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  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011 Whamcloud, Inc.
  33  *
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * lustre/ldlm/ldlm_lock.c
  40  *
  41  * Author: Peter Braam <braam@clusterfs.com>
  42  * Author: Phil Schwan <phil@clusterfs.com>
  43  */
  44
  45 #define DEBUG_SUBSYSTEM S_LDLM
  46
  47 #ifdef __KERNEL__
  48 # include <libcfs/libcfs.h>
  49 # include <linux/lustre_intent.h>
  50 #else
  51 # include <liblustre.h>
  52 #endif
  53
  54 #include <obd_class.h>
  55 #include "ldlm_internal.h"
  56
  57 /* lock types */
  58 char *ldlm_lockname[] = {
  59         [0] "--",
  60         [LCK_EX] "EX",
  61         [LCK_PW] "PW",
  62         [LCK_PR] "PR",
  63         [LCK_CW] "CW",
  64         [LCK_CR] "CR",
  65         [LCK_NL] "NL",
  66         [LCK_GROUP] "GROUP",
  67         [LCK_COS] "COS"
  68 };
  69
  70 char *ldlm_typename[] = {
  71         [LDLM_PLAIN] "PLN",
  72         [LDLM_EXTENT] "EXT",
  73         [LDLM_FLOCK] "FLK",
  74         [LDLM_IBITS] "IBT",
  75 };
  76
  77 static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = {
  78         [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_wire_to_local,
  79         [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_wire_to_local,
  80         [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_wire_to_local,
  81         [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_wire_to_local,
  82 };
  83
  84 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
  85         [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_local_to_wire,
  86         [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_local_to_wire,
  87         [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_local_to_wire,
  88         [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_local_to_wire,
  89 };
  90
  91 /**
  92  * Converts lock policy from local format to on the wire lock_desc format
  93  */
  94 void ldlm_convert_policy_to_wire(ldlm_type_t type,
  95                                  const ldlm_policy_data_t *lpolicy,
  96                                  ldlm_wire_policy_data_t *wpolicy)
  97 {
  98         ldlm_policy_local_to_wire_t convert;
  99
 100         convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
 101
 102         convert(lpolicy, wpolicy);
 103 }
 104
 105 /**
 106  * Converts lock policy from on the wire lock_desc format to local format
 107  */
 108 void ldlm_convert_policy_to_local(ldlm_type_t type,
 109                                   const ldlm_wire_policy_data_t *wpolicy,
 110                                   ldlm_policy_data_t *lpolicy)
 111 {
 112         ldlm_policy_wire_to_local_t convert;
 113
 114         convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE];
 115
 116         convert(wpolicy, lpolicy);
 117 }
 118
 119 char *ldlm_it2str(int it)
 120 {
 121         switch (it) {
 122         case IT_OPEN:
 123                 return "open";
 124         case IT_CREAT:
 125                 return "creat";
 126         case (IT_OPEN | IT_CREAT):
 127                 return "open|creat";
 128         case IT_READDIR:
 129                 return "readdir";
 130         case IT_GETATTR:
 131                 return "getattr";
 132         case IT_LOOKUP:
 133                 return "lookup";
 134         case IT_UNLINK:
 135                 return "unlink";
 136         case IT_GETXATTR:
 137                 return "getxattr";
 138         case IT_LAYOUT:
 139                 return "layout";
 140         default:
 141                 CERROR("Unknown intent %d\n", it);
 142                 return "UNKNOWN";
 143         }
 144 }
 145
 146 extern cfs_mem_cache_t *ldlm_lock_slab;
 147
 148 static ldlm_processing_policy ldlm_processing_policy_table[] = {
 149         [LDLM_PLAIN] ldlm_process_plain_lock,
 150         [LDLM_EXTENT] ldlm_process_extent_lock,
 151 #ifdef __KERNEL__
 152         [LDLM_FLOCK] ldlm_process_flock_lock,
 153 #endif
 154         [LDLM_IBITS] ldlm_process_inodebits_lock,
 155 };
 156
 157 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
 158 {
 159         return ldlm_processing_policy_table[res->lr_type];
 160 }
 161
 162 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
 163 {
 164         ns->ns_policy = arg;
 165 }
 166
 167 /*
 168  * REFCOUNTED LOCK OBJECTS
 169  */
 170
 171
 172 /*
 173  * Lock refcounts, during creation:
 174  *   - one special one for allocation, dec'd only once in destroy
 175  *   - one for being a lock that's in-use
 176  *   - one for the addref associated with a new lock
 177  */
 178 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 179 {
 180         cfs_atomic_inc(&lock->l_refc);
 181         return lock;
 182 }
 183
 184 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
 185 {
 186         LASSERT(size == sizeof(*lock));
 187         OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
 188 }
 189
 190 void ldlm_lock_put(struct ldlm_lock *lock)
 191 {
 192         ENTRY;
 193
 194         LASSERT(lock->l_resource != LP_POISON);
 195         LASSERT(cfs_atomic_read(&lock->l_refc) > 0);
 196         if (cfs_atomic_dec_and_test(&lock->l_refc)) {
 197                 struct ldlm_resource *res;
 198
 199                 LDLM_DEBUG(lock,
 200                            "final lock_put on destroyed lock, freeing it.");
 201
 202                 res = lock->l_resource;
 203                 LASSERT(lock->l_destroyed);
 204                 LASSERT(cfs_list_empty(&lock->l_res_link));
 205                 LASSERT(cfs_list_empty(&lock->l_pending_chain));
 206
 207                 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
 208                                      LDLM_NSS_LOCKS);
 209                 lu_ref_del(&res->lr_reference, "lock", lock);
 210                 ldlm_resource_putref(res);
 211                 lock->l_resource = NULL;
 212                 if (lock->l_export) {
 213                         class_export_lock_put(lock->l_export, lock);
 214                         lock->l_export = NULL;
 215                 }
 216
 217                 if (lock->l_lvb_data != NULL)
 218                         OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
 219
 220                 ldlm_interval_free(ldlm_interval_detach(lock));
 221                 lu_ref_fini(&lock->l_reference);
 222                 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
 223                                 ldlm_lock_free);
 224         }
 225
 226         EXIT;
 227 }
 228
 229 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
 230 {
 231         int rc = 0;
 232         if (!cfs_list_empty(&lock->l_lru)) {
 233                 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
 234
 235                 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
 236                 cfs_list_del_init(&lock->l_lru);
 237                 if (lock->l_flags & LDLM_FL_SKIPPED)
 238                         lock->l_flags &= ~LDLM_FL_SKIPPED;
 239                 LASSERT(ns->ns_nr_unused > 0);
 240                 ns->ns_nr_unused--;
 241                 rc = 1;
 242         }
 243         return rc;
 244 }
 245
 246 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
 247 {
 248         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
 249         int rc;
 250
 251         ENTRY;
 252         if (lock->l_ns_srv) {
 253                 LASSERT(cfs_list_empty(&lock->l_lru));
 254                 RETURN(0);
 255         }
 256
 257         cfs_spin_lock(&ns->ns_lock);
 258         rc = ldlm_lock_remove_from_lru_nolock(lock);
 259         cfs_spin_unlock(&ns->ns_lock);
 260         EXIT;
 261         return rc;
 262 }
 263
 264 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
 265 {
 266         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
 267
 268         lock->l_last_used = cfs_time_current();
 269         LASSERT(cfs_list_empty(&lock->l_lru));
 270         LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
 271         cfs_list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 272         LASSERT(ns->ns_nr_unused >= 0);
 273         ns->ns_nr_unused++;
 274 }
 275
 276 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
 277 {
 278         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
 279
 280         ENTRY;
 281         cfs_spin_lock(&ns->ns_lock);
 282         ldlm_lock_add_to_lru_nolock(lock);
 283         cfs_spin_unlock(&ns->ns_lock);
 284         EXIT;
 285 }
 286
 287 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
 288 {
 289         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
 290
 291         ENTRY;
 292         if (lock->l_ns_srv) {
 293                 LASSERT(cfs_list_empty(&lock->l_lru));
 294                 EXIT;
 295                 return;
 296         }
 297
 298         cfs_spin_lock(&ns->ns_lock);
 299         if (!cfs_list_empty(&lock->l_lru)) {
 300                 ldlm_lock_remove_from_lru_nolock(lock);
 301                 ldlm_lock_add_to_lru_nolock(lock);
 302         }
 303         cfs_spin_unlock(&ns->ns_lock);
 304         EXIT;
 305 }
 306
 307 /* This used to have a 'strict' flag, which recovery would use to mark an
 308  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
 309  * shall explain why it's gone: with the new hash table scheme, once you call
 310  * ldlm_lock_destroy, you can never drop your final references on this lock.
 311  * Because it's not in the hash table anymore.  -phil */
 312 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
 313 {
 314         ENTRY;
 315
 316         if (lock->l_readers || lock->l_writers) {
 317                 LDLM_ERROR(lock, "lock still has references");
 318                 ldlm_lock_dump(D_ERROR, lock, 0);
 319                 LBUG();
 320         }
 321
 322         if (!cfs_list_empty(&lock->l_res_link)) {
 323                 LDLM_ERROR(lock, "lock still on resource");
 324                 ldlm_lock_dump(D_ERROR, lock, 0);
 325                 LBUG();
 326         }
 327
 328         if (lock->l_destroyed) {
 329                 LASSERT(cfs_list_empty(&lock->l_lru));
 330                 EXIT;
 331                 return 0;
 332         }
 333         lock->l_destroyed = 1;
 334
 335         if (lock->l_export && lock->l_export->exp_lock_hash &&
 336             !cfs_hlist_unhashed(&lock->l_exp_hash))
 337                 cfs_hash_del(lock->l_export->exp_lock_hash,
 338                              &lock->l_remote_handle, &lock->l_exp_hash);
 339
 340         ldlm_lock_remove_from_lru(lock);
 341         class_handle_unhash(&lock->l_handle);
 342
 343 #if 0
 344         /* Wake anyone waiting for this lock */
 345         /* FIXME: I should probably add yet another flag, instead of using
 346          * l_export to only call this on clients */
 347         if (lock->l_export)
 348                 class_export_put(lock->l_export);
 349         lock->l_export = NULL;
 350         if (lock->l_export && lock->l_completion_ast)
 351                 lock->l_completion_ast(lock, 0);
 352 #endif
 353         EXIT;
 354         return 1;
 355 }
 356
 357 void ldlm_lock_destroy(struct ldlm_lock *lock)
 358 {
 359         int first;
 360         ENTRY;
 361         lock_res_and_lock(lock);
 362         first = ldlm_lock_destroy_internal(lock);
 363         unlock_res_and_lock(lock);
 364
 365         /* drop reference from hashtable only for first destroy */
 366         if (first) {
 367                 lu_ref_del(&lock->l_reference, "hash", lock);
 368                 LDLM_LOCK_RELEASE(lock);
 369         }
 370         EXIT;
 371 }
 372
 373 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
 374 {
 375         int first;
 376         ENTRY;
 377         first = ldlm_lock_destroy_internal(lock);
 378         /* drop reference from hashtable only for first destroy */
 379         if (first) {
 380                 lu_ref_del(&lock->l_reference, "hash", lock);
 381                 LDLM_LOCK_RELEASE(lock);
 382         }
 383         EXIT;
 384 }
 385
 386 /* this is called by portals_handle2object with the handle lock taken */
 387 static void lock_handle_addref(void *lock)
 388 {
 389         LDLM_LOCK_GET((struct ldlm_lock *)lock);
 390 }
 391
 392 /*
 393  * usage: pass in a resource on which you have done ldlm_resource_get
 394  *        new lock will take over the refcount.
 395  * returns: lock with refcount 2 - one for current caller and one for remote
 396  */
 397 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
 398 {
 399         struct ldlm_lock *lock;
 400         ENTRY;
 401
 402         if (resource == NULL)
 403                 LBUG();
 404
 405         OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, CFS_ALLOC_IO);
 406         if (lock == NULL)
 407                 RETURN(NULL);
 408
 409         cfs_spin_lock_init(&lock->l_lock);
 410         lock->l_resource = resource;
 411         lu_ref_add(&resource->lr_reference, "lock", lock);
 412
 413         cfs_atomic_set(&lock->l_refc, 2);
 414         CFS_INIT_LIST_HEAD(&lock->l_res_link);
 415         CFS_INIT_LIST_HEAD(&lock->l_lru);
 416         CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
 417         CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
 418         CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
 419         CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
 420         cfs_waitq_init(&lock->l_waitq);
 421         lock->l_blocking_lock = NULL;
 422         CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
 423         CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
 424         CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
 425
 426         lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
 427                              LDLM_NSS_LOCKS);
 428         CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
 429         class_handle_hash(&lock->l_handle, lock_handle_addref);
 430
 431         lu_ref_init(&lock->l_reference);
 432         lu_ref_add(&lock->l_reference, "hash", lock);
 433         lock->l_callback_timeout = 0;
 434
 435 #if LUSTRE_TRACKS_LOCK_EXP_REFS
 436         CFS_INIT_LIST_HEAD(&lock->l_exp_refs_link);
 437         lock->l_exp_refs_nr = 0;
 438         lock->l_exp_refs_target = NULL;
 439 #endif
 440
 441         RETURN(lock);
 442 }
 443
 444 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
 445                               const struct ldlm_res_id *new_resid)
 446 {
 447         struct ldlm_resource *oldres = lock->l_resource;
 448         struct ldlm_resource *newres;
 449         int type;
 450         ENTRY;
 451
 452         LASSERT(ns_is_client(ns));
 453
 454         lock_res_and_lock(lock);
 455         if (memcmp(new_resid, &lock->l_resource->lr_name,
 456                    sizeof(lock->l_resource->lr_name)) == 0) {
 457                 /* Nothing to do */
 458                 unlock_res_and_lock(lock);
 459                 RETURN(0);
 460         }
 461
 462         LASSERT(new_resid->name[0] != 0);
 463
 464         /* This function assumes that the lock isn't on any lists */
 465         LASSERT(cfs_list_empty(&lock->l_res_link));
 466
 467         type = oldres->lr_type;
 468         unlock_res_and_lock(lock);
 469
 470         newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
 471         if (newres == NULL)
 472                 RETURN(-ENOMEM);
 473
 474         lu_ref_add(&newres->lr_reference, "lock", lock);
 475         /*
 476          * To flip the lock from the old to the new resource, lock, oldres and
 477          * newres have to be locked. Resource spin-locks are nested within
 478          * lock->l_lock, and are taken in the memory address order to avoid
 479          * dead-locks.
 480          */
 481         cfs_spin_lock(&lock->l_lock);
 482         oldres = lock->l_resource;
 483         if (oldres < newres) {
 484                 lock_res(oldres);
 485                 lock_res_nested(newres, LRT_NEW);
 486         } else {
 487                 lock_res(newres);
 488                 lock_res_nested(oldres, LRT_NEW);
 489         }
 490         LASSERT(memcmp(new_resid, &oldres->lr_name,
 491                        sizeof oldres->lr_name) != 0);
 492         lock->l_resource = newres;
 493         unlock_res(oldres);
 494         unlock_res_and_lock(lock);
 495
 496         /* ...and the flowers are still standing! */
 497         lu_ref_del(&oldres->lr_reference, "lock", lock);
 498         ldlm_resource_putref(oldres);
 499
 500         RETURN(0);
 501 }
 502
 503 /*
 504  *  HANDLES
 505  */
 506
 507 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
 508 {
 509         lockh->cookie = lock->l_handle.h_cookie;
 510 }
 511
 512 /* if flags: atomically get the lock and set the flags.
 513  *           Return NULL if flag already set
 514  */
 515
 516 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
 517                                      int flags)
 518 {
 519         struct ldlm_lock *lock;
 520         ENTRY;
 521
 522         LASSERT(handle);
 523
 524         lock = class_handle2object(handle->cookie);
 525         if (lock == NULL)
 526                 RETURN(NULL);
 527
 528         /* It's unlikely but possible that someone marked the lock as
 529          * destroyed after we did handle2object on it */
 530         if (flags == 0 && !lock->l_destroyed) {
 531                 lu_ref_add(&lock->l_reference, "handle", cfs_current());
 532                 RETURN(lock);
 533         }
 534
 535         lock_res_and_lock(lock);
 536
 537         LASSERT(lock->l_resource != NULL);
 538
 539         lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
 540         if (unlikely(lock->l_destroyed)) {
 541                 unlock_res_and_lock(lock);
 542                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
 543                 LDLM_LOCK_PUT(lock);
 544                 RETURN(NULL);
 545         }
 546
 547         if (flags && (lock->l_flags & flags)) {
 548                 unlock_res_and_lock(lock);
 549                 LDLM_LOCK_PUT(lock);
 550                 RETURN(NULL);
 551         }
 552
 553         if (flags)
 554                 lock->l_flags |= flags;
 555
 556         unlock_res_and_lock(lock);
 557         RETURN(lock);
 558 }
 559
 560 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 561 {
 562         struct obd_export *exp = lock->l_export?:lock->l_conn_export;
 563         /* INODEBITS_INTEROP: If the other side does not support
 564          * inodebits, reply with a plain lock descriptor.
 565          */
 566         if ((lock->l_resource->lr_type == LDLM_IBITS) &&
 567             (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
 568                 /* Make sure all the right bits are set in this lock we
 569                    are going to pass to client */
 570                 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
 571                          (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
 572                           MDS_INODELOCK_LAYOUT),
 573                          "Inappropriate inode lock bits during "
 574                          "conversion " LPU64 "\n",
 575                          lock->l_policy_data.l_inodebits.bits);
 576
 577                 ldlm_res2desc(lock->l_resource, &desc->l_resource);
 578                 desc->l_resource.lr_type = LDLM_PLAIN;
 579
 580                 /* Convert "new" lock mode to something old client can
 581                    understand */
 582                 if ((lock->l_req_mode == LCK_CR) ||
 583                     (lock->l_req_mode == LCK_CW))
 584                         desc->l_req_mode = LCK_PR;
 585                 else
 586                         desc->l_req_mode = lock->l_req_mode;
 587                 if ((lock->l_granted_mode == LCK_CR) ||
 588                     (lock->l_granted_mode == LCK_CW)) {
 589                         desc->l_granted_mode = LCK_PR;
 590                 } else {
 591                         /* We never grant PW/EX locks to clients */
 592                         LASSERT((lock->l_granted_mode != LCK_PW) &&
 593                                 (lock->l_granted_mode != LCK_EX));
 594                         desc->l_granted_mode = lock->l_granted_mode;
 595                 }
 596
 597                 /* We do not copy policy here, because there is no
 598                    policy for plain locks */
 599         } else {
 600                 ldlm_res2desc(lock->l_resource, &desc->l_resource);
 601                 desc->l_req_mode = lock->l_req_mode;
 602                 desc->l_granted_mode = lock->l_granted_mode;
 603                 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
 604                                             &lock->l_policy_data,
 605                                             &desc->l_policy_data);
 606         }
 607 }
 608
 609 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
 610                            cfs_list_t *work_list)
 611 {
 612         if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
 613                 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
 614                 lock->l_flags |= LDLM_FL_AST_SENT;
 615                 /* If the enqueuing client said so, tell the AST recipient to
 616                  * discard dirty data, rather than writing back. */
 617                 if (new->l_flags & LDLM_AST_DISCARD_DATA)
 618                         lock->l_flags |= LDLM_FL_DISCARD_DATA;
 619                 LASSERT(cfs_list_empty(&lock->l_bl_ast));
 620                 cfs_list_add(&lock->l_bl_ast, work_list);
 621                 LDLM_LOCK_GET(lock);
 622                 LASSERT(lock->l_blocking_lock == NULL);
 623                 lock->l_blocking_lock = LDLM_LOCK_GET(new);
 624         }
 625 }
 626
 627 void ldlm_add_cp_work_item(struct ldlm_lock *lock, cfs_list_t *work_list)
 628 {
 629         if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
 630                 lock->l_flags |= LDLM_FL_CP_REQD;
 631                 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
 632                 LASSERT(cfs_list_empty(&lock->l_cp_ast));
 633                 cfs_list_add(&lock->l_cp_ast, work_list);
 634                 LDLM_LOCK_GET(lock);
 635         }
 636 }
 637
 638 /* must be called with lr_lock held */
 639 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
 640                             cfs_list_t *work_list)
 641 {
 642         ENTRY;
 643         check_res_locked(lock->l_resource);
 644         if (new)
 645                 ldlm_add_bl_work_item(lock, new, work_list);
 646         else
 647                 ldlm_add_cp_work_item(lock, work_list);
 648         EXIT;
 649 }
 650
 651 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 652 {
 653         struct ldlm_lock *lock;
 654
 655         lock = ldlm_handle2lock(lockh);
 656         LASSERT(lock != NULL);
 657         ldlm_lock_addref_internal(lock, mode);
 658         LDLM_LOCK_PUT(lock);
 659 }
 660
 661 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
 662 {
 663         ldlm_lock_remove_from_lru(lock);
 664         if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
 665                 lock->l_readers++;
 666                 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
 667         }
 668         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
 669                 lock->l_writers++;
 670                 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
 671         }
 672         LDLM_LOCK_GET(lock);
 673         lu_ref_add_atomic(&lock->l_reference, "user", lock);
 674         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 675 }
 676
 677 /**
 678  * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
 679  * or destroyed.
 680  *
 681  * \retval 0 success, lock was addref-ed
 682  *
 683  * \retval -EAGAIN lock is being canceled.
 684  */
 685 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
 686 {
 687         struct ldlm_lock *lock;
 688         int               result;
 689
 690         result = -EAGAIN;
 691         lock = ldlm_handle2lock(lockh);
 692         if (lock != NULL) {
 693                 lock_res_and_lock(lock);
 694                 if (lock->l_readers != 0 || lock->l_writers != 0 ||
 695                     !(lock->l_flags & LDLM_FL_CBPENDING)) {
 696                         ldlm_lock_addref_internal_nolock(lock, mode);
 697                         result = 0;
 698                 }
 699                 unlock_res_and_lock(lock);
 700                 LDLM_LOCK_PUT(lock);
 701         }
 702         return result;
 703 }
 704
 705 /* only called for local locks */
 706 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 707 {
 708         lock_res_and_lock(lock);
 709         ldlm_lock_addref_internal_nolock(lock, mode);
 710         unlock_res_and_lock(lock);
 711 }
 712
 713 /* only called in ldlm_flock_destroy and for local locks.
 714  *  * for LDLM_FLOCK type locks, l_blocking_ast is null, and
 715  *   * ldlm_lock_remove_from_lru() does nothing, it is safe
 716  *    * for ldlm_flock_destroy usage by dropping some code */
 717 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
 718 {
 719         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 720         if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
 721                 LASSERT(lock->l_readers > 0);
 722                 lu_ref_del(&lock->l_reference, "reader", lock);
 723                 lock->l_readers--;
 724         }
 725         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
 726                 LASSERT(lock->l_writers > 0);
 727                 lu_ref_del(&lock->l_reference, "writer", lock);
 728                 lock->l_writers--;
 729         }
 730
 731         lu_ref_del(&lock->l_reference, "user", lock);
 732         LDLM_LOCK_RELEASE(lock);    /* matches the LDLM_LOCK_GET() in addref */
 733 }
 734
 735 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
 736 {
 737         struct ldlm_namespace *ns;
 738         ENTRY;
 739
 740         lock_res_and_lock(lock);
 741
 742         ns = ldlm_lock_to_ns(lock);
 743
 744         ldlm_lock_decref_internal_nolock(lock, mode);
 745
 746         if (lock->l_flags & LDLM_FL_LOCAL &&
 747             !lock->l_readers && !lock->l_writers) {
 748                 /* If this is a local lock on a server namespace and this was
 749                  * the last reference, cancel the lock. */
 750                 CDEBUG(D_INFO, "forcing cancel of local lock\n");
 751                 lock->l_flags |= LDLM_FL_CBPENDING;
 752         }
 753
 754         if (!lock->l_readers && !lock->l_writers &&
 755             (lock->l_flags & LDLM_FL_CBPENDING)) {
 756                 /* If we received a blocked AST and this was the last reference,
 757                  * run the callback. */
 758                 if (lock->l_ns_srv && lock->l_export)
 759                         CERROR("FL_CBPENDING set on non-local lock--just a "
 760                                "warning\n");
 761
 762                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 763
 764                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
 765                 ldlm_lock_remove_from_lru(lock);
 766                 unlock_res_and_lock(lock);
 767
 768                 if (lock->l_flags & LDLM_FL_FAIL_LOC)
 769                         OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
 770
 771                 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
 772                     ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
 773                         ldlm_handle_bl_callback(ns, NULL, lock);
 774         } else if (ns_is_client(ns) &&
 775                    !lock->l_readers && !lock->l_writers &&
 776                    !(lock->l_flags & LDLM_FL_NO_LRU) &&
 777                    !(lock->l_flags & LDLM_FL_BL_AST)) {
 778
 779                 LDLM_DEBUG(lock, "add lock into lru list");
 780
 781                 /* If this is a client-side namespace and this was the last
 782                  * reference, put it on the LRU. */
 783                 ldlm_lock_add_to_lru(lock);
 784                 unlock_res_and_lock(lock);
 785
 786                 if (lock->l_flags & LDLM_FL_FAIL_LOC)
 787                         OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
 788
 789                 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
 790                  * are not supported by the server, otherwise, it is done on
 791                  * enqueue. */
 792                 if (!exp_connect_cancelset(lock->l_conn_export) &&
 793                     !ns_connect_lru_resize(ns))
 794                         ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
 795         } else {
 796                 LDLM_DEBUG(lock, "do not add lock into lru list");
 797                 unlock_res_and_lock(lock);
 798         }
 799
 800         EXIT;
 801 }
 802
 803 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 804 {
 805         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 806         LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
 807         ldlm_lock_decref_internal(lock, mode);
 808         LDLM_LOCK_PUT(lock);
 809 }
 810
 811 /* This will drop a lock reference and mark it for destruction, but will not
 812  * necessarily cancel the lock before returning. */
 813 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
 814 {
 815         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 816         ENTRY;
 817
 818         LASSERT(lock != NULL);
 819
 820         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 821         lock_res_and_lock(lock);
 822         lock->l_flags |= LDLM_FL_CBPENDING;
 823         unlock_res_and_lock(lock);
 824         ldlm_lock_decref_internal(lock, mode);
 825         LDLM_LOCK_PUT(lock);
 826 }
 827
 828 struct sl_insert_point {
 829         cfs_list_t *res_link;
 830         cfs_list_t *mode_link;
 831         cfs_list_t *policy_link;
 832 };
 833
 834 /*
 835  * search_granted_lock
 836  *
 837  * Description:
 838  *      Finds a position to insert the new lock.
 839  * Parameters:
 840  *      queue [input]:  the granted list where search acts on;
 841  *      req [input]:    the lock whose position to be located;
 842  *      prev [output]:  positions within 3 lists to insert @req to
 843  * Return Value:
 844  *      filled @prev
 845  * NOTE: called by
 846  *  - ldlm_grant_lock_with_skiplist
 847  */
 848 static void search_granted_lock(cfs_list_t *queue,
 849                                 struct ldlm_lock *req,
 850                                 struct sl_insert_point *prev)
 851 {
 852         cfs_list_t *tmp;
 853         struct ldlm_lock *lock, *mode_end, *policy_end;
 854         ENTRY;
 855
 856         cfs_list_for_each(tmp, queue) {
 857                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
 858
 859                 mode_end = cfs_list_entry(lock->l_sl_mode.prev,
 860                                           struct ldlm_lock, l_sl_mode);
 861
 862                 if (lock->l_req_mode != req->l_req_mode) {
 863                         /* jump to last lock of mode group */
 864                         tmp = &mode_end->l_res_link;
 865                         continue;
 866                 }
 867
 868                 /* suitable mode group is found */
 869                 if (lock->l_resource->lr_type == LDLM_PLAIN) {
 870                         /* insert point is last lock of the mode group */
 871                         prev->res_link = &mode_end->l_res_link;
 872                         prev->mode_link = &mode_end->l_sl_mode;
 873                         prev->policy_link = &req->l_sl_policy;
 874                         EXIT;
 875                         return;
 876                 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
 877                         for (;;) {
 878                                 policy_end =
 879                                         cfs_list_entry(lock->l_sl_policy.prev,
 880                                                        struct ldlm_lock,
 881                                                        l_sl_policy);
 882
 883                                 if (lock->l_policy_data.l_inodebits.bits ==
 884                                     req->l_policy_data.l_inodebits.bits) {
 885                                         /* insert point is last lock of
 886                                          * the policy group */
 887                                         prev->res_link =
 888                                                 &policy_end->l_res_link;
 889                                         prev->mode_link =
 890                                                 &policy_end->l_sl_mode;
 891                                         prev->policy_link =
 892                                                 &policy_end->l_sl_policy;
 893                                         EXIT;
 894                                         return;
 895                                 }
 896
 897                                 if (policy_end == mode_end)
 898                                         /* done with mode group */
 899                                         break;
 900
 901                                 /* go to next policy group within mode group */
 902                                 tmp = policy_end->l_res_link.next;
 903                                 lock = cfs_list_entry(tmp, struct ldlm_lock,
 904                                                       l_res_link);
 905                         }  /* loop over policy groups within the mode group */
 906
 907                         /* insert point is last lock of the mode group,
 908                          * new policy group is started */
 909                         prev->res_link = &mode_end->l_res_link;
 910                         prev->mode_link = &mode_end->l_sl_mode;
 911                         prev->policy_link = &req->l_sl_policy;
 912                         EXIT;
 913                         return;
 914                 } else {
 915                         LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
 916                         LBUG();
 917                 }
 918         }
 919
 920         /* insert point is last lock on the queue,
 921          * new mode group and new policy group are started */
 922         prev->res_link = queue->prev;
 923         prev->mode_link = &req->l_sl_mode;
 924         prev->policy_link = &req->l_sl_policy;
 925         EXIT;
 926         return;
 927 }
 928
 929 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
 930                                        struct sl_insert_point *prev)
 931 {
 932         struct ldlm_resource *res = lock->l_resource;
 933         ENTRY;
 934
 935         check_res_locked(res);
 936
 937         ldlm_resource_dump(D_INFO, res);
 938         CDEBUG(D_OTHER, "About to add this lock:\n");
 939         ldlm_lock_dump(D_OTHER, lock, 0);
 940
 941         if (lock->l_destroyed) {
 942                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 943                 return;
 944         }
 945
 946         LASSERT(cfs_list_empty(&lock->l_res_link));
 947         LASSERT(cfs_list_empty(&lock->l_sl_mode));
 948         LASSERT(cfs_list_empty(&lock->l_sl_policy));
 949
 950         cfs_list_add(&lock->l_res_link, prev->res_link);
 951         cfs_list_add(&lock->l_sl_mode, prev->mode_link);
 952         cfs_list_add(&lock->l_sl_policy, prev->policy_link);
 953
 954         EXIT;
 955 }
 956
 957 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
 958 {
 959         struct sl_insert_point prev;
 960         ENTRY;
 961
 962         LASSERT(lock->l_req_mode == lock->l_granted_mode);
 963
 964         search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
 965         ldlm_granted_list_add_lock(lock, &prev);
 966         EXIT;
 967 }
 968
 969 /* NOTE: called by
 970  *  - ldlm_lock_enqueue
 971  *  - ldlm_reprocess_queue
 972  *  - ldlm_lock_convert
 973  *
 974  * must be called with lr_lock held
 975  */
 976 void ldlm_grant_lock(struct ldlm_lock *lock, cfs_list_t *work_list)
 977 {
 978         struct ldlm_resource *res = lock->l_resource;
 979         ENTRY;
 980
 981         check_res_locked(res);
 982
 983         lock->l_granted_mode = lock->l_req_mode;
 984         if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
 985                 ldlm_grant_lock_with_skiplist(lock);
 986         else if (res->lr_type == LDLM_EXTENT)
 987                 ldlm_extent_add_lock(res, lock);
 988         else
 989                 ldlm_resource_add_lock(res, &res->lr_granted, lock);
 990
 991         if (lock->l_granted_mode < res->lr_most_restr)
 992                 res->lr_most_restr = lock->l_granted_mode;
 993
 994         if (work_list && lock->l_completion_ast != NULL)
 995                 ldlm_add_ast_work_item(lock, NULL, work_list);
 996
 997         ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
 998         EXIT;
 999 }
1000
1001 /* returns a referenced lock or NULL.  See the flag descriptions below, in the
1002  * comment above ldlm_lock_match */
1003 static struct ldlm_lock *search_queue(cfs_list_t *queue,
1004                                       ldlm_mode_t *mode,
1005                                       ldlm_policy_data_t *policy,
1006                                       struct ldlm_lock *old_lock,
1007                                       int flags, int unref)
1008 {
1009         struct ldlm_lock *lock;
1010         cfs_list_t       *tmp;
1011
1012         cfs_list_for_each(tmp, queue) {
1013                 ldlm_mode_t match;
1014
1015                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1016
1017                 if (lock == old_lock)
1018                         break;
1019
1020                 /* llite sometimes wants to match locks that will be
1021                  * canceled when their users drop, but we allow it to match
1022                  * if it passes in CBPENDING and the lock still has users.
1023                  * this is generally only going to be used by children
1024                  * whose parents already hold a lock so forward progress
1025                  * can still happen. */
1026                 if (lock->l_flags & LDLM_FL_CBPENDING &&
1027                     !(flags & LDLM_FL_CBPENDING))
1028                         continue;
1029                 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
1030                     lock->l_readers == 0 && lock->l_writers == 0)
1031                         continue;
1032
1033                 if (!(lock->l_req_mode & *mode))
1034                         continue;
1035                 match = lock->l_req_mode;
1036
1037                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
1038                     (lock->l_policy_data.l_extent.start >
1039                      policy->l_extent.start ||
1040                      lock->l_policy_data.l_extent.end < policy->l_extent.end))
1041                         continue;
1042
1043                 if (unlikely(match == LCK_GROUP) &&
1044                     lock->l_resource->lr_type == LDLM_EXTENT &&
1045                     lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
1046                         continue;
1047
1048                 /* We match if we have existing lock with same or wider set
1049                    of bits. */
1050                 if (lock->l_resource->lr_type == LDLM_IBITS &&
1051                      ((lock->l_policy_data.l_inodebits.bits &
1052                       policy->l_inodebits.bits) !=
1053                       policy->l_inodebits.bits))
1054                         continue;
1055
1056                 if (!unref &&
1057                     (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED ||
1058                      lock->l_fail_value != 0))
1059                         continue;
1060
1061                 if ((flags & LDLM_FL_LOCAL_ONLY) &&
1062                     !(lock->l_flags & LDLM_FL_LOCAL))
1063                         continue;
1064
1065                 if (flags & LDLM_FL_TEST_LOCK) {
1066                         LDLM_LOCK_GET(lock);
1067                         ldlm_lock_touch_in_lru(lock);
1068                 } else {
1069                         ldlm_lock_addref_internal_nolock(lock, match);
1070                 }
1071                 *mode = match;
1072                 return lock;
1073         }
1074
1075         return NULL;
1076 }
1077
1078 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock, int rc)
1079 {
1080         if (lock->l_fail_value == 0) {
1081                 lock->l_fail_value = rc;
1082                 cfs_waitq_signal(&lock->l_waitq);
1083         }
1084 }
1085 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1086
1087 void ldlm_lock_fail_match(struct ldlm_lock *lock, int rc)
1088 {
1089         lock_res_and_lock(lock);
1090         ldlm_lock_fail_match_locked(lock, rc);
1091         unlock_res_and_lock(lock);
1092 }
1093 EXPORT_SYMBOL(ldlm_lock_fail_match);
1094
1095 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1096 {
1097         lock->l_flags |= LDLM_FL_LVB_READY;
1098         cfs_waitq_signal(&lock->l_waitq);
1099 }
1100
1101 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1102 {
1103         lock_res_and_lock(lock);
1104         ldlm_lock_allow_match_locked(lock);
1105         unlock_res_and_lock(lock);
1106 }
1107
1108 /* Can be called in two ways:
1109  *
1110  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1111  * for a duplicate of.
1112  *
1113  * Otherwise, all of the fields must be filled in, to match against.
1114  *
1115  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1116  *     server (ie, connh is NULL)
1117  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1118  *     list will be considered
1119  * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1120  *     to be canceled can still be matched as long as they still have reader
1121  *     or writer refernces
1122  * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1123  *     just tell us if we would have matched.
1124  *
1125  * Returns 1 if it finds an already-existing lock that is compatible; in this
1126  * case, lockh is filled in with a addref()ed lock
1127  *
1128  * we also check security context, if that failed we simply return 0 (to keep
1129  * caller code unchanged), the context failure will be discovered by caller
1130  * sometime later.
1131  */
1132 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1133                             const struct ldlm_res_id *res_id, ldlm_type_t type,
1134                             ldlm_policy_data_t *policy, ldlm_mode_t mode,
1135                             struct lustre_handle *lockh, int unref)
1136 {
1137         struct ldlm_resource *res;
1138         struct ldlm_lock *lock, *old_lock = NULL;
1139         int rc = 0;
1140         ENTRY;
1141
1142         if (ns == NULL) {
1143                 old_lock = ldlm_handle2lock(lockh);
1144                 LASSERT(old_lock);
1145
1146                 ns = ldlm_lock_to_ns(old_lock);
1147                 res_id = &old_lock->l_resource->lr_name;
1148                 type = old_lock->l_resource->lr_type;
1149                 mode = old_lock->l_req_mode;
1150         }
1151
1152         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1153         if (res == NULL) {
1154                 LASSERT(old_lock == NULL);
1155                 RETURN(0);
1156         }
1157
1158         LDLM_RESOURCE_ADDREF(res);
1159         lock_res(res);
1160
1161         lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1162                             flags, unref);
1163         if (lock != NULL)
1164                 GOTO(out, rc = 1);
1165         if (flags & LDLM_FL_BLOCK_GRANTED)
1166                 GOTO(out, rc = 0);
1167         lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1168                             flags, unref);
1169         if (lock != NULL)
1170                 GOTO(out, rc = 1);
1171         lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1172                             flags, unref);
1173         if (lock != NULL)
1174                 GOTO(out, rc = 1);
1175
1176         EXIT;
1177  out:
1178         unlock_res(res);
1179         LDLM_RESOURCE_DELREF(res);
1180         ldlm_resource_putref(res);
1181
1182         if (lock) {
1183                 ldlm_lock2handle(lock, lockh);
1184                 if ((flags & LDLM_FL_LVB_READY) &&
1185                     (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1186                         struct l_wait_info lwi;
1187                         if (lock->l_completion_ast) {
1188                                 int err = lock->l_completion_ast(lock,
1189                                                           LDLM_FL_WAIT_NOREPROC,
1190                                                                  NULL);
1191                                 if (err) {
1192                                         if (flags & LDLM_FL_TEST_LOCK)
1193                                                 LDLM_LOCK_RELEASE(lock);
1194                                         else
1195                                                 ldlm_lock_decref_internal(lock,
1196                                                                           mode);
1197                                         rc = 0;
1198                                         goto out2;
1199                                 }
1200                         }
1201
1202                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1203                                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1204
1205                         /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1206                         l_wait_event(lock->l_waitq,
1207                                      lock->l_flags & LDLM_FL_LVB_READY ||
1208                                      lock->l_fail_value != 0,
1209                                      &lwi);
1210                         if (!(lock->l_flags & LDLM_FL_LVB_READY)) {
1211                                 if (flags & LDLM_FL_TEST_LOCK)
1212                                         LDLM_LOCK_RELEASE(lock);
1213                                 else
1214                                         ldlm_lock_decref_internal(lock, mode);
1215                                 rc = 0;
1216                         }
1217                 }
1218         }
1219  out2:
1220         if (rc) {
1221                 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1222                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1223                                 res_id->name[2] : policy->l_extent.start,
1224                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1225                                 res_id->name[3] : policy->l_extent.end);
1226
1227                 /* check user's security context */
1228                 if (lock->l_conn_export &&
1229                     sptlrpc_import_check_ctx(
1230                                 class_exp2cliimp(lock->l_conn_export))) {
1231                         if (!(flags & LDLM_FL_TEST_LOCK))
1232                                 ldlm_lock_decref_internal(lock, mode);
1233                         rc = 0;
1234                 }
1235
1236                 if (flags & LDLM_FL_TEST_LOCK)
1237                         LDLM_LOCK_RELEASE(lock);
1238
1239         } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1240                 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1241                                   LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1242                                   type, mode, res_id->name[0], res_id->name[1],
1243                                   (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1244                                         res_id->name[2] :policy->l_extent.start,
1245                                   (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1246                                         res_id->name[3] : policy->l_extent.end);
1247         }
1248         if (old_lock)
1249                 LDLM_LOCK_PUT(old_lock);
1250
1251         return rc ? mode : 0;
1252 }
1253
1254 ldlm_mode_t ldlm_revalidate_lock_handle(struct lustre_handle *lockh,
1255                                         __u64 *bits)
1256 {
1257         struct ldlm_lock *lock;
1258         ldlm_mode_t mode = 0;
1259         ENTRY;
1260
1261         lock = ldlm_handle2lock(lockh);
1262         if (lock != NULL) {
1263                 lock_res_and_lock(lock);
1264                 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED ||
1265                     lock->l_fail_value != 0)
1266                         GOTO(out, mode);
1267
1268                 if (lock->l_flags & LDLM_FL_CBPENDING &&
1269                     lock->l_readers == 0 && lock->l_writers == 0)
1270                         GOTO(out, mode);
1271
1272                 if (bits)
1273                         *bits = lock->l_policy_data.l_inodebits.bits;
1274                 mode = lock->l_granted_mode;
1275                 ldlm_lock_addref_internal_nolock(lock, mode);
1276         }
1277
1278         EXIT;
1279
1280 out:
1281         if (lock != NULL) {
1282                 unlock_res_and_lock(lock);
1283                 LDLM_LOCK_PUT(lock);
1284         }
1285         return mode;
1286 }
1287 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1288
1289 /* Returns a referenced lock */
1290 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1291                                    const struct ldlm_res_id *res_id,
1292                                    ldlm_type_t type,
1293                                    ldlm_mode_t mode,
1294                                    const struct ldlm_callback_suite *cbs,
1295                                    void *data, __u32 lvb_len)
1296 {
1297         struct ldlm_lock *lock;
1298         struct ldlm_resource *res;
1299         ENTRY;
1300
1301         res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1302         if (res == NULL)
1303                 RETURN(NULL);
1304
1305         lock = ldlm_lock_new(res);
1306
1307         if (lock == NULL)
1308                 RETURN(NULL);
1309
1310         lock->l_req_mode = mode;
1311         lock->l_ast_data = data;
1312         lock->l_pid = cfs_curproc_pid();
1313         lock->l_ns_srv = ns_is_server(ns);
1314         if (cbs) {
1315                 lock->l_blocking_ast = cbs->lcs_blocking;
1316                 lock->l_completion_ast = cbs->lcs_completion;
1317                 lock->l_glimpse_ast = cbs->lcs_glimpse;
1318                 lock->l_weigh_ast = cbs->lcs_weigh;
1319         }
1320
1321         lock->l_tree_node = NULL;
1322         /* if this is the extent lock, allocate the interval tree node */
1323         if (type == LDLM_EXTENT) {
1324                 if (ldlm_interval_alloc(lock) == NULL)
1325                         GOTO(out, 0);
1326         }
1327
1328         if (lvb_len) {
1329                 lock->l_lvb_len = lvb_len;
1330                 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1331                 if (lock->l_lvb_data == NULL)
1332                         GOTO(out, 0);
1333         }
1334
1335         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1336                 GOTO(out, 0);
1337
1338         RETURN(lock);
1339
1340 out:
1341         ldlm_lock_destroy(lock);
1342         LDLM_LOCK_RELEASE(lock);
1343         return NULL;
1344 }
1345
1346 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1347                                struct ldlm_lock **lockp,
1348                                void *cookie, int *flags)
1349 {
1350         struct ldlm_lock *lock = *lockp;
1351         struct ldlm_resource *res = lock->l_resource;
1352         int local = ns_is_client(ldlm_res_to_ns(res));
1353         ldlm_processing_policy policy;
1354         ldlm_error_t rc = ELDLM_OK;
1355         struct ldlm_interval *node = NULL;
1356         ENTRY;
1357
1358         lock->l_last_activity = cfs_time_current_sec();
1359         /* policies are not executed on the client or during replay */
1360         if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1361             && !local && ns->ns_policy) {
1362                 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1363                                    NULL);
1364                 if (rc == ELDLM_LOCK_REPLACED) {
1365                         /* The lock that was returned has already been granted,
1366                          * and placed into lockp.  If it's not the same as the
1367                          * one we passed in, then destroy the old one and our
1368                          * work here is done. */
1369                         if (lock != *lockp) {
1370                                 ldlm_lock_destroy(lock);
1371                                 LDLM_LOCK_RELEASE(lock);
1372                         }
1373                         *flags |= LDLM_FL_LOCK_CHANGED;
1374                         RETURN(0);
1375                 } else if (rc != ELDLM_OK ||
1376                            (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1377                         ldlm_lock_destroy(lock);
1378                         RETURN(rc);
1379                 }
1380         }
1381
1382         /* For a replaying lock, it might be already in granted list. So
1383          * unlinking the lock will cause the interval node to be freed, we
1384          * have to allocate the interval node early otherwise we can't regrant
1385          * this lock in the future. - jay */
1386         if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1387                 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1388
1389         lock_res_and_lock(lock);
1390         if (local && lock->l_req_mode == lock->l_granted_mode) {
1391                 /* The server returned a blocked lock, but it was granted
1392                  * before we got a chance to actually enqueue it.  We don't
1393                  * need to do anything else. */
1394                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1395                             LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1396                 GOTO(out, ELDLM_OK);
1397         }
1398
1399         ldlm_resource_unlink_lock(lock);
1400         if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1401                 if (node == NULL) {
1402                         ldlm_lock_destroy_nolock(lock);
1403                         GOTO(out, rc = -ENOMEM);
1404                 }
1405
1406                 CFS_INIT_LIST_HEAD(&node->li_group);
1407                 ldlm_interval_attach(node, lock);
1408                 node = NULL;
1409         }
1410
1411         /* Some flags from the enqueue want to make it into the AST, via the
1412          * lock's l_flags. */
1413         lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1414
1415         /* This distinction between local lock trees is very important; a client
1416          * namespace only has information about locks taken by that client, and
1417          * thus doesn't have enough information to decide for itself if it can
1418          * be granted (below).  In this case, we do exactly what the server
1419          * tells us to do, as dictated by the 'flags'.
1420          *
1421          * We do exactly the same thing during recovery, when the server is
1422          * more or less trusting the clients not to lie.
1423          *
1424          * FIXME (bug 268): Detect obvious lies by checking compatibility in
1425          * granted/converting queues. */
1426         if (local) {
1427                 if (*flags & LDLM_FL_BLOCK_CONV)
1428                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1429                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1430                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1431                 else
1432                         ldlm_grant_lock(lock, NULL);
1433                 GOTO(out, ELDLM_OK);
1434         } else if (*flags & LDLM_FL_REPLAY) {
1435                 if (*flags & LDLM_FL_BLOCK_CONV) {
1436                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1437                         GOTO(out, ELDLM_OK);
1438                 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1439                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1440                         GOTO(out, ELDLM_OK);
1441                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1442                         ldlm_grant_lock(lock, NULL);
1443                         GOTO(out, ELDLM_OK);
1444                 }
1445                 /* If no flags, fall through to normal enqueue path. */
1446         }
1447
1448         policy = ldlm_processing_policy_table[res->lr_type];
1449         policy(lock, flags, 1, &rc, NULL);
1450         GOTO(out, rc);
1451 out:
1452         unlock_res_and_lock(lock);
1453         if (node)
1454                 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1455         return rc;
1456 }
1457
1458 /* Must be called with namespace taken: queue is waiting or converting. */
1459 int ldlm_reprocess_queue(struct ldlm_resource *res, cfs_list_t *queue,
1460                          cfs_list_t *work_list)
1461 {
1462         cfs_list_t *tmp, *pos;
1463         ldlm_processing_policy policy;
1464         int flags;
1465         int rc = LDLM_ITER_CONTINUE;
1466         ldlm_error_t err;
1467         ENTRY;
1468
1469         check_res_locked(res);
1470
1471         policy = ldlm_processing_policy_table[res->lr_type];
1472         LASSERT(policy);
1473
1474         cfs_list_for_each_safe(tmp, pos, queue) {
1475                 struct ldlm_lock *pending;
1476                 pending = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1477
1478                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1479
1480                 flags = 0;
1481                 rc = policy(pending, &flags, 0, &err, work_list);
1482                 if (rc != LDLM_ITER_CONTINUE)
1483                         break;
1484         }
1485
1486         RETURN(rc);
1487 }
1488
1489 static int
1490 ldlm_work_bl_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1491 {
1492         struct ldlm_lock_desc d;
1493         struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1494                                                 l_bl_ast);
1495         int rc;
1496         ENTRY;
1497
1498         /* nobody should touch l_bl_ast */
1499         lock_res_and_lock(lock);
1500         cfs_list_del_init(&lock->l_bl_ast);
1501
1502         LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1503         LASSERT(lock->l_bl_ast_run == 0);
1504         LASSERT(lock->l_blocking_lock);
1505         lock->l_bl_ast_run++;
1506         unlock_res_and_lock(lock);
1507
1508         ldlm_lock2desc(lock->l_blocking_lock, &d);
1509
1510         rc = lock->l_blocking_ast(lock, &d, (void *)arg,
1511                                   LDLM_CB_BLOCKING);
1512         LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1513         lock->l_blocking_lock = NULL;
1514         LDLM_LOCK_RELEASE(lock);
1515
1516         RETURN(rc);
1517 }
1518
1519 static int
1520 ldlm_work_cp_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1521 {
1522         struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock, l_cp_ast);
1523         ldlm_completion_callback completion_callback;
1524         int rc = 0;
1525         ENTRY;
1526
1527         /* It's possible to receive a completion AST before we've set
1528          * the l_completion_ast pointer: either because the AST arrived
1529          * before the reply, or simply because there's a small race
1530          * window between receiving the reply and finishing the local
1531          * enqueue. (bug 842)
1532          *
1533          * This can't happen with the blocking_ast, however, because we
1534          * will never call the local blocking_ast until we drop our
1535          * reader/writer reference, which we won't do until we get the
1536          * reply and finish enqueueing. */
1537
1538         /* nobody should touch l_cp_ast */
1539         lock_res_and_lock(lock);
1540         cfs_list_del_init(&lock->l_cp_ast);
1541         LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1542         /* save l_completion_ast since it can be changed by
1543          * mds_intent_policy(), see bug 14225 */
1544         completion_callback = lock->l_completion_ast;
1545         lock->l_flags &= ~LDLM_FL_CP_REQD;
1546         unlock_res_and_lock(lock);
1547
1548         if (completion_callback != NULL)
1549                 rc = completion_callback(lock, 0, (void *)arg);
1550         LDLM_LOCK_RELEASE(lock);
1551
1552         RETURN(rc);
1553 }
1554
1555 static int
1556 ldlm_work_revoke_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1557 {
1558         struct ldlm_lock_desc desc;
1559         struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1560                                                 l_rk_ast);
1561         int rc;
1562         ENTRY;
1563
1564         cfs_list_del_init(&lock->l_rk_ast);
1565
1566         /* the desc just pretend to exclusive */
1567         ldlm_lock2desc(lock, &desc);
1568         desc.l_req_mode = LCK_EX;
1569         desc.l_granted_mode = 0;
1570
1571         rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1572         LDLM_LOCK_RELEASE(lock);
1573
1574         RETURN(rc);
1575 }
1576
1577 int ldlm_run_ast_work(struct ldlm_namespace *ns, cfs_list_t *rpc_list,
1578                       ldlm_desc_ast_t ast_type)
1579 {
1580         struct ldlm_cb_set_arg arg = { 0 };
1581         struct l_wait_info     lwi = { 0 };
1582         cfs_list_t *tmp, *pos;
1583         int (*work_ast_lock)(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg);
1584         unsigned int max_ast_count;
1585         ENTRY;
1586
1587         if (cfs_list_empty(rpc_list))
1588                 RETURN(0);
1589
1590         cfs_atomic_set(&arg.restart, 0);
1591         cfs_atomic_set(&arg.rpcs, 0);
1592         cfs_waitq_init(&arg.waitq);
1593
1594         switch (ast_type) {
1595         case LDLM_WORK_BL_AST:
1596                 arg.type = LDLM_BL_CALLBACK;
1597                 work_ast_lock = ldlm_work_bl_ast_lock;
1598                 break;
1599         case LDLM_WORK_CP_AST:
1600                 arg.type = LDLM_CP_CALLBACK;
1601                 work_ast_lock = ldlm_work_cp_ast_lock;
1602                 break;
1603         case LDLM_WORK_REVOKE_AST:
1604                 arg.type = LDLM_BL_CALLBACK;
1605                 work_ast_lock = ldlm_work_revoke_ast_lock;
1606                 break;
1607         default:
1608                 LBUG();
1609         }
1610
1611         max_ast_count = ns->ns_max_parallel_ast ? : UINT_MAX;
1612         arg.threshold = max_ast_count;
1613
1614         cfs_list_for_each_safe(tmp, pos, rpc_list) {
1615                 (void)work_ast_lock(tmp, &arg);
1616                 if (cfs_atomic_read(&arg.rpcs) < max_ast_count)
1617                         continue;
1618
1619                 l_wait_event(arg.waitq,
1620                              cfs_atomic_read(&arg.rpcs) < arg.threshold,
1621                              &lwi);
1622         }
1623
1624         arg.threshold = 1;
1625         l_wait_event(arg.waitq, cfs_atomic_read(&arg.rpcs) == 0, &lwi);
1626
1627         RETURN(cfs_atomic_read(&arg.restart) ? -ERESTART : 0);
1628 }
1629
1630 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1631 {
1632         ldlm_reprocess_all(res);
1633         return LDLM_ITER_CONTINUE;
1634 }
1635
1636 static int ldlm_reprocess_res(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1637                               cfs_hlist_node_t *hnode, void *arg)
1638 {
1639         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1640         int    rc;
1641
1642         rc = reprocess_one_queue(res, arg);
1643
1644         return rc == LDLM_ITER_STOP;
1645 }
1646
1647 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1648 {
1649         ENTRY;
1650
1651         if (ns != NULL) {
1652                 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1653                                          ldlm_reprocess_res, NULL);
1654         }
1655         EXIT;
1656 }
1657
1658 void ldlm_reprocess_all(struct ldlm_resource *res)
1659 {
1660         CFS_LIST_HEAD(rpc_list);
1661         int rc;
1662         ENTRY;
1663
1664         /* Local lock trees don't get reprocessed. */
1665         if (ns_is_client(ldlm_res_to_ns(res))) {
1666                 EXIT;
1667                 return;
1668         }
1669
1670  restart:
1671         lock_res(res);
1672         rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1673         if (rc == LDLM_ITER_CONTINUE)
1674                 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1675         unlock_res(res);
1676
1677         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
1678                                LDLM_WORK_CP_AST);
1679         if (rc == -ERESTART) {
1680                 LASSERT(cfs_list_empty(&rpc_list));
1681                 goto restart;
1682         }
1683         EXIT;
1684 }
1685
1686 void ldlm_cancel_callback(struct ldlm_lock *lock)
1687 {
1688         check_res_locked(lock->l_resource);
1689         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1690                 lock->l_flags |= LDLM_FL_CANCEL;
1691                 if (lock->l_blocking_ast) {
1692                         // l_check_no_ns_lock(ns);
1693                         unlock_res_and_lock(lock);
1694                         lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1695                                              LDLM_CB_CANCELING);
1696                         lock_res_and_lock(lock);
1697                 } else {
1698                         LDLM_DEBUG(lock, "no blocking ast");
1699                 }
1700         }
1701         lock->l_flags |= LDLM_FL_BL_DONE;
1702 }
1703
1704 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1705 {
1706         if (req->l_resource->lr_type != LDLM_PLAIN &&
1707             req->l_resource->lr_type != LDLM_IBITS)
1708                 return;
1709
1710         cfs_list_del_init(&req->l_sl_policy);
1711         cfs_list_del_init(&req->l_sl_mode);
1712 }
1713
1714 void ldlm_lock_cancel(struct ldlm_lock *lock)
1715 {
1716         struct ldlm_resource *res;
1717         struct ldlm_namespace *ns;
1718         ENTRY;
1719
1720         lock_res_and_lock(lock);
1721
1722         res = lock->l_resource;
1723         ns  = ldlm_res_to_ns(res);
1724
1725         /* Please do not, no matter how tempting, remove this LBUG without
1726          * talking to me first. -phik */
1727         if (lock->l_readers || lock->l_writers) {
1728                 LDLM_ERROR(lock, "lock still has references");
1729                 LBUG();
1730         }
1731
1732         ldlm_del_waiting_lock(lock);
1733
1734         /* Releases cancel callback. */
1735         ldlm_cancel_callback(lock);
1736
1737         /* Yes, second time, just in case it was added again while we were
1738            running with no res lock in ldlm_cancel_callback */
1739         ldlm_del_waiting_lock(lock);
1740         ldlm_resource_unlink_lock(lock);
1741         ldlm_lock_destroy_nolock(lock);
1742
1743         if (lock->l_granted_mode == lock->l_req_mode)
1744                 ldlm_pool_del(&ns->ns_pool, lock);
1745
1746         /* Make sure we will not be called again for same lock what is possible
1747          * if not to zero out lock->l_granted_mode */
1748         lock->l_granted_mode = LCK_MINMODE;
1749         unlock_res_and_lock(lock);
1750
1751         EXIT;
1752 }
1753
1754 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1755 {
1756         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1757         int rc = -EINVAL;
1758         ENTRY;
1759
1760         if (lock) {
1761                 if (lock->l_ast_data == NULL)
1762                         lock->l_ast_data = data;
1763                 if (lock->l_ast_data == data)
1764                         rc = 0;
1765                 LDLM_LOCK_PUT(lock);
1766         }
1767         RETURN(rc);
1768 }
1769 EXPORT_SYMBOL(ldlm_lock_set_data);
1770
1771 int ldlm_cancel_locks_for_export_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1772                                     cfs_hlist_node_t *hnode, void *data)
1773
1774 {
1775         struct obd_export    *exp  = data;
1776         struct ldlm_lock     *lock = cfs_hash_object(hs, hnode);
1777         struct ldlm_resource *res;
1778
1779         res = ldlm_resource_getref(lock->l_resource);
1780         LDLM_LOCK_GET(lock);
1781
1782         LDLM_DEBUG(lock, "export %p", exp);
1783         ldlm_res_lvbo_update(res, NULL, 1);
1784         ldlm_lock_cancel(lock);
1785         ldlm_reprocess_all(res);
1786         ldlm_resource_putref(res);
1787         LDLM_LOCK_RELEASE(lock);
1788         return 0;
1789 }
1790
1791 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1792 {
1793         cfs_hash_for_each_empty(exp->exp_lock_hash,
1794                                 ldlm_cancel_locks_for_export_cb, exp);
1795 }
1796
1797 /**
1798  * Downgrade an exclusive lock.
1799  *
1800  * A fast variant of ldlm_lock_convert for convertion of exclusive
1801  * locks. The convertion is always successful.
1802  *
1803  * \param lock A lock to convert
1804  * \param new_mode new lock mode
1805  */
1806 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1807 {
1808         ENTRY;
1809
1810         LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1811         LASSERT(new_mode == LCK_COS);
1812
1813         lock_res_and_lock(lock);
1814         ldlm_resource_unlink_lock(lock);
1815         /*
1816          * Remove the lock from pool as it will be added again in
1817          * ldlm_grant_lock() called below.
1818          */
1819         ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
1820
1821         lock->l_req_mode = new_mode;
1822         ldlm_grant_lock(lock, NULL);
1823         unlock_res_and_lock(lock);
1824         ldlm_reprocess_all(lock->l_resource);
1825
1826         EXIT;
1827 }
1828
1829 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1830                                         __u32 *flags)
1831 {
1832         CFS_LIST_HEAD(rpc_list);
1833         struct ldlm_resource *res;
1834         struct ldlm_namespace *ns;
1835         int granted = 0;
1836         int old_mode, rc;
1837         struct sl_insert_point prev;
1838         ldlm_error_t err;
1839         struct ldlm_interval *node;
1840         ENTRY;
1841
1842         if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1843                 *flags |= LDLM_FL_BLOCK_GRANTED;
1844                 RETURN(lock->l_resource);
1845         }
1846
1847         /* I can't check the type of lock here because the bitlock of lock
1848          * is not held here, so do the allocation blindly. -jay */
1849         OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1850         if (node == NULL)  /* Actually, this causes EDEADLOCK to be returned */
1851                 RETURN(NULL);
1852
1853         LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1854                  "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1855
1856         lock_res_and_lock(lock);
1857
1858         res = lock->l_resource;
1859         ns  = ldlm_res_to_ns(res);
1860
1861         old_mode = lock->l_req_mode;
1862         lock->l_req_mode = new_mode;
1863         if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1864                 /* remember the lock position where the lock might be
1865                  * added back to the granted list later and also
1866                  * remember the join mode for skiplist fixing. */
1867                 prev.res_link = lock->l_res_link.prev;
1868                 prev.mode_link = lock->l_sl_mode.prev;
1869                 prev.policy_link = lock->l_sl_policy.prev;
1870                 ldlm_resource_unlink_lock(lock);
1871         } else {
1872                 ldlm_resource_unlink_lock(lock);
1873                 if (res->lr_type == LDLM_EXTENT) {
1874                         /* FIXME: ugly code, I have to attach the lock to a
1875                          * interval node again since perhaps it will be granted
1876                          * soon */
1877                         CFS_INIT_LIST_HEAD(&node->li_group);
1878                         ldlm_interval_attach(node, lock);
1879                         node = NULL;
1880                 }
1881         }
1882
1883         /*
1884          * Remove old lock from the pool before adding the lock with new
1885          * mode below in ->policy()
1886          */
1887         ldlm_pool_del(&ns->ns_pool, lock);
1888
1889         /* If this is a local resource, put it on the appropriate list. */
1890         if (ns_is_client(ldlm_res_to_ns(res))) {
1891                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1892                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1893                 } else {
1894                         /* This should never happen, because of the way the
1895                          * server handles conversions. */
1896                         LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1897                                    *flags);
1898                         LBUG();
1899
1900                         ldlm_grant_lock(lock, &rpc_list);
1901                         granted = 1;
1902                         /* FIXME: completion handling not with lr_lock held ! */
1903                         if (lock->l_completion_ast)
1904                                 lock->l_completion_ast(lock, 0, NULL);
1905                 }
1906         } else {
1907                 int pflags = 0;
1908                 ldlm_processing_policy policy;
1909                 policy = ldlm_processing_policy_table[res->lr_type];
1910                 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1911                 if (rc == LDLM_ITER_STOP) {
1912                         lock->l_req_mode = old_mode;
1913                         if (res->lr_type == LDLM_EXTENT)
1914                                 ldlm_extent_add_lock(res, lock);
1915                         else
1916                                 ldlm_granted_list_add_lock(lock, &prev);
1917
1918                         res = NULL;
1919                 } else {
1920                         *flags |= LDLM_FL_BLOCK_GRANTED;
1921                         granted = 1;
1922                 }
1923         }
1924         unlock_res_and_lock(lock);
1925
1926         if (granted)
1927                 ldlm_run_ast_work(ns, &rpc_list, LDLM_WORK_CP_AST);
1928         if (node)
1929                 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1930         RETURN(res);
1931 }
1932
1933 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1934 {
1935         struct obd_device *obd = NULL;
1936
1937         if (!((libcfs_debug | D_ERROR) & level))
1938                 return;
1939
1940         if (!lock) {
1941                 CDEBUG(level, "  NULL LDLM lock\n");
1942                 return;
1943         }
1944
1945         CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1946                lock, lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1947                pos, lock->l_pid);
1948         if (lock->l_conn_export != NULL)
1949                 obd = lock->l_conn_export->exp_obd;
1950         if (lock->l_export && lock->l_export->exp_connection) {
1951                 CDEBUG(level, "  Node: NID %s (rhandle: "LPX64")\n",
1952                      libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1953                      lock->l_remote_handle.cookie);
1954         } else if (obd == NULL) {
1955                 CDEBUG(level, "  Node: local\n");
1956         } else {
1957                 struct obd_import *imp = obd->u.cli.cl_import;
1958                 CDEBUG(level, "  Node: NID %s (rhandle: "LPX64")\n",
1959                        libcfs_nid2str(imp->imp_connection->c_peer.nid),
1960                        lock->l_remote_handle.cookie);
1961         }
1962         CDEBUG(level, "  Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1963                   lock->l_resource,
1964                   lock->l_resource->lr_name.name[0],
1965                   lock->l_resource->lr_name.name[1],
1966                   lock->l_resource->lr_name.name[2]);
1967         CDEBUG(level, "  Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1968                "write: %d flags: "LPX64"\n", ldlm_lockname[lock->l_req_mode],
1969                ldlm_lockname[lock->l_granted_mode],
1970                cfs_atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1971                lock->l_flags);
1972         if (lock->l_resource->lr_type == LDLM_EXTENT)
1973                 CDEBUG(level, "  Extent: "LPU64" -> "LPU64
1974                        " (req "LPU64"-"LPU64")\n",
1975                        lock->l_policy_data.l_extent.start,
1976                        lock->l_policy_data.l_extent.end,
1977                        lock->l_req_extent.start, lock->l_req_extent.end);
1978         else if (lock->l_resource->lr_type == LDLM_FLOCK)
1979                 CDEBUG(level, "  Pid: %d Extent: "LPU64" -> "LPU64"\n",
1980                        lock->l_policy_data.l_flock.pid,
1981                        lock->l_policy_data.l_flock.start,
1982                        lock->l_policy_data.l_flock.end);
1983        else if (lock->l_resource->lr_type == LDLM_IBITS)
1984                 CDEBUG(level, "  Bits: "LPX64"\n",
1985                        lock->l_policy_data.l_inodebits.bits);
1986 }
1987
1988 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1989 {
1990         struct ldlm_lock *lock;
1991
1992         if (!((libcfs_debug | D_ERROR) & level))
1993                 return;
1994
1995         lock = ldlm_handle2lock(lockh);
1996         if (lock == NULL)
1997                 return;
1998
1999         ldlm_lock_dump(D_OTHER, lock, 0);
2000
2001         LDLM_LOCK_PUT(lock);
2002 }
2003
2004 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
2005                       struct libcfs_debug_msg_data *data, const char *fmt,
2006                       ...)
2007 {
2008         va_list args;
2009         cfs_debug_limit_state_t *cdls = data->msg_cdls;
2010
2011         va_start(args, fmt);
2012
2013         if (lock->l_resource == NULL) {
2014                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2015                                    data->msg_fn, data->msg_line, fmt, args,
2016                        " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2017                        "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" remote: "
2018                        LPX64" expref: %d pid: %u timeout: %lu\n", lock,
2019                        lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2020                        lock->l_readers, lock->l_writers,
2021                        ldlm_lockname[lock->l_granted_mode],
2022                        ldlm_lockname[lock->l_req_mode],
2023                        lock->l_flags, lock->l_remote_handle.cookie,
2024                        lock->l_export ?
2025                        cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2026                        lock->l_pid, lock->l_callback_timeout);
2027                 va_end(args);
2028                 return;
2029         }
2030
2031         switch (lock->l_resource->lr_type) {
2032         case LDLM_EXTENT:
2033                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2034                                    data->msg_fn, data->msg_line, fmt, args,
2035                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2036                        "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
2037                        "] (req "LPU64"->"LPU64") flags: "LPX64" remote: "LPX64
2038                        " expref: %d pid: %u timeout %lu\n",
2039                        ldlm_lock_to_ns_name(lock), lock,
2040                        lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2041                        lock->l_readers, lock->l_writers,
2042                        ldlm_lockname[lock->l_granted_mode],
2043                        ldlm_lockname[lock->l_req_mode],
2044                        lock->l_resource->lr_name.name[0],
2045                        lock->l_resource->lr_name.name[1],
2046                        cfs_atomic_read(&lock->l_resource->lr_refcount),
2047                        ldlm_typename[lock->l_resource->lr_type],
2048                        lock->l_policy_data.l_extent.start,
2049                        lock->l_policy_data.l_extent.end,
2050                        lock->l_req_extent.start, lock->l_req_extent.end,
2051                        lock->l_flags, lock->l_remote_handle.cookie,
2052                        lock->l_export ?
2053                        cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2054                        lock->l_pid, lock->l_callback_timeout);
2055                 break;
2056
2057         case LDLM_FLOCK:
2058                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2059                                    data->msg_fn, data->msg_line, fmt, args,
2060                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2061                        "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
2062                        "["LPU64"->"LPU64"] flags: "LPX64" remote: "LPX64
2063                        " expref: %d pid: %u timeout: %lu\n",
2064                        ldlm_lock_to_ns_name(lock), lock,
2065                        lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2066                        lock->l_readers, lock->l_writers,
2067                        ldlm_lockname[lock->l_granted_mode],
2068                        ldlm_lockname[lock->l_req_mode],
2069                        lock->l_resource->lr_name.name[0],
2070                        lock->l_resource->lr_name.name[1],
2071                        cfs_atomic_read(&lock->l_resource->lr_refcount),
2072                        ldlm_typename[lock->l_resource->lr_type],
2073                        lock->l_policy_data.l_flock.pid,
2074                        lock->l_policy_data.l_flock.start,
2075                        lock->l_policy_data.l_flock.end,
2076                        lock->l_flags, lock->l_remote_handle.cookie,
2077                        lock->l_export ?
2078                        cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2079                        lock->l_pid, lock->l_callback_timeout);
2080                 break;
2081
2082         case LDLM_IBITS:
2083                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2084                                    data->msg_fn, data->msg_line, fmt, args,
2085                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2086                        "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
2087                        "flags: "LPX64" remote: "LPX64" expref: %d "
2088                        "pid: %u timeout: %lu\n",
2089                        ldlm_lock_to_ns_name(lock),
2090                        lock, lock->l_handle.h_cookie,
2091                        cfs_atomic_read (&lock->l_refc),
2092                        lock->l_readers, lock->l_writers,
2093                        ldlm_lockname[lock->l_granted_mode],
2094                        ldlm_lockname[lock->l_req_mode],
2095                        lock->l_resource->lr_name.name[0],
2096                        lock->l_resource->lr_name.name[1],
2097                        lock->l_policy_data.l_inodebits.bits,
2098                        cfs_atomic_read(&lock->l_resource->lr_refcount),
2099                        ldlm_typename[lock->l_resource->lr_type],
2100                        lock->l_flags, lock->l_remote_handle.cookie,
2101                        lock->l_export ?
2102                        cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2103                        lock->l_pid, lock->l_callback_timeout);
2104                 break;
2105
2106         default:
2107                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2108                                    data->msg_fn, data->msg_line, fmt, args,
2109                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2110                        "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" "
2111                        "remote: "LPX64" expref: %d pid: %u timeout %lu\n",
2112                        ldlm_lock_to_ns_name(lock),
2113                        lock, lock->l_handle.h_cookie,
2114                        cfs_atomic_read (&lock->l_refc),
2115                        lock->l_readers, lock->l_writers,
2116                        ldlm_lockname[lock->l_granted_mode],
2117                        ldlm_lockname[lock->l_req_mode],
2118                        lock->l_resource->lr_name.name[0],
2119                        lock->l_resource->lr_name.name[1],
2120                        cfs_atomic_read(&lock->l_resource->lr_refcount),
2121                        ldlm_typename[lock->l_resource->lr_type],
2122                        lock->l_flags, lock->l_remote_handle.cookie,
2123                        lock->l_export ?
2124                        cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2125                        lock->l_pid, lock->l_callback_timeout);
2126                 break;
2127         }
2128         va_end(args);
2129 }
2130 EXPORT_SYMBOL(_ldlm_lock_debug);