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 /* lock types */
  34 char *ldlm_lockname[] = {
  35         [0] "--",
  36         [LCK_EX] "EX",
  37         [LCK_PW] "PW",
  38         [LCK_PR] "PR",
  39         [LCK_CW] "CW",
  40         [LCK_CR] "CR",
  41         [LCK_NL] "NL"
  42 };
  43 char *ldlm_typename[] = {
  44         [LDLM_PLAIN] "PLN",
  45         [LDLM_EXTENT] "EXT",
  46         [LDLM_MDSINTENT] "INT"
  47 };
  48
  49 char *ldlm_it2str(int it)
  50 {
  51         switch (it) {
  52         case IT_OPEN:
  53                 return "open";
  54         case IT_CREAT:
  55                 return "creat";
  56         case (IT_OPEN | IT_CREAT):
  57                 return "open|creat";
  58         case IT_MKDIR:
  59                 return "mkdir";
  60         case IT_LINK:
  61                 return "link";
  62         case IT_SYMLINK:
  63                 return "symlink";
  64         case IT_UNLINK:
  65                 return "unlink";
  66         case IT_RMDIR:
  67                 return "rmdir";
  68         case IT_RENAME:
  69                 return "rename";
  70         case IT_RENAME2:
  71                 return "rename2";
  72         case IT_READDIR:
  73                 return "readdir";
  74         case IT_GETATTR:
  75                 return "getattr";
  76         case IT_SETATTR:
  77                 return "setattr";
  78         case IT_READLINK:
  79                 return "readlink";
  80         case IT_MKNOD:
  81                 return "mknod";
  82         case IT_LOOKUP:
  83                 return "lookup";
  84         default:
  85                 CERROR("Unknown intent %d\n", it);
  86                 return "UNKNOWN";
  87         }
  88 }
  89
  90 extern kmem_cache_t *ldlm_lock_slab;
  91
  92 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b);
  93
  94 ldlm_res_compat ldlm_res_compat_table[] = {
  95         [LDLM_PLAIN] ldlm_plain_compat,
  96         [LDLM_EXTENT] ldlm_extent_compat,
  97         [LDLM_MDSINTENT] ldlm_plain_compat
  98 };
  99
 100 ldlm_res_policy ldlm_res_policy_table[] = {
 101         [LDLM_PLAIN] NULL,
 102         [LDLM_EXTENT] ldlm_extent_policy,
 103         [LDLM_MDSINTENT] NULL
 104 };
 105
 106 void ldlm_register_intent(int (*arg) (struct ldlm_lock * lock, void *req_cookie,
 107                                       ldlm_mode_t mode, void *data))
 108 {
 109         ldlm_res_policy_table[LDLM_MDSINTENT] = arg;
 110 }
 111
 112 void ldlm_unregister_intent(void)
 113 {
 114         ldlm_res_policy_table[LDLM_MDSINTENT] = NULL;
 115 }
 116
 117 /*
 118  * REFCOUNTED LOCK OBJECTS
 119  */
 120
 121
 122 /*
 123  * Lock refcounts, during creation:
 124  *   - one special one for allocation, dec'd only once in destroy
 125  *   - one for being a lock that's in-use
 126  *   - one for the addref associated with a new lock
 127  */
 128 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 129 {
 130         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 131         lock->l_refc++;
 132         ldlm_resource_getref(lock->l_resource);
 133         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 134         return lock;
 135 }
 136
 137 void ldlm_lock_put(struct ldlm_lock *lock)
 138 {
 139         struct lustre_lock *nslock = &lock->l_resource->lr_namespace->ns_lock;
 140         ENTRY;
 141
 142         l_lock(nslock);
 143         lock->l_refc--;
 144         //LDLM_DEBUG(lock, "after refc--");
 145         if (lock->l_refc < 0)
 146                 LBUG();
 147
 148         ldlm_resource_put(lock->l_resource);
 149         if (lock->l_parent)
 150                 LDLM_LOCK_PUT(lock->l_parent);
 151
 152         if (lock->l_refc == 0 && (lock->l_flags & LDLM_FL_DESTROYED)) {
 153                 lock->l_resource = NULL;
 154                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
 155                 if (lock->l_export && lock->l_export->exp_connection)
 156                         ptlrpc_put_connection(lock->l_export->exp_connection);
 157                 CDEBUG(D_MALLOC, "kfreed 'lock': %d at %p (tot 1).\n",
 158                        sizeof(*lock), lock);
 159                 kmem_cache_free(ldlm_lock_slab, lock);
 160         }
 161         l_unlock(nslock);
 162         EXIT;
 163 }
 164
 165 void ldlm_lock_destroy(struct ldlm_lock *lock)
 166 {
 167         ENTRY;
 168         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 169
 170         if (!list_empty(&lock->l_children)) {
 171                 LDLM_DEBUG(lock, "still has children (%p)!",
 172                            lock->l_children.next);
 173                 ldlm_lock_dump(lock);
 174                 LBUG();
 175         }
 176         if (lock->l_readers || lock->l_writers) {
 177                 LDLM_DEBUG(lock, "lock still has references");
 178                 ldlm_lock_dump(lock);
 179                 LBUG();
 180         }
 181
 182         if (!list_empty(&lock->l_res_link)) {
 183                 ldlm_lock_dump(lock);
 184                 LBUG();
 185         }
 186
 187         if (lock->l_flags & LDLM_FL_DESTROYED) {
 188                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 189                 EXIT;
 190                 return;
 191         }
 192
 193         list_del(&lock->l_export_chain);
 194         lock->l_export = NULL;
 195         lock->l_flags |= LDLM_FL_DESTROYED;
 196
 197         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 198         LDLM_LOCK_PUT(lock);
 199         EXIT;
 200 }
 201
 202 /*
 203    usage: pass in a resource on which you have done get
 204           pass in a parent lock on which you have done a get
 205           do not put the resource or the parent
 206    returns: lock with refcount 1
 207 */
 208 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 209                                        struct ldlm_resource *resource)
 210 {
 211         struct ldlm_lock *lock;
 212         ENTRY;
 213
 214         if (resource == NULL)
 215                 LBUG();
 216
 217         lock = kmem_cache_alloc(ldlm_lock_slab, SLAB_KERNEL);
 218         if (lock == NULL)
 219                 RETURN(NULL);
 220         CDEBUG(D_MALLOC, "kmalloced 'lock': %d at "
 221                "%p (tot %d).\n", sizeof(*lock), lock, 1);
 222
 223         memset(lock, 0, sizeof(*lock));
 224         get_random_bytes(&lock->l_random, sizeof(__u64));
 225
 226         lock->l_resource = resource;
 227         /* this refcount matches the one of the resource passed
 228            in which is not being put away */
 229         lock->l_refc = 1;
 230         INIT_LIST_HEAD(&lock->l_children);
 231         INIT_LIST_HEAD(&lock->l_res_link);
 232         INIT_LIST_HEAD(&lock->l_export_chain);
 233         INIT_LIST_HEAD(&lock->l_pending_chain);
 234         init_waitqueue_head(&lock->l_waitq);
 235
 236         if (parent != NULL) {
 237                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 238                 lock->l_parent = parent;
 239                 list_add(&lock->l_childof, &parent->l_children);
 240                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 241         }
 242         /* this is the extra refcount, to prevent the lock
 243            evaporating */
 244         LDLM_LOCK_GET(lock);
 245         RETURN(lock);
 246 }
 247
 248 int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3])
 249 {
 250         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 251         struct ldlm_resource *oldres = lock->l_resource;
 252         int type, i;
 253         ENTRY;
 254
 255         l_lock(&ns->ns_lock);
 256         if (memcmp(new_resid, lock->l_resource->lr_name,
 257                    sizeof(lock->l_resource->lr_name)) == 0) {
 258                 /* Nothing to do */
 259                 l_unlock(&ns->ns_lock);
 260                 RETURN(0);
 261         }
 262
 263         type = lock->l_resource->lr_type;
 264         if (new_resid[0] == 0)
 265                 LBUG();
 266         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid, type, 1);
 267         if (lock->l_resource == NULL) {
 268                 LBUG();
 269                 RETURN(-ENOMEM);
 270         }
 271
 272         /* move references over */
 273         for (i = 0; i < lock->l_refc; i++) {
 274                 int rc;
 275                 ldlm_resource_getref(lock->l_resource);
 276                 rc = ldlm_resource_put(oldres);
 277                 if (rc == 1 && i != lock->l_refc - 1)
 278                         LBUG();
 279         }
 280         /* compensate for the initial get above.. */
 281         ldlm_resource_put(lock->l_resource);
 282
 283         l_unlock(&ns->ns_lock);
 284         RETURN(0);
 285 }
 286
 287 /*
 288  *  HANDLES
 289  */
 290
 291 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 292 {
 293         lockh->addr = (__u64) (unsigned long)lock;
 294         lockh->cookie = lock->l_random;
 295 }
 296
 297 struct ldlm_lock *ldlm_handle2lock(struct lustre_handle *handle)
 298 {
 299         struct ldlm_lock *lock = NULL, *retval = NULL;
 300         ENTRY;
 301
 302         if (!handle || !handle->addr)
 303                 RETURN(NULL);
 304
 305         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
 306         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock))
 307                 RETURN(NULL);
 308
 309         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 310         if (lock->l_random != handle->cookie)
 311                 GOTO(out, NULL);
 312
 313         if (lock->l_flags & LDLM_FL_DESTROYED)
 314                 GOTO(out, NULL);
 315
 316         retval = LDLM_LOCK_GET(lock);
 317         EXIT;
 318       out:
 319         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 320         return retval;
 321 }
 322
 323 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 324 {
 325         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 326 }
 327
 328 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 329 {
 330         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 331         desc->l_req_mode = lock->l_req_mode;
 332         desc->l_granted_mode = lock->l_granted_mode;
 333         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 334         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 335 }
 336
 337 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 338                                    struct ldlm_lock *new)
 339 {
 340         struct ldlm_ast_work *w;
 341         ENTRY;
 342
 343         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 344         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 345                 GOTO(out, 0);
 346
 347         OBD_ALLOC(w, sizeof(*w));
 348         if (!w) {
 349                 LBUG();
 350                 GOTO(out, 0);
 351         }
 352
 353         if (new) {
 354                 lock->l_flags |= LDLM_FL_AST_SENT;
 355                 w->w_blocking = 1;
 356                 ldlm_lock2desc(new, &w->w_desc);
 357         }
 358
 359         w->w_lock = LDLM_LOCK_GET(lock);
 360         list_add(&w->w_list, lock->l_resource->lr_tmp);
 361       out:
 362         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 363         return;
 364 }
 365
 366 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 367 {
 368         struct ldlm_lock *lock;
 369
 370         lock = ldlm_handle2lock(lockh);
 371         ldlm_lock_addref_internal(lock, mode);
 372         LDLM_LOCK_PUT(lock);
 373 }
 374
 375 /* only called for local locks */
 376 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 377 {
 378         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 379         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 380                 lock->l_readers++;
 381         else
 382                 lock->l_writers++;
 383         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 384         LDLM_LOCK_GET(lock);
 385         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 386 }
 387
 388 /* Args: unlocked lock */
 389 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 390 {
 391         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 392         ENTRY;
 393
 394         if (lock == NULL)
 395                 LBUG();
 396
 397         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 398         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 399         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 400                 lock->l_readers--;
 401         else
 402                 lock->l_writers--;
 403
 404         /* If we received a blocked AST and this was the last reference,
 405          * run the callback. */
 406         if (!lock->l_readers && !lock->l_writers &&
 407             (lock->l_flags & LDLM_FL_CBPENDING)) {
 408                 if (!lock->l_resource->lr_namespace->ns_client) {
 409                         CERROR("LDLM_FL_CBPENDING set on non-local lock!\n");
 410                         LBUG();
 411                 }
 412
 413                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 414                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 415
 416                 /* FIXME: need a real 'desc' here */
 417                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 418                                      lock->l_data_len);
 419         } else
 420                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 421
 422         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 423         LDLM_LOCK_PUT(lock);    /* matches the handle2lock above */
 424
 425         EXIT;
 426 }
 427
 428 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 429                                  struct list_head *queue)
 430 {
 431         struct list_head *tmp, *pos;
 432         int rc = 1;
 433
 434         list_for_each_safe(tmp, pos, queue) {
 435                 struct ldlm_lock *child;
 436                 ldlm_res_compat compat;
 437
 438                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 439                 if (lock == child)
 440                         continue;
 441
 442                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 443                 if (compat && compat(child, lock)) {
 444                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 445                         continue;
 446                 }
 447                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 448                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 449                         continue;
 450                 }
 451
 452                 rc = 0;
 453
 454                 if (send_cbs && child->l_blocking_ast != NULL) {
 455                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 456                                "blocking AST.\n", child);
 457                         ldlm_add_ast_work_item(child, lock);
 458                 }
 459         }
 460
 461         return rc;
 462 }
 463
 464 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 465 {
 466         int rc;
 467         ENTRY;
 468
 469         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 470         rc = ldlm_lock_compat_list(lock, send_cbs,
 471                                    &lock->l_resource->lr_granted);
 472         /* FIXME: should we be sending ASTs to converting? */
 473         if (rc)
 474                 rc = ldlm_lock_compat_list
 475                         (lock, send_cbs, &lock->l_resource->lr_converting);
 476
 477         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 478         RETURN(rc);
 479 }
 480
 481 /* NOTE: called by
 482    - ldlm_handle_enqueuque - resource
 483 */
 484 void ldlm_grant_lock(struct ldlm_lock *lock)
 485 {
 486         struct ldlm_resource *res = lock->l_resource;
 487         ENTRY;
 488
 489         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 490         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 491         lock->l_granted_mode = lock->l_req_mode;
 492
 493         if (lock->l_granted_mode < res->lr_most_restr)
 494                 res->lr_most_restr = lock->l_granted_mode;
 495
 496         if (lock->l_completion_ast) {
 497                 ldlm_add_ast_work_item(lock, NULL);
 498         }
 499         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 500         EXIT;
 501 }
 502
 503 /* returns a referenced lock or NULL */
 504 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 505                                       struct ldlm_extent *extent)
 506 {
 507         struct ldlm_lock *lock;
 508         struct list_head *tmp;
 509
 510         list_for_each(tmp, queue) {
 511                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 512
 513                 if (lock->l_flags & (LDLM_FL_CBPENDING | LDLM_FL_DESTROYED))
 514                         continue;
 515
 516                 /* lock_convert() takes the resource lock, so we're sure that
 517                  * req_mode, lr_type, and l_cookie won't change beneath us */
 518                 if (lock->l_req_mode != mode)
 519                         continue;
 520
 521                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 522                     (lock->l_extent.start > extent->start ||
 523                      lock->l_extent.end < extent->end))
 524                         continue;
 525
 526                 ldlm_lock_addref_internal(lock, mode);
 527                 return lock;
 528         }
 529
 530         return NULL;
 531 }
 532
 533 /* Must be called with no resource or lock locks held.
 534  *
 535  * Returns 1 if it finds an already-existing lock that is compatible; in this
 536  * case, lockh is filled in with a addref()ed lock
 537 */
 538 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 * res_id, __u32 type,
 539                     void *cookie, int cookielen, ldlm_mode_t mode,
 540                     struct lustre_handle *lockh)
 541 {
 542         struct ldlm_resource *res;
 543         struct ldlm_lock *lock;
 544         int rc = 0;
 545         ENTRY;
 546
 547         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 548         if (res == NULL)
 549                 RETURN(0);
 550
 551         ns = res->lr_namespace;
 552         l_lock(&ns->ns_lock);
 553
 554         if ((lock = search_queue(&res->lr_granted, mode, cookie)))
 555                 GOTO(out, rc = 1);
 556         if ((lock = search_queue(&res->lr_converting, mode, cookie)))
 557                 GOTO(out, rc = 1);
 558         if ((lock = search_queue(&res->lr_waiting, mode, cookie)))
 559                 GOTO(out, rc = 1);
 560
 561         EXIT;
 562       out:
 563         ldlm_resource_put(res);
 564         l_unlock(&ns->ns_lock);
 565
 566         if (lock) {
 567                 ldlm_lock2handle(lock, lockh);
 568                 if (lock->l_completion_ast)
 569                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 570         }
 571         if (rc)
 572                 LDLM_DEBUG(lock, "matched");
 573         else
 574                 LDLM_DEBUG_NOLOCK("not matched");
 575         return rc;
 576 }
 577
 578 /* Returns a referenced lock */
 579 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 580                                    struct lustre_handle *parent_lock_handle,
 581                                    __u64 * res_id, __u32 type,
 582                                    ldlm_mode_t mode, void *data, __u32 data_len)
 583 {
 584         struct ldlm_resource *res, *parent_res = NULL;
 585         struct ldlm_lock *lock, *parent_lock;
 586
 587         parent_lock = ldlm_handle2lock(parent_lock_handle);
 588         if (parent_lock)
 589                 parent_res = parent_lock->l_resource;
 590
 591         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 592         if (res == NULL)
 593                 RETURN(NULL);
 594
 595         lock = ldlm_lock_new(parent_lock, res);
 596         if (lock == NULL) {
 597                 ldlm_resource_put(res);
 598                 RETURN(NULL);
 599         }
 600
 601         lock->l_req_mode = mode;
 602         lock->l_data = data;
 603         lock->l_data_len = data_len;
 604
 605         return lock;
 606 }
 607
 608 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 609 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock * lock,
 610                                void *cookie, int cookie_len,
 611                                int *flags,
 612                                ldlm_completion_callback completion,
 613                                ldlm_blocking_callback blocking)
 614 {
 615         struct ldlm_resource *res;
 616         int local;
 617         ldlm_res_policy policy;
 618         ENTRY;
 619
 620         res = lock->l_resource;
 621         lock->l_blocking_ast = blocking;
 622
 623         if (res->lr_type == LDLM_EXTENT)
 624                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 625
 626         /* policies are not executed on the client */
 627         local = res->lr_namespace->ns_client;
 628         if (!local && (policy = ldlm_res_policy_table[res->lr_type])) {
 629                 int rc;
 630                 rc = policy(lock, cookie, lock->l_req_mode, NULL);
 631
 632                 if (rc == ELDLM_LOCK_CHANGED) {
 633                         res = lock->l_resource;
 634                         *flags |= LDLM_FL_LOCK_CHANGED;
 635                 } else if (rc == ELDLM_LOCK_ABORTED) {
 636                         ldlm_lock_destroy(lock);
 637                         RETURN(rc);
 638                 }
 639         }
 640
 641         lock->l_cookie = cookie;
 642         lock->l_cookie_len = cookie_len;
 643
 644         if (local && lock->l_req_mode == lock->l_granted_mode) {
 645                 /* The server returned a blocked lock, but it was granted before
 646                  * we got a chance to actually enqueue it.  We don't need to do
 647                  * anything else. */
 648                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 649                           LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 650                 GOTO(out, ELDLM_OK);
 651         }
 652
 653         /* This distinction between local lock trees is very important; a client
 654          * namespace only has information about locks taken by that client, and
 655          * thus doesn't have enough information to decide for itself if it can
 656          * be granted (below).  In this case, we do exactly what the server
 657          * tells us to do, as dictated by the 'flags' */
 658         ldlm_resource_unlink_lock(lock);
 659         if (local) {
 660                 if (*flags & LDLM_FL_BLOCK_CONV)
 661                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 662                                                lock);
 663                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 664                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 665                 else
 666                         ldlm_grant_lock(lock);
 667                 GOTO(out, ELDLM_OK);
 668         }
 669
 670         /* FIXME: We may want to optimize by checking lr_most_restr */
 671         if (!list_empty(&res->lr_converting)) {
 672                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 673                 *flags |= LDLM_FL_BLOCK_CONV;
 674                 GOTO(out, ELDLM_OK);
 675         }
 676         if (!list_empty(&res->lr_waiting)) {
 677                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 678                 *flags |= LDLM_FL_BLOCK_WAIT;
 679                 GOTO(out, ELDLM_OK);
 680         }
 681         if (!ldlm_lock_compat(lock, 0)) {
 682                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 683                 *flags |= LDLM_FL_BLOCK_GRANTED;
 684                 GOTO(out, ELDLM_OK);
 685         }
 686
 687         ldlm_grant_lock(lock);
 688         EXIT;
 689       out:
 690         /* Don't set 'completion_ast' until here so that if the lock is granted
 691          * immediately we don't do an unnecessary completion call. */
 692         lock->l_completion_ast = completion;
 693         return ELDLM_OK;
 694 }
 695
 696 /* Must be called with namespace taken: queue is waiting or converting. */
 697 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 698                                 struct list_head *queue)
 699 {
 700         struct list_head *tmp, *pos;
 701         ENTRY;
 702
 703         list_for_each_safe(tmp, pos, queue) {
 704                 struct ldlm_lock *pending;
 705                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 706
 707                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 708
 709                 if (!ldlm_lock_compat(pending, 1))
 710                         RETURN(1);
 711
 712                 list_del_init(&pending->l_res_link);
 713                 ldlm_grant_lock(pending);
 714         }
 715
 716         RETURN(0);
 717 }
 718
 719 void ldlm_run_ast_work(struct list_head *rpc_list)
 720 {
 721         struct list_head *tmp, *pos;
 722         int rc;
 723         ENTRY;
 724
 725         list_for_each_safe(tmp, pos, rpc_list) {
 726                 struct ldlm_ast_work *w =
 727                         list_entry(tmp, struct ldlm_ast_work, w_list);
 728
 729                 if (w->w_blocking)
 730                         rc = w->w_lock->l_blocking_ast
 731                                 (w->w_lock, &w->w_desc, w->w_data,
 732                                  w->w_datalen);
 733                 else
 734                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags);
 735                 if (rc)
 736                         CERROR("Failed AST - should clean & disconnect "
 737                                "client\n");
 738                 LDLM_LOCK_PUT(w->w_lock);
 739                 list_del(&w->w_list);
 740                 OBD_FREE(w, sizeof(*w));
 741         }
 742         EXIT;
 743 }
 744
 745 /* Must be called with resource->lr_lock not taken. */
 746 void ldlm_reprocess_all(struct ldlm_resource *res)
 747 {
 748         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 749         ENTRY;
 750
 751         /* Local lock trees don't get reprocessed. */
 752         if (res->lr_namespace->ns_client) {
 753                 EXIT;
 754                 return;
 755         }
 756
 757         l_lock(&res->lr_namespace->ns_lock);
 758         res->lr_tmp = &rpc_list;
 759
 760         ldlm_reprocess_queue(res, &res->lr_converting);
 761         if (list_empty(&res->lr_converting))
 762                 ldlm_reprocess_queue(res, &res->lr_waiting);
 763
 764         res->lr_tmp = NULL;
 765         l_unlock(&res->lr_namespace->ns_lock);
 766
 767         ldlm_run_ast_work(&rpc_list);
 768         EXIT;
 769 }
 770
 771 void ldlm_lock_cancel(struct ldlm_lock *lock)
 772 {
 773         struct ldlm_resource *res;
 774         struct ldlm_namespace *ns;
 775         ENTRY;
 776
 777         res = lock->l_resource;
 778         ns = res->lr_namespace;
 779
 780         l_lock(&ns->ns_lock);
 781         if (lock->l_readers || lock->l_writers)
 782                 CDEBUG(D_INFO, "lock still has references (%d readers, %d "
 783                        "writers)\n", lock->l_readers, lock->l_writers);
 784
 785         ldlm_resource_unlink_lock(lock);
 786         ldlm_lock_destroy(lock);
 787         l_unlock(&ns->ns_lock);
 788         EXIT;
 789 }
 790
 791 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 792                                         int *flags)
 793 {
 794         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 795         struct ldlm_resource *res;
 796         struct ldlm_namespace *ns;
 797         int granted = 0;
 798         ENTRY;
 799
 800         res = lock->l_resource;
 801         ns = res->lr_namespace;
 802
 803         l_lock(&ns->ns_lock);
 804
 805         lock->l_req_mode = new_mode;
 806         ldlm_resource_unlink_lock(lock);
 807
 808         /* If this is a local resource, put it on the appropriate list. */
 809         if (res->lr_namespace->ns_client) {
 810                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED))
 811                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 812                                                lock);
 813                 else {
 814                         res->lr_tmp = &rpc_list;
 815                         ldlm_grant_lock(lock);
 816                         res->lr_tmp = NULL;
 817                         granted = 1;
 818                         /* FIXME: completion handling not with ns_lock held ! */
 819                         if (lock->l_completion_ast)
 820                                 lock->l_completion_ast(lock, 0);
 821                 }
 822         } else {
 823                 /* FIXME: We should try the conversion right away and possibly
 824                  * return success without the need for an extra AST */
 825                 list_add_tail(&lock->l_res_link, &res->lr_converting);
 826                 *flags |= LDLM_FL_BLOCK_CONV;
 827         }
 828
 829         l_unlock(&ns->ns_lock);
 830
 831         if (granted)
 832                 ldlm_run_ast_work(&rpc_list);
 833         RETURN(res);
 834 }
 835
 836 void ldlm_lock_dump(struct ldlm_lock *lock)
 837 {
 838         char ver[128];
 839
 840         if (!(portal_debug & D_OTHER))
 841                 return;
 842
 843         if (RES_VERSION_SIZE != 4)
 844                 LBUG();
 845
 846         if (!lock) {
 847                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
 848                 return;
 849         }
 850
 851         snprintf(ver, sizeof(ver), "%x %x %x %x",
 852                  lock->l_version[0], lock->l_version[1],
 853                  lock->l_version[2], lock->l_version[3]);
 854
 855         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
 856         if (lock->l_export && lock->l_export->exp_connection)
 857                 CDEBUG(D_OTHER, "  Node: NID %x (rhandle: %Lx)\n",
 858                        lock->l_export->exp_connection->c_peer.peer_nid,
 859                        lock->l_remote_handle.addr);
 860         else
 861                 CDEBUG(D_OTHER, "  Node: local\n");
 862         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
 863         CDEBUG(D_OTHER, "  Resource: %p (%Ld)\n", lock->l_resource,
 864                lock->l_resource->lr_name[0]);
 865         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
 866                (int)lock->l_req_mode, (int)lock->l_granted_mode);
 867         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
 868                lock->l_readers, lock->l_writers);
 869         if (lock->l_resource->lr_type == LDLM_EXTENT)
 870                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
 871                        (unsigned long long)lock->l_extent.start,
 872                        (unsigned long long)lock->l_extent.end);
 873 }