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/random.h>
  19 #include <linux/lustre_dlm.h>
  20 #include <linux/lustre_mds.h>
  21
  22 extern kmem_cache_t *ldlm_lock_slab;
  23 int (*mds_reint_p)(int offset, struct ptlrpc_request *req) = NULL;
  24 int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req) = NULL;
  25
  26 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b);
  27 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
  28                               ldlm_mode_t mode, void *data);
  29
  30 ldlm_res_compat ldlm_res_compat_table [] = {
  31         [LDLM_PLAIN] ldlm_plain_compat,
  32         [LDLM_EXTENT] ldlm_extent_compat,
  33         [LDLM_MDSINTENT] ldlm_plain_compat
  34 };
  35
  36 ldlm_res_policy ldlm_res_policy_table [] = {
  37         [LDLM_PLAIN] NULL,
  38         [LDLM_EXTENT] ldlm_extent_policy,
  39         [LDLM_MDSINTENT] ldlm_intent_policy
  40 };
  41
  42 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
  43 {
  44         lockh->addr = (__u64)(unsigned long)lock;
  45         lockh->cookie = lock->l_random;
  46 }
  47
  48 struct ldlm_lock *ldlm_handle2lock(struct lustre_handle *handle)
  49 {
  50         struct ldlm_lock *lock = NULL;
  51         ENTRY;
  52
  53         if (!handle)
  54                 RETURN(NULL);
  55         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
  56
  57         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock))
  58                 RETURN(NULL);
  59
  60         l_lock(&lock->l_resource->lr_namespace->ns_lock);
  61         if (lock->l_random != handle->cookie)
  62                 GOTO(out, handle = NULL);
  63
  64         if (lock->l_flags & LDLM_FL_DESTROYED)
  65                 GOTO(out, handle = NULL);
  66
  67         lock->l_refc++;
  68         EXIT;
  69  out:
  70         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
  71         return  lock;
  72 }
  73
  74 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
  75 {
  76         l_lock(&lock->l_resource->lr_namespace->ns_lock);
  77         lock->l_refc++;
  78         ldlm_resource_getref(lock->l_resource);
  79         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
  80         return lock;
  81 }
  82
  83 void ldlm_lock_put(struct ldlm_lock *lock)
  84 {
  85         struct lustre_lock *nslock = &lock->l_resource->lr_namespace->ns_lock;
  86         ENTRY;
  87
  88         l_lock(nslock);
  89         lock->l_refc--;
  90         if (lock->l_refc < 0)
  91                 LBUG();
  92
  93         ldlm_resource_put(lock->l_resource);
  94         if (lock->l_parent)
  95                 ldlm_lock_put(lock->l_parent);
  96
  97         if (lock->l_refc == 0 && (lock->l_flags & LDLM_FL_DESTROYED)) {
  98                 if (lock->l_connection)
  99                         ptlrpc_put_connection(lock->l_connection);
 100                 kmem_cache_free(ldlm_lock_slab, lock);
 101         }
 102         l_unlock(nslock);
 103         EXIT;
 104         return;
 105 }
 106
 107 void ldlm_lock_destroy(struct ldlm_lock *lock)
 108 {
 109         ENTRY;
 110         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 111
 112         if (!list_empty(&lock->l_children)) {
 113                 CERROR("lock %p still has children (%p)!\n", lock,
 114                        lock->l_children.next);
 115                 ldlm_lock_dump(lock);
 116                 LBUG();
 117         }
 118         if (lock->l_readers || lock->l_writers) {
 119                 CDEBUG(D_INFO, "lock still has references (%d readers, %d "
 120                        "writers)\n", lock->l_readers, lock->l_writers);
 121                 LBUG();
 122         }
 123
 124         if (!list_empty(&lock->l_res_link))
 125                 LBUG();
 126
 127         lock->l_flags = LDLM_FL_DESTROYED;
 128         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 129         ldlm_lock_put(lock);
 130         EXIT;
 131         return;
 132 }
 133 /*
 134    usage: pass in a resource on which you have done get
 135           pass in a parent lock on which you have done a get
 136           do not put the resource or the parent
 137    returns: lock with refcount 1
 138 */
 139 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 140                                        struct ldlm_resource *resource)
 141 {
 142         struct ldlm_lock *lock;
 143         ENTRY;
 144
 145         if (resource == NULL)
 146                 LBUG();
 147
 148         lock = kmem_cache_alloc(ldlm_lock_slab, SLAB_KERNEL);
 149         if (lock == NULL)
 150                 RETURN(NULL);
 151
 152         memset(lock, 0, sizeof(*lock));
 153         get_random_bytes(&lock->l_cookie, sizeof(__u64));
 154
 155         lock->l_resource = resource;
 156         lock->l_refc = 1;
 157         INIT_LIST_HEAD(&lock->l_children);
 158         INIT_LIST_HEAD(&lock->l_res_link);
 159         init_waitqueue_head(&lock->l_waitq);
 160
 161         if (parent != NULL) {
 162                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 163                 lock->l_parent = parent;
 164                 list_add(&lock->l_childof, &parent->l_children);
 165                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 166         }
 167         RETURN(lock);
 168 }
 169
 170 int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3])
 171 {
 172         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 173         int type, i;
 174         ENTRY;
 175
 176         l_lock(&ns->ns_lock);
 177         type = lock->l_resource->lr_type;
 178
 179         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid, type, 1);
 180         if (lock->l_resource == NULL) {
 181                 LBUG();
 182                 RETURN(-ENOMEM);
 183         }
 184
 185         /* move references over */
 186         for (i = 0; i < lock->l_refc; i++) {
 187                 int rc;
 188                 ldlm_resource_getref(lock->l_resource);
 189                 rc = ldlm_resource_put(lock->l_resource);
 190                 if (rc == 1 && i != lock->l_refc - 1)
 191                         LBUG();
 192         }
 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                 __u64 new_resid[3] = {0, 0, 0}, old_res;
 214                 int bufcount = -1, rc, size[3] = {sizeof(struct ldlm_reply),
 215                                                   sizeof(struct mds_body),
 216                                                   sizeof(struct obdo)};
 217
 218                 it->opc = NTOH__u64(it->opc);
 219
 220                 LDLM_DEBUG(lock, "intent policy, opc: %Ld", it->opc);
 221
 222                 /* prepare reply */
 223                 switch(it->opc) {
 224                 case IT_GETATTR:
 225                         /* Note that in the negative case you may be returning
 226                          * a file and its obdo */
 227                 case IT_CREAT:
 228                 case IT_CREAT|IT_OPEN:
 229                 case IT_MKDIR:
 230                 case IT_SYMLINK:
 231                 case IT_MKNOD:
 232                 case IT_LINK:
 233                 case IT_OPEN:
 234                 case IT_RENAME:
 235                         bufcount = 3;
 236                         break;
 237                 case IT_UNLINK:
 238                         bufcount = 2;
 239                         size[1] = sizeof(struct obdo);
 240                         break;
 241                 case IT_RMDIR:
 242                         bufcount = 1;
 243                         break;
 244                 default:
 245                         LBUG();
 246                 }
 247
 248                 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 249                                      &req->rq_repmsg);
 250                 if (rc) {
 251                         rc = req->rq_status = -ENOMEM;
 252                         RETURN(rc);
 253                 }
 254
 255                 rep = lustre_msg_buf(req->rq_repmsg, 0);
 256                 rep->lock_policy_res1 = 1;
 257
 258                 /* execute policy */
 259                 switch (it->opc) {
 260                 case IT_CREAT:
 261                 case IT_CREAT|IT_OPEN:
 262                 case IT_MKDIR:
 263                 case IT_SETATTR:
 264                 case IT_SYMLINK:
 265                 case IT_MKNOD:
 266                 case IT_LINK:
 267                 case IT_UNLINK:
 268                 case IT_RMDIR:
 269                 case IT_RENAME2:
 270                         if (mds_reint_p == NULL)
 271                                 mds_reint_p =
 272                                         inter_module_get_request
 273                                         ("mds_reint", "mds");
 274                         if (IS_ERR(mds_reint_p)) {
 275                                 CERROR("MDSINTENT locks require the MDS "
 276                                        "module.\n");
 277                                 LBUG();
 278                                 RETURN(-EINVAL);
 279                         }
 280                         rc = mds_reint_p(2, req);
 281                         if (rc)
 282                                 LBUG();
 283                         break;
 284                 case IT_GETATTR:
 285                 case IT_READDIR:
 286                 case IT_RENAME:
 287                 case IT_OPEN:
 288                         if (mds_getattr_name_p == NULL)
 289                                 mds_getattr_name_p =
 290                                         inter_module_get_request
 291                                         ("mds_getattr_name", "mds");
 292                         if (IS_ERR(mds_getattr_name_p)) {
 293                                 CERROR("MDSINTENT locks require the MDS "
 294                                        "module.\n");
 295                                 LBUG();
 296                                 RETURN(-EINVAL);
 297                         }
 298                         rc = mds_getattr_name_p(2, req);
 299                         if (rc) {
 300                                 req->rq_status = rc;
 301                                 RETURN(rc);
 302                         }
 303                         break;
 304                 case IT_READDIR|IT_OPEN:
 305                         LBUG();
 306                         break;
 307                 default:
 308                         CERROR("Unhandled intent\n");
 309                         LBUG();
 310                 }
 311
 312                 if (it->opc == IT_UNLINK || it->opc == IT_RMDIR)
 313                         RETURN(ELDLM_LOCK_ABORTED);
 314
 315                 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
 316                 rep->lock_policy_res2 = req->rq_status;
 317                 new_resid[0] = mds_rep->ino;
 318                 old_res = lock->l_resource->lr_name[0];
 319
 320                 CDEBUG(D_INFO, "remote intent: locking %d instead of"
 321                        "%ld\n", mds_rep->ino, (long)old_res);
 322
 323                 ldlm_lock_change_resource(lock, new_resid);
 324                 if (lock->l_resource == NULL) {
 325                         LBUG();
 326                         RETURN(-ENOMEM);
 327                 }
 328                 LDLM_DEBUG(lock, "intent policy, old res %ld",
 329                            (long)old_res);
 330                 RETURN(ELDLM_LOCK_CHANGED);
 331         } else {
 332                 int size = sizeof(struct ldlm_reply);
 333                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 334                                      &req->rq_repmsg);
 335                 if (rc) {
 336                         CERROR("out of memory\n");
 337                         LBUG();
 338                         RETURN(-ENOMEM);
 339                 }
 340         }
 341         RETURN(rc);
 342 }
 343
 344 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 345 {
 346         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 347 }
 348
 349 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 350 {
 351         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 352         desc->l_req_mode = lock->l_req_mode;
 353         desc->l_granted_mode = lock->l_granted_mode;
 354         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 355         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 356 }
 357
 358 static int ldlm_send_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock *new)
 359 {
 360         struct lustre_handle lockh;
 361         struct ldlm_lock_desc desc;
 362         struct ptlrpc_request *req = NULL;
 363         ENTRY;
 364
 365
 366         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 367         if (lock->l_flags & LDLM_FL_AST_SENT) {
 368                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 369                 RETURN(0);
 370         }
 371
 372         lock->l_flags |= LDLM_FL_AST_SENT;
 373         /* FIXME: this should merely add the lock to the lr_tmp list */
 374         ldlm_lock2handle(lock, &lockh);
 375         ldlm_lock2desc(new, &desc);
 376         lock->l_blocking_ast(&lockh, &desc, lock->l_data, lock->l_data_len,
 377                              &req);
 378         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 379
 380         if (req != NULL) {
 381                 struct list_head *list = lock->l_resource->lr_tmp;
 382                 list_add(&req->rq_multi, list);
 383         }
 384         RETURN(1);
 385 }
 386
 387 /* Args: unlocked lock */
 388 void ldlm_lock_addref(struct ldlm_lock *lock, __u32 mode)
 389 {
 390         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 391         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 392                 lock->l_readers++;
 393         else
 394                 lock->l_writers++;
 395         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 396         ldlm_lock_get(lock);
 397 }
 398
 399 /* Args: unlocked lock */
 400 void ldlm_lock_decref(struct ldlm_lock *lock, __u32 mode)
 401 {
 402         ENTRY;
 403
 404         if (lock == NULL)
 405                 LBUG();
 406
 407         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 408         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 409                 lock->l_readers--;
 410         else
 411                 lock->l_writers--;
 412
 413         /* If we received a blocked AST and this was the last reference,
 414          * run the callback. */
 415         if (!lock->l_readers && !lock->l_writers &&
 416             (lock->l_flags & LDLM_FL_CBPENDING)) {
 417                 struct lustre_handle lockh;
 418
 419                 if (!lock->l_resource->lr_namespace->ns_client) {
 420                         CERROR("LDLM_FL_CBPENDING set on non-local lock!\n");
 421                         LBUG();
 422                 }
 423
 424                 CDEBUG(D_INFO, "final decref done on cbpending lock, "
 425                        "calling callback.\n");
 426                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 427
 428                 ldlm_lock2handle(lock, &lockh);
 429                 lock->l_blocking_ast(&lockh, NULL, lock->l_data,
 430                                      lock->l_data_len, NULL);
 431         } else
 432                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 433
 434         ldlm_lock_put(lock);
 435
 436         EXIT;
 437 }
 438
 439 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 440                              struct list_head *queue)
 441 {
 442         struct list_head *tmp, *pos;
 443         int rc = 1;
 444
 445         list_for_each_safe(tmp, pos, queue) {
 446                 struct ldlm_lock *child;
 447                 ldlm_res_compat compat;
 448
 449                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 450                 if (lock == child)
 451                         continue;
 452
 453                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 454                 if (compat && compat(child, lock)) {
 455                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 456                         continue;
 457                 }
 458                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 459                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 460                         continue;
 461                 }
 462
 463                 rc = 0;
 464
 465                 CDEBUG(D_OTHER, "compat function failed and lock modes "
 466                        "incompat\n");
 467                 if (send_cbs && child->l_blocking_ast != NULL) {
 468                         CDEBUG(D_OTHER, "incompatible; sending blocking "
 469                                "AST.\n");
 470                         /* It's very difficult to actually send the AST from
 471                          * here, because we'd have to drop the lock before going
 472                          * to sleep to wait for the reply.  Instead we build the
 473                          * packet and send it later. */
 474                         ldlm_send_blocking_ast(child, lock);
 475                 }
 476         }
 477
 478         return rc;
 479 }
 480
 481 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 482 {
 483         int rc;
 484         ENTRY;
 485
 486         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 487         rc = ldlm_lock_compat_list(lock, send_cbs, &lock->l_resource->lr_granted);
 488         /* FIXME: should we be sending ASTs to converting? */
 489         if (rc)
 490                 rc = ldlm_lock_compat_list
 491                         (lock, send_cbs, &lock->l_resource->lr_converting);
 492
 493         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 494         RETURN(rc);
 495 }
 496
 497 /* NOTE: called by
 498    - ldlm_handle_enqueuque - resource
 499 */
 500 void ldlm_grant_lock(struct ldlm_lock *lock)
 501 {
 502         struct ldlm_resource *res = lock->l_resource;
 503         struct ptlrpc_request *req = NULL;
 504         ENTRY;
 505
 506         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 507         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 508         lock->l_granted_mode = lock->l_req_mode;
 509
 510         if (lock->l_granted_mode < res->lr_most_restr)
 511                 res->lr_most_restr = lock->l_granted_mode;
 512
 513         if (lock->l_completion_ast) {
 514                 struct lustre_handle lockh;
 515
 516                 /* FIXME: this should merely add lock to lr_tmp list */
 517                 ldlm_lock2handle(lock, &lockh);
 518                 lock->l_completion_ast(&lockh, NULL, lock->l_data,
 519                                        lock->l_data_len, &req);
 520                 if (req != NULL) {
 521                         struct list_head *list = res->lr_tmp;
 522                         if (list == NULL) {
 523                                 LBUG();
 524                                 return;
 525                         }
 526                         list_add(&req->rq_multi, list);
 527                 }
 528         }
 529         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 530         EXIT;
 531 }
 532
 533 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 534                                       struct ldlm_extent *extent)
 535 {
 536         struct ldlm_lock *lock;
 537         struct list_head *tmp;
 538
 539         list_for_each(tmp, queue) {
 540                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 541
 542                 if (lock->l_flags & LDLM_FL_CBPENDING)
 543                         continue;
 544
 545                 /* lock_convert() takes the resource lock, so we're sure that
 546                  * req_mode, lr_type, and l_cookie won't change beneath us */
 547                 if (lock->l_req_mode != mode)
 548                         continue;
 549
 550                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 551                     (lock->l_extent.start > extent->start ||
 552                      lock->l_extent.end < extent->end))
 553                         continue;
 554
 555                 ldlm_lock_addref(lock, mode);
 556                 return lock;
 557         }
 558
 559         return NULL;
 560 }
 561
 562 /* Must be called with no resource or lock locks held.
 563  *
 564  * Returns 1 if it finds an already-existing lock that is compatible; in this
 565  * case, lockh is filled in with a addref()ed lock */
 566 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 *res_id, __u32 type,
 567                     void *cookie, int cookielen, ldlm_mode_t mode,
 568                     struct lustre_handle *lockh)
 569 {
 570         struct ldlm_resource *res;
 571         struct ldlm_lock *lock;
 572         int rc = 0;
 573         ENTRY;
 574
 575         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 576         if (res == NULL)
 577                 RETURN(0);
 578
 579         ns = res->lr_namespace;
 580         l_lock(&ns->ns_lock);
 581
 582         if ((lock = search_queue(&res->lr_granted, mode, cookie)))
 583                 GOTO(out, rc = 1);
 584         if ((lock = search_queue(&res->lr_converting, mode, cookie)))
 585                 GOTO(out, rc = 1);
 586         if ((lock = search_queue(&res->lr_waiting, mode, cookie)))
 587                 GOTO(out, rc = 1);
 588
 589         EXIT;
 590  out:
 591         ldlm_resource_put(res);
 592         l_unlock(&ns->ns_lock);
 593
 594         if (lock)
 595                 wait_event_interruptible(lock->l_waitq, lock->l_req_mode ==
 596                                          lock->l_granted_mode);
 597
 598         return rc;
 599 }
 600
 601 /* Must be called without the resource lock held.  Returns a referenced,
 602  * unlocked ldlm_lock. */
 603 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 604                                    struct lustre_handle *parent_lock_handle,
 605                                    __u64 *res_id, __u32 type,
 606                                    ldlm_mode_t mode,
 607                                    void *data,
 608                                    __u32 data_len)
 609 {
 610         struct ldlm_resource *res, *parent_res = NULL;
 611         struct ldlm_lock *lock, *parent_lock;
 612
 613         parent_lock = lustre_handle2object(parent_lock_handle);
 614         if (parent_lock)
 615                 parent_res = parent_lock->l_resource;
 616
 617         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 618         if (res == NULL)
 619                 RETURN(NULL);
 620
 621         lock = ldlm_lock_new(parent_lock, res);
 622         if (lock == NULL) {
 623                 ldlm_resource_put(res);
 624                 RETURN(NULL);
 625         }
 626
 627         lock->l_req_mode = mode;
 628         lock->l_data = data;
 629         lock->l_data_len = data_len;
 630         ldlm_lock_addref(lock, mode);
 631
 632         return lock;
 633 }
 634
 635 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 636 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock *lock,
 637                                void *cookie, int cookie_len,
 638                                int *flags,
 639                                ldlm_lock_callback completion,
 640                                ldlm_lock_callback blocking)
 641 {
 642         struct ldlm_resource *res;
 643         int local;
 644         ldlm_res_policy policy;
 645         ENTRY;
 646
 647         res = lock->l_resource;
 648         local = res->lr_namespace->ns_client;
 649
 650         lock->l_blocking_ast = blocking;
 651
 652         if (res->lr_type == LDLM_EXTENT)
 653                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 654
 655         /* policies are not executed on the client */
 656         if (!local && (policy = ldlm_res_policy_table[res->lr_type])) {
 657                 int rc;
 658
 659                 /* We do this dancing with refcounts and locks because the
 660                  * policy function could send an RPC */
 661                 ldlm_resource_getref(res);
 662
 663                 rc = policy(lock, cookie, lock->l_req_mode, NULL);
 664
 665                 if (ldlm_resource_put(res) && rc != ELDLM_LOCK_CHANGED)
 666                         /* ldlm_resource_put() should not destroy 'res' unless
 667                          * 'res' is no longer the resource for this lock. */
 668                         LBUG();
 669
 670                 if (rc == ELDLM_LOCK_CHANGED) {
 671                         res = lock->l_resource;
 672                         *flags |= LDLM_FL_LOCK_CHANGED;
 673                 } else if (rc == ELDLM_LOCK_ABORTED) {
 674                         ldlm_lock_destroy(lock);
 675                         RETURN(rc);
 676                 }
 677         }
 678
 679         lock->l_cookie = cookie;
 680         lock->l_cookie_len = cookie_len;
 681
 682         if (local && lock->l_req_mode == lock->l_granted_mode) {
 683                 /* The server returned a blocked lock, but it was granted before
 684                  * we got a chance to actually enqueue it.  We don't need to do
 685                  * anything else. */
 686                 GOTO(out, ELDLM_OK);
 687         }
 688
 689         /* If this is a local resource, put it on the appropriate list. */
 690         list_del_init(&lock->l_res_link);
 691         if (local) {
 692                 if (*flags & LDLM_FL_BLOCK_CONV)
 693                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 694                                                lock);
 695                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 696                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 697                 else
 698                         ldlm_grant_lock(lock);
 699                 GOTO(out, ELDLM_OK);
 700         }
 701
 702         /* FIXME: We may want to optimize by checking lr_most_restr */
 703         if (!list_empty(&res->lr_converting)) {
 704                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 705                 *flags |= LDLM_FL_BLOCK_CONV;
 706                 GOTO(out, ELDLM_OK);
 707         }
 708         if (!list_empty(&res->lr_waiting)) {
 709                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 710                 *flags |= LDLM_FL_BLOCK_WAIT;
 711                 GOTO(out, ELDLM_OK);
 712         }
 713         if (!ldlm_lock_compat(lock,0)) {
 714                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 715                 *flags |= LDLM_FL_BLOCK_GRANTED;
 716                 GOTO(out, ELDLM_OK);
 717         }
 718
 719         ldlm_grant_lock(lock);
 720         EXIT;
 721  out:
 722         /* Don't set 'completion_ast' until here so that if the lock is granted
 723          * immediately we don't do an unnecessary completion call. */
 724         lock->l_completion_ast = completion;
 725         return ELDLM_OK;
 726 }
 727
 728 /* Must be called with namespace taken: queue is waiting or converting. */
 729 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 730                                 struct list_head *queue)
 731 {
 732         struct list_head *tmp, *pos;
 733         ENTRY;
 734
 735         list_for_each_safe(tmp, pos, queue) {
 736                 struct ldlm_lock *pending;
 737                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 738
 739                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 740
 741                 if (!ldlm_lock_compat(pending, 1))
 742                         RETURN(1);
 743
 744                 list_del_init(&pending->l_res_link);
 745                 ldlm_grant_lock(pending);
 746
 747                 ldlm_lock_addref(pending, pending->l_req_mode);
 748                 ldlm_lock_decref(pending, pending->l_granted_mode);
 749         }
 750
 751         RETURN(0);
 752 }
 753
 754 static void ldlm_send_delayed_asts(struct list_head *rpc_list)
 755 {
 756         struct list_head *tmp, *pos;
 757         ENTRY;
 758
 759         list_for_each_safe(tmp, pos, rpc_list) {
 760                 int rc;
 761                 struct ptlrpc_request *req =
 762                         list_entry(tmp, struct ptlrpc_request, rq_multi);
 763
 764                 CDEBUG(D_INFO, "Sending callback.\n");
 765
 766                 rc = ptlrpc_queue_wait(req);
 767                 rc = ptlrpc_check_status(req, rc);
 768                 ptlrpc_free_req(req);
 769                 if (rc)
 770                         CERROR("Callback send failed: %d\n", rc);
 771         }
 772         EXIT;
 773 }
 774
 775 /* Must be called with resource->lr_lock not taken. */
 776 void ldlm_reprocess_all(struct ldlm_resource *res)
 777 {
 778         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 779         ENTRY;
 780
 781         /* Local lock trees don't get reprocessed. */
 782         if (res->lr_namespace->ns_client) {
 783                 EXIT;
 784                 return;
 785         }
 786
 787         l_lock(&res->lr_namespace->ns_lock);
 788         res->lr_tmp = &rpc_list;
 789
 790         ldlm_reprocess_queue(res, &res->lr_converting);
 791         if (list_empty(&res->lr_converting))
 792                 ldlm_reprocess_queue(res, &res->lr_waiting);
 793
 794         res->lr_tmp = NULL;
 795         l_unlock(&res->lr_namespace->ns_lock);
 796
 797         ldlm_send_delayed_asts(&rpc_list);
 798         EXIT;
 799 }
 800
 801 /* Must be called with lock and lock->l_resource unlocked */
 802 void ldlm_lock_cancel(struct ldlm_lock *lock)
 803 {
 804         struct ldlm_resource *res;
 805         struct ldlm_namespace *ns;
 806         ENTRY;
 807
 808         res = lock->l_resource;
 809         ns = res->lr_namespace;
 810
 811         l_lock(&ns->ns_lock);
 812         if (lock->l_readers || lock->l_writers)
 813                 CDEBUG(D_INFO, "lock still has references (%d readers, %d "
 814                        "writers)\n", lock->l_readers, lock->l_writers);
 815
 816         ldlm_resource_unlink_lock(lock);
 817         ldlm_lock_destroy(lock);
 818         l_unlock(&ns->ns_lock);
 819 }
 820
 821 /* Must be called with lock and lock->l_resource unlocked */
 822 struct ldlm_resource *ldlm_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 823 {
 824         struct ldlm_lock *lock;
 825         struct ldlm_resource *res;
 826         struct ldlm_namespace *ns;
 827         ENTRY;
 828
 829         lock = lustre_handle2object(lockh);
 830         res = lock->l_resource;
 831         ns = res->lr_namespace;
 832
 833         l_lock(&ns->ns_lock);
 834
 835         lock->l_req_mode = new_mode;
 836         ldlm_resource_unlink_lock(lock);
 837
 838         /* If this is a local resource, put it on the appropriate list. */
 839         if (res->lr_namespace->ns_client) {
 840                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED))
 841                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 842                                                lock);
 843                 else {
 844                         ldlm_grant_lock(lock);
 845                         /* FIXME: completion handling not with ns_lock held ! */
 846                         wake_up(&lock->l_waitq);
 847                 }
 848         } else {
 849                 list_add(&lock->l_res_link, res->lr_converting.prev);
 850         }
 851
 852         l_unlock(&ns->ns_lock);
 853
 854         RETURN(res);
 855 }
 856
 857 void ldlm_lock_dump(struct ldlm_lock *lock)
 858 {
 859         char ver[128];
 860
 861         if (!(portal_debug & D_OTHER))
 862                 return;
 863
 864         if (RES_VERSION_SIZE != 4)
 865                 LBUG();
 866
 867         if (!lock) {
 868                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
 869                 return;
 870         }
 871
 872         snprintf(ver, sizeof(ver), "%x %x %x %x",
 873                  lock->l_version[0], lock->l_version[1],
 874                  lock->l_version[2], lock->l_version[3]);
 875
 876         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
 877         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
 878         CDEBUG(D_OTHER, "  Resource: %p (%Ld)\n", lock->l_resource,
 879                lock->l_resource->lr_name[0]);
 880         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
 881                (int)lock->l_req_mode, (int)lock->l_granted_mode);
 882         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
 883                lock->l_readers, lock->l_writers);
 884         if (lock->l_resource->lr_type == LDLM_EXTENT)
 885                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
 886                        (unsigned long long)lock->l_extent.start,
 887                        (unsigned long long)lock->l_extent.end);
 888 }