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