lustre/ldlm/ldlm_lock.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #define DEBUG_SUBSYSTEM S_LDLM
  25
  26 #include <linux/slab.h>
  27 #include <linux/module.h>
  28 #include <linux/random.h>
  29 #include <linux/lustre_dlm.h>
  30 #include <linux/lustre_mds.h>
  31 #include <linux/obd_class.h>
  32
  33 //struct lustre_lock ldlm_everything_lock;
  34
  35 /* lock types */
  36 char *ldlm_lockname[] = {
  37         [0] "--",
  38         [LCK_EX] "EX",
  39         [LCK_PW] "PW",
  40         [LCK_PR] "PR",
  41         [LCK_CW] "CW",
  42         [LCK_CR] "CR",
  43         [LCK_NL] "NL"
  44 };
  45 char *ldlm_typename[] = {
  46         [LDLM_PLAIN] "PLN",
  47         [LDLM_EXTENT] "EXT",
  48 };
  49
  50 char *ldlm_it2str(int it)
  51 {
  52         switch (it) {
  53         case IT_OPEN:
  54                 return "open";
  55         case IT_CREAT:
  56                 return "creat";
  57         case (IT_OPEN | IT_CREAT):
  58                 return "open|creat";
  59         case IT_MKDIR:
  60                 return "mkdir";
  61         case IT_LINK:
  62                 return "link";
  63         case IT_LINK2:
  64                 return "link2";
  65         case IT_SYMLINK:
  66                 return "symlink";
  67         case IT_UNLINK:
  68                 return "unlink";
  69         case IT_RMDIR:
  70                 return "rmdir";
  71         case IT_RENAME:
  72                 return "rename";
  73         case IT_RENAME2:
  74                 return "rename2";
  75         case IT_READDIR:
  76                 return "readdir";
  77         case IT_GETATTR:
  78                 return "getattr";
  79         case IT_SETATTR:
  80                 return "setattr";
  81         case IT_READLINK:
  82                 return "readlink";
  83         case IT_MKNOD:
  84                 return "mknod";
  85         case IT_LOOKUP:
  86                 return "lookup";
  87         default:
  88                 CERROR("Unknown intent %d\n", it);
  89                 return "UNKNOWN";
  90         }
  91 }
  92
  93 extern kmem_cache_t *ldlm_lock_slab;
  94 struct lustre_lock ldlm_handle_lock;
  95
  96 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b);
  97
  98 ldlm_res_compat ldlm_res_compat_table[] = {
  99         [LDLM_PLAIN] ldlm_plain_compat,
 100         [LDLM_EXTENT] ldlm_extent_compat,
 101 };
 102
 103 static ldlm_res_policy ldlm_intent_policy_func;
 104
 105 static int ldlm_plain_policy(struct ldlm_lock *lock, void *req_cookie,
 106                              ldlm_mode_t mode, int flags, void *data)
 107 {
 108         if ((flags & LDLM_FL_HAS_INTENT) && ldlm_intent_policy_func) {
 109                 return ldlm_intent_policy_func(lock, req_cookie, mode, flags,
 110                                                data);
 111         }
 112
 113         return ELDLM_OK;
 114 }
 115
 116 ldlm_res_policy ldlm_res_policy_table[] = {
 117         [LDLM_PLAIN] ldlm_plain_policy,
 118         [LDLM_EXTENT] ldlm_extent_policy,
 119 };
 120
 121 void ldlm_register_intent(ldlm_res_policy arg)
 122 {
 123         ldlm_intent_policy_func = arg;
 124 }
 125
 126 void ldlm_unregister_intent(void)
 127 {
 128         ldlm_intent_policy_func = NULL;
 129 }
 130
 131 /*
 132  * REFCOUNTED LOCK OBJECTS
 133  */
 134
 135
 136 /*
 137  * Lock refcounts, during creation:
 138  *   - one special one for allocation, dec'd only once in destroy
 139  *   - one for being a lock that's in-use
 140  *   - one for the addref associated with a new lock
 141  */
 142 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 143 {
 144         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 145         atomic_inc(&lock->l_refc);
 146         ldlm_resource_getref(lock->l_resource);
 147         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 148         return lock;
 149 }
 150
 151 void ldlm_lock_put(struct ldlm_lock *lock)
 152 {
 153         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 154         ENTRY;
 155
 156         l_lock(&ns->ns_lock);
 157         atomic_dec(&lock->l_refc);
 158         LASSERT(atomic_read(&lock->l_refc) >= 0);
 159
 160         if (ldlm_resource_putref(lock->l_resource)) {
 161                 LASSERT(atomic_read(&lock->l_refc) == 0);
 162                 lock->l_resource = NULL;
 163         }
 164         if (lock->l_parent)
 165                 LDLM_LOCK_PUT(lock->l_parent);
 166
 167         if (atomic_read(&lock->l_refc) == 0) {
 168                 LASSERT(lock->l_destroyed);
 169                 l_unlock(&ns->ns_lock);
 170                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
 171
 172                 spin_lock(&ns->ns_counter_lock);
 173                 ns->ns_locks--;
 174                 spin_unlock(&ns->ns_counter_lock);
 175
 176                 lock->l_resource = NULL;
 177                 lock->l_random = DEAD_HANDLE_MAGIC;
 178                 memset(lock, 0x5a, sizeof(*lock));
 179                 kmem_cache_free(ldlm_lock_slab, lock);
 180                 CDEBUG(D_MALLOC, "kfreed 'lock': %d at %p (tot 0).\n",
 181                        sizeof(*lock), lock);
 182         } else
 183                 l_unlock(&ns->ns_lock);
 184
 185         EXIT;
 186 }
 187
 188 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
 189 {
 190         ENTRY;
 191         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 192         if (!list_empty(&lock->l_lru)) {
 193                 list_del_init(&lock->l_lru);
 194                 lock->l_resource->lr_namespace->ns_nr_unused--;
 195                 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
 196         }
 197         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 198         EXIT;
 199 }
 200
 201 void ldlm_lock_destroy(struct ldlm_lock *lock)
 202 {
 203         ENTRY;
 204         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 205
 206         if (!list_empty(&lock->l_children)) {
 207                 LDLM_DEBUG(lock, "still has children (%p)!",
 208                            lock->l_children.next);
 209                 ldlm_lock_dump(lock);
 210                 LBUG();
 211         }
 212         if (lock->l_readers || lock->l_writers) {
 213                 LDLM_DEBUG(lock, "lock still has references");
 214                 ldlm_lock_dump(lock);
 215         }
 216
 217         if (!list_empty(&lock->l_res_link)) {
 218                 ldlm_lock_dump(lock);
 219                 LBUG();
 220         }
 221
 222         if (lock->l_destroyed) {
 223                 LASSERT(list_empty(&lock->l_lru));
 224                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 225                 EXIT;
 226                 return;
 227         }
 228         lock->l_destroyed = 1;
 229
 230         list_del(&lock->l_export_chain);
 231         lock->l_export = NULL;
 232         ldlm_lock_remove_from_lru(lock);
 233
 234         /* Wake anyone waiting for this lock */
 235         /* FIXME: I should probably add yet another flag, instead of using
 236          * l_export to only call this on clients */
 237         if (lock->l_export && lock->l_completion_ast)
 238                 lock->l_completion_ast(lock, 0);
 239
 240         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 241         LDLM_LOCK_PUT(lock);
 242         EXIT;
 243 }
 244
 245 /*
 246    usage: pass in a resource on which you have done get
 247           pass in a parent lock on which you have done a get
 248           do not put the resource or the parent
 249    returns: lock with refcount 1
 250 */
 251 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 252                                        struct ldlm_resource *resource)
 253 {
 254         struct ldlm_lock *lock;
 255         ENTRY;
 256
 257         if (resource == NULL)
 258                 LBUG();
 259
 260         lock = kmem_cache_alloc(ldlm_lock_slab, SLAB_KERNEL);
 261         if (lock == NULL)
 262                 RETURN(NULL);
 263
 264         memset(lock, 0, sizeof(*lock));
 265         get_random_bytes(&lock->l_random, sizeof(__u64));
 266
 267         lock->l_resource = resource;
 268         /* this refcount matches the one of the resource passed
 269            in which is not being put away */
 270         atomic_set(&lock->l_refc, 1);
 271         INIT_LIST_HEAD(&lock->l_children);
 272         INIT_LIST_HEAD(&lock->l_res_link);
 273         INIT_LIST_HEAD(&lock->l_lru);
 274         INIT_LIST_HEAD(&lock->l_export_chain);
 275         INIT_LIST_HEAD(&lock->l_pending_chain);
 276         init_waitqueue_head(&lock->l_waitq);
 277
 278         spin_lock(&resource->lr_namespace->ns_counter_lock);
 279         resource->lr_namespace->ns_locks++;
 280         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 281
 282         if (parent != NULL) {
 283                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 284                 lock->l_parent = parent;
 285                 list_add(&lock->l_childof, &parent->l_children);
 286                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 287         }
 288
 289         CDEBUG(D_MALLOC, "kmalloced 'lock': %d at "
 290                "%p (tot %d).\n", sizeof(*lock), lock, 1);
 291         /* this is the extra refcount, to prevent the lock from evaporating */
 292         LDLM_LOCK_GET(lock);
 293         RETURN(lock);
 294 }
 295
 296 int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3])
 297 {
 298         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 299         struct ldlm_resource *oldres = lock->l_resource;
 300         int i, refc;
 301         ENTRY;
 302
 303         l_lock(&ns->ns_lock);
 304         if (memcmp(new_resid, lock->l_resource->lr_name,
 305                    sizeof(lock->l_resource->lr_name)) == 0) {
 306                 /* Nothing to do */
 307                 l_unlock(&ns->ns_lock);
 308                 RETURN(0);
 309         }
 310
 311         LASSERT(new_resid[0] != 0);
 312
 313         /* This function assumes that the lock isn't on any lists */
 314         LASSERT(list_empty(&lock->l_res_link));
 315
 316         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
 317                                              lock->l_resource->lr_type, 1);
 318         if (lock->l_resource == NULL) {
 319                 LBUG();
 320                 RETURN(-ENOMEM);
 321         }
 322
 323         /* move references over */
 324         refc = atomic_read(&lock->l_refc);
 325         for (i = 0; i < refc; i++) {
 326                 int rc;
 327                 ldlm_resource_getref(lock->l_resource);
 328                 rc = ldlm_resource_putref(oldres);
 329                 if (rc == 1 && i != refc - 1)
 330                         LBUG();
 331         }
 332         /* compensate for the initial get above.. */
 333         ldlm_resource_putref(lock->l_resource);
 334
 335         l_unlock(&ns->ns_lock);
 336         RETURN(0);
 337 }
 338
 339 /*
 340  *  HANDLES
 341  */
 342
 343 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 344 {
 345         lockh->addr = (__u64) (unsigned long)lock;
 346         lockh->cookie = lock->l_random;
 347 }
 348
 349 /*
 350  * if flags: atomically get the lock and set the flags.
 351  * Return NULL if flag already set
 352  */
 353
 354 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int strict,
 355                                      int flags)
 356 {
 357         struct ldlm_lock *lock = NULL, *retval = NULL;
 358         ENTRY;
 359
 360         LASSERT(handle);
 361
 362         if (!handle->addr)
 363                 RETURN(NULL);
 364
 365         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
 366         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock)) {
 367                 //CERROR("bogus lock %p\n", lock);
 368                 GOTO(out2, retval);
 369         }
 370
 371         if (lock->l_random != handle->cookie) {
 372                 //CERROR("bogus cookie: lock %p has "LPX64" vs. handle "LPX64
 373                 //       "\n", lock, lock->l_random, handle->cookie);
 374                 GOTO(out2, NULL);
 375         }
 376         if (!lock->l_resource) {
 377                 CERROR("trying to lock bogus resource: lock %p\n", lock);
 378                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 379                 GOTO(out2, retval);
 380         }
 381         if (!lock->l_resource->lr_namespace) {
 382                 CERROR("trying to lock bogus namespace: lock %p\n", lock);
 383                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 384                 GOTO(out2, retval);
 385         }
 386
 387         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 388         if (strict && lock->l_destroyed) {
 389                 CERROR("lock already destroyed: lock %p\n", lock);
 390                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 391                 GOTO(out, NULL);
 392         }
 393
 394         if (flags && (lock->l_flags & flags))
 395                 GOTO(out, NULL);
 396
 397         if (flags)
 398                 lock->l_flags |= flags;
 399
 400         retval = LDLM_LOCK_GET(lock);
 401         if (!retval)
 402                 CERROR("lock disappeared below us!!! %p\n", lock);
 403         EXIT;
 404  out:
 405         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 406  out2:
 407         return retval;
 408 }
 409
 410 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 411 {
 412         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 413 }
 414
 415 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 416 {
 417         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 418         desc->l_req_mode = lock->l_req_mode;
 419         desc->l_granted_mode = lock->l_granted_mode;
 420         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 421         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 422 }
 423
 424 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 425                                    struct ldlm_lock *new)
 426 {
 427         struct ldlm_ast_work *w;
 428         ENTRY;
 429
 430         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 431         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 432                 GOTO(out, 0);
 433
 434         OBD_ALLOC(w, sizeof(*w));
 435         if (!w) {
 436                 LBUG();
 437                 GOTO(out, 0);
 438         }
 439
 440         if (new) {
 441                 lock->l_flags |= LDLM_FL_AST_SENT;
 442                 w->w_blocking = 1;
 443                 ldlm_lock2desc(new, &w->w_desc);
 444         }
 445
 446         w->w_lock = LDLM_LOCK_GET(lock);
 447         list_add(&w->w_list, lock->l_resource->lr_tmp);
 448       out:
 449         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 450         return;
 451 }
 452
 453 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 454 {
 455         struct ldlm_lock *lock;
 456
 457         lock = ldlm_handle2lock(lockh);
 458         ldlm_lock_addref_internal(lock, mode);
 459         LDLM_LOCK_PUT(lock);
 460 }
 461
 462 /* only called for local locks */
 463 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 464 {
 465         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 466         ldlm_lock_remove_from_lru(lock);
 467         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 468                 lock->l_readers++;
 469         else
 470                 lock->l_writers++;
 471         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 472         LDLM_LOCK_GET(lock);
 473         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 474 }
 475
 476 /* Args: unlocked lock */
 477 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 478                                     __u64 *res_id, int flags);
 479
 480 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 481 {
 482         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0, 0);
 483         struct ldlm_namespace *ns;
 484         ENTRY;
 485
 486         if (lock == NULL)
 487                 LBUG();
 488
 489         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 490         ns = lock->l_resource->lr_namespace;
 491         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 492         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 493                 lock->l_readers--;
 494         else
 495                 lock->l_writers--;
 496
 497         /* If we received a blocked AST and this was the last reference,
 498          * run the callback. */
 499         if (!lock->l_readers && !lock->l_writers &&
 500             (lock->l_flags & LDLM_FL_CBPENDING)) {
 501                 if (!lock->l_resource->lr_namespace->ns_client &&
 502                     lock->l_export)
 503                         CERROR("FL_CBPENDING set on non-local lock--just a "
 504                                "warning\n");
 505
 506                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 507                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 508
 509                 /* FIXME: need a real 'desc' here */
 510                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 511                                      lock->l_data_len, LDLM_CB_BLOCKING);
 512         } else if (ns->ns_client && !lock->l_readers && !lock->l_writers) {
 513                 LASSERT(list_empty(&lock->l_lru));
 514                 LASSERT(ns->ns_nr_unused >= 0);
 515                 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 516                 ns->ns_nr_unused++;
 517                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 518                 ldlm_cancel_lru(ns);
 519         } else
 520                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 521
 522         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 523         LDLM_LOCK_PUT(lock);    /* matches the handle2lock above */
 524
 525         EXIT;
 526 }
 527
 528 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 529                                  struct list_head *queue)
 530 {
 531         struct list_head *tmp, *pos;
 532         int rc = 1;
 533
 534         list_for_each_safe(tmp, pos, queue) {
 535                 struct ldlm_lock *child;
 536                 ldlm_res_compat compat;
 537
 538                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 539                 if (lock == child)
 540                         continue;
 541
 542                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 543                 if (compat && compat(child, lock)) {
 544                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 545                         continue;
 546                 }
 547                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 548                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 549                         continue;
 550                 }
 551
 552                 rc = 0;
 553
 554                 if (send_cbs && child->l_blocking_ast != NULL) {
 555                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 556                                "blocking AST.\n", child);
 557                         ldlm_add_ast_work_item(child, lock);
 558                 }
 559         }
 560
 561         return rc;
 562 }
 563
 564 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 565 {
 566         int rc;
 567         ENTRY;
 568
 569         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 570         rc = ldlm_lock_compat_list(lock, send_cbs,
 571                                    &lock->l_resource->lr_granted);
 572         /* FIXME: should we be sending ASTs to converting? */
 573         if (rc)
 574                 rc = ldlm_lock_compat_list
 575                         (lock, send_cbs, &lock->l_resource->lr_converting);
 576
 577         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 578         RETURN(rc);
 579 }
 580
 581 /* NOTE: called by
 582    - ldlm_handle_enqueuque - resource
 583 */
 584 void ldlm_grant_lock(struct ldlm_lock *lock)
 585 {
 586         struct ldlm_resource *res = lock->l_resource;
 587         ENTRY;
 588
 589         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 590         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 591         lock->l_granted_mode = lock->l_req_mode;
 592
 593         if (lock->l_granted_mode < res->lr_most_restr)
 594                 res->lr_most_restr = lock->l_granted_mode;
 595
 596         if (lock->l_completion_ast) {
 597                 ldlm_add_ast_work_item(lock, NULL);
 598         }
 599         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 600         EXIT;
 601 }
 602
 603 /* returns a referenced lock or NULL */
 604 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 605                                       struct ldlm_extent *extent,
 606                                       struct ldlm_lock *old_lock)
 607 {
 608         struct ldlm_lock *lock;
 609         struct list_head *tmp;
 610
 611         list_for_each(tmp, queue) {
 612                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 613
 614                 if (lock == old_lock)
 615                         continue;
 616
 617                 if (lock->l_flags & LDLM_FL_CBPENDING)
 618                         continue;
 619
 620                 if (lock->l_req_mode != mode)
 621                         continue;
 622
 623                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 624                     (lock->l_extent.start > extent->start ||
 625                      lock->l_extent.end < extent->end))
 626                         continue;
 627
 628                 if (lock->l_destroyed)
 629                         continue;
 630
 631                 ldlm_lock_addref_internal(lock, mode);
 632                 return lock;
 633         }
 634
 635         return NULL;
 636 }
 637
 638 /* Can be called in two ways:
 639  *
 640  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
 641  * for a duplicate of.
 642  *
 643  * Otherwise, all of the fields must be filled in, to match against.
 644  *
 645  * Returns 1 if it finds an already-existing lock that is compatible; in this
 646  * case, lockh is filled in with a addref()ed lock
 647  */
 648 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 *res_id, __u32 type,
 649                     void *cookie, int cookielen, ldlm_mode_t mode,
 650                     struct lustre_handle *lockh)
 651 {
 652         struct ldlm_resource *res;
 653         struct ldlm_lock *lock, *old_lock = NULL;
 654         int rc = 0;
 655         ENTRY;
 656
 657         if (ns == NULL) {
 658                 old_lock = ldlm_handle2lock(lockh);
 659                 LASSERT(old_lock);
 660
 661                 ns = old_lock->l_resource->lr_namespace;
 662                 res_id = old_lock->l_resource->lr_name;
 663                 type = old_lock->l_resource->lr_type;
 664                 mode = old_lock->l_req_mode;
 665         }
 666
 667         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 668         if (res == NULL) {
 669                 LASSERT(old_lock == NULL);
 670                 RETURN(0);
 671         }
 672
 673         l_lock(&ns->ns_lock);
 674
 675         if ((lock = search_queue(&res->lr_granted, mode, cookie, old_lock)))
 676                 GOTO(out, rc = 1);
 677         if ((lock = search_queue(&res->lr_converting, mode, cookie, old_lock)))
 678                 GOTO(out, rc = 1);
 679         if ((lock = search_queue(&res->lr_waiting, mode, cookie, old_lock)))
 680                 GOTO(out, rc = 1);
 681
 682         EXIT;
 683        out:
 684         ldlm_resource_putref(res);
 685         l_unlock(&ns->ns_lock);
 686
 687         if (lock) {
 688                 ldlm_lock2handle(lock, lockh);
 689                 if (lock->l_completion_ast)
 690                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 691         }
 692         if (rc)
 693                 LDLM_DEBUG(lock, "matched");
 694         else
 695                 LDLM_DEBUG_NOLOCK("not matched");
 696
 697         if (old_lock)
 698                 LDLM_LOCK_PUT(old_lock);
 699
 700         return rc;
 701 }
 702
 703 /* Returns a referenced lock */
 704 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 705                                    struct lustre_handle *parent_lock_handle,
 706                                    __u64 * res_id, __u32 type,
 707                                    ldlm_mode_t mode, void *data, __u32 data_len)
 708 {
 709         struct ldlm_resource *res, *parent_res = NULL;
 710         struct ldlm_lock *lock, *parent_lock = NULL;
 711
 712         if (parent_lock_handle) {
 713                 parent_lock = ldlm_handle2lock(parent_lock_handle);
 714                 if (parent_lock)
 715                         parent_res = parent_lock->l_resource;
 716         }
 717
 718         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 719         if (res == NULL)
 720                 RETURN(NULL);
 721
 722         lock = ldlm_lock_new(parent_lock, res);
 723         if (lock == NULL) {
 724                 ldlm_resource_putref(res);
 725                 RETURN(NULL);
 726         }
 727
 728         lock->l_req_mode = mode;
 729         lock->l_data = data;
 730         lock->l_data_len = data_len;
 731
 732         return lock;
 733 }
 734
 735 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 736 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock * lock,
 737                                void *cookie, int cookie_len,
 738                                int *flags,
 739                                ldlm_completion_callback completion,
 740                                ldlm_blocking_callback blocking)
 741 {
 742         struct ldlm_resource *res;
 743         int local;
 744         ldlm_res_policy policy;
 745         ENTRY;
 746
 747         res = lock->l_resource;
 748         lock->l_blocking_ast = blocking;
 749
 750         if (res->lr_type == LDLM_EXTENT)
 751                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 752
 753         /* policies are not executed on the client or during replay */
 754         local = res->lr_namespace->ns_client;
 755         if (!local && !(*flags & LDLM_FL_REPLAY) &&
 756             (policy = ldlm_res_policy_table[res->lr_type])) {
 757                 int rc;
 758                 rc = policy(lock, cookie, lock->l_req_mode, *flags, NULL);
 759
 760                 if (rc == ELDLM_LOCK_CHANGED) {
 761                         res = lock->l_resource;
 762                         *flags |= LDLM_FL_LOCK_CHANGED;
 763                 } else if (rc == ELDLM_LOCK_ABORTED) {
 764                         ldlm_lock_destroy(lock);
 765                         RETURN(rc);
 766                 }
 767         }
 768
 769         l_lock(&res->lr_namespace->ns_lock);
 770         if (local && lock->l_req_mode == lock->l_granted_mode) {
 771                 /* The server returned a blocked lock, but it was granted before
 772                  * we got a chance to actually enqueue it.  We don't need to do
 773                  * anything else. */
 774                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 775                           LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 776                 GOTO(out, ELDLM_OK);
 777         }
 778
 779         /* This distinction between local lock trees is very important; a client
 780          * namespace only has information about locks taken by that client, and
 781          * thus doesn't have enough information to decide for itself if it can
 782          * be granted (below).  In this case, we do exactly what the server
 783          * tells us to do, as dictated by the 'flags'.
 784          *
 785          * We do exactly the same thing during recovery, when the server is
 786          * more or less trusting the clients not to lie.
 787          *
 788          * FIXME (bug 629283): Detect obvious lies by checking compatibility in
 789          * granted/converting queues. */
 790         ldlm_resource_unlink_lock(lock);
 791         if (local || (*flags & LDLM_FL_REPLAY)) {
 792                 if (*flags & LDLM_FL_BLOCK_CONV)
 793                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 794                                                lock);
 795                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 796                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 797                 else
 798                         ldlm_grant_lock(lock);
 799                 GOTO(out, ELDLM_OK);
 800         }
 801
 802         /* FIXME: We may want to optimize by checking lr_most_restr */
 803         if (!list_empty(&res->lr_converting)) {
 804                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 805                 *flags |= LDLM_FL_BLOCK_CONV;
 806                 GOTO(out, ELDLM_OK);
 807         }
 808         if (!list_empty(&res->lr_waiting)) {
 809                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 810                 *flags |= LDLM_FL_BLOCK_WAIT;
 811                 GOTO(out, ELDLM_OK);
 812         }
 813         if (!ldlm_lock_compat(lock, 0)) {
 814                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 815                 *flags |= LDLM_FL_BLOCK_GRANTED;
 816                 GOTO(out, ELDLM_OK);
 817         }
 818
 819         ldlm_grant_lock(lock);
 820         EXIT;
 821       out:
 822         l_unlock(&res->lr_namespace->ns_lock);
 823         /* Don't set 'completion_ast' until here so that if the lock is granted
 824          * immediately we don't do an unnecessary completion call. */
 825         lock->l_completion_ast = completion;
 826         return ELDLM_OK;
 827 }
 828
 829 /* Must be called with namespace taken: queue is waiting or converting. */
 830 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 831                                 struct list_head *queue)
 832 {
 833         struct list_head *tmp, *pos;
 834         ENTRY;
 835
 836         list_for_each_safe(tmp, pos, queue) {
 837                 struct ldlm_lock *pending;
 838                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 839
 840                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 841
 842                 if (!ldlm_lock_compat(pending, 1))
 843                         RETURN(1);
 844
 845                 list_del_init(&pending->l_res_link);
 846                 ldlm_grant_lock(pending);
 847         }
 848
 849         RETURN(0);
 850 }
 851
 852 void ldlm_run_ast_work(struct list_head *rpc_list)
 853 {
 854         struct list_head *tmp, *pos;
 855         int rc;
 856         ENTRY;
 857
 858         list_for_each_safe(tmp, pos, rpc_list) {
 859                 struct ldlm_ast_work *w =
 860                         list_entry(tmp, struct ldlm_ast_work, w_list);
 861
 862                 if (w->w_blocking)
 863                         rc = w->w_lock->l_blocking_ast
 864                                 (w->w_lock, &w->w_desc, w->w_data,
 865                                  w->w_datalen, LDLM_CB_BLOCKING);
 866                 else
 867                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags);
 868                 if (rc)
 869                         CERROR("Failed AST - should clean & disconnect "
 870                                "client\n");
 871                 LDLM_LOCK_PUT(w->w_lock);
 872                 list_del(&w->w_list);
 873                 OBD_FREE(w, sizeof(*w));
 874         }
 875         EXIT;
 876 }
 877
 878 /* Must be called with resource->lr_lock not taken. */
 879 void ldlm_reprocess_all(struct ldlm_resource *res)
 880 {
 881         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 882         ENTRY;
 883
 884         /* Local lock trees don't get reprocessed. */
 885         if (res->lr_namespace->ns_client) {
 886                 EXIT;
 887                 return;
 888         }
 889
 890         l_lock(&res->lr_namespace->ns_lock);
 891         res->lr_tmp = &rpc_list;
 892
 893         ldlm_reprocess_queue(res, &res->lr_converting);
 894         if (list_empty(&res->lr_converting))
 895                 ldlm_reprocess_queue(res, &res->lr_waiting);
 896
 897         res->lr_tmp = NULL;
 898         l_unlock(&res->lr_namespace->ns_lock);
 899
 900         ldlm_run_ast_work(&rpc_list);
 901         EXIT;
 902 }
 903
 904 void ldlm_cancel_callback(struct ldlm_lock *lock)
 905 {
 906         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 907         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
 908                 lock->l_flags |= LDLM_FL_CANCEL;
 909                 if (lock->l_blocking_ast)
 910                         lock->l_blocking_ast(lock, NULL, lock->l_data,
 911                                              lock->l_data_len,
 912                                              LDLM_CB_CANCELING);
 913                 else
 914                         LDLM_DEBUG(lock, "no blocking ast");
 915         }
 916         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 917 }
 918
 919 void ldlm_lock_cancel(struct ldlm_lock *lock)
 920 {
 921         struct ldlm_resource *res;
 922         struct ldlm_namespace *ns;
 923         ENTRY;
 924
 925         res = lock->l_resource;
 926         ns = res->lr_namespace;
 927
 928         l_lock(&ns->ns_lock);
 929         if (lock->l_readers || lock->l_writers)
 930                 LDLM_DEBUG(lock, "lock still has references");
 931
 932         ldlm_cancel_callback(lock);
 933
 934         ldlm_del_waiting_lock(lock);
 935         ldlm_resource_unlink_lock(lock);
 936         ldlm_lock_destroy(lock);
 937         l_unlock(&ns->ns_lock);
 938         EXIT;
 939 }
 940
 941 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data, int datalen)
 942 {
 943         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 944         ENTRY;
 945
 946         if (lock == NULL)
 947                 RETURN(-EINVAL);
 948
 949         lock->l_data = data;
 950         lock->l_data_len = datalen;
 951
 952         LDLM_LOCK_PUT(lock);
 953
 954         RETURN(0);
 955 }
 956
 957 void ldlm_cancel_locks_for_export(struct obd_export *exp)
 958 {
 959         struct list_head *iter, *n; /* MUST BE CALLED "n"! */
 960
 961         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
 962                 struct ldlm_lock *lock;
 963                 struct ldlm_resource *res;
 964                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 965                 res = ldlm_resource_getref(lock->l_resource);
 966                 LDLM_DEBUG(lock, "export %p", exp);
 967                 ldlm_lock_cancel(lock);
 968                 ldlm_reprocess_all(res);
 969                 ldlm_resource_putref(res);
 970         }
 971 }
 972
 973 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 974                                         int *flags)
 975 {
 976         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 977         struct ldlm_resource *res;
 978         struct ldlm_namespace *ns;
 979         int granted = 0;
 980         ENTRY;
 981
 982         res = lock->l_resource;
 983         ns = res->lr_namespace;
 984
 985         l_lock(&ns->ns_lock);
 986
 987         lock->l_req_mode = new_mode;
 988         ldlm_resource_unlink_lock(lock);
 989
 990         /* If this is a local resource, put it on the appropriate list. */
 991         if (res->lr_namespace->ns_client) {
 992                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED))
 993                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 994                                                lock);
 995                 else {
 996                         /* This should never happen, because of the way the
 997                          * server handles conversions. */
 998                         LBUG();
 999
1000                         res->lr_tmp = &rpc_list;
1001                         ldlm_grant_lock(lock);
1002                         res->lr_tmp = NULL;
1003                         granted = 1;
1004                         /* FIXME: completion handling not with ns_lock held ! */
1005                         if (lock->l_completion_ast)
1006                                 lock->l_completion_ast(lock, 0);
1007                 }
1008         } else {
1009                 /* FIXME: We should try the conversion right away and possibly
1010                  * return success without the need for an extra AST */
1011                 ldlm_resource_add_lock(res, res->lr_converting.prev, lock);
1012                 *flags |= LDLM_FL_BLOCK_CONV;
1013         }
1014
1015         l_unlock(&ns->ns_lock);
1016
1017         if (granted)
1018                 ldlm_run_ast_work(&rpc_list);
1019         RETURN(res);
1020 }
1021
1022 void ldlm_lock_dump(struct ldlm_lock *lock)
1023 {
1024         char ver[128];
1025
1026         if (!(portal_debug & D_OTHER))
1027                 return;
1028
1029         if (RES_VERSION_SIZE != 4)
1030                 LBUG();
1031
1032         if (!lock) {
1033                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
1034                 return;
1035         }
1036
1037         snprintf(ver, sizeof(ver), "%x %x %x %x",
1038                  lock->l_version[0], lock->l_version[1],
1039                  lock->l_version[2], lock->l_version[3]);
1040
1041         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
1042         if (lock->l_export && lock->l_export->exp_connection)
1043                 CDEBUG(D_OTHER, "  Node: NID %x (rhandle: "LPX64")\n",
1044                        lock->l_export->exp_connection->c_peer.peer_nid,
1045                        lock->l_remote_handle.addr);
1046         else
1047                 CDEBUG(D_OTHER, "  Node: local\n");
1048         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
1049         CDEBUG(D_OTHER, "  Resource: %p ("LPD64")\n", lock->l_resource,
1050                lock->l_resource->lr_name[0]);
1051         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
1052                (int)lock->l_req_mode, (int)lock->l_granted_mode);
1053         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
1054                lock->l_readers, lock->l_writers);
1055         if (lock->l_resource->lr_type == LDLM_EXTENT)
1056                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
1057                        (unsigned long long)lock->l_extent.start,
1058                        (unsigned long long)lock->l_extent.end);
1059 }