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