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 ldlm_namespace *ns = lock->l_resource->lr_namespace;
 140         ENTRY;
 141
 142         l_lock(&ns->ns_lock);
 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                 spin_lock(&ns->ns_counter_lock);
 154                 ns->ns_locks--;
 155                 spin_unlock(&ns->ns_counter_lock);
 156
 157                 lock->l_resource = NULL;
 158                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
 159                 if (lock->l_export && lock->l_export->exp_connection)
 160                         ptlrpc_put_connection(lock->l_export->exp_connection);
 161                 kmem_cache_free(ldlm_lock_slab, lock);
 162                 CDEBUG(D_MALLOC, "kfreed 'lock': %d at %p (tot 0).\n",
 163                        sizeof(*lock), lock);
 164         }
 165         l_unlock(&ns->ns_lock);
 166         EXIT;
 167 }
 168
 169 void ldlm_lock_destroy(struct ldlm_lock *lock)
 170 {
 171         ENTRY;
 172         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 173
 174         if (!list_empty(&lock->l_children)) {
 175                 LDLM_DEBUG(lock, "still has children (%p)!",
 176                            lock->l_children.next);
 177                 ldlm_lock_dump(lock);
 178                 LBUG();
 179         }
 180         if (lock->l_readers || lock->l_writers) {
 181                 LDLM_DEBUG(lock, "lock still has references");
 182                 ldlm_lock_dump(lock);
 183                 LBUG();
 184         }
 185
 186         if (!list_empty(&lock->l_res_link)) {
 187                 ldlm_lock_dump(lock);
 188                 LBUG();
 189         }
 190
 191         if (lock->l_flags & LDLM_FL_DESTROYED) {
 192                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 193                 EXIT;
 194                 return;
 195         }
 196
 197         list_del(&lock->l_export_chain);
 198         lock->l_export = NULL;
 199         lock->l_flags |= LDLM_FL_DESTROYED;
 200
 201         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 202         LDLM_LOCK_PUT(lock);
 203         EXIT;
 204 }
 205
 206 /*
 207    usage: pass in a resource on which you have done get
 208           pass in a parent lock on which you have done a get
 209           do not put the resource or the parent
 210    returns: lock with refcount 1
 211 */
 212 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 213                                        struct ldlm_resource *resource)
 214 {
 215         struct ldlm_lock *lock;
 216         ENTRY;
 217
 218         if (resource == NULL)
 219                 LBUG();
 220
 221         lock = kmem_cache_alloc(ldlm_lock_slab, SLAB_KERNEL);
 222         if (lock == NULL)
 223                 RETURN(NULL);
 224         CDEBUG(D_MALLOC, "kmalloced 'lock': %d at "
 225                "%p (tot %d).\n", sizeof(*lock), lock, 1);
 226
 227         memset(lock, 0, sizeof(*lock));
 228         get_random_bytes(&lock->l_random, sizeof(__u64));
 229
 230         lock->l_resource = resource;
 231         /* this refcount matches the one of the resource passed
 232            in which is not being put away */
 233         lock->l_refc = 1;
 234         INIT_LIST_HEAD(&lock->l_children);
 235         INIT_LIST_HEAD(&lock->l_res_link);
 236         INIT_LIST_HEAD(&lock->l_export_chain);
 237         INIT_LIST_HEAD(&lock->l_pending_chain);
 238         init_waitqueue_head(&lock->l_waitq);
 239
 240         spin_lock(&resource->lr_namespace->ns_counter_lock);
 241         resource->lr_namespace->ns_locks++;
 242         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 243
 244         if (parent != NULL) {
 245                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 246                 lock->l_parent = parent;
 247                 list_add(&lock->l_childof, &parent->l_children);
 248                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 249         }
 250         /* this is the extra refcount, to prevent the lock
 251            evaporating */
 252         LDLM_LOCK_GET(lock);
 253         RETURN(lock);
 254 }
 255
 256 int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3])
 257 {
 258         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 259         struct ldlm_resource *oldres = lock->l_resource;
 260         int type, i;
 261         ENTRY;
 262
 263         l_lock(&ns->ns_lock);
 264         if (memcmp(new_resid, lock->l_resource->lr_name,
 265                    sizeof(lock->l_resource->lr_name)) == 0) {
 266                 /* Nothing to do */
 267                 l_unlock(&ns->ns_lock);
 268                 RETURN(0);
 269         }
 270
 271         type = lock->l_resource->lr_type;
 272         if (new_resid[0] == 0)
 273                 LBUG();
 274         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid, type, 1);
 275         if (lock->l_resource == NULL) {
 276                 LBUG();
 277                 RETURN(-ENOMEM);
 278         }
 279
 280         /* move references over */
 281         for (i = 0; i < lock->l_refc; i++) {
 282                 int rc;
 283                 ldlm_resource_getref(lock->l_resource);
 284                 rc = ldlm_resource_put(oldres);
 285                 if (rc == 1 && i != lock->l_refc - 1)
 286                         LBUG();
 287         }
 288         /* compensate for the initial get above.. */
 289         ldlm_resource_put(lock->l_resource);
 290
 291         l_unlock(&ns->ns_lock);
 292         RETURN(0);
 293 }
 294
 295 /*
 296  *  HANDLES
 297  */
 298
 299 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 300 {
 301         lockh->addr = (__u64) (unsigned long)lock;
 302         lockh->cookie = lock->l_random;
 303 }
 304
 305 struct ldlm_lock *ldlm_handle2lock(struct lustre_handle *handle)
 306 {
 307         struct ldlm_lock *lock = NULL, *retval = NULL;
 308         ENTRY;
 309
 310         if (!handle || !handle->addr)
 311                 RETURN(NULL);
 312
 313         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
 314         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock))
 315                 RETURN(NULL);
 316
 317         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 318         if (lock->l_random != handle->cookie)
 319                 GOTO(out, NULL);
 320
 321         if (lock->l_flags & LDLM_FL_DESTROYED)
 322                 GOTO(out, NULL);
 323
 324         retval = LDLM_LOCK_GET(lock);
 325         EXIT;
 326       out:
 327         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 328         return retval;
 329 }
 330
 331 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 332 {
 333         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 334 }
 335
 336 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 337 {
 338         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 339         desc->l_req_mode = lock->l_req_mode;
 340         desc->l_granted_mode = lock->l_granted_mode;
 341         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 342         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 343 }
 344
 345 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 346                                    struct ldlm_lock *new)
 347 {
 348         struct ldlm_ast_work *w;
 349         ENTRY;
 350
 351         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 352         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 353                 GOTO(out, 0);
 354
 355         OBD_ALLOC(w, sizeof(*w));
 356         if (!w) {
 357                 LBUG();
 358                 GOTO(out, 0);
 359         }
 360
 361         if (new) {
 362                 lock->l_flags |= LDLM_FL_AST_SENT;
 363                 w->w_blocking = 1;
 364                 ldlm_lock2desc(new, &w->w_desc);
 365         }
 366
 367         w->w_lock = LDLM_LOCK_GET(lock);
 368         list_add(&w->w_list, lock->l_resource->lr_tmp);
 369       out:
 370         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 371         return;
 372 }
 373
 374 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 375 {
 376         struct ldlm_lock *lock;
 377
 378         lock = ldlm_handle2lock(lockh);
 379         ldlm_lock_addref_internal(lock, mode);
 380         LDLM_LOCK_PUT(lock);
 381 }
 382
 383 /* only called for local locks */
 384 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 385 {
 386         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 387         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 388                 lock->l_readers++;
 389         else
 390                 lock->l_writers++;
 391         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 392         LDLM_LOCK_GET(lock);
 393         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 394 }
 395
 396 /* Args: unlocked lock */
 397 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 398 {
 399         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 400         ENTRY;
 401
 402         if (lock == NULL)
 403                 LBUG();
 404
 405         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 406         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 407         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 408                 lock->l_readers--;
 409         else
 410                 lock->l_writers--;
 411
 412         /* If we received a blocked AST and this was the last reference,
 413          * run the callback. */
 414         if (!lock->l_readers && !lock->l_writers &&
 415             (lock->l_flags & LDLM_FL_CBPENDING)) {
 416                 if (!lock->l_resource->lr_namespace->ns_client) {
 417                         CERROR("LDLM_FL_CBPENDING set on non-local lock!\n");
 418                         LBUG();
 419                 }
 420
 421                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 422                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 423
 424                 /* FIXME: need a real 'desc' here */
 425                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 426                                      lock->l_data_len);
 427         } else
 428                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 429
 430         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 431         LDLM_LOCK_PUT(lock);    /* matches the handle2lock above */
 432
 433         EXIT;
 434 }
 435
 436 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 437                                  struct list_head *queue)
 438 {
 439         struct list_head *tmp, *pos;
 440         int rc = 1;
 441
 442         list_for_each_safe(tmp, pos, queue) {
 443                 struct ldlm_lock *child;
 444                 ldlm_res_compat compat;
 445
 446                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 447                 if (lock == child)
 448                         continue;
 449
 450                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 451                 if (compat && compat(child, lock)) {
 452                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 453                         continue;
 454                 }
 455                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 456                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 457                         continue;
 458                 }
 459
 460                 rc = 0;
 461
 462                 if (send_cbs && child->l_blocking_ast != NULL) {
 463                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 464                                "blocking AST.\n", child);
 465                         ldlm_add_ast_work_item(child, lock);
 466                 }
 467         }
 468
 469         return rc;
 470 }
 471
 472 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 473 {
 474         int rc;
 475         ENTRY;
 476
 477         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 478         rc = ldlm_lock_compat_list(lock, send_cbs,
 479                                    &lock->l_resource->lr_granted);
 480         /* FIXME: should we be sending ASTs to converting? */
 481         if (rc)
 482                 rc = ldlm_lock_compat_list
 483                         (lock, send_cbs, &lock->l_resource->lr_converting);
 484
 485         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 486         RETURN(rc);
 487 }
 488
 489 /* NOTE: called by
 490    - ldlm_handle_enqueuque - resource
 491 */
 492 void ldlm_grant_lock(struct ldlm_lock *lock)
 493 {
 494         struct ldlm_resource *res = lock->l_resource;
 495         ENTRY;
 496
 497         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 498         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 499         lock->l_granted_mode = lock->l_req_mode;
 500
 501         if (lock->l_granted_mode < res->lr_most_restr)
 502                 res->lr_most_restr = lock->l_granted_mode;
 503
 504         if (lock->l_completion_ast) {
 505                 ldlm_add_ast_work_item(lock, NULL);
 506         }
 507         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 508         EXIT;
 509 }
 510
 511 /* returns a referenced lock or NULL */
 512 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 513                                       struct ldlm_extent *extent)
 514 {
 515         struct ldlm_lock *lock;
 516         struct list_head *tmp;
 517
 518         list_for_each(tmp, queue) {
 519                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 520
 521                 if (lock->l_flags & (LDLM_FL_CBPENDING | LDLM_FL_DESTROYED))
 522                         continue;
 523
 524                 /* lock_convert() takes the resource lock, so we're sure that
 525                  * req_mode, lr_type, and l_cookie won't change beneath us */
 526                 if (lock->l_req_mode != mode)
 527                         continue;
 528
 529                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 530                     (lock->l_extent.start > extent->start ||
 531                      lock->l_extent.end < extent->end))
 532                         continue;
 533
 534                 ldlm_lock_addref_internal(lock, mode);
 535                 return lock;
 536         }
 537
 538         return NULL;
 539 }
 540
 541 /* Must be called with no resource or lock locks held.
 542  *
 543  * Returns 1 if it finds an already-existing lock that is compatible; in this
 544  * case, lockh is filled in with a addref()ed lock
 545 */
 546 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 * res_id, __u32 type,
 547                     void *cookie, int cookielen, ldlm_mode_t mode,
 548                     struct lustre_handle *lockh)
 549 {
 550         struct ldlm_resource *res;
 551         struct ldlm_lock *lock;
 552         int rc = 0;
 553         ENTRY;
 554
 555         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 556         if (res == NULL)
 557                 RETURN(0);
 558
 559         ns = res->lr_namespace;
 560         l_lock(&ns->ns_lock);
 561
 562         if ((lock = search_queue(&res->lr_granted, mode, cookie)))
 563                 GOTO(out, rc = 1);
 564         if ((lock = search_queue(&res->lr_converting, mode, cookie)))
 565                 GOTO(out, rc = 1);
 566         if ((lock = search_queue(&res->lr_waiting, mode, cookie)))
 567                 GOTO(out, rc = 1);
 568
 569         EXIT;
 570       out:
 571         ldlm_resource_put(res);
 572         l_unlock(&ns->ns_lock);
 573
 574         if (lock) {
 575                 ldlm_lock2handle(lock, lockh);
 576                 if (lock->l_completion_ast)
 577                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 578         }
 579         if (rc)
 580                 LDLM_DEBUG(lock, "matched");
 581         else
 582                 LDLM_DEBUG_NOLOCK("not matched");
 583         return rc;
 584 }
 585
 586 /* Returns a referenced lock */
 587 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 588                                    struct lustre_handle *parent_lock_handle,
 589                                    __u64 * res_id, __u32 type,
 590                                    ldlm_mode_t mode, void *data, __u32 data_len)
 591 {
 592         struct ldlm_resource *res, *parent_res = NULL;
 593         struct ldlm_lock *lock, *parent_lock;
 594
 595         parent_lock = ldlm_handle2lock(parent_lock_handle);
 596         if (parent_lock)
 597                 parent_res = parent_lock->l_resource;
 598
 599         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 600         if (res == NULL)
 601                 RETURN(NULL);
 602
 603         lock = ldlm_lock_new(parent_lock, res);
 604         if (lock == NULL) {
 605                 ldlm_resource_put(res);
 606                 RETURN(NULL);
 607         }
 608
 609         lock->l_req_mode = mode;
 610         lock->l_data = data;
 611         lock->l_data_len = data_len;
 612
 613         return lock;
 614 }
 615
 616 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 617 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock * lock,
 618                                void *cookie, int cookie_len,
 619                                int *flags,
 620                                ldlm_completion_callback completion,
 621                                ldlm_blocking_callback blocking)
 622 {
 623         struct ldlm_resource *res;
 624         int local;
 625         ldlm_res_policy policy;
 626         ENTRY;
 627
 628         res = lock->l_resource;
 629         lock->l_blocking_ast = blocking;
 630
 631         if (res->lr_type == LDLM_EXTENT)
 632                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 633
 634         /* policies are not executed on the client */
 635         local = res->lr_namespace->ns_client;
 636         if (!local && (policy = ldlm_res_policy_table[res->lr_type])) {
 637                 int rc;
 638                 rc = policy(lock, cookie, lock->l_req_mode, NULL);
 639
 640                 if (rc == ELDLM_LOCK_CHANGED) {
 641                         res = lock->l_resource;
 642                         *flags |= LDLM_FL_LOCK_CHANGED;
 643                 } else if (rc == ELDLM_LOCK_ABORTED) {
 644                         ldlm_lock_destroy(lock);
 645                         RETURN(rc);
 646                 }
 647         }
 648
 649         lock->l_cookie = cookie;
 650         lock->l_cookie_len = cookie_len;
 651
 652         l_lock(&res->lr_namespace->ns_lock);
 653         if (local && lock->l_req_mode == lock->l_granted_mode) {
 654                 /* The server returned a blocked lock, but it was granted before
 655                  * we got a chance to actually enqueue it.  We don't need to do
 656                  * anything else. */
 657                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 658                           LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 659                 GOTO(out, ELDLM_OK);
 660         }
 661
 662         /* This distinction between local lock trees is very important; a client
 663          * namespace only has information about locks taken by that client, and
 664          * thus doesn't have enough information to decide for itself if it can
 665          * be granted (below).  In this case, we do exactly what the server
 666          * tells us to do, as dictated by the 'flags' */
 667         ldlm_resource_unlink_lock(lock);
 668         if (local) {
 669                 if (*flags & LDLM_FL_BLOCK_CONV)
 670                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 671                                                lock);
 672                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 673                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 674                 else
 675                         ldlm_grant_lock(lock);
 676                 GOTO(out, ELDLM_OK);
 677         }
 678
 679         /* FIXME: We may want to optimize by checking lr_most_restr */
 680         if (!list_empty(&res->lr_converting)) {
 681                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 682                 *flags |= LDLM_FL_BLOCK_CONV;
 683                 GOTO(out, ELDLM_OK);
 684         }
 685         if (!list_empty(&res->lr_waiting)) {
 686                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 687                 *flags |= LDLM_FL_BLOCK_WAIT;
 688                 GOTO(out, ELDLM_OK);
 689         }
 690         if (!ldlm_lock_compat(lock, 0)) {
 691                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 692                 *flags |= LDLM_FL_BLOCK_GRANTED;
 693                 GOTO(out, ELDLM_OK);
 694         }
 695
 696         ldlm_grant_lock(lock);
 697         EXIT;
 698       out:
 699         l_unlock(&res->lr_namespace->ns_lock);
 700         /* Don't set 'completion_ast' until here so that if the lock is granted
 701          * immediately we don't do an unnecessary completion call. */
 702         lock->l_completion_ast = completion;
 703         return ELDLM_OK;
 704 }
 705
 706 /* Must be called with namespace taken: queue is waiting or converting. */
 707 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 708                                 struct list_head *queue)
 709 {
 710         struct list_head *tmp, *pos;
 711         ENTRY;
 712
 713         list_for_each_safe(tmp, pos, queue) {
 714                 struct ldlm_lock *pending;
 715                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 716
 717                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 718
 719                 if (!ldlm_lock_compat(pending, 1))
 720                         RETURN(1);
 721
 722                 list_del_init(&pending->l_res_link);
 723                 ldlm_grant_lock(pending);
 724         }
 725
 726         RETURN(0);
 727 }
 728
 729 void ldlm_run_ast_work(struct list_head *rpc_list)
 730 {
 731         struct list_head *tmp, *pos;
 732         int rc;
 733         ENTRY;
 734
 735         list_for_each_safe(tmp, pos, rpc_list) {
 736                 struct ldlm_ast_work *w =
 737                         list_entry(tmp, struct ldlm_ast_work, w_list);
 738
 739                 if (w->w_blocking)
 740                         rc = w->w_lock->l_blocking_ast
 741                                 (w->w_lock, &w->w_desc, w->w_data,
 742                                  w->w_datalen);
 743                 else
 744                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags);
 745                 if (rc)
 746                         CERROR("Failed AST - should clean & disconnect "
 747                                "client\n");
 748                 LDLM_LOCK_PUT(w->w_lock);
 749                 list_del(&w->w_list);
 750                 OBD_FREE(w, sizeof(*w));
 751         }
 752         EXIT;
 753 }
 754
 755 /* Must be called with resource->lr_lock not taken. */
 756 void ldlm_reprocess_all(struct ldlm_resource *res)
 757 {
 758         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 759         ENTRY;
 760
 761         /* Local lock trees don't get reprocessed. */
 762         if (res->lr_namespace->ns_client) {
 763                 EXIT;
 764                 return;
 765         }
 766
 767         l_lock(&res->lr_namespace->ns_lock);
 768         res->lr_tmp = &rpc_list;
 769
 770         ldlm_reprocess_queue(res, &res->lr_converting);
 771         if (list_empty(&res->lr_converting))
 772                 ldlm_reprocess_queue(res, &res->lr_waiting);
 773
 774         res->lr_tmp = NULL;
 775         l_unlock(&res->lr_namespace->ns_lock);
 776
 777         ldlm_run_ast_work(&rpc_list);
 778         EXIT;
 779 }
 780
 781 void ldlm_lock_cancel(struct ldlm_lock *lock)
 782 {
 783         struct ldlm_resource *res;
 784         struct ldlm_namespace *ns;
 785         ENTRY;
 786
 787         res = lock->l_resource;
 788         ns = res->lr_namespace;
 789
 790         l_lock(&ns->ns_lock);
 791         if (lock->l_readers || lock->l_writers)
 792                 CDEBUG(D_INFO, "lock still has references (%d readers, %d "
 793                        "writers)\n", lock->l_readers, lock->l_writers);
 794
 795         ldlm_resource_unlink_lock(lock);
 796         ldlm_lock_destroy(lock);
 797         l_unlock(&ns->ns_lock);
 798         EXIT;
 799 }
 800
 801 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 802                                         int *flags)
 803 {
 804         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 805         struct ldlm_resource *res;
 806         struct ldlm_namespace *ns;
 807         int granted = 0;
 808         ENTRY;
 809
 810         res = lock->l_resource;
 811         ns = res->lr_namespace;
 812
 813         l_lock(&ns->ns_lock);
 814
 815         lock->l_req_mode = new_mode;
 816         ldlm_resource_unlink_lock(lock);
 817
 818         /* If this is a local resource, put it on the appropriate list. */
 819         if (res->lr_namespace->ns_client) {
 820                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED))
 821                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 822                                                lock);
 823                 else {
 824                         res->lr_tmp = &rpc_list;
 825                         ldlm_grant_lock(lock);
 826                         res->lr_tmp = NULL;
 827                         granted = 1;
 828                         /* FIXME: completion handling not with ns_lock held ! */
 829                         if (lock->l_completion_ast)
 830                                 lock->l_completion_ast(lock, 0);
 831                 }
 832         } else {
 833                 /* FIXME: We should try the conversion right away and possibly
 834                  * return success without the need for an extra AST */
 835                 ldlm_resource_add_lock(res, res->lr_converting.prev, lock);
 836                 *flags |= LDLM_FL_BLOCK_CONV;
 837         }
 838
 839         l_unlock(&ns->ns_lock);
 840
 841         if (granted)
 842                 ldlm_run_ast_work(&rpc_list);
 843         RETURN(res);
 844 }
 845
 846 void ldlm_lock_dump(struct ldlm_lock *lock)
 847 {
 848         char ver[128];
 849
 850         if (!(portal_debug & D_OTHER))
 851                 return;
 852
 853         if (RES_VERSION_SIZE != 4)
 854                 LBUG();
 855
 856         if (!lock) {
 857                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
 858                 return;
 859         }
 860
 861         snprintf(ver, sizeof(ver), "%x %x %x %x",
 862                  lock->l_version[0], lock->l_version[1],
 863                  lock->l_version[2], lock->l_version[3]);
 864
 865         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
 866         if (lock->l_export && lock->l_export->exp_connection)
 867                 CDEBUG(D_OTHER, "  Node: NID %x (rhandle: "LPX64")\n",
 868                        lock->l_export->exp_connection->c_peer.peer_nid,
 869                        lock->l_remote_handle.addr);
 870         else
 871                 CDEBUG(D_OTHER, "  Node: local\n");
 872         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
 873         CDEBUG(D_OTHER, "  Resource: %p ("LPD64")\n", lock->l_resource,
 874                lock->l_resource->lr_name[0]);
 875         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
 876                (int)lock->l_req_mode, (int)lock->l_granted_mode);
 877         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
 878                lock->l_readers, lock->l_writers);
 879         if (lock->l_resource->lr_type == LDLM_EXTENT)
 880                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
 881                        (unsigned long long)lock->l_extent.start,
 882                        (unsigned long long)lock->l_extent.end);
 883 }