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/dcache.h>
  29 # include <linux/namei.h>
  30 # include <linux/module.h>
  31 # include <linux/lustre_dlm.h>
  32 #else
  33 # include <liblustre.h>
  34 # include <libcfs/kp30.h>
  35 #endif
  36
  37 #include <linux/obd_class.h>
  38 #include "ldlm_internal.h"
  39
  40 //struct lustre_lock ldlm_everything_lock;
  41
  42 /* lock types */
  43 char *ldlm_lockname[] = {
  44         [0] "--",
  45         [LCK_EX] "EX",
  46         [LCK_PW] "PW",
  47         [LCK_PR] "PR",
  48         [LCK_CW] "CW",
  49         [LCK_CR] "CR",
  50         [LCK_NL] "NL",
  51         [LCK_GROUP] "GROUP"
  52 };
  53 char *ldlm_typename[] = {
  54         [LDLM_PLAIN] "PLN",
  55         [LDLM_EXTENT] "EXT",
  56         [LDLM_FLOCK] "FLK",
  57         [LDLM_IBITS] "IBT",
  58 };
  59
  60 char *ldlm_it2str(int it)
  61 {
  62         switch (it) {
  63         case IT_OPEN:
  64                 return "open";
  65         case IT_CREAT:
  66                 return "creat";
  67         case (IT_OPEN | IT_CREAT):
  68                 return "open|creat";
  69         case IT_READDIR:
  70                 return "readdir";
  71         case IT_GETATTR:
  72                 return "getattr";
  73         case IT_LOOKUP:
  74                 return "lookup";
  75         case IT_UNLINK:
  76                 return "unlink";
  77         case IT_GETXATTR:
  78                 return "getxattr";
  79         case IT_CHDIR:
  80                 return "chdir";
  81         default:
  82                 CERROR("Unknown intent %d\n", it);
  83                 return "UNKNOWN";
  84         }
  85 }
  86
  87 extern kmem_cache_t *ldlm_lock_slab;
  88
  89 static ldlm_processing_policy ldlm_processing_policy_table[] = {
  90         [LDLM_PLAIN] ldlm_process_plain_lock,
  91         [LDLM_EXTENT] ldlm_process_extent_lock,
  92 #ifdef __KERNEL__
  93         [LDLM_FLOCK] ldlm_process_flock_lock,
  94 #endif
  95         [LDLM_IBITS] ldlm_process_inodebits_lock,
  96 };
  97
  98 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
  99 {
 100         return ldlm_processing_policy_table[res->lr_type];
 101 }
 102
 103 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
 104 {
 105         ns->ns_policy = arg;
 106 }
 107
 108 /*
 109  * REFCOUNTED LOCK OBJECTS
 110  */
 111
 112
 113 /*
 114  * Lock refcounts, during creation:
 115  *   - one special one for allocation, dec'd only once in destroy
 116  *   - one for being a lock that's in-use
 117  *   - one for the addref associated with a new lock
 118  */
 119 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 120 {
 121         atomic_inc(&lock->l_refc);
 122         return lock;
 123 }
 124
 125 void ldlm_lock_put(struct ldlm_lock *lock)
 126 {
 127         ENTRY;
 128
 129         LASSERT(lock->l_resource != LP_POISON);
 130         LASSERT(atomic_read(&lock->l_refc) > 0);
 131         if (atomic_dec_and_test(&lock->l_refc)) {
 132                 struct ldlm_resource *res = lock->l_resource;
 133                 struct ldlm_namespace *ns = res->lr_namespace;
 134
 135                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
 136
 137                 LASSERT(lock->l_resource != LP_POISON);
 138                 lock_res(res);
 139                 LASSERT(lock->l_destroyed);
 140                 LASSERT(list_empty(&lock->l_res_link));
 141
 142                 if (lock->l_parent)
 143                         LDLM_LOCK_PUT(lock->l_parent);
 144                 unlock_res(res);
 145
 146                 ldlm_resource_putref(lock->l_resource);
 147                 lock->l_resource = NULL;
 148                 if (lock->l_export)
 149                         class_export_put(lock->l_export);
 150                 atomic_dec(&ns->ns_locks);
 151
 152                 if (lock->l_lvb_data != NULL)
 153                         OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
 154
 155                 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
 156         }
 157
 158         EXIT;
 159 }
 160
 161 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
 162 {
 163         ENTRY;
 164         spin_lock(&lock->l_resource->lr_namespace->ns_unused_lock);
 165         if (!list_empty(&lock->l_lru)) {
 166                 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
 167                 list_del_init(&lock->l_lru);
 168                 lock->l_resource->lr_namespace->ns_nr_unused--;
 169                 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
 170         }
 171         spin_unlock(&lock->l_resource->lr_namespace->ns_unused_lock);
 172         EXIT;
 173 }
 174
 175 /* This used to have a 'strict' flact, which recovery would use to mark an
 176  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
 177  * shall explain why it's gone: with the new hash table scheme, once you call
 178  * ldlm_lock_destroy, you can never drop your final references on this lock.
 179  * Because it's not in the hash table anymore.  -phil */
 180 void ldlm_lock_destroy(struct ldlm_lock *lock)
 181 {
 182         ENTRY;
 183
 184         lock_res(lock->l_resource);
 185
 186         if (!list_empty(&lock->l_children)) {
 187                 LDLM_ERROR(lock, "still has children (%p)!",
 188                            lock->l_children.next);
 189                 ldlm_lock_dump(D_ERROR, lock, 0);
 190                 LBUG();
 191         }
 192         if (lock->l_readers || lock->l_writers) {
 193                 LDLM_ERROR(lock, "lock still has references");
 194                 ldlm_lock_dump(D_ERROR, lock, 0);
 195                 LBUG();
 196         }
 197
 198         if (!list_empty(&lock->l_res_link)) {
 199                 LDLM_ERROR(lock, "lock still on resource");
 200                 ldlm_lock_dump(D_ERROR, lock, 0);
 201                 LBUG();
 202         }
 203
 204         if (lock->l_destroyed) {
 205                 LASSERT(list_empty(&lock->l_lru));
 206                 unlock_res(lock->l_resource);
 207                 EXIT;
 208                 return;
 209         }
 210         lock->l_destroyed = 1;
 211
 212         if (lock->l_export) {
 213                 spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
 214                 if (!list_empty(&lock->l_export_chain))
 215                         list_del_init(&lock->l_export_chain);
 216                 spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
 217         } else {
 218                 LASSERT(list_empty(&lock->l_export_chain));
 219         }
 220
 221         ldlm_lock_remove_from_lru(lock);
 222         class_handle_unhash(&lock->l_handle);
 223
 224 #if 0
 225         /* Wake anyone waiting for this lock */
 226         /* FIXME: I should probably add yet another flag, instead of using
 227          * l_export to only call this on clients */
 228         if (lock->l_export)
 229                 class_export_put(lock->l_export);
 230         lock->l_export = NULL;
 231         if (lock->l_export && lock->l_completion_ast)
 232                 lock->l_completion_ast(lock, 0);
 233 #endif
 234
 235         unlock_res(lock->l_resource);
 236         LDLM_LOCK_PUT(lock);
 237         EXIT;
 238 }
 239
 240 /* this is called by portals_handle2object with the handle lock taken */
 241 static void lock_handle_addref(void *lock)
 242 {
 243         LDLM_LOCK_GET((struct ldlm_lock *)lock);
 244 }
 245
 246 /*
 247  * usage: pass in a resource on which you have done ldlm_resource_get
 248  *        pass in a parent lock on which you have done a ldlm_lock_get
 249  *        after return, ldlm_*_put the resource and parent
 250  * returns: lock with refcount 2 - one for current caller and one for remote
 251  */
 252 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 253                                        struct ldlm_resource *resource)
 254 {
 255         struct ldlm_lock *lock;
 256         ENTRY;
 257
 258         if (resource == NULL)
 259                 LBUG();
 260
 261         OBD_SLAB_ALLOC(lock, ldlm_lock_slab, SLAB_NOFS, sizeof(*lock));
 262         if (lock == NULL)
 263                 RETURN(NULL);
 264
 265         lock->l_resource = ldlm_resource_getref(resource);
 266
 267         atomic_set(&lock->l_refc, 2);
 268         INIT_LIST_HEAD(&lock->l_children);
 269         INIT_LIST_HEAD(&lock->l_res_link);
 270         INIT_LIST_HEAD(&lock->l_lru);
 271         INIT_LIST_HEAD(&lock->l_export_chain);
 272         INIT_LIST_HEAD(&lock->l_pending_chain);
 273         INIT_LIST_HEAD(&lock->l_tmp);
 274         INIT_LIST_HEAD(&lock->l_bl_ast);
 275         INIT_LIST_HEAD(&lock->l_cp_ast);
 276         init_waitqueue_head(&lock->l_waitq);
 277         lock->l_blocking_lock = NULL;
 278
 279         atomic_inc(&resource->lr_namespace->ns_locks);
 280
 281         if (parent != NULL) {
 282                 spin_lock(&resource->lr_namespace->ns_hash_lock);
 283                 lock->l_parent = LDLM_LOCK_GET(parent);
 284                 list_add(&lock->l_childof, &parent->l_children);
 285                 spin_unlock(&resource->lr_namespace->ns_hash_lock);
 286         }
 287
 288         INIT_LIST_HEAD(&lock->l_handle.h_link);
 289         class_handle_hash(&lock->l_handle, lock_handle_addref);
 290
 291         RETURN(lock);
 292 }
 293
 294 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
 295                               struct ldlm_res_id new_resid)
 296 {
 297         struct ldlm_resource *oldres = lock->l_resource;
 298         ENTRY;
 299
 300         lock_res(oldres);
 301         if (memcmp(&new_resid, &lock->l_resource->lr_name,
 302                    sizeof(lock->l_resource->lr_name)) == 0) {
 303                 /* Nothing to do */
 304                 unlock_res(oldres);
 305                 RETURN(0);
 306         }
 307
 308         LASSERT(new_resid.name[0] != 0);
 309
 310         /* This function assumes that the lock isn't on any lists */
 311         LASSERT(list_empty(&lock->l_res_link));
 312
 313         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
 314                                              lock->l_resource->lr_type,
 315                                              1);
 316         if (lock->l_resource == NULL) {
 317                 LBUG();
 318                 RETURN(-ENOMEM);
 319         }
 320
 321         unlock_res(oldres);
 322
 323         /* ...and the flowers are still standing! */
 324         ldlm_resource_putref(oldres);
 325
 326         RETURN(0);
 327 }
 328
 329 /*
 330  *  HANDLES
 331  */
 332
 333 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 334 {
 335         lockh->cookie = lock->l_handle.h_cookie;
 336 }
 337
 338 /* if flags: atomically get the lock and set the flags.
 339  *           Return NULL if flag already set
 340  */
 341
 342 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
 343 {
 344         struct ldlm_namespace *ns;
 345         struct ldlm_lock *lock = NULL, *retval = NULL;
 346         ENTRY;
 347
 348         LASSERT(handle);
 349
 350         lock = class_handle2object(handle->cookie);
 351         if (lock == NULL)
 352                 RETURN(NULL);
 353
 354         LASSERT(lock->l_resource != NULL);
 355         ns = lock->l_resource->lr_namespace;
 356         LASSERT(ns != NULL);
 357
 358         lock_res(lock->l_resource);
 359
 360         /* It's unlikely but possible that someone marked the lock as
 361          * destroyed after we did handle2object on it */
 362         if (lock->l_destroyed) {
 363                 unlock_res(lock->l_resource);
 364                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
 365                 LDLM_LOCK_PUT(lock);
 366                 GOTO(out, retval);
 367         }
 368
 369         if (flags && (lock->l_flags & flags)) {
 370                 unlock_res(lock->l_resource);
 371                 LDLM_LOCK_PUT(lock);
 372                 GOTO(out, retval);
 373         }
 374
 375         if (flags)
 376                 lock->l_flags |= flags;
 377
 378         unlock_res(lock->l_resource);
 379         retval = lock;
 380         EXIT;
 381  out:
 382         return retval;
 383 }
 384
 385 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
 386                                       struct lustre_handle *handle)
 387 {
 388         struct ldlm_lock *retval = NULL;
 389         retval = __ldlm_handle2lock(handle, 0);
 390         return retval;
 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_policy_data, &lock->l_policy_data,
 399                sizeof(desc->l_policy_data));
 400 }
 401
 402 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
 403                            struct list_head *work_list)
 404 {
 405         if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
 406                 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
 407                 lock->l_flags |= LDLM_FL_AST_SENT;
 408                 /* If the enqueuing client said so, tell the AST recipient to
 409                  * discard dirty data, rather than writing back. */
 410                 if (new->l_flags & LDLM_AST_DISCARD_DATA)
 411                         lock->l_flags |= LDLM_FL_DISCARD_DATA;
 412                 LASSERT(list_empty(&lock->l_bl_ast));
 413                 list_add(&lock->l_bl_ast, work_list);
 414                 LDLM_LOCK_GET(lock);
 415                 LASSERT(lock->l_blocking_lock == NULL);
 416                 lock->l_blocking_lock = LDLM_LOCK_GET(new);
 417         }
 418 }
 419
 420 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
 421 {
 422         if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
 423                 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
 424                 lock->l_flags |= LDLM_FL_CP_REQD;
 425                 LASSERT(list_empty(&lock->l_cp_ast));
 426                 list_add(&lock->l_cp_ast, work_list);
 427                 LDLM_LOCK_GET(lock);
 428         }
 429 }
 430
 431 /* must be called with lr_lock held */
 432 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
 433                                 struct list_head *work_list)
 434 {
 435         ENTRY;
 436         check_res_locked(lock->l_resource);
 437         if (new)
 438                 ldlm_add_bl_work_item(lock, new, work_list);
 439         else
 440                 ldlm_add_cp_work_item(lock, work_list);
 441         EXIT;
 442 }
 443
 444 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 445 {
 446         struct ldlm_lock *lock;
 447
 448         lock = ldlm_handle2lock(lockh);
 449         ldlm_lock_addref_internal(lock, mode);
 450         LDLM_LOCK_PUT(lock);
 451 }
 452
 453 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
 454 {
 455         ldlm_lock_remove_from_lru(lock);
 456         if (mode & (LCK_NL | LCK_CR | LCK_PR))
 457                 lock->l_readers++;
 458         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
 459                 lock->l_writers++;
 460         lock->l_last_used = jiffies;
 461         LDLM_LOCK_GET(lock);
 462         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 463 }
 464
 465 /* only called for local locks */
 466 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 467 {
 468         lock_res(lock->l_resource);
 469         ldlm_lock_addref_internal_nolock(lock, mode);
 470         unlock_res(lock->l_resource);
 471 }
 472
 473 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
 474 {
 475         struct ldlm_namespace *ns;
 476         ENTRY;
 477
 478         ns = lock->l_resource->lr_namespace;
 479
 480         lock_res(lock->l_resource);
 481
 482         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 483         if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
 484                 LASSERT(lock->l_readers > 0);
 485                 lock->l_readers--;
 486         }
 487         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
 488                 LASSERT(lock->l_writers > 0);
 489                 lock->l_writers--;
 490         }
 491
 492         if (lock->l_flags & LDLM_FL_LOCAL &&
 493             !lock->l_readers && !lock->l_writers) {
 494                 /* If this is a local lock on a server namespace and this was
 495                  * the last reference, cancel the lock. */
 496                 CDEBUG(D_INFO, "forcing cancel of local lock\n");
 497                 lock->l_flags |= LDLM_FL_CBPENDING;
 498         }
 499
 500         if (!lock->l_readers && !lock->l_writers &&
 501             (lock->l_flags & LDLM_FL_CBPENDING)) {
 502                 /* If we received a blocked AST and this was the last reference,
 503                  * run the callback. */
 504                 if (ns->ns_client == LDLM_NAMESPACE_SERVER && lock->l_export)
 505                         CERROR("FL_CBPENDING set on non-local lock--just a "
 506                                "warning\n");
 507
 508                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 509
 510                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
 511                 ldlm_lock_remove_from_lru(lock);
 512                 unlock_res(lock->l_resource);
 513                 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
 514                                 ldlm_bl_to_thread(ns, NULL, lock) != 0)
 515                         ldlm_handle_bl_callback(ns, NULL, lock);
 516         } else if (ns->ns_client == LDLM_NAMESPACE_CLIENT &&
 517                    !lock->l_readers && !lock->l_writers) {
 518                 /* If this is a client-side namespace and this was the last
 519                  * reference, put it on the LRU. */
 520                 LASSERT(list_empty(&lock->l_lru));
 521                 LASSERT(ns->ns_nr_unused >= 0);
 522                 spin_lock(&ns->ns_unused_lock);
 523                 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 524                 ns->ns_nr_unused++;
 525                 spin_unlock(&ns->ns_unused_lock);
 526                 unlock_res(lock->l_resource);
 527                 ldlm_cancel_lru(ns, LDLM_ASYNC);
 528         } else {
 529                 unlock_res(lock->l_resource);
 530         }
 531
 532         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 533
 534         EXIT;
 535 }
 536
 537 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 538 {
 539         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 540         LASSERT(lock != NULL);
 541         ldlm_lock_decref_internal(lock, mode);
 542         LDLM_LOCK_PUT(lock);
 543 }
 544
 545 /* This will drop a lock reference and mark it for destruction, but will not
 546  * necessarily cancel the lock before returning. */
 547 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
 548 {
 549         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 550         ENTRY;
 551
 552         LASSERT(lock != NULL);
 553
 554         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 555         lock_res(lock->l_resource);
 556         lock->l_flags |= LDLM_FL_CBPENDING;
 557         unlock_res(lock->l_resource);
 558         ldlm_lock_decref_internal(lock, mode);
 559         LDLM_LOCK_PUT(lock);
 560 }
 561
 562 /* NOTE: called by
 563  *  - ldlm_lock_enqueue
 564  *  - ldlm_reprocess_queue
 565  *  - ldlm_lock_convert
 566  *
 567  * must be called with lr_lock held
 568  */
 569 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
 570 {
 571         struct ldlm_resource *res = lock->l_resource;
 572         ENTRY;
 573
 574         check_res_locked(res);
 575
 576         lock->l_granted_mode = lock->l_req_mode;
 577         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 578
 579         if (lock->l_granted_mode < res->lr_most_restr)
 580                 res->lr_most_restr = lock->l_granted_mode;
 581
 582         if (work_list && lock->l_completion_ast != NULL)
 583                 ldlm_add_ast_work_item(lock, NULL, work_list);
 584
 585         EXIT;
 586 }
 587
 588 /* returns a referenced lock or NULL.  See the flag descriptions below, in the
 589  * comment above ldlm_lock_match */
 590 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 591                                       ldlm_policy_data_t *policy,
 592                                       struct ldlm_lock *old_lock, int flags)
 593 {
 594         struct ldlm_lock *lock;
 595         struct list_head *tmp;
 596
 597         list_for_each(tmp, queue) {
 598                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 599
 600                 if (lock == old_lock)
 601                         break;
 602
 603                 /* llite sometimes wants to match locks that will be
 604                  * canceled when their users drop, but we allow it to match
 605                  * if it passes in CBPENDING and the lock still has users.
 606                  * this is generally only going to be used by children
 607                  * whose parents already hold a lock so forward progress
 608                  * can still happen. */
 609                 if (lock->l_flags & LDLM_FL_CBPENDING &&
 610                     !(flags & LDLM_FL_CBPENDING))
 611                         continue;
 612                 if (lock->l_flags & LDLM_FL_CBPENDING &&
 613                     lock->l_readers == 0 && lock->l_writers == 0)
 614                         continue;
 615
 616                 if (!(lock->l_req_mode & mode))
 617                         continue;
 618
 619                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 620                     (lock->l_policy_data.l_extent.start >
 621                      policy->l_extent.start ||
 622                      lock->l_policy_data.l_extent.end < policy->l_extent.end))
 623                         continue;
 624
 625                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 626                     mode == LCK_GROUP &&
 627                     lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
 628                         continue;
 629
 630                 /* We match if we have existing lock with same or wider set
 631                    of bits. */
 632                 if (lock->l_resource->lr_type == LDLM_IBITS &&
 633                      ((lock->l_policy_data.l_inodebits.bits &
 634                       policy->l_inodebits.bits) !=
 635                       policy->l_inodebits.bits))
 636                         continue;
 637
 638                 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
 639                         continue;
 640
 641                 if ((flags & LDLM_FL_LOCAL_ONLY) &&
 642                     !(lock->l_flags & LDLM_FL_LOCAL))
 643                         continue;
 644
 645                 if (flags & LDLM_FL_TEST_LOCK)
 646                         LDLM_LOCK_GET(lock);
 647                 else
 648                         ldlm_lock_addref_internal_nolock(lock, mode);
 649                 return lock;
 650         }
 651
 652         return NULL;
 653 }
 654
 655 void ldlm_lock_allow_match(struct ldlm_lock *lock)
 656 {
 657         lock_res(lock->l_resource);
 658         lock->l_flags |= LDLM_FL_CAN_MATCH;
 659         wake_up(&lock->l_waitq);
 660         unlock_res(lock->l_resource);
 661 }
 662
 663 /* Can be called in two ways:
 664  *
 665  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
 666  * for a duplicate of.
 667  *
 668  * Otherwise, all of the fields must be filled in, to match against.
 669  *
 670  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
 671  *     server (ie, connh is NULL)
 672  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
 673  *     list will be considered
 674  * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
 675  *     to be canceled can still be matched as long as they still have reader
 676  *     or writer refernces
 677  * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
 678  *     just tell us if we would have matched.
 679  *
 680  * Returns 1 if it finds an already-existing lock that is compatible; in this
 681  * case, lockh is filled in with a addref()ed lock
 682  */
 683 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
 684                     struct ldlm_res_id *res_id, __u32 type,
 685                     ldlm_policy_data_t *policy, ldlm_mode_t mode,
 686                     struct lustre_handle *lockh)
 687 {
 688         struct ldlm_resource *res;
 689         struct ldlm_lock *lock, *old_lock = NULL;
 690         int rc = 0;
 691         ENTRY;
 692
 693         if (ns == NULL) {
 694                 old_lock = ldlm_handle2lock(lockh);
 695                 LASSERT(old_lock);
 696
 697                 ns = old_lock->l_resource->lr_namespace;
 698                 res_id = &old_lock->l_resource->lr_name;
 699                 type = old_lock->l_resource->lr_type;
 700                 mode = old_lock->l_req_mode;
 701         }
 702
 703         res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
 704         if (res == NULL) {
 705                 LASSERT(old_lock == NULL);
 706                 RETURN(0);
 707         }
 708
 709         lock_res(res);
 710
 711         lock = search_queue(&res->lr_granted, mode, policy, old_lock, flags);
 712         if (lock != NULL)
 713                 GOTO(out, rc = 1);
 714         if (flags & LDLM_FL_BLOCK_GRANTED)
 715                 GOTO(out, rc = 0);
 716         lock = search_queue(&res->lr_converting, mode, policy, old_lock, flags);
 717         if (lock != NULL)
 718                 GOTO(out, rc = 1);
 719         lock = search_queue(&res->lr_waiting, mode, policy, old_lock, flags);
 720         if (lock != NULL)
 721                 GOTO(out, rc = 1);
 722
 723         EXIT;
 724  out:
 725         unlock_res(res);
 726         ldlm_resource_putref(res);
 727
 728         if (lock) {
 729                 ldlm_lock2handle(lock, lockh);
 730                 if (!(lock->l_flags & LDLM_FL_CAN_MATCH)) {
 731                         struct l_wait_info lwi;
 732                         if (lock->l_completion_ast) {
 733                                 int err = lock->l_completion_ast(lock,
 734                                                                  LDLM_FL_WAIT_NOREPROC,
 735                                                                  NULL);
 736                                 if (err) {
 737                                         rc = 0;
 738                                         goto out2;
 739                                 }
 740                         }
 741
 742                         lwi = LWI_TIMEOUT_INTR(obd_timeout*HZ, NULL,NULL,NULL);
 743
 744                         /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
 745                         l_wait_event(lock->l_waitq,
 746                                      (lock->l_flags & LDLM_FL_CAN_MATCH), &lwi);
 747                 }
 748         }
 749
 750 out2:
 751         if (rc) {
 752                 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
 753                            type == LDLM_PLAIN ? res_id->name[2] :
 754                                 policy->l_extent.start,
 755                            type == LDLM_PLAIN ? res_id->name[3] :
 756                            policy->l_extent.end);
 757         } else if (!(flags & LDLM_FL_TEST_LOCK)) {/* less verbose for test-only */
 758                 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
 759                                   LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
 760                                   type, mode, res_id->name[0], res_id->name[1],
 761                                   type == LDLM_PLAIN ? res_id->name[2] :
 762                                   policy->l_extent.start,
 763                                   type == LDLM_PLAIN ? res_id->name[3] :
 764                                   policy->l_extent.end);
 765         }
 766
 767         if (old_lock)
 768                 LDLM_LOCK_PUT(old_lock);
 769         if (flags & LDLM_FL_TEST_LOCK && rc)
 770                 LDLM_LOCK_PUT(lock);
 771
 772         return rc;
 773 }
 774
 775 /* Returns a referenced lock */
 776 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 777                                    struct lustre_handle *parent_lock_handle,
 778                                    struct ldlm_res_id res_id, __u32 type,
 779                                    ldlm_mode_t mode,
 780                                    ldlm_blocking_callback blocking,
 781                                    ldlm_completion_callback completion,
 782                                    ldlm_glimpse_callback glimpse,
 783                                    void *data, __u32 lvb_len)
 784 {
 785         struct ldlm_resource *res, *parent_res = NULL;
 786         struct ldlm_lock *lock, *parent_lock = NULL;
 787         ENTRY;
 788
 789         if (parent_lock_handle) {
 790                 parent_lock = ldlm_handle2lock(parent_lock_handle);
 791                 if (parent_lock)
 792                         parent_res = parent_lock->l_resource;
 793         }
 794
 795         res = ldlm_resource_get(ns, parent_res, res_id,
 796                                 type, 1);
 797         if (res == NULL)
 798                 RETURN(NULL);
 799
 800         lock = ldlm_lock_new(parent_lock, res);
 801         ldlm_resource_putref(res);
 802         if (parent_lock != NULL)
 803                 LDLM_LOCK_PUT(parent_lock);
 804
 805         if (lock == NULL)
 806                 RETURN(NULL);
 807
 808         lock->l_req_mode = mode;
 809         lock->l_ast_data = data;
 810         lock->l_blocking_ast = blocking;
 811         lock->l_completion_ast = completion;
 812         lock->l_glimpse_ast = glimpse;
 813         lock->l_pid = current->pid;
 814
 815         if (lvb_len) {
 816                 lock->l_lvb_len = lvb_len;
 817                 OBD_ALLOC(lock->l_lvb_data, lvb_len);
 818                 if (lock->l_lvb_data == NULL) {
 819                         OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
 820                         RETURN(NULL);
 821                 }
 822         }
 823
 824         RETURN(lock);
 825 }
 826
 827 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
 828                                struct ldlm_lock **lockp,
 829                                void *cookie, int *flags)
 830 {
 831         struct ldlm_lock *lock = *lockp;
 832         struct ldlm_resource *res = lock->l_resource;
 833         int local = res->lr_namespace->ns_client;
 834         ldlm_processing_policy policy;
 835         ldlm_error_t rc = ELDLM_OK;
 836         ENTRY;
 837
 838         /* policies are not executed on the client or during replay */
 839         if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
 840             && !local && ns->ns_policy) {
 841                 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
 842                                    NULL);
 843                 if (rc == ELDLM_LOCK_REPLACED) {
 844                         /* The lock that was returned has already been granted,
 845                          * and placed into lockp.  If it's not the same as the
 846                          * one we passed in, then destroy the old one and our
 847                          * work here is done. */
 848                         if (lock != *lockp) {
 849                                 ldlm_lock_destroy(lock);
 850                                 LDLM_LOCK_PUT(lock);
 851                         }
 852                         *flags |= LDLM_FL_LOCK_CHANGED;
 853                         RETURN(0);
 854                 } else if (rc == ELDLM_LOCK_ABORTED ||
 855                            (rc == 0 && (*flags & LDLM_FL_INTENT_ONLY))) {
 856                         ldlm_lock_destroy(lock);
 857                         RETURN(rc);
 858                 }
 859                 LASSERT(rc == ELDLM_OK);
 860         }
 861
 862         lock_res(lock->l_resource);
 863         if (local && lock->l_req_mode == lock->l_granted_mode) {
 864                 /* The server returned a blocked lock, but it was granted before
 865                  * we got a chance to actually enqueue it.  We don't need to do
 866                  * anything else. */
 867                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 868                             LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 869                 GOTO(out, ELDLM_OK);
 870         }
 871
 872         /* Some flags from the enqueue want to make it into the AST, via the
 873          * lock's l_flags. */
 874         lock->l_flags |= (*flags & LDLM_AST_DISCARD_DATA);
 875
 876         /* This distinction between local lock trees is very important; a client
 877          * namespace only has information about locks taken by that client, and
 878          * thus doesn't have enough information to decide for itself if it can
 879          * be granted (below).  In this case, we do exactly what the server
 880          * tells us to do, as dictated by the 'flags'.
 881          *
 882          * We do exactly the same thing during recovery, when the server is
 883          * more or less trusting the clients not to lie.
 884          *
 885          * FIXME (bug 268): Detect obvious lies by checking compatibility in
 886          * granted/converting queues. */
 887         ldlm_resource_unlink_lock(lock);
 888         if (local) {
 889                 if (*flags & LDLM_FL_BLOCK_CONV)
 890                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 891                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 892                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 893                 else
 894                         ldlm_grant_lock(lock, NULL);
 895                 GOTO(out, ELDLM_OK);
 896         } else if (*flags & LDLM_FL_REPLAY) {
 897                 if (*flags & LDLM_FL_BLOCK_CONV) {
 898                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 899                         GOTO(out, ELDLM_OK);
 900                 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
 901                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 902                         GOTO(out, ELDLM_OK);
 903                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
 904                         ldlm_grant_lock(lock, NULL);
 905                         GOTO(out, ELDLM_OK);
 906                 }
 907                 /* If no flags, fall through to normal enqueue path. */
 908         }
 909
 910         policy = ldlm_processing_policy_table[res->lr_type];
 911         policy(lock, flags, 1, &rc, NULL);
 912         EXIT;
 913 out:
 914         unlock_res(lock->l_resource);
 915         return rc;
 916 }
 917
 918 /* Must be called with namespace taken: queue is waiting or converting. */
 919 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
 920                          struct list_head *work_list)
 921 {
 922         struct list_head *tmp, *pos;
 923         ldlm_processing_policy policy;
 924         int flags;
 925         int rc = LDLM_ITER_CONTINUE;
 926         ldlm_error_t err;
 927         ENTRY;
 928
 929         check_res_locked(res);
 930
 931         policy = ldlm_processing_policy_table[res->lr_type];
 932         LASSERT(policy);
 933
 934         list_for_each_safe(tmp, pos, queue) {
 935                 struct ldlm_lock *pending;
 936                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 937
 938                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 939
 940                 flags = 0;
 941                 rc = policy(pending, &flags, 0, &err, work_list);
 942                 if (rc != LDLM_ITER_CONTINUE)
 943                         break;
 944         }
 945
 946         RETURN(rc);
 947 }
 948
 949 int ldlm_run_bl_ast_work(struct list_head *rpc_list)
 950 {
 951         struct list_head *tmp, *pos;
 952         struct ldlm_lock_desc d;
 953         int rc = 0, retval = 0;
 954         ENTRY;
 955
 956         list_for_each_safe(tmp, pos, rpc_list) {
 957                 struct ldlm_lock *lock =
 958                         list_entry(tmp, struct ldlm_lock, l_bl_ast);
 959
 960                 /* nobody should touch l_bl_ast */
 961                 lock_res(lock->l_resource);
 962                 list_del_init(&lock->l_bl_ast);
 963
 964                 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
 965                 LASSERT(lock->l_bl_ast_run == 0);
 966                 LASSERT(lock->l_blocking_lock);
 967                 lock->l_bl_ast_run++;
 968                 unlock_res(lock->l_resource);
 969
 970                 ldlm_lock2desc(lock->l_blocking_lock, &d);
 971
 972                 LDLM_LOCK_PUT(lock->l_blocking_lock);
 973                 lock->l_blocking_lock = NULL;
 974                 rc = lock->l_blocking_ast(lock, &d, NULL, LDLM_CB_BLOCKING);
 975
 976                 if (rc == -ERESTART)
 977                         retval = rc;
 978                 else if (rc)
 979                         CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
 980                                "disconnect client\n");
 981                 LDLM_LOCK_PUT(lock);
 982         }
 983         RETURN(retval);
 984 }
 985
 986 int ldlm_run_cp_ast_work(struct list_head *rpc_list)
 987 {
 988         struct list_head *tmp, *pos;
 989         int rc = 0, retval = 0;
 990         ENTRY;
 991
 992         /* It's possible to receive a completion AST before we've set
 993          * the l_completion_ast pointer: either because the AST arrived
 994          * before the reply, or simply because there's a small race
 995          * window between receiving the reply and finishing the local
 996          * enqueue. (bug 842)
 997          *
 998          * This can't happen with the blocking_ast, however, because we
 999          * will never call the local blocking_ast until we drop our
1000          * reader/writer reference, which we won't do until we get the
1001          * reply and finish enqueueing. */
1002
1003         list_for_each_safe(tmp, pos, rpc_list) {
1004                 struct ldlm_lock *lock =
1005                         list_entry(tmp, struct ldlm_lock, l_cp_ast);
1006
1007                 /* nobody should touch l_cp_ast */
1008                 lock_res(lock->l_resource);
1009                 list_del_init(&lock->l_cp_ast);
1010                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1011                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1012                 unlock_res(lock->l_resource);
1013
1014                 if (lock->l_completion_ast != NULL)
1015                         rc = lock->l_completion_ast(lock, 0, 0);
1016
1017                 if (rc == -ERESTART)
1018                         retval = rc;
1019                 else if (rc)
1020                         CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
1021                                "disconnect client\n");
1022                 LDLM_LOCK_PUT(lock);
1023         }
1024         RETURN(retval);
1025 }
1026
1027 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1028 {
1029         ldlm_reprocess_all(res);
1030         return LDLM_ITER_CONTINUE;
1031 }
1032
1033 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1034 {
1035         struct list_head *tmp;
1036         int i, rc;
1037
1038         spin_lock(&ns->ns_hash_lock);
1039         for (i = 0; i < RES_HASH_SIZE; i++) {
1040                 tmp = ns->ns_hash[i].next;
1041                 while (tmp != &(ns->ns_hash[i])) {
1042                         struct ldlm_resource *res =
1043                                 list_entry(tmp, struct ldlm_resource, lr_hash);
1044
1045                         ldlm_resource_getref(res);
1046                         spin_unlock(&ns->ns_hash_lock);
1047
1048                         rc = reprocess_one_queue(res, NULL);
1049
1050                         spin_lock(&ns->ns_hash_lock);
1051                         tmp = tmp->next;
1052                         ldlm_resource_putref_locked(res);
1053
1054                         if (rc == LDLM_ITER_STOP)
1055                                 GOTO(out, rc);
1056                 }
1057         }
1058  out:
1059         spin_unlock(&ns->ns_hash_lock);
1060         EXIT;
1061 }
1062
1063 void ldlm_reprocess_all(struct ldlm_resource *res)
1064 {
1065         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1066         int rc;
1067         ENTRY;
1068
1069         /* Local lock trees don't get reprocessed. */
1070         if (res->lr_namespace->ns_client) {
1071                 EXIT;
1072                 return;
1073         }
1074
1075  restart:
1076         lock_res(res);
1077         rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1078         if (rc == LDLM_ITER_CONTINUE)
1079                 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1080         unlock_res(res);
1081
1082         rc = ldlm_run_cp_ast_work(&rpc_list);
1083         if (rc == -ERESTART) {
1084                 LASSERT(list_empty(&rpc_list));
1085                 goto restart;
1086         }
1087         EXIT;
1088 }
1089
1090 void ldlm_cancel_callback(struct ldlm_lock *lock)
1091 {
1092         check_res_locked(lock->l_resource);
1093
1094         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1095                 lock->l_flags |= LDLM_FL_CANCEL;
1096                 if (lock->l_blocking_ast) {
1097                         unlock_res(lock->l_resource);
1098                         lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1099                                              LDLM_CB_CANCELING);
1100                         lock_res(lock->l_resource);
1101                 } else {
1102                         LDLM_DEBUG(lock, "no blocking ast");
1103                 }
1104         }
1105 }
1106
1107 void ldlm_lock_cancel(struct ldlm_lock *lock)
1108 {
1109         struct ldlm_resource *res;
1110         struct ldlm_namespace *ns;
1111         ENTRY;
1112
1113         res = lock->l_resource;
1114         ns = res->lr_namespace;
1115
1116         ldlm_del_waiting_lock(lock);
1117         lock_res(res);
1118
1119         /* Please do not, no matter how tempting, remove this LBUG without
1120          * talking to me first. -phik */
1121         if (lock->l_readers || lock->l_writers) {
1122                 LDLM_ERROR(lock, "lock still has references");
1123                 LBUG();
1124         }
1125
1126         ldlm_cancel_callback(lock);
1127
1128         ldlm_resource_unlink_lock(lock);
1129         unlock_res(res);
1130
1131         ldlm_lock_destroy(lock);
1132
1133         EXIT;
1134 }
1135
1136 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1137 {
1138         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1139         ENTRY;
1140
1141         if (lock == NULL)
1142                 RETURN(-EINVAL);
1143
1144         lock->l_ast_data = data;
1145         LDLM_LOCK_PUT(lock);
1146         RETURN(0);
1147 }
1148
1149 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1150 {
1151         struct ldlm_lock *lock;
1152         struct ldlm_resource *res;
1153
1154         spin_lock(&exp->exp_ldlm_data.led_lock);
1155         while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1156                 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1157                                   struct ldlm_lock, l_export_chain);
1158                 res = ldlm_resource_getref(lock->l_resource);
1159                 LDLM_LOCK_GET(lock);
1160                 spin_unlock(&exp->exp_ldlm_data.led_lock);
1161
1162                 LDLM_DEBUG(lock, "export %p", exp);
1163                 ldlm_lock_cancel(lock);
1164                 ldlm_reprocess_all(res);
1165
1166                 ldlm_resource_putref(res);
1167                 LDLM_LOCK_PUT(lock);
1168                 spin_lock(&exp->exp_ldlm_data.led_lock);
1169         }
1170         spin_unlock(&exp->exp_ldlm_data.led_lock);
1171 }
1172
1173 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1174                                         int *flags)
1175 {
1176         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1177         struct ldlm_resource *res;
1178         struct ldlm_namespace *ns;
1179         int granted = 0;
1180         int old_mode, rc;
1181         ldlm_error_t err;
1182         ENTRY;
1183
1184         if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1185                 *flags |= LDLM_FL_BLOCK_GRANTED;
1186                 RETURN(lock->l_resource);
1187         }
1188
1189         LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1190                  "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1191
1192         res = lock->l_resource;
1193         ns = res->lr_namespace;
1194
1195         lock_res(res);
1196
1197         old_mode = lock->l_req_mode;
1198         lock->l_req_mode = new_mode;
1199         ldlm_resource_unlink_lock(lock);
1200
1201         /* If this is a local resource, put it on the appropriate list. */
1202         if (res->lr_namespace->ns_client) {
1203                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1204                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1205                 } else {
1206                         /* This should never happen, because of the way the
1207                          * server handles conversions. */
1208                         LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1209                                    *flags);
1210                         LBUG();
1211
1212                         ldlm_grant_lock(lock, &rpc_list);
1213                         granted = 1;
1214                         /* FIXME: completion handling not with ns_lock held ! */
1215                         if (lock->l_completion_ast)
1216                                 lock->l_completion_ast(lock, 0, NULL);
1217                 }
1218         } else {
1219                 int pflags = 0;
1220                 ldlm_processing_policy policy;
1221                 policy = ldlm_processing_policy_table[res->lr_type];
1222                 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1223                 if (rc == LDLM_ITER_STOP) {
1224                         lock->l_req_mode = old_mode;
1225                         ldlm_resource_add_lock(res, &res->lr_granted, lock);
1226                         res = NULL;
1227                 } else {
1228                         *flags |= LDLM_FL_BLOCK_GRANTED;
1229                         granted = 1;
1230                 }
1231         }
1232         unlock_res(lock->l_resource);
1233
1234         if (granted)
1235                 ldlm_run_cp_ast_work(&rpc_list);
1236         RETURN(res);
1237 }
1238
1239 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1240 {
1241         char str[PTL_NALFMT_SIZE];
1242         struct obd_device *obd = NULL;
1243
1244         if (!((portal_debug | D_ERROR) & level))
1245                 return;
1246
1247         if (!lock) {
1248                 CDEBUG(level, "  NULL LDLM lock\n");
1249                 return;
1250         }
1251
1252         CDEBUG(level, "  -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1253                lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1254                pos, lock->l_pid);
1255         if (lock->l_conn_export != NULL)
1256                 obd = lock->l_conn_export->exp_obd;
1257         if (lock->l_export && lock->l_export->exp_connection) {
1258                 CDEBUG(level, "  Node: NID %s on %s (rhandle: "LPX64")\n",
1259                        ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str),
1260                        lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1261                        lock->l_remote_handle.cookie);
1262         } else if (obd == NULL) {
1263                 CDEBUG(level, "  Node: local\n");
1264         } else {
1265                 struct obd_import *imp = obd->u.cli.cl_import;
1266                 CDEBUG(level, "  Node: NID %s on %s (rhandle: "LPX64")\n",
1267                        ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1268                        imp->imp_connection->c_peer.peer_ni->pni_name,
1269                        lock->l_remote_handle.cookie);
1270         }
1271         CDEBUG(level, "  Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource,
1272                lock->l_resource->lr_name.name[0],
1273                lock->l_resource->lr_name.name[1]);
1274         CDEBUG(level, "  Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1275                "write: %d\n", ldlm_lockname[lock->l_req_mode],
1276                ldlm_lockname[lock->l_granted_mode],
1277                atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers);
1278         if (lock->l_resource->lr_type == LDLM_EXTENT)
1279                 CDEBUG(level, "  Extent: "LPU64" -> "LPU64
1280                        " (req "LPU64"-"LPU64")\n",
1281                        lock->l_policy_data.l_extent.start,
1282                        lock->l_policy_data.l_extent.end,
1283                        lock->l_req_extent.start, lock->l_req_extent.end);
1284         else if (lock->l_resource->lr_type == LDLM_FLOCK)
1285                 CDEBUG(level, "  Pid: "LPU64" Extent: "LPU64" -> "LPU64"\n",
1286                        lock->l_policy_data.l_flock.pid,
1287                        lock->l_policy_data.l_flock.start,
1288                        lock->l_policy_data.l_flock.end);
1289         else if (lock->l_resource->lr_type == LDLM_IBITS)
1290                 CDEBUG(level, " Bits: "LPX64"\n",
1291                        lock->l_policy_data.l_inodebits.bits);
1292 }
1293
1294 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1295 {
1296         struct ldlm_lock *lock;
1297
1298         lock = ldlm_handle2lock(lockh);
1299         if (lock == NULL)
1300                 return;
1301
1302         ldlm_lock_dump(D_OTHER, lock, 0);
1303
1304         LDLM_LOCK_PUT(lock);
1305 }