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_inode_link);
 233         INIT_LIST_HEAD(&lock->l_export_chain);
 234         INIT_LIST_HEAD(&lock->l_pending_chain);
 235         init_waitqueue_head(&lock->l_waitq);
 236
 237         if (parent != NULL) {
 238                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 239                 lock->l_parent = parent;
 240                 list_add(&lock->l_childof, &parent->l_children);
 241                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 242         }
 243         /* this is the extra refcount, to prevent the lock
 244            evaporating */
 245         LDLM_LOCK_GET(lock);
 246         RETURN(lock);
 247 }
 248
 249 int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3])
 250 {
 251         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 252         struct ldlm_resource *oldres = lock->l_resource;
 253         int type, i;
 254         ENTRY;
 255
 256         l_lock(&ns->ns_lock);
 257         if (memcmp(new_resid, lock->l_resource->lr_name,
 258                    sizeof(lock->l_resource->lr_name)) == 0) {
 259                 /* Nothing to do */
 260                 l_unlock(&ns->ns_lock);
 261                 RETURN(0);
 262         }
 263
 264         type = lock->l_resource->lr_type;
 265         if (new_resid[0] == 0)
 266                 LBUG();
 267         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid, type, 1);
 268         if (lock->l_resource == NULL) {
 269                 LBUG();
 270                 RETURN(-ENOMEM);
 271         }
 272
 273         /* move references over */
 274         for (i = 0; i < lock->l_refc; i++) {
 275                 int rc;
 276                 ldlm_resource_getref(lock->l_resource);
 277                 rc = ldlm_resource_put(oldres);
 278                 if (rc == 1 && i != lock->l_refc - 1)
 279                         LBUG();
 280         }
 281         /* compensate for the initial get above.. */
 282         ldlm_resource_put(lock->l_resource);
 283
 284         l_unlock(&ns->ns_lock);
 285         RETURN(0);
 286 }
 287
 288 /*
 289  *  HANDLES
 290  */
 291
 292 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 293 {
 294         lockh->addr = (__u64) (unsigned long)lock;
 295         lockh->cookie = lock->l_random;
 296 }
 297
 298 struct ldlm_lock *ldlm_handle2lock(struct lustre_handle *handle)
 299 {
 300         struct ldlm_lock *lock = NULL, *retval = NULL;
 301         ENTRY;
 302
 303         if (!handle || !handle->addr)
 304                 RETURN(NULL);
 305
 306         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
 307         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock))
 308                 RETURN(NULL);
 309
 310         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 311         if (lock->l_random != handle->cookie)
 312                 GOTO(out, NULL);
 313
 314         if (lock->l_flags & LDLM_FL_DESTROYED)
 315                 GOTO(out, NULL);
 316
 317         retval = LDLM_LOCK_GET(lock);
 318         EXIT;
 319       out:
 320         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 321         return retval;
 322 }
 323
 324 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 325 {
 326         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 327 }
 328
 329 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 330 {
 331         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 332         desc->l_req_mode = lock->l_req_mode;
 333         desc->l_granted_mode = lock->l_granted_mode;
 334         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 335         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 336 }
 337
 338 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 339                                    struct ldlm_lock *new)
 340 {
 341         struct ldlm_ast_work *w;
 342         ENTRY;
 343
 344         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 345         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 346                 GOTO(out, 0);
 347
 348         OBD_ALLOC(w, sizeof(*w));
 349         if (!w) {
 350                 LBUG();
 351                 GOTO(out, 0);
 352         }
 353
 354         if (new) {
 355                 lock->l_flags |= LDLM_FL_AST_SENT;
 356                 w->w_blocking = 1;
 357                 ldlm_lock2desc(new, &w->w_desc);
 358         }
 359
 360         w->w_lock = LDLM_LOCK_GET(lock);
 361         list_add(&w->w_list, lock->l_resource->lr_tmp);
 362       out:
 363         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 364         return;
 365 }
 366
 367 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 368 {
 369         struct ldlm_lock *lock;
 370
 371         lock = ldlm_handle2lock(lockh);
 372         ldlm_lock_addref_internal(lock, mode);
 373         LDLM_LOCK_PUT(lock);
 374 }
 375
 376 /* only called for local locks */
 377 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 378 {
 379         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 380         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 381                 lock->l_readers++;
 382         else
 383                 lock->l_writers++;
 384         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 385         LDLM_LOCK_GET(lock);
 386         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 387 }
 388
 389 /* Args: unlocked lock */
 390 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 391 {
 392         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 393         ENTRY;
 394
 395         if (lock == NULL)
 396                 LBUG();
 397
 398         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 399         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 400         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 401                 lock->l_readers--;
 402         else
 403                 lock->l_writers--;
 404
 405         /* If we received a blocked AST and this was the last reference,
 406          * run the callback. */
 407         if (!lock->l_readers && !lock->l_writers &&
 408             (lock->l_flags & LDLM_FL_CBPENDING)) {
 409                 if (!lock->l_resource->lr_namespace->ns_client) {
 410                         CERROR("LDLM_FL_CBPENDING set on non-local lock!\n");
 411                         LBUG();
 412                 }
 413
 414                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 415                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 416
 417                 /* FIXME: need a real 'desc' here */
 418                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 419                                      lock->l_data_len);
 420         } else
 421                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 422
 423         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 424         LDLM_LOCK_PUT(lock);    /* matches the handle2lock above */
 425
 426         EXIT;
 427 }
 428
 429 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 430                                  struct list_head *queue)
 431 {
 432         struct list_head *tmp, *pos;
 433         int rc = 1;
 434
 435         list_for_each_safe(tmp, pos, queue) {
 436                 struct ldlm_lock *child;
 437                 ldlm_res_compat compat;
 438
 439                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 440                 if (lock == child)
 441                         continue;
 442
 443                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 444                 if (compat && compat(child, lock)) {
 445                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 446                         continue;
 447                 }
 448                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 449                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 450                         continue;
 451                 }
 452
 453                 rc = 0;
 454
 455                 if (send_cbs && child->l_blocking_ast != NULL) {
 456                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 457                                "blocking AST.\n", child);
 458                         ldlm_add_ast_work_item(child, lock);
 459                 }
 460         }
 461
 462         return rc;
 463 }
 464
 465 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 466 {
 467         int rc;
 468         ENTRY;
 469
 470         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 471         rc = ldlm_lock_compat_list(lock, send_cbs,
 472                                    &lock->l_resource->lr_granted);
 473         /* FIXME: should we be sending ASTs to converting? */
 474         if (rc)
 475                 rc = ldlm_lock_compat_list
 476                         (lock, send_cbs, &lock->l_resource->lr_converting);
 477
 478         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 479         RETURN(rc);
 480 }
 481
 482 /* NOTE: called by
 483    - ldlm_handle_enqueuque - resource
 484 */
 485 void ldlm_grant_lock(struct ldlm_lock *lock)
 486 {
 487         struct ldlm_resource *res = lock->l_resource;
 488         ENTRY;
 489
 490         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 491         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 492         lock->l_granted_mode = lock->l_req_mode;
 493
 494         if (lock->l_granted_mode < res->lr_most_restr)
 495                 res->lr_most_restr = lock->l_granted_mode;
 496
 497         if (lock->l_completion_ast) {
 498                 ldlm_add_ast_work_item(lock, NULL);
 499         }
 500         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 501         EXIT;
 502 }
 503
 504 /* returns a referenced lock or NULL */
 505 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 506                                       struct ldlm_extent *extent)
 507 {
 508         struct ldlm_lock *lock;
 509         struct list_head *tmp;
 510
 511         list_for_each(tmp, queue) {
 512                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 513
 514                 if (lock->l_flags & LDLM_FL_CBPENDING)
 515                         continue;
 516
 517                 /* lock_convert() takes the resource lock, so we're sure that
 518                  * req_mode, lr_type, and l_cookie won't change beneath us */
 519                 if (lock->l_req_mode != mode)
 520                         continue;
 521
 522                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 523                     (lock->l_extent.start > extent->start ||
 524                      lock->l_extent.end < extent->end))
 525                         continue;
 526
 527                 ldlm_lock_addref_internal(lock, mode);
 528                 return lock;
 529         }
 530
 531         return NULL;
 532 }
 533
 534 /* Must be called with no resource or lock locks held.
 535  *
 536  * Returns 1 if it finds an already-existing lock that is compatible; in this
 537  * case, lockh is filled in with a addref()ed lock
 538 */
 539 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 * res_id, __u32 type,
 540                     void *cookie, int cookielen, ldlm_mode_t mode,
 541                     struct lustre_handle *lockh)
 542 {
 543         struct ldlm_resource *res;
 544         struct ldlm_lock *lock;
 545         int rc = 0;
 546         ENTRY;
 547
 548         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 549         if (res == NULL)
 550                 RETURN(0);
 551
 552         ns = res->lr_namespace;
 553         l_lock(&ns->ns_lock);
 554
 555         if ((lock = search_queue(&res->lr_granted, mode, cookie)))
 556                 GOTO(out, rc = 1);
 557         if ((lock = search_queue(&res->lr_converting, mode, cookie)))
 558                 GOTO(out, rc = 1);
 559         if ((lock = search_queue(&res->lr_waiting, mode, cookie)))
 560                 GOTO(out, rc = 1);
 561
 562         EXIT;
 563       out:
 564         ldlm_resource_put(res);
 565         l_unlock(&ns->ns_lock);
 566
 567         if (lock) {
 568                 ldlm_lock2handle(lock, lockh);
 569                 if (lock->l_completion_ast)
 570                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 571         }
 572         if (rc)
 573                 LDLM_DEBUG(lock, "matched");
 574         else
 575                 LDLM_DEBUG_NOLOCK("not matched");
 576         return rc;
 577 }
 578
 579 /* Returns a referenced lock */
 580 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 581                                    struct lustre_handle *parent_lock_handle,
 582                                    __u64 * res_id, __u32 type,
 583                                    ldlm_mode_t mode, void *data, __u32 data_len)
 584 {
 585         struct ldlm_resource *res, *parent_res = NULL;
 586         struct ldlm_lock *lock, *parent_lock;
 587
 588         parent_lock = ldlm_handle2lock(parent_lock_handle);
 589         if (parent_lock)
 590                 parent_res = parent_lock->l_resource;
 591
 592         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 593         if (res == NULL)
 594                 RETURN(NULL);
 595
 596         lock = ldlm_lock_new(parent_lock, res);
 597         if (lock == NULL) {
 598                 ldlm_resource_put(res);
 599                 RETURN(NULL);
 600         }
 601
 602         lock->l_req_mode = mode;
 603         lock->l_data = data;
 604         lock->l_data_len = data_len;
 605
 606         return lock;
 607 }
 608
 609 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 610 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock * lock,
 611                                void *cookie, int cookie_len,
 612                                int *flags,
 613                                ldlm_completion_callback completion,
 614                                ldlm_blocking_callback blocking)
 615 {
 616         struct ldlm_resource *res;
 617         int local;
 618         ldlm_res_policy policy;
 619         ENTRY;
 620
 621         res = lock->l_resource;
 622         lock->l_blocking_ast = blocking;
 623
 624         if (res->lr_type == LDLM_EXTENT)
 625                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 626
 627         /* policies are not executed on the client */
 628         local = res->lr_namespace->ns_client;
 629         if (!local && (policy = ldlm_res_policy_table[res->lr_type])) {
 630                 int rc;
 631                 rc = policy(lock, cookie, lock->l_req_mode, NULL);
 632
 633                 if (rc == ELDLM_LOCK_CHANGED) {
 634                         res = lock->l_resource;
 635                         *flags |= LDLM_FL_LOCK_CHANGED;
 636                 } else if (rc == ELDLM_LOCK_ABORTED) {
 637                         ldlm_lock_destroy(lock);
 638                         RETURN(rc);
 639                 }
 640         }
 641
 642         lock->l_cookie = cookie;
 643         lock->l_cookie_len = cookie_len;
 644
 645         if (local && lock->l_req_mode == lock->l_granted_mode) {
 646                 /* The server returned a blocked lock, but it was granted before
 647                  * we got a chance to actually enqueue it.  We don't need to do
 648                  * anything else. */
 649                 GOTO(out, ELDLM_OK);
 650         }
 651
 652         /* This distinction between local lock trees is very important; a client
 653          * namespace only has information about locks taken by that client, and
 654          * thus doesn't have enough information to decide for itself if it can
 655          * be granted (below).  In this case, we do exactly what the server
 656          * tells us to do, as dictated by the 'flags' */
 657         ldlm_resource_unlink_lock(lock);
 658         if (local) {
 659                 if (*flags & LDLM_FL_BLOCK_CONV)
 660                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 661                                                lock);
 662                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 663                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 664                 else
 665                         ldlm_grant_lock(lock);
 666                 GOTO(out, ELDLM_OK);
 667         }
 668
 669         /* FIXME: We may want to optimize by checking lr_most_restr */
 670         if (!list_empty(&res->lr_converting)) {
 671                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 672                 *flags |= LDLM_FL_BLOCK_CONV;
 673                 GOTO(out, ELDLM_OK);
 674         }
 675         if (!list_empty(&res->lr_waiting)) {
 676                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 677                 *flags |= LDLM_FL_BLOCK_WAIT;
 678                 GOTO(out, ELDLM_OK);
 679         }
 680         if (!ldlm_lock_compat(lock, 0)) {
 681                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 682                 *flags |= LDLM_FL_BLOCK_GRANTED;
 683                 GOTO(out, ELDLM_OK);
 684         }
 685
 686         ldlm_grant_lock(lock);
 687         EXIT;
 688       out:
 689         /* Don't set 'completion_ast' until here so that if the lock is granted
 690          * immediately we don't do an unnecessary completion call. */
 691         lock->l_completion_ast = completion;
 692         return ELDLM_OK;
 693 }
 694
 695 /* Must be called with namespace taken: queue is waiting or converting. */
 696 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 697                                 struct list_head *queue)
 698 {
 699         struct list_head *tmp, *pos;
 700         ENTRY;
 701
 702         list_for_each_safe(tmp, pos, queue) {
 703                 struct ldlm_lock *pending;
 704                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 705
 706                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 707
 708                 if (!ldlm_lock_compat(pending, 1))
 709                         RETURN(1);
 710
 711                 list_del_init(&pending->l_res_link);
 712                 ldlm_grant_lock(pending);
 713         }
 714
 715         RETURN(0);
 716 }
 717
 718 void ldlm_run_ast_work(struct list_head *rpc_list)
 719 {
 720         struct list_head *tmp, *pos;
 721         int rc;
 722         ENTRY;
 723
 724         list_for_each_safe(tmp, pos, rpc_list) {
 725                 struct ldlm_ast_work *w =
 726                         list_entry(tmp, struct ldlm_ast_work, w_list);
 727
 728                 if (w->w_blocking)
 729                         rc = w->w_lock->l_blocking_ast
 730                                 (w->w_lock, &w->w_desc, w->w_data,
 731                                  w->w_datalen);
 732                 else
 733                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags);
 734                 if (rc)
 735                         CERROR("Failed AST - should clean & disconnect "
 736                                "client\n");
 737                 LDLM_LOCK_PUT(w->w_lock);
 738                 list_del(&w->w_list);
 739                 OBD_FREE(w, sizeof(*w));
 740         }
 741         EXIT;
 742 }
 743
 744 /* Must be called with resource->lr_lock not taken. */
 745 void ldlm_reprocess_all(struct ldlm_resource *res)
 746 {
 747         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 748         ENTRY;
 749
 750         /* Local lock trees don't get reprocessed. */
 751         if (res->lr_namespace->ns_client) {
 752                 EXIT;
 753                 return;
 754         }
 755
 756         l_lock(&res->lr_namespace->ns_lock);
 757         res->lr_tmp = &rpc_list;
 758
 759         ldlm_reprocess_queue(res, &res->lr_converting);
 760         if (list_empty(&res->lr_converting))
 761                 ldlm_reprocess_queue(res, &res->lr_waiting);
 762
 763         res->lr_tmp = NULL;
 764         l_unlock(&res->lr_namespace->ns_lock);
 765
 766         ldlm_run_ast_work(&rpc_list);
 767         EXIT;
 768 }
 769
 770 void ldlm_lock_cancel(struct ldlm_lock *lock)
 771 {
 772         struct ldlm_resource *res;
 773         struct ldlm_namespace *ns;
 774         ENTRY;
 775
 776         res = lock->l_resource;
 777         ns = res->lr_namespace;
 778
 779         l_lock(&ns->ns_lock);
 780         if (lock->l_readers || lock->l_writers)
 781                 CDEBUG(D_INFO, "lock still has references (%d readers, %d "
 782                        "writers)\n", lock->l_readers, lock->l_writers);
 783
 784         ldlm_resource_unlink_lock(lock);
 785         ldlm_lock_destroy(lock);
 786         l_unlock(&ns->ns_lock);
 787         EXIT;
 788 }
 789
 790 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 791                                         int *flags)
 792 {
 793         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 794         struct ldlm_resource *res;
 795         struct ldlm_namespace *ns;
 796         int granted = 0;
 797         ENTRY;
 798
 799         res = lock->l_resource;
 800         ns = res->lr_namespace;
 801
 802         l_lock(&ns->ns_lock);
 803
 804         lock->l_req_mode = new_mode;
 805         ldlm_resource_unlink_lock(lock);
 806
 807         /* If this is a local resource, put it on the appropriate list. */
 808         if (res->lr_namespace->ns_client) {
 809                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED))
 810                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 811                                                lock);
 812                 else {
 813                         res->lr_tmp = &rpc_list;
 814                         ldlm_grant_lock(lock);
 815                         res->lr_tmp = NULL;
 816                         granted = 1;
 817                         /* FIXME: completion handling not with ns_lock held ! */
 818                         if (lock->l_completion_ast)
 819                                 lock->l_completion_ast(lock, 0);
 820                 }
 821         } else
 822                 list_add_tail(&lock->l_res_link, &res->lr_converting);
 823
 824         l_unlock(&ns->ns_lock);
 825
 826         if (granted)
 827                 ldlm_run_ast_work(&rpc_list);
 828         RETURN(res);
 829 }
 830
 831 void ldlm_lock_dump(struct ldlm_lock *lock)
 832 {
 833         char ver[128];
 834
 835         if (!(portal_debug & D_OTHER))
 836                 return;
 837
 838         if (RES_VERSION_SIZE != 4)
 839                 LBUG();
 840
 841         if (!lock) {
 842                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
 843                 return;
 844         }
 845
 846         snprintf(ver, sizeof(ver), "%x %x %x %x",
 847                  lock->l_version[0], lock->l_version[1],
 848                  lock->l_version[2], lock->l_version[3]);
 849
 850         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
 851         if (lock->l_export && lock->l_export->exp_connection)
 852                 CDEBUG(D_OTHER, "  Node: NID %x (rhandle: %Lx)\n",
 853                        lock->l_export->exp_connection->c_peer.peer_nid,
 854                        lock->l_remote_handle.addr);
 855         else
 856                 CDEBUG(D_OTHER, "  Node: local\n");
 857         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
 858         CDEBUG(D_OTHER, "  Resource: %p (%Ld)\n", lock->l_resource,
 859                lock->l_resource->lr_name[0]);
 860         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
 861                (int)lock->l_req_mode, (int)lock->l_granted_mode);
 862         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
 863                lock->l_readers, lock->l_writers);
 864         if (lock->l_resource->lr_type == LDLM_EXTENT)
 865                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
 866                        (unsigned long long)lock->l_extent.start,
 867                        (unsigned long long)lock->l_extent.end);
 868 }