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 //struct lustre_lock ldlm_everything_lock;
  34
  35 /* lock types */
  36 char *ldlm_lockname[] = {
  37         [0] "--",
  38         [LCK_EX] "EX",
  39         [LCK_PW] "PW",
  40         [LCK_PR] "PR",
  41         [LCK_CW] "CW",
  42         [LCK_CR] "CR",
  43         [LCK_NL] "NL"
  44 };
  45 char *ldlm_typename[] = {
  46         [LDLM_PLAIN] "PLN",
  47         [LDLM_EXTENT] "EXT",
  48 };
  49
  50 char *ldlm_it2str(int it)
  51 {
  52         switch (it) {
  53         case IT_OPEN:
  54                 return "open";
  55         case IT_CREAT:
  56                 return "creat";
  57         case (IT_OPEN | IT_CREAT):
  58                 return "open|creat";
  59         case IT_MKDIR:
  60                 return "mkdir";
  61         case IT_LINK:
  62                 return "link";
  63         case IT_LINK2:
  64                 return "link2";
  65         case IT_SYMLINK:
  66                 return "symlink";
  67         case IT_UNLINK:
  68                 return "unlink";
  69         case IT_RMDIR:
  70                 return "rmdir";
  71         case IT_RENAME:
  72                 return "rename";
  73         case IT_RENAME2:
  74                 return "rename2";
  75         case IT_READDIR:
  76                 return "readdir";
  77         case IT_GETATTR:
  78                 return "getattr";
  79         case IT_SETATTR:
  80                 return "setattr";
  81         case IT_READLINK:
  82                 return "readlink";
  83         case IT_MKNOD:
  84                 return "mknod";
  85         case IT_LOOKUP:
  86                 return "lookup";
  87         default:
  88                 CERROR("Unknown intent %d\n", it);
  89                 return "UNKNOWN";
  90         }
  91 }
  92
  93 extern kmem_cache_t *ldlm_lock_slab;
  94
  95 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b);
  96
  97 ldlm_res_compat ldlm_res_compat_table[] = {
  98         [LDLM_PLAIN] ldlm_plain_compat,
  99         [LDLM_EXTENT] ldlm_extent_compat,
 100 };
 101
 102 static ldlm_res_policy ldlm_intent_policy_func;
 103
 104 static int ldlm_plain_policy(struct ldlm_lock *lock, void *req_cookie,
 105                              ldlm_mode_t mode, int flags, void *data)
 106 {
 107         if ((flags & LDLM_FL_HAS_INTENT) && ldlm_intent_policy_func) {
 108                 return ldlm_intent_policy_func(lock, req_cookie, mode, flags,
 109                                                data);
 110         }
 111
 112         return ELDLM_OK;
 113 }
 114
 115 ldlm_res_policy ldlm_res_policy_table[] = {
 116         [LDLM_PLAIN] ldlm_plain_policy,
 117         [LDLM_EXTENT] ldlm_extent_policy,
 118 };
 119
 120 void ldlm_register_intent(ldlm_res_policy arg)
 121 {
 122         ldlm_intent_policy_func = arg;
 123 }
 124
 125 void ldlm_unregister_intent(void)
 126 {
 127         ldlm_intent_policy_func = NULL;
 128 }
 129
 130 /*
 131  * REFCOUNTED LOCK OBJECTS
 132  */
 133
 134
 135 /*
 136  * Lock refcounts, during creation:
 137  *   - one special one for allocation, dec'd only once in destroy
 138  *   - one for being a lock that's in-use
 139  *   - one for the addref associated with a new lock
 140  */
 141 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 142 {
 143         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 144         lock->l_refc++;
 145         ldlm_resource_getref(lock->l_resource);
 146         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 147         return lock;
 148 }
 149
 150 void ldlm_lock_put(struct ldlm_lock *lock)
 151 {
 152         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 153         ENTRY;
 154
 155         l_lock(&ns->ns_lock);
 156         lock->l_refc--;
 157         //LDLM_DEBUG(lock, "after refc--");
 158         if (lock->l_refc < 0)
 159                 LBUG();
 160
 161         if (ldlm_resource_put(lock->l_resource)) {
 162                 LASSERT(lock->l_refc == 0);
 163                 lock->l_resource = NULL;
 164         }
 165         if (lock->l_parent)
 166                 LDLM_LOCK_PUT(lock->l_parent);
 167
 168         if (lock->l_refc == 0 && (lock->l_flags & LDLM_FL_DESTROYED)) {
 169                 l_unlock(&ns->ns_lock);
 170                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
 171
 172                 //spin_lock(&ldlm_handle_lock);
 173                 spin_lock(&ns->ns_counter_lock);
 174                 ns->ns_locks--;
 175                 spin_unlock(&ns->ns_counter_lock);
 176
 177                 lock->l_resource = NULL;
 178                 lock->l_random = DEAD_HANDLE_MAGIC;
 179                 if (lock->l_export && lock->l_export->exp_connection)
 180                         ptlrpc_put_connection(lock->l_export->exp_connection);
 181                 kmem_cache_free(ldlm_lock_slab, lock);
 182                 //spin_unlock(&ldlm_handle_lock);
 183                 CDEBUG(D_MALLOC, "kfreed 'lock': %d at %p (tot 0).\n",
 184                        sizeof(*lock), lock);
 185         } else
 186                 l_unlock(&ns->ns_lock);
 187
 188         EXIT;
 189 }
 190
 191 void ldlm_lock_destroy(struct ldlm_lock *lock)
 192 {
 193         ENTRY;
 194         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 195
 196         if (!list_empty(&lock->l_children)) {
 197                 LDLM_DEBUG(lock, "still has children (%p)!",
 198                            lock->l_children.next);
 199                 ldlm_lock_dump(lock);
 200                 LBUG();
 201         }
 202         if (lock->l_readers || lock->l_writers) {
 203                 LDLM_DEBUG(lock, "lock still has references");
 204                 ldlm_lock_dump(lock);
 205         }
 206
 207         if (!list_empty(&lock->l_res_link)) {
 208                 ldlm_lock_dump(lock);
 209                 LBUG();
 210         }
 211
 212         if (lock->l_flags & LDLM_FL_DESTROYED) {
 213                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 214                 EXIT;
 215                 return;
 216         }
 217
 218         list_del(&lock->l_export_chain);
 219         lock->l_export = NULL;
 220         lock->l_flags |= LDLM_FL_DESTROYED;
 221
 222         /* Wake anyone waiting for this lock */
 223         /* FIXME: I should probably add yet another flag, instead of using
 224          * l_export to only call this on clients */
 225         if (lock->l_export && lock->l_completion_ast)
 226                 lock->l_completion_ast(lock, 0);
 227
 228         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 229         LDLM_LOCK_PUT(lock);
 230         EXIT;
 231 }
 232
 233 /*
 234    usage: pass in a resource on which you have done get
 235           pass in a parent lock on which you have done a get
 236           do not put the resource or the parent
 237    returns: lock with refcount 1
 238 */
 239 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 240                                        struct ldlm_resource *resource)
 241 {
 242         struct ldlm_lock *lock;
 243         ENTRY;
 244
 245         if (resource == NULL)
 246                 LBUG();
 247
 248         lock = kmem_cache_alloc(ldlm_lock_slab, SLAB_KERNEL);
 249         if (lock == NULL)
 250                 RETURN(NULL);
 251
 252         memset(lock, 0, sizeof(*lock));
 253         get_random_bytes(&lock->l_random, sizeof(__u64));
 254
 255         lock->l_resource = resource;
 256         /* this refcount matches the one of the resource passed
 257            in which is not being put away */
 258         lock->l_refc = 1;
 259         INIT_LIST_HEAD(&lock->l_children);
 260         INIT_LIST_HEAD(&lock->l_res_link);
 261         INIT_LIST_HEAD(&lock->l_export_chain);
 262         INIT_LIST_HEAD(&lock->l_pending_chain);
 263         init_waitqueue_head(&lock->l_waitq);
 264
 265         spin_lock(&resource->lr_namespace->ns_counter_lock);
 266         resource->lr_namespace->ns_locks++;
 267         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 268
 269         if (parent != NULL) {
 270                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 271                 lock->l_parent = parent;
 272                 list_add(&lock->l_childof, &parent->l_children);
 273                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 274         }
 275
 276         CDEBUG(D_MALLOC, "kmalloced 'lock': %d at "
 277                "%p (tot %d).\n", sizeof(*lock), lock, 1);
 278         /* this is the extra refcount, to prevent the lock from evaporating */
 279         LDLM_LOCK_GET(lock);
 280         RETURN(lock);
 281 }
 282
 283 int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3])
 284 {
 285         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 286         struct ldlm_resource *oldres = lock->l_resource;
 287         int type, i;
 288         ENTRY;
 289
 290         l_lock(&ns->ns_lock);
 291         if (memcmp(new_resid, lock->l_resource->lr_name,
 292                    sizeof(lock->l_resource->lr_name)) == 0) {
 293                 /* Nothing to do */
 294                 l_unlock(&ns->ns_lock);
 295                 RETURN(0);
 296         }
 297
 298         type = lock->l_resource->lr_type;
 299         if (new_resid[0] == 0)
 300                 LBUG();
 301         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid, type, 1);
 302         if (lock->l_resource == NULL) {
 303                 LBUG();
 304                 RETURN(-ENOMEM);
 305         }
 306
 307         /* move references over */
 308         for (i = 0; i < lock->l_refc; i++) {
 309                 int rc;
 310                 ldlm_resource_getref(lock->l_resource);
 311                 rc = ldlm_resource_put(oldres);
 312                 if (rc == 1 && i != lock->l_refc - 1)
 313                         LBUG();
 314         }
 315         /* compensate for the initial get above.. */
 316         ldlm_resource_put(lock->l_resource);
 317
 318         l_unlock(&ns->ns_lock);
 319         RETURN(0);
 320 }
 321
 322 /*
 323  *  HANDLES
 324  */
 325
 326 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 327 {
 328         lockh->addr = (__u64) (unsigned long)lock;
 329         lockh->cookie = lock->l_random;
 330 }
 331
 332 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle,
 333                                      int strict)
 334 {
 335         struct ldlm_lock *lock = NULL, *retval = NULL;
 336         ENTRY;
 337
 338         if (!handle || !handle->addr)
 339                 RETURN(NULL);
 340
 341         //spin_lock(&ldlm_handle_lock);
 342         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
 343         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock)) {
 344                 CERROR("bogus lock %p\n", lock);
 345                 GOTO(out2, retval);
 346         }
 347
 348         if (lock->l_random != handle->cookie) {
 349                 CERROR("bogus cookie: lock %p has "LPX64" vs. handle "LPX64"\n",
 350                        lock, lock->l_random, handle->cookie);
 351                 GOTO(out2, NULL);
 352         }
 353         if (!lock->l_resource) {
 354                 CERROR("trying to lock bogus resource: lock %p\n", lock);
 355                 LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 356                 GOTO(out2, retval);
 357         }
 358         if (!lock->l_resource->lr_namespace) {
 359                 CERROR("trying to lock bogus namespace: lock %p\n", lock);
 360                 LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 361                 GOTO(out2, retval);
 362         }
 363
 364         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 365         if (strict && lock->l_flags & LDLM_FL_DESTROYED) {
 366                 CERROR("lock already destroyed: lock %p\n", lock);
 367                 LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 368                 GOTO(out, NULL);
 369         }
 370
 371         retval = LDLM_LOCK_GET(lock);
 372         if (!retval)
 373                 CERROR("lock disappeared below us!!! %p\n", lock);
 374         EXIT;
 375  out:
 376         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 377  out2:
 378         //spin_unlock(&ldlm_handle_lock);
 379         return retval;
 380 }
 381
 382 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 383 {
 384         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 385 }
 386
 387 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 388 {
 389         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 390         desc->l_req_mode = lock->l_req_mode;
 391         desc->l_granted_mode = lock->l_granted_mode;
 392         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 393         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 394 }
 395
 396 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 397                                    struct ldlm_lock *new)
 398 {
 399         struct ldlm_ast_work *w;
 400         ENTRY;
 401
 402         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 403         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 404                 GOTO(out, 0);
 405
 406         OBD_ALLOC(w, sizeof(*w));
 407         if (!w) {
 408                 LBUG();
 409                 GOTO(out, 0);
 410         }
 411
 412         if (new) {
 413                 lock->l_flags |= LDLM_FL_AST_SENT;
 414                 w->w_blocking = 1;
 415                 ldlm_lock2desc(new, &w->w_desc);
 416         }
 417
 418         w->w_lock = LDLM_LOCK_GET(lock);
 419         list_add(&w->w_list, lock->l_resource->lr_tmp);
 420       out:
 421         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 422         return;
 423 }
 424
 425 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 426 {
 427         struct ldlm_lock *lock;
 428
 429         lock = ldlm_handle2lock(lockh);
 430         ldlm_lock_addref_internal(lock, mode);
 431         LDLM_LOCK_PUT(lock);
 432 }
 433
 434 /* only called for local locks */
 435 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 436 {
 437         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 438         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 439                 lock->l_readers++;
 440         else
 441                 lock->l_writers++;
 442         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 443         LDLM_LOCK_GET(lock);
 444         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 445 }
 446
 447 /* Args: unlocked lock */
 448 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 449 {
 450         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 451         ENTRY;
 452
 453         if (lock == NULL)
 454                 LBUG();
 455
 456         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 457         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 458         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 459                 lock->l_readers--;
 460         else
 461                 lock->l_writers--;
 462
 463         /* If we received a blocked AST and this was the last reference,
 464          * run the callback. */
 465         if (!lock->l_readers && !lock->l_writers &&
 466             (lock->l_flags & LDLM_FL_CBPENDING)) {
 467                 if (!lock->l_resource->lr_namespace->ns_client &&
 468                     lock->l_export)
 469                         CERROR("FL_CBPENDING set on non-local lock--just a "
 470                                "warning\n");
 471
 472                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 473                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 474
 475                 /* FIXME: need a real 'desc' here */
 476                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 477                                      lock->l_data_len, LDLM_CB_BLOCKING);
 478         } else
 479                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 480
 481         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 482         LDLM_LOCK_PUT(lock);    /* matches the handle2lock above */
 483
 484         EXIT;
 485 }
 486
 487 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 488                                  struct list_head *queue)
 489 {
 490         struct list_head *tmp, *pos;
 491         int rc = 1;
 492
 493         list_for_each_safe(tmp, pos, queue) {
 494                 struct ldlm_lock *child;
 495                 ldlm_res_compat compat;
 496
 497                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 498                 if (lock == child)
 499                         continue;
 500
 501                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 502                 if (compat && compat(child, lock)) {
 503                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 504                         continue;
 505                 }
 506                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 507                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 508                         continue;
 509                 }
 510
 511                 rc = 0;
 512
 513                 if (send_cbs && child->l_blocking_ast != NULL) {
 514                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 515                                "blocking AST.\n", child);
 516                         ldlm_add_ast_work_item(child, lock);
 517                 }
 518         }
 519
 520         return rc;
 521 }
 522
 523 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 524 {
 525         int rc;
 526         ENTRY;
 527
 528         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 529         rc = ldlm_lock_compat_list(lock, send_cbs,
 530                                    &lock->l_resource->lr_granted);
 531         /* FIXME: should we be sending ASTs to converting? */
 532         if (rc)
 533                 rc = ldlm_lock_compat_list
 534                         (lock, send_cbs, &lock->l_resource->lr_converting);
 535
 536         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 537         RETURN(rc);
 538 }
 539
 540 /* NOTE: called by
 541    - ldlm_handle_enqueuque - resource
 542 */
 543 void ldlm_grant_lock(struct ldlm_lock *lock)
 544 {
 545         struct ldlm_resource *res = lock->l_resource;
 546         ENTRY;
 547
 548         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 549         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 550         lock->l_granted_mode = lock->l_req_mode;
 551
 552         if (lock->l_granted_mode < res->lr_most_restr)
 553                 res->lr_most_restr = lock->l_granted_mode;
 554
 555         if (lock->l_completion_ast) {
 556                 ldlm_add_ast_work_item(lock, NULL);
 557         }
 558         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 559         EXIT;
 560 }
 561
 562 /* returns a referenced lock or NULL */
 563 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 564                                       struct ldlm_extent *extent)
 565 {
 566         struct ldlm_lock *lock;
 567         struct list_head *tmp;
 568
 569         list_for_each(tmp, queue) {
 570                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 571
 572                 if (lock->l_flags & (LDLM_FL_CBPENDING | LDLM_FL_DESTROYED))
 573                         continue;
 574
 575                 /* lock_convert() takes the resource lock, so we're sure that
 576                  * req_mode and lr_type won't change beneath us */
 577                 if (lock->l_req_mode != mode)
 578                         continue;
 579
 580                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 581                     (lock->l_extent.start > extent->start ||
 582                      lock->l_extent.end < extent->end))
 583                         continue;
 584
 585                 ldlm_lock_addref_internal(lock, mode);
 586                 return lock;
 587         }
 588
 589         return NULL;
 590 }
 591
 592 /* Must be called with no resource or lock locks held.
 593  *
 594  * Returns 1 if it finds an already-existing lock that is compatible; in this
 595  * case, lockh is filled in with a addref()ed lock
 596 */
 597 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 * res_id, __u32 type,
 598                     void *cookie, int cookielen, ldlm_mode_t mode,
 599                     struct lustre_handle *lockh)
 600 {
 601         struct ldlm_resource *res;
 602         struct ldlm_lock *lock;
 603         int rc = 0;
 604         ENTRY;
 605
 606         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 607         if (res == NULL)
 608                 RETURN(0);
 609
 610         ns = res->lr_namespace;
 611         l_lock(&ns->ns_lock);
 612
 613         if ((lock = search_queue(&res->lr_granted, mode, cookie)))
 614                 GOTO(out, rc = 1);
 615         if ((lock = search_queue(&res->lr_converting, mode, cookie)))
 616                 GOTO(out, rc = 1);
 617         if ((lock = search_queue(&res->lr_waiting, mode, cookie)))
 618                 GOTO(out, rc = 1);
 619
 620         EXIT;
 621       out:
 622         ldlm_resource_put(res);
 623         l_unlock(&ns->ns_lock);
 624
 625         if (lock) {
 626                 ldlm_lock2handle(lock, lockh);
 627                 if (lock->l_completion_ast)
 628                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 629         }
 630         if (rc)
 631                 LDLM_DEBUG(lock, "matched");
 632         else
 633                 LDLM_DEBUG_NOLOCK("not matched");
 634         return rc;
 635 }
 636
 637 /* Returns a referenced lock */
 638 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 639                                    struct lustre_handle *parent_lock_handle,
 640                                    __u64 * res_id, __u32 type,
 641                                    ldlm_mode_t mode, void *data, __u32 data_len)
 642 {
 643         struct ldlm_resource *res, *parent_res = NULL;
 644         struct ldlm_lock *lock, *parent_lock;
 645
 646         parent_lock = ldlm_handle2lock(parent_lock_handle);
 647         if (parent_lock)
 648                 parent_res = parent_lock->l_resource;
 649
 650         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 651         if (res == NULL)
 652                 RETURN(NULL);
 653
 654         lock = ldlm_lock_new(parent_lock, res);
 655         if (lock == NULL) {
 656                 ldlm_resource_put(res);
 657                 RETURN(NULL);
 658         }
 659
 660         lock->l_req_mode = mode;
 661         lock->l_data = data;
 662         lock->l_data_len = data_len;
 663
 664         return lock;
 665 }
 666
 667 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 668 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock * lock,
 669                                void *cookie, int cookie_len,
 670                                int *flags,
 671                                ldlm_completion_callback completion,
 672                                ldlm_blocking_callback blocking)
 673 {
 674         struct ldlm_resource *res;
 675         int local;
 676         ldlm_res_policy policy;
 677         ENTRY;
 678
 679         res = lock->l_resource;
 680         lock->l_blocking_ast = blocking;
 681
 682         if (res->lr_type == LDLM_EXTENT)
 683                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 684
 685         /* policies are not executed on the client or during replay */
 686         local = res->lr_namespace->ns_client;
 687         if (!local && !(*flags & LDLM_FL_REPLAY) &&
 688             (policy = ldlm_res_policy_table[res->lr_type])) {
 689                 int rc;
 690                 rc = policy(lock, cookie, lock->l_req_mode, *flags, NULL);
 691
 692                 if (rc == ELDLM_LOCK_CHANGED) {
 693                         res = lock->l_resource;
 694                         *flags |= LDLM_FL_LOCK_CHANGED;
 695                 } else if (rc == ELDLM_LOCK_ABORTED) {
 696                         ldlm_lock_destroy(lock);
 697                         RETURN(rc);
 698                 }
 699         }
 700
 701         l_lock(&res->lr_namespace->ns_lock);
 702         if (local && lock->l_req_mode == lock->l_granted_mode) {
 703                 /* The server returned a blocked lock, but it was granted before
 704                  * we got a chance to actually enqueue it.  We don't need to do
 705                  * anything else. */
 706                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 707                           LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 708                 GOTO(out, ELDLM_OK);
 709         }
 710
 711         /* This distinction between local lock trees is very important; a client
 712          * namespace only has information about locks taken by that client, and
 713          * thus doesn't have enough information to decide for itself if it can
 714          * be granted (below).  In this case, we do exactly what the server
 715          * tells us to do, as dictated by the 'flags'.
 716          *
 717          * We do exactly the same thing during recovery, when the server is
 718          * more or less trusting the clients not to lie.
 719          *
 720          * FIXME (bug 629283): Detect obvious lies by checking compatibility in
 721          * granted/converting queues. */
 722         ldlm_resource_unlink_lock(lock);
 723         if (local || (*flags & LDLM_FL_REPLAY)) {
 724                 if (*flags & LDLM_FL_BLOCK_CONV)
 725                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 726                                                lock);
 727                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 728                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 729                 else
 730                         ldlm_grant_lock(lock);
 731                 GOTO(out, ELDLM_OK);
 732         }
 733
 734         /* FIXME: We may want to optimize by checking lr_most_restr */
 735         if (!list_empty(&res->lr_converting)) {
 736                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 737                 *flags |= LDLM_FL_BLOCK_CONV;
 738                 GOTO(out, ELDLM_OK);
 739         }
 740         if (!list_empty(&res->lr_waiting)) {
 741                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 742                 *flags |= LDLM_FL_BLOCK_WAIT;
 743                 GOTO(out, ELDLM_OK);
 744         }
 745         if (!ldlm_lock_compat(lock, 0)) {
 746                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 747                 *flags |= LDLM_FL_BLOCK_GRANTED;
 748                 GOTO(out, ELDLM_OK);
 749         }
 750
 751         ldlm_grant_lock(lock);
 752         EXIT;
 753       out:
 754         l_unlock(&res->lr_namespace->ns_lock);
 755         /* Don't set 'completion_ast' until here so that if the lock is granted
 756          * immediately we don't do an unnecessary completion call. */
 757         lock->l_completion_ast = completion;
 758         return ELDLM_OK;
 759 }
 760
 761 /* Must be called with namespace taken: queue is waiting or converting. */
 762 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 763                                 struct list_head *queue)
 764 {
 765         struct list_head *tmp, *pos;
 766         ENTRY;
 767
 768         list_for_each_safe(tmp, pos, queue) {
 769                 struct ldlm_lock *pending;
 770                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 771
 772                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 773
 774                 if (!ldlm_lock_compat(pending, 1))
 775                         RETURN(1);
 776
 777                 list_del_init(&pending->l_res_link);
 778                 ldlm_grant_lock(pending);
 779         }
 780
 781         RETURN(0);
 782 }
 783
 784 void ldlm_run_ast_work(struct list_head *rpc_list)
 785 {
 786         struct list_head *tmp, *pos;
 787         int rc;
 788         ENTRY;
 789
 790         list_for_each_safe(tmp, pos, rpc_list) {
 791                 struct ldlm_ast_work *w =
 792                         list_entry(tmp, struct ldlm_ast_work, w_list);
 793
 794                 if (w->w_blocking)
 795                         rc = w->w_lock->l_blocking_ast
 796                                 (w->w_lock, &w->w_desc, w->w_data,
 797                                  w->w_datalen, LDLM_CB_BLOCKING);
 798                 else
 799                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags);
 800                 if (rc)
 801                         CERROR("Failed AST - should clean & disconnect "
 802                                "client\n");
 803                 LDLM_LOCK_PUT(w->w_lock);
 804                 list_del(&w->w_list);
 805                 OBD_FREE(w, sizeof(*w));
 806         }
 807         EXIT;
 808 }
 809
 810 /* Must be called with resource->lr_lock not taken. */
 811 void ldlm_reprocess_all(struct ldlm_resource *res)
 812 {
 813         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 814         ENTRY;
 815
 816         /* Local lock trees don't get reprocessed. */
 817         if (res->lr_namespace->ns_client) {
 818                 EXIT;
 819                 return;
 820         }
 821
 822         l_lock(&res->lr_namespace->ns_lock);
 823         res->lr_tmp = &rpc_list;
 824
 825         ldlm_reprocess_queue(res, &res->lr_converting);
 826         if (list_empty(&res->lr_converting))
 827                 ldlm_reprocess_queue(res, &res->lr_waiting);
 828
 829         res->lr_tmp = NULL;
 830         l_unlock(&res->lr_namespace->ns_lock);
 831
 832         ldlm_run_ast_work(&rpc_list);
 833         EXIT;
 834 }
 835
 836 void ldlm_cancel_callback(struct ldlm_lock *lock)
 837 {
 838         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 839         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
 840                 lock->l_flags |= LDLM_FL_CANCEL;
 841                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 842                                      lock->l_data_len, LDLM_CB_CANCELING);
 843         }
 844         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 845 }
 846
 847 void ldlm_lock_cancel(struct ldlm_lock *lock)
 848 {
 849         struct ldlm_resource *res;
 850         struct ldlm_namespace *ns;
 851         ENTRY;
 852
 853         res = lock->l_resource;
 854         ns = res->lr_namespace;
 855
 856         l_lock(&ns->ns_lock);
 857         if (lock->l_readers || lock->l_writers)
 858                 CDEBUG(D_INFO, "lock still has references (%d readers, %d "
 859                        "writers)\n", lock->l_readers, lock->l_writers);
 860
 861         ldlm_cancel_callback(lock);
 862
 863         ldlm_del_waiting_lock(lock);
 864         ldlm_resource_unlink_lock(lock);
 865         ldlm_lock_destroy(lock);
 866         l_unlock(&ns->ns_lock);
 867         EXIT;
 868 }
 869
 870 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data, int datalen)
 871 {
 872         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 873         ENTRY;
 874
 875         if (lock == NULL)
 876                 RETURN(-EINVAL);
 877
 878         lock->l_data = data;
 879         lock->l_data_len = datalen;
 880
 881         LDLM_LOCK_PUT(lock);
 882
 883         RETURN(0);
 884 }
 885
 886 void ldlm_cancel_locks_for_export(struct obd_export *exp)
 887 {
 888         struct list_head *iter, *n; /* MUST BE CALLED "n"! */
 889
 890         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
 891                 struct ldlm_lock *lock;
 892                 struct ldlm_resource *res;
 893                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 894                 res = ldlm_resource_getref(lock->l_resource);
 895                 LDLM_DEBUG(lock, "export %p", exp);
 896                 ldlm_lock_cancel(lock);
 897                 ldlm_reprocess_all(res);
 898                 ldlm_resource_put(res);
 899         }
 900 }
 901
 902 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 903                                         int *flags)
 904 {
 905         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 906         struct ldlm_resource *res;
 907         struct ldlm_namespace *ns;
 908         int granted = 0;
 909         ENTRY;
 910
 911         res = lock->l_resource;
 912         ns = res->lr_namespace;
 913
 914         l_lock(&ns->ns_lock);
 915
 916         lock->l_req_mode = new_mode;
 917         ldlm_resource_unlink_lock(lock);
 918
 919         /* If this is a local resource, put it on the appropriate list. */
 920         if (res->lr_namespace->ns_client) {
 921                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED))
 922                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 923                                                lock);
 924                 else {
 925                         /* This should never happen, because of the way the
 926                          * server handles conversions. */
 927                         LBUG();
 928
 929                         res->lr_tmp = &rpc_list;
 930                         ldlm_grant_lock(lock);
 931                         res->lr_tmp = NULL;
 932                         granted = 1;
 933                         /* FIXME: completion handling not with ns_lock held ! */
 934                         if (lock->l_completion_ast)
 935                                 lock->l_completion_ast(lock, 0);
 936                 }
 937         } else {
 938                 /* FIXME: We should try the conversion right away and possibly
 939                  * return success without the need for an extra AST */
 940                 ldlm_resource_add_lock(res, res->lr_converting.prev, lock);
 941                 *flags |= LDLM_FL_BLOCK_CONV;
 942         }
 943
 944         l_unlock(&ns->ns_lock);
 945
 946         if (granted)
 947                 ldlm_run_ast_work(&rpc_list);
 948         RETURN(res);
 949 }
 950
 951 void ldlm_lock_dump(struct ldlm_lock *lock)
 952 {
 953         char ver[128];
 954
 955         if (!(portal_debug & D_OTHER))
 956                 return;
 957
 958         if (RES_VERSION_SIZE != 4)
 959                 LBUG();
 960
 961         if (!lock) {
 962                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
 963                 return;
 964         }
 965
 966         snprintf(ver, sizeof(ver), "%x %x %x %x",
 967                  lock->l_version[0], lock->l_version[1],
 968                  lock->l_version[2], lock->l_version[3]);
 969
 970         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
 971         if (lock->l_export && lock->l_export->exp_connection)
 972                 CDEBUG(D_OTHER, "  Node: NID %x (rhandle: "LPX64")\n",
 973                        lock->l_export->exp_connection->c_peer.peer_nid,
 974                        lock->l_remote_handle.addr);
 975         else
 976                 CDEBUG(D_OTHER, "  Node: local\n");
 977         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
 978         CDEBUG(D_OTHER, "  Resource: %p ("LPD64")\n", lock->l_resource,
 979                lock->l_resource->lr_name[0]);
 980         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
 981                (int)lock->l_req_mode, (int)lock->l_granted_mode);
 982         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
 983                lock->l_readers, lock->l_writers);
 984         if (lock->l_resource->lr_type == LDLM_EXTENT)
 985                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
 986                        (unsigned long long)lock->l_extent.start,
 987                        (unsigned long long)lock->l_extent.end);
 988 }