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