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 #ifdef __KERNEL__
  27 # include <linux/slab.h>
  28 # include <linux/module.h>
  29 # include <linux/lustre_dlm.h>
  30 # include <linux/lustre_mds.h>
  31 #else
  32 # include <liblustre.h>
  33 # include <linux/kp30.h>
  34 #endif
  35
  36 #include <linux/obd_class.h>
  37 #include "ldlm_internal.h"
  38
  39 //struct lustre_lock ldlm_everything_lock;
  40
  41 /* lock types */
  42 char *ldlm_lockname[] = {
  43         [0] "--",
  44         [LCK_EX] "EX",
  45         [LCK_PW] "PW",
  46         [LCK_PR] "PR",
  47         [LCK_CW] "CW",
  48         [LCK_CR] "CR",
  49         [LCK_NL] "NL"
  50 };
  51 char *ldlm_typename[] = {
  52         [LDLM_PLAIN] "PLN",
  53         [LDLM_EXTENT] "EXT",
  54 };
  55
  56 #ifdef __KERNEL__
  57 char *ldlm_it2str(int it)
  58 {
  59         switch (it) {
  60         case IT_OPEN:
  61                 return "open";
  62         case IT_CREAT:
  63                 return "creat";
  64         case (IT_OPEN | IT_CREAT):
  65                 return "open|creat";
  66         case IT_READDIR:
  67                 return "readdir";
  68         case IT_GETATTR:
  69                 return "getattr";
  70         case IT_LOOKUP:
  71                 return "lookup";
  72         case IT_UNLINK:
  73                 return "unlink";
  74         case IT_GETXATTR:
  75                 return "getxattr";
  76         default:
  77                 CERROR("Unknown intent %d\n", it);
  78                 return "UNKNOWN";
  79         }
  80 }
  81 #endif
  82
  83 extern kmem_cache_t *ldlm_lock_slab;
  84 struct lustre_lock ldlm_handle_lock;
  85
  86 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b);
  87
  88 static ldlm_res_compat ldlm_res_compat_table[] = {
  89         [LDLM_PLAIN] ldlm_plain_compat,
  90         [LDLM_EXTENT] ldlm_extent_compat,
  91 };
  92
  93 static ldlm_res_policy ldlm_intent_policy_func;
  94
  95 static int ldlm_plain_policy(struct ldlm_namespace *ns, struct ldlm_lock **lock,
  96                              void *req_cookie, ldlm_mode_t mode, int flags,
  97                              void *data)
  98 {
  99         if ((flags & LDLM_FL_HAS_INTENT) && ldlm_intent_policy_func) {
 100                 return ldlm_intent_policy_func(ns, lock, req_cookie, mode,
 101                                                flags, data);
 102         }
 103
 104         return ELDLM_OK;
 105 }
 106
 107 static ldlm_res_policy ldlm_res_policy_table[] = {
 108         [LDLM_PLAIN] ldlm_plain_policy,
 109         [LDLM_EXTENT] ldlm_extent_policy,
 110 };
 111
 112 void ldlm_register_intent(ldlm_res_policy arg)
 113 {
 114         ldlm_intent_policy_func = arg;
 115 }
 116
 117 void ldlm_unregister_intent(void)
 118 {
 119         ldlm_intent_policy_func = 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         atomic_inc(&lock->l_refc);
 136         return lock;
 137 }
 138
 139 void ldlm_lock_put(struct ldlm_lock *lock)
 140 {
 141         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 142         ENTRY;
 143
 144         if (atomic_dec_and_test(&lock->l_refc)) {
 145                 l_lock(&ns->ns_lock);
 146                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
 147                 LASSERT(lock->l_destroyed);
 148                 LASSERT(list_empty(&lock->l_res_link));
 149
 150                 spin_lock(&ns->ns_counter_lock);
 151                 ns->ns_locks--;
 152                 spin_unlock(&ns->ns_counter_lock);
 153
 154                 ldlm_resource_putref(lock->l_resource);
 155                 lock->l_resource = NULL;
 156
 157                 if (lock->l_parent)
 158                         LDLM_LOCK_PUT(lock->l_parent);
 159
 160                 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
 161                 l_unlock(&ns->ns_lock);
 162         }
 163
 164         EXIT;
 165 }
 166
 167 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
 168 {
 169         ENTRY;
 170         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 171         if (!list_empty(&lock->l_lru)) {
 172                 list_del_init(&lock->l_lru);
 173                 lock->l_resource->lr_namespace->ns_nr_unused--;
 174                 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
 175         }
 176         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 177         EXIT;
 178 }
 179
 180 /* This used to have a 'strict' flact, which recovery would use to mark an
 181  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
 182  * shall explain why it's gone: with the new hash table scheme, once you call
 183  * ldlm_lock_destroy, you can never drop your final references on this lock.
 184  * Because it's not in the hash table anymore.  -phil */
 185 void ldlm_lock_destroy(struct ldlm_lock *lock)
 186 {
 187         ENTRY;
 188         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 189
 190         if (!list_empty(&lock->l_children)) {
 191                 LDLM_ERROR(lock, "still has children (%p)!",
 192                            lock->l_children.next);
 193                 ldlm_lock_dump(D_ERROR, lock);
 194                 LBUG();
 195         }
 196         if (lock->l_readers || lock->l_writers) {
 197                 LDLM_ERROR(lock, "lock still has references");
 198                 ldlm_lock_dump(D_ERROR, lock);
 199                 LBUG();
 200         }
 201
 202         if (!list_empty(&lock->l_res_link)) {
 203                 ldlm_lock_dump(D_ERROR, lock);
 204                 LBUG();
 205         }
 206
 207         if (lock->l_destroyed) {
 208                 LASSERT(list_empty(&lock->l_lru));
 209                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 210                 EXIT;
 211                 return;
 212         }
 213         lock->l_destroyed = 1;
 214
 215         list_del_init(&lock->l_export_chain);
 216         ldlm_lock_remove_from_lru(lock);
 217         class_handle_unhash(&lock->l_handle);
 218
 219 #if 0
 220         /* Wake anyone waiting for this lock */
 221         /* FIXME: I should probably add yet another flag, instead of using
 222          * l_export to only call this on clients */
 223         lock->l_export = NULL;
 224         if (lock->l_export && lock->l_completion_ast)
 225                 lock->l_completion_ast(lock, 0);
 226 #endif
 227
 228         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 229         LDLM_LOCK_PUT(lock);
 230         EXIT;
 231 }
 232
 233 /* this is called by portals_handle2object with the handle lock taken */
 234 static void lock_handle_addref(void *lock)
 235 {
 236         LDLM_LOCK_GET((struct ldlm_lock *)lock);
 237 }
 238
 239 /*
 240  * usage: pass in a resource on which you have done ldlm_resource_get
 241  *        pass in a parent lock on which you have done a ldlm_lock_get
 242  *        after return, ldlm_*_put the resource and parent
 243  * returns: lock with refcount 1
 244  */
 245 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 246                                        struct ldlm_resource *resource)
 247 {
 248         struct ldlm_lock *lock;
 249         ENTRY;
 250
 251         if (resource == NULL)
 252                 LBUG();
 253
 254         OBD_SLAB_ALLOC(lock, ldlm_lock_slab, SLAB_KERNEL, sizeof(*lock));
 255         if (lock == NULL)
 256                 RETURN(NULL);
 257
 258         lock->l_resource = ldlm_resource_getref(resource);
 259
 260         atomic_set(&lock->l_refc, 2);
 261         INIT_LIST_HEAD(&lock->l_children);
 262         INIT_LIST_HEAD(&lock->l_res_link);
 263         INIT_LIST_HEAD(&lock->l_lru);
 264         INIT_LIST_HEAD(&lock->l_export_chain);
 265         INIT_LIST_HEAD(&lock->l_pending_chain);
 266         init_waitqueue_head(&lock->l_waitq);
 267
 268         spin_lock(&resource->lr_namespace->ns_counter_lock);
 269         resource->lr_namespace->ns_locks++;
 270         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 271
 272         if (parent != NULL) {
 273                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 274                 lock->l_parent = LDLM_LOCK_GET(parent);
 275                 list_add(&lock->l_childof, &parent->l_children);
 276                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 277         }
 278
 279         INIT_LIST_HEAD(&lock->l_handle.h_link);
 280         class_handle_hash(&lock->l_handle, lock_handle_addref);
 281
 282         RETURN(lock);
 283 }
 284
 285 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
 286                               struct ldlm_res_id new_resid)
 287 {
 288         struct ldlm_resource *oldres = lock->l_resource;
 289         ENTRY;
 290
 291         l_lock(&ns->ns_lock);
 292         if (memcmp(&new_resid, &lock->l_resource->lr_name,
 293                    sizeof(lock->l_resource->lr_name)) == 0) {
 294                 /* Nothing to do */
 295                 l_unlock(&ns->ns_lock);
 296                 RETURN(0);
 297         }
 298
 299         LASSERT(new_resid.name[0] != 0);
 300
 301         /* This function assumes that the lock isn't on any lists */
 302         LASSERT(list_empty(&lock->l_res_link));
 303
 304         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
 305                                              lock->l_resource->lr_type, 1);
 306         if (lock->l_resource == NULL) {
 307                 LBUG();
 308                 RETURN(-ENOMEM);
 309         }
 310
 311         /* ...and the flowers are still standing! */
 312         ldlm_resource_putref(oldres);
 313
 314         l_unlock(&ns->ns_lock);
 315         RETURN(0);
 316 }
 317
 318 /*
 319  *  HANDLES
 320  */
 321
 322 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 323 {
 324         lockh->cookie = lock->l_handle.h_cookie;
 325 }
 326
 327 /* if flags: atomically get the lock and set the flags.
 328  *           Return NULL if flag already set
 329  */
 330
 331 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
 332 {
 333         struct ldlm_namespace *ns;
 334         struct ldlm_lock *lock = NULL, *retval = NULL;
 335         ENTRY;
 336
 337         LASSERT(handle);
 338
 339         lock = class_handle2object(handle->cookie);
 340         if (lock == NULL)
 341                 RETURN(NULL);
 342
 343         LASSERT(lock->l_resource != NULL);
 344         ns = lock->l_resource->lr_namespace;
 345         LASSERT(ns != NULL);
 346
 347         l_lock(&ns->ns_lock);
 348
 349         /* It's unlikely but possible that someone marked the lock as
 350          * destroyed after we did handle2object on it */
 351         if (lock->l_destroyed) {
 352                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
 353                 LDLM_LOCK_PUT(lock);
 354                 GOTO(out, retval);
 355         }
 356
 357         if (flags && (lock->l_flags & flags)) {
 358                 LDLM_LOCK_PUT(lock);
 359                 GOTO(out, retval);
 360         }
 361
 362         if (flags)
 363                 lock->l_flags |= flags;
 364
 365         retval = lock;
 366         EXIT;
 367  out:
 368         l_unlock(&ns->ns_lock);
 369         return retval;
 370 }
 371
 372 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
 373                                       struct lustre_handle *handle)
 374 {
 375         struct ldlm_lock *retval = NULL;
 376
 377         l_lock(&ns->ns_lock);
 378         retval = __ldlm_handle2lock(handle, 0);
 379         l_unlock(&ns->ns_lock);
 380
 381         return retval;
 382 }
 383
 384 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 385 {
 386         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 387 }
 388
 389 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 390 {
 391         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 392         desc->l_req_mode = lock->l_req_mode;
 393         desc->l_granted_mode = lock->l_granted_mode;
 394         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 395         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 396 }
 397
 398 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 399                                    struct ldlm_lock *new,
 400                                    void *data, int datalen)
 401 {
 402         struct ldlm_ast_work *w;
 403         ENTRY;
 404
 405         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 406         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 407                 GOTO(out, 0);
 408
 409         OBD_ALLOC(w, sizeof(*w));
 410         if (!w) {
 411                 LBUG();
 412                 GOTO(out, 0);
 413         }
 414
 415         w->w_data = data;
 416         w->w_datalen = datalen;
 417         if (new) {
 418                 lock->l_flags |= LDLM_FL_AST_SENT;
 419                 w->w_blocking = 1;
 420                 ldlm_lock2desc(new, &w->w_desc);
 421         }
 422
 423         w->w_lock = LDLM_LOCK_GET(lock);
 424         list_add(&w->w_list, lock->l_resource->lr_tmp);
 425         EXIT;
 426  out:
 427         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 428         return;
 429 }
 430
 431 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 432 {
 433         struct ldlm_lock *lock;
 434
 435         lock = ldlm_handle2lock(lockh);
 436         ldlm_lock_addref_internal(lock, mode);
 437         LDLM_LOCK_PUT(lock);
 438 }
 439
 440 /* only called for local locks */
 441 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 442 {
 443         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 444         ldlm_lock_remove_from_lru(lock);
 445         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 446                 lock->l_readers++;
 447         else
 448                 lock->l_writers++;
 449         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 450         LDLM_LOCK_GET(lock);
 451         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 452 }
 453
 454 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
 455 {
 456         struct ldlm_namespace *ns;
 457         ENTRY;
 458
 459         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 460         ns = lock->l_resource->lr_namespace;
 461         l_lock(&ns->ns_lock);
 462         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR) {
 463                 LASSERT(lock->l_readers > 0);
 464                 lock->l_readers--;
 465         } else {
 466                 LASSERT(lock->l_writers > 0);
 467                 lock->l_writers--;
 468         }
 469
 470         if (lock->l_flags & LDLM_FL_LOCAL &&
 471             !lock->l_readers && !lock->l_writers) {
 472                 /* If this is a local lock on a server namespace and this was
 473                  * the last reference, cancel the lock. */
 474                 CDEBUG(D_INFO, "forcing cancel of local lock\n");
 475                 lock->l_flags |= LDLM_FL_CBPENDING;
 476         }
 477
 478         if (!lock->l_readers && !lock->l_writers &&
 479             (lock->l_flags & LDLM_FL_CBPENDING)) {
 480                 /* If we received a blocked AST and this was the last reference,
 481                  * run the callback. */
 482                 if (!ns->ns_client && lock->l_export)
 483                         CERROR("FL_CBPENDING set on non-local lock--just a "
 484                                "warning\n");
 485
 486                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 487                 l_unlock(&ns->ns_lock);
 488
 489                 /* FIXME: need a real 'desc' here */
 490                 if (lock->l_blocking_ast != NULL)
 491                         lock->l_blocking_ast(lock, NULL, lock->l_data,
 492                                              LDLM_CB_BLOCKING);
 493                 else
 494                         LDLM_DEBUG(lock, "No blocking AST?");
 495         } else if (ns->ns_client && !lock->l_readers && !lock->l_writers) {
 496                 /* If this is a client-side namespace and this was the last
 497                  * reference, put it on the LRU. */
 498                 LASSERT(list_empty(&lock->l_lru));
 499                 LASSERT(ns->ns_nr_unused >= 0);
 500                 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 501                 ns->ns_nr_unused++;
 502                 l_unlock(&ns->ns_lock);
 503                 ldlm_cancel_lru(ns);
 504         } else {
 505                 l_unlock(&ns->ns_lock);
 506         }
 507
 508         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 509
 510         EXIT;
 511 }
 512
 513 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 514 {
 515         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 516         LASSERT(lock != NULL);
 517         ldlm_lock_decref_internal(lock, mode);
 518         LDLM_LOCK_PUT(lock);
 519 }
 520
 521 /* This will drop a lock reference and mark it for destruction, but will not
 522  * necessarily cancel the lock before returning. */
 523 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
 524 {
 525         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 526         ENTRY;
 527
 528         LASSERT(lock != NULL);
 529
 530         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 531         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 532         lock->l_flags |= LDLM_FL_CBPENDING;
 533         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 534         ldlm_lock_decref_internal(lock, mode);
 535         LDLM_LOCK_PUT(lock);
 536 }
 537
 538 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 539                                  struct list_head *queue)
 540 {
 541         struct list_head *tmp, *pos;
 542         int rc = 1;
 543
 544         list_for_each_safe(tmp, pos, queue) {
 545                 struct ldlm_lock *child;
 546                 ldlm_res_compat compat;
 547
 548                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 549                 if (lock == child)
 550                         continue;
 551
 552                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 553                 if (compat && compat(child, lock)) {
 554                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 555                         continue;
 556                 }
 557                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 558                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 559                         continue;
 560                 }
 561
 562                 rc = 0;
 563
 564                 if (send_cbs && child->l_blocking_ast != NULL) {
 565                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 566                                "blocking AST.\n", child);
 567                         ldlm_add_ast_work_item(child, lock, NULL, 0);
 568                 }
 569         }
 570
 571         return rc;
 572 }
 573
 574 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 575 {
 576         int rc;
 577         ENTRY;
 578
 579         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 580         rc = ldlm_lock_compat_list(lock, send_cbs,
 581                                    &lock->l_resource->lr_granted);
 582         /* FIXME: should we be sending ASTs to converting? */
 583         if (rc)
 584                 rc = ldlm_lock_compat_list
 585                         (lock, send_cbs, &lock->l_resource->lr_converting);
 586
 587         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 588         RETURN(rc);
 589 }
 590
 591 /* NOTE: called by
 592  *  - ldlm_lock_enqueue
 593  *  - ldlm_reprocess_queue
 594  *  - ldlm_lock_convert
 595  */
 596 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen)
 597 {
 598         struct ldlm_resource *res = lock->l_resource;
 599         ENTRY;
 600
 601         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 602         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 603         lock->l_granted_mode = lock->l_req_mode;
 604
 605         if (lock->l_granted_mode < res->lr_most_restr)
 606                 res->lr_most_restr = lock->l_granted_mode;
 607
 608         if (lock->l_completion_ast != NULL)
 609                 ldlm_add_ast_work_item(lock, NULL, data, datalen);
 610
 611         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 612         EXIT;
 613 }
 614
 615 /* returns a referenced lock or NULL.  See the flag descriptions below, in the
 616  * comment above ldlm_lock_match */
 617 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 618                                       struct ldlm_extent *extent,
 619                                       struct ldlm_lock *old_lock, void *data,
 620                                       int flags)
 621 {
 622         struct ldlm_lock *lock;
 623         struct list_head *tmp;
 624
 625         list_for_each(tmp, queue) {
 626                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 627
 628                 if (lock == old_lock)
 629                         break;
 630
 631                 /* llite sometimes wants to match locks that will be
 632                  * canceled when their users drop, but we allow it to match
 633                  * if it passes in CBPENDING and the lock still has users.
 634                  * this is generally only going to be used by children
 635                  * whose parents already hold a lock so forward progress
 636                  * can still happen. */
 637                 if (lock->l_flags & LDLM_FL_CBPENDING &&
 638                     !(flags & LDLM_FL_CBPENDING))
 639                         continue;
 640                 if (lock->l_flags & LDLM_FL_CBPENDING &&
 641                     lock->l_readers == 0 && lock->l_writers == 0)
 642                         continue;
 643
 644                 if (lock->l_req_mode != mode)
 645                         continue;
 646
 647                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 648                     (lock->l_extent.start > extent->start ||
 649                      lock->l_extent.end < extent->end))
 650                         continue;
 651
 652                 if (lock->l_destroyed)
 653                         continue;
 654
 655                 if ((flags & LDLM_FL_LOCAL_ONLY) &&
 656                     !(lock->l_flags & LDLM_FL_LOCAL))
 657                         continue;
 658
 659                 if ((flags & LDLM_FL_MATCH_DATA) && lock->l_data != data) {
 660                         LDLM_DEBUG(lock, "data mismatch: have %p, want %p",
 661                                    lock->l_data, data);
 662                         continue;
 663                 }
 664
 665                 ldlm_lock_addref_internal(lock, mode);
 666                 return lock;
 667         }
 668
 669         return NULL;
 670 }
 671
 672 /* Can be called in two ways:
 673  *
 674  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
 675  * for a duplicate of.
 676  *
 677  * Otherwise, all of the fields must be filled in, to match against.
 678  *
 679  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
 680  *     server (ie, connh is NULL)
 681  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
 682  *     list will be considered
 683  * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
 684  *     to be canceled can still be matched as long as they still have reader
 685  *     or writer refernces
 686  * If 'flags' contains LDLM_FL_MATCH_DATA, then only match a lock if the opaque
 687  *     data is the same.
 688  *
 689  * Returns 1 if it finds an already-existing lock that is compatible; in this
 690  * case, lockh is filled in with a addref()ed lock
 691  */
 692 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
 693                     struct ldlm_res_id *res_id, __u32 type, void *cookie,
 694                     int cookielen, ldlm_mode_t mode, void *data,
 695                     struct lustre_handle *lockh)
 696 {
 697         struct ldlm_resource *res;
 698         struct ldlm_lock *lock, *old_lock = NULL;
 699         int rc = 0;
 700         ENTRY;
 701
 702         if (ns == NULL) {
 703                 old_lock = ldlm_handle2lock(lockh);
 704                 LASSERT(old_lock);
 705
 706                 ns = old_lock->l_resource->lr_namespace;
 707                 res_id = &old_lock->l_resource->lr_name;
 708                 type = old_lock->l_resource->lr_type;
 709                 mode = old_lock->l_req_mode;
 710         }
 711
 712         res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
 713         if (res == NULL) {
 714                 LASSERT(old_lock == NULL);
 715                 RETURN(0);
 716         }
 717
 718         l_lock(&ns->ns_lock);
 719
 720         lock = search_queue(&res->lr_granted, mode, cookie, old_lock, data,
 721                             flags);
 722         if (lock != NULL)
 723                 GOTO(out, rc = 1);
 724         if (flags & LDLM_FL_BLOCK_GRANTED)
 725                 GOTO(out, rc = 0);
 726         lock = search_queue(&res->lr_converting, mode, cookie, old_lock, data,
 727                             flags);
 728         if (lock != NULL)
 729                 GOTO(out, rc = 1);
 730         lock = search_queue(&res->lr_waiting, mode, cookie, old_lock, data,
 731                             flags);
 732         if (lock != NULL)
 733                 GOTO(out, rc = 1);
 734
 735         EXIT;
 736  out:
 737         ldlm_resource_putref(res);
 738         l_unlock(&ns->ns_lock);
 739
 740         if (lock) {
 741                 ldlm_lock2handle(lock, lockh);
 742                 if (lock->l_completion_ast)
 743                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 744                                                NULL);
 745         }
 746         if (rc)
 747                 LDLM_DEBUG(lock, "matched");
 748         else
 749                 LDLM_DEBUG_NOLOCK("not matched");
 750
 751         if (old_lock)
 752                 LDLM_LOCK_PUT(old_lock);
 753
 754         return rc;
 755 }
 756
 757 /* Returns a referenced lock */
 758 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 759                                    struct lustre_handle *parent_lock_handle,
 760                                    struct ldlm_res_id res_id, __u32 type,
 761                                    ldlm_mode_t mode,
 762                                    ldlm_blocking_callback blocking,
 763                                    void *data)
 764 {
 765         struct ldlm_resource *res, *parent_res = NULL;
 766         struct ldlm_lock *lock, *parent_lock = NULL;
 767         ENTRY;
 768
 769         if (parent_lock_handle) {
 770                 parent_lock = ldlm_handle2lock(parent_lock_handle);
 771                 if (parent_lock)
 772                         parent_res = parent_lock->l_resource;
 773         }
 774
 775         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 776         if (res == NULL)
 777                 RETURN(NULL);
 778
 779         lock = ldlm_lock_new(parent_lock, res);
 780         ldlm_resource_putref(res);
 781         if (parent_lock != NULL)
 782                 LDLM_LOCK_PUT(parent_lock);
 783
 784         if (lock == NULL)
 785                 RETURN(NULL);
 786
 787         lock->l_req_mode = mode;
 788         lock->l_data = data;
 789         lock->l_blocking_ast = blocking;
 790
 791         RETURN(lock);
 792 }
 793
 794 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
 795                                struct ldlm_lock **lockp,
 796                                void *cookie, int cookie_len,
 797                                int *flags,
 798                                ldlm_completion_callback completion)
 799 {
 800         struct ldlm_resource *res;
 801         struct ldlm_lock *lock = *lockp;
 802         int local;
 803         ldlm_res_policy policy;
 804         ENTRY;
 805
 806         res = lock->l_resource;
 807
 808         if (res->lr_type == LDLM_EXTENT)
 809                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 810
 811         /* policies are not executed on the client or during replay */
 812         local = res->lr_namespace->ns_client;
 813         if (!local && !(*flags & LDLM_FL_REPLAY) &&
 814             (policy = ldlm_res_policy_table[res->lr_type])) {
 815                 int rc;
 816                 rc = policy(ns, lockp, cookie, lock->l_req_mode, *flags, NULL);
 817                 if (rc == ELDLM_LOCK_CHANGED) {
 818                         res = lock->l_resource;
 819                         *flags |= LDLM_FL_LOCK_CHANGED;
 820                 } else if (rc == ELDLM_LOCK_REPLACED) {
 821                         /* The lock that was returned has already been granted,
 822                          * and placed into lockp.  Destroy the old one and our
 823                          * work here is done. */
 824                         ldlm_lock_destroy(lock);
 825                         LDLM_LOCK_PUT(lock);
 826                         *flags |= LDLM_FL_LOCK_CHANGED;
 827                         RETURN(0);
 828                 } else if (rc == ELDLM_LOCK_ABORTED) {
 829                         ldlm_lock_destroy(lock);
 830                         RETURN(rc);
 831                 }
 832         }
 833
 834         l_lock(&ns->ns_lock);
 835         if (local && lock->l_req_mode == lock->l_granted_mode) {
 836                 /* The server returned a blocked lock, but it was granted before
 837                  * we got a chance to actually enqueue it.  We don't need to do
 838                  * anything else. */
 839                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 840                             LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 841                 GOTO(out, ELDLM_OK);
 842         }
 843
 844         /* This distinction between local lock trees is very important; a client
 845          * namespace only has information about locks taken by that client, and
 846          * thus doesn't have enough information to decide for itself if it can
 847          * be granted (below).  In this case, we do exactly what the server
 848          * tells us to do, as dictated by the 'flags'.
 849          *
 850          * We do exactly the same thing during recovery, when the server is
 851          * more or less trusting the clients not to lie.
 852          *
 853          * FIXME (bug 268): Detect obvious lies by checking compatibility in
 854          * granted/converting queues. */
 855         ldlm_resource_unlink_lock(lock);
 856         if (local) {
 857                 if (*flags & LDLM_FL_BLOCK_CONV)
 858                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 859                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 860                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 861                 else
 862                         ldlm_grant_lock(lock, NULL, 0);
 863                 GOTO(out, ELDLM_OK);
 864         } else if (*flags & LDLM_FL_REPLAY) {
 865                 if (*flags & LDLM_FL_BLOCK_CONV) {
 866                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 867                         GOTO(out, ELDLM_OK);
 868                 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
 869                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 870                         GOTO(out, ELDLM_OK);
 871                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
 872                         ldlm_grant_lock(lock, NULL, 0);
 873                         GOTO(out, ELDLM_OK);
 874                 }
 875                 /* If no flags, fall through to normal enqueue path. */
 876         }
 877
 878         /* FIXME: We may want to optimize by checking lr_most_restr */
 879         if (!list_empty(&res->lr_converting)) {
 880                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 881                 *flags |= LDLM_FL_BLOCK_CONV;
 882                 GOTO(out, ELDLM_OK);
 883         }
 884         if (!list_empty(&res->lr_waiting)) {
 885                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 886                 *flags |= LDLM_FL_BLOCK_WAIT;
 887                 GOTO(out, ELDLM_OK);
 888         }
 889         if (!ldlm_lock_compat(lock, 0)) {
 890                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 891                 *flags |= LDLM_FL_BLOCK_GRANTED;
 892                 GOTO(out, ELDLM_OK);
 893         }
 894         ldlm_grant_lock(lock, NULL, 0);
 895         EXIT;
 896       out:
 897         /* Don't set 'completion_ast' until here so that if the lock is granted
 898          * immediately we don't do an unnecessary completion call. */
 899         lock->l_completion_ast = completion;
 900         l_unlock(&ns->ns_lock);
 901         return ELDLM_OK;
 902 }
 903
 904 /* Must be called with namespace taken: queue is waiting or converting. */
 905 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 906                                 struct list_head *queue)
 907 {
 908         struct list_head *tmp, *pos;
 909         ENTRY;
 910
 911         list_for_each_safe(tmp, pos, queue) {
 912                 struct ldlm_lock *pending;
 913                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 914
 915                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 916
 917                 if (!ldlm_lock_compat(pending, 1))
 918                         RETURN(1);
 919
 920                 list_del_init(&pending->l_res_link);
 921                 ldlm_grant_lock(pending, NULL, 0);
 922         }
 923
 924         RETURN(0);
 925 }
 926
 927 int ldlm_run_ast_work(struct list_head *rpc_list)
 928 {
 929         struct list_head *tmp, *pos;
 930         int rc, retval = 0;
 931         ENTRY;
 932
 933         list_for_each_safe(tmp, pos, rpc_list) {
 934                 struct ldlm_ast_work *w =
 935                         list_entry(tmp, struct ldlm_ast_work, w_list);
 936
 937                 /* It's possible to receive a completion AST before we've set
 938                  * the l_completion_ast pointer: either because the AST arrived
 939                  * before the reply, or simply because there's a small race
 940                  * window between receiving the reply and finishing the local
 941                  * enqueue. (bug 842)
 942                  *
 943                  * This can't happen with the blocking_ast, however, because we
 944                  * will never call the local blocking_ast until we drop our
 945                  * reader/writer reference, which we won't do until we get the
 946                  * reply and finish enqueueing. */
 947                 if (w->w_blocking) {
 948                         LASSERT(w->w_lock->l_blocking_ast != NULL);
 949                         rc = w->w_lock->l_blocking_ast
 950                                 (w->w_lock, &w->w_desc, w->w_data,
 951                                  LDLM_CB_BLOCKING);
 952                 } else if (w->w_lock->l_completion_ast != NULL) {
 953                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags,
 954                                                          w->w_data);
 955                 } else {
 956                         rc = 0;
 957                 }
 958                 if (rc == -ERESTART)
 959                         retval = rc;
 960                 else if (rc)
 961                         CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
 962                                "disconnect client\n");
 963                 LDLM_LOCK_PUT(w->w_lock);
 964                 list_del(&w->w_list);
 965                 OBD_FREE(w, sizeof(*w));
 966         }
 967         RETURN(retval);
 968 }
 969
 970 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
 971 {
 972         ldlm_reprocess_all(res);
 973         return LDLM_ITER_CONTINUE;
 974 }
 975
 976 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
 977 {
 978         (void)ldlm_namespace_foreach_res(ns, reprocess_one_queue, NULL);
 979 }
 980
 981 void ldlm_reprocess_all(struct ldlm_resource *res)
 982 {
 983         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 984         int rc;
 985         ENTRY;
 986
 987         /* Local lock trees don't get reprocessed. */
 988         if (res->lr_namespace->ns_client) {
 989                 EXIT;
 990                 return;
 991         }
 992
 993  restart:
 994         l_lock(&res->lr_namespace->ns_lock);
 995         res->lr_tmp = &rpc_list;
 996
 997         ldlm_reprocess_queue(res, &res->lr_converting);
 998         if (list_empty(&res->lr_converting))
 999                 ldlm_reprocess_queue(res, &res->lr_waiting);
1000
1001         res->lr_tmp = NULL;
1002         l_unlock(&res->lr_namespace->ns_lock);
1003
1004         rc = ldlm_run_ast_work(&rpc_list);
1005         if (rc == -ERESTART)
1006                 goto restart;
1007         EXIT;
1008 }
1009
1010 void ldlm_cancel_callback(struct ldlm_lock *lock)
1011 {
1012         l_lock(&lock->l_resource->lr_namespace->ns_lock);
1013         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1014                 lock->l_flags |= LDLM_FL_CANCEL;
1015                 if (lock->l_blocking_ast) {
1016                         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1017                         lock->l_blocking_ast(lock, NULL, lock->l_data,
1018                                              LDLM_CB_CANCELING);
1019                         return;
1020                 } else {
1021                         LDLM_DEBUG(lock, "no blocking ast");
1022                 }
1023         }
1024         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1025 }
1026
1027 void ldlm_lock_cancel(struct ldlm_lock *lock)
1028 {
1029         struct ldlm_resource *res;
1030         struct ldlm_namespace *ns;
1031         ENTRY;
1032
1033         ldlm_del_waiting_lock(lock);
1034
1035         res = lock->l_resource;
1036         ns = res->lr_namespace;
1037
1038         l_lock(&ns->ns_lock);
1039         /* Please do not, no matter how tempting, remove this LBUG without
1040          * talking to me first. -phik */
1041         if (lock->l_readers || lock->l_writers) {
1042                 LDLM_DEBUG(lock, "lock still has references");
1043                 ldlm_lock_dump(D_OTHER, lock);
1044                 LBUG();
1045         }
1046
1047         ldlm_cancel_callback(lock); /* XXX FIXME bug 1030 */
1048
1049         ldlm_resource_unlink_lock(lock);
1050         ldlm_lock_destroy(lock);
1051         l_unlock(&ns->ns_lock);
1052         EXIT;
1053 }
1054
1055 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1056 {
1057         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1058         ENTRY;
1059
1060         if (lock == NULL)
1061                 RETURN(-EINVAL);
1062
1063         lock->l_data = data;
1064
1065         LDLM_LOCK_PUT(lock);
1066
1067         RETURN(0);
1068 }
1069
1070 /* This function is only called from one thread (per export); no locking around
1071  * the list ops needed */
1072 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1073 {
1074         struct list_head *iter, *n;
1075
1076         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
1077                 struct ldlm_lock *lock;
1078                 struct ldlm_resource *res;
1079                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1080                 res = ldlm_resource_getref(lock->l_resource);
1081                 LDLM_DEBUG(lock, "export %p", exp);
1082                 ldlm_lock_cancel(lock);
1083                 ldlm_reprocess_all(res);
1084                 ldlm_resource_putref(res);
1085         }
1086 }
1087
1088 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1089                                         int *flags)
1090 {
1091         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1092         struct ldlm_resource *res;
1093         struct ldlm_namespace *ns;
1094         int granted = 0;
1095         ENTRY;
1096
1097         LBUG();
1098
1099         res = lock->l_resource;
1100         ns = res->lr_namespace;
1101
1102         l_lock(&ns->ns_lock);
1103
1104         lock->l_req_mode = new_mode;
1105         ldlm_resource_unlink_lock(lock);
1106
1107         /* If this is a local resource, put it on the appropriate list. */
1108         if (res->lr_namespace->ns_client) {
1109                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1110                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1111                 } else {
1112                         /* This should never happen, because of the way the
1113                          * server handles conversions. */
1114                         LBUG();
1115
1116                         res->lr_tmp = &rpc_list;
1117                         ldlm_grant_lock(lock, NULL, 0);
1118                         res->lr_tmp = NULL;
1119                         granted = 1;
1120                         /* FIXME: completion handling not with ns_lock held ! */
1121                         if (lock->l_completion_ast)
1122                                 lock->l_completion_ast(lock, 0, NULL);
1123                 }
1124         } else {
1125                 /* FIXME: We should try the conversion right away and possibly
1126                  * return success without the need for an extra AST */
1127                 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1128                 *flags |= LDLM_FL_BLOCK_CONV;
1129         }
1130
1131         l_unlock(&ns->ns_lock);
1132
1133         if (granted)
1134                 ldlm_run_ast_work(&rpc_list);
1135         RETURN(res);
1136 }
1137
1138 void ldlm_lock_dump(int level, struct ldlm_lock *lock)
1139 {
1140         char ver[128];
1141         struct obd_device *obd;
1142
1143         if (!((portal_debug | D_ERROR) & level))
1144                 return;
1145
1146         if (RES_VERSION_SIZE != 4)
1147                 LBUG();
1148
1149         if (!lock) {
1150                 CDEBUG(level, "  NULL LDLM lock\n");
1151                 return;
1152         }
1153
1154         snprintf(ver, sizeof(ver), "%x %x %x %x",
1155                  lock->l_version[0], lock->l_version[1],
1156                  lock->l_version[2], lock->l_version[3]);
1157
1158         CDEBUG(level, "  -- Lock dump: %p (%s) (rc: %d)\n", lock, ver,
1159                atomic_read(&lock->l_refc));
1160         obd = class_conn2obd(lock->l_connh);
1161         if (lock->l_export && lock->l_export->exp_connection) {
1162                 CDEBUG(level, "  Node: NID "LPX64" on %s (rhandle: "LPX64")\n",
1163                        lock->l_export->exp_connection->c_peer.peer_nid,
1164                        lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1165                        lock->l_remote_handle.cookie);
1166         } else if (obd == NULL) {
1167                 CDEBUG(level, "  Node: local\n");
1168         } else {
1169                 struct obd_import *imp = obd->u.cli.cl_import;
1170                 CDEBUG(level, "  Node: NID "LPX64" on %s (rhandle: "LPX64")\n",
1171                        imp->imp_connection->c_peer.peer_nid,
1172                        imp->imp_connection->c_peer.peer_ni->pni_name,
1173                        lock->l_remote_handle.cookie);
1174         }
1175         CDEBUG(level, "  Parent: %p\n", lock->l_parent);
1176         CDEBUG(level, "  Resource: %p ("LPD64")\n", lock->l_resource,
1177                lock->l_resource->lr_name.name[0]);
1178         CDEBUG(level, "  Requested mode: %d, granted mode: %d\n",
1179                (int)lock->l_req_mode, (int)lock->l_granted_mode);
1180         CDEBUG(level, "  Readers: %u ; Writers; %u\n",
1181                lock->l_readers, lock->l_writers);
1182         if (lock->l_resource->lr_type == LDLM_EXTENT)
1183                 CDEBUG(level, "  Extent: "LPU64" -> "LPU64"\n",
1184                        lock->l_extent.start, lock->l_extent.end);
1185 }
1186
1187 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1188 {
1189         struct ldlm_lock *lock;
1190
1191         lock = ldlm_handle2lock(lockh);
1192         if (lock == NULL)
1193                 return;
1194
1195         ldlm_lock_dump(D_OTHER, lock);
1196
1197         LDLM_LOCK_PUT(lock);
1198 }