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         default:
  75                 CERROR("Unknown intent %d\n", it);
  76                 return "UNKNOWN";
  77         }
  78 }
  79 #endif
  80
  81 extern kmem_cache_t *ldlm_lock_slab;
  82 struct lustre_lock ldlm_handle_lock;
  83
  84 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b);
  85
  86 static ldlm_res_compat ldlm_res_compat_table[] = {
  87         [LDLM_PLAIN] ldlm_plain_compat,
  88         [LDLM_EXTENT] ldlm_extent_compat,
  89 };
  90
  91 static ldlm_res_policy ldlm_intent_policy_func;
  92
  93 static int ldlm_plain_policy(struct ldlm_namespace *ns, struct ldlm_lock **lock,
  94                              void *req_cookie, ldlm_mode_t mode, int flags,
  95                              void *data)
  96 {
  97         if ((flags & LDLM_FL_HAS_INTENT) && ldlm_intent_policy_func) {
  98                 return ldlm_intent_policy_func(ns, lock, req_cookie, mode,
  99                                                flags, data);
 100         }
 101
 102         return ELDLM_OK;
 103 }
 104
 105 static ldlm_res_policy ldlm_res_policy_table[] = {
 106         [LDLM_PLAIN] ldlm_plain_policy,
 107         [LDLM_EXTENT] ldlm_extent_policy,
 108 };
 109
 110 void ldlm_register_intent(ldlm_res_policy arg)
 111 {
 112         ldlm_intent_policy_func = arg;
 113 }
 114
 115 void ldlm_unregister_intent(void)
 116 {
 117         ldlm_intent_policy_func = NULL;
 118 }
 119
 120 /*
 121  * REFCOUNTED LOCK OBJECTS
 122  */
 123
 124
 125 /*
 126  * Lock refcounts, during creation:
 127  *   - one special one for allocation, dec'd only once in destroy
 128  *   - one for being a lock that's in-use
 129  *   - one for the addref associated with a new lock
 130  */
 131 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 132 {
 133         atomic_inc(&lock->l_refc);
 134         return lock;
 135 }
 136
 137 void ldlm_lock_put(struct ldlm_lock *lock)
 138 {
 139         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 140         ENTRY;
 141
 142         if (atomic_dec_and_test(&lock->l_refc)) {
 143                 l_lock(&ns->ns_lock);
 144                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
 145                 LASSERT(lock->l_destroyed);
 146                 LASSERT(list_empty(&lock->l_res_link));
 147
 148                 spin_lock(&ns->ns_counter_lock);
 149                 ns->ns_locks--;
 150                 spin_unlock(&ns->ns_counter_lock);
 151
 152                 ldlm_resource_putref(lock->l_resource);
 153                 lock->l_resource = NULL;
 154
 155                 if (lock->l_parent)
 156                         LDLM_LOCK_PUT(lock->l_parent);
 157
 158                 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
 159                 l_unlock(&ns->ns_lock);
 160         }
 161
 162         EXIT;
 163 }
 164
 165 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
 166 {
 167         ENTRY;
 168         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 169         if (!list_empty(&lock->l_lru)) {
 170                 list_del_init(&lock->l_lru);
 171                 lock->l_resource->lr_namespace->ns_nr_unused--;
 172                 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
 173         }
 174         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 175         EXIT;
 176 }
 177
 178 /* This used to have a 'strict' flact, which recovery would use to mark an
 179  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
 180  * shall explain why it's gone: with the new hash table scheme, once you call
 181  * ldlm_lock_destroy, you can never drop your final references on this lock.
 182  * Because it's not in the hash table anymore.  -phil */
 183 void ldlm_lock_destroy(struct ldlm_lock *lock)
 184 {
 185         ENTRY;
 186         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 187
 188         if (!list_empty(&lock->l_children)) {
 189                 LDLM_ERROR(lock, "still has children (%p)!",
 190                            lock->l_children.next);
 191                 ldlm_lock_dump(D_ERROR, lock);
 192                 LBUG();
 193         }
 194         if (lock->l_readers || lock->l_writers) {
 195                 LDLM_ERROR(lock, "lock still has references");
 196                 ldlm_lock_dump(D_ERROR, lock);
 197                 LBUG();
 198         }
 199
 200         if (!list_empty(&lock->l_res_link)) {
 201                 ldlm_lock_dump(D_ERROR, lock);
 202                 LBUG();
 203         }
 204
 205         if (lock->l_destroyed) {
 206                 LASSERT(list_empty(&lock->l_lru));
 207                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 208                 EXIT;
 209                 return;
 210         }
 211         lock->l_destroyed = 1;
 212
 213         list_del_init(&lock->l_export_chain);
 214         ldlm_lock_remove_from_lru(lock);
 215         class_handle_unhash(&lock->l_handle);
 216
 217 #if 0
 218         /* Wake anyone waiting for this lock */
 219         /* FIXME: I should probably add yet another flag, instead of using
 220          * l_export to only call this on clients */
 221         lock->l_export = NULL;
 222         if (lock->l_export && lock->l_completion_ast)
 223                 lock->l_completion_ast(lock, 0);
 224 #endif
 225
 226         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 227         LDLM_LOCK_PUT(lock);
 228         EXIT;
 229 }
 230
 231 /* this is called by portals_handle2object with the handle lock taken */
 232 static void lock_handle_addref(void *lock)
 233 {
 234         LDLM_LOCK_GET((struct ldlm_lock *)lock);
 235 }
 236
 237 /*
 238  * usage: pass in a resource on which you have done ldlm_resource_get
 239  *        pass in a parent lock on which you have done a ldlm_lock_get
 240  *        after return, ldlm_*_put the resource and parent
 241  * returns: lock with refcount 1
 242  */
 243 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 244                                        struct ldlm_resource *resource)
 245 {
 246         struct ldlm_lock *lock;
 247         ENTRY;
 248
 249         if (resource == NULL)
 250                 LBUG();
 251
 252         OBD_SLAB_ALLOC(lock, ldlm_lock_slab, SLAB_KERNEL, sizeof(*lock));
 253         if (lock == NULL)
 254                 RETURN(NULL);
 255
 256         lock->l_resource = ldlm_resource_getref(resource);
 257
 258         atomic_set(&lock->l_refc, 2);
 259         INIT_LIST_HEAD(&lock->l_children);
 260         INIT_LIST_HEAD(&lock->l_res_link);
 261         INIT_LIST_HEAD(&lock->l_lru);
 262         INIT_LIST_HEAD(&lock->l_export_chain);
 263         INIT_LIST_HEAD(&lock->l_pending_chain);
 264         init_waitqueue_head(&lock->l_waitq);
 265
 266         spin_lock(&resource->lr_namespace->ns_counter_lock);
 267         resource->lr_namespace->ns_locks++;
 268         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 269
 270         if (parent != NULL) {
 271                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 272                 lock->l_parent = LDLM_LOCK_GET(parent);
 273                 list_add(&lock->l_childof, &parent->l_children);
 274                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 275         }
 276
 277         INIT_LIST_HEAD(&lock->l_handle.h_link);
 278         class_handle_hash(&lock->l_handle, lock_handle_addref);
 279
 280         RETURN(lock);
 281 }
 282
 283 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
 284                               struct ldlm_res_id new_resid)
 285 {
 286         struct ldlm_resource *oldres = lock->l_resource;
 287         ENTRY;
 288
 289         l_lock(&ns->ns_lock);
 290         if (memcmp(&new_resid, &lock->l_resource->lr_name,
 291                    sizeof(lock->l_resource->lr_name)) == 0) {
 292                 /* Nothing to do */
 293                 l_unlock(&ns->ns_lock);
 294                 RETURN(0);
 295         }
 296
 297         LASSERT(new_resid.name[0] != 0);
 298
 299         /* This function assumes that the lock isn't on any lists */
 300         LASSERT(list_empty(&lock->l_res_link));
 301
 302         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
 303                                              lock->l_resource->lr_type, 1);
 304         if (lock->l_resource == NULL) {
 305                 LBUG();
 306                 RETURN(-ENOMEM);
 307         }
 308
 309         /* ...and the flowers are still standing! */
 310         ldlm_resource_putref(oldres);
 311
 312         l_unlock(&ns->ns_lock);
 313         RETURN(0);
 314 }
 315
 316 /*
 317  *  HANDLES
 318  */
 319
 320 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 321 {
 322         lockh->cookie = lock->l_handle.h_cookie;
 323 }
 324
 325 /* if flags: atomically get the lock and set the flags.
 326  *           Return NULL if flag already set
 327  */
 328
 329 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
 330 {
 331         struct ldlm_lock *lock = NULL, *retval = NULL;
 332         ENTRY;
 333
 334         LASSERT(handle);
 335
 336         lock = class_handle2object(handle->cookie);
 337         if (lock == NULL)
 338                 RETURN(NULL);
 339
 340         LASSERT(lock->l_resource != NULL);
 341         LASSERT(lock->l_resource->lr_namespace != NULL);
 342
 343         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 344
 345         /* It's unlikely but possible that someone marked the lock as
 346          * destroyed after we did handle2object on it */
 347         if (lock->l_destroyed) {
 348                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
 349                 LDLM_LOCK_PUT(lock);
 350                 GOTO(out, retval);
 351         }
 352
 353         if (flags && (lock->l_flags & flags)) {
 354                 LDLM_LOCK_PUT(lock);
 355                 GOTO(out, retval);
 356         }
 357
 358         if (flags)
 359                 lock->l_flags |= flags;
 360
 361         retval = lock;
 362         EXIT;
 363  out:
 364         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 365         return retval;
 366 }
 367
 368 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
 369                                       struct lustre_handle *handle)
 370 {
 371         struct ldlm_lock *retval = NULL;
 372
 373         l_lock(&ns->ns_lock);
 374         retval = __ldlm_handle2lock(handle, 0);
 375         l_unlock(&ns->ns_lock);
 376
 377         return retval;
 378 }
 379
 380 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 381 {
 382         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 383 }
 384
 385 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 386 {
 387         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 388         desc->l_req_mode = lock->l_req_mode;
 389         desc->l_granted_mode = lock->l_granted_mode;
 390         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 391         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 392 }
 393
 394 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 395                                    struct ldlm_lock *new,
 396                                    void *data, int datalen)
 397 {
 398         struct ldlm_ast_work *w;
 399         ENTRY;
 400
 401         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 402         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 403                 GOTO(out, 0);
 404
 405         OBD_ALLOC(w, sizeof(*w));
 406         if (!w) {
 407                 LBUG();
 408                 GOTO(out, 0);
 409         }
 410
 411         w->w_data = data;
 412         w->w_datalen = datalen;
 413         if (new) {
 414                 lock->l_flags |= LDLM_FL_AST_SENT;
 415                 w->w_blocking = 1;
 416                 ldlm_lock2desc(new, &w->w_desc);
 417         }
 418
 419         w->w_lock = LDLM_LOCK_GET(lock);
 420         list_add(&w->w_list, lock->l_resource->lr_tmp);
 421         EXIT;
 422  out:
 423         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 424         return;
 425 }
 426
 427 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 428 {
 429         struct ldlm_lock *lock;
 430
 431         lock = ldlm_handle2lock(lockh);
 432         ldlm_lock_addref_internal(lock, mode);
 433         LDLM_LOCK_PUT(lock);
 434 }
 435
 436 /* only called for local locks */
 437 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 438 {
 439         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 440         ldlm_lock_remove_from_lru(lock);
 441         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 442                 lock->l_readers++;
 443         else
 444                 lock->l_writers++;
 445         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 446         LDLM_LOCK_GET(lock);
 447         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 448 }
 449
 450 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
 451 {
 452         struct ldlm_namespace *ns;
 453         ENTRY;
 454
 455         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 456         ns = lock->l_resource->lr_namespace;
 457         l_lock(&ns->ns_lock);
 458         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR) {
 459                 LASSERT(lock->l_readers > 0);
 460                 lock->l_readers--;
 461         } else {
 462                 LASSERT(lock->l_writers > 0);
 463                 lock->l_writers--;
 464         }
 465
 466         if (lock->l_flags & LDLM_FL_LOCAL &&
 467             !lock->l_readers && !lock->l_writers) {
 468                 /* If this is a local lock on a server namespace and this was
 469                  * the last reference, cancel the lock. */
 470                 CDEBUG(D_INFO, "forcing cancel of local lock\n");
 471                 lock->l_flags |= LDLM_FL_CBPENDING;
 472         }
 473
 474         if (!lock->l_readers && !lock->l_writers &&
 475             (lock->l_flags & LDLM_FL_CBPENDING)) {
 476                 /* If we received a blocked AST and this was the last reference,
 477                  * run the callback. */
 478                 if (!ns->ns_client && lock->l_export)
 479                         CERROR("FL_CBPENDING set on non-local lock--just a "
 480                                "warning\n");
 481
 482                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 483                 l_unlock(&ns->ns_lock);
 484
 485                 /* FIXME: need a real 'desc' here */
 486                 if (lock->l_blocking_ast != NULL)
 487                         lock->l_blocking_ast(lock, NULL, lock->l_data,
 488                                              LDLM_CB_BLOCKING);
 489                 else
 490                         LDLM_DEBUG(lock, "No blocking AST?");
 491         } else if (ns->ns_client && !lock->l_readers && !lock->l_writers) {
 492                 /* If this is a client-side namespace and this was the last
 493                  * reference, put it on the LRU. */
 494                 LASSERT(list_empty(&lock->l_lru));
 495                 LASSERT(ns->ns_nr_unused >= 0);
 496                 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 497                 ns->ns_nr_unused++;
 498                 l_unlock(&ns->ns_lock);
 499                 ldlm_cancel_lru(ns);
 500         } else {
 501                 l_unlock(&ns->ns_lock);
 502         }
 503
 504         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 505
 506         EXIT;
 507 }
 508
 509 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 510 {
 511         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 512         LASSERT(lock != NULL);
 513         ldlm_lock_decref_internal(lock, mode);
 514         LDLM_LOCK_PUT(lock);
 515 }
 516
 517 /* This will drop a lock reference and mark it for destruction, but will not
 518  * necessarily cancel the lock before returning. */
 519 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
 520 {
 521         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 522         ENTRY;
 523
 524         LASSERT(lock != NULL);
 525
 526         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 527         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 528         lock->l_flags |= LDLM_FL_CBPENDING;
 529         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 530         ldlm_lock_decref_internal(lock, mode);
 531         LDLM_LOCK_PUT(lock);
 532 }
 533
 534 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 535                                  struct list_head *queue)
 536 {
 537         struct list_head *tmp, *pos;
 538         int rc = 1;
 539
 540         list_for_each_safe(tmp, pos, queue) {
 541                 struct ldlm_lock *child;
 542                 ldlm_res_compat compat;
 543
 544                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 545                 if (lock == child)
 546                         continue;
 547
 548                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 549                 if (compat && compat(child, lock)) {
 550                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 551                         continue;
 552                 }
 553                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 554                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 555                         continue;
 556                 }
 557
 558                 rc = 0;
 559
 560                 if (send_cbs && child->l_blocking_ast != NULL) {
 561                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 562                                "blocking AST.\n", child);
 563                         ldlm_add_ast_work_item(child, lock, NULL, 0);
 564                 }
 565         }
 566
 567         return rc;
 568 }
 569
 570 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 571 {
 572         int rc;
 573         ENTRY;
 574
 575         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 576         rc = ldlm_lock_compat_list(lock, send_cbs,
 577                                    &lock->l_resource->lr_granted);
 578         /* FIXME: should we be sending ASTs to converting? */
 579         if (rc)
 580                 rc = ldlm_lock_compat_list
 581                         (lock, send_cbs, &lock->l_resource->lr_converting);
 582
 583         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 584         RETURN(rc);
 585 }
 586
 587 /* NOTE: called by
 588  *  - ldlm_lock_enqueue
 589  *  - ldlm_reprocess_queue
 590  *  - ldlm_lock_convert
 591  */
 592 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen)
 593 {
 594         struct ldlm_resource *res = lock->l_resource;
 595         ENTRY;
 596
 597         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 598         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 599         lock->l_granted_mode = lock->l_req_mode;
 600
 601         if (lock->l_granted_mode < res->lr_most_restr)
 602                 res->lr_most_restr = lock->l_granted_mode;
 603
 604         if (lock->l_completion_ast != NULL)
 605                 ldlm_add_ast_work_item(lock, NULL, data, datalen);
 606
 607         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 608         EXIT;
 609 }
 610
 611 /* returns a referenced lock or NULL.  See the flag descriptions below, in the
 612  * comment above ldlm_lock_match */
 613 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 614                                       struct ldlm_extent *extent,
 615                                       struct ldlm_lock *old_lock, void *data,
 616                                       int flags)
 617 {
 618         struct ldlm_lock *lock;
 619         struct list_head *tmp;
 620
 621         list_for_each(tmp, queue) {
 622                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 623
 624                 if (lock == old_lock)
 625                         break;
 626
 627                 /* llite sometimes wants to match locks that will be
 628                  * canceled when their users drop, but we allow it to match
 629                  * if it passes in CBPENDING and the lock still has users.
 630                  * this is generally only going to be used by children
 631                  * whose parents already hold a lock so forward progress
 632                  * can still happen. */
 633                 if (lock->l_flags & LDLM_FL_CBPENDING &&
 634                     !(flags & LDLM_FL_CBPENDING))
 635                         continue;
 636                 if (lock->l_flags & LDLM_FL_CBPENDING &&
 637                     lock->l_readers == 0 && lock->l_writers == 0)
 638                         continue;
 639
 640                 if (lock->l_req_mode != mode)
 641                         continue;
 642
 643                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 644                     (lock->l_extent.start > extent->start ||
 645                      lock->l_extent.end < extent->end))
 646                         continue;
 647
 648                 if (lock->l_destroyed)
 649                         continue;
 650
 651                 if ((flags & LDLM_FL_LOCAL_ONLY) &&
 652                     !(lock->l_flags & LDLM_FL_LOCAL))
 653                         continue;
 654
 655                 if ((flags & LDLM_FL_MATCH_DATA) && lock->l_data != data)
 656                         continue;
 657
 658                 ldlm_lock_addref_internal(lock, mode);
 659                 return lock;
 660         }
 661
 662         return NULL;
 663 }
 664
 665 /* Can be called in two ways:
 666  *
 667  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
 668  * for a duplicate of.
 669  *
 670  * Otherwise, all of the fields must be filled in, to match against.
 671  *
 672  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
 673  *     server (ie, connh is NULL)
 674  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
 675  *     list will be considered
 676  * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
 677  *     to be canceled can still be matched as long as they still have reader
 678  *     or writer refernces
 679  * If 'flags' contains LDLM_FL_MATCH_DATA, then only match a lock if the opaque
 680  *     data is the same.
 681  *
 682  * Returns 1 if it finds an already-existing lock that is compatible; in this
 683  * case, lockh is filled in with a addref()ed lock
 684  */
 685 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
 686                     struct ldlm_res_id *res_id, __u32 type, void *cookie,
 687                     int cookielen, ldlm_mode_t mode, void *data,
 688                     struct lustre_handle *lockh)
 689 {
 690         struct ldlm_resource *res;
 691         struct ldlm_lock *lock, *old_lock = NULL;
 692         int rc = 0;
 693         ENTRY;
 694
 695         if (ns == NULL) {
 696                 old_lock = ldlm_handle2lock(lockh);
 697                 LASSERT(old_lock);
 698
 699                 ns = old_lock->l_resource->lr_namespace;
 700                 res_id = &old_lock->l_resource->lr_name;
 701                 type = old_lock->l_resource->lr_type;
 702                 mode = old_lock->l_req_mode;
 703         }
 704
 705         res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
 706         if (res == NULL) {
 707                 LASSERT(old_lock == NULL);
 708                 RETURN(0);
 709         }
 710
 711         l_lock(&ns->ns_lock);
 712
 713         lock = search_queue(&res->lr_granted, mode, cookie, old_lock, data,
 714                             flags);
 715         if (lock != NULL)
 716                 GOTO(out, rc = 1);
 717         if (flags & LDLM_FL_BLOCK_GRANTED)
 718                 GOTO(out, rc = 0);
 719         lock = search_queue(&res->lr_converting, mode, cookie, old_lock, data,
 720                             flags);
 721         if (lock != NULL)
 722                 GOTO(out, rc = 1);
 723         lock = search_queue(&res->lr_waiting, mode, cookie, old_lock, data,
 724                             flags);
 725         if (lock != NULL)
 726                 GOTO(out, rc = 1);
 727
 728         EXIT;
 729  out:
 730         ldlm_resource_putref(res);
 731         l_unlock(&ns->ns_lock);
 732
 733         if (lock) {
 734                 ldlm_lock2handle(lock, lockh);
 735                 if (lock->l_completion_ast)
 736                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 737                                                NULL);
 738         }
 739         if (rc)
 740                 LDLM_DEBUG(lock, "matched");
 741         else
 742                 LDLM_DEBUG_NOLOCK("not matched");
 743
 744         if (old_lock)
 745                 LDLM_LOCK_PUT(old_lock);
 746
 747         return rc;
 748 }
 749
 750 /* Returns a referenced lock */
 751 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 752                                    struct lustre_handle *parent_lock_handle,
 753                                    struct ldlm_res_id res_id, __u32 type,
 754                                    ldlm_mode_t mode,
 755                                    ldlm_blocking_callback blocking,
 756                                    void *data)
 757 {
 758         struct ldlm_resource *res, *parent_res = NULL;
 759         struct ldlm_lock *lock, *parent_lock = NULL;
 760         ENTRY;
 761
 762         if (parent_lock_handle) {
 763                 parent_lock = ldlm_handle2lock(parent_lock_handle);
 764                 if (parent_lock)
 765                         parent_res = parent_lock->l_resource;
 766         }
 767
 768         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 769         if (res == NULL)
 770                 RETURN(NULL);
 771
 772         lock = ldlm_lock_new(parent_lock, res);
 773         ldlm_resource_putref(res);
 774         if (parent_lock != NULL)
 775                 LDLM_LOCK_PUT(parent_lock);
 776
 777         if (lock == NULL)
 778                 RETURN(NULL);
 779
 780         lock->l_req_mode = mode;
 781         lock->l_data = data;
 782         lock->l_blocking_ast = blocking;
 783
 784         RETURN(lock);
 785 }
 786
 787 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
 788                                struct ldlm_lock **lockp,
 789                                void *cookie, int cookie_len,
 790                                int *flags,
 791                                ldlm_completion_callback completion)
 792 {
 793         struct ldlm_resource *res;
 794         struct ldlm_lock *lock = *lockp;
 795         int local;
 796         ldlm_res_policy policy;
 797         ENTRY;
 798
 799         res = lock->l_resource;
 800
 801         if (res->lr_type == LDLM_EXTENT)
 802                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 803
 804         /* policies are not executed on the client or during replay */
 805         local = res->lr_namespace->ns_client;
 806         if (!local && !(*flags & LDLM_FL_REPLAY) &&
 807             (policy = ldlm_res_policy_table[res->lr_type])) {
 808                 int rc;
 809                 rc = policy(ns, lockp, cookie, lock->l_req_mode, *flags, NULL);
 810                 if (rc == ELDLM_LOCK_CHANGED) {
 811                         res = lock->l_resource;
 812                         *flags |= LDLM_FL_LOCK_CHANGED;
 813                 } else if (rc == ELDLM_LOCK_REPLACED) {
 814                         /* The lock that was returned has already been granted,
 815                          * and placed into lockp.  Destroy the old one and our
 816                          * work here is done. */
 817                         ldlm_lock_destroy(lock);
 818                         LDLM_LOCK_PUT(lock);
 819                         *flags |= LDLM_FL_LOCK_CHANGED;
 820                         RETURN(0);
 821                 } else if (rc == ELDLM_LOCK_ABORTED) {
 822                         ldlm_lock_destroy(lock);
 823                         RETURN(rc);
 824                 }
 825         }
 826
 827         l_lock(&ns->ns_lock);
 828         if (local && lock->l_req_mode == lock->l_granted_mode) {
 829                 /* The server returned a blocked lock, but it was granted before
 830                  * we got a chance to actually enqueue it.  We don't need to do
 831                  * anything else. */
 832                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 833                             LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 834                 GOTO(out, ELDLM_OK);
 835         }
 836
 837         /* This distinction between local lock trees is very important; a client
 838          * namespace only has information about locks taken by that client, and
 839          * thus doesn't have enough information to decide for itself if it can
 840          * be granted (below).  In this case, we do exactly what the server
 841          * tells us to do, as dictated by the 'flags'.
 842          *
 843          * We do exactly the same thing during recovery, when the server is
 844          * more or less trusting the clients not to lie.
 845          *
 846          * FIXME (bug 268): Detect obvious lies by checking compatibility in
 847          * granted/converting queues. */
 848         ldlm_resource_unlink_lock(lock);
 849         if (local) {
 850                 if (*flags & LDLM_FL_BLOCK_CONV)
 851                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 852                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 853                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 854                 else
 855                         ldlm_grant_lock(lock, NULL, 0);
 856                 GOTO(out, ELDLM_OK);
 857         } else if (*flags & LDLM_FL_REPLAY) {
 858                 if (*flags & LDLM_FL_BLOCK_CONV) {
 859                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 860                         GOTO(out, ELDLM_OK);
 861                 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
 862                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 863                         GOTO(out, ELDLM_OK);
 864                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
 865                         ldlm_grant_lock(lock, NULL, 0);
 866                         GOTO(out, ELDLM_OK);
 867                 }
 868                 /* If no flags, fall through to normal enqueue path. */
 869         }
 870
 871         /* FIXME: We may want to optimize by checking lr_most_restr */
 872         if (!list_empty(&res->lr_converting)) {
 873                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 874                 *flags |= LDLM_FL_BLOCK_CONV;
 875                 GOTO(out, ELDLM_OK);
 876         }
 877         if (!list_empty(&res->lr_waiting)) {
 878                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 879                 *flags |= LDLM_FL_BLOCK_WAIT;
 880                 GOTO(out, ELDLM_OK);
 881         }
 882         if (!ldlm_lock_compat(lock, 0)) {
 883                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 884                 *flags |= LDLM_FL_BLOCK_GRANTED;
 885                 GOTO(out, ELDLM_OK);
 886         }
 887         ldlm_grant_lock(lock, NULL, 0);
 888         EXIT;
 889       out:
 890         /* Don't set 'completion_ast' until here so that if the lock is granted
 891          * immediately we don't do an unnecessary completion call. */
 892         lock->l_completion_ast = completion;
 893         l_unlock(&ns->ns_lock);
 894         return ELDLM_OK;
 895 }
 896
 897 /* Must be called with namespace taken: queue is waiting or converting. */
 898 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 899                                 struct list_head *queue)
 900 {
 901         struct list_head *tmp, *pos;
 902         ENTRY;
 903
 904         list_for_each_safe(tmp, pos, queue) {
 905                 struct ldlm_lock *pending;
 906                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 907
 908                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 909
 910                 if (!ldlm_lock_compat(pending, 1))
 911                         RETURN(1);
 912
 913                 list_del_init(&pending->l_res_link);
 914                 ldlm_grant_lock(pending, NULL, 0);
 915         }
 916
 917         RETURN(0);
 918 }
 919
 920 int ldlm_run_ast_work(struct list_head *rpc_list)
 921 {
 922         struct list_head *tmp, *pos;
 923         int rc, retval = 0;
 924         ENTRY;
 925
 926         list_for_each_safe(tmp, pos, rpc_list) {
 927                 struct ldlm_ast_work *w =
 928                         list_entry(tmp, struct ldlm_ast_work, w_list);
 929
 930                 /* It's possible to receive a completion AST before we've set
 931                  * the l_completion_ast pointer: either because the AST arrived
 932                  * before the reply, or simply because there's a small race
 933                  * window between receiving the reply and finishing the local
 934                  * enqueue. (bug 842)
 935                  *
 936                  * This can't happen with the blocking_ast, however, because we
 937                  * will never call the local blocking_ast until we drop our
 938                  * reader/writer reference, which we won't do until we get the
 939                  * reply and finish enqueueing. */
 940                 if (w->w_blocking) {
 941                         LASSERT(w->w_lock->l_blocking_ast != NULL);
 942                         rc = w->w_lock->l_blocking_ast
 943                                 (w->w_lock, &w->w_desc, w->w_data,
 944                                  LDLM_CB_BLOCKING);
 945                 } else if (w->w_lock->l_completion_ast != NULL) {
 946                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags,
 947                                                          w->w_data);
 948                 } else {
 949                         rc = 0;
 950                 }
 951                 if (rc == -ERESTART)
 952                         retval = rc;
 953                 else if (rc)
 954                         CERROR("Failed AST - should clean & disconnect "
 955                                "client\n");
 956                 LDLM_LOCK_PUT(w->w_lock);
 957                 list_del(&w->w_list);
 958                 OBD_FREE(w, sizeof(*w));
 959         }
 960         RETURN(retval);
 961 }
 962
 963 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
 964 {
 965         ldlm_reprocess_all(res);
 966         return LDLM_ITER_CONTINUE;
 967 }
 968
 969 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
 970 {
 971         (void)ldlm_namespace_foreach_res(ns, reprocess_one_queue, NULL);
 972 }
 973
 974 void ldlm_reprocess_all(struct ldlm_resource *res)
 975 {
 976         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 977         int rc;
 978         ENTRY;
 979
 980         /* Local lock trees don't get reprocessed. */
 981         if (res->lr_namespace->ns_client) {
 982                 EXIT;
 983                 return;
 984         }
 985
 986  restart:
 987         l_lock(&res->lr_namespace->ns_lock);
 988         res->lr_tmp = &rpc_list;
 989
 990         ldlm_reprocess_queue(res, &res->lr_converting);
 991         if (list_empty(&res->lr_converting))
 992                 ldlm_reprocess_queue(res, &res->lr_waiting);
 993
 994         res->lr_tmp = NULL;
 995         l_unlock(&res->lr_namespace->ns_lock);
 996
 997         rc = ldlm_run_ast_work(&rpc_list);
 998         if (rc == -ERESTART)
 999                 goto restart;
1000         EXIT;
1001 }
1002
1003 void ldlm_cancel_callback(struct ldlm_lock *lock)
1004 {
1005         l_lock(&lock->l_resource->lr_namespace->ns_lock);
1006         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1007                 lock->l_flags |= LDLM_FL_CANCEL;
1008                 if (lock->l_blocking_ast) {
1009                         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1010                         lock->l_blocking_ast(lock, NULL, lock->l_data,
1011                                              LDLM_CB_CANCELING);
1012                         return;
1013                 } else {
1014                         LDLM_DEBUG(lock, "no blocking ast");
1015                 }
1016         }
1017         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1018 }
1019
1020 void ldlm_lock_cancel(struct ldlm_lock *lock)
1021 {
1022         struct ldlm_resource *res;
1023         struct ldlm_namespace *ns;
1024         ENTRY;
1025
1026         ldlm_del_waiting_lock(lock);
1027
1028         res = lock->l_resource;
1029         ns = res->lr_namespace;
1030
1031         l_lock(&ns->ns_lock);
1032         /* Please do not, no matter how tempting, remove this LBUG without
1033          * talking to me first. -phik */
1034         if (lock->l_readers || lock->l_writers) {
1035                 LDLM_DEBUG(lock, "lock still has references");
1036                 ldlm_lock_dump(D_OTHER, lock);
1037                 LBUG();
1038         }
1039
1040         ldlm_cancel_callback(lock); /* XXX FIXME bug 1030 */
1041
1042         ldlm_resource_unlink_lock(lock);
1043         ldlm_lock_destroy(lock);
1044         l_unlock(&ns->ns_lock);
1045         EXIT;
1046 }
1047
1048 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1049 {
1050         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1051         ENTRY;
1052
1053         if (lock == NULL)
1054                 RETURN(-EINVAL);
1055
1056         lock->l_data = data;
1057
1058         LDLM_LOCK_PUT(lock);
1059
1060         RETURN(0);
1061 }
1062
1063 /* This function is only called from one thread (per export); no locking around
1064  * the list ops needed */
1065 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1066 {
1067         struct list_head *iter, *n;
1068
1069         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
1070                 struct ldlm_lock *lock;
1071                 struct ldlm_resource *res;
1072                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1073                 res = ldlm_resource_getref(lock->l_resource);
1074                 LDLM_DEBUG(lock, "export %p", exp);
1075                 ldlm_lock_cancel(lock);
1076                 ldlm_reprocess_all(res);
1077                 ldlm_resource_putref(res);
1078         }
1079 }
1080
1081 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1082                                         int *flags)
1083 {
1084         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1085         struct ldlm_resource *res;
1086         struct ldlm_namespace *ns;
1087         int granted = 0;
1088         ENTRY;
1089
1090         LBUG();
1091
1092         res = lock->l_resource;
1093         ns = res->lr_namespace;
1094
1095         l_lock(&ns->ns_lock);
1096
1097         lock->l_req_mode = new_mode;
1098         ldlm_resource_unlink_lock(lock);
1099
1100         /* If this is a local resource, put it on the appropriate list. */
1101         if (res->lr_namespace->ns_client) {
1102                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1103                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1104                 } else {
1105                         /* This should never happen, because of the way the
1106                          * server handles conversions. */
1107                         LBUG();
1108
1109                         res->lr_tmp = &rpc_list;
1110                         ldlm_grant_lock(lock, NULL, 0);
1111                         res->lr_tmp = NULL;
1112                         granted = 1;
1113                         /* FIXME: completion handling not with ns_lock held ! */
1114                         if (lock->l_completion_ast)
1115                                 lock->l_completion_ast(lock, 0, NULL);
1116                 }
1117         } else {
1118                 /* FIXME: We should try the conversion right away and possibly
1119                  * return success without the need for an extra AST */
1120                 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1121                 *flags |= LDLM_FL_BLOCK_CONV;
1122         }
1123
1124         l_unlock(&ns->ns_lock);
1125
1126         if (granted)
1127                 ldlm_run_ast_work(&rpc_list);
1128         RETURN(res);
1129 }
1130
1131 void ldlm_lock_dump(int level, struct ldlm_lock *lock)
1132 {
1133         char ver[128];
1134         struct obd_device *obd;
1135
1136         if (!((portal_debug | D_ERROR) & level))
1137                 return;
1138
1139         if (RES_VERSION_SIZE != 4)
1140                 LBUG();
1141
1142         if (!lock) {
1143                 CDEBUG(level, "  NULL LDLM lock\n");
1144                 return;
1145         }
1146
1147         snprintf(ver, sizeof(ver), "%x %x %x %x",
1148                  lock->l_version[0], lock->l_version[1],
1149                  lock->l_version[2], lock->l_version[3]);
1150
1151         CDEBUG(level, "  -- Lock dump: %p (%s) (rc: %d)\n", lock, ver,
1152                atomic_read(&lock->l_refc));
1153         obd = class_conn2obd(lock->l_connh);
1154         if (lock->l_export && lock->l_export->exp_connection) {
1155                 CDEBUG(level, "  Node: NID "LPX64" on %s (rhandle: "LPX64")\n",
1156                        lock->l_export->exp_connection->c_peer.peer_nid,
1157                        lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1158                        lock->l_remote_handle.cookie);
1159         } else if (obd == NULL) {
1160                 CDEBUG(level, "  Node: local\n");
1161         } else {
1162                 struct obd_import *imp = obd->u.cli.cl_import;
1163                 CDEBUG(level, "  Node: NID "LPX64" on %s (rhandle: "LPX64")\n",
1164                        imp->imp_connection->c_peer.peer_nid,
1165                        imp->imp_connection->c_peer.peer_ni->pni_name,
1166                        lock->l_remote_handle.cookie);
1167         }
1168         CDEBUG(level, "  Parent: %p\n", lock->l_parent);
1169         CDEBUG(level, "  Resource: %p ("LPD64")\n", lock->l_resource,
1170                lock->l_resource->lr_name.name[0]);
1171         CDEBUG(level, "  Requested mode: %d, granted mode: %d\n",
1172                (int)lock->l_req_mode, (int)lock->l_granted_mode);
1173         CDEBUG(level, "  Readers: %u ; Writers; %u\n",
1174                lock->l_readers, lock->l_writers);
1175         if (lock->l_resource->lr_type == LDLM_EXTENT)
1176                 CDEBUG(level, "  Extent: "LPU64" -> "LPU64"\n",
1177                        lock->l_extent.start, lock->l_extent.end);
1178 }
1179
1180 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1181 {
1182         struct ldlm_lock *lock;
1183
1184         lock = ldlm_handle2lock(lockh);
1185         if (lock == NULL)
1186                 return;
1187
1188         ldlm_lock_dump(D_OTHER, lock);
1189
1190         LDLM_LOCK_PUT(lock);
1191 }