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