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                 RETURN(NULL);
 314
 315         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
 316         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock))
 317                 RETURN(NULL);
 318
 319         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 320         if (lock->l_random != handle->cookie)
 321                 GOTO(out, NULL);
 322
 323         if (lock->l_flags & LDLM_FL_DESTROYED)
 324                 GOTO(out, NULL);
 325
 326         retval = LDLM_LOCK_GET(lock);
 327         EXIT;
 328       out:
 329         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 330         return retval;
 331 }
 332
 333 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 334 {
 335         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 336 }
 337
 338 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 339 {
 340         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 341         desc->l_req_mode = lock->l_req_mode;
 342         desc->l_granted_mode = lock->l_granted_mode;
 343         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 344         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 345 }
 346
 347 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 348                                    struct ldlm_lock *new)
 349 {
 350         struct ldlm_ast_work *w;
 351         ENTRY;
 352
 353         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 354         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 355                 GOTO(out, 0);
 356
 357         OBD_ALLOC(w, sizeof(*w));
 358         if (!w) {
 359                 LBUG();
 360                 GOTO(out, 0);
 361         }
 362
 363         if (new) {
 364                 lock->l_flags |= LDLM_FL_AST_SENT;
 365                 w->w_blocking = 1;
 366                 ldlm_lock2desc(new, &w->w_desc);
 367         }
 368
 369         w->w_lock = LDLM_LOCK_GET(lock);
 370         list_add(&w->w_list, lock->l_resource->lr_tmp);
 371       out:
 372         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 373         return;
 374 }
 375
 376 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 377 {
 378         struct ldlm_lock *lock;
 379
 380         lock = ldlm_handle2lock(lockh);
 381         ldlm_lock_addref_internal(lock, mode);
 382         LDLM_LOCK_PUT(lock);
 383 }
 384
 385 /* only called for local locks */
 386 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 387 {
 388         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 389         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 390                 lock->l_readers++;
 391         else
 392                 lock->l_writers++;
 393         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 394         LDLM_LOCK_GET(lock);
 395         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 396 }
 397
 398 /* Args: unlocked lock */
 399 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 400 {
 401         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 402         ENTRY;
 403
 404         if (lock == NULL)
 405                 LBUG();
 406
 407         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 408         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 409         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 410                 lock->l_readers--;
 411         else
 412                 lock->l_writers--;
 413
 414         /* If we received a blocked AST and this was the last reference,
 415          * run the callback. */
 416         if (!lock->l_readers && !lock->l_writers &&
 417             (lock->l_flags & LDLM_FL_CBPENDING)) {
 418                 if (!lock->l_resource->lr_namespace->ns_client &&
 419                     lock->l_export)
 420                         CERROR("FL_CBPENDING set on non-local lock--just a "
 421                                "warning\n");
 422
 423                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 424                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 425
 426                 /* FIXME: need a real 'desc' here */
 427                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 428                                      lock->l_data_len, LDLM_CB_BLOCKING);
 429         } else
 430                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 431
 432         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 433         LDLM_LOCK_PUT(lock);    /* matches the handle2lock above */
 434
 435         EXIT;
 436 }
 437
 438 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 439                                  struct list_head *queue)
 440 {
 441         struct list_head *tmp, *pos;
 442         int rc = 1;
 443
 444         list_for_each_safe(tmp, pos, queue) {
 445                 struct ldlm_lock *child;
 446                 ldlm_res_compat compat;
 447
 448                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 449                 if (lock == child)
 450                         continue;
 451
 452                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 453                 if (compat && compat(child, lock)) {
 454                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 455                         continue;
 456                 }
 457                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 458                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 459                         continue;
 460                 }
 461
 462                 rc = 0;
 463
 464                 if (send_cbs && child->l_blocking_ast != NULL) {
 465                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 466                                "blocking AST.\n", child);
 467                         ldlm_add_ast_work_item(child, lock);
 468                 }
 469         }
 470
 471         return rc;
 472 }
 473
 474 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 475 {
 476         int rc;
 477         ENTRY;
 478
 479         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 480         rc = ldlm_lock_compat_list(lock, send_cbs,
 481                                    &lock->l_resource->lr_granted);
 482         /* FIXME: should we be sending ASTs to converting? */
 483         if (rc)
 484                 rc = ldlm_lock_compat_list
 485                         (lock, send_cbs, &lock->l_resource->lr_converting);
 486
 487         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 488         RETURN(rc);
 489 }
 490
 491 /* NOTE: called by
 492    - ldlm_handle_enqueuque - resource
 493 */
 494 void ldlm_grant_lock(struct ldlm_lock *lock)
 495 {
 496         struct ldlm_resource *res = lock->l_resource;
 497         ENTRY;
 498
 499         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 500         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 501         lock->l_granted_mode = lock->l_req_mode;
 502
 503         if (lock->l_granted_mode < res->lr_most_restr)
 504                 res->lr_most_restr = lock->l_granted_mode;
 505
 506         if (lock->l_completion_ast) {
 507                 ldlm_add_ast_work_item(lock, NULL);
 508         }
 509         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 510         EXIT;
 511 }
 512
 513 /* returns a referenced lock or NULL */
 514 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 515                                       struct ldlm_extent *extent)
 516 {
 517         struct ldlm_lock *lock;
 518         struct list_head *tmp;
 519
 520         list_for_each(tmp, queue) {
 521                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 522
 523                 if (lock->l_flags & (LDLM_FL_CBPENDING | LDLM_FL_DESTROYED))
 524                         continue;
 525
 526                 /* lock_convert() takes the resource lock, so we're sure that
 527                  * req_mode, lr_type, and l_cookie won't change beneath us */
 528                 if (lock->l_req_mode != mode)
 529                         continue;
 530
 531                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 532                     (lock->l_extent.start > extent->start ||
 533                      lock->l_extent.end < extent->end))
 534                         continue;
 535
 536                 ldlm_lock_addref_internal(lock, mode);
 537                 return lock;
 538         }
 539
 540         return NULL;
 541 }
 542
 543 /* Must be called with no resource or lock locks held.
 544  *
 545  * Returns 1 if it finds an already-existing lock that is compatible; in this
 546  * case, lockh is filled in with a addref()ed lock
 547 */
 548 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 * res_id, __u32 type,
 549                     void *cookie, int cookielen, ldlm_mode_t mode,
 550                     struct lustre_handle *lockh)
 551 {
 552         struct ldlm_resource *res;
 553         struct ldlm_lock *lock;
 554         int rc = 0;
 555         ENTRY;
 556
 557         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 558         if (res == NULL)
 559                 RETURN(0);
 560
 561         ns = res->lr_namespace;
 562         l_lock(&ns->ns_lock);
 563
 564         if ((lock = search_queue(&res->lr_granted, mode, cookie)))
 565                 GOTO(out, rc = 1);
 566         if ((lock = search_queue(&res->lr_converting, mode, cookie)))
 567                 GOTO(out, rc = 1);
 568         if ((lock = search_queue(&res->lr_waiting, mode, cookie)))
 569                 GOTO(out, rc = 1);
 570
 571         EXIT;
 572       out:
 573         ldlm_resource_put(res);
 574         l_unlock(&ns->ns_lock);
 575
 576         if (lock) {
 577                 ldlm_lock2handle(lock, lockh);
 578                 if (lock->l_completion_ast)
 579                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 580         }
 581         if (rc)
 582                 LDLM_DEBUG(lock, "matched");
 583         else
 584                 LDLM_DEBUG_NOLOCK("not matched");
 585         return rc;
 586 }
 587
 588 /* Returns a referenced lock */
 589 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 590                                    struct lustre_handle *parent_lock_handle,
 591                                    __u64 * res_id, __u32 type,
 592                                    ldlm_mode_t mode, void *data, __u32 data_len)
 593 {
 594         struct ldlm_resource *res, *parent_res = NULL;
 595         struct ldlm_lock *lock, *parent_lock;
 596
 597         parent_lock = ldlm_handle2lock(parent_lock_handle);
 598         if (parent_lock)
 599                 parent_res = parent_lock->l_resource;
 600
 601         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 602         if (res == NULL)
 603                 RETURN(NULL);
 604
 605         lock = ldlm_lock_new(parent_lock, res);
 606         if (lock == NULL) {
 607                 ldlm_resource_put(res);
 608                 RETURN(NULL);
 609         }
 610
 611         lock->l_req_mode = mode;
 612         lock->l_data = data;
 613         lock->l_data_len = data_len;
 614
 615         return lock;
 616 }
 617
 618 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 619 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock * lock,
 620                                void *cookie, int cookie_len,
 621                                int *flags,
 622                                ldlm_completion_callback completion,
 623                                ldlm_blocking_callback blocking)
 624 {
 625         struct ldlm_resource *res;
 626         int local;
 627         ldlm_res_policy policy;
 628         ENTRY;
 629
 630         res = lock->l_resource;
 631         lock->l_blocking_ast = blocking;
 632
 633         if (res->lr_type == LDLM_EXTENT)
 634                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 635
 636         /* policies are not executed on the client */
 637         local = res->lr_namespace->ns_client;
 638         if (!local && (policy = ldlm_res_policy_table[res->lr_type])) {
 639                 int rc;
 640                 rc = policy(lock, cookie, lock->l_req_mode, NULL);
 641
 642                 if (rc == ELDLM_LOCK_CHANGED) {
 643                         res = lock->l_resource;
 644                         *flags |= LDLM_FL_LOCK_CHANGED;
 645                 } else if (rc == ELDLM_LOCK_ABORTED) {
 646                         ldlm_lock_destroy(lock);
 647                         RETURN(rc);
 648                 }
 649         }
 650
 651         lock->l_cookie = cookie;
 652         lock->l_cookie_len = cookie_len;
 653
 654         l_lock(&res->lr_namespace->ns_lock);
 655         if (local && lock->l_req_mode == lock->l_granted_mode) {
 656                 /* The server returned a blocked lock, but it was granted before
 657                  * we got a chance to actually enqueue it.  We don't need to do
 658                  * anything else. */
 659                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 660                           LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 661                 GOTO(out, ELDLM_OK);
 662         }
 663
 664         /* This distinction between local lock trees is very important; a client
 665          * namespace only has information about locks taken by that client, and
 666          * thus doesn't have enough information to decide for itself if it can
 667          * be granted (below).  In this case, we do exactly what the server
 668          * tells us to do, as dictated by the 'flags' */
 669         ldlm_resource_unlink_lock(lock);
 670         if (local) {
 671                 if (*flags & LDLM_FL_BLOCK_CONV)
 672                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 673                                                lock);
 674                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 675                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 676                 else
 677                         ldlm_grant_lock(lock);
 678                 GOTO(out, ELDLM_OK);
 679         }
 680
 681         /* FIXME: We may want to optimize by checking lr_most_restr */
 682         if (!list_empty(&res->lr_converting)) {
 683                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 684                 *flags |= LDLM_FL_BLOCK_CONV;
 685                 GOTO(out, ELDLM_OK);
 686         }
 687         if (!list_empty(&res->lr_waiting)) {
 688                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 689                 *flags |= LDLM_FL_BLOCK_WAIT;
 690                 GOTO(out, ELDLM_OK);
 691         }
 692         if (!ldlm_lock_compat(lock, 0)) {
 693                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 694                 *flags |= LDLM_FL_BLOCK_GRANTED;
 695                 GOTO(out, ELDLM_OK);
 696         }
 697
 698         ldlm_grant_lock(lock);
 699         EXIT;
 700       out:
 701         l_unlock(&res->lr_namespace->ns_lock);
 702         /* Don't set 'completion_ast' until here so that if the lock is granted
 703          * immediately we don't do an unnecessary completion call. */
 704         lock->l_completion_ast = completion;
 705         return ELDLM_OK;
 706 }
 707
 708 /* Must be called with namespace taken: queue is waiting or converting. */
 709 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 710                                 struct list_head *queue)
 711 {
 712         struct list_head *tmp, *pos;
 713         ENTRY;
 714
 715         list_for_each_safe(tmp, pos, queue) {
 716                 struct ldlm_lock *pending;
 717                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 718
 719                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 720
 721                 if (!ldlm_lock_compat(pending, 1))
 722                         RETURN(1);
 723
 724                 list_del_init(&pending->l_res_link);
 725                 ldlm_grant_lock(pending);
 726         }
 727
 728         RETURN(0);
 729 }
 730
 731 void ldlm_run_ast_work(struct list_head *rpc_list)
 732 {
 733         struct list_head *tmp, *pos;
 734         int rc;
 735         ENTRY;
 736
 737         list_for_each_safe(tmp, pos, rpc_list) {
 738                 struct ldlm_ast_work *w =
 739                         list_entry(tmp, struct ldlm_ast_work, w_list);
 740
 741                 if (w->w_blocking)
 742                         rc = w->w_lock->l_blocking_ast
 743                                 (w->w_lock, &w->w_desc, w->w_data,
 744                                  w->w_datalen, LDLM_CB_BLOCKING);
 745                 else
 746                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags);
 747                 if (rc)
 748                         CERROR("Failed AST - should clean & disconnect "
 749                                "client\n");
 750                 LDLM_LOCK_PUT(w->w_lock);
 751                 list_del(&w->w_list);
 752                 OBD_FREE(w, sizeof(*w));
 753         }
 754         EXIT;
 755 }
 756
 757 /* Must be called with resource->lr_lock not taken. */
 758 void ldlm_reprocess_all(struct ldlm_resource *res)
 759 {
 760         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 761         ENTRY;
 762
 763         /* Local lock trees don't get reprocessed. */
 764         if (res->lr_namespace->ns_client) {
 765                 EXIT;
 766                 return;
 767         }
 768
 769         l_lock(&res->lr_namespace->ns_lock);
 770         res->lr_tmp = &rpc_list;
 771
 772         ldlm_reprocess_queue(res, &res->lr_converting);
 773         if (list_empty(&res->lr_converting))
 774                 ldlm_reprocess_queue(res, &res->lr_waiting);
 775
 776         res->lr_tmp = NULL;
 777         l_unlock(&res->lr_namespace->ns_lock);
 778
 779         ldlm_run_ast_work(&rpc_list);
 780         EXIT;
 781 }
 782
 783 void ldlm_cancel_callback(struct ldlm_lock *lock)
 784 {
 785         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 786         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
 787                 lock->l_flags |= LDLM_FL_CANCEL;
 788                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 789                                      lock->l_data_len, LDLM_CB_CANCELING);
 790         }
 791         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 792 }
 793
 794 void ldlm_lock_cancel(struct ldlm_lock *lock)
 795 {
 796         struct ldlm_resource *res;
 797         struct ldlm_namespace *ns;
 798         ENTRY;
 799
 800         res = lock->l_resource;
 801         ns = res->lr_namespace;
 802
 803         l_lock(&ns->ns_lock);
 804         if (lock->l_readers || lock->l_writers)
 805                 CDEBUG(D_INFO, "lock still has references (%d readers, %d "
 806                        "writers)\n", lock->l_readers, lock->l_writers);
 807
 808         ldlm_cancel_callback(lock);
 809
 810         ldlm_del_waiting_lock(lock);
 811         ldlm_resource_unlink_lock(lock);
 812         ldlm_lock_destroy(lock);
 813         l_unlock(&ns->ns_lock);
 814         EXIT;
 815 }
 816
 817 void ldlm_cancel_locks_for_export(struct obd_export *exp)
 818 {
 819         struct list_head *iter, *n; /* MUST BE CALLED "n"! */
 820
 821         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
 822                 struct ldlm_lock *lock;
 823                 struct ldlm_resource *res;
 824                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 825                 res = ldlm_resource_getref(lock->l_resource);
 826                 LDLM_DEBUG(lock, "cancelling lock for export %p", exp);
 827                 ldlm_lock_cancel(lock);
 828                 ldlm_reprocess_all(res);
 829                 ldlm_resource_put(res);
 830         }
 831 }
 832
 833 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 834                                         int *flags)
 835 {
 836         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 837         struct ldlm_resource *res;
 838         struct ldlm_namespace *ns;
 839         int granted = 0;
 840         ENTRY;
 841
 842         res = lock->l_resource;
 843         ns = res->lr_namespace;
 844
 845         l_lock(&ns->ns_lock);
 846
 847         lock->l_req_mode = new_mode;
 848         ldlm_resource_unlink_lock(lock);
 849
 850         /* If this is a local resource, put it on the appropriate list. */
 851         if (res->lr_namespace->ns_client) {
 852                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED))
 853                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 854                                                lock);
 855                 else {
 856                         /* This should never happen, because of the way the
 857                          * server handles conversions. */
 858                         LBUG();
 859
 860                         res->lr_tmp = &rpc_list;
 861                         ldlm_grant_lock(lock);
 862                         res->lr_tmp = NULL;
 863                         granted = 1;
 864                         /* FIXME: completion handling not with ns_lock held ! */
 865                         if (lock->l_completion_ast)
 866                                 lock->l_completion_ast(lock, 0);
 867                 }
 868         } else {
 869                 /* FIXME: We should try the conversion right away and possibly
 870                  * return success without the need for an extra AST */
 871                 ldlm_resource_add_lock(res, res->lr_converting.prev, lock);
 872                 *flags |= LDLM_FL_BLOCK_CONV;
 873         }
 874
 875         l_unlock(&ns->ns_lock);
 876
 877         if (granted)
 878                 ldlm_run_ast_work(&rpc_list);
 879         RETURN(res);
 880 }
 881
 882 void ldlm_lock_dump(struct ldlm_lock *lock)
 883 {
 884         char ver[128];
 885
 886         if (!(portal_debug & D_OTHER))
 887                 return;
 888
 889         if (RES_VERSION_SIZE != 4)
 890                 LBUG();
 891
 892         if (!lock) {
 893                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
 894                 return;
 895         }
 896
 897         snprintf(ver, sizeof(ver), "%x %x %x %x",
 898                  lock->l_version[0], lock->l_version[1],
 899                  lock->l_version[2], lock->l_version[3]);
 900
 901         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
 902         if (lock->l_export && lock->l_export->exp_connection)
 903                 CDEBUG(D_OTHER, "  Node: NID %x (rhandle: "LPX64")\n",
 904                        lock->l_export->exp_connection->c_peer.peer_nid,
 905                        lock->l_remote_handle.addr);
 906         else
 907                 CDEBUG(D_OTHER, "  Node: local\n");
 908         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
 909         CDEBUG(D_OTHER, "  Resource: %p ("LPD64")\n", lock->l_resource,
 910                lock->l_resource->lr_name[0]);
 911         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
 912                (int)lock->l_req_mode, (int)lock->l_granted_mode);
 913         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
 914                lock->l_readers, lock->l_writers);
 915         if (lock->l_resource->lr_type == LDLM_EXTENT)
 916                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
 917                        (unsigned long long)lock->l_extent.start,
 918                        (unsigned long long)lock->l_extent.end);
 919 }