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  *
   6  * This code is issued under the GNU General Public License.
   7  * See the file COPYING in this distribution
   8  *
   9  * by Cluster File Systems, Inc.
  10  * authors, Peter Braam <braam@clusterfs.com> &
  11  * Phil Schwan <phil@clusterfs.com>
  12  */
  13
  14 #define DEBUG_SUBSYSTEM S_LDLM
  15
  16 #include <linux/slab.h>
  17 #include <linux/module.h>
  18 #include <linux/lustre_dlm.h>
  19 #include <linux/lustre_mds.h>
  20
  21 static kmem_cache_t *ldlm_lock_slab;
  22 int (*mds_reint_p)(int offset, struct ptlrpc_request *req) = NULL;
  23 int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req) = NULL;
  24
  25 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b);
  26 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
  27                               ldlm_mode_t mode, void *data);
  28
  29 ldlm_res_compat ldlm_res_compat_table [] = {
  30         [LDLM_PLAIN] ldlm_plain_compat,
  31         [LDLM_EXTENT] ldlm_extent_compat,
  32         [LDLM_MDSINTENT] ldlm_plain_compat
  33 };
  34
  35 ldlm_res_policy ldlm_res_policy_table [] = {
  36         [LDLM_PLAIN] NULL,
  37         [LDLM_EXTENT] ldlm_extent_policy,
  38         [LDLM_MDSINTENT] ldlm_intent_policy
  39 };
  40
  41 void ldlm_lock2handle(struct ldlm_lock *lock, struct ldlm_handle *lockh)
  42 {
  43         handle->addr = (__u64)(unsigned long)lock;
  44         handle->cookie = lock->l_cookie;
  45 }
  46
  47 struct *ldlm_handle2lock(struct ldlm_handle *handle)
  48 {
  49         struct ldlm_lock *lock = NULL;
  50         ENTRY;
  51
  52         if (!handle)
  53                 RETURN(NULL);
  54         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
  55
  56         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock))
  57                 RETURN(NULL);
  58
  59         l_lock(&lock->l_resource->lr_namespace->ns_lock);
  60         if (lock->l_cookie != handle->cookie)
  61                 GOTO(out, handle = NULL);
  62
  63         if (lock->l_flags & LDLM_FL_DESTROYED)
  64                 GOTO(out, handle = NULL);
  65
  66         lock->l_refc++;
  67         EXIT;
  68  out:
  69         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
  70         return  handle;
  71 }
  72
  73 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
  74 {
  75         l_lock(&lock->l_resource->lr_namespace->ns_lock);
  76         lock->l_refc++;
  77         ldlm_resource_getref(lock->l_resource);
  78         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
  79         return lock;
  80 }
  81
  82 void ldlm_lock_put(struct ldlm_lock *lock)
  83 {
  84         struct l_lock *nslock = &lock->l_resource->lr_namespace->ns_lock;
  85         ENTRY;
  86
  87         l_lock(&nslock);
  88         lock->l_refc--;
  89         if (lock->l_refc < 0)
  90                 LBUG();
  91
  92         ldlm_resource_put(lock->l_resource);
  93         if (lock->l_parent)
  94                 ldlm_lock_put(lock->l_parent);
  95
  96         if (lock->l_refc == 0 && (lock->l_flags & LDLM_FL_DESTROYED)) {
  97                 if (lock->l_connection)
  98                         ptlrpc_put_connection(lock->l_connection);
  99                 kmem_cache_free(ldlm_lock_slab, lock);
 100         }
 101         l_unlock(&nslock);
 102         EXIT;
 103         return;
 104 }
 105
 106 void ldlm_lock_destroy(struct ldlm_lock *lock)
 107 {
 108         ENTRY;
 109         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 110
 111         if (!list_empty(&lock->l_children)) {
 112                 CERROR("lock %p still has children (%p)!\n", lock,
 113                        lock->l_children.next);
 114                 ldlm_lock_dump(lock);
 115                 LBUG();
 116         }
 117         if (lock->l_readers || lock->l_writers) {
 118                 CDEBUG(D_INFO, "lock still has references (%d readers, %d "
 119                        "writers)\n", lock->l_readers, lock->l_writers);
 120                 LBUG();
 121         }
 122
 123         if (!list_empty(lock->l_res))
 124                 LBUG();
 125
 126         lock->l_flags = LDLM_FL_DESTROYED;
 127         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 128         ldlm_lock_put(lock);
 129         EXIT;
 130         return;
 131 }
 132 /*
 133    usage: pass in a resource on which you have done get
 134           pass in a parent lock on which you have done a get
 135           do not put the resource or the parent
 136    returns: lock with refcount 1
 137 */
 138 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 139                                        struct ldlm_resource *resource)
 140 {
 141         struct ldlm_lock *lock;
 142         ENTRY;
 143
 144         if (resource == NULL)
 145                 LBUG();
 146
 147         lock = kmem_cache_alloc(ldlm_lock_slab, SLAB_KERNEL);
 148         if (lock == NULL)
 149                 RETURN(NULL);
 150
 151         memset(lock, 0, sizeof(*lock));
 152         get_random_bytes(&lock->l_cookie, sizeof(__u64));
 153
 154         lock->l_resource = resource;
 155         lock->l_refc = 1;
 156         INIT_LIST_HEAD(&lock->l_children);
 157         INIT_LIST_HEAD(&lock->l_res_link);
 158         init_waitqueue_head(&lock->l_waitq);
 159
 160         if (parent != NULL) {
 161                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 162                 lock->l_parent = parent;
 163                 list_add(&lock->l_childof, &parent->l_children);
 164                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 165         }
 166         RETURN(lock);
 167 }
 168
 169 int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3])
 170 {
 171         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 172         int type, i;
 173         ENTRY;
 174
 175         l_lock(&ns->ns_lock);
 176         type = lock->l_resource->lr_type;
 177
 178         for (i = 0; i < lock->l_refc; i++) {
 179                 int rc;
 180                 rc = ldlm_resource_put(lock->l_resource);
 181                 if (rc == 1 && i != lock->l_refc - 1)
 182                         LBUG();
 183         }
 184
 185         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid, type, 1);
 186         if (lock->l_resource == NULL) {
 187                 LBUG();
 188                 RETURN(-ENOMEM);
 189         }
 190
 191         for (i = 1; i < lock->l_refc; i++)
 192                 ldlm_resource_addref(lock->l_resource);
 193
 194         l_unlock(&ns->ns_lock);
 195         RETURN(0);
 196 }
 197
 198 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
 199                               ldlm_mode_t mode, void *data)
 200 {
 201         struct ptlrpc_request *req = req_cookie;
 202         int rc = 0;
 203         ENTRY;
 204
 205         if (!req_cookie)
 206                 RETURN(0);
 207
 208         if (req->rq_reqmsg->bufcount > 1) {
 209                 /* an intent needs to be considered */
 210                 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
 211                 struct mds_body *mds_rep;
 212                 struct ldlm_reply *rep;
 213                 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 214                 __u32 type = lock->l_resource->lr_type;
 215                 __u64 new_resid[3] = {0, 0, 0}, old_res;
 216                 int bufcount = -1, rc, size[3] = {sizeof(struct ldlm_reply),
 217                                                   sizeof(struct mds_body),
 218                                                   sizeof(struct obdo)};
 219
 220                 it->opc = NTOH__u64(it->opc);
 221
 222                 LDLM_DEBUG(lock, "intent policy, opc: %Ld", it->opc);
 223
 224                 /* prepare reply */
 225                 switch(it->opc) {
 226                 case IT_GETATTR:
 227                         /* Note that in the negative case you may be returning
 228                          * a file and its obdo */
 229                 case IT_CREAT:
 230                 case IT_CREAT|IT_OPEN:
 231                 case IT_MKDIR:
 232                 case IT_SYMLINK:
 233                 case IT_MKNOD:
 234                 case IT_LINK:
 235                 case IT_OPEN:
 236                 case IT_RENAME:
 237                         bufcount = 3;
 238                         break;
 239                 case IT_UNLINK:
 240                         bufcount = 2;
 241                         size[1] = sizeof(struct obdo);
 242                         break;
 243                 case IT_RMDIR:
 244                         bufcount = 1;
 245                         break;
 246                 default:
 247                         LBUG();
 248                 }
 249
 250                 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 251                                      &req->rq_repmsg);
 252                 if (rc) {
 253                         rc = req->rq_status = -ENOMEM;
 254                         RETURN(rc);
 255                 }
 256
 257                 rep = lustre_msg_buf(req->rq_repmsg, 0);
 258                 rep->lock_policy_res1 = 1;
 259
 260                 /* execute policy */
 261                 switch (it->opc) {
 262                 case IT_CREAT:
 263                 case IT_CREAT|IT_OPEN:
 264                 case IT_MKDIR:
 265                 case IT_SETATTR:
 266                 case IT_SYMLINK:
 267                 case IT_MKNOD:
 268                 case IT_LINK:
 269                 case IT_UNLINK:
 270                 case IT_RMDIR:
 271                 case IT_RENAME2:
 272                         if (mds_reint_p == NULL)
 273                                 mds_reint_p =
 274                                         inter_module_get_request
 275                                         ("mds_reint", "mds");
 276                         if (IS_ERR(mds_reint_p)) {
 277                                 CERROR("MDSINTENT locks require the MDS "
 278                                        "module.\n");
 279                                 LBUG();
 280                                 RETURN(-EINVAL);
 281                         }
 282                         rc = mds_reint_p(2, req);
 283                         if (rc)
 284                                 LBUG();
 285                         break;
 286                 case IT_GETATTR:
 287                 case IT_READDIR:
 288                 case IT_RENAME:
 289                 case IT_OPEN:
 290                         if (mds_getattr_name_p == NULL)
 291                                 mds_getattr_name_p =
 292                                         inter_module_get_request
 293                                         ("mds_getattr_name", "mds");
 294                         if (IS_ERR(mds_getattr_name_p)) {
 295                                 CERROR("MDSINTENT locks require the MDS "
 296                                        "module.\n");
 297                                 LBUG();
 298                                 RETURN(-EINVAL);
 299                         }
 300                         rc = mds_getattr_name_p(2, req);
 301                         if (rc) {
 302                                 req->rq_status = rc;
 303                                 RETURN(rc);
 304                         }
 305                         break;
 306                 case IT_READDIR|IT_OPEN:
 307                         LBUG();
 308                         break;
 309                 default:
 310                         CERROR("Unhandled intent\n");
 311                         LBUG();
 312                 }
 313
 314                 if (it->opc == IT_UNLINK || it->opc == IT_RMDIR)
 315                         RETURN(ELDLM_LOCK_ABORTED);
 316
 317                 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
 318                 rep->lock_policy_res2 = req->rq_status;
 319                 new_resid[0] = mds_rep->ino;
 320                 old_res = lock->l_resource->lr_name[0];
 321
 322                 CDEBUG(D_INFO, "remote intent: locking %d instead of"
 323                        "%ld\n", mds_rep->ino, (long)old_res);
 324
 325                 ldlm_lock_change_resource(lock, new_resid);
 326                 if (lock->l_resource == NULL) {
 327                         LBUG();
 328                         RETURN(-ENOMEM);
 329                 }
 330                 LDLM_DEBUG(lock, "intent policy, old res %ld",
 331                            (long)old_res);
 332                 RETURN(ELDLM_LOCK_CHANGED);
 333         } else {
 334                 int size = sizeof(struct ldlm_reply);
 335                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 336                                      &req->rq_repmsg);
 337                 if (rc) {
 338                         CERROR("out of memory\n");
 339                         LBUG();
 340                         RETURN(-ENOMEM);
 341                 }
 342         }
 343         RETURN(rc);
 344 }
 345
 346 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 347 {
 348         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 349 }
 350
 351 /* Args: unreferenced, locked lock
 352  *
 353  * Caller must do its own ldlm_resource_put() on lock->l_resource */
 354 void ldlm_lock_free(struct ldlm_lock *lock)
 355 {
 356         if (!list_empty(&lock->l_children)) {
 357                 CERROR("lock %p still has children!\n", lock);
 358                 ldlm_lock_dump(lock);
 359                 LBUG();
 360         }
 361
 362         if (lock->l_readers || lock->l_writers)
 363                 CDEBUG(D_INFO, "lock still has references (%d readers, %d "
 364                        "writers)\n", lock->l_readers, lock->l_writers);
 365
 366         if (lock->l_connection)
 367                 ptlrpc_put_connection(lock->l_connection);
 368         kmem_cache_free(ldlm_lock_slab, lock);
 369 }
 370
 371 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 372 {
 373         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 374         desc->l_req_mode = lock->l_req_mode;
 375         desc->l_granted_mode = lock->l_granted_mode;
 376         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 377         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 378 }
 379
 380 static int ldlm_send_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock *new)
 381 {
 382         struct ptlrpc_request *req = NULL;
 383         ENTRY;
 384
 385         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 386         if (lock->l_flags & LDLM_FL_AST_SENT) {
 387                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 388                 RETURN(0);
 389         }
 390
 391         lock->l_flags |= LDLM_FL_AST_SENT;
 392
 393         lock->l_blocking_ast(lock, new, lock->l_data, lock->l_data_len, &req);
 394         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 395
 396         if (req != NULL) {
 397                 struct list_head *list = lock->l_resource->lr_tmp;
 398                 list_add(&req->rq_multi, list);
 399         }
 400         RETURN(1);
 401 }
 402
 403 /* Args: unlocked lock */
 404 void ldlm_lock_addref(struct ldlm_lock *lock, __u32 mode)
 405 {
 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         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 412         ldlm_lock_get(lock);
 413 }
 414
 415 /* Args: unlocked lock */
 416 void ldlm_lock_decref(struct ldlm_lock *lock, __u32 mode)
 417 {
 418         ENTRY;
 419
 420         if (lock == NULL)
 421                 LBUG();
 422
 423         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 424         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 425                 lock->l_readers--;
 426         else
 427                 lock->l_writers--;
 428
 429         /* If we received a blocked AST and this was the last reference,
 430          * run the callback. */
 431         if (!lock->l_readers && !lock->l_writers &&
 432             (lock->l_flags & LDLM_FL_BLOCKED_PENDING)) {
 433                 if (!lock->l_resource->lr_namespace->ns_client) {
 434                         CERROR("LDLM_FL_DYING set on non-local lock!\n");
 435                         LBUG();
 436                 }
 437
 438                 CDEBUG(D_INFO, "final decref done on dying lock, "
 439                        "calling callback.\n");
 440                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 441
 442                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 443                                      lock->l_data_len, NULL);
 444         } else
 445                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 446
 447         ldlm_lock_put(lock);
 448
 449         EXIT;
 450 }
 451
 452 static int _ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs,
 453                              struct list_head *queue)
 454 {
 455         struct list_head *tmp, *pos;
 456         int rc = 0;
 457
 458         list_for_each_safe(tmp, pos, queue) {
 459                 struct ldlm_lock *child;
 460                 ldlm_res_compat compat;
 461
 462                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 463                 if (lock == child)
 464                         continue;
 465
 466                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 467                 if (compat(child, lock)) {
 468                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 469                         continue;
 470                 }
 471                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 472                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 473                         continue;
 474                 }
 475
 476                 rc = 1;
 477
 478                 CDEBUG(D_OTHER, "compat function failed and lock modes "
 479                        "incompat\n");
 480                 if (send_cbs && child->l_blocking_ast != NULL) {
 481                         CDEBUG(D_OTHER, "incompatible; sending blocking "
 482                                "AST.\n");
 483                         /* It's very difficult to actually send the AST from
 484                          * here, because we'd have to drop the lock before going
 485                          * to sleep to wait for the reply.  Instead we build the
 486                          * packet and send it later. */
 487                         ldlm_send_blocking_ast(child, lock);
 488                 }
 489         }
 490
 491         return rc;
 492 }
 493
 494 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 495 {
 496         int rc;
 497         ENTRY;
 498
 499         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 500         rc = _ldlm_lock_compat(lock, send_cbs, &lock->l_resource->lr_granted);
 501         /* FIXME: should we be sending ASTs to converting? */
 502         rc |= _ldlm_lock_compat(lock, send_cbs,
 503                                 &lock->l_resource->lr_converting);
 504         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 505
 506         RETURN(rc);
 507 }
 508
 509 void ldlm_grant_lock(struct ldlm_resource *res, struct ldlm_lock *lock)
 510 {
 511         struct ptlrpc_request *req = NULL;
 512         ENTRY;
 513
 514         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 515         lock->l_granted_mode = lock->l_req_mode;
 516
 517         if (lock->l_granted_mode < res->lr_most_restr)
 518                 res->lr_most_restr = lock->l_granted_mode;
 519
 520         if (lock->l_completion_ast) {
 521                 lock->l_completion_ast(lock, NULL, lock->l_data,
 522                                        lock->l_data_len, &req);
 523                 if (req != NULL) {
 524                         struct list_head *list = res->lr_tmp;
 525                         if (list == NULL) {
 526                                 LBUG();
 527                                 return;
 528                         }
 529                         list_add(&req->rq_multi, list);
 530                 }
 531         }
 532         EXIT;
 533 }
 534
 535 static int search_queue(struct list_head *queue, ldlm_mode_t mode,
 536                         struct ldlm_extent *extent, struct lustre_handle *lockh)
 537 {
 538         struct list_head *tmp;
 539
 540         list_for_each(tmp, queue) {
 541                 struct ldlm_lock *lock;
 542                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 543
 544                 if (lock->l_flags & LDLM_FL_DYING)
 545                         continue;
 546
 547                 /* lock_convert() takes the resource lock, so we're sure that
 548                  * req_mode, lr_type, and l_cookie won't change beneath us */
 549                 if (lock->l_req_mode != mode)
 550                         continue;
 551
 552                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 553                     (lock->l_extent.start > extent->start ||
 554                      lock->l_extent.end < extent->end))
 555                         continue;
 556
 557                 ldlm_lock_addref(lock, mode);
 558                 ldlm_object2handle(lock, lockh);
 559                 return 1;
 560         }
 561
 562         return 0;
 563 }
 564
 565 /* Must be called with no resource or lock locks held.
 566  *
 567  * Returns 1 if it finds an already-existing lock that is compatible; in this
 568  * case, lockh is filled in with a addref()ed lock */
 569 int ldlm_local_lock_match(struct ldlm_namespace *ns, __u64 *res_id, __u32 type,
 570                           void *cookie, int cookielen, ldlm_mode_t mode,
 571                           struct lustre_handle *lockh)
 572 {
 573         struct ldlm_resource *res;
 574         struct ldlm_namespace *ns;
 575         int rc = 0;
 576         ENTRY;
 577
 578         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 579         if (res == NULL)
 580                 RETURN(0);
 581
 582         ns = res->lr_namespace;
 583         l_lock(&ns->ns_lock);
 584
 585         if (search_queue(&res->lr_granted, mode, cookie, lockh))
 586                 GOTO(out, rc = 1);
 587         if (search_queue(&res->lr_converting, mode, cookie, lockh))
 588                 GOTO(out, rc = 1);
 589         if (search_queue(&res->lr_waiting, mode, cookie, lockh))
 590                 GOTO(out, rc = 1);
 591
 592         EXIT;
 593  out:
 594         ldlm_resource_put(res);
 595         l_unlock(&ns->ns_lock);
 596
 597         wait_event_interruptible(lock->l_waitq, lock->l_req_mode ==
 598                                  lock->l_granted_mode);
 599
 600         return rc;
 601 }
 602
 603 /* Must be called without the resource lock held.  Returns a referenced,
 604  * unlocked ldlm_lock. */
 605 struct ldlm_lock *
 606 ldlm_local_lock_create(struct ldlm_namespace *ns,
 607                        struct lustre_handle *parent_lock_handle,
 608                        __u64 *res_id, __u32 type,
 609                        ldlm_mode_t mode,
 610                        void *data,
 611                        __u32 data_len)
 612 {
 613         struct ldlm_resource *res, *parent_res = NULL;
 614         struct ldlm_lock *lock, *parent_lock;
 615
 616         parent_lock = lustre_handle2object(parent_lock_handle);
 617         if (parent_lock)
 618                 parent_res = parent_lock->l_resource;
 619
 620         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 621         if (res == NULL)
 622                 RETURN(NULL);
 623
 624         lock = ldlm_lock_new(parent_lock, res);
 625         if (lock == NULL) {
 626                 ldlm_resource_put(res);
 627                 RETURN(NULL);
 628         }
 629
 630         lock->l_req_mode = mode;
 631         lock->l_data = data;
 632         lock->l_data_len = data_len;
 633         ldlm_lock_addref(lock, mode);
 634
 635         return lock;
 636 }
 637
 638 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 639 ldlm_error_t ldlm_local_lock_enqueue(struct lustre_handle *lockh,
 640                                      void *cookie, int cookie_len,
 641                                      int *flags,
 642                                      ldlm_lock_callback completion,
 643                                      ldlm_lock_callback blocking)
 644 {
 645         struct ldlm_resource *res;
 646         struct ldlm_lock *lock;
 647         int incompat = 0, local;
 648         ldlm_res_policy policy;
 649         ENTRY;
 650
 651         lock = lustre_handle2object(lockh);
 652         res = lock->l_resource;
 653         local = res->lr_namespace->ns_client;
 654
 655         lock->l_blocking_ast = blocking;
 656
 657         if (res->lr_type == LDLM_EXTENT)
 658                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 659
 660         /* policies are not executed on the client */
 661         if (!local && (policy = ldlm_res_policy_table[res->lr_type])) {
 662                 int rc;
 663
 664                 /* We do this dancing with refcounts and locks because the
 665                  * policy function could send an RPC */
 666                 ldlm_resource_addref(res);
 667
 668                 rc = policy(lock, cookie, lock->l_req_mode, NULL);
 669
 670                 if (ldlm_resource_put(res) && rc != ELDLM_LOCK_CHANGED)
 671                         /* ldlm_resource_put() should not destroy 'res' unless
 672                          * 'res' is no longer the resource for this lock. */
 673                         LBUG();
 674
 675                 if (rc == ELDLM_LOCK_CHANGED) {
 676                         res = lock->l_resource;
 677                         *flags |= LDLM_FL_LOCK_CHANGED;
 678                 } else if (rc == ELDLM_LOCK_ABORTED) {
 679                         ldlm_lock_destroy(lock);
 680                         RETURN(rc);
 681                 }
 682         }
 683
 684         lock->l_cookie = cookie;
 685         lock->l_cookie_len = cookie_len;
 686
 687         if (local && lock->l_req_mode == lock->l_granted_mode) {
 688                 /* The server returned a blocked lock, but it was granted before
 689                  * we got a chance to actually enqueue it.  We don't need to do
 690                  * anything else. */
 691                 GOTO(out, ELDLM_OK);
 692         }
 693
 694         /* If this is a local resource, put it on the appropriate list. */
 695         list_del_init(&lock->l_res_link);
 696         if (local) {
 697                 if (*flags & LDLM_FL_BLOCK_CONV)
 698                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 699                                                lock);
 700                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 701                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 702                 else
 703                         ldlm_grant_lock(res, lock);
 704                 GOTO(out, ELDLM_OK);
 705         }
 706
 707         /* FIXME: We may want to optimize by checking lr_most_restr */
 708         if (!list_empty(&res->lr_converting)) {
 709                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 710                 *flags |= LDLM_FL_BLOCK_CONV;
 711                 GOTO(out, ELDLM_OK);
 712         }
 713         if (!list_empty(&res->lr_waiting)) {
 714                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 715                 *flags |= LDLM_FL_BLOCK_WAIT;
 716                 GOTO(out, ELDLM_OK);
 717         }
 718         incompat = ldlm_lock_compat(lock, 0);
 719         if (incompat) {
 720                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 721                 *flags |= LDLM_FL_BLOCK_GRANTED;
 722                 GOTO(out, ELDLM_OK);
 723         }
 724
 725         ldlm_grant_lock(res, lock);
 726         EXIT;
 727  out:
 728         /* Don't set 'completion_ast' until here so that if the lock is granted
 729          * immediately we don't do an unnecessary completion call. */
 730         lock->l_completion_ast = completion;
 731         return ELDLM_OK;
 732 }
 733
 734 /* Must be called with resource->lr_lock taken. */
 735 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 736                                 struct list_head *converting)
 737 {
 738         struct list_head *tmp, *pos;
 739         ENTRY;
 740
 741         list_for_each_safe(tmp, pos, converting) {
 742                 struct ldlm_lock *pending;
 743                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 744
 745                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 746
 747                 if (ldlm_lock_compat(pending, 1))
 748                         RETURN(1);
 749
 750                 list_del_init(&pending->l_res_link);
 751                 ldlm_grant_lock(res, pending);
 752
 753                 ldlm_lock_addref(pending, pending->l_req_mode);
 754                 ldlm_lock_decref(pending, pending->l_granted_mode);
 755         }
 756
 757         RETURN(0);
 758 }
 759
 760 static void ldlm_send_delayed_asts(struct list_head *rpc_list)
 761 {
 762         struct list_head *tmp, *pos;
 763         ENTRY;
 764
 765         list_for_each_safe(tmp, pos, rpc_list) {
 766                 int rc;
 767                 struct ptlrpc_request *req =
 768                         list_entry(tmp, struct ptlrpc_request, rq_multi);
 769
 770                 CDEBUG(D_INFO, "Sending callback.\n");
 771
 772                 rc = ptlrpc_queue_wait(req);
 773                 rc = ptlrpc_check_status(req, rc);
 774                 ptlrpc_free_req(req);
 775                 if (rc)
 776                         CERROR("Callback send failed: %d\n", rc);
 777         }
 778         EXIT;
 779 }
 780
 781 /* Must be called with resource->lr_lock not taken. */
 782 void ldlm_reprocess_all(struct ldlm_resource *res)
 783 {
 784         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 785         ENTRY;
 786
 787         /* Local lock trees don't get reprocessed. */
 788         if (res->lr_namespace->ns_client) {
 789                 EXIT;
 790                 return;
 791         }
 792
 793         l_lock(&res->lr_namespace->ns_lock);
 794         res->lr_tmp = &rpc_list;
 795
 796         ldlm_reprocess_queue(res, &res->lr_converting);
 797         if (list_empty(&res->lr_converting))
 798                 ldlm_reprocess_queue(res, &res->lr_waiting);
 799
 800         res->lr_tmp = NULL;
 801         l_unlock(&res->lr_namespace->ns_lock);
 802
 803         ldlm_send_delayed_asts(&rpc_list);
 804         EXIT;
 805 }
 806
 807 /* Must be called with lock and lock->l_resource unlocked */
 808 struct ldlm_resource *ldlm_local_lock_cancel(struct ldlm_lock *lock)
 809 {
 810         struct ldlm_resource *res;
 811         struct ldlm_namespace *ns;
 812         ENTRY;
 813
 814         res = lock->l_resource;
 815         ns = res->lr_namespace;
 816
 817         l_lock(&ns->ns_lock);
 818         if (lock->l_readers || lock->l_writers)
 819                 CDEBUG(D_INFO, "lock still has references (%d readers, %d "
 820                        "writers)\n", lock->l_readers, lock->l_writers);
 821
 822         ldlm_resource_del_lock(lock);
 823         ldlm_lock_destroy(lock);
 824
 825         l_unlock(&ns->ns_lock);
 826
 827         RETURN(res);
 828 }
 829
 830 /* Must be called with lock and lock->l_resource unlocked */
 831 struct ldlm_resource *ldlm_local_lock_convert(struct lustre_handle *lockh,
 832                                               int new_mode, int *flags)
 833 {
 834         struct ldlm_lock *lock;
 835         struct ldlm_resource *res;
 836         struct ldlm_namespace *ns;
 837         ENTRY;
 838
 839         lock = lustre_handle2object(lockh);
 840         res = lock->l_resource;
 841         ns = res->lr_namespace;
 842
 843         l_lock(&ns->ns_lock);
 844
 845         lock->l_req_mode = new_mode;
 846         ldlm_resource_del_lock(lock);
 847
 848         /* If this is a local resource, put it on the appropriate list. */
 849         if (res->lr_namespace->ns_client) {
 850                 if (*flags & LDLM_FL_BLOCK_CONV)
 851                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 852                                                lock);
 853                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 854                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 855                 else
 856                         ldlm_grant_lock(res, lock);
 857         } else {
 858                 list_add(&lock->l_res_link, res->lr_converting.prev);
 859         }
 860
 861         l_unlock(&ns->ns_lock);
 862
 863         RETURN(res);
 864 }
 865
 866 void ldlm_lock_dump(struct ldlm_lock *lock)
 867 {
 868         char ver[128];
 869
 870         if (!(portal_debug & D_OTHER))
 871                 return;
 872
 873         if (RES_VERSION_SIZE != 4)
 874                 LBUG();
 875
 876         if (!lock) {
 877                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
 878                 return;
 879         }
 880
 881         snprintf(ver, sizeof(ver), "%x %x %x %x",
 882                  lock->l_version[0], lock->l_version[1],
 883                  lock->l_version[2], lock->l_version[3]);
 884
 885         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
 886         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
 887         CDEBUG(D_OTHER, "  Resource: %p (%Ld)\n", lock->l_resource,
 888                lock->l_resource->lr_name[0]);
 889         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
 890                (int)lock->l_req_mode, (int)lock->l_granted_mode);
 891         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
 892                lock->l_readers, lock->l_writers);
 893         if (lock->l_resource->lr_type == LDLM_EXTENT)
 894                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
 895                        (unsigned long long)lock->l_extent.start,
 896                        (unsigned long long)lock->l_extent.end);
 897 }