lustre/ldlm/ldlm_lock.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #define DEBUG_SUBSYSTEM S_LDLM
  25
  26 #include <linux/slab.h>
  27 #include <linux/module.h>
  28 #include <linux/random.h>
  29 #include <linux/lustre_dlm.h>
  30 #include <linux/lustre_mds.h>
  31 #include <linux/obd_class.h>
  32
  33 /* this lock protects ldlm_handle2lock's integrity */
  34 static spinlock_t ldlm_handle_lock = SPIN_LOCK_UNLOCKED;
  35
  36 /* lock types */
  37 char *ldlm_lockname[] = {
  38         [0] "--",
  39         [LCK_EX] "EX",
  40         [LCK_PW] "PW",
  41         [LCK_PR] "PR",
  42         [LCK_CW] "CW",
  43         [LCK_CR] "CR",
  44         [LCK_NL] "NL"
  45 };
  46 char *ldlm_typename[] = {
  47         [LDLM_PLAIN] "PLN",
  48         [LDLM_EXTENT] "EXT",
  49         [LDLM_MDSINTENT] "INT"
  50 };
  51
  52 char *ldlm_it2str(int it)
  53 {
  54         switch (it) {
  55         case IT_OPEN:
  56                 return "open";
  57         case IT_CREAT:
  58                 return "creat";
  59         case (IT_OPEN | IT_CREAT):
  60                 return "open|creat";
  61         case IT_MKDIR:
  62                 return "mkdir";
  63         case IT_LINK:
  64                 return "link";
  65         case IT_LINK2:
  66                 return "link2";
  67         case IT_SYMLINK:
  68                 return "symlink";
  69         case IT_UNLINK:
  70                 return "unlink";
  71         case IT_RMDIR:
  72                 return "rmdir";
  73         case IT_RENAME:
  74                 return "rename";
  75         case IT_RENAME2:
  76                 return "rename2";
  77         case IT_READDIR:
  78                 return "readdir";
  79         case IT_GETATTR:
  80                 return "getattr";
  81         case IT_SETATTR:
  82                 return "setattr";
  83         case IT_READLINK:
  84                 return "readlink";
  85         case IT_MKNOD:
  86                 return "mknod";
  87         case IT_LOOKUP:
  88                 return "lookup";
  89         default:
  90                 CERROR("Unknown intent %d\n", it);
  91                 return "UNKNOWN";
  92         }
  93 }
  94
  95 extern kmem_cache_t *ldlm_lock_slab;
  96
  97 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b);
  98
  99 ldlm_res_compat ldlm_res_compat_table[] = {
 100         [LDLM_PLAIN] ldlm_plain_compat,
 101         [LDLM_EXTENT] ldlm_extent_compat,
 102         [LDLM_MDSINTENT] ldlm_plain_compat
 103 };
 104
 105 ldlm_res_policy ldlm_res_policy_table[] = {
 106         [LDLM_PLAIN] NULL,
 107         [LDLM_EXTENT] ldlm_extent_policy,
 108         [LDLM_MDSINTENT] NULL
 109 };
 110
 111 void ldlm_register_intent(int (*arg) (struct ldlm_lock * lock, void *req_cookie,
 112                                       ldlm_mode_t mode, void *data))
 113 {
 114         ldlm_res_policy_table[LDLM_MDSINTENT] = arg;
 115 }
 116
 117 void ldlm_unregister_intent(void)
 118 {
 119         ldlm_res_policy_table[LDLM_MDSINTENT] = NULL;
 120 }
 121
 122 /*
 123  * REFCOUNTED LOCK OBJECTS
 124  */
 125
 126
 127 /*
 128  * Lock refcounts, during creation:
 129  *   - one special one for allocation, dec'd only once in destroy
 130  *   - one for being a lock that's in-use
 131  *   - one for the addref associated with a new lock
 132  */
 133 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 134 {
 135         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 136         lock->l_refc++;
 137         ldlm_resource_getref(lock->l_resource);
 138         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 139         return lock;
 140 }
 141
 142 void ldlm_lock_put(struct ldlm_lock *lock)
 143 {
 144         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 145         ENTRY;
 146
 147         l_lock(&ns->ns_lock);
 148         lock->l_refc--;
 149         //LDLM_DEBUG(lock, "after refc--");
 150         if (lock->l_refc < 0)
 151                 LBUG();
 152
 153         if (ldlm_resource_put(lock->l_resource))
 154                 lock->l_resource = NULL;
 155         if (lock->l_parent)
 156                 LDLM_LOCK_PUT(lock->l_parent);
 157
 158         if (lock->l_refc == 0 && (lock->l_flags & LDLM_FL_DESTROYED)) {
 159                 l_unlock(&ns->ns_lock);
 160                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
 161
 162                 //spin_lock(&ldlm_handle_lock);
 163                 spin_lock(&ns->ns_counter_lock);
 164                 ns->ns_locks--;
 165                 spin_unlock(&ns->ns_counter_lock);
 166
 167                 lock->l_resource = NULL;
 168                 if (lock->l_export && lock->l_export->exp_connection)
 169                         ptlrpc_put_connection(lock->l_export->exp_connection);
 170                 kmem_cache_free(ldlm_lock_slab, lock);
 171                 //spin_unlock(&ldlm_handle_lock);
 172                 CDEBUG(D_MALLOC, "kfreed 'lock': %d at %p (tot 0).\n",
 173                        sizeof(*lock), lock);
 174         } else
 175                 l_unlock(&ns->ns_lock);
 176
 177         EXIT;
 178 }
 179
 180 void ldlm_lock_destroy(struct ldlm_lock *lock)
 181 {
 182         ENTRY;
 183         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 184
 185         if (!list_empty(&lock->l_children)) {
 186                 LDLM_DEBUG(lock, "still has children (%p)!",
 187                            lock->l_children.next);
 188                 ldlm_lock_dump(lock);
 189                 LBUG();
 190         }
 191         if (lock->l_readers || lock->l_writers) {
 192                 LDLM_DEBUG(lock, "lock still has references");
 193                 ldlm_lock_dump(lock);
 194                 LBUG();
 195         }
 196
 197         if (!list_empty(&lock->l_res_link)) {
 198                 ldlm_lock_dump(lock);
 199                 LBUG();
 200         }
 201
 202         if (lock->l_flags & LDLM_FL_DESTROYED) {
 203                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 204                 EXIT;
 205                 return;
 206         }
 207
 208         list_del(&lock->l_export_chain);
 209         lock->l_export = NULL;
 210         lock->l_flags |= LDLM_FL_DESTROYED;
 211
 212         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 213         LDLM_LOCK_PUT(lock);
 214         EXIT;
 215 }
 216
 217 /*
 218    usage: pass in a resource on which you have done get
 219           pass in a parent lock on which you have done a get
 220           do not put the resource or the parent
 221    returns: lock with refcount 1
 222 */
 223 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 224                                        struct ldlm_resource *resource)
 225 {
 226         struct ldlm_lock *lock;
 227         ENTRY;
 228
 229         if (resource == NULL)
 230                 LBUG();
 231
 232         lock = kmem_cache_alloc(ldlm_lock_slab, SLAB_KERNEL);
 233         if (lock == NULL)
 234                 RETURN(NULL);
 235
 236         memset(lock, 0, sizeof(*lock));
 237         get_random_bytes(&lock->l_random, sizeof(__u64));
 238
 239         lock->l_resource = resource;
 240         /* this refcount matches the one of the resource passed
 241            in which is not being put away */
 242         lock->l_refc = 1;
 243         INIT_LIST_HEAD(&lock->l_children);
 244         INIT_LIST_HEAD(&lock->l_res_link);
 245         INIT_LIST_HEAD(&lock->l_export_chain);
 246         INIT_LIST_HEAD(&lock->l_pending_chain);
 247         init_waitqueue_head(&lock->l_waitq);
 248
 249         spin_lock(&resource->lr_namespace->ns_counter_lock);
 250         resource->lr_namespace->ns_locks++;
 251         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 252
 253         if (parent != NULL) {
 254                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 255                 lock->l_parent = parent;
 256                 list_add(&lock->l_childof, &parent->l_children);
 257                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 258         }
 259
 260         CDEBUG(D_MALLOC, "kmalloced 'lock': %d at "
 261                "%p (tot %d).\n", sizeof(*lock), lock, 1);
 262         /* this is the extra refcount, to prevent the lock from evaporating */
 263         LDLM_LOCK_GET(lock);
 264         RETURN(lock);
 265 }
 266
 267 int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3])
 268 {
 269         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 270         struct ldlm_resource *oldres = lock->l_resource;
 271         int type, i;
 272         ENTRY;
 273
 274         l_lock(&ns->ns_lock);
 275         if (memcmp(new_resid, lock->l_resource->lr_name,
 276                    sizeof(lock->l_resource->lr_name)) == 0) {
 277                 /* Nothing to do */
 278                 l_unlock(&ns->ns_lock);
 279                 RETURN(0);
 280         }
 281
 282         type = lock->l_resource->lr_type;
 283         if (new_resid[0] == 0)
 284                 LBUG();
 285         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid, type, 1);
 286         if (lock->l_resource == NULL) {
 287                 LBUG();
 288                 RETURN(-ENOMEM);
 289         }
 290
 291         /* move references over */
 292         for (i = 0; i < lock->l_refc; i++) {
 293                 int rc;
 294                 ldlm_resource_getref(lock->l_resource);
 295                 rc = ldlm_resource_put(oldres);
 296                 if (rc == 1 && i != lock->l_refc - 1)
 297                         LBUG();
 298         }
 299         /* compensate for the initial get above.. */
 300         ldlm_resource_put(lock->l_resource);
 301
 302         l_unlock(&ns->ns_lock);
 303         RETURN(0);
 304 }
 305
 306 /*
 307  *  HANDLES
 308  */
 309
 310 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 311 {
 312         lockh->addr = (__u64) (unsigned long)lock;
 313         lockh->cookie = lock->l_random;
 314 }
 315
 316 struct ldlm_lock *ldlm_handle2lock(struct lustre_handle *handle)
 317 {
 318         struct ldlm_lock *lock = NULL, *retval = NULL;
 319         ENTRY;
 320
 321         if (!handle || !handle->addr) {
 322                 CERROR("bogus handle %p->"LPX64"\n", handle,
 323                        handle ? handle->addr : -1);
 324                 RETURN(NULL);
 325         }
 326
 327         //spin_lock(&ldlm_handle_lock);
 328         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
 329         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock)) {
 330                 CERROR("bogus lock %p\n", lock);
 331                 GOTO(out2, retval);
 332         }
 333
 334         if (!lock->l_resource) {
 335                 CERROR("trying to lock bogus resource: lock %p\n", lock);
 336                 LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 337                 GOTO(out2, retval);
 338         }
 339         if (!lock->l_resource->lr_namespace) {
 340                 CERROR("trying to lock bogus namespace: lock %p\n", lock);
 341                 LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 342                 GOTO(out2, retval);
 343         }
 344         if (lock->l_random != handle->cookie) {
 345                 CERROR("bogus cookie: lock "LPX64", handle "LPX64"\n",
 346                        lock->l_random, handle->cookie);
 347                 GOTO(out, NULL);
 348         }
 349
 350         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 351         if (lock->l_flags & LDLM_FL_DESTROYED) {
 352                 CERROR("lock already destroyed: lock %p\n", lock);
 353                 LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 354                 GOTO(out, NULL);
 355         }
 356
 357         retval = LDLM_LOCK_GET(lock);
 358         if (!retval)
 359                 CERROR("lock disappeared below us!!! %p\n", lock);
 360         EXIT;
 361  out:
 362         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 363  out2:
 364         //spin_unlock(&ldlm_handle_lock);
 365         return retval;
 366 }
 367
 368 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 369 {
 370         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 371 }
 372
 373 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 374 {
 375         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 376         desc->l_req_mode = lock->l_req_mode;
 377         desc->l_granted_mode = lock->l_granted_mode;
 378         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 379         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 380 }
 381
 382 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 383                                    struct ldlm_lock *new)
 384 {
 385         struct ldlm_ast_work *w;
 386         ENTRY;
 387
 388         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 389         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 390                 GOTO(out, 0);
 391
 392         OBD_ALLOC(w, sizeof(*w));
 393         if (!w) {
 394                 LBUG();
 395                 GOTO(out, 0);
 396         }
 397
 398         if (new) {
 399                 lock->l_flags |= LDLM_FL_AST_SENT;
 400                 w->w_blocking = 1;
 401                 ldlm_lock2desc(new, &w->w_desc);
 402         }
 403
 404         w->w_lock = LDLM_LOCK_GET(lock);
 405         list_add(&w->w_list, lock->l_resource->lr_tmp);
 406       out:
 407         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 408         return;
 409 }
 410
 411 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 412 {
 413         struct ldlm_lock *lock;
 414
 415         lock = ldlm_handle2lock(lockh);
 416         ldlm_lock_addref_internal(lock, mode);
 417         LDLM_LOCK_PUT(lock);
 418 }
 419
 420 /* only called for local locks */
 421 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 422 {
 423         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 424         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 425                 lock->l_readers++;
 426         else
 427                 lock->l_writers++;
 428         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 429         LDLM_LOCK_GET(lock);
 430         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 431 }
 432
 433 /* Args: unlocked lock */
 434 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 435 {
 436         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 437         ENTRY;
 438
 439         if (lock == NULL)
 440                 LBUG();
 441
 442         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 443         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 444         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 445                 lock->l_readers--;
 446         else
 447                 lock->l_writers--;
 448
 449         /* If we received a blocked AST and this was the last reference,
 450          * run the callback. */
 451         if (!lock->l_readers && !lock->l_writers &&
 452             (lock->l_flags & LDLM_FL_CBPENDING)) {
 453                 if (!lock->l_resource->lr_namespace->ns_client &&
 454                     lock->l_export)
 455                         CERROR("FL_CBPENDING set on non-local lock--just a "
 456                                "warning\n");
 457
 458                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 459                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 460
 461                 /* FIXME: need a real 'desc' here */
 462                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 463                                      lock->l_data_len, LDLM_CB_BLOCKING);
 464         } else
 465                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 466
 467         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 468         LDLM_LOCK_PUT(lock);    /* matches the handle2lock above */
 469
 470         EXIT;
 471 }
 472
 473 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 474                                  struct list_head *queue)
 475 {
 476         struct list_head *tmp, *pos;
 477         int rc = 1;
 478
 479         list_for_each_safe(tmp, pos, queue) {
 480                 struct ldlm_lock *child;
 481                 ldlm_res_compat compat;
 482
 483                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 484                 if (lock == child)
 485                         continue;
 486
 487                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 488                 if (compat && compat(child, lock)) {
 489                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 490                         continue;
 491                 }
 492                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 493                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 494                         continue;
 495                 }
 496
 497                 rc = 0;
 498
 499                 if (send_cbs && child->l_blocking_ast != NULL) {
 500                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 501                                "blocking AST.\n", child);
 502                         ldlm_add_ast_work_item(child, lock);
 503                 }
 504         }
 505
 506         return rc;
 507 }
 508
 509 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 510 {
 511         int rc;
 512         ENTRY;
 513
 514         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 515         rc = ldlm_lock_compat_list(lock, send_cbs,
 516                                    &lock->l_resource->lr_granted);
 517         /* FIXME: should we be sending ASTs to converting? */
 518         if (rc)
 519                 rc = ldlm_lock_compat_list
 520                         (lock, send_cbs, &lock->l_resource->lr_converting);
 521
 522         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 523         RETURN(rc);
 524 }
 525
 526 /* NOTE: called by
 527    - ldlm_handle_enqueuque - resource
 528 */
 529 void ldlm_grant_lock(struct ldlm_lock *lock)
 530 {
 531         struct ldlm_resource *res = lock->l_resource;
 532         ENTRY;
 533
 534         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 535         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 536         lock->l_granted_mode = lock->l_req_mode;
 537
 538         if (lock->l_granted_mode < res->lr_most_restr)
 539                 res->lr_most_restr = lock->l_granted_mode;
 540
 541         if (lock->l_completion_ast) {
 542                 ldlm_add_ast_work_item(lock, NULL);
 543         }
 544         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 545         EXIT;
 546 }
 547
 548 /* returns a referenced lock or NULL */
 549 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 550                                       struct ldlm_extent *extent)
 551 {
 552         struct ldlm_lock *lock;
 553         struct list_head *tmp;
 554
 555         list_for_each(tmp, queue) {
 556                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 557
 558                 if (lock->l_flags & (LDLM_FL_CBPENDING | LDLM_FL_DESTROYED))
 559                         continue;
 560
 561                 /* lock_convert() takes the resource lock, so we're sure that
 562                  * req_mode, lr_type, and l_cookie won't change beneath us */
 563                 if (lock->l_req_mode != mode)
 564                         continue;
 565
 566                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 567                     (lock->l_extent.start > extent->start ||
 568                      lock->l_extent.end < extent->end))
 569                         continue;
 570
 571                 ldlm_lock_addref_internal(lock, mode);
 572                 return lock;
 573         }
 574
 575         return NULL;
 576 }
 577
 578 /* Must be called with no resource or lock locks held.
 579  *
 580  * Returns 1 if it finds an already-existing lock that is compatible; in this
 581  * case, lockh is filled in with a addref()ed lock
 582 */
 583 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 * res_id, __u32 type,
 584                     void *cookie, int cookielen, ldlm_mode_t mode,
 585                     struct lustre_handle *lockh)
 586 {
 587         struct ldlm_resource *res;
 588         struct ldlm_lock *lock;
 589         int rc = 0;
 590         ENTRY;
 591
 592         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 593         if (res == NULL)
 594                 RETURN(0);
 595
 596         ns = res->lr_namespace;
 597         l_lock(&ns->ns_lock);
 598
 599         if ((lock = search_queue(&res->lr_granted, mode, cookie)))
 600                 GOTO(out, rc = 1);
 601         if ((lock = search_queue(&res->lr_converting, mode, cookie)))
 602                 GOTO(out, rc = 1);
 603         if ((lock = search_queue(&res->lr_waiting, mode, cookie)))
 604                 GOTO(out, rc = 1);
 605
 606         EXIT;
 607       out:
 608         ldlm_resource_put(res);
 609         l_unlock(&ns->ns_lock);
 610
 611         if (lock) {
 612                 ldlm_lock2handle(lock, lockh);
 613                 if (lock->l_completion_ast)
 614                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 615         }
 616         if (rc)
 617                 LDLM_DEBUG(lock, "matched");
 618         else
 619                 LDLM_DEBUG_NOLOCK("not matched");
 620         return rc;
 621 }
 622
 623 /* Returns a referenced lock */
 624 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 625                                    struct lustre_handle *parent_lock_handle,
 626                                    __u64 * res_id, __u32 type,
 627                                    ldlm_mode_t mode, void *data, __u32 data_len)
 628 {
 629         struct ldlm_resource *res, *parent_res = NULL;
 630         struct ldlm_lock *lock, *parent_lock;
 631
 632         parent_lock = ldlm_handle2lock(parent_lock_handle);
 633         if (parent_lock)
 634                 parent_res = parent_lock->l_resource;
 635
 636         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 637         if (res == NULL)
 638                 RETURN(NULL);
 639
 640         lock = ldlm_lock_new(parent_lock, res);
 641         if (lock == NULL) {
 642                 ldlm_resource_put(res);
 643                 RETURN(NULL);
 644         }
 645
 646         lock->l_req_mode = mode;
 647         lock->l_data = data;
 648         lock->l_data_len = data_len;
 649
 650         return lock;
 651 }
 652
 653 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 654 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock * lock,
 655                                void *cookie, int cookie_len,
 656                                int *flags,
 657                                ldlm_completion_callback completion,
 658                                ldlm_blocking_callback blocking)
 659 {
 660         struct ldlm_resource *res;
 661         int local;
 662         ldlm_res_policy policy;
 663         ENTRY;
 664
 665         res = lock->l_resource;
 666         lock->l_blocking_ast = blocking;
 667
 668         if (res->lr_type == LDLM_EXTENT)
 669                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 670
 671         /* policies are not executed on the client */
 672         local = res->lr_namespace->ns_client;
 673         if (!local && (policy = ldlm_res_policy_table[res->lr_type])) {
 674                 int rc;
 675                 rc = policy(lock, cookie, lock->l_req_mode, NULL);
 676
 677                 if (rc == ELDLM_LOCK_CHANGED) {
 678                         res = lock->l_resource;
 679                         *flags |= LDLM_FL_LOCK_CHANGED;
 680                 } else if (rc == ELDLM_LOCK_ABORTED) {
 681                         ldlm_lock_destroy(lock);
 682                         RETURN(rc);
 683                 }
 684         }
 685
 686         lock->l_cookie = cookie;
 687         lock->l_cookie_len = cookie_len;
 688
 689         l_lock(&res->lr_namespace->ns_lock);
 690         if (local && lock->l_req_mode == lock->l_granted_mode) {
 691                 /* The server returned a blocked lock, but it was granted before
 692                  * we got a chance to actually enqueue it.  We don't need to do
 693                  * anything else. */
 694                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 695                           LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 696                 GOTO(out, ELDLM_OK);
 697         }
 698
 699         /* This distinction between local lock trees is very important; a client
 700          * namespace only has information about locks taken by that client, and
 701          * thus doesn't have enough information to decide for itself if it can
 702          * be granted (below).  In this case, we do exactly what the server
 703          * tells us to do, as dictated by the 'flags' */
 704         ldlm_resource_unlink_lock(lock);
 705         if (local) {
 706                 if (*flags & LDLM_FL_BLOCK_CONV)
 707                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 708                                                lock);
 709                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 710                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 711                 else
 712                         ldlm_grant_lock(lock);
 713                 GOTO(out, ELDLM_OK);
 714         }
 715
 716         /* FIXME: We may want to optimize by checking lr_most_restr */
 717         if (!list_empty(&res->lr_converting)) {
 718                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 719                 *flags |= LDLM_FL_BLOCK_CONV;
 720                 GOTO(out, ELDLM_OK);
 721         }
 722         if (!list_empty(&res->lr_waiting)) {
 723                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 724                 *flags |= LDLM_FL_BLOCK_WAIT;
 725                 GOTO(out, ELDLM_OK);
 726         }
 727         if (!ldlm_lock_compat(lock, 0)) {
 728                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 729                 *flags |= LDLM_FL_BLOCK_GRANTED;
 730                 GOTO(out, ELDLM_OK);
 731         }
 732
 733         ldlm_grant_lock(lock);
 734         EXIT;
 735       out:
 736         l_unlock(&res->lr_namespace->ns_lock);
 737         /* Don't set 'completion_ast' until here so that if the lock is granted
 738          * immediately we don't do an unnecessary completion call. */
 739         lock->l_completion_ast = completion;
 740         return ELDLM_OK;
 741 }
 742
 743 /* Must be called with namespace taken: queue is waiting or converting. */
 744 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 745                                 struct list_head *queue)
 746 {
 747         struct list_head *tmp, *pos;
 748         ENTRY;
 749
 750         list_for_each_safe(tmp, pos, queue) {
 751                 struct ldlm_lock *pending;
 752                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 753
 754                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 755
 756                 if (!ldlm_lock_compat(pending, 1))
 757                         RETURN(1);
 758
 759                 list_del_init(&pending->l_res_link);
 760                 ldlm_grant_lock(pending);
 761         }
 762
 763         RETURN(0);
 764 }
 765
 766 void ldlm_run_ast_work(struct list_head *rpc_list)
 767 {
 768         struct list_head *tmp, *pos;
 769         int rc;
 770         ENTRY;
 771
 772         list_for_each_safe(tmp, pos, rpc_list) {
 773                 struct ldlm_ast_work *w =
 774                         list_entry(tmp, struct ldlm_ast_work, w_list);
 775
 776                 if (w->w_blocking)
 777                         rc = w->w_lock->l_blocking_ast
 778                                 (w->w_lock, &w->w_desc, w->w_data,
 779                                  w->w_datalen, LDLM_CB_BLOCKING);
 780                 else
 781                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags);
 782                 if (rc)
 783                         CERROR("Failed AST - should clean & disconnect "
 784                                "client\n");
 785                 LDLM_LOCK_PUT(w->w_lock);
 786                 list_del(&w->w_list);
 787                 OBD_FREE(w, sizeof(*w));
 788         }
 789         EXIT;
 790 }
 791
 792 /* Must be called with resource->lr_lock not taken. */
 793 void ldlm_reprocess_all(struct ldlm_resource *res)
 794 {
 795         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 796         ENTRY;
 797
 798         /* Local lock trees don't get reprocessed. */
 799         if (res->lr_namespace->ns_client) {
 800                 EXIT;
 801                 return;
 802         }
 803
 804         l_lock(&res->lr_namespace->ns_lock);
 805         res->lr_tmp = &rpc_list;
 806
 807         ldlm_reprocess_queue(res, &res->lr_converting);
 808         if (list_empty(&res->lr_converting))
 809                 ldlm_reprocess_queue(res, &res->lr_waiting);
 810
 811         res->lr_tmp = NULL;
 812         l_unlock(&res->lr_namespace->ns_lock);
 813
 814         ldlm_run_ast_work(&rpc_list);
 815         EXIT;
 816 }
 817
 818 void ldlm_cancel_callback(struct ldlm_lock *lock)
 819 {
 820         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 821         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
 822                 lock->l_flags |= LDLM_FL_CANCEL;
 823                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 824                                      lock->l_data_len, LDLM_CB_CANCELING);
 825         }
 826         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 827 }
 828
 829 void ldlm_lock_cancel(struct ldlm_lock *lock)
 830 {
 831         struct ldlm_resource *res;
 832         struct ldlm_namespace *ns;
 833         ENTRY;
 834
 835         res = lock->l_resource;
 836         ns = res->lr_namespace;
 837
 838         l_lock(&ns->ns_lock);
 839         if (lock->l_readers || lock->l_writers)
 840                 CDEBUG(D_INFO, "lock still has references (%d readers, %d "
 841                        "writers)\n", lock->l_readers, lock->l_writers);
 842
 843         ldlm_cancel_callback(lock);
 844
 845         ldlm_del_waiting_lock(lock);
 846         ldlm_resource_unlink_lock(lock);
 847         ldlm_lock_destroy(lock);
 848         l_unlock(&ns->ns_lock);
 849         EXIT;
 850 }
 851
 852 void ldlm_cancel_locks_for_export(struct obd_export *exp)
 853 {
 854         struct list_head *iter, *n; /* MUST BE CALLED "n"! */
 855
 856         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
 857                 struct ldlm_lock *lock;
 858                 struct ldlm_resource *res;
 859                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 860                 res = ldlm_resource_getref(lock->l_resource);
 861                 LDLM_DEBUG(lock, "cancelling lock for export %p", exp);
 862                 ldlm_lock_cancel(lock);
 863                 ldlm_reprocess_all(res);
 864                 ldlm_resource_put(res);
 865         }
 866 }
 867
 868 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 869                                         int *flags)
 870 {
 871         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 872         struct ldlm_resource *res;
 873         struct ldlm_namespace *ns;
 874         int granted = 0;
 875         ENTRY;
 876
 877         res = lock->l_resource;
 878         ns = res->lr_namespace;
 879
 880         l_lock(&ns->ns_lock);
 881
 882         lock->l_req_mode = new_mode;
 883         ldlm_resource_unlink_lock(lock);
 884
 885         /* If this is a local resource, put it on the appropriate list. */
 886         if (res->lr_namespace->ns_client) {
 887                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED))
 888                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 889                                                lock);
 890                 else {
 891                         /* This should never happen, because of the way the
 892                          * server handles conversions. */
 893                         LBUG();
 894
 895                         res->lr_tmp = &rpc_list;
 896                         ldlm_grant_lock(lock);
 897                         res->lr_tmp = NULL;
 898                         granted = 1;
 899                         /* FIXME: completion handling not with ns_lock held ! */
 900                         if (lock->l_completion_ast)
 901                                 lock->l_completion_ast(lock, 0);
 902                 }
 903         } else {
 904                 /* FIXME: We should try the conversion right away and possibly
 905                  * return success without the need for an extra AST */
 906                 ldlm_resource_add_lock(res, res->lr_converting.prev, lock);
 907                 *flags |= LDLM_FL_BLOCK_CONV;
 908         }
 909
 910         l_unlock(&ns->ns_lock);
 911
 912         if (granted)
 913                 ldlm_run_ast_work(&rpc_list);
 914         RETURN(res);
 915 }
 916
 917 void ldlm_lock_dump(struct ldlm_lock *lock)
 918 {
 919         char ver[128];
 920
 921         if (!(portal_debug & D_OTHER))
 922                 return;
 923
 924         if (RES_VERSION_SIZE != 4)
 925                 LBUG();
 926
 927         if (!lock) {
 928                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
 929                 return;
 930         }
 931
 932         snprintf(ver, sizeof(ver), "%x %x %x %x",
 933                  lock->l_version[0], lock->l_version[1],
 934                  lock->l_version[2], lock->l_version[3]);
 935
 936         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
 937         if (lock->l_export && lock->l_export->exp_connection)
 938                 CDEBUG(D_OTHER, "  Node: NID %x (rhandle: "LPX64")\n",
 939                        lock->l_export->exp_connection->c_peer.peer_nid,
 940                        lock->l_remote_handle.addr);
 941         else
 942                 CDEBUG(D_OTHER, "  Node: local\n");
 943         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
 944         CDEBUG(D_OTHER, "  Resource: %p ("LPD64")\n", lock->l_resource,
 945                lock->l_resource->lr_name[0]);
 946         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
 947                (int)lock->l_req_mode, (int)lock->l_granted_mode);
 948         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
 949                lock->l_readers, lock->l_writers);
 950         if (lock->l_resource->lr_type == LDLM_EXTENT)
 951                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
 952                        (unsigned long long)lock->l_extent.start,
 953                        (unsigned long long)lock->l_extent.end);
 954 }