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, 2003 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/lustre_dlm.h>
  29 #include <linux/lustre_mds.h>
  30 #include <linux/obd_class.h>
  31
  32 //struct lustre_lock ldlm_everything_lock;
  33
  34 /* lock types */
  35 char *ldlm_lockname[] = {
  36         [0] "--",
  37         [LCK_EX] "EX",
  38         [LCK_PW] "PW",
  39         [LCK_PR] "PR",
  40         [LCK_CW] "CW",
  41         [LCK_CR] "CR",
  42         [LCK_NL] "NL"
  43 };
  44 char *ldlm_typename[] = {
  45         [LDLM_PLAIN] "PLN",
  46         [LDLM_EXTENT] "EXT",
  47 };
  48
  49 char *ldlm_it2str(int it)
  50 {
  51         switch (it) {
  52         case IT_OPEN:
  53                 return "open";
  54         case IT_CREAT:
  55                 return "creat";
  56         case (IT_OPEN | IT_CREAT):
  57                 return "open|creat";
  58         case IT_READDIR:
  59                 return "readdir";
  60         case IT_GETATTR:
  61                 return "getattr";
  62         case IT_TRUNC:
  63                 return "truncate";
  64         case IT_SETATTR:
  65                 return "setattr";
  66         case IT_LOOKUP:
  67                 return "lookup";
  68         case IT_UNLINK:
  69                 return "unlink";
  70         default:
  71                 CERROR("Unknown intent %d\n", it);
  72                 return "UNKNOWN";
  73         }
  74 }
  75
  76 extern kmem_cache_t *ldlm_lock_slab;
  77 struct lustre_lock ldlm_handle_lock;
  78
  79 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b);
  80
  81 ldlm_res_compat ldlm_res_compat_table[] = {
  82         [LDLM_PLAIN] ldlm_plain_compat,
  83         [LDLM_EXTENT] ldlm_extent_compat,
  84 };
  85
  86 static ldlm_res_policy ldlm_intent_policy_func;
  87
  88 static int ldlm_plain_policy(struct ldlm_namespace *ns, struct ldlm_lock **lock,
  89                              void *req_cookie, ldlm_mode_t mode, int flags,
  90                              void *data)
  91 {
  92         if ((flags & LDLM_FL_HAS_INTENT) && ldlm_intent_policy_func) {
  93                 return ldlm_intent_policy_func(ns, lock, req_cookie, mode,
  94                                                flags, data);
  95         }
  96
  97         return ELDLM_OK;
  98 }
  99
 100 ldlm_res_policy ldlm_res_policy_table[] = {
 101         [LDLM_PLAIN] ldlm_plain_policy,
 102         [LDLM_EXTENT] ldlm_extent_policy,
 103 };
 104
 105 void ldlm_register_intent(ldlm_res_policy arg)
 106 {
 107         ldlm_intent_policy_func = arg;
 108 }
 109
 110 void ldlm_unregister_intent(void)
 111 {
 112         ldlm_intent_policy_func = NULL;
 113 }
 114
 115 /*
 116  * REFCOUNTED LOCK OBJECTS
 117  */
 118
 119
 120 /*
 121  * Lock refcounts, during creation:
 122  *   - one special one for allocation, dec'd only once in destroy
 123  *   - one for being a lock that's in-use
 124  *   - one for the addref associated with a new lock
 125  */
 126 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 127 {
 128         atomic_inc(&lock->l_refc);
 129         return lock;
 130 }
 131
 132 void ldlm_lock_put(struct ldlm_lock *lock)
 133 {
 134         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 135         ENTRY;
 136
 137         if (atomic_dec_and_test(&lock->l_refc)) {
 138                 l_lock(&ns->ns_lock);
 139                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
 140                 LASSERT(lock->l_destroyed);
 141                 LASSERT(list_empty(&lock->l_res_link));
 142
 143                 spin_lock(&ns->ns_counter_lock);
 144                 ns->ns_locks--;
 145                 spin_unlock(&ns->ns_counter_lock);
 146
 147                 ldlm_resource_putref(lock->l_resource);
 148                 lock->l_resource = NULL;
 149
 150                 if (lock->l_parent)
 151                         LDLM_LOCK_PUT(lock->l_parent);
 152
 153                 PORTAL_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
 154                 l_unlock(&ns->ns_lock);
 155         }
 156
 157         EXIT;
 158 }
 159
 160 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
 161 {
 162         ENTRY;
 163         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 164         if (!list_empty(&lock->l_lru)) {
 165                 list_del_init(&lock->l_lru);
 166                 lock->l_resource->lr_namespace->ns_nr_unused--;
 167                 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
 168         }
 169         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 170         EXIT;
 171 }
 172
 173 /* This used to have a 'strict' flact, which recovery would use to mark an
 174  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
 175  * shall explain why it's gone: with the new hash table scheme, once you call
 176  * ldlm_lock_destroy, you can never drop your final references on this lock.
 177  * Because it's not in the hash table anymore.  -phil */
 178 void ldlm_lock_destroy(struct ldlm_lock *lock)
 179 {
 180         ENTRY;
 181         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 182
 183         if (!list_empty(&lock->l_children)) {
 184                 LDLM_ERROR(lock, "still has children (%p)!",
 185                            lock->l_children.next);
 186                 ldlm_lock_dump(D_ERROR, lock);
 187                 LBUG();
 188         }
 189         if (lock->l_readers || lock->l_writers) {
 190                 LDLM_ERROR(lock, "lock still has references");
 191                 ldlm_lock_dump(D_ERROR, lock);
 192                 LBUG();
 193         }
 194
 195         if (!list_empty(&lock->l_res_link)) {
 196                 ldlm_lock_dump(D_ERROR, lock);
 197                 LBUG();
 198         }
 199
 200         if (lock->l_destroyed) {
 201                 LASSERT(list_empty(&lock->l_lru));
 202                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 203                 EXIT;
 204                 return;
 205         }
 206         lock->l_destroyed = 1;
 207
 208         list_del_init(&lock->l_export_chain);
 209         ldlm_lock_remove_from_lru(lock);
 210         portals_handle_unhash(&lock->l_handle);
 211
 212 #if 0
 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         lock->l_export = NULL;
 217         if (lock->l_export && lock->l_completion_ast)
 218                 lock->l_completion_ast(lock, 0);
 219 #endif
 220
 221         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 222         LDLM_LOCK_PUT(lock);
 223         EXIT;
 224 }
 225
 226 /* this is called by portals_handle2object with the handle lock taken */
 227 static void lock_handle_addref(void *lock)
 228 {
 229         LDLM_LOCK_GET((struct ldlm_lock *)lock);
 230 }
 231
 232 /*
 233  * usage: pass in a resource on which you have done ldlm_resource_get
 234  *        pass in a parent lock on which you have done a ldlm_lock_get
 235  *        after return, ldlm_*_put the resource and parent
 236  * returns: lock with refcount 1
 237  */
 238 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 239                                        struct ldlm_resource *resource)
 240 {
 241         struct ldlm_lock *lock;
 242         ENTRY;
 243
 244         if (resource == NULL)
 245                 LBUG();
 246
 247         PORTAL_SLAB_ALLOC(lock, ldlm_lock_slab, sizeof(*lock));
 248         if (lock == NULL)
 249                 RETURN(NULL);
 250
 251         lock->l_resource = ldlm_resource_getref(resource);
 252
 253         atomic_set(&lock->l_refc, 2);
 254         INIT_LIST_HEAD(&lock->l_children);
 255         INIT_LIST_HEAD(&lock->l_res_link);
 256         INIT_LIST_HEAD(&lock->l_lru);
 257         INIT_LIST_HEAD(&lock->l_export_chain);
 258         INIT_LIST_HEAD(&lock->l_pending_chain);
 259         init_waitqueue_head(&lock->l_waitq);
 260
 261         spin_lock(&resource->lr_namespace->ns_counter_lock);
 262         resource->lr_namespace->ns_locks++;
 263         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 264
 265         if (parent != NULL) {
 266                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 267                 lock->l_parent = LDLM_LOCK_GET(parent);
 268                 list_add(&lock->l_childof, &parent->l_children);
 269                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 270         }
 271
 272         INIT_LIST_HEAD(&lock->l_handle.h_link);
 273         portals_handle_hash(&lock->l_handle, lock_handle_addref);
 274
 275         RETURN(lock);
 276 }
 277
 278 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
 279                               struct ldlm_res_id new_resid)
 280 {
 281         struct ldlm_resource *oldres = lock->l_resource;
 282         ENTRY;
 283
 284         l_lock(&ns->ns_lock);
 285         if (memcmp(&new_resid, &lock->l_resource->lr_name,
 286                    sizeof(lock->l_resource->lr_name)) == 0) {
 287                 /* Nothing to do */
 288                 l_unlock(&ns->ns_lock);
 289                 RETURN(0);
 290         }
 291
 292         LASSERT(new_resid.name[0] != 0);
 293
 294         /* This function assumes that the lock isn't on any lists */
 295         LASSERT(list_empty(&lock->l_res_link));
 296
 297         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
 298                                              lock->l_resource->lr_type, 1);
 299         if (lock->l_resource == NULL) {
 300                 LBUG();
 301                 RETURN(-ENOMEM);
 302         }
 303
 304         /* ...and the flowers are still standing! */
 305         ldlm_resource_putref(oldres);
 306
 307         l_unlock(&ns->ns_lock);
 308         RETURN(0);
 309 }
 310
 311 /*
 312  *  HANDLES
 313  */
 314
 315 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 316 {
 317         memset(&lockh->addr, 0x69, sizeof(lockh->addr));
 318         lockh->cookie = lock->l_handle.h_cookie;
 319 }
 320
 321 /* if flags: atomically get the lock and set the flags.
 322  *           Return NULL if flag already set
 323  */
 324
 325 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
 326 {
 327         struct ldlm_lock *lock = NULL, *retval = NULL;
 328         ENTRY;
 329
 330         LASSERT(handle);
 331
 332         lock = portals_handle2object(handle->cookie);
 333         if (lock == NULL)
 334                 RETURN(NULL);
 335
 336         LASSERT(lock->l_resource != NULL);
 337         LASSERT(lock->l_resource->lr_namespace != NULL);
 338
 339         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 340
 341         /* It's unlikely but possible that someone marked the lock as
 342          * destroyed after we did handle2object on it */
 343         if (lock->l_destroyed) {
 344                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
 345                 LDLM_LOCK_PUT(lock);
 346                 GOTO(out, retval);
 347         }
 348
 349         if (flags && (lock->l_flags & flags)) {
 350                 LDLM_LOCK_PUT(lock);
 351                 GOTO(out, retval);
 352         }
 353
 354         if (flags)
 355                 lock->l_flags |= flags;
 356
 357         retval = lock;
 358         EXIT;
 359  out:
 360         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 361         return retval;
 362 }
 363
 364 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
 365                                       struct lustre_handle *handle)
 366 {
 367         struct ldlm_lock *retval = NULL;
 368
 369         l_lock(&ns->ns_lock);
 370         retval = __ldlm_handle2lock(handle, 0);
 371         l_unlock(&ns->ns_lock);
 372
 373         return retval;
 374 }
 375
 376 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 377 {
 378         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 379 }
 380
 381 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 382 {
 383         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 384         desc->l_req_mode = lock->l_req_mode;
 385         desc->l_granted_mode = lock->l_granted_mode;
 386         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 387         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 388 }
 389
 390 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 391                                    struct ldlm_lock *new,
 392                                    void *data, int datalen)
 393 {
 394         struct ldlm_ast_work *w;
 395         ENTRY;
 396
 397         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 398         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 399                 GOTO(out, 0);
 400
 401         OBD_ALLOC(w, sizeof(*w));
 402         if (!w) {
 403                 LBUG();
 404                 GOTO(out, 0);
 405         }
 406
 407         w->w_data = data;
 408         w->w_datalen = datalen;
 409         if (new) {
 410                 lock->l_flags |= LDLM_FL_AST_SENT;
 411                 w->w_blocking = 1;
 412                 ldlm_lock2desc(new, &w->w_desc);
 413         }
 414
 415         w->w_lock = LDLM_LOCK_GET(lock);
 416         list_add(&w->w_list, lock->l_resource->lr_tmp);
 417         EXIT;
 418  out:
 419         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 420         return;
 421 }
 422
 423 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 424 {
 425         struct ldlm_lock *lock;
 426
 427         lock = ldlm_handle2lock(lockh);
 428         ldlm_lock_addref_internal(lock, mode);
 429         LDLM_LOCK_PUT(lock);
 430 }
 431
 432 /* only called for local locks */
 433 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 434 {
 435         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 436         ldlm_lock_remove_from_lru(lock);
 437         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 438                 lock->l_readers++;
 439         else
 440                 lock->l_writers++;
 441         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 442         LDLM_LOCK_GET(lock);
 443         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 444 }
 445
 446 /* Args: unlocked lock */
 447 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 448                                     struct ldlm_res_id, int flags);
 449
 450 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
 451 {
 452         struct ldlm_namespace *ns;
 453         ENTRY;
 454
 455         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 456         ns = lock->l_resource->lr_namespace;
 457         l_lock(&ns->ns_lock);
 458         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR) {
 459                 LASSERT(lock->l_readers > 0);
 460                 lock->l_readers--;
 461         } else {
 462                 LASSERT(lock->l_writers > 0);
 463                 lock->l_writers--;
 464         }
 465
 466         if (lock->l_flags & LDLM_FL_LOCAL &&
 467             !lock->l_readers && !lock->l_writers) {
 468                 /* If this is a local lock on a server namespace and this was
 469                  * the last reference, cancel the lock. */
 470                 CDEBUG(D_INFO, "forcing cancel of local lock\n");
 471                 lock->l_flags |= LDLM_FL_CBPENDING;
 472         }
 473
 474         if (!lock->l_readers && !lock->l_writers &&
 475             (lock->l_flags & LDLM_FL_CBPENDING)) {
 476                 /* If we received a blocked AST and this was the last reference,
 477                  * run the callback. */
 478                 if (!ns->ns_client && lock->l_export)
 479                         CERROR("FL_CBPENDING set on non-local lock--just a "
 480                                "warning\n");
 481
 482                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 483                 l_unlock(&ns->ns_lock);
 484
 485                 /* FIXME: need a real 'desc' here */
 486                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 487                                      LDLM_CB_BLOCKING);
 488         } else if (ns->ns_client && !lock->l_readers && !lock->l_writers) {
 489                 /* If this is a client-side namespace and this was the last
 490                  * reference, put it on the LRU. */
 491                 LASSERT(list_empty(&lock->l_lru));
 492                 LASSERT(ns->ns_nr_unused >= 0);
 493                 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 494                 ns->ns_nr_unused++;
 495                 l_unlock(&ns->ns_lock);
 496                 ldlm_cancel_lru(ns);
 497         } else {
 498                 l_unlock(&ns->ns_lock);
 499         }
 500
 501         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 502
 503         EXIT;
 504 }
 505
 506 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 507 {
 508         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 509         LASSERT(lock != NULL);
 510         ldlm_lock_decref_internal(lock, mode);
 511         LDLM_LOCK_PUT(lock);
 512 }
 513
 514 /* This will drop a lock reference and mark it for destruction, but will not
 515  * necessarily cancel the lock before returning. */
 516 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
 517 {
 518         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 519         ENTRY;
 520
 521         LASSERT(lock != NULL);
 522
 523         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 524         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 525         lock->l_flags |= LDLM_FL_CBPENDING;
 526         ldlm_lock_decref_internal(lock, mode);
 527         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 528         LDLM_LOCK_PUT(lock);
 529 }
 530
 531 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 532                                  struct list_head *queue)
 533 {
 534         struct list_head *tmp, *pos;
 535         int rc = 1;
 536
 537         list_for_each_safe(tmp, pos, queue) {
 538                 struct ldlm_lock *child;
 539                 ldlm_res_compat compat;
 540
 541                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 542                 if (lock == child)
 543                         continue;
 544
 545                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 546                 if (compat && compat(child, lock)) {
 547                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 548                         continue;
 549                 }
 550                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 551                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 552                         continue;
 553                 }
 554
 555                 rc = 0;
 556
 557                 if (send_cbs && child->l_blocking_ast != NULL) {
 558                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 559                                "blocking AST.\n", child);
 560                         ldlm_add_ast_work_item(child, lock, NULL, 0);
 561                 }
 562         }
 563
 564         return rc;
 565 }
 566
 567 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 568 {
 569         int rc;
 570         ENTRY;
 571
 572         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 573         rc = ldlm_lock_compat_list(lock, send_cbs,
 574                                    &lock->l_resource->lr_granted);
 575         /* FIXME: should we be sending ASTs to converting? */
 576         if (rc)
 577                 rc = ldlm_lock_compat_list
 578                         (lock, send_cbs, &lock->l_resource->lr_converting);
 579
 580         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 581         RETURN(rc);
 582 }
 583
 584 /* NOTE: called by
 585  *  - ldlm_lock_enqueue
 586  *  - ldlm_reprocess_queue
 587  *  - ldlm_lock_convert
 588  */
 589 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen)
 590 {
 591         struct ldlm_resource *res = lock->l_resource;
 592         ENTRY;
 593
 594         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 595         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 596         lock->l_granted_mode = lock->l_req_mode;
 597
 598         if (lock->l_granted_mode < res->lr_most_restr)
 599                 res->lr_most_restr = lock->l_granted_mode;
 600
 601         if (lock->l_completion_ast != NULL)
 602                 ldlm_add_ast_work_item(lock, NULL, data, datalen);
 603
 604         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 605         EXIT;
 606 }
 607
 608 /* returns a referenced lock or NULL.  See the flag descriptions below, in the
 609  * comment above ldlm_lock_match */
 610 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 611                                       struct ldlm_extent *extent,
 612                                       struct ldlm_lock *old_lock, int flags)
 613 {
 614         struct ldlm_lock *lock;
 615         struct list_head *tmp;
 616
 617         list_for_each(tmp, queue) {
 618                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 619
 620                 if (lock == old_lock)
 621                         break;
 622
 623                 if (lock->l_flags & LDLM_FL_CBPENDING)
 624                         continue;
 625
 626                 if (lock->l_req_mode != mode)
 627                         continue;
 628
 629                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 630                     (lock->l_extent.start > extent->start ||
 631                      lock->l_extent.end < extent->end))
 632                         continue;
 633
 634                 if (lock->l_destroyed)
 635                         continue;
 636
 637                 if ((flags & LDLM_FL_LOCAL_ONLY) &&
 638                     !(lock->l_flags & LDLM_FL_LOCAL))
 639                         continue;
 640
 641                 ldlm_lock_addref_internal(lock, mode);
 642                 return lock;
 643         }
 644
 645         return NULL;
 646 }
 647
 648 /* Can be called in two ways:
 649  *
 650  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
 651  * for a duplicate of.
 652  *
 653  * Otherwise, all of the fields must be filled in, to match against.
 654  *
 655  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
 656  *     server (ie, connh is NULL)
 657  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
 658  *     list will be considered
 659  *
 660  * Returns 1 if it finds an already-existing lock that is compatible; in this
 661  * case, lockh is filled in with a addref()ed lock
 662  */
 663 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
 664                     struct ldlm_res_id *res_id, __u32 type, void *cookie,
 665                     int cookielen, ldlm_mode_t mode,struct lustre_handle *lockh)
 666 {
 667         struct ldlm_resource *res;
 668         struct ldlm_lock *lock, *old_lock = NULL;
 669         int rc = 0;
 670         ENTRY;
 671
 672         if (ns == NULL) {
 673                 old_lock = ldlm_handle2lock(lockh);
 674                 LASSERT(old_lock);
 675
 676                 ns = old_lock->l_resource->lr_namespace;
 677                 res_id = &old_lock->l_resource->lr_name;
 678                 type = old_lock->l_resource->lr_type;
 679                 mode = old_lock->l_req_mode;
 680         }
 681
 682         res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
 683         if (res == NULL) {
 684                 LASSERT(old_lock == NULL);
 685                 RETURN(0);
 686         }
 687
 688         l_lock(&ns->ns_lock);
 689
 690         lock = search_queue(&res->lr_granted, mode, cookie, old_lock, flags);
 691         if (lock != NULL)
 692                 GOTO(out, rc = 1);
 693         if (flags & LDLM_FL_BLOCK_GRANTED)
 694                 GOTO(out, rc = 0);
 695         lock = search_queue(&res->lr_converting, mode, cookie, old_lock, flags);
 696         if (lock != NULL)
 697                 GOTO(out, rc = 1);
 698         lock = search_queue(&res->lr_waiting, mode, cookie, old_lock, flags);
 699         if (lock != NULL)
 700                 GOTO(out, rc = 1);
 701
 702         EXIT;
 703        out:
 704         ldlm_resource_putref(res);
 705         l_unlock(&ns->ns_lock);
 706
 707         if (lock) {
 708                 ldlm_lock2handle(lock, lockh);
 709                 if (lock->l_completion_ast)
 710                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC, NULL);
 711         }
 712         if (rc)
 713                 LDLM_DEBUG(lock, "matched");
 714         else
 715                 LDLM_DEBUG_NOLOCK("not matched");
 716
 717         if (old_lock)
 718                 LDLM_LOCK_PUT(old_lock);
 719
 720         return rc;
 721 }
 722
 723 /* Returns a referenced lock */
 724 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 725                                    struct lustre_handle *parent_lock_handle,
 726                                    struct ldlm_res_id res_id, __u32 type,
 727                                    ldlm_mode_t mode, void *data, void *cp_data)
 728 {
 729         struct ldlm_resource *res, *parent_res = NULL;
 730         struct ldlm_lock *lock, *parent_lock = NULL;
 731         ENTRY;
 732
 733         if (parent_lock_handle) {
 734                 parent_lock = ldlm_handle2lock(parent_lock_handle);
 735                 if (parent_lock)
 736                         parent_res = parent_lock->l_resource;
 737         }
 738
 739         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 740         if (res == NULL)
 741                 RETURN(NULL);
 742
 743         lock = ldlm_lock_new(parent_lock, res);
 744         ldlm_resource_putref(res);
 745         if (parent_lock != NULL)
 746                 LDLM_LOCK_PUT(parent_lock);
 747
 748         if (lock == NULL)
 749                 RETURN(NULL);
 750
 751         lock->l_req_mode = mode;
 752         lock->l_data = data;
 753         lock->l_cp_data = cp_data;
 754
 755         RETURN(lock);
 756 }
 757
 758 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
 759                                struct ldlm_lock **lockp,
 760                                void *cookie, int cookie_len,
 761                                int *flags,
 762                                ldlm_completion_callback completion,
 763                                ldlm_blocking_callback blocking)
 764 {
 765         struct ldlm_resource *res;
 766         struct ldlm_lock *lock = *lockp;
 767         int local;
 768         ldlm_res_policy policy;
 769         ENTRY;
 770
 771         res = lock->l_resource;
 772         lock->l_blocking_ast = blocking;
 773
 774         if (res->lr_type == LDLM_EXTENT)
 775                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 776
 777         /* policies are not executed on the client or during replay */
 778         local = res->lr_namespace->ns_client;
 779         if (!local && !(*flags & LDLM_FL_REPLAY) &&
 780             (policy = ldlm_res_policy_table[res->lr_type])) {
 781                 int rc;
 782                 rc = policy(ns, lockp, cookie, lock->l_req_mode, *flags, NULL);
 783                 if (rc == ELDLM_LOCK_CHANGED) {
 784                         res = lock->l_resource;
 785                         *flags |= LDLM_FL_LOCK_CHANGED;
 786                 } else if (rc == ELDLM_LOCK_REPLACED) {
 787                         /* The lock that was returned has already been granted,
 788                          * and placed into lockp.  Destroy the old one and our
 789                          * work here is done. */
 790                         ldlm_lock_destroy(lock);
 791                         LDLM_LOCK_PUT(lock);
 792                         *flags |= LDLM_FL_LOCK_CHANGED;
 793                         RETURN(0);
 794                 } else if (rc == ELDLM_LOCK_ABORTED) {
 795                         ldlm_lock_destroy(lock);
 796                         RETURN(rc);
 797                 }
 798         }
 799
 800         l_lock(&ns->ns_lock);
 801         if (local && lock->l_req_mode == lock->l_granted_mode) {
 802                 /* The server returned a blocked lock, but it was granted before
 803                  * we got a chance to actually enqueue it.  We don't need to do
 804                  * anything else. */
 805                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 806                             LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 807                 GOTO(out, ELDLM_OK);
 808         }
 809
 810         /* This distinction between local lock trees is very important; a client
 811          * namespace only has information about locks taken by that client, and
 812          * thus doesn't have enough information to decide for itself if it can
 813          * be granted (below).  In this case, we do exactly what the server
 814          * tells us to do, as dictated by the 'flags'.
 815          *
 816          * We do exactly the same thing during recovery, when the server is
 817          * more or less trusting the clients not to lie.
 818          *
 819          * FIXME (bug 268): Detect obvious lies by checking compatibility in
 820          * granted/converting queues. */
 821         ldlm_resource_unlink_lock(lock);
 822         if (local) {
 823                 if (*flags & LDLM_FL_BLOCK_CONV)
 824                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 825                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 826                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 827                 else
 828                         ldlm_grant_lock(lock, NULL, 0);
 829                 GOTO(out, ELDLM_OK);
 830         } else if (*flags & LDLM_FL_REPLAY) {
 831                 if (*flags & LDLM_FL_BLOCK_CONV) {
 832                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 833                         GOTO(out, ELDLM_OK);
 834                 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
 835                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 836                         GOTO(out, ELDLM_OK);
 837                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
 838                         ldlm_grant_lock(lock, NULL, 0);
 839                         GOTO(out, ELDLM_OK);
 840                 }
 841                 /* If no flags, fall through to normal enqueue path. */
 842         }
 843
 844         /* FIXME: We may want to optimize by checking lr_most_restr */
 845         if (!list_empty(&res->lr_converting)) {
 846                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 847                 *flags |= LDLM_FL_BLOCK_CONV;
 848                 GOTO(out, ELDLM_OK);
 849         }
 850         if (!list_empty(&res->lr_waiting)) {
 851                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 852                 *flags |= LDLM_FL_BLOCK_WAIT;
 853                 GOTO(out, ELDLM_OK);
 854         }
 855         if (!ldlm_lock_compat(lock, 0)) {
 856                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 857                 *flags |= LDLM_FL_BLOCK_GRANTED;
 858                 GOTO(out, ELDLM_OK);
 859         }
 860
 861         if (lock->l_granted_cb != NULL && lock->l_data != NULL) {
 862                 /* We just -know- */
 863                 struct ptlrpc_request *req = lock->l_data;
 864                 lock->l_granted_cb(lock, req->rq_repmsg, 0);
 865         }
 866         ldlm_grant_lock(lock, NULL, 0);
 867         EXIT;
 868       out:
 869         l_unlock(&ns->ns_lock);
 870         /* Don't set 'completion_ast' until here so that if the lock is granted
 871          * immediately we don't do an unnecessary completion call. */
 872         lock->l_completion_ast = completion;
 873         return ELDLM_OK;
 874 }
 875
 876 /* Must be called with namespace taken: queue is waiting or converting. */
 877 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 878                                 struct list_head *queue)
 879 {
 880         struct list_head *tmp, *pos;
 881         ENTRY;
 882
 883         list_for_each_safe(tmp, pos, queue) {
 884                 struct ldlm_lock *pending;
 885                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 886
 887                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 888
 889                 if (!ldlm_lock_compat(pending, 1))
 890                         RETURN(1);
 891
 892                 list_del_init(&pending->l_res_link);
 893                 ldlm_grant_lock(pending, NULL, 0);
 894         }
 895
 896         RETURN(0);
 897 }
 898
 899 int ldlm_run_ast_work(struct list_head *rpc_list)
 900 {
 901         struct list_head *tmp, *pos;
 902         int rc, retval = 0;
 903         ENTRY;
 904
 905         list_for_each_safe(tmp, pos, rpc_list) {
 906                 struct ldlm_ast_work *w =
 907                         list_entry(tmp, struct ldlm_ast_work, w_list);
 908
 909                 if (w->w_blocking)
 910                         rc = w->w_lock->l_blocking_ast
 911                                 (w->w_lock, &w->w_desc, w->w_data,
 912                                  LDLM_CB_BLOCKING);
 913                 else
 914                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags,
 915                                                          w->w_data);
 916                 if (rc == -ERESTART)
 917                         retval = rc;
 918                 else if (rc)
 919                         CERROR("Failed AST - should clean & disconnect "
 920                                "client\n");
 921                 LDLM_LOCK_PUT(w->w_lock);
 922                 list_del(&w->w_list);
 923                 OBD_FREE(w, sizeof(*w));
 924         }
 925         RETURN(retval);
 926 }
 927
 928 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
 929 {
 930         ldlm_reprocess_all(res);
 931         return LDLM_ITER_CONTINUE;
 932 }
 933
 934 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
 935 {
 936         (void)ldlm_namespace_foreach_res(ns, reprocess_one_queue, NULL);
 937 }
 938
 939 void ldlm_reprocess_all(struct ldlm_resource *res)
 940 {
 941         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 942         int rc;
 943         ENTRY;
 944
 945         /* Local lock trees don't get reprocessed. */
 946         if (res->lr_namespace->ns_client) {
 947                 EXIT;
 948                 return;
 949         }
 950
 951  restart:
 952         l_lock(&res->lr_namespace->ns_lock);
 953         res->lr_tmp = &rpc_list;
 954
 955         ldlm_reprocess_queue(res, &res->lr_converting);
 956         if (list_empty(&res->lr_converting))
 957                 ldlm_reprocess_queue(res, &res->lr_waiting);
 958
 959         res->lr_tmp = NULL;
 960         l_unlock(&res->lr_namespace->ns_lock);
 961
 962         rc = ldlm_run_ast_work(&rpc_list);
 963         if (rc == -ERESTART)
 964                 goto restart;
 965         EXIT;
 966 }
 967
 968 void ldlm_cancel_callback(struct ldlm_lock *lock)
 969 {
 970         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 971         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
 972                 lock->l_flags |= LDLM_FL_CANCEL;
 973                 if (lock->l_blocking_ast)
 974                         lock->l_blocking_ast(lock, NULL, lock->l_data,
 975                                              LDLM_CB_CANCELING);
 976                 else
 977                         LDLM_DEBUG(lock, "no blocking ast");
 978         }
 979         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 980 }
 981
 982 void ldlm_lock_cancel(struct ldlm_lock *lock)
 983 {
 984         struct ldlm_resource *res;
 985         struct ldlm_namespace *ns;
 986         ENTRY;
 987
 988         ldlm_del_waiting_lock(lock);
 989
 990         res = lock->l_resource;
 991         ns = res->lr_namespace;
 992
 993         l_lock(&ns->ns_lock);
 994         /* Please do not, no matter how tempting, remove this LBUG without
 995          * talking to me first. -phik */
 996         if (lock->l_readers || lock->l_writers) {
 997                 LDLM_DEBUG(lock, "lock still has references");
 998                 ldlm_lock_dump(D_OTHER, lock);
 999                 LBUG();
1000         }
1001
1002         ldlm_cancel_callback(lock);
1003
1004         ldlm_resource_unlink_lock(lock);
1005         ldlm_lock_destroy(lock);
1006         l_unlock(&ns->ns_lock);
1007         EXIT;
1008 }
1009
1010 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data, void *cp_data)
1011 {
1012         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1013         ENTRY;
1014
1015         if (lock == NULL)
1016                 RETURN(-EINVAL);
1017
1018         lock->l_data = data;
1019         lock->l_cp_data = cp_data;
1020
1021         LDLM_LOCK_PUT(lock);
1022
1023         RETURN(0);
1024 }
1025
1026 /* This function is only called from one thread (per export); no locking around
1027  * the list ops needed */
1028 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1029 {
1030         struct list_head *iter, *n;
1031
1032         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
1033                 struct ldlm_lock *lock;
1034                 struct ldlm_resource *res;
1035                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1036                 res = ldlm_resource_getref(lock->l_resource);
1037                 LDLM_DEBUG(lock, "export %p", exp);
1038                 ldlm_lock_cancel(lock);
1039                 ldlm_reprocess_all(res);
1040                 ldlm_resource_putref(res);
1041         }
1042 }
1043
1044 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1045                                         int *flags)
1046 {
1047         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1048         struct ldlm_resource *res;
1049         struct ldlm_namespace *ns;
1050         int granted = 0;
1051         ENTRY;
1052
1053         LBUG();
1054
1055         res = lock->l_resource;
1056         ns = res->lr_namespace;
1057
1058         l_lock(&ns->ns_lock);
1059
1060         lock->l_req_mode = new_mode;
1061         ldlm_resource_unlink_lock(lock);
1062
1063         /* If this is a local resource, put it on the appropriate list. */
1064         if (res->lr_namespace->ns_client) {
1065                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1066                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1067                 } else {
1068                         /* This should never happen, because of the way the
1069                          * server handles conversions. */
1070                         LBUG();
1071
1072                         res->lr_tmp = &rpc_list;
1073                         ldlm_grant_lock(lock, NULL, 0);
1074                         res->lr_tmp = NULL;
1075                         granted = 1;
1076                         /* FIXME: completion handling not with ns_lock held ! */
1077                         if (lock->l_completion_ast)
1078                                 lock->l_completion_ast(lock, 0, NULL);
1079                 }
1080         } else {
1081                 /* FIXME: We should try the conversion right away and possibly
1082                  * return success without the need for an extra AST */
1083                 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1084                 *flags |= LDLM_FL_BLOCK_CONV;
1085         }
1086
1087         l_unlock(&ns->ns_lock);
1088
1089         if (granted)
1090                 ldlm_run_ast_work(&rpc_list);
1091         RETURN(res);
1092 }
1093
1094 void ldlm_lock_dump(int level, struct ldlm_lock *lock)
1095 {
1096         char ver[128];
1097
1098         if (!((portal_debug | D_ERROR) & level))
1099                 return;
1100
1101         if (RES_VERSION_SIZE != 4)
1102                 LBUG();
1103
1104         if (!lock) {
1105                 CDEBUG(level, "  NULL LDLM lock\n");
1106                 return;
1107         }
1108
1109         snprintf(ver, sizeof(ver), "%x %x %x %x",
1110                  lock->l_version[0], lock->l_version[1],
1111                  lock->l_version[2], lock->l_version[3]);
1112
1113         CDEBUG(level, "  -- Lock dump: %p (%s) (rc: %d)\n", lock, ver,
1114                atomic_read(&lock->l_refc));
1115         if (lock->l_export && lock->l_export->exp_connection)
1116                 CDEBUG(level, "  Node: NID %x (rhandle: "LPX64")\n",
1117                        lock->l_export->exp_connection->c_peer.peer_nid,
1118                        lock->l_remote_handle.cookie);
1119         else
1120                 CDEBUG(level, "  Node: local\n");
1121         CDEBUG(level, "  Parent: %p\n", lock->l_parent);
1122         CDEBUG(level, "  Resource: %p ("LPD64")\n", lock->l_resource,
1123                lock->l_resource->lr_name.name[0]);
1124         CDEBUG(level, "  Requested mode: %d, granted mode: %d\n",
1125                (int)lock->l_req_mode, (int)lock->l_granted_mode);
1126         CDEBUG(level, "  Readers: %u ; Writers; %u\n",
1127                lock->l_readers, lock->l_writers);
1128         if (lock->l_resource->lr_type == LDLM_EXTENT)
1129                 CDEBUG(level, "  Extent: "LPU64" -> "LPU64"\n",
1130                        lock->l_extent.start, lock->l_extent.end);
1131 }
1132
1133 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1134 {
1135         struct ldlm_lock *lock;
1136
1137         lock = ldlm_handle2lock(lockh);
1138         if (lock == NULL)
1139                 return;
1140
1141         ldlm_lock_dump(D_OTHER, lock);
1142
1143         LDLM_LOCK_PUT(lock);
1144 }