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                 if (lock->l_export && lock->l_export->exp_connection)
 179                         ptlrpc_put_connection(lock->l_export->exp_connection);
 180                 memset(lock, 0x5a, sizeof(*lock));
 181                 kmem_cache_free(ldlm_lock_slab, lock);
 182                 CDEBUG(D_MALLOC, "kfreed 'lock': %d at %p (tot 0).\n",
 183                        sizeof(*lock), lock);
 184         } else
 185                 l_unlock(&ns->ns_lock);
 186
 187         EXIT;
 188 }
 189
 190 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
 191 {
 192         ENTRY;
 193         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 194         if (!list_empty(&lock->l_lru)) {
 195                 list_del_init(&lock->l_lru);
 196                 lock->l_resource->lr_namespace->ns_nr_unused--;
 197                 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
 198         }
 199         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 200         EXIT;
 201 }
 202
 203 void ldlm_lock_destroy(struct ldlm_lock *lock)
 204 {
 205         ENTRY;
 206         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 207
 208         if (!list_empty(&lock->l_children)) {
 209                 LDLM_DEBUG(lock, "still has children (%p)!",
 210                            lock->l_children.next);
 211                 ldlm_lock_dump(lock);
 212                 LBUG();
 213         }
 214         if (lock->l_readers || lock->l_writers) {
 215                 LDLM_DEBUG(lock, "lock still has references");
 216                 ldlm_lock_dump(lock);
 217         }
 218
 219         if (!list_empty(&lock->l_res_link)) {
 220                 ldlm_lock_dump(lock);
 221                 LBUG();
 222         }
 223
 224         if (lock->l_destroyed) {
 225                 LASSERT(list_empty(&lock->l_lru));
 226                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 227                 EXIT;
 228                 return;
 229         }
 230         lock->l_destroyed = 1;
 231
 232         list_del(&lock->l_export_chain);
 233         lock->l_export = NULL;
 234         ldlm_lock_remove_from_lru(lock);
 235
 236         /* Wake anyone waiting for this lock */
 237         /* FIXME: I should probably add yet another flag, instead of using
 238          * l_export to only call this on clients */
 239         if (lock->l_export && lock->l_completion_ast)
 240                 lock->l_completion_ast(lock, 0);
 241
 242         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 243         LDLM_LOCK_PUT(lock);
 244         EXIT;
 245 }
 246
 247 /*
 248    usage: pass in a resource on which you have done get
 249           pass in a parent lock on which you have done a get
 250           do not put the resource or the parent
 251    returns: lock with refcount 1
 252 */
 253 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 254                                        struct ldlm_resource *resource)
 255 {
 256         struct ldlm_lock *lock;
 257         ENTRY;
 258
 259         if (resource == NULL)
 260                 LBUG();
 261
 262         lock = kmem_cache_alloc(ldlm_lock_slab, SLAB_KERNEL);
 263         if (lock == NULL)
 264                 RETURN(NULL);
 265
 266         memset(lock, 0, sizeof(*lock));
 267         get_random_bytes(&lock->l_random, sizeof(__u64));
 268
 269         lock->l_resource = resource;
 270         /* this refcount matches the one of the resource passed
 271            in which is not being put away */
 272         atomic_set(&lock->l_refc, 1);
 273         INIT_LIST_HEAD(&lock->l_children);
 274         INIT_LIST_HEAD(&lock->l_res_link);
 275         INIT_LIST_HEAD(&lock->l_lru);
 276         INIT_LIST_HEAD(&lock->l_export_chain);
 277         INIT_LIST_HEAD(&lock->l_pending_chain);
 278         init_waitqueue_head(&lock->l_waitq);
 279
 280         spin_lock(&resource->lr_namespace->ns_counter_lock);
 281         resource->lr_namespace->ns_locks++;
 282         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 283
 284         if (parent != NULL) {
 285                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 286                 lock->l_parent = parent;
 287                 list_add(&lock->l_childof, &parent->l_children);
 288                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 289         }
 290
 291         CDEBUG(D_MALLOC, "kmalloced 'lock': %d at "
 292                "%p (tot %d).\n", sizeof(*lock), lock, 1);
 293         /* this is the extra refcount, to prevent the lock from evaporating */
 294         LDLM_LOCK_GET(lock);
 295         RETURN(lock);
 296 }
 297
 298 int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3])
 299 {
 300         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 301         struct ldlm_resource *oldres = lock->l_resource;
 302         int i, refc;
 303         ENTRY;
 304
 305         l_lock(&ns->ns_lock);
 306         if (memcmp(new_resid, lock->l_resource->lr_name,
 307                    sizeof(lock->l_resource->lr_name)) == 0) {
 308                 /* Nothing to do */
 309                 l_unlock(&ns->ns_lock);
 310                 RETURN(0);
 311         }
 312
 313         LASSERT(new_resid[0] != 0);
 314
 315         /* This function assumes that the lock isn't on any lists */
 316         LASSERT(list_empty(&lock->l_res_link));
 317
 318         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
 319                                              lock->l_resource->lr_type, 1);
 320         if (lock->l_resource == NULL) {
 321                 LBUG();
 322                 RETURN(-ENOMEM);
 323         }
 324
 325         /* move references over */
 326         refc = atomic_read(&lock->l_refc);
 327         for (i = 0; i < refc; i++) {
 328                 int rc;
 329                 ldlm_resource_getref(lock->l_resource);
 330                 rc = ldlm_resource_putref(oldres);
 331                 if (rc == 1 && i != refc - 1)
 332                         LBUG();
 333         }
 334         /* compensate for the initial get above.. */
 335         ldlm_resource_putref(lock->l_resource);
 336
 337         l_unlock(&ns->ns_lock);
 338         RETURN(0);
 339 }
 340
 341 /*
 342  *  HANDLES
 343  */
 344
 345 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 346 {
 347         lockh->addr = (__u64) (unsigned long)lock;
 348         lockh->cookie = lock->l_random;
 349 }
 350
 351 /*
 352  * if flags: atomically get the lock and set the flags.
 353  * Return NULL if flag already set
 354  */
 355
 356 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int strict,
 357                                      int flags)
 358 {
 359         struct ldlm_lock *lock = NULL, *retval = NULL;
 360         ENTRY;
 361
 362         LASSERT(handle);
 363
 364         if (!handle->addr)
 365                 RETURN(NULL);
 366
 367         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
 368         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock)) {
 369                 //CERROR("bogus lock %p\n", lock);
 370                 GOTO(out2, retval);
 371         }
 372
 373         if (lock->l_random != handle->cookie) {
 374                 //CERROR("bogus cookie: lock %p has "LPX64" vs. handle "LPX64
 375                 //       "\n", lock, lock->l_random, handle->cookie);
 376                 GOTO(out2, NULL);
 377         }
 378         if (!lock->l_resource) {
 379                 CERROR("trying to lock bogus resource: lock %p\n", lock);
 380                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 381                 GOTO(out2, retval);
 382         }
 383         if (!lock->l_resource->lr_namespace) {
 384                 CERROR("trying to lock bogus namespace: lock %p\n", lock);
 385                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 386                 GOTO(out2, retval);
 387         }
 388
 389         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 390         if (strict && lock->l_destroyed) {
 391                 CERROR("lock already destroyed: lock %p\n", lock);
 392                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 393                 GOTO(out, NULL);
 394         }
 395
 396         if (flags && (lock->l_flags & flags))
 397                 GOTO(out, NULL);
 398
 399         if (flags)
 400                 lock->l_flags |= flags;
 401
 402         retval = LDLM_LOCK_GET(lock);
 403         if (!retval)
 404                 CERROR("lock disappeared below us!!! %p\n", lock);
 405         EXIT;
 406  out:
 407         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 408  out2:
 409         return retval;
 410 }
 411
 412 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 413 {
 414         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 415 }
 416
 417 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 418 {
 419         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 420         desc->l_req_mode = lock->l_req_mode;
 421         desc->l_granted_mode = lock->l_granted_mode;
 422         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 423         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 424 }
 425
 426 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 427                                    struct ldlm_lock *new)
 428 {
 429         struct ldlm_ast_work *w;
 430         ENTRY;
 431
 432         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 433         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 434                 GOTO(out, 0);
 435
 436         OBD_ALLOC(w, sizeof(*w));
 437         if (!w) {
 438                 LBUG();
 439                 GOTO(out, 0);
 440         }
 441
 442         if (new) {
 443                 lock->l_flags |= LDLM_FL_AST_SENT;
 444                 w->w_blocking = 1;
 445                 ldlm_lock2desc(new, &w->w_desc);
 446         }
 447
 448         w->w_lock = LDLM_LOCK_GET(lock);
 449         list_add(&w->w_list, lock->l_resource->lr_tmp);
 450       out:
 451         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 452         return;
 453 }
 454
 455 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 456 {
 457         struct ldlm_lock *lock;
 458
 459         lock = ldlm_handle2lock(lockh);
 460         ldlm_lock_addref_internal(lock, mode);
 461         LDLM_LOCK_PUT(lock);
 462 }
 463
 464 /* only called for local locks */
 465 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 466 {
 467         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 468         ldlm_lock_remove_from_lru(lock);
 469         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 470                 lock->l_readers++;
 471         else
 472                 lock->l_writers++;
 473         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 474         LDLM_LOCK_GET(lock);
 475         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 476 }
 477
 478 /* Args: unlocked lock */
 479 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 480                                     __u64 *res_id, int flags);
 481
 482 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 483 {
 484         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0, 0);
 485         struct ldlm_namespace *ns;
 486         ENTRY;
 487
 488         if (lock == NULL)
 489                 LBUG();
 490
 491         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 492         ns = lock->l_resource->lr_namespace;
 493         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 494         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 495                 lock->l_readers--;
 496         else
 497                 lock->l_writers--;
 498
 499         /* If we received a blocked AST and this was the last reference,
 500          * run the callback. */
 501         if (!lock->l_readers && !lock->l_writers &&
 502             (lock->l_flags & LDLM_FL_CBPENDING)) {
 503                 if (!lock->l_resource->lr_namespace->ns_client &&
 504                     lock->l_export)
 505                         CERROR("FL_CBPENDING set on non-local lock--just a "
 506                                "warning\n");
 507
 508                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 509                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 510
 511                 /* FIXME: need a real 'desc' here */
 512                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 513                                      lock->l_data_len, LDLM_CB_BLOCKING);
 514         } else if (ns->ns_client && !lock->l_readers && !lock->l_writers) {
 515                 LASSERT(list_empty(&lock->l_lru));
 516                 LASSERT(ns->ns_nr_unused >= 0);
 517                 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 518                 ns->ns_nr_unused++;
 519                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 520                 ldlm_cancel_lru(ns);
 521         } else
 522                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 523
 524         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 525         LDLM_LOCK_PUT(lock);    /* matches the handle2lock above */
 526
 527         EXIT;
 528 }
 529
 530 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 531                                  struct list_head *queue)
 532 {
 533         struct list_head *tmp, *pos;
 534         int rc = 1;
 535
 536         list_for_each_safe(tmp, pos, queue) {
 537                 struct ldlm_lock *child;
 538                 ldlm_res_compat compat;
 539
 540                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 541                 if (lock == child)
 542                         continue;
 543
 544                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 545                 if (compat && compat(child, lock)) {
 546                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 547                         continue;
 548                 }
 549                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 550                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 551                         continue;
 552                 }
 553
 554                 rc = 0;
 555
 556                 if (send_cbs && child->l_blocking_ast != NULL) {
 557                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 558                                "blocking AST.\n", child);
 559                         ldlm_add_ast_work_item(child, lock);
 560                 }
 561         }
 562
 563         return rc;
 564 }
 565
 566 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 567 {
 568         int rc;
 569         ENTRY;
 570
 571         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 572         rc = ldlm_lock_compat_list(lock, send_cbs,
 573                                    &lock->l_resource->lr_granted);
 574         /* FIXME: should we be sending ASTs to converting? */
 575         if (rc)
 576                 rc = ldlm_lock_compat_list
 577                         (lock, send_cbs, &lock->l_resource->lr_converting);
 578
 579         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 580         RETURN(rc);
 581 }
 582
 583 /* NOTE: called by
 584    - ldlm_handle_enqueuque - resource
 585 */
 586 void ldlm_grant_lock(struct ldlm_lock *lock)
 587 {
 588         struct ldlm_resource *res = lock->l_resource;
 589         ENTRY;
 590
 591         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 592         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 593         lock->l_granted_mode = lock->l_req_mode;
 594
 595         if (lock->l_granted_mode < res->lr_most_restr)
 596                 res->lr_most_restr = lock->l_granted_mode;
 597
 598         if (lock->l_completion_ast) {
 599                 ldlm_add_ast_work_item(lock, NULL);
 600         }
 601         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 602         EXIT;
 603 }
 604
 605 /* returns a referenced lock or NULL */
 606 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 607                                       struct ldlm_extent *extent,
 608                                       struct ldlm_lock *old_lock)
 609 {
 610         struct ldlm_lock *lock;
 611         struct list_head *tmp;
 612
 613         list_for_each(tmp, queue) {
 614                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 615
 616                 if (lock == old_lock)
 617                         continue;
 618
 619                 if (lock->l_flags & LDLM_FL_CBPENDING)
 620                         continue;
 621
 622                 if (lock->l_req_mode != mode)
 623                         continue;
 624
 625                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 626                     (lock->l_extent.start > extent->start ||
 627                      lock->l_extent.end < extent->end))
 628                         continue;
 629
 630                 if (lock->l_destroyed)
 631                         continue;
 632
 633                 ldlm_lock_addref_internal(lock, mode);
 634                 return lock;
 635         }
 636
 637         return NULL;
 638 }
 639
 640 /* Can be called in two ways:
 641  *
 642  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
 643  * for a duplicate of.
 644  *
 645  * Otherwise, all of the fields must be filled in, to match against.
 646  *
 647  * Returns 1 if it finds an already-existing lock that is compatible; in this
 648  * case, lockh is filled in with a addref()ed lock
 649  */
 650 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 *res_id, __u32 type,
 651                     void *cookie, int cookielen, ldlm_mode_t mode,
 652                     struct lustre_handle *lockh)
 653 {
 654         struct ldlm_resource *res;
 655         struct ldlm_lock *lock, *old_lock = NULL;
 656         int rc = 0;
 657         ENTRY;
 658
 659         if (ns == NULL) {
 660                 old_lock = ldlm_handle2lock(lockh);
 661                 LASSERT(old_lock);
 662
 663                 ns = old_lock->l_resource->lr_namespace;
 664                 res_id = old_lock->l_resource->lr_name;
 665                 type = old_lock->l_resource->lr_type;
 666                 mode = old_lock->l_req_mode;
 667         }
 668
 669         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 670         if (res == NULL) {
 671                 LASSERT(old_lock == NULL);
 672                 RETURN(0);
 673         }
 674
 675         l_lock(&ns->ns_lock);
 676
 677         if ((lock = search_queue(&res->lr_granted, mode, cookie, old_lock)))
 678                 GOTO(out, rc = 1);
 679         if ((lock = search_queue(&res->lr_converting, mode, cookie, old_lock)))
 680                 GOTO(out, rc = 1);
 681         if ((lock = search_queue(&res->lr_waiting, mode, cookie, old_lock)))
 682                 GOTO(out, rc = 1);
 683
 684         EXIT;
 685        out:
 686         ldlm_resource_putref(res);
 687         l_unlock(&ns->ns_lock);
 688
 689         if (lock) {
 690                 ldlm_lock2handle(lock, lockh);
 691                 if (lock->l_completion_ast)
 692                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 693         }
 694         if (rc)
 695                 LDLM_DEBUG(lock, "matched");
 696         else
 697                 LDLM_DEBUG_NOLOCK("not matched");
 698
 699         if (old_lock)
 700                 LDLM_LOCK_PUT(old_lock);
 701
 702         return rc;
 703 }
 704
 705 /* Returns a referenced lock */
 706 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 707                                    struct lustre_handle *parent_lock_handle,
 708                                    __u64 * res_id, __u32 type,
 709                                    ldlm_mode_t mode, void *data, __u32 data_len)
 710 {
 711         struct ldlm_resource *res, *parent_res = NULL;
 712         struct ldlm_lock *lock, *parent_lock = NULL;
 713
 714         if (parent_lock_handle) {
 715                 parent_lock = ldlm_handle2lock(parent_lock_handle);
 716                 if (parent_lock)
 717                         parent_res = parent_lock->l_resource;
 718         }
 719
 720         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 721         if (res == NULL)
 722                 RETURN(NULL);
 723
 724         lock = ldlm_lock_new(parent_lock, res);
 725         if (lock == NULL) {
 726                 ldlm_resource_putref(res);
 727                 RETURN(NULL);
 728         }
 729
 730         lock->l_req_mode = mode;
 731         lock->l_data = data;
 732         lock->l_data_len = data_len;
 733
 734         return lock;
 735 }
 736
 737 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 738 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock * lock,
 739                                void *cookie, int cookie_len,
 740                                int *flags,
 741                                ldlm_completion_callback completion,
 742                                ldlm_blocking_callback blocking)
 743 {
 744         struct ldlm_resource *res;
 745         int local;
 746         ldlm_res_policy policy;
 747         ENTRY;
 748
 749         res = lock->l_resource;
 750         lock->l_blocking_ast = blocking;
 751
 752         if (res->lr_type == LDLM_EXTENT)
 753                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 754
 755         /* policies are not executed on the client or during replay */
 756         local = res->lr_namespace->ns_client;
 757         if (!local && !(*flags & LDLM_FL_REPLAY) &&
 758             (policy = ldlm_res_policy_table[res->lr_type])) {
 759                 int rc;
 760                 rc = policy(lock, cookie, lock->l_req_mode, *flags, NULL);
 761
 762                 if (rc == ELDLM_LOCK_CHANGED) {
 763                         res = lock->l_resource;
 764                         *flags |= LDLM_FL_LOCK_CHANGED;
 765                 } else if (rc == ELDLM_LOCK_ABORTED) {
 766                         ldlm_lock_destroy(lock);
 767                         RETURN(rc);
 768                 }
 769         }
 770
 771         l_lock(&res->lr_namespace->ns_lock);
 772         if (local && lock->l_req_mode == lock->l_granted_mode) {
 773                 /* The server returned a blocked lock, but it was granted before
 774                  * we got a chance to actually enqueue it.  We don't need to do
 775                  * anything else. */
 776                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 777                           LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 778                 GOTO(out, ELDLM_OK);
 779         }
 780
 781         /* This distinction between local lock trees is very important; a client
 782          * namespace only has information about locks taken by that client, and
 783          * thus doesn't have enough information to decide for itself if it can
 784          * be granted (below).  In this case, we do exactly what the server
 785          * tells us to do, as dictated by the 'flags'.
 786          *
 787          * We do exactly the same thing during recovery, when the server is
 788          * more or less trusting the clients not to lie.
 789          *
 790          * FIXME (bug 629283): Detect obvious lies by checking compatibility in
 791          * granted/converting queues. */
 792         ldlm_resource_unlink_lock(lock);
 793         if (local || (*flags & LDLM_FL_REPLAY)) {
 794                 if (*flags & LDLM_FL_BLOCK_CONV)
 795                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 796                                                lock);
 797                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 798                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 799                 else
 800                         ldlm_grant_lock(lock);
 801                 GOTO(out, ELDLM_OK);
 802         }
 803
 804         /* FIXME: We may want to optimize by checking lr_most_restr */
 805         if (!list_empty(&res->lr_converting)) {
 806                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 807                 *flags |= LDLM_FL_BLOCK_CONV;
 808                 GOTO(out, ELDLM_OK);
 809         }
 810         if (!list_empty(&res->lr_waiting)) {
 811                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 812                 *flags |= LDLM_FL_BLOCK_WAIT;
 813                 GOTO(out, ELDLM_OK);
 814         }
 815         if (!ldlm_lock_compat(lock, 0)) {
 816                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 817                 *flags |= LDLM_FL_BLOCK_GRANTED;
 818                 GOTO(out, ELDLM_OK);
 819         }
 820
 821         ldlm_grant_lock(lock);
 822         EXIT;
 823       out:
 824         l_unlock(&res->lr_namespace->ns_lock);
 825         /* Don't set 'completion_ast' until here so that if the lock is granted
 826          * immediately we don't do an unnecessary completion call. */
 827         lock->l_completion_ast = completion;
 828         return ELDLM_OK;
 829 }
 830
 831 /* Must be called with namespace taken: queue is waiting or converting. */
 832 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 833                                 struct list_head *queue)
 834 {
 835         struct list_head *tmp, *pos;
 836         ENTRY;
 837
 838         list_for_each_safe(tmp, pos, queue) {
 839                 struct ldlm_lock *pending;
 840                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 841
 842                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 843
 844                 if (!ldlm_lock_compat(pending, 1))
 845                         RETURN(1);
 846
 847                 list_del_init(&pending->l_res_link);
 848                 ldlm_grant_lock(pending);
 849         }
 850
 851         RETURN(0);
 852 }
 853
 854 void ldlm_run_ast_work(struct list_head *rpc_list)
 855 {
 856         struct list_head *tmp, *pos;
 857         int rc;
 858         ENTRY;
 859
 860         list_for_each_safe(tmp, pos, rpc_list) {
 861                 struct ldlm_ast_work *w =
 862                         list_entry(tmp, struct ldlm_ast_work, w_list);
 863
 864                 if (w->w_blocking)
 865                         rc = w->w_lock->l_blocking_ast
 866                                 (w->w_lock, &w->w_desc, w->w_data,
 867                                  w->w_datalen, LDLM_CB_BLOCKING);
 868                 else
 869                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags);
 870                 if (rc)
 871                         CERROR("Failed AST - should clean & disconnect "
 872                                "client\n");
 873                 LDLM_LOCK_PUT(w->w_lock);
 874                 list_del(&w->w_list);
 875                 OBD_FREE(w, sizeof(*w));
 876         }
 877         EXIT;
 878 }
 879
 880 /* Must be called with resource->lr_lock not taken. */
 881 void ldlm_reprocess_all(struct ldlm_resource *res)
 882 {
 883         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 884         ENTRY;
 885
 886         /* Local lock trees don't get reprocessed. */
 887         if (res->lr_namespace->ns_client) {
 888                 EXIT;
 889                 return;
 890         }
 891
 892         l_lock(&res->lr_namespace->ns_lock);
 893         res->lr_tmp = &rpc_list;
 894
 895         ldlm_reprocess_queue(res, &res->lr_converting);
 896         if (list_empty(&res->lr_converting))
 897                 ldlm_reprocess_queue(res, &res->lr_waiting);
 898
 899         res->lr_tmp = NULL;
 900         l_unlock(&res->lr_namespace->ns_lock);
 901
 902         ldlm_run_ast_work(&rpc_list);
 903         EXIT;
 904 }
 905
 906 void ldlm_cancel_callback(struct ldlm_lock *lock)
 907 {
 908         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 909         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
 910                 lock->l_flags |= LDLM_FL_CANCEL;
 911                 if (lock->l_blocking_ast)
 912                         lock->l_blocking_ast(lock, NULL, lock->l_data,
 913                                              lock->l_data_len,
 914                                              LDLM_CB_CANCELING);
 915                 else
 916                         LDLM_DEBUG(lock, "no blocking ast");
 917         }
 918         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 919 }
 920
 921 void ldlm_lock_cancel(struct ldlm_lock *lock)
 922 {
 923         struct ldlm_resource *res;
 924         struct ldlm_namespace *ns;
 925         ENTRY;
 926
 927         res = lock->l_resource;
 928         ns = res->lr_namespace;
 929
 930         l_lock(&ns->ns_lock);
 931         if (lock->l_readers || lock->l_writers)
 932                 LDLM_DEBUG(lock, "lock still has references");
 933
 934         ldlm_cancel_callback(lock);
 935
 936         ldlm_del_waiting_lock(lock);
 937         ldlm_resource_unlink_lock(lock);
 938         ldlm_lock_destroy(lock);
 939         l_unlock(&ns->ns_lock);
 940         EXIT;
 941 }
 942
 943 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data, int datalen)
 944 {
 945         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 946         ENTRY;
 947
 948         if (lock == NULL)
 949                 RETURN(-EINVAL);
 950
 951         lock->l_data = data;
 952         lock->l_data_len = datalen;
 953
 954         LDLM_LOCK_PUT(lock);
 955
 956         RETURN(0);
 957 }
 958
 959 void ldlm_cancel_locks_for_export(struct obd_export *exp)
 960 {
 961         struct list_head *iter, *n; /* MUST BE CALLED "n"! */
 962
 963         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
 964                 struct ldlm_lock *lock;
 965                 struct ldlm_resource *res;
 966                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 967                 res = ldlm_resource_getref(lock->l_resource);
 968                 LDLM_DEBUG(lock, "export %p", exp);
 969                 ldlm_lock_cancel(lock);
 970                 ldlm_reprocess_all(res);
 971                 ldlm_resource_putref(res);
 972         }
 973 }
 974
 975 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 976                                         int *flags)
 977 {
 978         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 979         struct ldlm_resource *res;
 980         struct ldlm_namespace *ns;
 981         int granted = 0;
 982         ENTRY;
 983
 984         res = lock->l_resource;
 985         ns = res->lr_namespace;
 986
 987         l_lock(&ns->ns_lock);
 988
 989         lock->l_req_mode = new_mode;
 990         ldlm_resource_unlink_lock(lock);
 991
 992         /* If this is a local resource, put it on the appropriate list. */
 993         if (res->lr_namespace->ns_client) {
 994                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED))
 995                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 996                                                lock);
 997                 else {
 998                         /* This should never happen, because of the way the
 999                          * server handles conversions. */
1000                         LBUG();
1001
1002                         res->lr_tmp = &rpc_list;
1003                         ldlm_grant_lock(lock);
1004                         res->lr_tmp = NULL;
1005                         granted = 1;
1006                         /* FIXME: completion handling not with ns_lock held ! */
1007                         if (lock->l_completion_ast)
1008                                 lock->l_completion_ast(lock, 0);
1009                 }
1010         } else {
1011                 /* FIXME: We should try the conversion right away and possibly
1012                  * return success without the need for an extra AST */
1013                 ldlm_resource_add_lock(res, res->lr_converting.prev, lock);
1014                 *flags |= LDLM_FL_BLOCK_CONV;
1015         }
1016
1017         l_unlock(&ns->ns_lock);
1018
1019         if (granted)
1020                 ldlm_run_ast_work(&rpc_list);
1021         RETURN(res);
1022 }
1023
1024 void ldlm_lock_dump(struct ldlm_lock *lock)
1025 {
1026         char ver[128];
1027
1028         if (!(portal_debug & D_OTHER))
1029                 return;
1030
1031         if (RES_VERSION_SIZE != 4)
1032                 LBUG();
1033
1034         if (!lock) {
1035                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
1036                 return;
1037         }
1038
1039         snprintf(ver, sizeof(ver), "%x %x %x %x",
1040                  lock->l_version[0], lock->l_version[1],
1041                  lock->l_version[2], lock->l_version[3]);
1042
1043         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
1044         if (lock->l_export && lock->l_export->exp_connection)
1045                 CDEBUG(D_OTHER, "  Node: NID %x (rhandle: "LPX64")\n",
1046                        lock->l_export->exp_connection->c_peer.peer_nid,
1047                        lock->l_remote_handle.addr);
1048         else
1049                 CDEBUG(D_OTHER, "  Node: local\n");
1050         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
1051         CDEBUG(D_OTHER, "  Resource: %p ("LPD64")\n", lock->l_resource,
1052                lock->l_resource->lr_name[0]);
1053         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
1054                (int)lock->l_req_mode, (int)lock->l_granted_mode);
1055         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
1056                lock->l_readers, lock->l_writers);
1057         if (lock->l_resource->lr_type == LDLM_EXTENT)
1058                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
1059                        (unsigned long long)lock->l_extent.start,
1060                        (unsigned long long)lock->l_extent.end);
1061 }