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