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                 if (lock->l_export)
 157                         class_export_put(lock->l_export);
 158
 159                 if (lock->l_parent)
 160                         LDLM_LOCK_PUT(lock->l_parent);
 161
 162                 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
 163                 l_unlock(&ns->ns_lock);
 164         }
 165
 166         EXIT;
 167 }
 168
 169 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
 170 {
 171         ENTRY;
 172         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 173         if (!list_empty(&lock->l_lru)) {
 174                 list_del_init(&lock->l_lru);
 175                 lock->l_resource->lr_namespace->ns_nr_unused--;
 176                 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
 177         }
 178         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 179         EXIT;
 180 }
 181
 182 /* This used to have a 'strict' flact, which recovery would use to mark an
 183  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
 184  * shall explain why it's gone: with the new hash table scheme, once you call
 185  * ldlm_lock_destroy, you can never drop your final references on this lock.
 186  * Because it's not in the hash table anymore.  -phil */
 187 void ldlm_lock_destroy(struct ldlm_lock *lock)
 188 {
 189         ENTRY;
 190         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 191
 192         if (!list_empty(&lock->l_children)) {
 193                 LDLM_ERROR(lock, "still has children (%p)!",
 194                            lock->l_children.next);
 195                 ldlm_lock_dump(D_ERROR, lock);
 196                 LBUG();
 197         }
 198         if (lock->l_readers || lock->l_writers) {
 199                 LDLM_ERROR(lock, "lock still has references");
 200                 ldlm_lock_dump(D_ERROR, lock);
 201                 LBUG();
 202         }
 203
 204         if (!list_empty(&lock->l_res_link)) {
 205                 ldlm_lock_dump(D_ERROR, lock);
 206                 LBUG();
 207         }
 208
 209         if (lock->l_destroyed) {
 210                 LASSERT(list_empty(&lock->l_lru));
 211                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 212                 EXIT;
 213                 return;
 214         }
 215         lock->l_destroyed = 1;
 216
 217         list_del_init(&lock->l_export_chain);
 218         ldlm_lock_remove_from_lru(lock);
 219         class_handle_unhash(&lock->l_handle);
 220
 221 #if 0
 222         /* Wake anyone waiting for this lock */
 223         /* FIXME: I should probably add yet another flag, instead of using
 224          * l_export to only call this on clients */
 225         if (lock->l_export)
 226                 class_export_put(lock->l_export);
 227         lock->l_export = NULL;
 228         if (lock->l_export && lock->l_completion_ast)
 229                 lock->l_completion_ast(lock, 0);
 230 #endif
 231
 232         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 233         LDLM_LOCK_PUT(lock);
 234         EXIT;
 235 }
 236
 237 /* this is called by portals_handle2object with the handle lock taken */
 238 static void lock_handle_addref(void *lock)
 239 {
 240         LDLM_LOCK_GET((struct ldlm_lock *)lock);
 241 }
 242
 243 /*
 244  * usage: pass in a resource on which you have done ldlm_resource_get
 245  *        pass in a parent lock on which you have done a ldlm_lock_get
 246  *        after return, ldlm_*_put the resource and parent
 247  * returns: lock with refcount 1
 248  */
 249 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 250                                        struct ldlm_resource *resource)
 251 {
 252         struct ldlm_lock *lock;
 253         ENTRY;
 254
 255         if (resource == NULL)
 256                 LBUG();
 257
 258         OBD_SLAB_ALLOC(lock, ldlm_lock_slab, SLAB_KERNEL, sizeof(*lock));
 259         if (lock == NULL)
 260                 RETURN(NULL);
 261
 262         lock->l_resource = ldlm_resource_getref(resource);
 263
 264         atomic_set(&lock->l_refc, 2);
 265         INIT_LIST_HEAD(&lock->l_children);
 266         INIT_LIST_HEAD(&lock->l_res_link);
 267         INIT_LIST_HEAD(&lock->l_lru);
 268         INIT_LIST_HEAD(&lock->l_export_chain);
 269         INIT_LIST_HEAD(&lock->l_pending_chain);
 270         init_waitqueue_head(&lock->l_waitq);
 271
 272         spin_lock(&resource->lr_namespace->ns_counter_lock);
 273         resource->lr_namespace->ns_locks++;
 274         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 275
 276         if (parent != NULL) {
 277                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 278                 lock->l_parent = LDLM_LOCK_GET(parent);
 279                 list_add(&lock->l_childof, &parent->l_children);
 280                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 281         }
 282
 283         INIT_LIST_HEAD(&lock->l_handle.h_link);
 284         class_handle_hash(&lock->l_handle, lock_handle_addref);
 285
 286         RETURN(lock);
 287 }
 288
 289 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
 290                               struct ldlm_res_id new_resid)
 291 {
 292         struct ldlm_resource *oldres = lock->l_resource;
 293         ENTRY;
 294
 295         l_lock(&ns->ns_lock);
 296         if (memcmp(&new_resid, &lock->l_resource->lr_name,
 297                    sizeof(lock->l_resource->lr_name)) == 0) {
 298                 /* Nothing to do */
 299                 l_unlock(&ns->ns_lock);
 300                 RETURN(0);
 301         }
 302
 303         LASSERT(new_resid.name[0] != 0);
 304
 305         /* This function assumes that the lock isn't on any lists */
 306         LASSERT(list_empty(&lock->l_res_link));
 307
 308         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
 309                                              lock->l_resource->lr_type, 1);
 310         if (lock->l_resource == NULL) {
 311                 LBUG();
 312                 RETURN(-ENOMEM);
 313         }
 314
 315         /* ...and the flowers are still standing! */
 316         ldlm_resource_putref(oldres);
 317
 318         l_unlock(&ns->ns_lock);
 319         RETURN(0);
 320 }
 321
 322 /*
 323  *  HANDLES
 324  */
 325
 326 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 327 {
 328         lockh->cookie = lock->l_handle.h_cookie;
 329 }
 330
 331 /* if flags: atomically get the lock and set the flags.
 332  *           Return NULL if flag already set
 333  */
 334
 335 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
 336 {
 337         struct ldlm_namespace *ns;
 338         struct ldlm_lock *lock = NULL, *retval = NULL;
 339         ENTRY;
 340
 341         LASSERT(handle);
 342
 343         lock = class_handle2object(handle->cookie);
 344         if (lock == NULL)
 345                 RETURN(NULL);
 346
 347         LASSERT(lock->l_resource != NULL);
 348         ns = lock->l_resource->lr_namespace;
 349         LASSERT(ns != NULL);
 350
 351         l_lock(&ns->ns_lock);
 352
 353         /* It's unlikely but possible that someone marked the lock as
 354          * destroyed after we did handle2object on it */
 355         if (lock->l_destroyed) {
 356                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
 357                 LDLM_LOCK_PUT(lock);
 358                 GOTO(out, retval);
 359         }
 360
 361         if (flags && (lock->l_flags & flags)) {
 362                 LDLM_LOCK_PUT(lock);
 363                 GOTO(out, retval);
 364         }
 365
 366         if (flags)
 367                 lock->l_flags |= flags;
 368
 369         retval = lock;
 370         EXIT;
 371  out:
 372         l_unlock(&ns->ns_lock);
 373         return retval;
 374 }
 375
 376 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
 377                                       struct lustre_handle *handle)
 378 {
 379         struct ldlm_lock *retval = NULL;
 380
 381         l_lock(&ns->ns_lock);
 382         retval = __ldlm_handle2lock(handle, 0);
 383         l_unlock(&ns->ns_lock);
 384
 385         return retval;
 386 }
 387
 388 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 389 {
 390         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 391 }
 392
 393 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 394 {
 395         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 396         desc->l_req_mode = lock->l_req_mode;
 397         desc->l_granted_mode = lock->l_granted_mode;
 398         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 399         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 400 }
 401
 402 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 403                                    struct ldlm_lock *new,
 404                                    void *data, int datalen)
 405 {
 406         struct ldlm_ast_work *w;
 407         ENTRY;
 408
 409         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 410         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 411                 GOTO(out, 0);
 412
 413         OBD_ALLOC(w, sizeof(*w));
 414         if (!w) {
 415                 LBUG();
 416                 GOTO(out, 0);
 417         }
 418
 419         w->w_data = data;
 420         w->w_datalen = datalen;
 421         if (new) {
 422                 lock->l_flags |= LDLM_FL_AST_SENT;
 423                 w->w_blocking = 1;
 424                 ldlm_lock2desc(new, &w->w_desc);
 425         }
 426
 427         w->w_lock = LDLM_LOCK_GET(lock);
 428         list_add(&w->w_list, lock->l_resource->lr_tmp);
 429         EXIT;
 430  out:
 431         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 432         return;
 433 }
 434
 435 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 436 {
 437         struct ldlm_lock *lock;
 438
 439         lock = ldlm_handle2lock(lockh);
 440         ldlm_lock_addref_internal(lock, mode);
 441         LDLM_LOCK_PUT(lock);
 442 }
 443
 444 /* only called for local locks */
 445 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 446 {
 447         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 448         ldlm_lock_remove_from_lru(lock);
 449         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 450                 lock->l_readers++;
 451         else
 452                 lock->l_writers++;
 453         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 454         LDLM_LOCK_GET(lock);
 455         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 456 }
 457
 458 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
 459 {
 460         struct ldlm_namespace *ns;
 461         ENTRY;
 462
 463         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 464         ns = lock->l_resource->lr_namespace;
 465         l_lock(&ns->ns_lock);
 466         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR) {
 467                 LASSERT(lock->l_readers > 0);
 468                 lock->l_readers--;
 469         } else {
 470                 LASSERT(lock->l_writers > 0);
 471                 lock->l_writers--;
 472         }
 473
 474         if (lock->l_flags & LDLM_FL_LOCAL &&
 475             !lock->l_readers && !lock->l_writers) {
 476                 /* If this is a local lock on a server namespace and this was
 477                  * the last reference, cancel the lock. */
 478                 CDEBUG(D_INFO, "forcing cancel of local lock\n");
 479                 lock->l_flags |= LDLM_FL_CBPENDING;
 480         }
 481
 482         if (!lock->l_readers && !lock->l_writers &&
 483             (lock->l_flags & LDLM_FL_CBPENDING)) {
 484                 /* If we received a blocked AST and this was the last reference,
 485                  * run the callback. */
 486                 if (!ns->ns_client && lock->l_export)
 487                         CERROR("FL_CBPENDING set on non-local lock--just a "
 488                                "warning\n");
 489
 490                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 491                 l_unlock(&ns->ns_lock);
 492
 493                 /* FIXME: need a real 'desc' here */
 494                 if (lock->l_blocking_ast != NULL)
 495                         lock->l_blocking_ast(lock, NULL, lock->l_data,
 496                                              LDLM_CB_BLOCKING);
 497                 else
 498                         LDLM_DEBUG(lock, "No blocking AST?");
 499         } else if (ns->ns_client && !lock->l_readers && !lock->l_writers) {
 500                 /* If this is a client-side namespace and this was the last
 501                  * reference, put it on the LRU. */
 502                 LASSERT(list_empty(&lock->l_lru));
 503                 LASSERT(ns->ns_nr_unused >= 0);
 504                 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 505                 ns->ns_nr_unused++;
 506                 l_unlock(&ns->ns_lock);
 507                 ldlm_cancel_lru(ns);
 508         } else {
 509                 l_unlock(&ns->ns_lock);
 510         }
 511
 512         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 513
 514         EXIT;
 515 }
 516
 517 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 518 {
 519         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 520         LASSERT(lock != NULL);
 521         ldlm_lock_decref_internal(lock, mode);
 522         LDLM_LOCK_PUT(lock);
 523 }
 524
 525 /* This will drop a lock reference and mark it for destruction, but will not
 526  * necessarily cancel the lock before returning. */
 527 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
 528 {
 529         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 530         ENTRY;
 531
 532         LASSERT(lock != NULL);
 533
 534         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 535         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 536         lock->l_flags |= LDLM_FL_CBPENDING;
 537         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 538         ldlm_lock_decref_internal(lock, mode);
 539         LDLM_LOCK_PUT(lock);
 540 }
 541
 542 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 543                                  struct list_head *queue)
 544 {
 545         struct list_head *tmp, *pos;
 546         int rc = 1;
 547
 548         list_for_each_safe(tmp, pos, queue) {
 549                 struct ldlm_lock *child;
 550                 ldlm_res_compat compat;
 551
 552                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 553                 if (lock == child)
 554                         continue;
 555
 556                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 557                 if (compat && compat(child, lock)) {
 558                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 559                         continue;
 560                 }
 561                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 562                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 563                         continue;
 564                 }
 565
 566                 rc = 0;
 567
 568                 if (send_cbs && child->l_blocking_ast != NULL) {
 569                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 570                                "blocking AST.\n", child);
 571                         ldlm_add_ast_work_item(child, lock, NULL, 0);
 572                 }
 573         }
 574
 575         return rc;
 576 }
 577
 578 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 579 {
 580         int rc;
 581         ENTRY;
 582
 583         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 584         rc = ldlm_lock_compat_list(lock, send_cbs,
 585                                    &lock->l_resource->lr_granted);
 586         /* FIXME: should we be sending ASTs to converting? */
 587         if (rc)
 588                 rc = ldlm_lock_compat_list
 589                         (lock, send_cbs, &lock->l_resource->lr_converting);
 590
 591         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 592         RETURN(rc);
 593 }
 594
 595 /* NOTE: called by
 596  *  - ldlm_lock_enqueue
 597  *  - ldlm_reprocess_queue
 598  *  - ldlm_lock_convert
 599  */
 600 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen)
 601 {
 602         struct ldlm_resource *res = lock->l_resource;
 603         ENTRY;
 604
 605         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 606         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 607         lock->l_granted_mode = lock->l_req_mode;
 608
 609         if (lock->l_granted_mode < res->lr_most_restr)
 610                 res->lr_most_restr = lock->l_granted_mode;
 611
 612         if (lock->l_completion_ast != NULL)
 613                 ldlm_add_ast_work_item(lock, NULL, data, datalen);
 614
 615         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 616         EXIT;
 617 }
 618
 619 /* returns a referenced lock or NULL.  See the flag descriptions below, in the
 620  * comment above ldlm_lock_match */
 621 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 622                                       struct ldlm_extent *extent,
 623                                       struct ldlm_lock *old_lock, void *data,
 624                                       int flags)
 625 {
 626         struct ldlm_lock *lock;
 627         struct list_head *tmp;
 628
 629         list_for_each(tmp, queue) {
 630                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 631
 632                 if (lock == old_lock)
 633                         break;
 634
 635                 /* llite sometimes wants to match locks that will be
 636                  * canceled when their users drop, but we allow it to match
 637                  * if it passes in CBPENDING and the lock still has users.
 638                  * this is generally only going to be used by children
 639                  * whose parents already hold a lock so forward progress
 640                  * can still happen. */
 641                 if (lock->l_flags & LDLM_FL_CBPENDING &&
 642                     !(flags & LDLM_FL_CBPENDING))
 643                         continue;
 644                 if (lock->l_flags & LDLM_FL_CBPENDING &&
 645                     lock->l_readers == 0 && lock->l_writers == 0)
 646                         continue;
 647
 648                 if (lock->l_req_mode != mode)
 649                         continue;
 650
 651                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 652                     (lock->l_extent.start > extent->start ||
 653                      lock->l_extent.end < extent->end))
 654                         continue;
 655
 656                 if (lock->l_destroyed)
 657                         continue;
 658
 659                 if ((flags & LDLM_FL_LOCAL_ONLY) &&
 660                     !(lock->l_flags & LDLM_FL_LOCAL))
 661                         continue;
 662
 663                 if ((flags & LDLM_FL_MATCH_DATA) && lock->l_data != data) {
 664                         LDLM_DEBUG(lock, "data mismatch: have %p, want %p",
 665                                    lock->l_data, data);
 666                         continue;
 667                 }
 668
 669                 ldlm_lock_addref_internal(lock, mode);
 670                 return lock;
 671         }
 672
 673         return NULL;
 674 }
 675
 676 /* Can be called in two ways:
 677  *
 678  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
 679  * for a duplicate of.
 680  *
 681  * Otherwise, all of the fields must be filled in, to match against.
 682  *
 683  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
 684  *     server (ie, connh is NULL)
 685  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
 686  *     list will be considered
 687  * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
 688  *     to be canceled can still be matched as long as they still have reader
 689  *     or writer refernces
 690  * If 'flags' contains LDLM_FL_MATCH_DATA, then only match a lock if the opaque
 691  *     data is the same.
 692  *
 693  * Returns 1 if it finds an already-existing lock that is compatible; in this
 694  * case, lockh is filled in with a addref()ed lock
 695  */
 696 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
 697                     struct ldlm_res_id *res_id, __u32 type, void *cookie,
 698                     int cookielen, ldlm_mode_t mode, void *data,
 699                     struct lustre_handle *lockh)
 700 {
 701         struct ldlm_resource *res;
 702         struct ldlm_lock *lock, *old_lock = NULL;
 703         int rc = 0;
 704         ENTRY;
 705
 706         if (ns == NULL) {
 707                 old_lock = ldlm_handle2lock(lockh);
 708                 LASSERT(old_lock);
 709
 710                 ns = old_lock->l_resource->lr_namespace;
 711                 res_id = &old_lock->l_resource->lr_name;
 712                 type = old_lock->l_resource->lr_type;
 713                 mode = old_lock->l_req_mode;
 714         }
 715
 716         res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
 717         if (res == NULL) {
 718                 LASSERT(old_lock == NULL);
 719                 RETURN(0);
 720         }
 721
 722         l_lock(&ns->ns_lock);
 723
 724         lock = search_queue(&res->lr_granted, mode, cookie, old_lock, data,
 725                             flags);
 726         if (lock != NULL)
 727                 GOTO(out, rc = 1);
 728         if (flags & LDLM_FL_BLOCK_GRANTED)
 729                 GOTO(out, rc = 0);
 730         lock = search_queue(&res->lr_converting, mode, cookie, old_lock, data,
 731                             flags);
 732         if (lock != NULL)
 733                 GOTO(out, rc = 1);
 734         lock = search_queue(&res->lr_waiting, mode, cookie, old_lock, data,
 735                             flags);
 736         if (lock != NULL)
 737                 GOTO(out, rc = 1);
 738
 739         EXIT;
 740  out:
 741         ldlm_resource_putref(res);
 742         l_unlock(&ns->ns_lock);
 743
 744         if (lock) {
 745                 ldlm_lock2handle(lock, lockh);
 746                 if (lock->l_completion_ast)
 747                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 748                                                NULL);
 749         }
 750         if (rc)
 751                 LDLM_DEBUG(lock, "matched");
 752         else
 753                 LDLM_DEBUG_NOLOCK("not matched");
 754
 755         if (old_lock)
 756                 LDLM_LOCK_PUT(old_lock);
 757
 758         return rc;
 759 }
 760
 761 /* Returns a referenced lock */
 762 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 763                                    struct lustre_handle *parent_lock_handle,
 764                                    struct ldlm_res_id res_id, __u32 type,
 765                                    ldlm_mode_t mode,
 766                                    ldlm_blocking_callback blocking,
 767                                    void *data)
 768 {
 769         struct ldlm_resource *res, *parent_res = NULL;
 770         struct ldlm_lock *lock, *parent_lock = NULL;
 771         ENTRY;
 772
 773         if (parent_lock_handle) {
 774                 parent_lock = ldlm_handle2lock(parent_lock_handle);
 775                 if (parent_lock)
 776                         parent_res = parent_lock->l_resource;
 777         }
 778
 779         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 780         if (res == NULL)
 781                 RETURN(NULL);
 782
 783         lock = ldlm_lock_new(parent_lock, res);
 784         ldlm_resource_putref(res);
 785         if (parent_lock != NULL)
 786                 LDLM_LOCK_PUT(parent_lock);
 787
 788         if (lock == NULL)
 789                 RETURN(NULL);
 790
 791         lock->l_req_mode = mode;
 792         lock->l_data = data;
 793         lock->l_blocking_ast = blocking;
 794
 795         RETURN(lock);
 796 }
 797
 798 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
 799                                struct ldlm_lock **lockp,
 800                                void *cookie, int cookie_len,
 801                                int *flags,
 802                                ldlm_completion_callback completion)
 803 {
 804         struct ldlm_resource *res;
 805         struct ldlm_lock *lock = *lockp;
 806         int local;
 807         ldlm_res_policy policy;
 808         ENTRY;
 809
 810         res = lock->l_resource;
 811
 812         if (res->lr_type == LDLM_EXTENT)
 813                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 814
 815         /* policies are not executed on the client or during replay */
 816         local = res->lr_namespace->ns_client;
 817         if (!local && !(*flags & LDLM_FL_REPLAY) &&
 818             (policy = ldlm_res_policy_table[res->lr_type])) {
 819                 int rc;
 820                 rc = policy(ns, lockp, cookie, lock->l_req_mode, *flags, NULL);
 821                 if (rc == ELDLM_LOCK_CHANGED) {
 822                         res = lock->l_resource;
 823                         *flags |= LDLM_FL_LOCK_CHANGED;
 824                 } else if (rc == ELDLM_LOCK_REPLACED) {
 825                         /* The lock that was returned has already been granted,
 826                          * and placed into lockp.  Destroy the old one and our
 827                          * work here is done. */
 828                         ldlm_lock_destroy(lock);
 829                         LDLM_LOCK_PUT(lock);
 830                         *flags |= LDLM_FL_LOCK_CHANGED;
 831                         RETURN(0);
 832                 } else if (rc == ELDLM_LOCK_ABORTED) {
 833                         ldlm_lock_destroy(lock);
 834                         RETURN(rc);
 835                 }
 836         }
 837
 838         l_lock(&ns->ns_lock);
 839         if (local && lock->l_req_mode == lock->l_granted_mode) {
 840                 /* The server returned a blocked lock, but it was granted before
 841                  * we got a chance to actually enqueue it.  We don't need to do
 842                  * anything else. */
 843                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 844                             LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 845                 GOTO(out, ELDLM_OK);
 846         }
 847
 848         /* This distinction between local lock trees is very important; a client
 849          * namespace only has information about locks taken by that client, and
 850          * thus doesn't have enough information to decide for itself if it can
 851          * be granted (below).  In this case, we do exactly what the server
 852          * tells us to do, as dictated by the 'flags'.
 853          *
 854          * We do exactly the same thing during recovery, when the server is
 855          * more or less trusting the clients not to lie.
 856          *
 857          * FIXME (bug 268): Detect obvious lies by checking compatibility in
 858          * granted/converting queues. */
 859         ldlm_resource_unlink_lock(lock);
 860         if (local) {
 861                 if (*flags & LDLM_FL_BLOCK_CONV)
 862                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 863                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 864                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 865                 else
 866                         ldlm_grant_lock(lock, NULL, 0);
 867                 GOTO(out, ELDLM_OK);
 868         } else if (*flags & LDLM_FL_REPLAY) {
 869                 if (*flags & LDLM_FL_BLOCK_CONV) {
 870                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 871                         GOTO(out, ELDLM_OK);
 872                 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
 873                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 874                         GOTO(out, ELDLM_OK);
 875                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
 876                         ldlm_grant_lock(lock, NULL, 0);
 877                         GOTO(out, ELDLM_OK);
 878                 }
 879                 /* If no flags, fall through to normal enqueue path. */
 880         }
 881
 882         /* FIXME: We may want to optimize by checking lr_most_restr */
 883         if (!list_empty(&res->lr_converting)) {
 884                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 885                 *flags |= LDLM_FL_BLOCK_CONV;
 886                 GOTO(out, ELDLM_OK);
 887         }
 888         if (!list_empty(&res->lr_waiting)) {
 889                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 890                 *flags |= LDLM_FL_BLOCK_WAIT;
 891                 GOTO(out, ELDLM_OK);
 892         }
 893         if (!ldlm_lock_compat(lock, 0)) {
 894                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 895                 *flags |= LDLM_FL_BLOCK_GRANTED;
 896                 GOTO(out, ELDLM_OK);
 897         }
 898         ldlm_grant_lock(lock, NULL, 0);
 899         EXIT;
 900       out:
 901         /* Don't set 'completion_ast' until here so that if the lock is granted
 902          * immediately we don't do an unnecessary completion call. */
 903         lock->l_completion_ast = completion;
 904         l_unlock(&ns->ns_lock);
 905         return ELDLM_OK;
 906 }
 907
 908 /* Must be called with namespace taken: queue is waiting or converting. */
 909 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 910                                 struct list_head *queue)
 911 {
 912         struct list_head *tmp, *pos;
 913         ENTRY;
 914
 915         list_for_each_safe(tmp, pos, queue) {
 916                 struct ldlm_lock *pending;
 917                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 918
 919                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 920
 921                 if (!ldlm_lock_compat(pending, 1))
 922                         RETURN(1);
 923
 924                 list_del_init(&pending->l_res_link);
 925                 ldlm_grant_lock(pending, NULL, 0);
 926         }
 927
 928         RETURN(0);
 929 }
 930
 931 int ldlm_run_ast_work(struct list_head *rpc_list)
 932 {
 933         struct list_head *tmp, *pos;
 934         int rc, retval = 0;
 935         ENTRY;
 936
 937         list_for_each_safe(tmp, pos, rpc_list) {
 938                 struct ldlm_ast_work *w =
 939                         list_entry(tmp, struct ldlm_ast_work, w_list);
 940
 941                 /* It's possible to receive a completion AST before we've set
 942                  * the l_completion_ast pointer: either because the AST arrived
 943                  * before the reply, or simply because there's a small race
 944                  * window between receiving the reply and finishing the local
 945                  * enqueue. (bug 842)
 946                  *
 947                  * This can't happen with the blocking_ast, however, because we
 948                  * will never call the local blocking_ast until we drop our
 949                  * reader/writer reference, which we won't do until we get the
 950                  * reply and finish enqueueing. */
 951                 if (w->w_blocking) {
 952                         LASSERT(w->w_lock->l_blocking_ast != NULL);
 953                         rc = w->w_lock->l_blocking_ast
 954                                 (w->w_lock, &w->w_desc, w->w_data,
 955                                  LDLM_CB_BLOCKING);
 956                 } else if (w->w_lock->l_completion_ast != NULL) {
 957                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags,
 958                                                          w->w_data);
 959                 } else {
 960                         rc = 0;
 961                 }
 962                 if (rc == -ERESTART)
 963                         retval = rc;
 964                 else if (rc)
 965                         CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
 966                                "disconnect client\n");
 967                 LDLM_LOCK_PUT(w->w_lock);
 968                 list_del(&w->w_list);
 969                 OBD_FREE(w, sizeof(*w));
 970         }
 971         RETURN(retval);
 972 }
 973
 974 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
 975 {
 976         ldlm_reprocess_all(res);
 977         return LDLM_ITER_CONTINUE;
 978 }
 979
 980 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
 981 {
 982         (void)ldlm_namespace_foreach_res(ns, reprocess_one_queue, NULL);
 983 }
 984
 985 void ldlm_reprocess_all(struct ldlm_resource *res)
 986 {
 987         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 988         int rc;
 989         ENTRY;
 990
 991         /* Local lock trees don't get reprocessed. */
 992         if (res->lr_namespace->ns_client) {
 993                 EXIT;
 994                 return;
 995         }
 996
 997  restart:
 998         l_lock(&res->lr_namespace->ns_lock);
 999         res->lr_tmp = &rpc_list;
1000
1001         ldlm_reprocess_queue(res, &res->lr_converting);
1002         if (list_empty(&res->lr_converting))
1003                 ldlm_reprocess_queue(res, &res->lr_waiting);
1004
1005         res->lr_tmp = NULL;
1006         l_unlock(&res->lr_namespace->ns_lock);
1007
1008         rc = ldlm_run_ast_work(&rpc_list);
1009         if (rc == -ERESTART)
1010                 goto restart;
1011         EXIT;
1012 }
1013
1014 void ldlm_cancel_callback(struct ldlm_lock *lock)
1015 {
1016         l_lock(&lock->l_resource->lr_namespace->ns_lock);
1017         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1018                 lock->l_flags |= LDLM_FL_CANCEL;
1019                 if (lock->l_blocking_ast) {
1020                         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1021                         lock->l_blocking_ast(lock, NULL, lock->l_data,
1022                                              LDLM_CB_CANCELING);
1023                         return;
1024                 } else {
1025                         LDLM_DEBUG(lock, "no blocking ast");
1026                 }
1027         }
1028         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1029 }
1030
1031 void ldlm_lock_cancel(struct ldlm_lock *lock)
1032 {
1033         struct ldlm_resource *res;
1034         struct ldlm_namespace *ns;
1035         ENTRY;
1036
1037         ldlm_del_waiting_lock(lock);
1038
1039         res = lock->l_resource;
1040         ns = res->lr_namespace;
1041
1042         l_lock(&ns->ns_lock);
1043         /* Please do not, no matter how tempting, remove this LBUG without
1044          * talking to me first. -phik */
1045         if (lock->l_readers || lock->l_writers) {
1046                 LDLM_DEBUG(lock, "lock still has references");
1047                 ldlm_lock_dump(D_OTHER, lock);
1048                 LBUG();
1049         }
1050
1051         ldlm_cancel_callback(lock); /* XXX FIXME bug 1030 */
1052
1053         ldlm_resource_unlink_lock(lock);
1054         ldlm_lock_destroy(lock);
1055         l_unlock(&ns->ns_lock);
1056         EXIT;
1057 }
1058
1059 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1060 {
1061         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1062         ENTRY;
1063
1064         if (lock == NULL)
1065                 RETURN(-EINVAL);
1066
1067         lock->l_data = data;
1068
1069         LDLM_LOCK_PUT(lock);
1070
1071         RETURN(0);
1072 }
1073
1074 /* This function is only called from one thread (per export); no locking around
1075  * the list ops needed */
1076 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1077 {
1078         struct list_head *iter, *n;
1079
1080         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
1081                 struct ldlm_lock *lock;
1082                 struct ldlm_resource *res;
1083                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1084                 res = ldlm_resource_getref(lock->l_resource);
1085                 LDLM_DEBUG(lock, "export %p", exp);
1086                 ldlm_lock_cancel(lock);
1087                 ldlm_reprocess_all(res);
1088                 ldlm_resource_putref(res);
1089         }
1090 }
1091
1092 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1093                                         int *flags)
1094 {
1095         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1096         struct ldlm_resource *res;
1097         struct ldlm_namespace *ns;
1098         int granted = 0;
1099         ENTRY;
1100
1101         LBUG();
1102
1103         res = lock->l_resource;
1104         ns = res->lr_namespace;
1105
1106         l_lock(&ns->ns_lock);
1107
1108         lock->l_req_mode = new_mode;
1109         ldlm_resource_unlink_lock(lock);
1110
1111         /* If this is a local resource, put it on the appropriate list. */
1112         if (res->lr_namespace->ns_client) {
1113                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1114                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1115                 } else {
1116                         /* This should never happen, because of the way the
1117                          * server handles conversions. */
1118                         LBUG();
1119
1120                         res->lr_tmp = &rpc_list;
1121                         ldlm_grant_lock(lock, NULL, 0);
1122                         res->lr_tmp = NULL;
1123                         granted = 1;
1124                         /* FIXME: completion handling not with ns_lock held ! */
1125                         if (lock->l_completion_ast)
1126                                 lock->l_completion_ast(lock, 0, NULL);
1127                 }
1128         } else {
1129                 /* FIXME: We should try the conversion right away and possibly
1130                  * return success without the need for an extra AST */
1131                 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1132                 *flags |= LDLM_FL_BLOCK_CONV;
1133         }
1134
1135         l_unlock(&ns->ns_lock);
1136
1137         if (granted)
1138                 ldlm_run_ast_work(&rpc_list);
1139         RETURN(res);
1140 }
1141
1142 void ldlm_lock_dump(int level, struct ldlm_lock *lock)
1143 {
1144         char ver[128];
1145         struct obd_device *obd;
1146
1147         if (!((portal_debug | D_ERROR) & level))
1148                 return;
1149
1150         if (RES_VERSION_SIZE != 4)
1151                 LBUG();
1152
1153         if (!lock) {
1154                 CDEBUG(level, "  NULL LDLM lock\n");
1155                 return;
1156         }
1157
1158         snprintf(ver, sizeof(ver), "%x %x %x %x",
1159                  lock->l_version[0], lock->l_version[1],
1160                  lock->l_version[2], lock->l_version[3]);
1161
1162         CDEBUG(level, "  -- Lock dump: %p (%s) (rc: %d)\n", lock, ver,
1163                atomic_read(&lock->l_refc));
1164         obd = class_conn2obd(lock->l_connh);
1165         if (lock->l_export && lock->l_export->exp_connection) {
1166                 CDEBUG(level, "  Node: NID "LPX64" on %s (rhandle: "LPX64")\n",
1167                        lock->l_export->exp_connection->c_peer.peer_nid,
1168                        lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1169                        lock->l_remote_handle.cookie);
1170         } else if (obd == NULL) {
1171                 CDEBUG(level, "  Node: local\n");
1172         } else {
1173                 struct obd_import *imp = obd->u.cli.cl_import;
1174                 CDEBUG(level, "  Node: NID "LPX64" on %s (rhandle: "LPX64")\n",
1175                        imp->imp_connection->c_peer.peer_nid,
1176                        imp->imp_connection->c_peer.peer_ni->pni_name,
1177                        lock->l_remote_handle.cookie);
1178         }
1179         CDEBUG(level, "  Parent: %p\n", lock->l_parent);
1180         CDEBUG(level, "  Resource: %p ("LPD64")\n", lock->l_resource,
1181                lock->l_resource->lr_name.name[0]);
1182         CDEBUG(level, "  Requested mode: %d, granted mode: %d\n",
1183                (int)lock->l_req_mode, (int)lock->l_granted_mode);
1184         CDEBUG(level, "  Readers: %u ; Writers; %u\n",
1185                lock->l_readers, lock->l_writers);
1186         if (lock->l_resource->lr_type == LDLM_EXTENT)
1187                 CDEBUG(level, "  Extent: "LPU64" -> "LPU64"\n",
1188                        lock->l_extent.start, lock->l_extent.end);
1189 }
1190
1191 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1192 {
1193         struct ldlm_lock *lock;
1194
1195         lock = ldlm_handle2lock(lockh);
1196         if (lock == NULL)
1197                 return;
1198
1199         ldlm_lock_dump(D_OTHER, lock);
1200
1201         LDLM_LOCK_PUT(lock);
1202 }