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 //struct lustre_lock ldlm_everything_lock;
  34
  35 /* lock types */
  36 char *ldlm_lockname[] = {
  37         [0] "--",
  38         [LCK_EX] "EX",
  39         [LCK_PW] "PW",
  40         [LCK_PR] "PR",
  41         [LCK_CW] "CW",
  42         [LCK_CR] "CR",
  43         [LCK_NL] "NL"
  44 };
  45 char *ldlm_typename[] = {
  46         [LDLM_PLAIN] "PLN",
  47         [LDLM_EXTENT] "EXT",
  48         [LDLM_MDSINTENT] "INT"
  49 };
  50
  51 char *ldlm_it2str(int it)
  52 {
  53         switch (it) {
  54         case IT_OPEN:
  55                 return "open";
  56         case IT_CREAT:
  57                 return "creat";
  58         case (IT_OPEN | IT_CREAT):
  59                 return "open|creat";
  60         case IT_MKDIR:
  61                 return "mkdir";
  62         case IT_LINK:
  63                 return "link";
  64         case IT_LINK2:
  65                 return "link2";
  66         case IT_SYMLINK:
  67                 return "symlink";
  68         case IT_UNLINK:
  69                 return "unlink";
  70         case IT_RMDIR:
  71                 return "rmdir";
  72         case IT_RENAME:
  73                 return "rename";
  74         case IT_RENAME2:
  75                 return "rename2";
  76         case IT_READDIR:
  77                 return "readdir";
  78         case IT_GETATTR:
  79                 return "getattr";
  80         case IT_SETATTR:
  81                 return "setattr";
  82         case IT_READLINK:
  83                 return "readlink";
  84         case IT_MKNOD:
  85                 return "mknod";
  86         case IT_LOOKUP:
  87                 return "lookup";
  88         default:
  89                 CERROR("Unknown intent %d\n", it);
  90                 return "UNKNOWN";
  91         }
  92 }
  93
  94 extern kmem_cache_t *ldlm_lock_slab;
  95
  96 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b);
  97
  98 ldlm_res_compat ldlm_res_compat_table[] = {
  99         [LDLM_PLAIN] ldlm_plain_compat,
 100         [LDLM_EXTENT] ldlm_extent_compat,
 101         [LDLM_MDSINTENT] ldlm_plain_compat
 102 };
 103
 104 ldlm_res_policy ldlm_res_policy_table[] = {
 105         [LDLM_PLAIN] NULL,
 106         [LDLM_EXTENT] ldlm_extent_policy,
 107         [LDLM_MDSINTENT] NULL
 108 };
 109
 110 void ldlm_register_intent(int (*arg) (struct ldlm_lock * lock, void *req_cookie,
 111                                       ldlm_mode_t mode, int flags, void *data))
 112 {
 113         ldlm_res_policy_table[LDLM_MDSINTENT] = arg;
 114 }
 115
 116 void ldlm_unregister_intent(void)
 117 {
 118         ldlm_res_policy_table[LDLM_MDSINTENT] = NULL;
 119 }
 120
 121 /*
 122  * REFCOUNTED LOCK OBJECTS
 123  */
 124
 125
 126 /*
 127  * Lock refcounts, during creation:
 128  *   - one special one for allocation, dec'd only once in destroy
 129  *   - one for being a lock that's in-use
 130  *   - one for the addref associated with a new lock
 131  */
 132 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 133 {
 134         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 135         lock->l_refc++;
 136         ldlm_resource_getref(lock->l_resource);
 137         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 138         return lock;
 139 }
 140
 141 void ldlm_lock_put(struct ldlm_lock *lock)
 142 {
 143         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 144         ENTRY;
 145
 146         l_lock(&ns->ns_lock);
 147         lock->l_refc--;
 148         //LDLM_DEBUG(lock, "after refc--");
 149         if (lock->l_refc < 0)
 150                 LBUG();
 151
 152         if (ldlm_resource_put(lock->l_resource)) {
 153                 LASSERT(lock->l_refc == 0);
 154                 lock->l_resource = NULL;
 155         }
 156         if (lock->l_parent)
 157                 LDLM_LOCK_PUT(lock->l_parent);
 158
 159         if (lock->l_refc == 0 && (lock->l_flags & LDLM_FL_DESTROYED)) {
 160                 l_unlock(&ns->ns_lock);
 161                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
 162
 163                 //spin_lock(&ldlm_handle_lock);
 164                 spin_lock(&ns->ns_counter_lock);
 165                 ns->ns_locks--;
 166                 spin_unlock(&ns->ns_counter_lock);
 167
 168                 lock->l_resource = NULL;
 169                 lock->l_random = DEAD_HANDLE_MAGIC;
 170                 if (lock->l_export && lock->l_export->exp_connection)
 171                         ptlrpc_put_connection(lock->l_export->exp_connection);
 172                 kmem_cache_free(ldlm_lock_slab, lock);
 173                 //spin_unlock(&ldlm_handle_lock);
 174                 CDEBUG(D_MALLOC, "kfreed 'lock': %d at %p (tot 0).\n",
 175                        sizeof(*lock), lock);
 176         } else
 177                 l_unlock(&ns->ns_lock);
 178
 179         EXIT;
 180 }
 181
 182 void ldlm_lock_destroy(struct ldlm_lock *lock)
 183 {
 184         ENTRY;
 185         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 186
 187         if (!list_empty(&lock->l_children)) {
 188                 LDLM_DEBUG(lock, "still has children (%p)!",
 189                            lock->l_children.next);
 190                 ldlm_lock_dump(lock);
 191                 LBUG();
 192         }
 193         if (lock->l_readers || lock->l_writers) {
 194                 LDLM_DEBUG(lock, "lock still has references");
 195                 ldlm_lock_dump(lock);
 196         }
 197
 198         if (!list_empty(&lock->l_res_link)) {
 199                 ldlm_lock_dump(lock);
 200                 LBUG();
 201         }
 202
 203         if (lock->l_flags & LDLM_FL_DESTROYED) {
 204                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 205                 EXIT;
 206                 return;
 207         }
 208
 209         list_del(&lock->l_export_chain);
 210         lock->l_export = NULL;
 211         lock->l_flags |= LDLM_FL_DESTROYED;
 212
 213         /* Wake anyone waiting for this lock */
 214         /* FIXME: I should probably add yet another flag, instead of using
 215          * l_export to only call this on clients */
 216         if (lock->l_export && lock->l_completion_ast)
 217                 lock->l_completion_ast(lock, 0);
 218
 219         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 220         LDLM_LOCK_PUT(lock);
 221         EXIT;
 222 }
 223
 224 /*
 225    usage: pass in a resource on which you have done get
 226           pass in a parent lock on which you have done a get
 227           do not put the resource or the parent
 228    returns: lock with refcount 1
 229 */
 230 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 231                                        struct ldlm_resource *resource)
 232 {
 233         struct ldlm_lock *lock;
 234         ENTRY;
 235
 236         if (resource == NULL)
 237                 LBUG();
 238
 239         lock = kmem_cache_alloc(ldlm_lock_slab, SLAB_KERNEL);
 240         if (lock == NULL)
 241                 RETURN(NULL);
 242
 243         memset(lock, 0, sizeof(*lock));
 244         get_random_bytes(&lock->l_random, sizeof(__u64));
 245
 246         lock->l_resource = resource;
 247         /* this refcount matches the one of the resource passed
 248            in which is not being put away */
 249         lock->l_refc = 1;
 250         INIT_LIST_HEAD(&lock->l_children);
 251         INIT_LIST_HEAD(&lock->l_res_link);
 252         INIT_LIST_HEAD(&lock->l_export_chain);
 253         INIT_LIST_HEAD(&lock->l_pending_chain);
 254         init_waitqueue_head(&lock->l_waitq);
 255
 256         spin_lock(&resource->lr_namespace->ns_counter_lock);
 257         resource->lr_namespace->ns_locks++;
 258         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 259
 260         if (parent != NULL) {
 261                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 262                 lock->l_parent = parent;
 263                 list_add(&lock->l_childof, &parent->l_children);
 264                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 265         }
 266
 267         CDEBUG(D_MALLOC, "kmalloced 'lock': %d at "
 268                "%p (tot %d).\n", sizeof(*lock), lock, 1);
 269         /* this is the extra refcount, to prevent the lock from evaporating */
 270         LDLM_LOCK_GET(lock);
 271         RETURN(lock);
 272 }
 273
 274 int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3])
 275 {
 276         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 277         struct ldlm_resource *oldres = lock->l_resource;
 278         int type, i;
 279         ENTRY;
 280
 281         l_lock(&ns->ns_lock);
 282         if (memcmp(new_resid, lock->l_resource->lr_name,
 283                    sizeof(lock->l_resource->lr_name)) == 0) {
 284                 /* Nothing to do */
 285                 l_unlock(&ns->ns_lock);
 286                 RETURN(0);
 287         }
 288
 289         type = lock->l_resource->lr_type;
 290         if (new_resid[0] == 0)
 291                 LBUG();
 292         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid, type, 1);
 293         if (lock->l_resource == NULL) {
 294                 LBUG();
 295                 RETURN(-ENOMEM);
 296         }
 297
 298         /* move references over */
 299         for (i = 0; i < lock->l_refc; i++) {
 300                 int rc;
 301                 ldlm_resource_getref(lock->l_resource);
 302                 rc = ldlm_resource_put(oldres);
 303                 if (rc == 1 && i != lock->l_refc - 1)
 304                         LBUG();
 305         }
 306         /* compensate for the initial get above.. */
 307         ldlm_resource_put(lock->l_resource);
 308
 309         l_unlock(&ns->ns_lock);
 310         RETURN(0);
 311 }
 312
 313 /*
 314  *  HANDLES
 315  */
 316
 317 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 318 {
 319         lockh->addr = (__u64) (unsigned long)lock;
 320         lockh->cookie = lock->l_random;
 321 }
 322
 323 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle,
 324                                      int strict)
 325 {
 326         struct ldlm_lock *lock = NULL, *retval = NULL;
 327         ENTRY;
 328
 329         if (!handle || !handle->addr)
 330                 RETURN(NULL);
 331
 332         //spin_lock(&ldlm_handle_lock);
 333         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
 334         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock)) {
 335                 CERROR("bogus lock %p\n", lock);
 336                 GOTO(out2, retval);
 337         }
 338
 339         if (lock->l_random != handle->cookie) {
 340                 CERROR("bogus cookie: lock %p has "LPX64" vs. handle "LPX64"\n",
 341                        lock, lock->l_random, handle->cookie);
 342                 GOTO(out2, NULL);
 343         }
 344         if (!lock->l_resource) {
 345                 CERROR("trying to lock bogus resource: lock %p\n", lock);
 346                 LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 347                 GOTO(out2, retval);
 348         }
 349         if (!lock->l_resource->lr_namespace) {
 350                 CERROR("trying to lock bogus namespace: lock %p\n", lock);
 351                 LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 352                 GOTO(out2, retval);
 353         }
 354
 355         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 356         if (strict && lock->l_flags & LDLM_FL_DESTROYED) {
 357                 CERROR("lock already destroyed: lock %p\n", lock);
 358                 LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 359                 GOTO(out, NULL);
 360         }
 361
 362         retval = LDLM_LOCK_GET(lock);
 363         if (!retval)
 364                 CERROR("lock disappeared below us!!! %p\n", lock);
 365         EXIT;
 366  out:
 367         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 368  out2:
 369         //spin_unlock(&ldlm_handle_lock);
 370         return retval;
 371 }
 372
 373 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 374 {
 375         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 376 }
 377
 378 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 379 {
 380         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 381         desc->l_req_mode = lock->l_req_mode;
 382         desc->l_granted_mode = lock->l_granted_mode;
 383         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 384         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 385 }
 386
 387 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 388                                    struct ldlm_lock *new)
 389 {
 390         struct ldlm_ast_work *w;
 391         ENTRY;
 392
 393         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 394         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 395                 GOTO(out, 0);
 396
 397         OBD_ALLOC(w, sizeof(*w));
 398         if (!w) {
 399                 LBUG();
 400                 GOTO(out, 0);
 401         }
 402
 403         if (new) {
 404                 lock->l_flags |= LDLM_FL_AST_SENT;
 405                 w->w_blocking = 1;
 406                 ldlm_lock2desc(new, &w->w_desc);
 407         }
 408
 409         w->w_lock = LDLM_LOCK_GET(lock);
 410         list_add(&w->w_list, lock->l_resource->lr_tmp);
 411       out:
 412         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 413         return;
 414 }
 415
 416 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 417 {
 418         struct ldlm_lock *lock;
 419
 420         lock = ldlm_handle2lock(lockh);
 421         ldlm_lock_addref_internal(lock, mode);
 422         LDLM_LOCK_PUT(lock);
 423 }
 424
 425 /* only called for local locks */
 426 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 427 {
 428         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 429         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 430                 lock->l_readers++;
 431         else
 432                 lock->l_writers++;
 433         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 434         LDLM_LOCK_GET(lock);
 435         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 436 }
 437
 438 /* Args: unlocked lock */
 439 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 440 {
 441         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 442         ENTRY;
 443
 444         if (lock == NULL)
 445                 LBUG();
 446
 447         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 448         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 449         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 450                 lock->l_readers--;
 451         else
 452                 lock->l_writers--;
 453
 454         /* If we received a blocked AST and this was the last reference,
 455          * run the callback. */
 456         if (!lock->l_readers && !lock->l_writers &&
 457             (lock->l_flags & LDLM_FL_CBPENDING)) {
 458                 if (!lock->l_resource->lr_namespace->ns_client &&
 459                     lock->l_export)
 460                         CERROR("FL_CBPENDING set on non-local lock--just a "
 461                                "warning\n");
 462
 463                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 464                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 465
 466                 /* FIXME: need a real 'desc' here */
 467                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 468                                      lock->l_data_len, LDLM_CB_BLOCKING);
 469         } else
 470                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 471
 472         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 473         LDLM_LOCK_PUT(lock);    /* matches the handle2lock above */
 474
 475         EXIT;
 476 }
 477
 478 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 479                                  struct list_head *queue)
 480 {
 481         struct list_head *tmp, *pos;
 482         int rc = 1;
 483
 484         list_for_each_safe(tmp, pos, queue) {
 485                 struct ldlm_lock *child;
 486                 ldlm_res_compat compat;
 487
 488                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 489                 if (lock == child)
 490                         continue;
 491
 492                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 493                 if (compat && compat(child, lock)) {
 494                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 495                         continue;
 496                 }
 497                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 498                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 499                         continue;
 500                 }
 501
 502                 rc = 0;
 503
 504                 if (send_cbs && child->l_blocking_ast != NULL) {
 505                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 506                                "blocking AST.\n", child);
 507                         ldlm_add_ast_work_item(child, lock);
 508                 }
 509         }
 510
 511         return rc;
 512 }
 513
 514 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 515 {
 516         int rc;
 517         ENTRY;
 518
 519         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 520         rc = ldlm_lock_compat_list(lock, send_cbs,
 521                                    &lock->l_resource->lr_granted);
 522         /* FIXME: should we be sending ASTs to converting? */
 523         if (rc)
 524                 rc = ldlm_lock_compat_list
 525                         (lock, send_cbs, &lock->l_resource->lr_converting);
 526
 527         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 528         RETURN(rc);
 529 }
 530
 531 /* NOTE: called by
 532    - ldlm_handle_enqueuque - resource
 533 */
 534 void ldlm_grant_lock(struct ldlm_lock *lock)
 535 {
 536         struct ldlm_resource *res = lock->l_resource;
 537         ENTRY;
 538
 539         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 540         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 541         lock->l_granted_mode = lock->l_req_mode;
 542
 543         if (lock->l_granted_mode < res->lr_most_restr)
 544                 res->lr_most_restr = lock->l_granted_mode;
 545
 546         if (lock->l_completion_ast) {
 547                 ldlm_add_ast_work_item(lock, NULL);
 548         }
 549         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 550         EXIT;
 551 }
 552
 553 /* returns a referenced lock or NULL */
 554 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 555                                       struct ldlm_extent *extent)
 556 {
 557         struct ldlm_lock *lock;
 558         struct list_head *tmp;
 559
 560         list_for_each(tmp, queue) {
 561                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 562
 563                 if (lock->l_flags & (LDLM_FL_CBPENDING | LDLM_FL_DESTROYED))
 564                         continue;
 565
 566                 /* lock_convert() takes the resource lock, so we're sure that
 567                  * req_mode and lr_type won't change beneath us */
 568                 if (lock->l_req_mode != mode)
 569                         continue;
 570
 571                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 572                     (lock->l_extent.start > extent->start ||
 573                      lock->l_extent.end < extent->end))
 574                         continue;
 575
 576                 ldlm_lock_addref_internal(lock, mode);
 577                 return lock;
 578         }
 579
 580         return NULL;
 581 }
 582
 583 /* Must be called with no resource or lock locks held.
 584  *
 585  * Returns 1 if it finds an already-existing lock that is compatible; in this
 586  * case, lockh is filled in with a addref()ed lock
 587 */
 588 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 * res_id, __u32 type,
 589                     void *cookie, int cookielen, ldlm_mode_t mode,
 590                     struct lustre_handle *lockh)
 591 {
 592         struct ldlm_resource *res;
 593         struct ldlm_lock *lock;
 594         int rc = 0;
 595         ENTRY;
 596
 597         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 598         if (res == NULL)
 599                 RETURN(0);
 600
 601         ns = res->lr_namespace;
 602         l_lock(&ns->ns_lock);
 603
 604         if ((lock = search_queue(&res->lr_granted, mode, cookie)))
 605                 GOTO(out, rc = 1);
 606         if ((lock = search_queue(&res->lr_converting, mode, cookie)))
 607                 GOTO(out, rc = 1);
 608         if ((lock = search_queue(&res->lr_waiting, mode, cookie)))
 609                 GOTO(out, rc = 1);
 610
 611         EXIT;
 612       out:
 613         ldlm_resource_put(res);
 614         l_unlock(&ns->ns_lock);
 615
 616         if (lock) {
 617                 ldlm_lock2handle(lock, lockh);
 618                 if (lock->l_completion_ast)
 619                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 620         }
 621         if (rc)
 622                 LDLM_DEBUG(lock, "matched");
 623         else
 624                 LDLM_DEBUG_NOLOCK("not matched");
 625         return rc;
 626 }
 627
 628 /* Returns a referenced lock */
 629 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 630                                    struct lustre_handle *parent_lock_handle,
 631                                    __u64 * res_id, __u32 type,
 632                                    ldlm_mode_t mode, void *data, __u32 data_len)
 633 {
 634         struct ldlm_resource *res, *parent_res = NULL;
 635         struct ldlm_lock *lock, *parent_lock;
 636
 637         parent_lock = ldlm_handle2lock(parent_lock_handle);
 638         if (parent_lock)
 639                 parent_res = parent_lock->l_resource;
 640
 641         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 642         if (res == NULL)
 643                 RETURN(NULL);
 644
 645         lock = ldlm_lock_new(parent_lock, res);
 646         if (lock == NULL) {
 647                 ldlm_resource_put(res);
 648                 RETURN(NULL);
 649         }
 650
 651         lock->l_req_mode = mode;
 652         lock->l_data = data;
 653         lock->l_data_len = data_len;
 654
 655         return lock;
 656 }
 657
 658 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 659 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock * lock,
 660                                void *cookie, int cookie_len,
 661                                int *flags,
 662                                ldlm_completion_callback completion,
 663                                ldlm_blocking_callback blocking)
 664 {
 665         struct ldlm_resource *res;
 666         int local;
 667         ldlm_res_policy policy;
 668         ENTRY;
 669
 670         res = lock->l_resource;
 671         lock->l_blocking_ast = blocking;
 672
 673         if (res->lr_type == LDLM_EXTENT)
 674                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 675
 676         /* policies are not executed on the client or during replay */
 677         local = res->lr_namespace->ns_client;
 678         if (!local && !(*flags & LDLM_FL_REPLAY) &&
 679             (policy = ldlm_res_policy_table[res->lr_type])) {
 680                 int rc;
 681                 rc = policy(lock, cookie, lock->l_req_mode, *flags, NULL);
 682
 683                 if (rc == ELDLM_LOCK_CHANGED) {
 684                         res = lock->l_resource;
 685                         *flags |= LDLM_FL_LOCK_CHANGED;
 686                 } else if (rc == ELDLM_LOCK_ABORTED) {
 687                         ldlm_lock_destroy(lock);
 688                         RETURN(rc);
 689                 }
 690         }
 691
 692         l_lock(&res->lr_namespace->ns_lock);
 693         if (local && lock->l_req_mode == lock->l_granted_mode) {
 694                 /* The server returned a blocked lock, but it was granted before
 695                  * we got a chance to actually enqueue it.  We don't need to do
 696                  * anything else. */
 697                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 698                           LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 699                 GOTO(out, ELDLM_OK);
 700         }
 701
 702         /* This distinction between local lock trees is very important; a client
 703          * namespace only has information about locks taken by that client, and
 704          * thus doesn't have enough information to decide for itself if it can
 705          * be granted (below).  In this case, we do exactly what the server
 706          * tells us to do, as dictated by the 'flags'.
 707          *
 708          * We do exactly the same thing during recovery, when the server is
 709          * more or less trusting the clients not to lie.
 710          *
 711          * FIXME (bug 629283): Detect obvious lies by checking compatibility in
 712          * granted/converting queues. */
 713         ldlm_resource_unlink_lock(lock);
 714         if (local || (*flags & LDLM_FL_REPLAY)) {
 715                 if (*flags & LDLM_FL_BLOCK_CONV)
 716                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 717                                                lock);
 718                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 719                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 720                 else
 721                         ldlm_grant_lock(lock);
 722                 GOTO(out, ELDLM_OK);
 723         }
 724
 725         /* FIXME: We may want to optimize by checking lr_most_restr */
 726         if (!list_empty(&res->lr_converting)) {
 727                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 728                 *flags |= LDLM_FL_BLOCK_CONV;
 729                 GOTO(out, ELDLM_OK);
 730         }
 731         if (!list_empty(&res->lr_waiting)) {
 732                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 733                 *flags |= LDLM_FL_BLOCK_WAIT;
 734                 GOTO(out, ELDLM_OK);
 735         }
 736         if (!ldlm_lock_compat(lock, 0)) {
 737                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 738                 *flags |= LDLM_FL_BLOCK_GRANTED;
 739                 GOTO(out, ELDLM_OK);
 740         }
 741
 742         ldlm_grant_lock(lock);
 743         EXIT;
 744       out:
 745         l_unlock(&res->lr_namespace->ns_lock);
 746         /* Don't set 'completion_ast' until here so that if the lock is granted
 747          * immediately we don't do an unnecessary completion call. */
 748         lock->l_completion_ast = completion;
 749         return ELDLM_OK;
 750 }
 751
 752 /* Must be called with namespace taken: queue is waiting or converting. */
 753 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 754                                 struct list_head *queue)
 755 {
 756         struct list_head *tmp, *pos;
 757         ENTRY;
 758
 759         list_for_each_safe(tmp, pos, queue) {
 760                 struct ldlm_lock *pending;
 761                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 762
 763                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 764
 765                 if (!ldlm_lock_compat(pending, 1))
 766                         RETURN(1);
 767
 768                 list_del_init(&pending->l_res_link);
 769                 ldlm_grant_lock(pending);
 770         }
 771
 772         RETURN(0);
 773 }
 774
 775 void ldlm_run_ast_work(struct list_head *rpc_list)
 776 {
 777         struct list_head *tmp, *pos;
 778         int rc;
 779         ENTRY;
 780
 781         list_for_each_safe(tmp, pos, rpc_list) {
 782                 struct ldlm_ast_work *w =
 783                         list_entry(tmp, struct ldlm_ast_work, w_list);
 784
 785                 if (w->w_blocking)
 786                         rc = w->w_lock->l_blocking_ast
 787                                 (w->w_lock, &w->w_desc, w->w_data,
 788                                  w->w_datalen, LDLM_CB_BLOCKING);
 789                 else
 790                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags);
 791                 if (rc)
 792                         CERROR("Failed AST - should clean & disconnect "
 793                                "client\n");
 794                 LDLM_LOCK_PUT(w->w_lock);
 795                 list_del(&w->w_list);
 796                 OBD_FREE(w, sizeof(*w));
 797         }
 798         EXIT;
 799 }
 800
 801 /* Must be called with resource->lr_lock not taken. */
 802 void ldlm_reprocess_all(struct ldlm_resource *res)
 803 {
 804         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 805         ENTRY;
 806
 807         /* Local lock trees don't get reprocessed. */
 808         if (res->lr_namespace->ns_client) {
 809                 EXIT;
 810                 return;
 811         }
 812
 813         l_lock(&res->lr_namespace->ns_lock);
 814         res->lr_tmp = &rpc_list;
 815
 816         ldlm_reprocess_queue(res, &res->lr_converting);
 817         if (list_empty(&res->lr_converting))
 818                 ldlm_reprocess_queue(res, &res->lr_waiting);
 819
 820         res->lr_tmp = NULL;
 821         l_unlock(&res->lr_namespace->ns_lock);
 822
 823         ldlm_run_ast_work(&rpc_list);
 824         EXIT;
 825 }
 826
 827 void ldlm_cancel_callback(struct ldlm_lock *lock)
 828 {
 829         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 830         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
 831                 lock->l_flags |= LDLM_FL_CANCEL;
 832                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 833                                      lock->l_data_len, LDLM_CB_CANCELING);
 834         }
 835         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 836 }
 837
 838 void ldlm_lock_cancel(struct ldlm_lock *lock)
 839 {
 840         struct ldlm_resource *res;
 841         struct ldlm_namespace *ns;
 842         ENTRY;
 843
 844         res = lock->l_resource;
 845         ns = res->lr_namespace;
 846
 847         l_lock(&ns->ns_lock);
 848         if (lock->l_readers || lock->l_writers)
 849                 CDEBUG(D_INFO, "lock still has references (%d readers, %d "
 850                        "writers)\n", lock->l_readers, lock->l_writers);
 851
 852         ldlm_cancel_callback(lock);
 853
 854         ldlm_del_waiting_lock(lock);
 855         ldlm_resource_unlink_lock(lock);
 856         ldlm_lock_destroy(lock);
 857         l_unlock(&ns->ns_lock);
 858         EXIT;
 859 }
 860
 861 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data, int datalen)
 862 {
 863         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 864         ENTRY;
 865
 866         if (lock == NULL)
 867                 RETURN(-EINVAL);
 868
 869         lock->l_data = data;
 870         lock->l_data_len = datalen;
 871
 872         LDLM_LOCK_PUT(lock);
 873
 874         RETURN(0);
 875 }
 876
 877 void ldlm_cancel_locks_for_export(struct obd_export *exp)
 878 {
 879         struct list_head *iter, *n; /* MUST BE CALLED "n"! */
 880
 881         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
 882                 struct ldlm_lock *lock;
 883                 struct ldlm_resource *res;
 884                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 885                 res = ldlm_resource_getref(lock->l_resource);
 886                 LDLM_DEBUG(lock, "export %p", exp);
 887                 ldlm_lock_cancel(lock);
 888                 ldlm_reprocess_all(res);
 889                 ldlm_resource_put(res);
 890         }
 891 }
 892
 893 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 894                                         int *flags)
 895 {
 896         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 897         struct ldlm_resource *res;
 898         struct ldlm_namespace *ns;
 899         int granted = 0;
 900         ENTRY;
 901
 902         res = lock->l_resource;
 903         ns = res->lr_namespace;
 904
 905         l_lock(&ns->ns_lock);
 906
 907         lock->l_req_mode = new_mode;
 908         ldlm_resource_unlink_lock(lock);
 909
 910         /* If this is a local resource, put it on the appropriate list. */
 911         if (res->lr_namespace->ns_client) {
 912                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED))
 913                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 914                                                lock);
 915                 else {
 916                         /* This should never happen, because of the way the
 917                          * server handles conversions. */
 918                         LBUG();
 919
 920                         res->lr_tmp = &rpc_list;
 921                         ldlm_grant_lock(lock);
 922                         res->lr_tmp = NULL;
 923                         granted = 1;
 924                         /* FIXME: completion handling not with ns_lock held ! */
 925                         if (lock->l_completion_ast)
 926                                 lock->l_completion_ast(lock, 0);
 927                 }
 928         } else {
 929                 /* FIXME: We should try the conversion right away and possibly
 930                  * return success without the need for an extra AST */
 931                 ldlm_resource_add_lock(res, res->lr_converting.prev, lock);
 932                 *flags |= LDLM_FL_BLOCK_CONV;
 933         }
 934
 935         l_unlock(&ns->ns_lock);
 936
 937         if (granted)
 938                 ldlm_run_ast_work(&rpc_list);
 939         RETURN(res);
 940 }
 941
 942 void ldlm_lock_dump(struct ldlm_lock *lock)
 943 {
 944         char ver[128];
 945
 946         if (!(portal_debug & D_OTHER))
 947                 return;
 948
 949         if (RES_VERSION_SIZE != 4)
 950                 LBUG();
 951
 952         if (!lock) {
 953                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
 954                 return;
 955         }
 956
 957         snprintf(ver, sizeof(ver), "%x %x %x %x",
 958                  lock->l_version[0], lock->l_version[1],
 959                  lock->l_version[2], lock->l_version[3]);
 960
 961         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
 962         if (lock->l_export && lock->l_export->exp_connection)
 963                 CDEBUG(D_OTHER, "  Node: NID %x (rhandle: "LPX64")\n",
 964                        lock->l_export->exp_connection->c_peer.peer_nid,
 965                        lock->l_remote_handle.addr);
 966         else
 967                 CDEBUG(D_OTHER, "  Node: local\n");
 968         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
 969         CDEBUG(D_OTHER, "  Resource: %p ("LPD64")\n", lock->l_resource,
 970                lock->l_resource->lr_name[0]);
 971         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
 972                (int)lock->l_req_mode, (int)lock->l_granted_mode);
 973         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
 974                lock->l_readers, lock->l_writers);
 975         if (lock->l_resource->lr_type == LDLM_EXTENT)
 976                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
 977                        (unsigned long long)lock->l_extent.start,
 978                        (unsigned long long)lock->l_extent.end);
 979 }