, there should be a block entry in there
 803  * for this name, and we should *not* attempt to master
 804  * the lock here.   need to wait around for that node
 805  * to assert_master (or die).
 806  *
 807  */
 808 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
 809                                           const char *lockid,
 810                                           int namelen,
 811                                           int flags)
 812 {
 813         struct dlm_lock_resource *tmpres=NULL, *res=NULL;
 814         struct dlm_master_list_entry *mle = NULL;
 815         struct dlm_master_list_entry *alloc_mle = NULL;
 816         int blocked = 0;
 817         int ret, nodenum;
 818         struct dlm_node_iter iter;
 819         unsigned int hash;
 820         int tries = 0;
 821         int bit, wait_on_recovery = 0;
 822         int drop_inflight_if_nonlocal = 0;
 823
 824         BUG_ON(!lockid);
 825
 826         hash = dlm_lockid_hash(lockid, namelen);
 827
 828         mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
 829
 830 lookup:
 831         spin_lock(&dlm->spinlock);
 832         tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
 833         if (tmpres) {
 834                 int dropping_ref = 0;
 835
 836                 spin_lock(&tmpres->spinlock);
 837                 if (tmpres->owner == dlm->node_num) {
 838                         BUG_ON(tmpres->state & DLM_LOCK_RES_DROPPING_REF);
 839                         dlm_lockres_grab_inflight_ref(dlm, tmpres);
 840                 } else if (tmpres->state & DLM_LOCK_RES_DROPPING_REF)
 841                         dropping_ref = 1;
 842                 spin_unlock(&tmpres->spinlock);
 843                 spin_unlock(&dlm->spinlock);
 844
 845                 /* wait until done messaging the master, drop our ref to allow
 846                  * the lockres to be purged, start over. */
 847                 if (dropping_ref) {
 848                         spin_lock(&tmpres->spinlock);
 849                         __dlm_wait_on_lockres_flags(tmpres, DLM_LOCK_RES_DROPPING_REF);
 850                         spin_unlock(&tmpres->spinlock);
 851                         dlm_lockres_put(tmpres);
 852                         tmpres = NULL;
 853                         goto lookup;
 854                 }
 855
 856                 mlog(0, "found in hash!\n");
 857                 if (res)
 858                         dlm_lockres_put(res);
 859                 res = tmpres;
 860                 goto leave;
 861         }
 862
 863         if (!res) {
 864                 spin_unlock(&dlm->spinlock);
 865                 mlog(0, "allocating a new resource\n");
 866                 /* nothing found and we need to allocate one. */
 867                 alloc_mle = (struct dlm_master_list_entry *)
 868                         kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
 869                 if (!alloc_mle)
 870                         goto leave;
 871                 res = dlm_new_lockres(dlm, lockid, namelen);
 872                 if (!res)
 873                         goto leave;
 874                 goto lookup;
 875         }
 876
 877         mlog(0, "no lockres found, allocated our own: %p\n", res);
 878
 879         if (flags & LKM_LOCAL) {
 880                 /* caller knows it's safe to assume it's not mastered elsewhere
 881                  * DONE!  return right away */
 882                 spin_lock(&res->spinlock);
 883                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
 884                 __dlm_insert_lockres(dlm, res);
 885                 dlm_lockres_grab_inflight_ref(dlm, res);
 886                 spin_unlock(&res->spinlock);
 887                 spin_unlock(&dlm->spinlock);
 888                 /* lockres still marked IN_PROGRESS */
 889                 goto wake_waiters;
 890         }
 891
 892         /* check master list to see if another node has started mastering it */
 893         spin_lock(&dlm->master_lock);
 894
 895         /* if we found a block, wait for lock to be mastered by another node */
 896         blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
 897         if (blocked) {
 898                 int mig;
 899                 if (mle->type == DLM_MLE_MASTER) {
 900                         mlog(ML_ERROR, "master entry for nonexistent lock!\n");
 901                         BUG();
 902                 }
 903                 mig = (mle->type == DLM_MLE_MIGRATION);
 904                 /* if there is a migration in progress, let the migration
 905                  * finish before continuing.  we can wait for the absence
 906                  * of the MIGRATION mle: either the migrate finished or
 907                  * one of the nodes died and the mle was cleaned up.
 908                  * if there is a BLOCK here, but it already has a master
 909                  * set, we are too late.  the master does not have a ref
 910                  * for us in the refmap.  detach the mle and drop it.
 911                  * either way, go back to the top and start over. */
 912                 if (mig || mle->master != O2NM_MAX_NODES) {
 913                         BUG_ON(mig && mle->master == dlm->node_num);
 914                         /* we arrived too late.  the master does not
 915                          * have a ref for us. retry. */
 916                         mlog(0, "%s:%.*s: late on %s\n",
 917                              dlm->name, namelen, lockid,
 918                              mig ?  "MIGRATION" : "BLOCK");
 919                         spin_unlock(&dlm->master_lock);
 920                         spin_unlock(&dlm->spinlock);
 921
 922                         /* master is known, detach */
 923                         if (!mig)
 924                                 dlm_mle_detach_hb_events(dlm, mle);
 925                         dlm_put_mle(mle);
 926                         mle = NULL;
 927                         /* this is lame, but we cant wait on either
 928                          * the mle or lockres waitqueue here */
 929                         if (mig)
 930                                 msleep(100);
 931                         goto lookup;
 932                 }
 933         } else {
 934                 /* go ahead and try to master lock on this node */
 935                 mle = alloc_mle;
 936                 /* make sure this does not get freed below */
 937                 alloc_mle = NULL;
 938                 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
 939                 set_bit(dlm->node_num, mle->maybe_map);
 940                 list_add(&mle->list, &dlm->master_list);
 941
 942                 /* still holding the dlm spinlock, check the recovery map
 943                  * to see if there are any nodes that still need to be
 944                  * considered.  these will not appear in the mle nodemap
 945                  * but they might own this lockres.  wait on them. */
 946                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
 947                 if (bit < O2NM_MAX_NODES) {
 948                         mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to "
 949                              "recover before lock mastery can begin\n",
 950                              dlm->name, namelen, (char *)lockid, bit);
 951                         wait_on_recovery = 1;
 952                 }
 953         }
 954
 955         /* at this point there is either a DLM_MLE_BLOCK or a
 956          * DLM_MLE_MASTER on the master list, so it's safe to add the
 957          * lockres to the hashtable.  anyone who finds the lock will
 958          * still have to wait on the IN_PROGRESS. */
 959
 960         /* finally add the lockres to its hash bucket */
 961         __dlm_insert_lockres(dlm, res);
 962         /* since this lockres is new it doesnt not require the spinlock */
 963         dlm_lockres_grab_inflight_ref_new(dlm, res);
 964
 965         /* if this node does not become the master make sure to drop
 966          * this inflight reference below */
 967         drop_inflight_if_nonlocal = 1;
 968
 969         /* get an extra ref on the mle in case this is a BLOCK
 970          * if so, the creator of the BLOCK may try to put the last
 971          * ref at this time in the assert master handler, so we
 972          * need an extra one to keep from a bad ptr deref. */
 973         dlm_get_mle_inuse(mle);
 974         spin_unlock(&dlm->master_lock);
 975         spin_unlock(&dlm->spinlock);
 976
 977 redo_request:
 978         while (wait_on_recovery) {
 979                 /* any cluster changes that occurred after dropping the
 980                  * dlm spinlock would be detectable be a change on the mle,
 981                  * so we only need to clear out the recovery map once. */
 982                 if (dlm_is_recovery_lock(lockid, namelen)) {
 983                         mlog(ML_NOTICE, "%s: recovery map is not empty, but "
 984                              "must master $RECOVERY lock now\n", dlm->name);
 985                         if (!dlm_pre_master_reco_lockres(dlm, res))
 986                                 wait_on_recovery = 0;
 987                         else {
 988                                 mlog(0, "%s: waiting 500ms for heartbeat state "
 989                                     "change\n", dlm->name);
 990                                 msleep(500);
 991                         }
 992                         continue;
 993                 }
 994
 995                 dlm_kick_recovery_thread(dlm);
 996                 msleep(1000);
 997                 dlm_wait_for_recovery(dlm);
 998
 999                 spin_lock(&dlm->spinlock);
1000                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
1001                 if (bit < O2NM_MAX_NODES) {
1002                         mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to "
1003                              "recover before lock mastery can begin\n",
1004                              dlm->name, namelen, (char *)lockid, bit);
1005                         wait_on_recovery = 1;
1006                 } else
1007                         wait_on_recovery = 0;
1008                 spin_unlock(&dlm->spinlock);
1009
1010                 if (wait_on_recovery)
1011                         dlm_wait_for_node_recovery(dlm, bit, 10000);
1012         }
1013
1014         /* must wait for lock to be mastered elsewhere */
1015         if (blocked)
1016                 goto wait;
1017
1018         ret = -EINVAL;
1019         dlm_node_iter_init(mle->vote_map, &iter);
1020         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1021                 ret = dlm_do_master_request(res, mle, nodenum);
1022                 if (ret < 0)
1023                         mlog_errno(ret);
1024                 if (mle->master != O2NM_MAX_NODES) {
1025                         /* found a master ! */
1026                         if (mle->master <= nodenum)
1027                                 break;
1028                         /* if our master request has not reached the master
1029                          * yet, keep going until it does.  this is how the
1030                          * master will know that asserts are needed back to
1031                          * the lower nodes. */
1032                         mlog(0, "%s:%.*s: requests only up to %u but master "
1033                              "is %u, keep going\n", dlm->name, namelen,
1034                              lockid, nodenum, mle->master);
1035                 }
1036         }
1037
1038 wait:
1039         /* keep going until the response map includes all nodes */
1040         ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
1041         if (ret < 0) {
1042                 wait_on_recovery = 1;
1043                 mlog(0, "%s:%.*s: node map changed, redo the "
1044                      "master request now, blocked=%d\n",
1045                      dlm->name, res->lockname.len,
1046                      res->lockname.name, blocked);
1047                 if (++tries > 20) {
1048                         mlog(ML_ERROR, "%s:%.*s: spinning on "
1049                              "dlm_wait_for_lock_mastery, blocked=%d\n",
1050                              dlm->name, res->lockname.len,
1051                              res->lockname.name, blocked);
1052                         dlm_print_one_lock_resource(res);
1053                         dlm_print_one_mle(mle);
1054                         tries = 0;
1055                 }
1056                 goto redo_request;
1057         }
1058
1059         mlog(0, "lockres mastered by %u\n", res->owner);
1060         /* make sure we never continue without this */
1061         BUG_ON(res->owner == O2NM_MAX_NODES);
1062
1063         /* master is known, detach if not already detached */
1064         dlm_mle_detach_hb_events(dlm, mle);
1065         dlm_put_mle(mle);
1066         /* put the extra ref */
1067         dlm_put_mle_inuse(mle);
1068
1069 wake_waiters:
1070         spin_lock(&res->spinlock);
1071         if (res->owner != dlm->node_num && drop_inflight_if_nonlocal)
1072                 dlm_lockres_drop_inflight_ref(dlm, res);
1073         res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1074         spin_unlock(&res->spinlock);
1075         wake_up(&res->wq);
1076
1077 leave:
1078         /* need to free the unused mle */
1079         if (alloc_mle)
1080                 kmem_cache_free(dlm_mle_cache, alloc_mle);
1081
1082         return res;
1083 }
1084
1085
1086 #define DLM_MASTERY_TIMEOUT_MS   5000
1087
1088 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
1089                                      struct dlm_lock_resource *res,
1090                                      struct dlm_master_list_entry *mle,
1091                                      int *blocked)
1092 {
1093         u8 m;
1094         int ret, bit;
1095         int map_changed, voting_done;
1096         int assert, sleep;
1097
1098 recheck:
1099         ret = 0;
1100         assert = 0;
1101
1102         /* check if another node has already become the owner */
1103         spin_lock(&res->spinlock);
1104         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1105                 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1106                      res->lockname.len, res->lockname.name, res->owner);
1107                 spin_unlock(&res->spinlock);
1108                 /* this will cause the master to re-assert across
1109                  * the whole cluster, freeing up mles */
1110                 if (res->owner != dlm->node_num) {
1111                         ret = dlm_do_master_request(res, mle, res->owner);
1112                         if (ret < 0) {
1113                                 /* give recovery a chance to run */
1114                                 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1115                                 msleep(500);
1116                                 goto recheck;
1117                         }
1118                 }
1119                 ret = 0;
1120                 goto leave;
1121         }
1122         spin_unlock(&res->spinlock);
1123
1124         spin_lock(&mle->spinlock);
1125         m = mle->master;
1126         map_changed = (memcmp(mle->vote_map, mle->node_map,
1127                               sizeof(mle->vote_map)) != 0);
1128         voting_done = (memcmp(mle->vote_map, mle->response_map,
1129                              sizeof(mle->vote_map)) == 0);
1130
1131         /* restart if we hit any errors */
1132         if (map_changed) {
1133                 int b;
1134                 mlog(0, "%s: %.*s: node map changed, restarting\n",
1135                      dlm->name, res->lockname.len, res->lockname.name);
1136                 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1137                 b = (mle->type == DLM_MLE_BLOCK);
1138                 if ((*blocked && !b) || (!*blocked && b)) {
1139                         mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1140                              dlm->name, res->lockname.len, res->lockname.name,
1141                              *blocked, b);
1142                         *blocked = b;
1143                 }
1144                 spin_unlock(&mle->spinlock);
1145                 if (ret < 0) {
1146                         mlog_errno(ret);
1147                         goto leave;
1148                 }
1149                 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1150                      "rechecking now\n", dlm->name, res->lockname.len,
1151                      res->lockname.name);
1152                 goto recheck;
1153         } else {
1154                 if (!voting_done) {
1155                         mlog(0, "map not changed and voting not done "
1156                              "for %s:%.*s\n", dlm->name, res->lockname.len,
1157                              res->lockname.name);
1158                 }
1159         }
1160
1161         if (m != O2NM_MAX_NODES) {
1162                 /* another node has done an assert!
1163                  * all done! */
1164                 sleep = 0;
1165         } else {
1166                 sleep = 1;
1167                 /* have all nodes responded? */
1168                 if (voting_done && !*blocked) {
1169                         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1170                         if (dlm->node_num <= bit) {
1171                                 /* my node number is lowest.
1172                                  * now tell other nodes that I am
1173                                  * mastering this. */
1174                                 mle->master = dlm->node_num;
1175                                 /* ref was grabbed in get_lock_resource
1176                                  * will be dropped in dlmlock_master */
1177                                 assert = 1;
1178                                 sleep = 0;
1179                         }
1180                         /* if voting is done, but we have not received
1181                          * an assert master yet, we must sleep */
1182                 }
1183         }
1184
1185         spin_unlock(&mle->spinlock);
1186
1187         /* sleep if we haven't finished voting yet */
1188         if (sleep) {
1189                 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1190
1191                 /*
1192                 if (atomic_read(&mle->mle_refs.refcount) < 2)
1193                         mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1194                         atomic_read(&mle->mle_refs.refcount),
1195                         res->lockname.len, res->lockname.name);
1196                 */
1197                 atomic_set(&mle->woken, 0);
1198                 (void)wait_event_timeout(mle->wq,
1199                                          (atomic_read(&mle->woken) == 1),
1200                                          timeo);
1201                 if (res->owner == O2NM_MAX_NODES) {
1202                         mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1203                              res->lockname.len, res->lockname.name);
1204                         goto recheck;
1205                 }
1206                 mlog(0, "done waiting, master is %u\n", res->owner);
1207                 ret = 0;
1208                 goto leave;
1209         }
1210
1211         ret = 0;   /* done */
1212         if (assert) {
1213                 m = dlm->node_num;
1214                 mlog(0, "about to master %.*s here, this=%u\n",
1215                      res->lockname.len, res->lockname.name, m);
1216                 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1217                 if (ret) {
1218                         /* This is a failure in the network path,
1219                          * not in the response to the assert_master
1220                          * (any nonzero response is a BUG on this node).
1221                          * Most likely a socket just got disconnected
1222                          * due to node death. */
1223                         mlog_errno(ret);
1224                 }
1225                 /* no longer need to restart lock mastery.
1226                  * all living nodes have been contacted. */
1227                 ret = 0;
1228         }
1229
1230         /* set the lockres owner */
1231         spin_lock(&res->spinlock);
1232         /* mastery reference obtained either during
1233          * assert_master_handler or in get_lock_resource */
1234         dlm_change_lockres_owner(dlm, res, m);
1235         spin_unlock(&res->spinlock);
1236
1237 leave:
1238         return ret;
1239 }
1240
1241 struct dlm_bitmap_diff_iter
1242 {
1243         int curnode;
1244         unsigned long *orig_bm;
1245         unsigned long *cur_bm;
1246         unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1247 };
1248
1249 enum dlm_node_state_change
1250 {
1251         NODE_DOWN = -1,
1252         NODE_NO_CHANGE = 0,
1253         NODE_UP
1254 };
1255
1256 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1257                                       unsigned long *orig_bm,
1258                                       unsigned long *cur_bm)
1259 {
1260         unsigned long p1, p2;
1261         int i;
1262
1263         iter->curnode = -1;
1264         iter->orig_bm = orig_bm;
1265         iter->cur_bm = cur_bm;
1266
1267         for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1268                 p1 = *(iter->orig_bm + i);
1269                 p2 = *(iter->cur_bm + i);
1270                 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1271         }
1272 }
1273
1274 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1275                                      enum dlm_node_state_change *state)
1276 {
1277         int bit;
1278
1279         if (iter->curnode >= O2NM_MAX_NODES)
1280                 return -ENOENT;
1281
1282         bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1283                             iter->curnode+1);
1284         if (bit >= O2NM_MAX_NODES) {
1285                 iter->curnode = O2NM_MAX_NODES;
1286                 return -ENOENT;
1287         }
1288
1289         /* if it was there in the original then this node died */
1290         if (test_bit(bit, iter->orig_bm))
1291                 *state = NODE_DOWN;
1292         else
1293                 *state = NODE_UP;
1294
1295         iter->curnode = bit;
1296         return bit;
1297 }
1298
1299
1300 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1301                                     struct dlm_lock_resource *res,
1302                                     struct dlm_master_list_entry *mle,
1303                                     int blocked)
1304 {
1305         struct dlm_bitmap_diff_iter bdi;
1306         enum dlm_node_state_change sc;
1307         int node;
1308         int ret = 0;
1309
1310         mlog(0, "something happened such that the "
1311              "master process may need to be restarted!\n");
1312
1313         assert_spin_locked(&mle->spinlock);
1314
1315         dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1316         node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1317         while (node >= 0) {
1318                 if (sc == NODE_UP) {
1319                         /* a node came up.  clear any old vote from
1320                          * the response map and set it in the vote map
1321                          * then restart the mastery. */
1322                         mlog(ML_NOTICE, "node %d up while restarting\n", node);
1323
1324                         /* redo the master request, but only for the new node */
1325                         mlog(0, "sending request to new node\n");
1326                         clear_bit(node, mle->response_map);
1327                         set_bit(node, mle->vote_map);
1328                 } else {
1329                         mlog(ML_ERROR, "node down! %d\n", node);
1330                         if (blocked) {
1331                                 int lowest = find_next_bit(mle->maybe_map,
1332                                                        O2NM_MAX_NODES, 0);
1333
1334                                 /* act like it was never there */
1335                                 clear_bit(node, mle->maybe_map);
1336
1337                                 if (node == lowest) {
1338                                         mlog(0, "expected master %u died"
1339                                             " while this node was blocked "
1340                                             "waiting on it!\n", node);
1341                                         lowest = find_next_bit(mle->maybe_map,
1342                                                         O2NM_MAX_NODES,
1343                                                         lowest+1);
1344                                         if (lowest < O2NM_MAX_NODES) {
1345                                                 mlog(0, "%s:%.*s:still "
1346                                                      "blocked. waiting on %u "
1347                                                      "now\n", dlm->name,
1348                                                      res->lockname.len,
1349                                                      res->lockname.name,
1350                                                      lowest);
1351                                         } else {
1352                                                 /* mle is an MLE_BLOCK, but
1353                                                  * there is now nothing left to
1354                                                  * block on.  we need to return
1355                                                  * all the way back out and try
1356                                                  * again with an MLE_MASTER.
1357                                                  * dlm_do_local_recovery_cleanup
1358                                                  * has already run, so the mle
1359                                                  * refcount is ok */
1360                                                 mlog(0, "%s:%.*s: no "
1361                                                      "longer blocking. try to "
1362                                                      "master this here\n",
1363                                                      dlm->name,
1364                                                      res->lockname.len,
1365                                                      res->lockname.name);
1366                                                 mle->type = DLM_MLE_MASTER;
1367                                                 mle->u.res = res;
1368                                         }
1369                                 }
1370                         }
1371
1372                         /* now blank out everything, as if we had never
1373                          * contacted anyone */
1374                         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1375                         memset(mle->response_map, 0, sizeof(mle->response_map));
1376                         /* reset the vote_map to the current node_map */
1377                         memcpy(mle->vote_map, mle->node_map,
1378                                sizeof(mle->node_map));
1379                         /* put myself into the maybe map */
1380                         if (mle->type != DLM_MLE_BLOCK)
1381                                 set_bit(dlm->node_num, mle->maybe_map);
1382                 }
1383                 ret = -EAGAIN;
1384                 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1385         }
1386         return ret;
1387 }
1388
1389
1390 /*
1391  * DLM_MASTER_REQUEST_MSG
1392  *
1393  * returns: 0 on success,
1394  *          -errno on a network error
1395  *
1396  * on error, the caller should assume the target node is "dead"
1397  *
1398  */
1399
1400 static int dlm_do_master_request(struct dlm_lock_resource *res,
1401                                  struct dlm_master_list_entry *mle, int to)
1402 {
1403         struct dlm_ctxt *dlm = mle->dlm;
1404         struct dlm_master_request request;
1405         int ret, response=0, resend;
1406
1407         memset(&request, 0, sizeof(request));
1408         request.node_idx = dlm->node_num;
1409
1410         BUG_ON(mle->type == DLM_MLE_MIGRATION);
1411
1412         if (mle->type != DLM_MLE_MASTER) {
1413                 request.namelen = mle->u.name.len;
1414                 memcpy(request.name, mle->u.name.name, request.namelen);
1415         } else {
1416                 request.namelen = mle->u.res->lockname.len;
1417                 memcpy(request.name, mle->u.res->lockname.name,
1418                         request.namelen);
1419         }
1420
1421 again:
1422         ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1423                                  sizeof(request), to, &response);
1424         if (ret < 0)  {
1425                 if (ret == -ESRCH) {
1426                         /* should never happen */
1427                         mlog(ML_ERROR, "TCP stack not ready!\n");
1428                         BUG();
1429                 } else if (ret == -EINVAL) {
1430                         mlog(ML_ERROR, "bad args passed to o2net!\n");
1431                         BUG();
1432                 } else if (ret == -ENOMEM) {
1433                         mlog(ML_ERROR, "out of memory while trying to send "
1434                              "network message!  retrying\n");
1435                         /* this is totally crude */
1436                         msleep(50);
1437                         goto again;
1438                 } else if (!dlm_is_host_down(ret)) {
1439                         /* not a network error. bad. */
1440                         mlog_errno(ret);
1441                         mlog(ML_ERROR, "unhandled error!");
1442                         BUG();
1443                 }
1444                 /* all other errors should be network errors,
1445                  * and likely indicate node death */
1446                 mlog(ML_ERROR, "link to %d went down!\n", to);
1447                 goto out;
1448         }
1449
1450         ret = 0;
1451         resend = 0;
1452         spin_lock(&mle->spinlock);
1453         switch (response) {
1454                 case DLM_MASTER_RESP_YES:
1455                         set_bit(to, mle->response_map);
1456                         mlog(0, "node %u is the master, response=YES\n", to);
1457                         mlog(0, "%s:%.*s: master node %u now knows I have a "
1458                              "reference\n", dlm->name, res->lockname.len,
1459                              res->lockname.name, to);
1460                         mle->master = to;
1461                         break;
1462                 case DLM_MASTER_RESP_NO:
1463                         mlog(0, "node %u not master, response=NO\n", to);
1464                         set_bit(to, mle->response_map);
1465                         break;
1466                 case DLM_MASTER_RESP_MAYBE:
1467                         mlog(0, "node %u not master, response=MAYBE\n", to);
1468                         set_bit(to, mle->response_map);
1469                         set_bit(to, mle->maybe_map);
1470                         break;
1471                 case DLM_MASTER_RESP_ERROR:
1472                         mlog(0, "node %u hit an error, resending\n", to);
1473                         resend = 1;
1474                         response = 0;
1475                         break;
1476                 default:
1477                         mlog(ML_ERROR, "bad response! %u\n", response);
1478                         BUG();
1479         }
1480         spin_unlock(&mle->spinlock);
1481         if (resend) {
1482                 /* this is also totally crude */
1483                 msleep(50);
1484                 goto again;
1485         }
1486
1487 out:
1488         return ret;
1489 }
1490
1491 /*
1492  * locks that can be taken here:
1493  * dlm->spinlock
1494  * res->spinlock
1495  * mle->spinlock
1496  * dlm->master_list
1497  *
1498  * if possible, TRIM THIS DOWN!!!
1499  */
1500 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1501                                void **ret_data)
1502 {
1503         u8 response = DLM_MASTER_RESP_MAYBE;
1504         struct dlm_ctxt *dlm = data;
1505         struct dlm_lock_resource *res = NULL;
1506         struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1507         struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1508         char *name;
1509         unsigned int namelen, hash;
1510         int found, ret;
1511         int set_maybe;
1512         int dispatch_assert = 0;
1513
1514         if (!dlm_grab(dlm))
1515                 return DLM_MASTER_RESP_NO;
1516
1517         if (!dlm_domain_fully_joined(dlm)) {
1518                 response = DLM_MASTER_RESP_NO;
1519                 goto send_response;
1520         }
1521
1522         name = request->name;
1523         namelen = request->namelen;
1524         hash = dlm_lockid_hash(name, namelen);
1525
1526         if (namelen > DLM_LOCKID_NAME_MAX) {
1527                 response = DLM_IVBUFLEN;
1528                 goto send_response;
1529         }
1530
1531 way_up_top:
1532         spin_lock(&dlm->spinlock);
1533         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1534         if (res) {
1535                 spin_unlock(&dlm->spinlock);
1536
1537                 /* take care of the easy cases up front */
1538                 spin_lock(&res->spinlock);
1539                 if (res->state & (DLM_LOCK_RES_RECOVERING|
1540                                   DLM_LOCK_RES_MIGRATING)) {
1541                         spin_unlock(&res->spinlock);
1542                         mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1543                              "being recovered/migrated\n");
1544                         response = DLM_MASTER_RESP_ERROR;
1545                         if (mle)
1546                                 kmem_cache_free(dlm_mle_cache, mle);
1547                         goto send_response;
1548                 }
1549
1550                 if (res->owner == dlm->node_num) {
1551                         mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1552                              dlm->name, namelen, name, request->node_idx);
1553                         dlm_lockres_set_refmap_bit(request->node_idx, res);
1554                         spin_unlock(&res->spinlock);
1555                         response = DLM_MASTER_RESP_YES;
1556                         if (mle)
1557                                 kmem_cache_free(dlm_mle_cache, mle);
1558
1559                         /* this node is the owner.
1560                          * there is some extra work that needs to
1561                          * happen now.  the requesting node has
1562                          * caused all nodes up to this one to
1563                          * create mles.  this node now needs to
1564                          * go back and clean those up. */
1565                         dispatch_assert = 1;
1566                         goto send_response;
1567                 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1568                         spin_unlock(&res->spinlock);
1569                         // mlog(0, "node %u is the master\n", res->owner);
1570                         response = DLM_MASTER_RESP_NO;
1571                         if (mle)
1572                                 kmem_cache_free(dlm_mle_cache, mle);
1573                         goto send_response;
1574                 }
1575
1576                 /* ok, there is no owner.  either this node is
1577                  * being blocked, or it is actively trying to
1578                  * master this lock. */
1579                 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1580                         mlog(ML_ERROR, "lock with no owner should be "
1581                              "in-progress!\n");
1582                         BUG();
1583                 }
1584
1585                 // mlog(0, "lockres is in progress...\n");
1586                 spin_lock(&dlm->master_lock);
1587                 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1588                 if (!found) {
1589                         mlog(ML_ERROR, "no mle found for this lock!\n");
1590                         BUG();
1591                 }
1592                 set_maybe = 1;
1593                 spin_lock(&tmpmle->spinlock);
1594                 if (tmpmle->type == DLM_MLE_BLOCK) {
1595                         // mlog(0, "this node is waiting for "
1596                         // "lockres to be mastered\n");
1597                         response = DLM_MASTER_RESP_NO;
1598                 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1599                         mlog(0, "node %u is master, but trying to migrate to "
1600                              "node %u.\n", tmpmle->master, tmpmle->new_master);
1601                         if (tmpmle->master == dlm->node_num) {
1602                                 mlog(ML_ERROR, "no owner on lockres, but this "
1603                                      "node is trying to migrate it to %u?!\n",
1604                                      tmpmle->new_master);
1605                                 BUG();
1606                         } else {
1607                                 /* the real master can respond on its own */
1608                                 response = DLM_MASTER_RESP_NO;
1609                         }
1610                 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1611                         set_maybe = 0;
1612                         if (tmpmle->master == dlm->node_num) {
1613                                 response = DLM_MASTER_RESP_YES;
1614                                 /* this node will be the owner.
1615                                  * go back and clean the mles on any
1616                                  * other nodes */
1617                                 dispatch_assert = 1;
1618                                 dlm_lockres_set_refmap_bit(request->node_idx, res);
1619                                 mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1620                                      dlm->name, namelen, name,
1621                                      request->node_idx);
1622                         } else
1623                                 response = DLM_MASTER_RESP_NO;
1624                 } else {
1625                         // mlog(0, "this node is attempting to "
1626                         // "master lockres\n");
1627                         response = DLM_MASTER_RESP_MAYBE;
1628                 }
1629                 if (set_maybe)
1630                         set_bit(request->node_idx, tmpmle->maybe_map);
1631                 spin_unlock(&tmpmle->spinlock);
1632
1633                 spin_unlock(&dlm->master_lock);
1634                 spin_unlock(&res->spinlock);
1635
1636                 /* keep the mle attached to heartbeat events */
1637                 dlm_put_mle(tmpmle);
1638                 if (mle)
1639                         kmem_cache_free(dlm_mle_cache, mle);
1640                 goto send_response;
1641         }
1642
1643         /*
1644          * lockres doesn't exist on this node
1645          * if there is an MLE_BLOCK, return NO
1646          * if there is an MLE_MASTER, return MAYBE
1647          * otherwise, add an MLE_BLOCK, return NO
1648          */
1649         spin_lock(&dlm->master_lock);
1650         found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1651         if (!found) {
1652                 /* this lockid has never been seen on this node yet */
1653                 // mlog(0, "no mle found\n");
1654                 if (!mle) {
1655                         spin_unlock(&dlm->master_lock);
1656                         spin_unlock(&dlm->spinlock);
1657
1658                         mle = (struct dlm_master_list_entry *)
1659                                 kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1660                         if (!mle) {
1661                                 response = DLM_MASTER_RESP_ERROR;
1662                                 mlog_errno(-ENOMEM);
1663                                 goto send_response;
1664                         }
1665                         goto way_up_top;
1666                 }
1667
1668                 // mlog(0, "this is second time thru, already allocated, "
1669                 // "add the block.\n");
1670                 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1671                 set_bit(request->node_idx, mle->maybe_map);
1672                 list_add(&mle->list, &dlm->master_list);
1673                 response = DLM_MASTER_RESP_NO;
1674         } else {
1675                 // mlog(0, "mle was found\n");
1676                 set_maybe = 1;
1677                 spin_lock(&tmpmle->spinlock);
1678                 if (tmpmle->master == dlm->node_num) {
1679                         mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1680                         BUG();
1681                 }
1682                 if (tmpmle->type == DLM_MLE_BLOCK)
1683                         response = DLM_MASTER_RESP_NO;
1684                 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1685                         mlog(0, "migration mle was found (%u->%u)\n",
1686                              tmpmle->master, tmpmle->new_master);
1687                         /* real master can respond on its own */
1688                         response = DLM_MASTER_RESP_NO;
1689                 } else
1690                         response = DLM_MASTER_RESP_MAYBE;
1691                 if (set_maybe)
1692                         set_bit(request->node_idx, tmpmle->maybe_map);
1693                 spin_unlock(&tmpmle->spinlock);
1694         }
1695         spin_unlock(&dlm->master_lock);
1696         spin_unlock(&dlm->spinlock);
1697
1698         if (found) {
1699                 /* keep the mle attached to heartbeat events */
1700                 dlm_put_mle(tmpmle);
1701         }
1702 send_response:
1703         /*
1704          * __dlm_lookup_lockres() grabbed a reference to this lockres.
1705          * The reference is released by dlm_assert_master_worker() under
1706          * the call to dlm_dispatch_assert_master().  If
1707          * dlm_assert_master_worker() isn't called, we drop it here.
1708          */
1709         if (dispatch_assert) {
1710                 if (response != DLM_MASTER_RESP_YES)
1711                         mlog(ML_ERROR, "invalid response %d\n", response);
1712                 if (!res) {
1713                         mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1714                         BUG();
1715                 }
1716                 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1717                              dlm->node_num, res->lockname.len, res->lockname.name);
1718                 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1719                                                  DLM_ASSERT_MASTER_MLE_CLEANUP);
1720                 if (ret < 0) {
1721                         mlog(ML_ERROR, "failed to dispatch assert master work\n");
1722                         response = DLM_MASTER_RESP_ERROR;
1723                         dlm_lockres_put(res);
1724                 }
1725         } else {
1726                 if (res)
1727                         dlm_lockres_put(res);
1728         }
1729
1730         dlm_put(dlm);
1731         return response;
1732 }
1733
1734 /*
1735  * DLM_ASSERT_MASTER_MSG
1736  */
1737
1738
1739 /*
1740  * NOTE: this can be used for debugging
1741  * can periodically run all locks owned by this node
1742  * and re-assert across the cluster...
1743  */
1744 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1745                                 struct dlm_lock_resource *res,
1746                                 void *nodemap, u32 flags)
1747 {
1748         struct dlm_assert_master assert;
1749         int to, tmpret;
1750         struct dlm_node_iter iter;
1751         int ret = 0;
1752         int reassert;
1753         const char *lockname = res->lockname.name;
1754         unsigned int namelen = res->lockname.len;
1755
1756         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1757
1758         spin_lock(&res->spinlock);
1759         res->state |= DLM_LOCK_RES_SETREF_INPROG;
1760         spin_unlock(&res->spinlock);
1761
1762 again:
1763         reassert = 0;
1764
1765         /* note that if this nodemap is empty, it returns 0 */
1766         dlm_node_iter_init(nodemap, &iter);
1767         while ((to = dlm_node_iter_next(&iter)) >= 0) {
1768                 int r = 0;
1769                 struct dlm_master_list_entry *mle = NULL;
1770
1771                 mlog(0, "sending assert master to %d (%.*s)\n", to,
1772                      namelen, lockname);
1773                 memset(&assert, 0, sizeof(assert));
1774                 assert.node_idx = dlm->node_num;
1775                 assert.namelen = namelen;
1776                 memcpy(assert.name, lockname, namelen);
1777                 assert.flags = cpu_to_be32(flags);
1778
1779                 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1780                                             &assert, sizeof(assert), to, &r);
1781                 if (tmpret < 0) {
1782                         mlog(0, "assert_master returned %d!\n", tmpret);
1783                         if (!dlm_is_host_down(tmpret)) {
1784                                 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1785                                 BUG();
1786                         }
1787                         /* a node died.  finish out the rest of the nodes. */
1788                         mlog(0, "link to %d went down!\n", to);
1789                         /* any nonzero status return will do */
1790                         ret = tmpret;
1791                         r = 0;
1792                 } else if (r < 0) {
1793                         /* ok, something horribly messed.  kill thyself. */
1794                         mlog(ML_ERROR,"during assert master of %.*s to %u, "
1795                              "got %d.\n", namelen, lockname, to, r);
1796                         spin_lock(&dlm->spinlock);
1797                         spin_lock(&dlm->master_lock);
1798                         if (dlm_find_mle(dlm, &mle, (char *)lockname,
1799                                          namelen)) {
1800                                 dlm_print_one_mle(mle);
1801                                 __dlm_put_mle(mle);
1802                         }
1803                         spin_unlock(&dlm->master_lock);
1804                         spin_unlock(&dlm->spinlock);
1805                         BUG();
1806                 }
1807
1808                 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1809                     !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1810                                 mlog(ML_ERROR, "%.*s: very strange, "
1811                                      "master MLE but no lockres on %u\n",
1812                                      namelen, lockname, to);
1813                 }
1814
1815                 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1816                         mlog(0, "%.*s: node %u create mles on other "
1817                              "nodes and requests a re-assert\n",
1818                              namelen, lockname, to);
1819                         reassert = 1;
1820                 }
1821                 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1822                         mlog(0, "%.*s: node %u has a reference to this "
1823                              "lockres, set the bit in the refmap\n",
1824                              namelen, lockname, to);
1825                         spin_lock(&res->spinlock);
1826                         dlm_lockres_set_refmap_bit(to, res);
1827                         spin_unlock(&res->spinlock);
1828                 }
1829         }
1830
1831         if (reassert)
1832                 goto again;
1833
1834         spin_lock(&res->spinlock);
1835         res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1836         spin_unlock(&res->spinlock);
1837         wake_up(&res->wq);
1838
1839         return ret;
1840 }
1841
1842 /*
1843  * locks that can be taken here:
1844  * dlm->spinlock
1845  * res->spinlock
1846  * mle->spinlock
1847  * dlm->master_list
1848  *
1849  * if possible, TRIM THIS DOWN!!!
1850  */
1851 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1852                               void **ret_data)
1853 {
1854         struct dlm_ctxt *dlm = data;
1855         struct dlm_master_list_entry *mle = NULL;
1856         struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1857         struct dlm_lock_resource *res = NULL;
1858         char *name;
1859         unsigned int namelen, hash;
1860         u32 flags;
1861         int master_request = 0, have_lockres_ref = 0;
1862         int ret = 0;
1863
1864         if (!dlm_grab(dlm))
1865                 return 0;
1866
1867         name = assert->name;
1868         namelen = assert->namelen;
1869         hash = dlm_lockid_hash(name, namelen);
1870         flags = be32_to_cpu(assert->flags);
1871
1872         if (namelen > DLM_LOCKID_NAME_MAX) {
1873                 mlog(ML_ERROR, "Invalid name length!");
1874                 goto done;
1875         }
1876
1877         spin_lock(&dlm->spinlock);
1878
1879         if (flags)
1880                 mlog(0, "assert_master with flags: %u\n", flags);
1881
1882         /* find the MLE */
1883         spin_lock(&dlm->master_lock);
1884         if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1885                 /* not an error, could be master just re-asserting */
1886                 mlog(0, "just got an assert_master from %u, but no "
1887                      "MLE for it! (%.*s)\n", assert->node_idx,
1888                      namelen, name);
1889         } else {
1890                 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1891                 if (bit >= O2NM_MAX_NODES) {
1892                         /* not necessarily an error, though less likely.
1893                          * could be master just re-asserting. */
1894                         mlog(0, "no bits set in the maybe_map, but %u "
1895                              "is asserting! (%.*s)\n", assert->node_idx,
1896                              namelen, name);
1897                 } else if (bit != assert->node_idx) {
1898                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1899                                 mlog(0, "master %u was found, %u should "
1900                                      "back off\n", assert->node_idx, bit);
1901                         } else {
1902                                 /* with the fix for bug 569, a higher node
1903                                  * number winning the mastery will respond
1904                                  * YES to mastery requests, but this node
1905                                  * had no way of knowing.  let it pass. */
1906                                 mlog(0, "%u is the lowest node, "
1907                                      "%u is asserting. (%.*s)  %u must "
1908                                      "have begun after %u won.\n", bit,
1909                                      assert->node_idx, namelen, name, bit,
1910                                      assert->node_idx);
1911                         }
1912                 }
1913                 if (mle->type == DLM_MLE_MIGRATION) {
1914                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1915                                 mlog(0, "%s:%.*s: got cleanup assert"
1916                                      " from %u for migration\n",
1917                                      dlm->name, namelen, name,
1918                                      assert->node_idx);
1919                         } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1920                                 mlog(0, "%s:%.*s: got unrelated assert"
1921                                      " from %u for migration, ignoring\n",
1922                                      dlm->name, namelen, name,
1923                                      assert->node_idx);
1924                                 __dlm_put_mle(mle);
1925                                 spin_unlock(&dlm->master_lock);
1926                                 spin_unlock(&dlm->spinlock);
1927                                 goto done;
1928                         }
1929                 }
1930         }
1931         spin_unlock(&dlm->master_lock);
1932
1933         /* ok everything checks out with the MLE
1934          * now check to see if there is a lockres */
1935         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1936         if (res) {
1937                 spin_lock(&res->spinlock);
1938                 if (res->state & DLM_LOCK_RES_RECOVERING)  {
1939                         mlog(ML_ERROR, "%u asserting but %.*s is "
1940                              "RECOVERING!\n", assert->node_idx, namelen, name);
1941                         goto kill;
1942                 }
1943                 if (!mle) {
1944                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1945                             res->owner != assert->node_idx) {
1946                                 mlog(ML_ERROR, "assert_master from "
1947                                           "%u, but current owner is "
1948                                           "%u! (%.*s)\n",
1949                                        assert->node_idx, res->owner,
1950                                        namelen, name);
1951                                 goto kill;
1952                         }
1953                 } else if (mle->type != DLM_MLE_MIGRATION) {
1954                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1955                                 /* owner is just re-asserting */
1956                                 if (res->owner == assert->node_idx) {
1957                                         mlog(0, "owner %u re-asserting on "
1958                                              "lock %.*s\n", assert->node_idx,
1959                                              namelen, name);
1960                                         goto ok;
1961                                 }
1962                                 mlog(ML_ERROR, "got assert_master from "
1963                                      "node %u, but %u is the owner! "
1964                                      "(%.*s)\n", assert->node_idx,
1965                                      res->owner, namelen, name);
1966                                 goto kill;
1967                         }
1968                         if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1969                                 mlog(ML_ERROR, "got assert from %u, but lock "
1970                                      "with no owner should be "
1971                                      "in-progress! (%.*s)\n",
1972                                      assert->node_idx,
1973                                      namelen, name);
1974                                 goto kill;
1975                         }
1976                 } else /* mle->type == DLM_MLE_MIGRATION */ {
1977                         /* should only be getting an assert from new master */
1978                         if (assert->node_idx != mle->new_master) {
1979                                 mlog(ML_ERROR, "got assert from %u, but "
1980                                      "new master is %u, and old master "
1981                                      "was %u (%.*s)\n",
1982                                      assert->node_idx, mle->new_master,
1983                                      mle->master, namelen, name);
1984                                 goto kill;
1985                         }
1986
1987                 }
1988 ok:
1989                 spin_unlock(&res->spinlock);
1990         }
1991         spin_unlock(&dlm->spinlock);
1992
1993         // mlog(0, "woo!  got an assert_master from node %u!\n",
1994         //           assert->node_idx);
1995         if (mle) {
1996                 int extra_ref = 0;
1997                 int nn = -1;
1998                 int rr, err = 0;
1999
2000                 spin_lock(&mle->spinlock);
2001                 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
2002                         extra_ref = 1;
2003                 else {
2004                         /* MASTER mle: if any bits set in the response map
2005                          * then the calling node needs to re-assert to clear
2006                          * up nodes that this node contacted */
2007                         while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
2008                                                     nn+1)) < O2NM_MAX_NODES) {
2009                                 if (nn != dlm->node_num && nn != assert->node_idx)
2010                                         master_request = 1;
2011                         }
2012                 }
2013                 mle->master = assert->node_idx;
2014                 atomic_set(&mle->woken, 1);
2015                 wake_up(&mle->wq);
2016                 spin_unlock(&mle->spinlock);
2017
2018                 if (res) {
2019                         int wake = 0;
2020                         spin_lock(&res->spinlock);
2021                         if (mle->type == DLM_MLE_MIGRATION) {
2022                                 mlog(0, "finishing off migration of lockres %.*s, "
2023                                         "from %u to %u\n",
2024                                         res->lockname.len, res->lockname.name,
2025                                         dlm->node_num, mle->new_master);
2026                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2027                                 wake = 1;
2028                                 dlm_change_lockres_owner(dlm, res, mle->new_master);
2029                                 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
2030                         } else {
2031                                 dlm_change_lockres_owner(dlm, res, mle->master);
2032                         }
2033                         spin_unlock(&res->spinlock);
2034                         have_lockres_ref = 1;
2035                         if (wake)
2036                                 wake_up(&res->wq);
2037                 }
2038
2039                 /* master is known, detach if not already detached.
2040                  * ensures that only one assert_master call will happen
2041                  * on this mle. */
2042                 spin_lock(&dlm->spinlock);
2043                 spin_lock(&dlm->master_lock);
2044
2045                 rr = atomic_read(&mle->mle_refs.refcount);
2046                 if (mle->inuse > 0) {
2047                         if (extra_ref && rr < 3)
2048                                 err = 1;
2049                         else if (!extra_ref && rr < 2)
2050                                 err = 1;
2051                 } else {
2052                         if (extra_ref && rr < 2)
2053                                 err = 1;
2054                         else if (!extra_ref && rr < 1)
2055                                 err = 1;
2056                 }
2057                 if (err) {
2058                         mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
2059                              "that will mess up this node, refs=%d, extra=%d, "
2060                              "inuse=%d\n", dlm->name, namelen, name,
2061                              assert->node_idx, rr, extra_ref, mle->inuse);
2062                         dlm_print_one_mle(mle);
2063                 }
2064                 list_del_init(&mle->list);
2065                 __dlm_mle_detach_hb_events(dlm, mle);
2066                 __dlm_put_mle(mle);
2067                 if (extra_ref) {
2068                         /* the assert master message now balances the extra
2069                          * ref given by the master / migration request message.
2070                          * if this is the last put, it will be removed
2071                          * from the list. */
2072                         __dlm_put_mle(mle);
2073                 }
2074                 spin_unlock(&dlm->master_lock);
2075                 spin_unlock(&dlm->spinlock);
2076         } else if (res) {
2077                 if (res->owner != assert->node_idx) {
2078                         mlog(0, "assert_master from %u, but current "
2079                              "owner is %u (%.*s), no mle\n", assert->node_idx,
2080                              res->owner, namelen, name);
2081                 }
2082         }
2083
2084 done:
2085         ret = 0;
2086         if (res) {
2087                 spin_lock(&res->spinlock);
2088                 res->state |= DLM_LOCK_RES_SETREF_INPROG;
2089                 spin_unlock(&res->spinlock);
2090                 *ret_data = (void *)res;
2091         }
2092         dlm_put(dlm);
2093         if (master_request) {
2094                 mlog(0, "need to tell master to reassert\n");
2095                 /* positive. negative would shoot down the node. */
2096                 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2097                 if (!have_lockres_ref) {
2098                         mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2099                              "mle present here for %s:%.*s, but no lockres!\n",
2100                              assert->node_idx, dlm->name, namelen, name);
2101                 }
2102         }
2103         if (have_lockres_ref) {
2104                 /* let the master know we have a reference to the lockres */
2105                 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2106                 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2107                      dlm->name, namelen, name, assert->node_idx);
2108         }
2109         return ret;
2110
2111 kill:
2112         /* kill the caller! */
2113         mlog(ML_ERROR, "Bad message received from another node.  Dumping state "
2114              "and killing the other node now!  This node is OK and can continue.\n");
2115         __dlm_print_one_lock_resource(res);
2116         spin_unlock(&res->spinlock);
2117         spin_unlock(&dlm->spinlock);
2118         *ret_data = (void *)res;
2119         dlm_put(dlm);
2120         return -EINVAL;
2121 }
2122
2123 void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2124 {
2125         struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2126
2127         if (ret_data) {
2128                 spin_lock(&res->spinlock);
2129                 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2130                 spin_unlock(&res->spinlock);
2131                 wake_up(&res->wq);
2132                 dlm_lockres_put(res);
2133         }
2134         return;
2135 }
2136
2137 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2138                                struct dlm_lock_resource *res,
2139                                int ignore_higher, u8 request_from, u32 flags)
2140 {
2141         struct dlm_work_item *item;
2142         item = kzalloc(sizeof(*item), GFP_NOFS);
2143         if (!item)
2144                 return -ENOMEM;
2145
2146
2147         /* queue up work for dlm_assert_master_worker */
2148         dlm_grab(dlm);  /* get an extra ref for the work item */
2149         dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2150         item->u.am.lockres = res; /* already have a ref */
2151         /* can optionally ignore node numbers higher than this node */
2152         item->u.am.ignore_higher = ignore_higher;
2153         item->u.am.request_from = request_from;
2154         item->u.am.flags = flags;
2155
2156         if (ignore_higher)
2157                 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2158                      res->lockname.name);
2159
2160         spin_lock(&dlm->work_lock);
2161         list_add_tail(&item->list, &dlm->work_list);
2162         spin_unlock(&dlm->work_lock);
2163
2164         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2165         return 0;
2166 }
2167
2168 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2169 {
2170         struct dlm_ctxt *dlm = data;
2171         int ret = 0;
2172         struct dlm_lock_resource *res;
2173         unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2174         int ignore_higher;
2175         int bit;
2176         u8 request_from;
2177         u32 flags;
2178
2179         dlm = item->dlm;
2180         res = item->u.am.lockres;
2181         ignore_higher = item->u.am.ignore_higher;
2182         request_from = item->u.am.request_from;
2183         flags = item->u.am.flags;
2184
2185         spin_lock(&dlm->spinlock);
2186         memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2187         spin_unlock(&dlm->spinlock);
2188
2189         clear_bit(dlm->node_num, nodemap);
2190         if (ignore_higher) {
2191                 /* if is this just to clear up mles for nodes below
2192                  * this node, do not send the message to the original
2193                  * caller or any node number higher than this */
2194                 clear_bit(request_from, nodemap);
2195                 bit = dlm->node_num;
2196                 while (1) {
2197                         bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2198                                             bit+1);
2199                         if (bit >= O2NM_MAX_NODES)
2200                                 break;
2201                         clear_bit(bit, nodemap);
2202                 }
2203         }
2204
2205         /*
2206          * If we're migrating this lock to someone else, we are no
2207          * longer allowed to assert out own mastery.  OTOH, we need to
2208          * prevent migration from starting while we're still asserting
2209          * our dominance.  The reserved ast delays migration.
2210          */
2211         spin_lock(&res->spinlock);
2212         if (res->state & DLM_LOCK_RES_MIGRATING) {
2213                 mlog(0, "Someone asked us to assert mastery, but we're "
2214                      "in the middle of migration.  Skipping assert, "
2215                      "the new master will handle that.\n");
2216                 spin_unlock(&res->spinlock);
2217                 goto put;
2218         } else
2219                 __dlm_lockres_reserve_ast(res);
2220         spin_unlock(&res->spinlock);
2221
2222         /* this call now finishes out the nodemap
2223          * even if one or more nodes die */
2224         mlog(0, "worker about to master %.*s here, this=%u\n",
2225                      res->lockname.len, res->lockname.name, dlm->node_num);
2226         ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2227         if (ret < 0) {
2228                 /* no need to restart, we are done */
2229                 if (!dlm_is_host_down(ret))
2230                         mlog_errno(ret);
2231         }
2232
2233         /* Ok, we've asserted ourselves.  Let's let migration start. */
2234         dlm_lockres_release_ast(dlm, res);
2235
2236 put:
2237         dlm_lockres_put(res);
2238
2239         mlog(0, "finished with dlm_assert_master_worker\n");
2240 }
2241
2242 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2243  * We cannot wait for node recovery to complete to begin mastering this
2244  * lockres because this lockres is used to kick off recovery! ;-)
2245  * So, do a pre-check on all living nodes to see if any of those nodes
2246  * think that $RECOVERY is currently mastered by a dead node.  If so,
2247  * we wait a short time to allow that node to get notified by its own
2248  * heartbeat stack, then check again.  All $RECOVERY lock resources
2249  * mastered by dead nodes are purged when the hearbeat callback is
2250  * fired, so we can know for sure that it is safe to continue once
2251  * the node returns a live node or no node.  */
2252 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2253                                        struct dlm_lock_resource *res)
2254 {
2255         struct dlm_node_iter iter;
2256         int nodenum;
2257         int ret = 0;
2258         u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2259
2260         spin_lock(&dlm->spinlock);
2261         dlm_node_iter_init(dlm->domain_map, &iter);
2262         spin_unlock(&dlm->spinlock);
2263
2264         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2265                 /* do not send to self */
2266                 if (nodenum == dlm->node_num)
2267                         continue;
2268                 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2269                 if (ret < 0) {
2270                         mlog_errno(ret);
2271                         if (!dlm_is_host_down(ret))
2272                                 BUG();
2273                         /* host is down, so answer for that node would be
2274                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
2275                         ret = 0;
2276                 }
2277
2278                 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2279                         /* check to see if this master is in the recovery map */
2280                         spin_lock(&dlm->spinlock);
2281                         if (test_bit(master, dlm->recovery_map)) {
2282                                 mlog(ML_NOTICE, "%s: node %u has not seen "
2283                                      "node %u go down yet, and thinks the "
2284                                      "dead node is mastering the recovery "
2285                                      "lock.  must wait.\n", dlm->name,
2286                                      nodenum, master);
2287                                 ret = -EAGAIN;
2288                         }
2289                         spin_unlock(&dlm->spinlock);
2290                         mlog(0, "%s: reco lock master is %u\n", dlm->name,
2291                              master);
2292                         break;
2293                 }
2294         }
2295         return ret;
2296 }
2297
2298 /*
2299  * DLM_DEREF_LOCKRES_MSG
2300  */
2301
2302 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2303 {
2304         struct dlm_deref_lockres deref;
2305         int ret = 0, r;
2306         const char *lockname;
2307         unsigned int namelen;
2308
2309         lockname = res->lockname.name;
2310         namelen = res->lockname.len;
2311         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2312
2313         mlog(0, "%s:%.*s: sending deref to %d\n",
2314              dlm->name, namelen, lockname, res->owner);
2315         memset(&deref, 0, sizeof(deref));
2316         deref.node_idx = dlm->node_num;
2317         deref.namelen = namelen;
2318         memcpy(deref.name, lockname, namelen);
2319
2320         ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2321                                  &deref, sizeof(deref), res->owner, &r);
2322         if (ret < 0)
2323                 mlog_errno(ret);
2324         else if (r < 0) {
2325                 /* BAD.  other node says I did not have a ref. */
2326                 mlog(ML_ERROR,"while dropping ref on %s:%.*s "
2327                     "(master=%u) got %d.\n", dlm->name, namelen,
2328                     lockname, res->owner, r);
2329                 dlm_print_one_lock_resource(res);
2330                 BUG();
2331         }
2332         return ret;
2333 }
2334
2335 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2336                               void **ret_data)
2337 {
2338         struct dlm_ctxt *dlm = data;
2339         struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2340         struct dlm_lock_resource *res = NULL;
2341         char *name;
2342         unsigned int namelen;
2343         int ret = -EINVAL;
2344         u8 node;
2345         unsigned int hash;
2346         struct dlm_work_item *item;
2347         int cleared = 0;
2348         int dispatch = 0;
2349
2350         if (!dlm_grab(dlm))
2351                 return 0;
2352
2353         name = deref->name;
2354         namelen = deref->namelen;
2355         node = deref->node_idx;
2356
2357         if (namelen > DLM_LOCKID_NAME_MAX) {
2358                 mlog(ML_ERROR, "Invalid name length!");
2359                 goto done;
2360         }
2361         if (deref->node_idx >= O2NM_MAX_NODES) {
2362                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2363                 goto done;
2364         }
2365
2366         hash = dlm_lockid_hash(name, namelen);
2367
2368         spin_lock(&dlm->spinlock);
2369         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2370         if (!res) {
2371                 spin_unlock(&dlm->spinlock);
2372                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2373                      dlm->name, namelen, name);
2374                 goto done;
2375         }
2376         spin_unlock(&dlm->spinlock);
2377
2378         spin_lock(&res->spinlock);
2379         if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2380                 dispatch = 1;
2381         else {
2382                 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2383                 if (test_bit(node, res->refmap)) {
2384                         dlm_lockres_clear_refmap_bit(node, res);
2385                         cleared = 1;
2386                 }
2387         }
2388         spin_unlock(&res->spinlock);
2389
2390         if (!dispatch) {
2391                 if (cleared)
2392                         dlm_lockres_calc_usage(dlm, res);
2393                 else {
2394                         mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2395                         "but it is already dropped!\n", dlm->name,
2396                         res->lockname.len, res->lockname.name, node);
2397                         dlm_print_one_lock_resource(res);
2398                 }
2399                 ret = 0;
2400                 goto done;
2401         }
2402
2403         item = kzalloc(sizeof(*item), GFP_NOFS);
2404         if (!item) {
2405                 ret = -ENOMEM;
2406                 mlog_errno(ret);
2407                 goto done;
2408         }
2409
2410         dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2411         item->u.dl.deref_res = res;
2412         item->u.dl.deref_node = node;
2413
2414         spin_lock(&dlm->work_lock);
2415         list_add_tail(&item->list, &dlm->work_list);
2416         spin_unlock(&dlm->work_lock);
2417
2418         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2419         return 0;
2420
2421 done:
2422         if (res)
2423                 dlm_lockres_put(res);
2424         dlm_put(dlm);
2425
2426         return ret;
2427 }
2428
2429 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2430 {
2431         struct dlm_ctxt *dlm;
2432         struct dlm_lock_resource *res;
2433         u8 node;
2434         u8 cleared = 0;
2435
2436         dlm = item->dlm;
2437         res = item->u.dl.deref_res;
2438         node = item->u.dl.deref_node;
2439
2440         spin_lock(&res->spinlock);
2441         BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2442         if (test_bit(node, res->refmap)) {
2443                 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2444                 dlm_lockres_clear_refmap_bit(node, res);
2445                 cleared = 1;
2446         }
2447         spin_unlock(&res->spinlock);
2448
2449         if (cleared) {
2450                 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2451                      dlm->name, res->lockname.len, res->lockname.name, node);
2452                 dlm_lockres_calc_usage(dlm, res);
2453         } else {
2454                 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2455                      "but it is already dropped!\n", dlm->name,
2456                      res->lockname.len, res->lockname.name, node);
2457                 dlm_print_one_lock_resource(res);
2458         }
2459
2460         dlm_lockres_put(res);
2461 }
2462
2463 /* Checks whether the lockres can be migrated. Returns 0 if yes, < 0
2464  * if not. If 0, numlocks is set to the number of locks in the lockres.
2465  */
2466 static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
2467                                       struct dlm_lock_resource *res,
2468                                       int *numlocks)
2469 {
2470         int ret;
2471         int i;
2472         int count = 0;
2473         struct list_head *queue;
2474         struct dlm_lock *lock;
2475
2476         assert_spin_locked(&res->spinlock);
2477
2478         ret = -EINVAL;
2479         if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
2480                 mlog(0, "cannot migrate lockres with unknown owner!\n");
2481                 goto leave;
2482         }
2483
2484         if (res->owner != dlm->node_num) {
2485                 mlog(0, "cannot migrate lockres this node doesn't own!\n");
2486                 goto leave;
2487         }
2488
2489         ret = 0;
2490         queue = &res->granted;
2491         for (i = 0; i < 3; i++) {
2492                 list_for_each_entry(lock, queue, list) {
2493                         ++count;
2494                         if (lock->ml.node == dlm->node_num) {
2495                                 mlog(0, "found a lock owned by this node still "
2496                                      "on the %s queue!  will not migrate this "
2497                                      "lockres\n", (i == 0 ? "granted" :
2498                                                    (i == 1 ? "converting" :
2499                                                     "blocked")));
2500                                 ret = -ENOTEMPTY;
2501                                 goto leave;
2502                         }
2503                 }
2504                 queue++;
2505         }
2506
2507         *numlocks = count;
2508         mlog(0, "migrateable lockres having %d locks\n", *numlocks);
2509
2510 leave:
2511         return ret;
2512 }
2513
2514 /*
2515  * DLM_MIGRATE_LOCKRES
2516  */
2517
2518
2519 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2520                                struct dlm_lock_resource *res,
2521                                u8 target)
2522 {
2523         struct dlm_master_list_entry *mle = NULL;
2524         struct dlm_master_list_entry *oldmle = NULL;
2525         struct dlm_migratable_lockres *mres = NULL;
2526         int ret = 0;
2527         const char *name;
2528         unsigned int namelen;
2529         int mle_added = 0;
2530         int numlocks;
2531         int wake = 0;
2532
2533         if (!dlm_grab(dlm))
2534                 return -EINVAL;
2535
2536         name = res->lockname.name;
2537         namelen = res->lockname.len;
2538
2539         mlog(0, "migrating %.*s to %u\n", namelen, name, target);
2540
2541         /*
2542          * ensure this lockres is a proper candidate for migration
2543          */
2544         spin_lock(&res->spinlock);
2545         ret = dlm_is_lockres_migrateable(dlm, res, &numlocks);
2546         if (ret < 0) {
2547                 spin_unlock(&res->spinlock);
2548                 goto leave;
2549         }
2550         spin_unlock(&res->spinlock);
2551
2552         /* no work to do */
2553         if (numlocks == 0) {
2554                 mlog(0, "no locks were found on this lockres! done!\n");
2555                 goto leave;
2556         }
2557
2558         /*
2559          * preallocate up front
2560          * if this fails, abort
2561          */
2562
2563         ret = -ENOMEM;
2564         mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2565         if (!mres) {
2566                 mlog_errno(ret);
2567                 goto leave;
2568         }
2569
2570         mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
2571                                                                 GFP_NOFS);
2572         if (!mle) {
2573                 mlog_errno(ret);
2574                 goto leave;
2575         }
2576         ret = 0;
2577
2578         /*
2579          * find a node to migrate the lockres to
2580          */
2581
2582         mlog(0, "picking a migration node\n");
2583         spin_lock(&dlm->spinlock);
2584         /* pick a new node */
2585         if (!test_bit(target, dlm->domain_map) ||
2586             target >= O2NM_MAX_NODES) {
2587                 target = dlm_pick_migration_target(dlm, res);
2588         }
2589         mlog(0, "node %u chosen for migration\n", target);
2590
2591         if (target >= O2NM_MAX_NODES ||
2592             !test_bit(target, dlm->domain_map)) {
2593                 /* target chosen is not alive */
2594                 ret = -EINVAL;
2595         }
2596
2597         if (ret) {
2598                 spin_unlock(&dlm->spinlock);
2599                 goto fail;
2600         }
2601
2602         mlog(0, "continuing with target = %u\n", target);
2603
2604         /*
2605          * clear any existing master requests and
2606          * add the migration mle to the list
2607          */
2608         spin_lock(&dlm->master_lock);
2609         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2610                                     namelen, target, dlm->node_num);
2611         spin_unlock(&dlm->master_lock);
2612         spin_unlock(&dlm->spinlock);
2613
2614         if (ret == -EEXIST) {
2615                 mlog(0, "another process is already migrating it\n");
2616                 goto fail;
2617         }
2618         mle_added = 1;
2619
2620         /*
2621          * set the MIGRATING flag and flush asts
2622          * if we fail after this we need to re-dirty the lockres
2623          */
2624         if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2625                 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2626                      "the target went down.\n", res->lockname.len,
2627                      res->lockname.name, target);
2628                 spin_lock(&res->spinlock);
2629                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2630                 wake = 1;
2631                 spin_unlock(&res->spinlock);
2632                 ret = -EINVAL;
2633         }
2634
2635 fail:
2636         if (oldmle) {
2637                 /* master is known, detach if not already detached */
2638                 dlm_mle_detach_hb_events(dlm, oldmle);
2639                 dlm_put_mle(oldmle);
2640         }
2641
2642         if (ret < 0) {
2643                 if (mle_added) {
2644                         dlm_mle_detach_hb_events(dlm, mle);
2645                         dlm_put_mle(mle);
2646                 } else if (mle) {
2647                         kmem_cache_free(dlm_mle_cache, mle);
2648                 }
2649                 goto leave;
2650         }
2651
2652         /*
2653          * at this point, we have a migration target, an mle
2654          * in the master list, and the MIGRATING flag set on
2655          * the lockres
2656          */
2657
2658         /* now that remote nodes are spinning on the MIGRATING flag,
2659          * ensure that all assert_master work is flushed. */
2660         flush_workqueue(dlm->dlm_worker);
2661
2662         /* get an extra reference on the mle.
2663          * otherwise the assert_master from the new
2664          * master will destroy this.
2665          * also, make sure that all callers of dlm_get_mle
2666          * take both dlm->spinlock and dlm->master_lock */
2667         spin_lock(&dlm->spinlock);
2668         spin_lock(&dlm->master_lock);
2669         dlm_get_mle_inuse(mle);
2670         spin_unlock(&dlm->master_lock);
2671         spin_unlock(&dlm->spinlock);
2672
2673         /* notify new node and send all lock state */
2674         /* call send_one_lockres with migration flag.
2675          * this serves as notice to the target node that a
2676          * migration is starting. */
2677         ret = dlm_send_one_lockres(dlm, res, mres, target,
2678                                    DLM_MRES_MIGRATION);
2679
2680         if (ret < 0) {
2681                 mlog(0, "migration to node %u failed with %d\n",
2682                      target, ret);
2683                 /* migration failed, detach and clean up mle */
2684                 dlm_mle_detach_hb_events(dlm, mle);
2685                 dlm_put_mle(mle);
2686                 dlm_put_mle_inuse(mle);
2687                 spin_lock(&res->spinlock);
2688                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2689                 wake = 1;
2690                 spin_unlock(&res->spinlock);
2691                 goto leave;
2692         }
2693
2694         /* at this point, the target sends a message to all nodes,
2695          * (using dlm_do_migrate_request).  this node is skipped since
2696          * we had to put an mle in the list to begin the process.  this
2697          * node now waits for target to do an assert master.  this node
2698          * will be the last one notified, ensuring that the migration
2699          * is complete everywhere.  if the target dies while this is
2700          * going on, some nodes could potentially see the target as the
2701          * master, so it is important that my recovery finds the migration
2702          * mle and sets the master to UNKNONWN. */
2703
2704
2705         /* wait for new node to assert master */
2706         while (1) {
2707                 ret = wait_event_interruptible_timeout(mle->wq,
2708                                         (atomic_read(&mle->woken) == 1),
2709                                         msecs_to_jiffies(5000));
2710
2711                 if (ret >= 0) {
2712                         if (atomic_read(&mle->woken) == 1 ||
2713                             res->owner == target)
2714                                 break;
2715
2716                         mlog(0, "%s:%.*s: timed out during migration\n",
2717                              dlm->name, res->lockname.len, res->lockname.name);
2718                         /* avoid hang during shutdown when migrating lockres
2719                          * to a node which also goes down */
2720                         if (dlm_is_node_dead(dlm, target)) {
2721                                 mlog(0, "%s:%.*s: expected migration "
2722                                      "target %u is no longer up, restarting\n",
2723                                      dlm->name, res->lockname.len,
2724                                      res->lockname.name, target);
2725                                 ret = -EINVAL;
2726                                 /* migration failed, detach and clean up mle */
2727                                 dlm_mle_detach_hb_events(dlm, mle);
2728                                 dlm_put_mle(mle);
2729                                 dlm_put_mle_inuse(mle);
2730                                 spin_lock(&res->spinlock);
2731                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2732                                 wake = 1;
2733                                 spin_unlock(&res->spinlock);
2734                                 goto leave;
2735                         }
2736                 } else
2737                         mlog(0, "%s:%.*s: caught signal during migration\n",
2738                              dlm->name, res->lockname.len, res->lockname.name);
2739         }
2740
2741         /* all done, set the owner, clear the flag */
2742         spin_lock(&res->spinlock);
2743         dlm_set_lockres_owner(dlm, res, target);
2744         res->state &= ~DLM_LOCK_RES_MIGRATING;
2745         dlm_remove_nonlocal_locks(dlm, res);
2746         spin_unlock(&res->spinlock);
2747         wake_up(&res->wq);
2748
2749         /* master is known, detach if not already detached */
2750         dlm_mle_detach_hb_events(dlm, mle);
2751         dlm_put_mle_inuse(mle);
2752         ret = 0;
2753
2754         dlm_lockres_calc_usage(dlm, res);
2755
2756 leave:
2757         /* re-dirty the lockres if we failed */
2758         if (ret < 0)
2759                 dlm_kick_thread(dlm, res);
2760
2761         /* wake up waiters if the MIGRATING flag got set
2762          * but migration failed */
2763         if (wake)
2764                 wake_up(&res->wq);
2765
2766         /* TODO: cleanup */
2767         if (mres)
2768                 free_page((unsigned long)mres);
2769
2770         dlm_put(dlm);
2771
2772         mlog(0, "returning %d\n", ret);
2773         return ret;
2774 }
2775
2776 #define DLM_MIGRATION_RETRY_MS  100
2777
2778 /* Should be called only after beginning the domain leave process.
2779  * There should not be any remaining locks on nonlocal lock resources,
2780  * and there should be no local locks left on locally mastered resources.
2781  *
2782  * Called with the dlm spinlock held, may drop it to do migration, but
2783  * will re-acquire before exit.
2784  *
2785  * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped */
2786 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2787 {
2788         int ret;
2789         int lock_dropped = 0;
2790         int numlocks;
2791
2792         spin_lock(&res->spinlock);
2793         if (res->owner != dlm->node_num) {
2794                 if (!__dlm_lockres_unused(res)) {
2795                         mlog(ML_ERROR, "%s:%.*s: this node is not master, "
2796                              "trying to free this but locks remain\n",
2797                              dlm->name, res->lockname.len, res->lockname.name);
2798                 }
2799                 spin_unlock(&res->spinlock);
2800                 goto leave;
2801         }
2802
2803         /* No need to migrate a lockres having no locks */
2804         ret = dlm_is_lockres_migrateable(dlm, res, &numlocks);
2805         if (ret >= 0 && numlocks == 0) {
2806                 spin_unlock(&res->spinlock);
2807                 goto leave;
2808         }
2809         spin_unlock(&res->spinlock);
2810
2811         /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2812         spin_unlock(&dlm->spinlock);
2813         lock_dropped = 1;
2814         while (1) {
2815                 ret = dlm_migrate_lockres(dlm, res, O2NM_MAX_NODES);
2816                 if (ret >= 0)
2817                         break;
2818                 if (ret == -ENOTEMPTY) {
2819                         mlog(ML_ERROR, "lockres %.*s still has local locks!\n",
2820                                 res->lockname.len, res->lockname.name);
2821                         BUG();
2822                 }
2823
2824                 mlog(0, "lockres %.*s: migrate failed, "
2825                      "retrying\n", res->lockname.len,
2826                      res->lockname.name);
2827                 msleep(DLM_MIGRATION_RETRY_MS);
2828         }
2829         spin_lock(&dlm->spinlock);
2830 leave:
2831         return lock_dropped;
2832 }
2833
2834 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2835 {
2836         int ret;
2837         spin_lock(&dlm->ast_lock);
2838         spin_lock(&lock->spinlock);
2839         ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2840         spin_unlock(&lock->spinlock);
2841         spin_unlock(&dlm->ast_lock);
2842         return ret;
2843 }
2844
2845 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2846                                      struct dlm_lock_resource *res,
2847                                      u8 mig_target)
2848 {
2849         int can_proceed;
2850         spin_lock(&res->spinlock);
2851         can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2852         spin_unlock(&res->spinlock);
2853
2854         /* target has died, so make the caller break out of the
2855          * wait_event, but caller must recheck the domain_map */
2856         spin_lock(&dlm->spinlock);
2857         if (!test_bit(mig_target, dlm->domain_map))
2858                 can_proceed = 1;
2859         spin_unlock(&dlm->spinlock);
2860         return can_proceed;
2861 }
2862
2863 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2864                                 struct dlm_lock_resource *res)
2865 {
2866         int ret;
2867         spin_lock(&res->spinlock);
2868         ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2869         spin_unlock(&res->spinlock);
2870         return ret;
2871 }
2872
2873
2874 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2875                                        struct dlm_lock_resource *res,
2876                                        u8 target)
2877 {
2878         int ret = 0;
2879
2880         mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2881                res->lockname.len, res->lockname.name, dlm->node_num,
2882                target);
2883         /* need to set MIGRATING flag on lockres.  this is done by
2884          * ensuring that all asts have been flushed for this lockres. */
2885         spin_lock(&res->spinlock);
2886         BUG_ON(res->migration_pending);
2887         res->migration_pending = 1;
2888         /* strategy is to reserve an extra ast then release
2889          * it below, letting the release do all of the work */
2890         __dlm_lockres_reserve_ast(res);
2891         spin_unlock(&res->spinlock);
2892
2893         /* now flush all the pending asts */
2894         dlm_kick_thread(dlm, res);
2895         /* before waiting on DIRTY, block processes which may
2896          * try to dirty the lockres before MIGRATING is set */
2897         spin_lock(&res->spinlock);
2898         BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2899         res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2900         spin_unlock(&res->spinlock);
2901         /* now wait on any pending asts and the DIRTY state */
2902         wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2903         dlm_lockres_release_ast(dlm, res);
2904
2905         mlog(0, "about to wait on migration_wq, dirty=%s\n",
2906                res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2907         /* if the extra ref we just put was the final one, this
2908          * will pass thru immediately.  otherwise, we need to wait
2909          * for the last ast to finish. */
2910 again:
2911         ret = wait_event_interruptible_timeout(dlm->migration_wq,
2912                    dlm_migration_can_proceed(dlm, res, target),
2913                    msecs_to_jiffies(1000));
2914         if (ret < 0) {
2915                 mlog(0, "woken again: migrating? %s, dead? %s\n",
2916                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2917                        test_bit(target, dlm->domain_map) ? "no":"yes");
2918         } else {
2919                 mlog(0, "all is well: migrating? %s, dead? %s\n",
2920                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2921                        test_bit(target, dlm->domain_map) ? "no":"yes");
2922         }
2923         if (!dlm_migration_can_proceed(dlm, res, target)) {
2924                 mlog(0, "trying again...\n");
2925                 goto again;
2926         }
2927         /* now that we are sure the MIGRATING state is there, drop
2928          * the unneded state which blocked threads trying to DIRTY */
2929         spin_lock(&res->spinlock);
2930         BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2931         BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2932         res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2933         spin_unlock(&res->spinlock);
2934
2935         /* did the target go down or die? */
2936         spin_lock(&dlm->spinlock);
2937         if (!test_bit(target, dlm->domain_map)) {
2938                 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2939                      target);
2940                 ret = -EHOSTDOWN;
2941         }
2942         spin_unlock(&dlm->spinlock);
2943
2944         /*
2945          * at this point:
2946          *
2947          *   o the DLM_LOCK_RES_MIGRATING flag is set
2948          *   o there are no pending asts on this lockres
2949          *   o all processes trying to reserve an ast on this
2950          *     lockres must wait for the MIGRATING flag to clear
2951          */
2952         return ret;
2953 }
2954
2955 /* last step in the migration process.
2956  * original master calls this to free all of the dlm_lock
2957  * structures that used to be for other nodes. */
2958 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2959                                       struct dlm_lock_resource *res)
2960 {
2961         struct list_head *queue = &res->granted;
2962         int i, bit;
2963         struct dlm_lock *lock, *next;
2964
2965         assert_spin_locked(&res->spinlock);
2966
2967         BUG_ON(res->owner == dlm->node_num);
2968
2969         for (i=0; i<3; i++) {
2970                 list_for_each_entry_safe(lock, next, queue, list) {
2971                         if (lock->ml.node != dlm->node_num) {
2972                                 mlog(0, "putting lock for node %u\n",
2973                                      lock->ml.node);
2974                                 /* be extra careful */
2975                                 BUG_ON(!list_empty(&lock->ast_list));
2976                                 BUG_ON(!list_empty(&lock->bast_list));
2977                                 BUG_ON(lock->ast_pending);
2978                                 BUG_ON(lock->bast_pending);
2979                                 dlm_lockres_clear_refmap_bit(lock->ml.node, res);
2980                                 list_del_init(&lock->list);
2981                                 dlm_lock_put(lock);
2982                                 /* In a normal unlock, we would have added a
2983                                  * DLM_UNLOCK_FREE_LOCK action. Force it. */
2984                                 dlm_lock_put(lock);
2985                         }
2986                 }
2987                 queue++;
2988         }
2989         bit = 0;
2990         while (1) {
2991                 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2992                 if (bit >= O2NM_MAX_NODES)
2993                         break;
2994                 /* do not clear the local node reference, if there is a
2995                  * process holding this, let it drop the ref itself */
2996                 if (bit != dlm->node_num) {
2997                         mlog(0, "%s:%.*s: node %u had a ref to this "
2998                              "migrating lockres, clearing\n", dlm->name,
2999                              res->lockname.len, res->lockname.name, bit);
3000                         dlm_lockres_clear_refmap_bit(bit, res);
3001                 }
3002                 bit++;
3003         }
3004 }
3005
3006 /* for now this is not too intelligent.  we will
3007  * need stats to make this do the right thing.
3008  * this just finds the first lock on one of the
3009  * queues and uses that node as the target. */
3010 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
3011                                     struct dlm_lock_resource *res)
3012 {
3013         int i;
3014         struct list_head *queue = &res->granted;
3015         struct dlm_lock *lock;
3016         int nodenum;
3017
3018         assert_spin_locked(&dlm->spinlock);
3019
3020         spin_lock(&res->spinlock);
3021         for (i=0; i<3; i++) {
3022                 list_for_each_entry(lock, queue, list) {
3023                         /* up to the caller to make sure this node
3024                          * is alive */
3025                         if (lock->ml.node != dlm->node_num) {
3026                                 spin_unlock(&res->spinlock);
3027                                 return lock->ml.node;
3028                         }
3029                 }
3030                 queue++;
3031         }
3032         spin_unlock(&res->spinlock);
3033         mlog(0, "have not found a suitable target yet! checking domain map\n");
3034
3035         /* ok now we're getting desperate.  pick anyone alive. */
3036         nodenum = -1;
3037         while (1) {
3038                 nodenum = find_next_bit(dlm->domain_map,
3039                                         O2NM_MAX_NODES, nodenum+1);
3040                 mlog(0, "found %d in domain map\n", nodenum);
3041                 if (nodenum >= O2NM_MAX_NODES)
3042                         break;
3043                 if (nodenum != dlm->node_num) {
3044                         mlog(0, "picking %d\n", nodenum);
3045                         return nodenum;
3046                 }
3047         }
3048
3049         mlog(0, "giving up.  no master to migrate to\n");
3050         return DLM_LOCK_RES_OWNER_UNKNOWN;
3051 }
3052
3053
3054
3055 /* this is called by the new master once all lockres
3056  * data has been received */
3057 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
3058                                   struct dlm_lock_resource *res,
3059                                   u8 master, u8 new_master,
3060                                   struct dlm_node_iter *iter)
3061 {
3062         struct dlm_migrate_request migrate;
3063         int ret, status = 0;
3064         int nodenum;
3065
3066         memset(&migrate, 0, sizeof(migrate));
3067         migrate.namelen = res->lockname.len;
3068         memcpy(migrate.name, res->lockname.name, migrate.namelen);
3069         migrate.new_master = new_master;
3070         migrate.master = master;
3071
3072         ret = 0;
3073
3074         /* send message to all nodes, except the master and myself */
3075         while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
3076                 if (nodenum == master ||
3077                     nodenum == new_master)
3078                         continue;
3079
3080                 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
3081                                          &migrate, sizeof(migrate), nodenum,
3082                                          &status);
3083                 if (ret < 0)
3084                         mlog_errno(ret);
3085                 else if (status < 0) {
3086                         mlog(0, "migrate request (node %u) returned %d!\n",
3087                              nodenum, status);
3088                         ret = status;
3089                 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
3090                         /* during the migration request we short-circuited
3091                          * the mastery of the lockres.  make sure we have
3092                          * a mastery ref for nodenum */
3093                         mlog(0, "%s:%.*s: need ref for node %u\n",
3094                              dlm->name, res->lockname.len, res->lockname.name,
3095                              nodenum);
3096                         spin_lock(&res->spinlock);
3097                         dlm_lockres_set_refmap_bit(nodenum, res);
3098                         spin_unlock(&res->spinlock);
3099                 }
3100         }
3101
3102         if (ret < 0)
3103                 mlog_errno(ret);
3104
3105         mlog(0, "returning ret=%d\n", ret);
3106         return ret;
3107 }
3108
3109
3110 /* if there is an existing mle for this lockres, we now know who the master is.
3111  * (the one who sent us *this* message) we can clear it up right away.
3112  * since the process that put the mle on the list still has a reference to it,
3113  * we can unhash it now, set the master and wake the process.  as a result,
3114  * we will have no mle in the list to start with.  now we can add an mle for
3115  * the migration and this should be the only one found for those scanning the
3116  * list.  */
3117 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
3118                                 void **ret_data)
3119 {
3120         struct dlm_ctxt *dlm = data;
3121         struct dlm_lock_resource *res = NULL;
3122         struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
3123         struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
3124         const char *name;
3125         unsigned int namelen, hash;
3126         int ret = 0;
3127
3128         if (!dlm_grab(dlm))
3129                 return -EINVAL;
3130
3131         name = migrate->name;
3132         namelen = migrate->namelen;
3133         hash = dlm_lockid_hash(name, namelen);
3134
3135         /* preallocate.. if this fails, abort */
3136         mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
3137                                                          GFP_NOFS);
3138
3139         if (!mle) {
3140                 ret = -ENOMEM;
3141                 goto leave;
3142         }
3143
3144         /* check for pre-existing lock */
3145         spin_lock(&dlm->spinlock);
3146         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
3147         spin_lock(&dlm->master_lock);
3148
3149         if (res) {
3150                 spin_lock(&res->spinlock);
3151                 if (res->state & DLM_LOCK_RES_RECOVERING) {
3152                         /* if all is working ok, this can only mean that we got
3153                         * a migrate request from a node that we now see as
3154                         * dead.  what can we do here?  drop it to the floor? */
3155                         spin_unlock(&res->spinlock);
3156                         mlog(ML_ERROR, "Got a migrate request, but the "
3157                              "lockres is marked as recovering!");
3158                         kmem_cache_free(dlm_mle_cache, mle);
3159                         ret = -EINVAL; /* need a better solution */
3160                         goto unlock;
3161                 }
3162                 res->state |= DLM_LOCK_RES_MIGRATING;
3163                 spin_unlock(&res->spinlock);
3164         }
3165
3166         /* ignore status.  only nonzero status would BUG. */
3167         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3168                                     name, namelen,
3169                                     migrate->new_master,
3170                                     migrate->master);
3171
3172 unlock:
3173         spin_unlock(&dlm->master_lock);
3174         spin_unlock(&dlm->spinlock);
3175
3176         if (oldmle) {
3177                 /* master is known, detach if not already detached */
3178                 dlm_mle_detach_hb_events(dlm, oldmle);
3179                 dlm_put_mle(oldmle);
3180         }
3181
3182         if (res)
3183                 dlm_lockres_put(res);
3184 leave:
3185         dlm_put(dlm);
3186         return ret;
3187 }
3188
3189 /* must be holding dlm->spinlock and dlm->master_lock
3190  * when adding a migration mle, we can clear any other mles
3191  * in the master list because we know with certainty that
3192  * the master is "master".  so we remove any old mle from
3193  * the list after setting it's master field, and then add
3194  * the new migration mle.  this way we can hold with the rule
3195  * of having only one mle for a given lock name at all times. */
3196 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3197                                  struct dlm_lock_resource *res,
3198                                  struct dlm_master_list_entry *mle,
3199                                  struct dlm_master_list_entry **oldmle,
3200                                  const char *name, unsigned int namelen,
3201                                  u8 new_master, u8 master)
3202 {
3203         int found;
3204         int ret = 0;
3205
3206         *oldmle = NULL;
3207
3208         mlog_entry_void();
3209
3210         assert_spin_locked(&dlm->spinlock);
3211         assert_spin_locked(&dlm->master_lock);
3212
3213         /* caller is responsible for any ref taken here on oldmle */
3214         found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3215         if (found) {
3216                 struct dlm_master_list_entry *tmp = *oldmle;
3217                 spin_lock(&tmp->spinlock);
3218                 if (tmp->type == DLM_MLE_MIGRATION) {
3219                         if (master == dlm->node_num) {
3220                                 /* ah another process raced me to it */
3221                                 mlog(0, "tried to migrate %.*s, but some "
3222                                      "process beat me to it\n",
3223                                      namelen, name);
3224                                 ret = -EEXIST;
3225                         } else {
3226                                 /* bad.  2 NODES are trying to migrate! */
3227                                 mlog(ML_ERROR, "migration error  mle: "
3228                                      "master=%u new_master=%u // request: "
3229                                      "master=%u new_master=%u // "
3230                                      "lockres=%.*s\n",
3231                                      tmp->master, tmp->new_master,
3232                                      master, new_master,
3233                                      namelen, name);
3234                                 BUG();
3235                         }
3236                 } else {
3237                         /* this is essentially what assert_master does */
3238                         tmp->master = master;
3239                         atomic_set(&tmp->woken, 1);
3240                         wake_up(&tmp->wq);
3241                         /* remove it from the list so that only one
3242                          * mle will be found */
3243                         list_del_init(&tmp->list);
3244                         /* this was obviously WRONG.  mle is uninited here.  should be tmp. */
3245                         __dlm_mle_detach_hb_events(dlm, tmp);
3246                         ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3247                         mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3248                             "telling master to get ref for cleared out mle "
3249                             "during migration\n", dlm->name, namelen, name,
3250                             master, new_master);
3251                 }
3252                 spin_unlock(&tmp->spinlock);
3253         }
3254
3255         /* now add a migration mle to the tail of the list */
3256         dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3257         mle->new_master = new_master;
3258         /* the new master will be sending an assert master for this.
3259          * at that point we will get the refmap reference */
3260         mle->master = master;
3261         /* do this for consistency with other mle types */
3262         set_bit(new_master, mle->maybe_map);
3263         list_add(&mle->list, &dlm->master_list);
3264
3265         return ret;
3266 }
3267
3268
3269 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3270 {
3271         struct dlm_master_list_entry *mle, *next;
3272         struct dlm_lock_resource *res;
3273         unsigned int hash;
3274
3275         mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node);
3276 top:
3277         assert_spin_locked(&dlm->spinlock);
3278
3279         /* clean the master list */
3280         spin_lock(&dlm->master_lock);
3281         list_for_each_entry_safe(mle, next, &dlm->master_list, list) {
3282                 BUG_ON(mle->type != DLM_MLE_BLOCK &&
3283                        mle->type != DLM_MLE_MASTER &&
3284                        mle->type != DLM_MLE_MIGRATION);
3285
3286                 /* MASTER mles are initiated locally.  the waiting
3287                  * process will notice the node map change
3288                  * shortly.  let that happen as normal. */
3289                 if (mle->type == DLM_MLE_MASTER)
3290                         continue;
3291
3292
3293                 /* BLOCK mles are initiated by other nodes.
3294                  * need to clean up if the dead node would have
3295                  * been the master. */
3296                 if (mle->type == DLM_MLE_BLOCK) {
3297                         int bit;
3298
3299                         spin_lock(&mle->spinlock);
3300                         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3301                         if (bit != dead_node) {
3302                                 mlog(0, "mle found, but dead node %u would "
3303                                      "not have been master\n", dead_node);
3304                                 spin_unlock(&mle->spinlock);
3305                         } else {
3306                                 /* must drop the refcount by one since the
3307                                  * assert_master will never arrive.  this
3308                                  * may result in the mle being unlinked and
3309                                  * freed, but there may still be a process
3310                                  * waiting in the dlmlock path which is fine. */
3311                                 mlog(0, "node %u was expected master\n",
3312                                      dead_node);
3313                                 atomic_set(&mle->woken, 1);
3314                                 spin_unlock(&mle->spinlock);
3315                                 wake_up(&mle->wq);
3316                                 /* do not need events any longer, so detach
3317                                  * from heartbeat */
3318                                 __dlm_mle_detach_hb_events(dlm, mle);
3319                                 __dlm_put_mle(mle);
3320                         }
3321                         continue;
3322                 }
3323
3324                 /* everything else is a MIGRATION mle */
3325
3326                 /* the rule for MIGRATION mles is that the master
3327                  * becomes UNKNOWN if *either* the original or
3328                  * the new master dies.  all UNKNOWN lockreses
3329                  * are sent to whichever node becomes the recovery
3330                  * master.  the new master is responsible for
3331                  * determining if there is still a master for
3332                  * this lockres, or if he needs to take over
3333                  * mastery.  either way, this node should expect
3334                  * another message to resolve this. */
3335                 if (mle->master != dead_node &&
3336                     mle->new_master != dead_node)
3337                         continue;
3338
3339                 /* if we have reached this point, this mle needs to
3340                  * be removed from the list and freed. */
3341
3342                 /* remove from the list early.  NOTE: unlinking
3343                  * list_head while in list_for_each_safe */
3344                 __dlm_mle_detach_hb_events(dlm, mle);
3345                 spin_lock(&mle->spinlock);
3346                 list_del_init(&mle->list);
3347                 atomic_set(&mle->woken, 1);
3348                 spin_unlock(&mle->spinlock);
3349                 wake_up(&mle->wq);
3350
3351                 mlog(0, "%s: node %u died during migration from "
3352                      "%u to %u!\n", dlm->name, dead_node,
3353                      mle->master, mle->new_master);
3354                 /* if there is a lockres associated with this
3355                  * mle, find it and set its owner to UNKNOWN */
3356                 hash = dlm_lockid_hash(mle->u.name.name, mle->u.name.len);
3357                 res = __dlm_lookup_lockres(dlm, mle->u.name.name,
3358                                            mle->u.name.len, hash);
3359                 if (res) {
3360                         /* unfortunately if we hit this rare case, our
3361                          * lock ordering is messed.  we need to drop
3362                          * the master lock so that we can take the
3363                          * lockres lock, meaning that we will have to
3364                          * restart from the head of list. */
3365                         spin_unlock(&dlm->master_lock);
3366
3367                         /* move lockres onto recovery list */
3368                         spin_lock(&res->spinlock);
3369                         dlm_set_lockres_owner(dlm, res,
3370                                         DLM_LOCK_RES_OWNER_UNKNOWN);
3371                         dlm_move_lockres_to_recovery_list(dlm, res);
3372                         spin_unlock(&res->spinlock);
3373                         dlm_lockres_put(res);
3374
3375                         /* about to get rid of mle, detach from heartbeat */
3376                         __dlm_mle_detach_hb_events(dlm, mle);
3377
3378                         /* dump the mle */
3379                         spin_lock(&dlm->master_lock);
3380                         __dlm_put_mle(mle);
3381                         spin_unlock(&dlm->master_lock);
3382
3383                         /* restart */
3384                         goto top;
3385                 }
3386
3387                 /* this may be the last reference */
3388                 __dlm_put_mle(mle);
3389         }
3390         spin_unlock(&dlm->master_lock);
3391 }
3392
3393
3394 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3395                          u8 old_master)
3396 {
3397         struct dlm_node_iter iter;
3398         int ret = 0;
3399
3400         spin_lock(&dlm->spinlock);
3401         dlm_node_iter_init(dlm->domain_map, &iter);
3402         clear_bit(old_master, iter.node_map);
3403         clear_bit(dlm->node_num, iter.node_map);
3404         spin_unlock(&dlm->spinlock);
3405
3406         /* ownership of the lockres is changing.  account for the
3407          * mastery reference here since old_master will briefly have
3408          * a reference after the migration completes */
3409         spin_lock(&res->spinlock);
3410         dlm_lockres_set_refmap_bit(old_master, res);
3411         spin_unlock(&res->spinlock);
3412
3413         mlog(0, "now time to do a migrate request to other nodes\n");
3414         ret = dlm_do_migrate_request(dlm, res, old_master,
3415                                      dlm->node_num, &iter);
3416         if (ret < 0) {
3417                 mlog_errno(ret);
3418                 goto leave;
3419         }
3420
3421         mlog(0, "doing assert master of %.*s to all except the original node\n",
3422              res->lockname.len, res->lockname.name);
3423         /* this call now finishes out the nodemap
3424          * even if one or more nodes die */
3425         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3426                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3427         if (ret < 0) {
3428                 /* no longer need to retry.  all living nodes contacted. */
3429                 mlog_errno(ret);
3430                 ret = 0;
3431         }
3432
3433         memset(iter.node_map, 0, sizeof(iter.node_map));
3434         set_bit(old_master, iter.node_map);
3435         mlog(0, "doing assert master of %.*s back to %u\n",
3436              res->lockname.len, res->lockname.name, old_master);
3437         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3438                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3439         if (ret < 0) {
3440                 mlog(0, "assert master to original master failed "
3441                      "with %d.\n", ret);
3442                 /* the only nonzero status here would be because of
3443                  * a dead original node.  we're done. */
3444                 ret = 0;
3445         }
3446
3447         /* all done, set the owner, clear the flag */
3448         spin_lock(&res->spinlock);
3449         dlm_set_lockres_owner(dlm, res, dlm->node_num);
3450         res->state &= ~DLM_LOCK_RES_MIGRATING;
3451         spin_unlock(&res->spinlock);
3452         /* re-dirty it on the new master */
3453         dlm_kick_thread(dlm, res);
3454         wake_up(&res->wq);
3455 leave:
3456         return ret;
3457 }
3458
3459 /*
3460  * LOCKRES AST REFCOUNT
3461  * this is integral to migration
3462  */
3463
3464 /* for future intent to call an ast, reserve one ahead of time.
3465  * this should be called only after waiting on the lockres
3466  * with dlm_wait_on_lockres, and while still holding the
3467  * spinlock after the call. */
3468 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3469 {
3470         assert_spin_locked(&res->spinlock);
3471         if (res->state & DLM_LOCK_RES_MIGRATING) {
3472                 __dlm_print_one_lock_resource(res);
3473         }
3474         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3475
3476         atomic_inc(&res->asts_reserved);
3477 }
3478
3479 /*
3480  * used to drop the reserved ast, either because it went unused,
3481  * or because the ast/bast was actually called.
3482  *
3483  * also, if there is a pending migration on this lockres,
3484  * and this was the last pending ast on the lockres,
3485  * atomically set the MIGRATING flag before we drop the lock.
3486  * this is how we ensure that migration can proceed with no
3487  * asts in progress.  note that it is ok if the state of the
3488  * queues is such that a lock should be granted in the future
3489  * or that a bast should be fired, because the new master will
3490  * shuffle the lists on this lockres as soon as it is migrated.
3491  */
3492 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3493                              struct dlm_lock_resource *res)
3494 {
3495         if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3496                 return;
3497
3498         if (!res->migration_pending) {
3499                 spin_unlock(&res->spinlock);
3500                 return;
3501         }
3502
3503         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3504         res->migration_pending = 0;
3505         res->state |= DLM_LOCK_RES_MIGRATING;
3506         spin_unlock(&res->spinlock);
3507         wake_up(&res->wq);
3508         wake_up(&dlm->migration_wq);
3509 }