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                 memset(lock, 0x5a, sizeof(*lock));
 182                 kmem_cache_free(ldlm_lock_slab, lock);
 183                 //spin_unlock(&ldlm_handle_lock);
 184                 CDEBUG(D_MALLOC, "kfreed 'lock': %d at %p (tot 0).\n",
 185                        sizeof(*lock), lock);
 186         } else
 187                 l_unlock(&ns->ns_lock);
 188
 189         EXIT;
 190 }
 191
 192 void ldlm_lock_destroy(struct ldlm_lock *lock)
 193 {
 194         ENTRY;
 195         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 196
 197         if (!list_empty(&lock->l_children)) {
 198                 LDLM_DEBUG(lock, "still has children (%p)!",
 199                            lock->l_children.next);
 200                 ldlm_lock_dump(lock);
 201                 LBUG();
 202         }
 203         if (lock->l_readers || lock->l_writers) {
 204                 LDLM_DEBUG(lock, "lock still has references");
 205                 ldlm_lock_dump(lock);
 206         }
 207
 208         if (!list_empty(&lock->l_res_link)) {
 209                 ldlm_lock_dump(lock);
 210                 LBUG();
 211         }
 212
 213         if (lock->l_flags & LDLM_FL_DESTROYED) {
 214                 LASSERT(list_empty(&lock->l_lru));
 215                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 216                 EXIT;
 217                 return;
 218         }
 219
 220         list_del_init(&lock->l_lru);
 221         list_del(&lock->l_export_chain);
 222         lock->l_export = NULL;
 223         lock->l_flags |= LDLM_FL_DESTROYED;
 224
 225         /* Wake anyone waiting for this lock */
 226         /* FIXME: I should probably add yet another flag, instead of using
 227          * l_export to only call this on clients */
 228         if (lock->l_export && lock->l_completion_ast)
 229                 lock->l_completion_ast(lock, 0);
 230
 231         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 232         LDLM_LOCK_PUT(lock);
 233         EXIT;
 234 }
 235
 236 /*
 237    usage: pass in a resource on which you have done get
 238           pass in a parent lock on which you have done a get
 239           do not put the resource or the parent
 240    returns: lock with refcount 1
 241 */
 242 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 243                                        struct ldlm_resource *resource)
 244 {
 245         struct ldlm_lock *lock;
 246         ENTRY;
 247
 248         if (resource == NULL)
 249                 LBUG();
 250
 251         lock = kmem_cache_alloc(ldlm_lock_slab, SLAB_KERNEL);
 252         if (lock == NULL)
 253                 RETURN(NULL);
 254
 255         memset(lock, 0, sizeof(*lock));
 256         get_random_bytes(&lock->l_random, sizeof(__u64));
 257
 258         lock->l_resource = resource;
 259         /* this refcount matches the one of the resource passed
 260            in which is not being put away */
 261         lock->l_refc = 1;
 262         INIT_LIST_HEAD(&lock->l_children);
 263         INIT_LIST_HEAD(&lock->l_res_link);
 264         INIT_LIST_HEAD(&lock->l_lru);
 265         INIT_LIST_HEAD(&lock->l_export_chain);
 266         INIT_LIST_HEAD(&lock->l_pending_chain);
 267         init_waitqueue_head(&lock->l_waitq);
 268
 269         spin_lock(&resource->lr_namespace->ns_counter_lock);
 270         resource->lr_namespace->ns_locks++;
 271         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 272
 273         if (parent != NULL) {
 274                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 275                 lock->l_parent = parent;
 276                 list_add(&lock->l_childof, &parent->l_children);
 277                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 278         }
 279
 280         CDEBUG(D_MALLOC, "kmalloced 'lock': %d at "
 281                "%p (tot %d).\n", sizeof(*lock), lock, 1);
 282         /* this is the extra refcount, to prevent the lock from evaporating */
 283         LDLM_LOCK_GET(lock);
 284         RETURN(lock);
 285 }
 286
 287 int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3])
 288 {
 289         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 290         struct ldlm_resource *oldres = lock->l_resource;
 291         int i;
 292         ENTRY;
 293
 294         l_lock(&ns->ns_lock);
 295         if (memcmp(new_resid, lock->l_resource->lr_name,
 296                    sizeof(lock->l_resource->lr_name)) == 0) {
 297                 /* Nothing to do */
 298                 l_unlock(&ns->ns_lock);
 299                 RETURN(0);
 300         }
 301
 302         LASSERT(new_resid[0] != 0);
 303
 304         /* This function assumes that the lock isn't on any lists */
 305         LASSERT(list_empty(&lock->l_res_link));
 306
 307         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
 308                                              lock->l_resource->lr_type, 1);
 309         if (lock->l_resource == NULL) {
 310                 LBUG();
 311                 RETURN(-ENOMEM);
 312         }
 313
 314         /* move references over */
 315         for (i = 0; i < lock->l_refc; i++) {
 316                 int rc;
 317                 ldlm_resource_getref(lock->l_resource);
 318                 rc = ldlm_resource_put(oldres);
 319                 if (rc == 1 && i != lock->l_refc - 1)
 320                         LBUG();
 321         }
 322         /* compensate for the initial get above.. */
 323         ldlm_resource_put(lock->l_resource);
 324
 325         l_unlock(&ns->ns_lock);
 326         RETURN(0);
 327 }
 328
 329 /*
 330  *  HANDLES
 331  */
 332
 333 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 334 {
 335         lockh->addr = (__u64) (unsigned long)lock;
 336         lockh->cookie = lock->l_random;
 337 }
 338
 339 /*
 340  * if flags: atomically get the lock and set the flags.
 341  * Return NULL if flag already set
 342  */
 343
 344 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle,
 345                                      int strict, int flags)
 346 {
 347         struct ldlm_lock *lock = NULL, *retval = NULL;
 348         ENTRY;
 349
 350         LASSERT(handle);
 351
 352         if (!handle->addr)
 353                 RETURN(NULL);
 354
 355         //spin_lock(&ldlm_handle_lock);
 356         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
 357         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock)) {
 358                 //CERROR("bogus lock %p\n", lock);
 359                 GOTO(out2, retval);
 360         }
 361
 362         if (lock->l_random != handle->cookie) {
 363                 //CERROR("bogus cookie: lock %p has "LPX64" vs. handle "LPX64"\n",
 364                 //       lock, lock->l_random, handle->cookie);
 365                 GOTO(out2, NULL);
 366         }
 367         if (!lock->l_resource) {
 368                 CERROR("trying to lock bogus resource: lock %p\n", lock);
 369                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 370                 GOTO(out2, retval);
 371         }
 372         if (!lock->l_resource->lr_namespace) {
 373                 CERROR("trying to lock bogus namespace: lock %p\n", lock);
 374                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 375                 GOTO(out2, retval);
 376         }
 377
 378         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 379         if (strict && lock->l_flags & LDLM_FL_DESTROYED) {
 380                 CERROR("lock already destroyed: lock %p\n", lock);
 381                 //LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 382                 GOTO(out, NULL);
 383         }
 384
 385         if (flags && (lock->l_flags & flags))
 386                 GOTO(out, NULL);
 387
 388         if (flags)
 389                 lock->l_flags |= flags;
 390
 391         retval = LDLM_LOCK_GET(lock);
 392         if (!retval)
 393                 CERROR("lock disappeared below us!!! %p\n", lock);
 394         EXIT;
 395  out:
 396         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 397  out2:
 398         //spin_unlock(&ldlm_handle_lock);
 399         return retval;
 400 }
 401
 402 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 403 {
 404         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 405 }
 406
 407 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 408 {
 409         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 410         desc->l_req_mode = lock->l_req_mode;
 411         desc->l_granted_mode = lock->l_granted_mode;
 412         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 413         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 414 }
 415
 416 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 417                                    struct ldlm_lock *new)
 418 {
 419         struct ldlm_ast_work *w;
 420         ENTRY;
 421
 422         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 423         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 424                 GOTO(out, 0);
 425
 426         OBD_ALLOC(w, sizeof(*w));
 427         if (!w) {
 428                 LBUG();
 429                 GOTO(out, 0);
 430         }
 431
 432         if (new) {
 433                 lock->l_flags |= LDLM_FL_AST_SENT;
 434                 w->w_blocking = 1;
 435                 ldlm_lock2desc(new, &w->w_desc);
 436         }
 437
 438         w->w_lock = LDLM_LOCK_GET(lock);
 439         list_add(&w->w_list, lock->l_resource->lr_tmp);
 440       out:
 441         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 442         return;
 443 }
 444
 445 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 446 {
 447         struct ldlm_lock *lock;
 448
 449         lock = ldlm_handle2lock(lockh);
 450         ldlm_lock_addref_internal(lock, mode);
 451         LDLM_LOCK_PUT(lock);
 452 }
 453
 454 /* only called for local locks */
 455 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 456 {
 457         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 458
 459         if (!list_empty(&lock->l_lru)) {
 460                 list_del_init(&lock->l_lru);
 461                 lock->l_resource->lr_namespace->ns_nr_unused--;
 462                 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
 463         }
 464
 465         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 466                 lock->l_readers++;
 467         else
 468                 lock->l_writers++;
 469         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 470         LDLM_LOCK_GET(lock);
 471         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 472 }
 473
 474 /* Args: unlocked lock */
 475 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 476                                     __u64 *res_id, int flags);
 477
 478 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 479 {
 480         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0, 0);
 481         struct ldlm_namespace *ns;
 482         ENTRY;
 483
 484         if (lock == NULL)
 485                 LBUG();
 486
 487         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 488         ns = lock->l_resource->lr_namespace;
 489         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 490         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 491                 lock->l_readers--;
 492         else
 493                 lock->l_writers--;
 494
 495         /* If we received a blocked AST and this was the last reference,
 496          * run the callback. */
 497         if (!lock->l_readers && !lock->l_writers &&
 498             (lock->l_flags & LDLM_FL_CBPENDING)) {
 499                 if (!lock->l_resource->lr_namespace->ns_client &&
 500                     lock->l_export)
 501                         CERROR("FL_CBPENDING set on non-local lock--just a "
 502                                "warning\n");
 503
 504                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 505                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 506
 507                 /* FIXME: need a real 'desc' here */
 508                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 509                                      lock->l_data_len, LDLM_CB_BLOCKING);
 510         } else if (ns->ns_client && !lock->l_readers && !lock->l_writers) {
 511                 LASSERT(list_empty(&lock->l_lru));
 512                 LASSERT(ns->ns_nr_unused >= 0);
 513                 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 514                 ns->ns_nr_unused++;
 515                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 516                 ldlm_cancel_lru(ns);
 517         } else
 518                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 519
 520         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 521         LDLM_LOCK_PUT(lock);    /* matches the handle2lock above */
 522
 523         EXIT;
 524 }
 525
 526 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 527                                  struct list_head *queue)
 528 {
 529         struct list_head *tmp, *pos;
 530         int rc = 1;
 531
 532         list_for_each_safe(tmp, pos, queue) {
 533                 struct ldlm_lock *child;
 534                 ldlm_res_compat compat;
 535
 536                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 537                 if (lock == child)
 538                         continue;
 539
 540                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 541                 if (compat && compat(child, lock)) {
 542                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 543                         continue;
 544                 }
 545                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 546                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 547                         continue;
 548                 }
 549
 550                 rc = 0;
 551
 552                 if (send_cbs && child->l_blocking_ast != NULL) {
 553                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 554                                "blocking AST.\n", child);
 555                         ldlm_add_ast_work_item(child, lock);
 556                 }
 557         }
 558
 559         return rc;
 560 }
 561
 562 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 563 {
 564         int rc;
 565         ENTRY;
 566
 567         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 568         rc = ldlm_lock_compat_list(lock, send_cbs,
 569                                    &lock->l_resource->lr_granted);
 570         /* FIXME: should we be sending ASTs to converting? */
 571         if (rc)
 572                 rc = ldlm_lock_compat_list
 573                         (lock, send_cbs, &lock->l_resource->lr_converting);
 574
 575         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 576         RETURN(rc);
 577 }
 578
 579 /* NOTE: called by
 580    - ldlm_handle_enqueuque - resource
 581 */
 582 void ldlm_grant_lock(struct ldlm_lock *lock)
 583 {
 584         struct ldlm_resource *res = lock->l_resource;
 585         ENTRY;
 586
 587         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 588         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 589         lock->l_granted_mode = lock->l_req_mode;
 590
 591         if (lock->l_granted_mode < res->lr_most_restr)
 592                 res->lr_most_restr = lock->l_granted_mode;
 593
 594         if (lock->l_completion_ast) {
 595                 ldlm_add_ast_work_item(lock, NULL);
 596         }
 597         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 598         EXIT;
 599 }
 600
 601 /* returns a referenced lock or NULL */
 602 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 603                                       struct ldlm_extent *extent,
 604                                       struct ldlm_lock *old_lock)
 605 {
 606         struct ldlm_lock *lock;
 607         struct list_head *tmp;
 608
 609         list_for_each(tmp, queue) {
 610                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 611
 612                 if (lock == old_lock)
 613                         continue;
 614
 615                 if (lock->l_flags & (LDLM_FL_CBPENDING | LDLM_FL_DESTROYED))
 616                         continue;
 617
 618                 if (lock->l_req_mode != mode)
 619                         continue;
 620
 621                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 622                     (lock->l_extent.start > extent->start ||
 623                      lock->l_extent.end < extent->end))
 624                         continue;
 625
 626                 ldlm_lock_addref_internal(lock, mode);
 627                 return lock;
 628         }
 629
 630         return NULL;
 631 }
 632
 633 /* Can be called in two ways:
 634  *
 635  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
 636  * for a duplicate of.
 637  *
 638  * Otherwise, all of the fields must be filled in, to match against.
 639  *
 640  * Returns 1 if it finds an already-existing lock that is compatible; in this
 641  * case, lockh is filled in with a addref()ed lock
 642  */
 643 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 *res_id, __u32 type,
 644                     void *cookie, int cookielen, ldlm_mode_t mode,
 645                     struct lustre_handle *lockh)
 646 {
 647         struct ldlm_resource *res;
 648         struct ldlm_lock *lock, *old_lock = NULL;
 649         int rc = 0;
 650         ENTRY;
 651
 652         if (ns == NULL) {
 653                 old_lock = ldlm_handle2lock(lockh);
 654                 LASSERT(old_lock);
 655
 656                 ns = old_lock->l_resource->lr_namespace;
 657                 res_id = old_lock->l_resource->lr_name;
 658                 type = old_lock->l_resource->lr_type;
 659                 mode = old_lock->l_req_mode;
 660         }
 661
 662         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 663         if (res == NULL) {
 664                 LASSERT(old_lock == NULL);
 665                 RETURN(0);
 666         }
 667
 668         l_lock(&ns->ns_lock);
 669
 670         if ((lock = search_queue(&res->lr_granted, mode, cookie, old_lock)))
 671                 GOTO(out, rc = 1);
 672         if ((lock = search_queue(&res->lr_converting, mode, cookie, old_lock)))
 673                 GOTO(out, rc = 1);
 674         if ((lock = search_queue(&res->lr_waiting, mode, cookie, old_lock)))
 675                 GOTO(out, rc = 1);
 676
 677         EXIT;
 678        out:
 679         ldlm_resource_put(res);
 680         l_unlock(&ns->ns_lock);
 681
 682         if (lock) {
 683                 ldlm_lock2handle(lock, lockh);
 684                 if (lock->l_completion_ast)
 685                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 686         }
 687         if (rc)
 688                 LDLM_DEBUG(lock, "matched");
 689         else
 690                 LDLM_DEBUG_NOLOCK("not matched");
 691
 692         if (old_lock)
 693                 LDLM_LOCK_PUT(old_lock);
 694
 695         return rc;
 696 }
 697
 698 /* Returns a referenced lock */
 699 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 700                                    struct lustre_handle *parent_lock_handle,
 701                                    __u64 * res_id, __u32 type,
 702                                    ldlm_mode_t mode, void *data, __u32 data_len)
 703 {
 704         struct ldlm_resource *res, *parent_res = NULL;
 705         struct ldlm_lock *lock, *parent_lock = NULL;
 706
 707         if (parent_lock_handle) {
 708                 parent_lock = ldlm_handle2lock(parent_lock_handle);
 709                 if (parent_lock)
 710                         parent_res = parent_lock->l_resource;
 711         }
 712
 713         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 714         if (res == NULL)
 715                 RETURN(NULL);
 716
 717         lock = ldlm_lock_new(parent_lock, res);
 718         if (lock == NULL) {
 719                 ldlm_resource_put(res);
 720                 RETURN(NULL);
 721         }
 722
 723         lock->l_req_mode = mode;
 724         lock->l_data = data;
 725         lock->l_data_len = data_len;
 726
 727         return lock;
 728 }
 729
 730 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 731 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock * lock,
 732                                void *cookie, int cookie_len,
 733                                int *flags,
 734                                ldlm_completion_callback completion,
 735                                ldlm_blocking_callback blocking)
 736 {
 737         struct ldlm_resource *res;
 738         int local;
 739         ldlm_res_policy policy;
 740         ENTRY;
 741
 742         res = lock->l_resource;
 743         lock->l_blocking_ast = blocking;
 744
 745         if (res->lr_type == LDLM_EXTENT)
 746                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 747
 748         /* policies are not executed on the client or during replay */
 749         local = res->lr_namespace->ns_client;
 750         if (!local && !(*flags & LDLM_FL_REPLAY) &&
 751             (policy = ldlm_res_policy_table[res->lr_type])) {
 752                 int rc;
 753                 rc = policy(lock, cookie, lock->l_req_mode, *flags, NULL);
 754
 755                 if (rc == ELDLM_LOCK_CHANGED) {
 756                         res = lock->l_resource;
 757                         *flags |= LDLM_FL_LOCK_CHANGED;
 758                 } else if (rc == ELDLM_LOCK_ABORTED) {
 759                         ldlm_lock_destroy(lock);
 760                         RETURN(rc);
 761                 }
 762         }
 763
 764         l_lock(&res->lr_namespace->ns_lock);
 765         if (local && lock->l_req_mode == lock->l_granted_mode) {
 766                 /* The server returned a blocked lock, but it was granted before
 767                  * we got a chance to actually enqueue it.  We don't need to do
 768                  * anything else. */
 769                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 770                           LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 771                 GOTO(out, ELDLM_OK);
 772         }
 773
 774         /* This distinction between local lock trees is very important; a client
 775          * namespace only has information about locks taken by that client, and
 776          * thus doesn't have enough information to decide for itself if it can
 777          * be granted (below).  In this case, we do exactly what the server
 778          * tells us to do, as dictated by the 'flags'.
 779          *
 780          * We do exactly the same thing during recovery, when the server is
 781          * more or less trusting the clients not to lie.
 782          *
 783          * FIXME (bug 629283): Detect obvious lies by checking compatibility in
 784          * granted/converting queues. */
 785         ldlm_resource_unlink_lock(lock);
 786         if (local || (*flags & LDLM_FL_REPLAY)) {
 787                 if (*flags & LDLM_FL_BLOCK_CONV)
 788                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 789                                                lock);
 790                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 791                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 792                 else
 793                         ldlm_grant_lock(lock);
 794                 GOTO(out, ELDLM_OK);
 795         }
 796
 797         /* FIXME: We may want to optimize by checking lr_most_restr */
 798         if (!list_empty(&res->lr_converting)) {
 799                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 800                 *flags |= LDLM_FL_BLOCK_CONV;
 801                 GOTO(out, ELDLM_OK);
 802         }
 803         if (!list_empty(&res->lr_waiting)) {
 804                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 805                 *flags |= LDLM_FL_BLOCK_WAIT;
 806                 GOTO(out, ELDLM_OK);
 807         }
 808         if (!ldlm_lock_compat(lock, 0)) {
 809                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 810                 *flags |= LDLM_FL_BLOCK_GRANTED;
 811                 GOTO(out, ELDLM_OK);
 812         }
 813
 814         ldlm_grant_lock(lock);
 815         EXIT;
 816       out:
 817         l_unlock(&res->lr_namespace->ns_lock);
 818         /* Don't set 'completion_ast' until here so that if the lock is granted
 819          * immediately we don't do an unnecessary completion call. */
 820         lock->l_completion_ast = completion;
 821         return ELDLM_OK;
 822 }
 823
 824 /* Must be called with namespace taken: queue is waiting or converting. */
 825 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 826                                 struct list_head *queue)
 827 {
 828         struct list_head *tmp, *pos;
 829         ENTRY;
 830
 831         list_for_each_safe(tmp, pos, queue) {
 832                 struct ldlm_lock *pending;
 833                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 834
 835                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 836
 837                 if (!ldlm_lock_compat(pending, 1))
 838                         RETURN(1);
 839
 840                 list_del_init(&pending->l_res_link);
 841                 ldlm_grant_lock(pending);
 842         }
 843
 844         RETURN(0);
 845 }
 846
 847 void ldlm_run_ast_work(struct list_head *rpc_list)
 848 {
 849         struct list_head *tmp, *pos;
 850         int rc;
 851         ENTRY;
 852
 853         list_for_each_safe(tmp, pos, rpc_list) {
 854                 struct ldlm_ast_work *w =
 855                         list_entry(tmp, struct ldlm_ast_work, w_list);
 856
 857                 if (w->w_blocking)
 858                         rc = w->w_lock->l_blocking_ast
 859                                 (w->w_lock, &w->w_desc, w->w_data,
 860                                  w->w_datalen, LDLM_CB_BLOCKING);
 861                 else
 862                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags);
 863                 if (rc)
 864                         CERROR("Failed AST - should clean & disconnect "
 865                                "client\n");
 866                 LDLM_LOCK_PUT(w->w_lock);
 867                 list_del(&w->w_list);
 868                 OBD_FREE(w, sizeof(*w));
 869         }
 870         EXIT;
 871 }
 872
 873 /* Must be called with resource->lr_lock not taken. */
 874 void ldlm_reprocess_all(struct ldlm_resource *res)
 875 {
 876         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 877         ENTRY;
 878
 879         /* Local lock trees don't get reprocessed. */
 880         if (res->lr_namespace->ns_client) {
 881                 EXIT;
 882                 return;
 883         }
 884
 885         l_lock(&res->lr_namespace->ns_lock);
 886         res->lr_tmp = &rpc_list;
 887
 888         ldlm_reprocess_queue(res, &res->lr_converting);
 889         if (list_empty(&res->lr_converting))
 890                 ldlm_reprocess_queue(res, &res->lr_waiting);
 891
 892         res->lr_tmp = NULL;
 893         l_unlock(&res->lr_namespace->ns_lock);
 894
 895         ldlm_run_ast_work(&rpc_list);
 896         EXIT;
 897 }
 898
 899 void ldlm_cancel_callback(struct ldlm_lock *lock)
 900 {
 901         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 902         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
 903                 lock->l_flags |= LDLM_FL_CANCEL;
 904                 if (lock->l_blocking_ast)
 905                         lock->l_blocking_ast(lock, NULL, lock->l_data,
 906                                              lock->l_data_len,
 907                                              LDLM_CB_CANCELING);
 908                 else
 909                         LDLM_DEBUG(lock, "no blocking ast");
 910         }
 911         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 912 }
 913
 914 void ldlm_lock_cancel(struct ldlm_lock *lock)
 915 {
 916         struct ldlm_resource *res;
 917         struct ldlm_namespace *ns;
 918         ENTRY;
 919
 920         res = lock->l_resource;
 921         ns = res->lr_namespace;
 922
 923         l_lock(&ns->ns_lock);
 924         if (lock->l_readers || lock->l_writers)
 925                 LDLM_DEBUG(lock, "lock still has references");
 926
 927         ldlm_cancel_callback(lock);
 928
 929         ldlm_del_waiting_lock(lock);
 930         ldlm_resource_unlink_lock(lock);
 931         ldlm_lock_destroy(lock);
 932         l_unlock(&ns->ns_lock);
 933         EXIT;
 934 }
 935
 936 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data, int datalen)
 937 {
 938         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 939         ENTRY;
 940
 941         if (lock == NULL)
 942                 RETURN(-EINVAL);
 943
 944         lock->l_data = data;
 945         lock->l_data_len = datalen;
 946
 947         LDLM_LOCK_PUT(lock);
 948
 949         RETURN(0);
 950 }
 951
 952 void ldlm_cancel_locks_for_export(struct obd_export *exp)
 953 {
 954         struct list_head *iter, *n; /* MUST BE CALLED "n"! */
 955
 956         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
 957                 struct ldlm_lock *lock;
 958                 struct ldlm_resource *res;
 959                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 960                 res = ldlm_resource_getref(lock->l_resource);
 961                 LDLM_DEBUG(lock, "export %p", exp);
 962                 ldlm_lock_cancel(lock);
 963                 ldlm_reprocess_all(res);
 964                 ldlm_resource_put(res);
 965         }
 966 }
 967
 968 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 969                                         int *flags)
 970 {
 971         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 972         struct ldlm_resource *res;
 973         struct ldlm_namespace *ns;
 974         int granted = 0;
 975         ENTRY;
 976
 977         res = lock->l_resource;
 978         ns = res->lr_namespace;
 979
 980         l_lock(&ns->ns_lock);
 981
 982         lock->l_req_mode = new_mode;
 983         ldlm_resource_unlink_lock(lock);
 984
 985         /* If this is a local resource, put it on the appropriate list. */
 986         if (res->lr_namespace->ns_client) {
 987                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED))
 988                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 989                                                lock);
 990                 else {
 991                         /* This should never happen, because of the way the
 992                          * server handles conversions. */
 993                         LBUG();
 994
 995                         res->lr_tmp = &rpc_list;
 996                         ldlm_grant_lock(lock);
 997                         res->lr_tmp = NULL;
 998                         granted = 1;
 999                         /* FIXME: completion handling not with ns_lock held ! */
1000                         if (lock->l_completion_ast)
1001                                 lock->l_completion_ast(lock, 0);
1002                 }
1003         } else {
1004                 /* FIXME: We should try the conversion right away and possibly
1005                  * return success without the need for an extra AST */
1006                 ldlm_resource_add_lock(res, res->lr_converting.prev, lock);
1007                 *flags |= LDLM_FL_BLOCK_CONV;
1008         }
1009
1010         l_unlock(&ns->ns_lock);
1011
1012         if (granted)
1013                 ldlm_run_ast_work(&rpc_list);
1014         RETURN(res);
1015 }
1016
1017 void ldlm_lock_dump(struct ldlm_lock *lock)
1018 {
1019         char ver[128];
1020
1021         if (!(portal_debug & D_OTHER))
1022                 return;
1023
1024         if (RES_VERSION_SIZE != 4)
1025                 LBUG();
1026
1027         if (!lock) {
1028                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
1029                 return;
1030         }
1031
1032         snprintf(ver, sizeof(ver), "%x %x %x %x",
1033                  lock->l_version[0], lock->l_version[1],
1034                  lock->l_version[2], lock->l_version[3]);
1035
1036         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
1037         if (lock->l_export && lock->l_export->exp_connection)
1038                 CDEBUG(D_OTHER, "  Node: NID %x (rhandle: "LPX64")\n",
1039                        lock->l_export->exp_connection->c_peer.peer_nid,
1040                        lock->l_remote_handle.addr);
1041         else
1042                 CDEBUG(D_OTHER, "  Node: local\n");
1043         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
1044         CDEBUG(D_OTHER, "  Resource: %p ("LPD64")\n", lock->l_resource,
1045                lock->l_resource->lr_name[0]);
1046         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
1047                (int)lock->l_req_mode, (int)lock->l_granted_mode);
1048         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
1049                lock->l_readers, lock->l_writers);
1050         if (lock->l_resource->lr_type == LDLM_EXTENT)
1051                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
1052                        (unsigned long long)lock->l_extent.start,
1053                        (unsigned long long)lock->l_extent.end);
1054 }