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 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle,
 339                                      int strict)
 340 {
 341         struct ldlm_lock *lock = NULL, *retval = NULL;
 342         ENTRY;
 343
 344         if (!handle || !handle->addr)
 345                 RETURN(NULL);
 346
 347         //spin_lock(&ldlm_handle_lock);
 348         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
 349         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock)) {
 350                 //CERROR("bogus lock %p\n", lock);
 351                 GOTO(out2, retval);
 352         }
 353
 354         if (lock->l_random != handle->cookie) {
 355                 CERROR("bogus cookie: lock %p has "LPX64" vs. handle "LPX64"\n",
 356                        lock, lock->l_random, handle->cookie);
 357                 GOTO(out2, NULL);
 358         }
 359         if (!lock->l_resource) {
 360                 CERROR("trying to lock bogus resource: lock %p\n", lock);
 361                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 362                 GOTO(out2, retval);
 363         }
 364         if (!lock->l_resource->lr_namespace) {
 365                 CERROR("trying to lock bogus namespace: lock %p\n", lock);
 366                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 367                 GOTO(out2, retval);
 368         }
 369
 370         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 371         if (strict && lock->l_flags & LDLM_FL_DESTROYED) {
 372                 CERROR("lock already destroyed: lock %p\n", lock);
 373                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 374                 GOTO(out, NULL);
 375         }
 376
 377         retval = LDLM_LOCK_GET(lock);
 378         if (!retval)
 379                 CERROR("lock disappeared below us!!! %p\n", lock);
 380         EXIT;
 381  out:
 382         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 383  out2:
 384         //spin_unlock(&ldlm_handle_lock);
 385         return retval;
 386 }
 387
 388 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 389 {
 390         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 391 }
 392
 393 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 394 {
 395         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 396         desc->l_req_mode = lock->l_req_mode;
 397         desc->l_granted_mode = lock->l_granted_mode;
 398         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 399         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 400 }
 401
 402 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 403                                    struct ldlm_lock *new)
 404 {
 405         struct ldlm_ast_work *w;
 406         ENTRY;
 407
 408         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 409         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 410                 GOTO(out, 0);
 411
 412         OBD_ALLOC(w, sizeof(*w));
 413         if (!w) {
 414                 LBUG();
 415                 GOTO(out, 0);
 416         }
 417
 418         if (new) {
 419                 lock->l_flags |= LDLM_FL_AST_SENT;
 420                 w->w_blocking = 1;
 421                 ldlm_lock2desc(new, &w->w_desc);
 422         }
 423
 424         w->w_lock = LDLM_LOCK_GET(lock);
 425         list_add(&w->w_list, lock->l_resource->lr_tmp);
 426       out:
 427         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 428         return;
 429 }
 430
 431 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 432 {
 433         struct ldlm_lock *lock;
 434
 435         lock = ldlm_handle2lock(lockh);
 436         ldlm_lock_addref_internal(lock, mode);
 437         LDLM_LOCK_PUT(lock);
 438 }
 439
 440 /* only called for local locks */
 441 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 442 {
 443         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 444
 445         if (!list_empty(&lock->l_lru)) {
 446                 list_del_init(&lock->l_lru);
 447                 lock->l_resource->lr_namespace->ns_nr_unused--;
 448                 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
 449         }
 450
 451         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 452                 lock->l_readers++;
 453         else
 454                 lock->l_writers++;
 455         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 456         LDLM_LOCK_GET(lock);
 457         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 458 }
 459
 460 /* Args: unlocked lock */
 461 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 462                                     __u64 *res_id, int flags);
 463
 464 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 465 {
 466         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 467         struct ldlm_namespace *ns;
 468         ENTRY;
 469
 470         if (lock == NULL)
 471                 LBUG();
 472
 473         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 474         ns = lock->l_resource->lr_namespace;
 475         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 476         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 477                 lock->l_readers--;
 478         else
 479                 lock->l_writers--;
 480
 481         /* If we received a blocked AST and this was the last reference,
 482          * run the callback. */
 483         if (!lock->l_readers && !lock->l_writers &&
 484             (lock->l_flags & LDLM_FL_CBPENDING)) {
 485                 if (!lock->l_resource->lr_namespace->ns_client &&
 486                     lock->l_export)
 487                         CERROR("FL_CBPENDING set on non-local lock--just a "
 488                                "warning\n");
 489
 490                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 491                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 492
 493                 /* FIXME: need a real 'desc' here */
 494                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 495                                      lock->l_data_len, LDLM_CB_BLOCKING);
 496         } else if (ns->ns_client && !lock->l_readers && !lock->l_writers) {
 497                 LASSERT(list_empty(&lock->l_lru));
 498                 LASSERT(ns->ns_nr_unused >= 0);
 499                 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 500                 ns->ns_nr_unused++;
 501                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 502                 ldlm_cancel_lru(ns);
 503         } else
 504                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 505
 506         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 507         LDLM_LOCK_PUT(lock);    /* matches the handle2lock above */
 508
 509         EXIT;
 510 }
 511
 512 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 513                                  struct list_head *queue)
 514 {
 515         struct list_head *tmp, *pos;
 516         int rc = 1;
 517
 518         list_for_each_safe(tmp, pos, queue) {
 519                 struct ldlm_lock *child;
 520                 ldlm_res_compat compat;
 521
 522                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 523                 if (lock == child)
 524                         continue;
 525
 526                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 527                 if (compat && compat(child, lock)) {
 528                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 529                         continue;
 530                 }
 531                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 532                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 533                         continue;
 534                 }
 535
 536                 rc = 0;
 537
 538                 if (send_cbs && child->l_blocking_ast != NULL) {
 539                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 540                                "blocking AST.\n", child);
 541                         ldlm_add_ast_work_item(child, lock);
 542                 }
 543         }
 544
 545         return rc;
 546 }
 547
 548 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 549 {
 550         int rc;
 551         ENTRY;
 552
 553         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 554         rc = ldlm_lock_compat_list(lock, send_cbs,
 555                                    &lock->l_resource->lr_granted);
 556         /* FIXME: should we be sending ASTs to converting? */
 557         if (rc)
 558                 rc = ldlm_lock_compat_list
 559                         (lock, send_cbs, &lock->l_resource->lr_converting);
 560
 561         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 562         RETURN(rc);
 563 }
 564
 565 /* NOTE: called by
 566    - ldlm_handle_enqueuque - resource
 567 */
 568 void ldlm_grant_lock(struct ldlm_lock *lock)
 569 {
 570         struct ldlm_resource *res = lock->l_resource;
 571         ENTRY;
 572
 573         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 574         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 575         lock->l_granted_mode = lock->l_req_mode;
 576
 577         if (lock->l_granted_mode < res->lr_most_restr)
 578                 res->lr_most_restr = lock->l_granted_mode;
 579
 580         if (lock->l_completion_ast) {
 581                 ldlm_add_ast_work_item(lock, NULL);
 582         }
 583         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 584         EXIT;
 585 }
 586
 587 /* returns a referenced lock or NULL */
 588 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 589                                       struct ldlm_extent *extent,
 590                                       struct ldlm_lock *old_lock)
 591 {
 592         struct ldlm_lock *lock;
 593         struct list_head *tmp;
 594
 595         list_for_each(tmp, queue) {
 596                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 597
 598                 if (lock == old_lock)
 599                         continue;
 600
 601                 if (lock->l_flags & (LDLM_FL_CBPENDING | LDLM_FL_DESTROYED))
 602                         continue;
 603
 604                 if (lock->l_req_mode != mode)
 605                         continue;
 606
 607                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 608                     (lock->l_extent.start > extent->start ||
 609                      lock->l_extent.end < extent->end))
 610                         continue;
 611
 612                 ldlm_lock_addref_internal(lock, mode);
 613                 return lock;
 614         }
 615
 616         return NULL;
 617 }
 618
 619 /* Can be called in two ways:
 620  *
 621  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
 622  * for a duplicate of.
 623  *
 624  * Otherwise, all of the fields must be filled in, to match against.
 625  *
 626  * Returns 1 if it finds an already-existing lock that is compatible; in this
 627  * case, lockh is filled in with a addref()ed lock
 628  */
 629 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 *res_id, __u32 type,
 630                     void *cookie, int cookielen, ldlm_mode_t mode,
 631                     struct lustre_handle *lockh)
 632 {
 633         struct ldlm_resource *res;
 634         struct ldlm_lock *lock, *old_lock = NULL;
 635         int rc = 0;
 636         ENTRY;
 637
 638         if (ns == NULL) {
 639                 old_lock = ldlm_handle2lock(lockh);
 640                 LASSERT(old_lock);
 641
 642                 ns = old_lock->l_resource->lr_namespace;
 643                 res_id = old_lock->l_resource->lr_name;
 644                 type = old_lock->l_resource->lr_type;
 645                 mode = old_lock->l_req_mode;
 646         }
 647
 648         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 649         if (res == NULL) {
 650                 LASSERT(old_lock == NULL);
 651                 RETURN(0);
 652         }
 653
 654         l_lock(&ns->ns_lock);
 655
 656         if ((lock = search_queue(&res->lr_granted, mode, cookie, old_lock)))
 657                 GOTO(out, rc = 1);
 658         if ((lock = search_queue(&res->lr_converting, mode, cookie, old_lock)))
 659                 GOTO(out, rc = 1);
 660         if ((lock = search_queue(&res->lr_waiting, mode, cookie, old_lock)))
 661                 GOTO(out, rc = 1);
 662
 663         EXIT;
 664        out:
 665         ldlm_resource_put(res);
 666         l_unlock(&ns->ns_lock);
 667
 668         if (lock) {
 669                 ldlm_lock2handle(lock, lockh);
 670                 if (lock->l_completion_ast)
 671                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 672         }
 673         if (rc)
 674                 LDLM_DEBUG(lock, "matched");
 675         else
 676                 LDLM_DEBUG_NOLOCK("not matched");
 677
 678         if (old_lock)
 679                 LDLM_LOCK_PUT(old_lock);
 680
 681         return rc;
 682 }
 683
 684 /* Returns a referenced lock */
 685 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 686                                    struct lustre_handle *parent_lock_handle,
 687                                    __u64 * res_id, __u32 type,
 688                                    ldlm_mode_t mode, void *data, __u32 data_len)
 689 {
 690         struct ldlm_resource *res, *parent_res = NULL;
 691         struct ldlm_lock *lock, *parent_lock;
 692
 693         parent_lock = ldlm_handle2lock(parent_lock_handle);
 694         if (parent_lock)
 695                 parent_res = parent_lock->l_resource;
 696
 697         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 698         if (res == NULL)
 699                 RETURN(NULL);
 700
 701         lock = ldlm_lock_new(parent_lock, res);
 702         if (lock == NULL) {
 703                 ldlm_resource_put(res);
 704                 RETURN(NULL);
 705         }
 706
 707         lock->l_req_mode = mode;
 708         lock->l_data = data;
 709         lock->l_data_len = data_len;
 710
 711         return lock;
 712 }
 713
 714 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 715 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock * lock,
 716                                void *cookie, int cookie_len,
 717                                int *flags,
 718                                ldlm_completion_callback completion,
 719                                ldlm_blocking_callback blocking)
 720 {
 721         struct ldlm_resource *res;
 722         int local;
 723         ldlm_res_policy policy;
 724         ENTRY;
 725
 726         res = lock->l_resource;
 727         lock->l_blocking_ast = blocking;
 728
 729         if (res->lr_type == LDLM_EXTENT)
 730                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 731
 732         /* policies are not executed on the client or during replay */
 733         local = res->lr_namespace->ns_client;
 734         if (!local && !(*flags & LDLM_FL_REPLAY) &&
 735             (policy = ldlm_res_policy_table[res->lr_type])) {
 736                 int rc;
 737                 rc = policy(lock, cookie, lock->l_req_mode, *flags, NULL);
 738
 739                 if (rc == ELDLM_LOCK_CHANGED) {
 740                         res = lock->l_resource;
 741                         *flags |= LDLM_FL_LOCK_CHANGED;
 742                 } else if (rc == ELDLM_LOCK_ABORTED) {
 743                         ldlm_lock_destroy(lock);
 744                         RETURN(rc);
 745                 }
 746         }
 747
 748         l_lock(&res->lr_namespace->ns_lock);
 749         if (local && lock->l_req_mode == lock->l_granted_mode) {
 750                 /* The server returned a blocked lock, but it was granted before
 751                  * we got a chance to actually enqueue it.  We don't need to do
 752                  * anything else. */
 753                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 754                           LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 755                 GOTO(out, ELDLM_OK);
 756         }
 757
 758         /* This distinction between local lock trees is very important; a client
 759          * namespace only has information about locks taken by that client, and
 760          * thus doesn't have enough information to decide for itself if it can
 761          * be granted (below).  In this case, we do exactly what the server
 762          * tells us to do, as dictated by the 'flags'.
 763          *
 764          * We do exactly the same thing during recovery, when the server is
 765          * more or less trusting the clients not to lie.
 766          *
 767          * FIXME (bug 629283): Detect obvious lies by checking compatibility in
 768          * granted/converting queues. */
 769         ldlm_resource_unlink_lock(lock);
 770         if (local || (*flags & LDLM_FL_REPLAY)) {
 771                 if (*flags & LDLM_FL_BLOCK_CONV)
 772                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 773                                                lock);
 774                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 775                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 776                 else
 777                         ldlm_grant_lock(lock);
 778                 GOTO(out, ELDLM_OK);
 779         }
 780
 781         /* FIXME: We may want to optimize by checking lr_most_restr */
 782         if (!list_empty(&res->lr_converting)) {
 783                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 784                 *flags |= LDLM_FL_BLOCK_CONV;
 785                 GOTO(out, ELDLM_OK);
 786         }
 787         if (!list_empty(&res->lr_waiting)) {
 788                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 789                 *flags |= LDLM_FL_BLOCK_WAIT;
 790                 GOTO(out, ELDLM_OK);
 791         }
 792         if (!ldlm_lock_compat(lock, 0)) {
 793                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 794                 *flags |= LDLM_FL_BLOCK_GRANTED;
 795                 GOTO(out, ELDLM_OK);
 796         }
 797
 798         ldlm_grant_lock(lock);
 799         EXIT;
 800       out:
 801         l_unlock(&res->lr_namespace->ns_lock);
 802         /* Don't set 'completion_ast' until here so that if the lock is granted
 803          * immediately we don't do an unnecessary completion call. */
 804         lock->l_completion_ast = completion;
 805         return ELDLM_OK;
 806 }
 807
 808 /* Must be called with namespace taken: queue is waiting or converting. */
 809 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 810                                 struct list_head *queue)
 811 {
 812         struct list_head *tmp, *pos;
 813         ENTRY;
 814
 815         list_for_each_safe(tmp, pos, queue) {
 816                 struct ldlm_lock *pending;
 817                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 818
 819                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 820
 821                 if (!ldlm_lock_compat(pending, 1))
 822                         RETURN(1);
 823
 824                 list_del_init(&pending->l_res_link);
 825                 ldlm_grant_lock(pending);
 826         }
 827
 828         RETURN(0);
 829 }
 830
 831 void ldlm_run_ast_work(struct list_head *rpc_list)
 832 {
 833         struct list_head *tmp, *pos;
 834         int rc;
 835         ENTRY;
 836
 837         list_for_each_safe(tmp, pos, rpc_list) {
 838                 struct ldlm_ast_work *w =
 839                         list_entry(tmp, struct ldlm_ast_work, w_list);
 840
 841                 if (w->w_blocking)
 842                         rc = w->w_lock->l_blocking_ast
 843                                 (w->w_lock, &w->w_desc, w->w_data,
 844                                  w->w_datalen, LDLM_CB_BLOCKING);
 845                 else
 846                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags);
 847                 if (rc)
 848                         CERROR("Failed AST - should clean & disconnect "
 849                                "client\n");
 850                 LDLM_LOCK_PUT(w->w_lock);
 851                 list_del(&w->w_list);
 852                 OBD_FREE(w, sizeof(*w));
 853         }
 854         EXIT;
 855 }
 856
 857 /* Must be called with resource->lr_lock not taken. */
 858 void ldlm_reprocess_all(struct ldlm_resource *res)
 859 {
 860         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 861         ENTRY;
 862
 863         /* Local lock trees don't get reprocessed. */
 864         if (res->lr_namespace->ns_client) {
 865                 EXIT;
 866                 return;
 867         }
 868
 869         l_lock(&res->lr_namespace->ns_lock);
 870         res->lr_tmp = &rpc_list;
 871
 872         ldlm_reprocess_queue(res, &res->lr_converting);
 873         if (list_empty(&res->lr_converting))
 874                 ldlm_reprocess_queue(res, &res->lr_waiting);
 875
 876         res->lr_tmp = NULL;
 877         l_unlock(&res->lr_namespace->ns_lock);
 878
 879         ldlm_run_ast_work(&rpc_list);
 880         EXIT;
 881 }
 882
 883 void ldlm_cancel_callback(struct ldlm_lock *lock)
 884 {
 885         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 886         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
 887                 lock->l_flags |= LDLM_FL_CANCEL;
 888                 if (lock->l_blocking_ast)
 889                         lock->l_blocking_ast(lock, NULL, lock->l_data,
 890                                              lock->l_data_len,
 891                                              LDLM_CB_CANCELING);
 892                 else
 893                         LDLM_DEBUG(lock, "no blocking ast");
 894         }
 895         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 896 }
 897
 898 void ldlm_lock_cancel(struct ldlm_lock *lock)
 899 {
 900         struct ldlm_resource *res;
 901         struct ldlm_namespace *ns;
 902         ENTRY;
 903
 904         res = lock->l_resource;
 905         ns = res->lr_namespace;
 906
 907         l_lock(&ns->ns_lock);
 908         if (lock->l_readers || lock->l_writers)
 909                 LDLM_DEBUG(lock, "lock still has references");
 910
 911         ldlm_cancel_callback(lock);
 912
 913         ldlm_del_waiting_lock(lock);
 914         ldlm_resource_unlink_lock(lock);
 915         ldlm_lock_destroy(lock);
 916         l_unlock(&ns->ns_lock);
 917         EXIT;
 918 }
 919
 920 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data, int datalen)
 921 {
 922         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 923         ENTRY;
 924
 925         if (lock == NULL)
 926                 RETURN(-EINVAL);
 927
 928         lock->l_data = data;
 929         lock->l_data_len = datalen;
 930
 931         LDLM_LOCK_PUT(lock);
 932
 933         RETURN(0);
 934 }
 935
 936 void ldlm_cancel_locks_for_export(struct obd_export *exp)
 937 {
 938         struct list_head *iter, *n; /* MUST BE CALLED "n"! */
 939
 940         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
 941                 struct ldlm_lock *lock;
 942                 struct ldlm_resource *res;
 943                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 944                 res = ldlm_resource_getref(lock->l_resource);
 945                 LDLM_DEBUG(lock, "export %p", exp);
 946                 ldlm_lock_cancel(lock);
 947                 ldlm_reprocess_all(res);
 948                 ldlm_resource_put(res);
 949         }
 950 }
 951
 952 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 953                                         int *flags)
 954 {
 955         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 956         struct ldlm_resource *res;
 957         struct ldlm_namespace *ns;
 958         int granted = 0;
 959         ENTRY;
 960
 961         res = lock->l_resource;
 962         ns = res->lr_namespace;
 963
 964         l_lock(&ns->ns_lock);
 965
 966         lock->l_req_mode = new_mode;
 967         ldlm_resource_unlink_lock(lock);
 968
 969         /* If this is a local resource, put it on the appropriate list. */
 970         if (res->lr_namespace->ns_client) {
 971                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED))
 972                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 973                                                lock);
 974                 else {
 975                         /* This should never happen, because of the way the
 976                          * server handles conversions. */
 977                         LBUG();
 978
 979                         res->lr_tmp = &rpc_list;
 980                         ldlm_grant_lock(lock);
 981                         res->lr_tmp = NULL;
 982                         granted = 1;
 983                         /* FIXME: completion handling not with ns_lock held ! */
 984                         if (lock->l_completion_ast)
 985                                 lock->l_completion_ast(lock, 0);
 986                 }
 987         } else {
 988                 /* FIXME: We should try the conversion right away and possibly
 989                  * return success without the need for an extra AST */
 990                 ldlm_resource_add_lock(res, res->lr_converting.prev, lock);
 991                 *flags |= LDLM_FL_BLOCK_CONV;
 992         }
 993
 994         l_unlock(&ns->ns_lock);
 995
 996         if (granted)
 997                 ldlm_run_ast_work(&rpc_list);
 998         RETURN(res);
 999 }
1000
1001 void ldlm_lock_dump(struct ldlm_lock *lock)
1002 {
1003         char ver[128];
1004
1005         if (!(portal_debug & D_OTHER))
1006                 return;
1007
1008         if (RES_VERSION_SIZE != 4)
1009                 LBUG();
1010
1011         if (!lock) {
1012                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
1013                 return;
1014         }
1015
1016         snprintf(ver, sizeof(ver), "%x %x %x %x",
1017                  lock->l_version[0], lock->l_version[1],
1018                  lock->l_version[2], lock->l_version[3]);
1019
1020         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
1021         if (lock->l_export && lock->l_export->exp_connection)
1022                 CDEBUG(D_OTHER, "  Node: NID %x (rhandle: "LPX64")\n",
1023                        lock->l_export->exp_connection->c_peer.peer_nid,
1024                        lock->l_remote_handle.addr);
1025         else
1026                 CDEBUG(D_OTHER, "  Node: local\n");
1027         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
1028         CDEBUG(D_OTHER, "  Resource: %p ("LPD64")\n", lock->l_resource,
1029                lock->l_resource->lr_name[0]);
1030         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
1031                (int)lock->l_req_mode, (int)lock->l_granted_mode);
1032         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
1033                lock->l_readers, lock->l_writers);
1034         if (lock->l_resource->lr_type == LDLM_EXTENT)
1035                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
1036                        (unsigned long long)lock->l_extent.start,
1037                        (unsigned long long)lock->l_extent.end);
1038 }