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 };
  49
  50 char *ldlm_it2str(int it)
  51 {
  52         switch (it) {
  53         case IT_OPEN:
  54                 return "open";
  55         case IT_CREAT:
  56                 return "creat";
  57         case (IT_OPEN | IT_CREAT):
  58                 return "open|creat";
  59         case IT_MKDIR:
  60                 return "mkdir";
  61         case IT_LINK:
  62                 return "link";
  63         case IT_LINK2:
  64                 return "link2";
  65         case IT_SYMLINK:
  66                 return "symlink";
  67         case IT_UNLINK:
  68                 return "unlink";
  69         case IT_RMDIR:
  70                 return "rmdir";
  71         case IT_RENAME:
  72                 return "rename";
  73         case IT_RENAME2:
  74                 return "rename2";
  75         case IT_READDIR:
  76                 return "readdir";
  77         case IT_GETATTR:
  78                 return "getattr";
  79         case IT_SETATTR:
  80                 return "setattr";
  81         case IT_READLINK:
  82                 return "readlink";
  83         case IT_MKNOD:
  84                 return "mknod";
  85         case IT_LOOKUP:
  86                 return "lookup";
  87         default:
  88                 CERROR("Unknown intent %d\n", it);
  89                 return "UNKNOWN";
  90         }
  91 }
  92
  93 extern kmem_cache_t *ldlm_lock_slab;
  94
  95 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b);
  96
  97 ldlm_res_compat ldlm_res_compat_table[] = {
  98         [LDLM_PLAIN] ldlm_plain_compat,
  99         [LDLM_EXTENT] ldlm_extent_compat,
 100 };
 101
 102 static ldlm_res_policy ldlm_intent_policy_func;
 103
 104 static int ldlm_plain_policy(struct ldlm_lock *lock, void *req_cookie,
 105                              ldlm_mode_t mode, int flags, void *data)
 106 {
 107         if ((flags & LDLM_FL_HAS_INTENT) && ldlm_intent_policy_func) {
 108                 return ldlm_intent_policy_func(lock, req_cookie, mode, flags,
 109                                                data);
 110         }
 111
 112         return ELDLM_OK;
 113 }
 114
 115 ldlm_res_policy ldlm_res_policy_table[] = {
 116         [LDLM_PLAIN] ldlm_plain_policy,
 117         [LDLM_EXTENT] ldlm_extent_policy,
 118 };
 119
 120 void ldlm_register_intent(ldlm_res_policy arg)
 121 {
 122         ldlm_intent_policy_func = arg;
 123 }
 124
 125 void ldlm_unregister_intent(void)
 126 {
 127         ldlm_intent_policy_func = NULL;
 128 }
 129
 130 /*
 131  * REFCOUNTED LOCK OBJECTS
 132  */
 133
 134
 135 /*
 136  * Lock refcounts, during creation:
 137  *   - one special one for allocation, dec'd only once in destroy
 138  *   - one for being a lock that's in-use
 139  *   - one for the addref associated with a new lock
 140  */
 141 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 142 {
 143         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 144         lock->l_refc++;
 145         ldlm_resource_getref(lock->l_resource);
 146         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 147         return lock;
 148 }
 149
 150 void ldlm_lock_put(struct ldlm_lock *lock)
 151 {
 152         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 153         ENTRY;
 154
 155         l_lock(&ns->ns_lock);
 156         lock->l_refc--;
 157         //LDLM_DEBUG(lock, "after refc--");
 158         if (lock->l_refc < 0)
 159                 LBUG();
 160
 161         if (ldlm_resource_put(lock->l_resource)) {
 162                 LASSERT(lock->l_refc == 0);
 163                 lock->l_resource = NULL;
 164         }
 165         if (lock->l_parent)
 166                 LDLM_LOCK_PUT(lock->l_parent);
 167
 168         if (lock->l_refc == 0 && (lock->l_flags & LDLM_FL_DESTROYED)) {
 169                 l_unlock(&ns->ns_lock);
 170                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
 171
 172                 //spin_lock(&ldlm_handle_lock);
 173                 spin_lock(&ns->ns_counter_lock);
 174                 ns->ns_locks--;
 175                 spin_unlock(&ns->ns_counter_lock);
 176
 177                 lock->l_resource = NULL;
 178                 lock->l_random = DEAD_HANDLE_MAGIC;
 179                 if (lock->l_export && lock->l_export->exp_connection)
 180                         ptlrpc_put_connection(lock->l_export->exp_connection);
 181                 kmem_cache_free(ldlm_lock_slab, lock);
 182                 //spin_unlock(&ldlm_handle_lock);
 183                 CDEBUG(D_MALLOC, "kfreed 'lock': %d at %p (tot 0).\n",
 184                        sizeof(*lock), lock);
 185         } else
 186                 l_unlock(&ns->ns_lock);
 187
 188         EXIT;
 189 }
 190
 191 void ldlm_lock_destroy(struct ldlm_lock *lock)
 192 {
 193         ENTRY;
 194         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 195
 196         if (!list_empty(&lock->l_children)) {
 197                 LDLM_DEBUG(lock, "still has children (%p)!",
 198                            lock->l_children.next);
 199                 ldlm_lock_dump(lock);
 200                 LBUG();
 201         }
 202         if (lock->l_readers || lock->l_writers) {
 203                 LDLM_DEBUG(lock, "lock still has references");
 204                 ldlm_lock_dump(lock);
 205         }
 206
 207         if (!list_empty(&lock->l_res_link)) {
 208                 ldlm_lock_dump(lock);
 209                 LBUG();
 210         }
 211
 212         if (lock->l_flags & LDLM_FL_DESTROYED) {
 213                 LASSERT(list_empty(&lock->l_lru));
 214                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 215                 EXIT;
 216                 return;
 217         }
 218
 219         list_del_init(&lock->l_lru);
 220         list_del(&lock->l_export_chain);
 221         lock->l_export = NULL;
 222         lock->l_flags |= LDLM_FL_DESTROYED;
 223
 224         /* Wake anyone waiting for this lock */
 225         /* FIXME: I should probably add yet another flag, instead of using
 226          * l_export to only call this on clients */
 227         if (lock->l_export && lock->l_completion_ast)
 228                 lock->l_completion_ast(lock, 0);
 229
 230         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 231         LDLM_LOCK_PUT(lock);
 232         EXIT;
 233 }
 234
 235 /*
 236    usage: pass in a resource on which you have done get
 237           pass in a parent lock on which you have done a get
 238           do not put the resource or the parent
 239    returns: lock with refcount 1
 240 */
 241 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 242                                        struct ldlm_resource *resource)
 243 {
 244         struct ldlm_lock *lock;
 245         ENTRY;
 246
 247         if (resource == NULL)
 248                 LBUG();
 249
 250         lock = kmem_cache_alloc(ldlm_lock_slab, SLAB_KERNEL);
 251         if (lock == NULL)
 252                 RETURN(NULL);
 253
 254         memset(lock, 0, sizeof(*lock));
 255         get_random_bytes(&lock->l_random, sizeof(__u64));
 256
 257         lock->l_resource = resource;
 258         /* this refcount matches the one of the resource passed
 259            in which is not being put away */
 260         lock->l_refc = 1;
 261         INIT_LIST_HEAD(&lock->l_children);
 262         INIT_LIST_HEAD(&lock->l_res_link);
 263         INIT_LIST_HEAD(&lock->l_lru);
 264         INIT_LIST_HEAD(&lock->l_export_chain);
 265         INIT_LIST_HEAD(&lock->l_pending_chain);
 266         init_waitqueue_head(&lock->l_waitq);
 267
 268         spin_lock(&resource->lr_namespace->ns_counter_lock);
 269         resource->lr_namespace->ns_locks++;
 270         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 271
 272         if (parent != NULL) {
 273                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 274                 lock->l_parent = parent;
 275                 list_add(&lock->l_childof, &parent->l_children);
 276                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 277         }
 278
 279         CDEBUG(D_MALLOC, "kmalloced 'lock': %d at "
 280                "%p (tot %d).\n", sizeof(*lock), lock, 1);
 281         /* this is the extra refcount, to prevent the lock from evaporating */
 282         LDLM_LOCK_GET(lock);
 283         RETURN(lock);
 284 }
 285
 286 int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3])
 287 {
 288         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 289         struct ldlm_resource *oldres = lock->l_resource;
 290         int i;
 291         ENTRY;
 292
 293         l_lock(&ns->ns_lock);
 294         if (memcmp(new_resid, lock->l_resource->lr_name,
 295                    sizeof(lock->l_resource->lr_name)) == 0) {
 296                 /* Nothing to do */
 297                 l_unlock(&ns->ns_lock);
 298                 RETURN(0);
 299         }
 300
 301         LASSERT(new_resid[0] != 0);
 302
 303         /* This function assumes that the lock isn't on any lists */
 304         LASSERT(list_empty(&lock->l_res_link));
 305
 306         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
 307                                              lock->l_resource->lr_type, 1);
 308         if (lock->l_resource == NULL) {
 309                 LBUG();
 310                 RETURN(-ENOMEM);
 311         }
 312
 313         /* move references over */
 314         for (i = 0; i < lock->l_refc; i++) {
 315                 int rc;
 316                 ldlm_resource_getref(lock->l_resource);
 317                 rc = ldlm_resource_put(oldres);
 318                 if (rc == 1 && i != lock->l_refc - 1)
 319                         LBUG();
 320         }
 321         /* compensate for the initial get above.. */
 322         ldlm_resource_put(lock->l_resource);
 323
 324         l_unlock(&ns->ns_lock);
 325         RETURN(0);
 326 }
 327
 328 /*
 329  *  HANDLES
 330  */
 331
 332 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 333 {
 334         lockh->addr = (__u64) (unsigned long)lock;
 335         lockh->cookie = lock->l_random;
 336 }
 337
 338 /*
 339  * if flags: atomically get the lock and set the flags.
 340  * Return NULL if flag already set
 341  */
 342
 343 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle,
 344                                      int strict, int flags)
 345 {
 346         struct ldlm_lock *lock = NULL, *retval = NULL;
 347         ENTRY;
 348
 349         if (!handle || !handle->addr)
 350                 RETURN(NULL);
 351
 352         //spin_lock(&ldlm_handle_lock);
 353         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
 354         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock)) {
 355                 //CERROR("bogus lock %p\n", lock);
 356                 GOTO(out2, retval);
 357         }
 358
 359         if (lock->l_random != handle->cookie) {
 360                 //CERROR("bogus cookie: lock %p has "LPX64" vs. handle "LPX64"\n",
 361                 //       lock, lock->l_random, handle->cookie);
 362                 GOTO(out2, NULL);
 363         }
 364         if (!lock->l_resource) {
 365                 CERROR("trying to lock bogus resource: lock %p\n", lock);
 366                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 367                 GOTO(out2, retval);
 368         }
 369         if (!lock->l_resource->lr_namespace) {
 370                 CERROR("trying to lock bogus namespace: lock %p\n", lock);
 371                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 372                 GOTO(out2, retval);
 373         }
 374
 375         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 376         if (strict && lock->l_flags & LDLM_FL_DESTROYED) {
 377                 CERROR("lock already destroyed: lock %p\n", lock);
 378                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 379                 GOTO(out, NULL);
 380         }
 381
 382         if (flags && (lock->l_flags & flags))
 383                 GOTO(out, NULL);
 384
 385         if (flags)
 386                 lock->l_flags |= flags;
 387
 388         retval = LDLM_LOCK_GET(lock);
 389         if (!retval)
 390                 CERROR("lock disappeared below us!!! %p\n", lock);
 391         EXIT;
 392  out:
 393         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 394  out2:
 395         //spin_unlock(&ldlm_handle_lock);
 396         return retval;
 397 }
 398
 399 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 400 {
 401         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 402 }
 403
 404 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 405 {
 406         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 407         desc->l_req_mode = lock->l_req_mode;
 408         desc->l_granted_mode = lock->l_granted_mode;
 409         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 410         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 411 }
 412
 413 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 414                                    struct ldlm_lock *new)
 415 {
 416         struct ldlm_ast_work *w;
 417         ENTRY;
 418
 419         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 420         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 421                 GOTO(out, 0);
 422
 423         OBD_ALLOC(w, sizeof(*w));
 424         if (!w) {
 425                 LBUG();
 426                 GOTO(out, 0);
 427         }
 428
 429         if (new) {
 430                 lock->l_flags |= LDLM_FL_AST_SENT;
 431                 w->w_blocking = 1;
 432                 ldlm_lock2desc(new, &w->w_desc);
 433         }
 434
 435         w->w_lock = LDLM_LOCK_GET(lock);
 436         list_add(&w->w_list, lock->l_resource->lr_tmp);
 437       out:
 438         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 439         return;
 440 }
 441
 442 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 443 {
 444         struct ldlm_lock *lock;
 445
 446         lock = ldlm_handle2lock(lockh);
 447         ldlm_lock_addref_internal(lock, mode);
 448         LDLM_LOCK_PUT(lock);
 449 }
 450
 451 /* only called for local locks */
 452 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 453 {
 454         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 455
 456         if (!list_empty(&lock->l_lru)) {
 457                 list_del_init(&lock->l_lru);
 458                 lock->l_resource->lr_namespace->ns_nr_unused--;
 459                 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
 460         }
 461
 462         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 463                 lock->l_readers++;
 464         else
 465                 lock->l_writers++;
 466         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 467         LDLM_LOCK_GET(lock);
 468         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 469 }
 470
 471 /* Args: unlocked lock */
 472 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 473                                     __u64 *res_id, int flags);
 474
 475 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 476 {
 477         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0, 0);
 478         struct ldlm_namespace *ns;
 479         ENTRY;
 480
 481         if (lock == NULL)
 482                 LBUG();
 483
 484         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 485         ns = lock->l_resource->lr_namespace;
 486         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 487         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 488                 lock->l_readers--;
 489         else
 490                 lock->l_writers--;
 491
 492         /* If we received a blocked AST and this was the last reference,
 493          * run the callback. */
 494         if (!lock->l_readers && !lock->l_writers &&
 495             (lock->l_flags & LDLM_FL_CBPENDING)) {
 496                 if (!lock->l_resource->lr_namespace->ns_client &&
 497                     lock->l_export)
 498                         CERROR("FL_CBPENDING set on non-local lock--just a "
 499                                "warning\n");
 500
 501                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 502                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 503
 504                 /* FIXME: need a real 'desc' here */
 505                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 506                                      lock->l_data_len, LDLM_CB_BLOCKING);
 507         } else if (ns->ns_client && !lock->l_readers && !lock->l_writers) {
 508                 LASSERT(list_empty(&lock->l_lru));
 509                 LASSERT(ns->ns_nr_unused >= 0);
 510                 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 511                 ns->ns_nr_unused++;
 512                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 513                 ldlm_cancel_lru(ns);
 514         } else
 515                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 516
 517         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 518         LDLM_LOCK_PUT(lock);    /* matches the handle2lock above */
 519
 520         EXIT;
 521 }
 522
 523 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 524                                  struct list_head *queue)
 525 {
 526         struct list_head *tmp, *pos;
 527         int rc = 1;
 528
 529         list_for_each_safe(tmp, pos, queue) {
 530                 struct ldlm_lock *child;
 531                 ldlm_res_compat compat;
 532
 533                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 534                 if (lock == child)
 535                         continue;
 536
 537                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 538                 if (compat && compat(child, lock)) {
 539                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 540                         continue;
 541                 }
 542                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 543                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 544                         continue;
 545                 }
 546
 547                 rc = 0;
 548
 549                 if (send_cbs && child->l_blocking_ast != NULL) {
 550                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 551                                "blocking AST.\n", child);
 552                         ldlm_add_ast_work_item(child, lock);
 553                 }
 554         }
 555
 556         return rc;
 557 }
 558
 559 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 560 {
 561         int rc;
 562         ENTRY;
 563
 564         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 565         rc = ldlm_lock_compat_list(lock, send_cbs,
 566                                    &lock->l_resource->lr_granted);
 567         /* FIXME: should we be sending ASTs to converting? */
 568         if (rc)
 569                 rc = ldlm_lock_compat_list
 570                         (lock, send_cbs, &lock->l_resource->lr_converting);
 571
 572         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 573         RETURN(rc);
 574 }
 575
 576 /* NOTE: called by
 577    - ldlm_handle_enqueuque - resource
 578 */
 579 void ldlm_grant_lock(struct ldlm_lock *lock)
 580 {
 581         struct ldlm_resource *res = lock->l_resource;
 582         ENTRY;
 583
 584         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 585         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 586         lock->l_granted_mode = lock->l_req_mode;
 587
 588         if (lock->l_granted_mode < res->lr_most_restr)
 589                 res->lr_most_restr = lock->l_granted_mode;
 590
 591         if (lock->l_completion_ast) {
 592                 ldlm_add_ast_work_item(lock, NULL);
 593         }
 594         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 595         EXIT;
 596 }
 597
 598 /* returns a referenced lock or NULL */
 599 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 600                                       struct ldlm_extent *extent,
 601                                       struct ldlm_lock *old_lock)
 602 {
 603         struct ldlm_lock *lock;
 604         struct list_head *tmp;
 605
 606         list_for_each(tmp, queue) {
 607                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 608
 609                 if (lock == old_lock)
 610                         continue;
 611
 612                 if (lock->l_flags & (LDLM_FL_CBPENDING | LDLM_FL_DESTROYED))
 613                         continue;
 614
 615                 if (lock->l_req_mode != mode)
 616                         continue;
 617
 618                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 619                     (lock->l_extent.start > extent->start ||
 620                      lock->l_extent.end < extent->end))
 621                         continue;
 622
 623                 ldlm_lock_addref_internal(lock, mode);
 624                 return lock;
 625         }
 626
 627         return NULL;
 628 }
 629
 630 /* Can be called in two ways:
 631  *
 632  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
 633  * for a duplicate of.
 634  *
 635  * Otherwise, all of the fields must be filled in, to match against.
 636  *
 637  * Returns 1 if it finds an already-existing lock that is compatible; in this
 638  * case, lockh is filled in with a addref()ed lock
 639  */
 640 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 *res_id, __u32 type,
 641                     void *cookie, int cookielen, ldlm_mode_t mode,
 642                     struct lustre_handle *lockh)
 643 {
 644         struct ldlm_resource *res;
 645         struct ldlm_lock *lock, *old_lock = NULL;
 646         int rc = 0;
 647         ENTRY;
 648
 649         if (ns == NULL) {
 650                 old_lock = ldlm_handle2lock(lockh);
 651                 LASSERT(old_lock);
 652
 653                 ns = old_lock->l_resource->lr_namespace;
 654                 res_id = old_lock->l_resource->lr_name;
 655                 type = old_lock->l_resource->lr_type;
 656                 mode = old_lock->l_req_mode;
 657         }
 658
 659         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 660         if (res == NULL) {
 661                 LASSERT(old_lock == NULL);
 662                 RETURN(0);
 663         }
 664
 665         l_lock(&ns->ns_lock);
 666
 667         if ((lock = search_queue(&res->lr_granted, mode, cookie, old_lock)))
 668                 GOTO(out, rc = 1);
 669         if ((lock = search_queue(&res->lr_converting, mode, cookie, old_lock)))
 670                 GOTO(out, rc = 1);
 671         if ((lock = search_queue(&res->lr_waiting, mode, cookie, old_lock)))
 672                 GOTO(out, rc = 1);
 673
 674         EXIT;
 675        out:
 676         ldlm_resource_put(res);
 677         l_unlock(&ns->ns_lock);
 678
 679         if (lock) {
 680                 ldlm_lock2handle(lock, lockh);
 681                 if (lock->l_completion_ast)
 682                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 683         }
 684         if (rc)
 685                 LDLM_DEBUG(lock, "matched");
 686         else
 687                 LDLM_DEBUG_NOLOCK("not matched");
 688
 689         if (old_lock)
 690                 LDLM_LOCK_PUT(old_lock);
 691
 692         return rc;
 693 }
 694
 695 /* Returns a referenced lock */
 696 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 697                                    struct lustre_handle *parent_lock_handle,
 698                                    __u64 * res_id, __u32 type,
 699                                    ldlm_mode_t mode, void *data, __u32 data_len)
 700 {
 701         struct ldlm_resource *res, *parent_res = NULL;
 702         struct ldlm_lock *lock, *parent_lock;
 703
 704         parent_lock = ldlm_handle2lock(parent_lock_handle);
 705         if (parent_lock)
 706                 parent_res = parent_lock->l_resource;
 707
 708         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 709         if (res == NULL)
 710                 RETURN(NULL);
 711
 712         lock = ldlm_lock_new(parent_lock, res);
 713         if (lock == NULL) {
 714                 ldlm_resource_put(res);
 715                 RETURN(NULL);
 716         }
 717
 718         lock->l_req_mode = mode;
 719         lock->l_data = data;
 720         lock->l_data_len = data_len;
 721
 722         return lock;
 723 }
 724
 725 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 726 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock * lock,
 727                                void *cookie, int cookie_len,
 728                                int *flags,
 729                                ldlm_completion_callback completion,
 730                                ldlm_blocking_callback blocking)
 731 {
 732         struct ldlm_resource *res;
 733         int local;
 734         ldlm_res_policy policy;
 735         ENTRY;
 736
 737         res = lock->l_resource;
 738         lock->l_blocking_ast = blocking;
 739
 740         if (res->lr_type == LDLM_EXTENT)
 741                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 742
 743         /* policies are not executed on the client or during replay */
 744         local = res->lr_namespace->ns_client;
 745         if (!local && !(*flags & LDLM_FL_REPLAY) &&
 746             (policy = ldlm_res_policy_table[res->lr_type])) {
 747                 int rc;
 748                 rc = policy(lock, cookie, lock->l_req_mode, *flags, NULL);
 749
 750                 if (rc == ELDLM_LOCK_CHANGED) {
 751                         res = lock->l_resource;
 752                         *flags |= LDLM_FL_LOCK_CHANGED;
 753                 } else if (rc == ELDLM_LOCK_ABORTED) {
 754                         ldlm_lock_destroy(lock);
 755                         RETURN(rc);
 756                 }
 757         }
 758
 759         l_lock(&res->lr_namespace->ns_lock);
 760         if (local && lock->l_req_mode == lock->l_granted_mode) {
 761                 /* The server returned a blocked lock, but it was granted before
 762                  * we got a chance to actually enqueue it.  We don't need to do
 763                  * anything else. */
 764                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 765                           LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 766                 GOTO(out, ELDLM_OK);
 767         }
 768
 769         /* This distinction between local lock trees is very important; a client
 770          * namespace only has information about locks taken by that client, and
 771          * thus doesn't have enough information to decide for itself if it can
 772          * be granted (below).  In this case, we do exactly what the server
 773          * tells us to do, as dictated by the 'flags'.
 774          *
 775          * We do exactly the same thing during recovery, when the server is
 776          * more or less trusting the clients not to lie.
 777          *
 778          * FIXME (bug 629283): Detect obvious lies by checking compatibility in
 779          * granted/converting queues. */
 780         ldlm_resource_unlink_lock(lock);
 781         if (local || (*flags & LDLM_FL_REPLAY)) {
 782                 if (*flags & LDLM_FL_BLOCK_CONV)
 783                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 784                                                lock);
 785                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 786                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 787                 else
 788                         ldlm_grant_lock(lock);
 789                 GOTO(out, ELDLM_OK);
 790         }
 791
 792         /* FIXME: We may want to optimize by checking lr_most_restr */
 793         if (!list_empty(&res->lr_converting)) {
 794                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 795                 *flags |= LDLM_FL_BLOCK_CONV;
 796                 GOTO(out, ELDLM_OK);
 797         }
 798         if (!list_empty(&res->lr_waiting)) {
 799                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 800                 *flags |= LDLM_FL_BLOCK_WAIT;
 801                 GOTO(out, ELDLM_OK);
 802         }
 803         if (!ldlm_lock_compat(lock, 0)) {
 804                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 805                 *flags |= LDLM_FL_BLOCK_GRANTED;
 806                 GOTO(out, ELDLM_OK);
 807         }
 808
 809         ldlm_grant_lock(lock);
 810         EXIT;
 811       out:
 812         l_unlock(&res->lr_namespace->ns_lock);
 813         /* Don't set 'completion_ast' until here so that if the lock is granted
 814          * immediately we don't do an unnecessary completion call. */
 815         lock->l_completion_ast = completion;
 816         return ELDLM_OK;
 817 }
 818
 819 /* Must be called with namespace taken: queue is waiting or converting. */
 820 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 821                                 struct list_head *queue)
 822 {
 823         struct list_head *tmp, *pos;
 824         ENTRY;
 825
 826         list_for_each_safe(tmp, pos, queue) {
 827                 struct ldlm_lock *pending;
 828                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 829
 830                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 831
 832                 if (!ldlm_lock_compat(pending, 1))
 833                         RETURN(1);
 834
 835                 list_del_init(&pending->l_res_link);
 836                 ldlm_grant_lock(pending);
 837         }
 838
 839         RETURN(0);
 840 }
 841
 842 void ldlm_run_ast_work(struct list_head *rpc_list)
 843 {
 844         struct list_head *tmp, *pos;
 845         int rc;
 846         ENTRY;
 847
 848         list_for_each_safe(tmp, pos, rpc_list) {
 849                 struct ldlm_ast_work *w =
 850                         list_entry(tmp, struct ldlm_ast_work, w_list);
 851
 852                 if (w->w_blocking)
 853                         rc = w->w_lock->l_blocking_ast
 854                                 (w->w_lock, &w->w_desc, w->w_data,
 855                                  w->w_datalen, LDLM_CB_BLOCKING);
 856                 else
 857                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags);
 858                 if (rc)
 859                         CERROR("Failed AST - should clean & disconnect "
 860                                "client\n");
 861                 LDLM_LOCK_PUT(w->w_lock);
 862                 list_del(&w->w_list);
 863                 OBD_FREE(w, sizeof(*w));
 864         }
 865         EXIT;
 866 }
 867
 868 /* Must be called with resource->lr_lock not taken. */
 869 void ldlm_reprocess_all(struct ldlm_resource *res)
 870 {
 871         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 872         ENTRY;
 873
 874         /* Local lock trees don't get reprocessed. */
 875         if (res->lr_namespace->ns_client) {
 876                 EXIT;
 877                 return;
 878         }
 879
 880         l_lock(&res->lr_namespace->ns_lock);
 881         res->lr_tmp = &rpc_list;
 882
 883         ldlm_reprocess_queue(res, &res->lr_converting);
 884         if (list_empty(&res->lr_converting))
 885                 ldlm_reprocess_queue(res, &res->lr_waiting);
 886
 887         res->lr_tmp = NULL;
 888         l_unlock(&res->lr_namespace->ns_lock);
 889
 890         ldlm_run_ast_work(&rpc_list);
 891         EXIT;
 892 }
 893
 894 void ldlm_cancel_callback(struct ldlm_lock *lock)
 895 {
 896         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 897         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
 898                 lock->l_flags |= LDLM_FL_CANCEL;
 899                 if (lock->l_blocking_ast)
 900                         lock->l_blocking_ast(lock, NULL, lock->l_data,
 901                                              lock->l_data_len,
 902                                              LDLM_CB_CANCELING);
 903                 else
 904                         LDLM_DEBUG(lock, "no blocking ast");
 905         }
 906         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 907 }
 908
 909 void ldlm_lock_cancel(struct ldlm_lock *lock)
 910 {
 911         struct ldlm_resource *res;
 912         struct ldlm_namespace *ns;
 913         ENTRY;
 914
 915         res = lock->l_resource;
 916         ns = res->lr_namespace;
 917
 918         l_lock(&ns->ns_lock);
 919         if (lock->l_readers || lock->l_writers)
 920                 LDLM_DEBUG(lock, "lock still has references");
 921
 922         ldlm_cancel_callback(lock);
 923
 924         ldlm_del_waiting_lock(lock);
 925         ldlm_resource_unlink_lock(lock);
 926         ldlm_lock_destroy(lock);
 927         l_unlock(&ns->ns_lock);
 928         EXIT;
 929 }
 930
 931 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data, int datalen)
 932 {
 933         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 934         ENTRY;
 935
 936         if (lock == NULL)
 937                 RETURN(-EINVAL);
 938
 939         lock->l_data = data;
 940         lock->l_data_len = datalen;
 941
 942         LDLM_LOCK_PUT(lock);
 943
 944         RETURN(0);
 945 }
 946
 947 void ldlm_cancel_locks_for_export(struct obd_export *exp)
 948 {
 949         struct list_head *iter, *n; /* MUST BE CALLED "n"! */
 950
 951         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
 952                 struct ldlm_lock *lock;
 953                 struct ldlm_resource *res;
 954                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 955                 res = ldlm_resource_getref(lock->l_resource);
 956                 LDLM_DEBUG(lock, "export %p", exp);
 957                 ldlm_lock_cancel(lock);
 958                 ldlm_reprocess_all(res);
 959                 ldlm_resource_put(res);
 960         }
 961 }
 962
 963 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 964                                         int *flags)
 965 {
 966         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 967         struct ldlm_resource *res;
 968         struct ldlm_namespace *ns;
 969         int granted = 0;
 970         ENTRY;
 971
 972         res = lock->l_resource;
 973         ns = res->lr_namespace;
 974
 975         l_lock(&ns->ns_lock);
 976
 977         lock->l_req_mode = new_mode;
 978         ldlm_resource_unlink_lock(lock);
 979
 980         /* If this is a local resource, put it on the appropriate list. */
 981         if (res->lr_namespace->ns_client) {
 982                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED))
 983                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 984                                                lock);
 985                 else {
 986                         /* This should never happen, because of the way the
 987                          * server handles conversions. */
 988                         LBUG();
 989
 990                         res->lr_tmp = &rpc_list;
 991                         ldlm_grant_lock(lock);
 992                         res->lr_tmp = NULL;
 993                         granted = 1;
 994                         /* FIXME: completion handling not with ns_lock held ! */
 995                         if (lock->l_completion_ast)
 996                                 lock->l_completion_ast(lock, 0);
 997                 }
 998         } else {
 999                 /* FIXME: We should try the conversion right away and possibly
1000                  * return success without the need for an extra AST */
1001                 ldlm_resource_add_lock(res, res->lr_converting.prev, lock);
1002                 *flags |= LDLM_FL_BLOCK_CONV;
1003         }
1004
1005         l_unlock(&ns->ns_lock);
1006
1007         if (granted)
1008                 ldlm_run_ast_work(&rpc_list);
1009         RETURN(res);
1010 }
1011
1012 void ldlm_lock_dump(struct ldlm_lock *lock)
1013 {
1014         char ver[128];
1015
1016         if (!(portal_debug & D_OTHER))
1017                 return;
1018
1019         if (RES_VERSION_SIZE != 4)
1020                 LBUG();
1021
1022         if (!lock) {
1023                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
1024                 return;
1025         }
1026
1027         snprintf(ver, sizeof(ver), "%x %x %x %x",
1028                  lock->l_version[0], lock->l_version[1],
1029                  lock->l_version[2], lock->l_version[3]);
1030
1031         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
1032         if (lock->l_export && lock->l_export->exp_connection)
1033                 CDEBUG(D_OTHER, "  Node: NID %x (rhandle: "LPX64")\n",
1034                        lock->l_export->exp_connection->c_peer.peer_nid,
1035                        lock->l_remote_handle.addr);
1036         else
1037                 CDEBUG(D_OTHER, "  Node: local\n");
1038         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
1039         CDEBUG(D_OTHER, "  Resource: %p ("LPD64")\n", lock->l_resource,
1040                lock->l_resource->lr_name[0]);
1041         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
1042                (int)lock->l_req_mode, (int)lock->l_granted_mode);
1043         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
1044                lock->l_readers, lock->l_writers);
1045         if (lock->l_resource->lr_type == LDLM_EXTENT)
1046                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
1047                        (unsigned long long)lock->l_extent.start,
1048                        (unsigned long long)lock->l_extent.end);
1049 }