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