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