lustre/ldlm/ldlm_lock.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  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_wire18_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_wire18_to_local,
  78         [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_wire_to_local,
  79 };
  80
  81 static ldlm_policy_wire_to_local_t ldlm_policy_wire21_to_local[] = {
  82         [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_wire_to_local,
  83         [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_wire_to_local,
  84         [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_wire21_to_local,
  85         [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_wire_to_local,
  86 };
  87
  88 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
  89         [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_local_to_wire,
  90         [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_local_to_wire,
  91         [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_local_to_wire,
  92         [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_local_to_wire,
  93 };
  94
  95 /**
  96  * Converts lock policy from local format to on the wire lock_desc format
  97  */
  98 void ldlm_convert_policy_to_wire(ldlm_type_t type,
  99                                  const ldlm_policy_data_t *lpolicy,
 100                                  ldlm_wire_policy_data_t *wpolicy)
 101 {
 102         ldlm_policy_local_to_wire_t convert;
 103
 104         convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
 105
 106         convert(lpolicy, wpolicy);
 107 }
 108
 109 /**
 110  * Converts lock policy from on the wire lock_desc format to local format
 111  */
 112 void ldlm_convert_policy_to_local(struct obd_export *exp, ldlm_type_t type,
 113                                   const ldlm_wire_policy_data_t *wpolicy,
 114                                   ldlm_policy_data_t *lpolicy)
 115 {
 116         ldlm_policy_wire_to_local_t convert;
 117         int new_client;
 118
 119         /** some badnes for 2.0.0 clients, but 2.0.0 isn't supported */
 120         new_client = (exp->exp_connect_flags & OBD_CONNECT_FULL20) != 0;
 121         if (new_client)
 122                convert = ldlm_policy_wire21_to_local[type - LDLM_MIN_TYPE];
 123         else
 124                convert = ldlm_policy_wire18_to_local[type - LDLM_MIN_TYPE];
 125
 126         convert(wpolicy, lpolicy);
 127 }
 128
 129 char *ldlm_it2str(int it)
 130 {
 131         switch (it) {
 132         case IT_OPEN:
 133                 return "open";
 134         case IT_CREAT:
 135                 return "creat";
 136         case (IT_OPEN | IT_CREAT):
 137                 return "open|creat";
 138         case IT_READDIR:
 139                 return "readdir";
 140         case IT_GETATTR:
 141                 return "getattr";
 142         case IT_LOOKUP:
 143                 return "lookup";
 144         case IT_UNLINK:
 145                 return "unlink";
 146         case IT_GETXATTR:
 147                 return "getxattr";
 148         case IT_LAYOUT:
 149                 return "layout";
 150         default:
 151                 CERROR("Unknown intent %d\n", it);
 152                 return "UNKNOWN";
 153         }
 154 }
 155
 156 extern cfs_mem_cache_t *ldlm_lock_slab;
 157
 158 #ifdef HAVE_SERVER_SUPPORT
 159 static ldlm_processing_policy ldlm_processing_policy_table[] = {
 160         [LDLM_PLAIN] ldlm_process_plain_lock,
 161         [LDLM_EXTENT] ldlm_process_extent_lock,
 162 # ifdef __KERNEL__
 163         [LDLM_FLOCK] ldlm_process_flock_lock,
 164 # endif
 165         [LDLM_IBITS] ldlm_process_inodebits_lock,
 166 };
 167
 168 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
 169 {
 170         return ldlm_processing_policy_table[res->lr_type];
 171 }
 172 #endif /* HAVE_SERVER_SUPPORT */
 173
 174 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
 175 {
 176         ns->ns_policy = arg;
 177 }
 178
 179 /*
 180  * REFCOUNTED LOCK OBJECTS
 181  */
 182
 183
 184 /*
 185  * Lock refcounts, during creation:
 186  *   - one special one for allocation, dec'd only once in destroy
 187  *   - one for being a lock that's in-use
 188  *   - one for the addref associated with a new lock
 189  */
 190 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 191 {
 192         cfs_atomic_inc(&lock->l_refc);
 193         return lock;
 194 }
 195
 196 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
 197 {
 198         LASSERT(size == sizeof(*lock));
 199         OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
 200 }
 201
 202 void ldlm_lock_put(struct ldlm_lock *lock)
 203 {
 204         ENTRY;
 205
 206         LASSERT(lock->l_resource != LP_POISON);
 207         LASSERT(cfs_atomic_read(&lock->l_refc) > 0);
 208         if (cfs_atomic_dec_and_test(&lock->l_refc)) {
 209                 struct ldlm_resource *res;
 210
 211                 LDLM_DEBUG(lock,
 212                            "final lock_put on destroyed lock, freeing it.");
 213
 214                 res = lock->l_resource;
 215                 LASSERT(lock->l_destroyed);
 216                 LASSERT(cfs_list_empty(&lock->l_res_link));
 217                 LASSERT(cfs_list_empty(&lock->l_pending_chain));
 218
 219                 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
 220                                      LDLM_NSS_LOCKS);
 221                 lu_ref_del(&res->lr_reference, "lock", lock);
 222                 ldlm_resource_putref(res);
 223                 lock->l_resource = NULL;
 224                 if (lock->l_export) {
 225                         class_export_lock_put(lock->l_export, lock);
 226                         lock->l_export = NULL;
 227                 }
 228
 229                 if (lock->l_lvb_data != NULL)
 230                         OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
 231
 232                 ldlm_interval_free(ldlm_interval_detach(lock));
 233                 lu_ref_fini(&lock->l_reference);
 234                 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
 235                                 ldlm_lock_free);
 236         }
 237
 238         EXIT;
 239 }
 240
 241 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
 242 {
 243         int rc = 0;
 244         if (!cfs_list_empty(&lock->l_lru)) {
 245                 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
 246
 247                 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
 248                 cfs_list_del_init(&lock->l_lru);
 249                 if (lock->l_flags & LDLM_FL_SKIPPED)
 250                         lock->l_flags &= ~LDLM_FL_SKIPPED;
 251                 LASSERT(ns->ns_nr_unused > 0);
 252                 ns->ns_nr_unused--;
 253                 rc = 1;
 254         }
 255         return rc;
 256 }
 257
 258 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
 259 {
 260         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
 261         int rc;
 262
 263         ENTRY;
 264         if (lock->l_ns_srv) {
 265                 LASSERT(cfs_list_empty(&lock->l_lru));
 266                 RETURN(0);
 267         }
 268
 269         cfs_spin_lock(&ns->ns_lock);
 270         rc = ldlm_lock_remove_from_lru_nolock(lock);
 271         cfs_spin_unlock(&ns->ns_lock);
 272         EXIT;
 273         return rc;
 274 }
 275
 276 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
 277 {
 278         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
 279
 280         lock->l_last_used = cfs_time_current();
 281         LASSERT(cfs_list_empty(&lock->l_lru));
 282         LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
 283         cfs_list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 284         LASSERT(ns->ns_nr_unused >= 0);
 285         ns->ns_nr_unused++;
 286 }
 287
 288 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
 289 {
 290         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
 291
 292         ENTRY;
 293         cfs_spin_lock(&ns->ns_lock);
 294         ldlm_lock_add_to_lru_nolock(lock);
 295         cfs_spin_unlock(&ns->ns_lock);
 296         EXIT;
 297 }
 298
 299 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
 300 {
 301         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
 302
 303         ENTRY;
 304         if (lock->l_ns_srv) {
 305                 LASSERT(cfs_list_empty(&lock->l_lru));
 306                 EXIT;
 307                 return;
 308         }
 309
 310         cfs_spin_lock(&ns->ns_lock);
 311         if (!cfs_list_empty(&lock->l_lru)) {
 312                 ldlm_lock_remove_from_lru_nolock(lock);
 313                 ldlm_lock_add_to_lru_nolock(lock);
 314         }
 315         cfs_spin_unlock(&ns->ns_lock);
 316         EXIT;
 317 }
 318
 319 /* This used to have a 'strict' flag, which recovery would use to mark an
 320  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
 321  * shall explain why it's gone: with the new hash table scheme, once you call
 322  * ldlm_lock_destroy, you can never drop your final references on this lock.
 323  * Because it's not in the hash table anymore.  -phil */
 324 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
 325 {
 326         ENTRY;
 327
 328         if (lock->l_readers || lock->l_writers) {
 329                 LDLM_ERROR(lock, "lock still has references");
 330                 ldlm_lock_dump(D_ERROR, lock, 0);
 331                 LBUG();
 332         }
 333
 334         if (!cfs_list_empty(&lock->l_res_link)) {
 335                 LDLM_ERROR(lock, "lock still on resource");
 336                 ldlm_lock_dump(D_ERROR, lock, 0);
 337                 LBUG();
 338         }
 339
 340         if (lock->l_destroyed) {
 341                 LASSERT(cfs_list_empty(&lock->l_lru));
 342                 EXIT;
 343                 return 0;
 344         }
 345         lock->l_destroyed = 1;
 346
 347         if (lock->l_export && lock->l_export->exp_lock_hash &&
 348             !cfs_hlist_unhashed(&lock->l_exp_hash))
 349                 cfs_hash_del(lock->l_export->exp_lock_hash,
 350                              &lock->l_remote_handle, &lock->l_exp_hash);
 351
 352         ldlm_lock_remove_from_lru(lock);
 353         class_handle_unhash(&lock->l_handle);
 354
 355 #if 0
 356         /* Wake anyone waiting for this lock */
 357         /* FIXME: I should probably add yet another flag, instead of using
 358          * l_export to only call this on clients */
 359         if (lock->l_export)
 360                 class_export_put(lock->l_export);
 361         lock->l_export = NULL;
 362         if (lock->l_export && lock->l_completion_ast)
 363                 lock->l_completion_ast(lock, 0);
 364 #endif
 365         EXIT;
 366         return 1;
 367 }
 368
 369 void ldlm_lock_destroy(struct ldlm_lock *lock)
 370 {
 371         int first;
 372         ENTRY;
 373         lock_res_and_lock(lock);
 374         first = ldlm_lock_destroy_internal(lock);
 375         unlock_res_and_lock(lock);
 376
 377         /* drop reference from hashtable only for first destroy */
 378         if (first) {
 379                 lu_ref_del(&lock->l_reference, "hash", lock);
 380                 LDLM_LOCK_RELEASE(lock);
 381         }
 382         EXIT;
 383 }
 384
 385 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
 386 {
 387         int first;
 388         ENTRY;
 389         first = ldlm_lock_destroy_internal(lock);
 390         /* drop reference from hashtable only for first destroy */
 391         if (first) {
 392                 lu_ref_del(&lock->l_reference, "hash", lock);
 393                 LDLM_LOCK_RELEASE(lock);
 394         }
 395         EXIT;
 396 }
 397
 398 /* this is called by portals_handle2object with the handle lock taken */
 399 static void lock_handle_addref(void *lock)
 400 {
 401         LDLM_LOCK_GET((struct ldlm_lock *)lock);
 402 }
 403
 404 /*
 405  * usage: pass in a resource on which you have done ldlm_resource_get
 406  *        new lock will take over the refcount.
 407  * returns: lock with refcount 2 - one for current caller and one for remote
 408  */
 409 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
 410 {
 411         struct ldlm_lock *lock;
 412         ENTRY;
 413
 414         if (resource == NULL)
 415                 LBUG();
 416
 417         OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, CFS_ALLOC_IO);
 418         if (lock == NULL)
 419                 RETURN(NULL);
 420
 421         cfs_spin_lock_init(&lock->l_lock);
 422         lock->l_resource = resource;
 423         lu_ref_add(&resource->lr_reference, "lock", lock);
 424
 425         cfs_atomic_set(&lock->l_refc, 2);
 426         CFS_INIT_LIST_HEAD(&lock->l_res_link);
 427         CFS_INIT_LIST_HEAD(&lock->l_lru);
 428         CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
 429         CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
 430         CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
 431         CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
 432         cfs_waitq_init(&lock->l_waitq);
 433         lock->l_blocking_lock = NULL;
 434         CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
 435         CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
 436         CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
 437
 438         lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
 439                              LDLM_NSS_LOCKS);
 440         CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
 441         class_handle_hash(&lock->l_handle, lock_handle_addref);
 442
 443         lu_ref_init(&lock->l_reference);
 444         lu_ref_add(&lock->l_reference, "hash", lock);
 445         lock->l_callback_timeout = 0;
 446
 447 #if LUSTRE_TRACKS_LOCK_EXP_REFS
 448         CFS_INIT_LIST_HEAD(&lock->l_exp_refs_link);
 449         lock->l_exp_refs_nr = 0;
 450         lock->l_exp_refs_target = NULL;
 451 #endif
 452         CFS_INIT_LIST_HEAD(&lock->l_exp_list);
 453
 454         RETURN(lock);
 455 }
 456
 457 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
 458                               const struct ldlm_res_id *new_resid)
 459 {
 460         struct ldlm_resource *oldres = lock->l_resource;
 461         struct ldlm_resource *newres;
 462         int type;
 463         ENTRY;
 464
 465         LASSERT(ns_is_client(ns));
 466
 467         lock_res_and_lock(lock);
 468         if (memcmp(new_resid, &lock->l_resource->lr_name,
 469                    sizeof(lock->l_resource->lr_name)) == 0) {
 470                 /* Nothing to do */
 471                 unlock_res_and_lock(lock);
 472                 RETURN(0);
 473         }
 474
 475         LASSERT(new_resid->name[0] != 0);
 476
 477         /* This function assumes that the lock isn't on any lists */
 478         LASSERT(cfs_list_empty(&lock->l_res_link));
 479
 480         type = oldres->lr_type;
 481         unlock_res_and_lock(lock);
 482
 483         newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
 484         if (newres == NULL)
 485                 RETURN(-ENOMEM);
 486
 487         lu_ref_add(&newres->lr_reference, "lock", lock);
 488         /*
 489          * To flip the lock from the old to the new resource, lock, oldres and
 490          * newres have to be locked. Resource spin-locks are nested within
 491          * lock->l_lock, and are taken in the memory address order to avoid
 492          * dead-locks.
 493          */
 494         cfs_spin_lock(&lock->l_lock);
 495         oldres = lock->l_resource;
 496         if (oldres < newres) {
 497                 lock_res(oldres);
 498                 lock_res_nested(newres, LRT_NEW);
 499         } else {
 500                 lock_res(newres);
 501                 lock_res_nested(oldres, LRT_NEW);
 502         }
 503         LASSERT(memcmp(new_resid, &oldres->lr_name,
 504                        sizeof oldres->lr_name) != 0);
 505         lock->l_resource = newres;
 506         unlock_res(oldres);
 507         unlock_res_and_lock(lock);
 508
 509         /* ...and the flowers are still standing! */
 510         lu_ref_del(&oldres->lr_reference, "lock", lock);
 511         ldlm_resource_putref(oldres);
 512
 513         RETURN(0);
 514 }
 515
 516 /*
 517  *  HANDLES
 518  */
 519
 520 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
 521 {
 522         lockh->cookie = lock->l_handle.h_cookie;
 523 }
 524
 525 /* if flags: atomically get the lock and set the flags.
 526  *           Return NULL if flag already set
 527  */
 528
 529 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
 530                                      int flags)
 531 {
 532         struct ldlm_lock *lock;
 533         ENTRY;
 534
 535         LASSERT(handle);
 536
 537         lock = class_handle2object(handle->cookie);
 538         if (lock == NULL)
 539                 RETURN(NULL);
 540
 541         /* It's unlikely but possible that someone marked the lock as
 542          * destroyed after we did handle2object on it */
 543         if (flags == 0 && !lock->l_destroyed) {
 544                 lu_ref_add(&lock->l_reference, "handle", cfs_current());
 545                 RETURN(lock);
 546         }
 547
 548         lock_res_and_lock(lock);
 549
 550         LASSERT(lock->l_resource != NULL);
 551
 552         lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
 553         if (unlikely(lock->l_destroyed)) {
 554                 unlock_res_and_lock(lock);
 555                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
 556                 LDLM_LOCK_PUT(lock);
 557                 RETURN(NULL);
 558         }
 559
 560         if (flags && (lock->l_flags & flags)) {
 561                 unlock_res_and_lock(lock);
 562                 LDLM_LOCK_PUT(lock);
 563                 RETURN(NULL);
 564         }
 565
 566         if (flags)
 567                 lock->l_flags |= flags;
 568
 569         unlock_res_and_lock(lock);
 570         RETURN(lock);
 571 }
 572
 573 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 574 {
 575         struct obd_export *exp = lock->l_export?:lock->l_conn_export;
 576         /* INODEBITS_INTEROP: If the other side does not support
 577          * inodebits, reply with a plain lock descriptor.
 578          */
 579         if ((lock->l_resource->lr_type == LDLM_IBITS) &&
 580             (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
 581                 /* Make sure all the right bits are set in this lock we
 582                    are going to pass to client */
 583                 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
 584                          (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
 585                           MDS_INODELOCK_LAYOUT),
 586                          "Inappropriate inode lock bits during "
 587                          "conversion " LPU64 "\n",
 588                          lock->l_policy_data.l_inodebits.bits);
 589
 590                 ldlm_res2desc(lock->l_resource, &desc->l_resource);
 591                 desc->l_resource.lr_type = LDLM_PLAIN;
 592
 593                 /* Convert "new" lock mode to something old client can
 594                    understand */
 595                 if ((lock->l_req_mode == LCK_CR) ||
 596                     (lock->l_req_mode == LCK_CW))
 597                         desc->l_req_mode = LCK_PR;
 598                 else
 599                         desc->l_req_mode = lock->l_req_mode;
 600                 if ((lock->l_granted_mode == LCK_CR) ||
 601                     (lock->l_granted_mode == LCK_CW)) {
 602                         desc->l_granted_mode = LCK_PR;
 603                 } else {
 604                         /* We never grant PW/EX locks to clients */
 605                         LASSERT((lock->l_granted_mode != LCK_PW) &&
 606                                 (lock->l_granted_mode != LCK_EX));
 607                         desc->l_granted_mode = lock->l_granted_mode;
 608                 }
 609
 610                 /* We do not copy policy here, because there is no
 611                    policy for plain locks */
 612         } else {
 613                 ldlm_res2desc(lock->l_resource, &desc->l_resource);
 614                 desc->l_req_mode = lock->l_req_mode;
 615                 desc->l_granted_mode = lock->l_granted_mode;
 616                 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
 617                                             &lock->l_policy_data,
 618                                             &desc->l_policy_data);
 619         }
 620 }
 621
 622 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
 623                            cfs_list_t *work_list)
 624 {
 625         if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
 626                 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
 627                 lock->l_flags |= LDLM_FL_AST_SENT;
 628                 /* If the enqueuing client said so, tell the AST recipient to
 629                  * discard dirty data, rather than writing back. */
 630                 if (new->l_flags & LDLM_AST_DISCARD_DATA)
 631                         lock->l_flags |= LDLM_FL_DISCARD_DATA;
 632                 LASSERT(cfs_list_empty(&lock->l_bl_ast));
 633                 cfs_list_add(&lock->l_bl_ast, work_list);
 634                 LDLM_LOCK_GET(lock);
 635                 LASSERT(lock->l_blocking_lock == NULL);
 636                 lock->l_blocking_lock = LDLM_LOCK_GET(new);
 637         }
 638 }
 639
 640 void ldlm_add_cp_work_item(struct ldlm_lock *lock, cfs_list_t *work_list)
 641 {
 642         if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
 643                 lock->l_flags |= LDLM_FL_CP_REQD;
 644                 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
 645                 LASSERT(cfs_list_empty(&lock->l_cp_ast));
 646                 cfs_list_add(&lock->l_cp_ast, work_list);
 647                 LDLM_LOCK_GET(lock);
 648         }
 649 }
 650
 651 /* must be called with lr_lock held */
 652 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
 653                             cfs_list_t *work_list)
 654 {
 655         ENTRY;
 656         check_res_locked(lock->l_resource);
 657         if (new)
 658                 ldlm_add_bl_work_item(lock, new, work_list);
 659         else
 660                 ldlm_add_cp_work_item(lock, work_list);
 661         EXIT;
 662 }
 663
 664 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 665 {
 666         struct ldlm_lock *lock;
 667
 668         lock = ldlm_handle2lock(lockh);
 669         LASSERT(lock != NULL);
 670         ldlm_lock_addref_internal(lock, mode);
 671         LDLM_LOCK_PUT(lock);
 672 }
 673
 674 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
 675 {
 676         ldlm_lock_remove_from_lru(lock);
 677         if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
 678                 lock->l_readers++;
 679                 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
 680         }
 681         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
 682                 lock->l_writers++;
 683                 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
 684         }
 685         LDLM_LOCK_GET(lock);
 686         lu_ref_add_atomic(&lock->l_reference, "user", lock);
 687         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 688 }
 689
 690 /**
 691  * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
 692  * or destroyed.
 693  *
 694  * \retval 0 success, lock was addref-ed
 695  *
 696  * \retval -EAGAIN lock is being canceled.
 697  */
 698 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
 699 {
 700         struct ldlm_lock *lock;
 701         int               result;
 702
 703         result = -EAGAIN;
 704         lock = ldlm_handle2lock(lockh);
 705         if (lock != NULL) {
 706                 lock_res_and_lock(lock);
 707                 if (lock->l_readers != 0 || lock->l_writers != 0 ||
 708                     !(lock->l_flags & LDLM_FL_CBPENDING)) {
 709                         ldlm_lock_addref_internal_nolock(lock, mode);
 710                         result = 0;
 711                 }
 712                 unlock_res_and_lock(lock);
 713                 LDLM_LOCK_PUT(lock);
 714         }
 715         return result;
 716 }
 717
 718 /* only called for local locks */
 719 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 720 {
 721         lock_res_and_lock(lock);
 722         ldlm_lock_addref_internal_nolock(lock, mode);
 723         unlock_res_and_lock(lock);
 724 }
 725
 726 /* only called in ldlm_flock_destroy and for local locks.
 727  *  * for LDLM_FLOCK type locks, l_blocking_ast is null, and
 728  *   * ldlm_lock_remove_from_lru() does nothing, it is safe
 729  *    * for ldlm_flock_destroy usage by dropping some code */
 730 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
 731 {
 732         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 733         if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
 734                 LASSERT(lock->l_readers > 0);
 735                 lu_ref_del(&lock->l_reference, "reader", lock);
 736                 lock->l_readers--;
 737         }
 738         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
 739                 LASSERT(lock->l_writers > 0);
 740                 lu_ref_del(&lock->l_reference, "writer", lock);
 741                 lock->l_writers--;
 742         }
 743
 744         lu_ref_del(&lock->l_reference, "user", lock);
 745         LDLM_LOCK_RELEASE(lock);    /* matches the LDLM_LOCK_GET() in addref */
 746 }
 747
 748 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
 749 {
 750         struct ldlm_namespace *ns;
 751         ENTRY;
 752
 753         lock_res_and_lock(lock);
 754
 755         ns = ldlm_lock_to_ns(lock);
 756
 757         ldlm_lock_decref_internal_nolock(lock, mode);
 758
 759         if (lock->l_flags & LDLM_FL_LOCAL &&
 760             !lock->l_readers && !lock->l_writers) {
 761                 /* If this is a local lock on a server namespace and this was
 762                  * the last reference, cancel the lock. */
 763                 CDEBUG(D_INFO, "forcing cancel of local lock\n");
 764                 lock->l_flags |= LDLM_FL_CBPENDING;
 765         }
 766
 767         if (!lock->l_readers && !lock->l_writers &&
 768             (lock->l_flags & LDLM_FL_CBPENDING)) {
 769                 /* If we received a blocked AST and this was the last reference,
 770                  * run the callback. */
 771                 if (lock->l_ns_srv && lock->l_export)
 772                         CERROR("FL_CBPENDING set on non-local lock--just a "
 773                                "warning\n");
 774
 775                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 776
 777                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
 778                 ldlm_lock_remove_from_lru(lock);
 779                 unlock_res_and_lock(lock);
 780
 781                 if (lock->l_flags & LDLM_FL_FAIL_LOC)
 782                         OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
 783
 784                 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
 785                     ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
 786                         ldlm_handle_bl_callback(ns, NULL, lock);
 787         } else if (ns_is_client(ns) &&
 788                    !lock->l_readers && !lock->l_writers &&
 789                    !(lock->l_flags & LDLM_FL_NO_LRU) &&
 790                    !(lock->l_flags & LDLM_FL_BL_AST)) {
 791
 792                 LDLM_DEBUG(lock, "add lock into lru list");
 793
 794                 /* If this is a client-side namespace and this was the last
 795                  * reference, put it on the LRU. */
 796                 ldlm_lock_add_to_lru(lock);
 797                 unlock_res_and_lock(lock);
 798
 799                 if (lock->l_flags & LDLM_FL_FAIL_LOC)
 800                         OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
 801
 802                 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
 803                  * are not supported by the server, otherwise, it is done on
 804                  * enqueue. */
 805                 if (!exp_connect_cancelset(lock->l_conn_export) &&
 806                     !ns_connect_lru_resize(ns))
 807                         ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
 808         } else {
 809                 LDLM_DEBUG(lock, "do not add lock into lru list");
 810                 unlock_res_and_lock(lock);
 811         }
 812
 813         EXIT;
 814 }
 815
 816 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 817 {
 818         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 819         LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
 820         ldlm_lock_decref_internal(lock, mode);
 821         LDLM_LOCK_PUT(lock);
 822 }
 823
 824 /* This will drop a lock reference and mark it for destruction, but will not
 825  * necessarily cancel the lock before returning. */
 826 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
 827 {
 828         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 829         ENTRY;
 830
 831         LASSERT(lock != NULL);
 832
 833         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 834         lock_res_and_lock(lock);
 835         lock->l_flags |= LDLM_FL_CBPENDING;
 836         unlock_res_and_lock(lock);
 837         ldlm_lock_decref_internal(lock, mode);
 838         LDLM_LOCK_PUT(lock);
 839 }
 840
 841 struct sl_insert_point {
 842         cfs_list_t *res_link;
 843         cfs_list_t *mode_link;
 844         cfs_list_t *policy_link;
 845 };
 846
 847 /*
 848  * search_granted_lock
 849  *
 850  * Description:
 851  *      Finds a position to insert the new lock.
 852  * Parameters:
 853  *      queue [input]:  the granted list where search acts on;
 854  *      req [input]:    the lock whose position to be located;
 855  *      prev [output]:  positions within 3 lists to insert @req to
 856  * Return Value:
 857  *      filled @prev
 858  * NOTE: called by
 859  *  - ldlm_grant_lock_with_skiplist
 860  */
 861 static void search_granted_lock(cfs_list_t *queue,
 862                                 struct ldlm_lock *req,
 863                                 struct sl_insert_point *prev)
 864 {
 865         cfs_list_t *tmp;
 866         struct ldlm_lock *lock, *mode_end, *policy_end;
 867         ENTRY;
 868
 869         cfs_list_for_each(tmp, queue) {
 870                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
 871
 872                 mode_end = cfs_list_entry(lock->l_sl_mode.prev,
 873                                           struct ldlm_lock, l_sl_mode);
 874
 875                 if (lock->l_req_mode != req->l_req_mode) {
 876                         /* jump to last lock of mode group */
 877                         tmp = &mode_end->l_res_link;
 878                         continue;
 879                 }
 880
 881                 /* suitable mode group is found */
 882                 if (lock->l_resource->lr_type == LDLM_PLAIN) {
 883                         /* insert point is last lock of the mode group */
 884                         prev->res_link = &mode_end->l_res_link;
 885                         prev->mode_link = &mode_end->l_sl_mode;
 886                         prev->policy_link = &req->l_sl_policy;
 887                         EXIT;
 888                         return;
 889                 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
 890                         for (;;) {
 891                                 policy_end =
 892                                         cfs_list_entry(lock->l_sl_policy.prev,
 893                                                        struct ldlm_lock,
 894                                                        l_sl_policy);
 895
 896                                 if (lock->l_policy_data.l_inodebits.bits ==
 897                                     req->l_policy_data.l_inodebits.bits) {
 898                                         /* insert point is last lock of
 899                                          * the policy group */
 900                                         prev->res_link =
 901                                                 &policy_end->l_res_link;
 902                                         prev->mode_link =
 903                                                 &policy_end->l_sl_mode;
 904                                         prev->policy_link =
 905                                                 &policy_end->l_sl_policy;
 906                                         EXIT;
 907                                         return;
 908                                 }
 909
 910                                 if (policy_end == mode_end)
 911                                         /* done with mode group */
 912                                         break;
 913
 914                                 /* go to next policy group within mode group */
 915                                 tmp = policy_end->l_res_link.next;
 916                                 lock = cfs_list_entry(tmp, struct ldlm_lock,
 917                                                       l_res_link);
 918                         }  /* loop over policy groups within the mode group */
 919
 920                         /* insert point is last lock of the mode group,
 921                          * new policy group is started */
 922                         prev->res_link = &mode_end->l_res_link;
 923                         prev->mode_link = &mode_end->l_sl_mode;
 924                         prev->policy_link = &req->l_sl_policy;
 925                         EXIT;
 926                         return;
 927                 } else {
 928                         LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
 929                         LBUG();
 930                 }
 931         }
 932
 933         /* insert point is last lock on the queue,
 934          * new mode group and new policy group are started */
 935         prev->res_link = queue->prev;
 936         prev->mode_link = &req->l_sl_mode;
 937         prev->policy_link = &req->l_sl_policy;
 938         EXIT;
 939         return;
 940 }
 941
 942 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
 943                                        struct sl_insert_point *prev)
 944 {
 945         struct ldlm_resource *res = lock->l_resource;
 946         ENTRY;
 947
 948         check_res_locked(res);
 949
 950         ldlm_resource_dump(D_INFO, res);
 951         CDEBUG(D_OTHER, "About to add this lock:\n");
 952         ldlm_lock_dump(D_OTHER, lock, 0);
 953
 954         if (lock->l_destroyed) {
 955                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 956                 return;
 957         }
 958
 959         LASSERT(cfs_list_empty(&lock->l_res_link));
 960         LASSERT(cfs_list_empty(&lock->l_sl_mode));
 961         LASSERT(cfs_list_empty(&lock->l_sl_policy));
 962
 963         cfs_list_add(&lock->l_res_link, prev->res_link);
 964         cfs_list_add(&lock->l_sl_mode, prev->mode_link);
 965         cfs_list_add(&lock->l_sl_policy, prev->policy_link);
 966
 967         EXIT;
 968 }
 969
 970 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
 971 {
 972         struct sl_insert_point prev;
 973         ENTRY;
 974
 975         LASSERT(lock->l_req_mode == lock->l_granted_mode);
 976
 977         search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
 978         ldlm_granted_list_add_lock(lock, &prev);
 979         EXIT;
 980 }
 981
 982 /* NOTE: called by
 983  *  - ldlm_lock_enqueue
 984  *  - ldlm_reprocess_queue
 985  *  - ldlm_lock_convert
 986  *
 987  * must be called with lr_lock held
 988  */
 989 void ldlm_grant_lock(struct ldlm_lock *lock, cfs_list_t *work_list)
 990 {
 991         struct ldlm_resource *res = lock->l_resource;
 992         ENTRY;
 993
 994         check_res_locked(res);
 995
 996         lock->l_granted_mode = lock->l_req_mode;
 997         if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
 998                 ldlm_grant_lock_with_skiplist(lock);
 999         else if (res->lr_type == LDLM_EXTENT)
1000                 ldlm_extent_add_lock(res, lock);
1001         else
1002                 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1003
1004         if (lock->l_granted_mode < res->lr_most_restr)
1005                 res->lr_most_restr = lock->l_granted_mode;
1006
1007         if (work_list && lock->l_completion_ast != NULL)
1008                 ldlm_add_ast_work_item(lock, NULL, work_list);
1009
1010         ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1011         EXIT;
1012 }
1013
1014 /* returns a referenced lock or NULL.  See the flag descriptions below, in the
1015  * comment above ldlm_lock_match */
1016 static struct ldlm_lock *search_queue(cfs_list_t *queue,
1017                                       ldlm_mode_t *mode,
1018                                       ldlm_policy_data_t *policy,
1019                                       struct ldlm_lock *old_lock,
1020                                       int flags, int unref)
1021 {
1022         struct ldlm_lock *lock;
1023         cfs_list_t       *tmp;
1024
1025         cfs_list_for_each(tmp, queue) {
1026                 ldlm_mode_t match;
1027
1028                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1029
1030                 if (lock == old_lock)
1031                         break;
1032
1033                 /* llite sometimes wants to match locks that will be
1034                  * canceled when their users drop, but we allow it to match
1035                  * if it passes in CBPENDING and the lock still has users.
1036                  * this is generally only going to be used by children
1037                  * whose parents already hold a lock so forward progress
1038                  * can still happen. */
1039                 if (lock->l_flags & LDLM_FL_CBPENDING &&
1040                     !(flags & LDLM_FL_CBPENDING))
1041                         continue;
1042                 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
1043                     lock->l_readers == 0 && lock->l_writers == 0)
1044                         continue;
1045
1046                 if (!(lock->l_req_mode & *mode))
1047                         continue;
1048                 match = lock->l_req_mode;
1049
1050                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
1051                     (lock->l_policy_data.l_extent.start >
1052                      policy->l_extent.start ||
1053                      lock->l_policy_data.l_extent.end < policy->l_extent.end))
1054                         continue;
1055
1056                 if (unlikely(match == LCK_GROUP) &&
1057                     lock->l_resource->lr_type == LDLM_EXTENT &&
1058                     lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
1059                         continue;
1060
1061                 /* We match if we have existing lock with same or wider set
1062                    of bits. */
1063                 if (lock->l_resource->lr_type == LDLM_IBITS &&
1064                      ((lock->l_policy_data.l_inodebits.bits &
1065                       policy->l_inodebits.bits) !=
1066                       policy->l_inodebits.bits))
1067                         continue;
1068
1069                 if (!unref &&
1070                     (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED ||
1071                      lock->l_failed))
1072                         continue;
1073
1074                 if ((flags & LDLM_FL_LOCAL_ONLY) &&
1075                     !(lock->l_flags & LDLM_FL_LOCAL))
1076                         continue;
1077
1078                 if (flags & LDLM_FL_TEST_LOCK) {
1079                         LDLM_LOCK_GET(lock);
1080                         ldlm_lock_touch_in_lru(lock);
1081                 } else {
1082                         ldlm_lock_addref_internal_nolock(lock, match);
1083                 }
1084                 *mode = match;
1085                 return lock;
1086         }
1087
1088         return NULL;
1089 }
1090
1091 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1092 {
1093         if (!lock->l_failed) {
1094                 lock->l_failed = 1;
1095                 cfs_waitq_broadcast(&lock->l_waitq);
1096         }
1097 }
1098 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1099
1100 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1101 {
1102         lock_res_and_lock(lock);
1103         ldlm_lock_fail_match_locked(lock);
1104         unlock_res_and_lock(lock);
1105 }
1106 EXPORT_SYMBOL(ldlm_lock_fail_match);
1107
1108 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1109 {
1110         lock->l_flags |= LDLM_FL_LVB_READY;
1111         cfs_waitq_broadcast(&lock->l_waitq);
1112 }
1113
1114 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1115 {
1116         lock_res_and_lock(lock);
1117         ldlm_lock_allow_match_locked(lock);
1118         unlock_res_and_lock(lock);
1119 }
1120
1121 /* Can be called in two ways:
1122  *
1123  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1124  * for a duplicate of.
1125  *
1126  * Otherwise, all of the fields must be filled in, to match against.
1127  *
1128  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1129  *     server (ie, connh is NULL)
1130  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1131  *     list will be considered
1132  * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1133  *     to be canceled can still be matched as long as they still have reader
1134  *     or writer refernces
1135  * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1136  *     just tell us if we would have matched.
1137  *
1138  * Returns 1 if it finds an already-existing lock that is compatible; in this
1139  * case, lockh is filled in with a addref()ed lock
1140  *
1141  * we also check security context, if that failed we simply return 0 (to keep
1142  * caller code unchanged), the context failure will be discovered by caller
1143  * sometime later.
1144  */
1145 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1146                             const struct ldlm_res_id *res_id, ldlm_type_t type,
1147                             ldlm_policy_data_t *policy, ldlm_mode_t mode,
1148                             struct lustre_handle *lockh, int unref)
1149 {
1150         struct ldlm_resource *res;
1151         struct ldlm_lock *lock, *old_lock = NULL;
1152         int rc = 0;
1153         ENTRY;
1154
1155         if (ns == NULL) {
1156                 old_lock = ldlm_handle2lock(lockh);
1157                 LASSERT(old_lock);
1158
1159                 ns = ldlm_lock_to_ns(old_lock);
1160                 res_id = &old_lock->l_resource->lr_name;
1161                 type = old_lock->l_resource->lr_type;
1162                 mode = old_lock->l_req_mode;
1163         }
1164
1165         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1166         if (res == NULL) {
1167                 LASSERT(old_lock == NULL);
1168                 RETURN(0);
1169         }
1170
1171         LDLM_RESOURCE_ADDREF(res);
1172         lock_res(res);
1173
1174         lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1175                             flags, unref);
1176         if (lock != NULL)
1177                 GOTO(out, rc = 1);
1178         if (flags & LDLM_FL_BLOCK_GRANTED)
1179                 GOTO(out, rc = 0);
1180         lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1181                             flags, unref);
1182         if (lock != NULL)
1183                 GOTO(out, rc = 1);
1184         lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1185                             flags, unref);
1186         if (lock != NULL)
1187                 GOTO(out, rc = 1);
1188
1189         EXIT;
1190  out:
1191         unlock_res(res);
1192         LDLM_RESOURCE_DELREF(res);
1193         ldlm_resource_putref(res);
1194
1195         if (lock) {
1196                 ldlm_lock2handle(lock, lockh);
1197                 if ((flags & LDLM_FL_LVB_READY) &&
1198                     (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1199                         struct l_wait_info lwi;
1200                         if (lock->l_completion_ast) {
1201                                 int err = lock->l_completion_ast(lock,
1202                                                           LDLM_FL_WAIT_NOREPROC,
1203                                                                  NULL);
1204                                 if (err) {
1205                                         if (flags & LDLM_FL_TEST_LOCK)
1206                                                 LDLM_LOCK_RELEASE(lock);
1207                                         else
1208                                                 ldlm_lock_decref_internal(lock,
1209                                                                           mode);
1210                                         rc = 0;
1211                                         goto out2;
1212                                 }
1213                         }
1214
1215                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1216                                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1217
1218                         /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1219                         l_wait_event(lock->l_waitq,
1220                                      lock->l_flags & LDLM_FL_LVB_READY ||
1221                                      lock->l_failed,
1222                                      &lwi);
1223                         if (!(lock->l_flags & LDLM_FL_LVB_READY)) {
1224                                 if (flags & LDLM_FL_TEST_LOCK)
1225                                         LDLM_LOCK_RELEASE(lock);
1226                                 else
1227                                         ldlm_lock_decref_internal(lock, mode);
1228                                 rc = 0;
1229                         }
1230                 }
1231         }
1232  out2:
1233         if (rc) {
1234                 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1235                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1236                                 res_id->name[2] : policy->l_extent.start,
1237                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1238                                 res_id->name[3] : policy->l_extent.end);
1239
1240                 /* check user's security context */
1241                 if (lock->l_conn_export &&
1242                     sptlrpc_import_check_ctx(
1243                                 class_exp2cliimp(lock->l_conn_export))) {
1244                         if (!(flags & LDLM_FL_TEST_LOCK))
1245                                 ldlm_lock_decref_internal(lock, mode);
1246                         rc = 0;
1247                 }
1248
1249                 if (flags & LDLM_FL_TEST_LOCK)
1250                         LDLM_LOCK_RELEASE(lock);
1251
1252         } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1253                 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1254                                   LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1255                                   type, mode, res_id->name[0], res_id->name[1],
1256                                   (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1257                                         res_id->name[2] :policy->l_extent.start,
1258                                   (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1259                                         res_id->name[3] : policy->l_extent.end);
1260         }
1261         if (old_lock)
1262                 LDLM_LOCK_PUT(old_lock);
1263
1264         return rc ? mode : 0;
1265 }
1266
1267 ldlm_mode_t ldlm_revalidate_lock_handle(struct lustre_handle *lockh,
1268                                         __u64 *bits)
1269 {
1270         struct ldlm_lock *lock;
1271         ldlm_mode_t mode = 0;
1272         ENTRY;
1273
1274         lock = ldlm_handle2lock(lockh);
1275         if (lock != NULL) {
1276                 lock_res_and_lock(lock);
1277                 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED ||
1278                     lock->l_failed)
1279                         GOTO(out, mode);
1280
1281                 if (lock->l_flags & LDLM_FL_CBPENDING &&
1282                     lock->l_readers == 0 && lock->l_writers == 0)
1283                         GOTO(out, mode);
1284
1285                 if (bits)
1286                         *bits = lock->l_policy_data.l_inodebits.bits;
1287                 mode = lock->l_granted_mode;
1288                 ldlm_lock_addref_internal_nolock(lock, mode);
1289         }
1290
1291         EXIT;
1292
1293 out:
1294         if (lock != NULL) {
1295                 unlock_res_and_lock(lock);
1296                 LDLM_LOCK_PUT(lock);
1297         }
1298         return mode;
1299 }
1300 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1301
1302 /* Returns a referenced lock */
1303 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1304                                    const struct ldlm_res_id *res_id,
1305                                    ldlm_type_t type,
1306                                    ldlm_mode_t mode,
1307                                    const struct ldlm_callback_suite *cbs,
1308                                    void *data, __u32 lvb_len)
1309 {
1310         struct ldlm_lock *lock;
1311         struct ldlm_resource *res;
1312         ENTRY;
1313
1314         res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1315         if (res == NULL)
1316                 RETURN(NULL);
1317
1318         lock = ldlm_lock_new(res);
1319
1320         if (lock == NULL)
1321                 RETURN(NULL);
1322
1323         lock->l_req_mode = mode;
1324         lock->l_ast_data = data;
1325         lock->l_pid = cfs_curproc_pid();
1326         lock->l_ns_srv = !!ns_is_server(ns);
1327         if (cbs) {
1328                 lock->l_blocking_ast = cbs->lcs_blocking;
1329                 lock->l_completion_ast = cbs->lcs_completion;
1330                 lock->l_glimpse_ast = cbs->lcs_glimpse;
1331                 lock->l_weigh_ast = cbs->lcs_weigh;
1332         }
1333
1334         lock->l_tree_node = NULL;
1335         /* if this is the extent lock, allocate the interval tree node */
1336         if (type == LDLM_EXTENT) {
1337                 if (ldlm_interval_alloc(lock) == NULL)
1338                         GOTO(out, 0);
1339         }
1340
1341         if (lvb_len) {
1342                 lock->l_lvb_len = lvb_len;
1343                 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1344                 if (lock->l_lvb_data == NULL)
1345                         GOTO(out, 0);
1346         }
1347
1348         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1349                 GOTO(out, 0);
1350
1351         RETURN(lock);
1352
1353 out:
1354         ldlm_lock_destroy(lock);
1355         LDLM_LOCK_RELEASE(lock);
1356         return NULL;
1357 }
1358
1359 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1360                                struct ldlm_lock **lockp,
1361                                void *cookie, int *flags)
1362 {
1363         struct ldlm_lock *lock = *lockp;
1364         struct ldlm_resource *res = lock->l_resource;
1365         int local = ns_is_client(ldlm_res_to_ns(res));
1366 #ifdef HAVE_SERVER_SUPPORT
1367         ldlm_processing_policy policy;
1368 #endif
1369         ldlm_error_t rc = ELDLM_OK;
1370         struct ldlm_interval *node = NULL;
1371         ENTRY;
1372
1373         lock->l_last_activity = cfs_time_current_sec();
1374         /* policies are not executed on the client or during replay */
1375         if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1376             && !local && ns->ns_policy) {
1377                 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1378                                    NULL);
1379                 if (rc == ELDLM_LOCK_REPLACED) {
1380                         /* The lock that was returned has already been granted,
1381                          * and placed into lockp.  If it's not the same as the
1382                          * one we passed in, then destroy the old one and our
1383                          * work here is done. */
1384                         if (lock != *lockp) {
1385                                 ldlm_lock_destroy(lock);
1386                                 LDLM_LOCK_RELEASE(lock);
1387                         }
1388                         *flags |= LDLM_FL_LOCK_CHANGED;
1389                         RETURN(0);
1390                 } else if (rc != ELDLM_OK ||
1391                            (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1392                         ldlm_lock_destroy(lock);
1393                         RETURN(rc);
1394                 }
1395         }
1396
1397         /* For a replaying lock, it might be already in granted list. So
1398          * unlinking the lock will cause the interval node to be freed, we
1399          * have to allocate the interval node early otherwise we can't regrant
1400          * this lock in the future. - jay */
1401         if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1402                 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1403
1404         lock_res_and_lock(lock);
1405         if (local && lock->l_req_mode == lock->l_granted_mode) {
1406                 /* The server returned a blocked lock, but it was granted
1407                  * before we got a chance to actually enqueue it.  We don't
1408                  * need to do anything else. */
1409                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1410                             LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1411                 GOTO(out, ELDLM_OK);
1412         }
1413
1414         ldlm_resource_unlink_lock(lock);
1415         if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1416                 if (node == NULL) {
1417                         ldlm_lock_destroy_nolock(lock);
1418                         GOTO(out, rc = -ENOMEM);
1419                 }
1420
1421                 CFS_INIT_LIST_HEAD(&node->li_group);
1422                 ldlm_interval_attach(node, lock);
1423                 node = NULL;
1424         }
1425
1426         /* Some flags from the enqueue want to make it into the AST, via the
1427          * lock's l_flags. */
1428         lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1429
1430         /* This distinction between local lock trees is very important; a client
1431          * namespace only has information about locks taken by that client, and
1432          * thus doesn't have enough information to decide for itself if it can
1433          * be granted (below).  In this case, we do exactly what the server
1434          * tells us to do, as dictated by the 'flags'.
1435          *
1436          * We do exactly the same thing during recovery, when the server is
1437          * more or less trusting the clients not to lie.
1438          *
1439          * FIXME (bug 268): Detect obvious lies by checking compatibility in
1440          * granted/converting queues. */
1441         if (local) {
1442                 if (*flags & LDLM_FL_BLOCK_CONV)
1443                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1444                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1445                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1446                 else
1447                         ldlm_grant_lock(lock, NULL);
1448                 GOTO(out, ELDLM_OK);
1449 #ifdef HAVE_SERVER_SUPPORT
1450         } else if (*flags & LDLM_FL_REPLAY) {
1451                 if (*flags & LDLM_FL_BLOCK_CONV) {
1452                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1453                         GOTO(out, ELDLM_OK);
1454                 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1455                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1456                         GOTO(out, ELDLM_OK);
1457                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1458                         ldlm_grant_lock(lock, NULL);
1459                         GOTO(out, ELDLM_OK);
1460                 }
1461                 /* If no flags, fall through to normal enqueue path. */
1462         }
1463
1464         policy = ldlm_processing_policy_table[res->lr_type];
1465         policy(lock, flags, 1, &rc, NULL);
1466         GOTO(out, rc);
1467 #else
1468         } else {
1469                 CERROR("This is client-side-only module, cannot handle "
1470                        "LDLM_NAMESPACE_SERVER resource type lock.\n");
1471                 LBUG();
1472         }
1473 #endif
1474
1475 out:
1476         unlock_res_and_lock(lock);
1477         if (node)
1478                 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1479         return rc;
1480 }
1481
1482 #ifdef HAVE_SERVER_SUPPORT
1483 /* Must be called with namespace taken: queue is waiting or converting. */
1484 int ldlm_reprocess_queue(struct ldlm_resource *res, cfs_list_t *queue,
1485                          cfs_list_t *work_list)
1486 {
1487         cfs_list_t *tmp, *pos;
1488         ldlm_processing_policy policy;
1489         int flags;
1490         int rc = LDLM_ITER_CONTINUE;
1491         ldlm_error_t err;
1492         ENTRY;
1493
1494         check_res_locked(res);
1495
1496         policy = ldlm_processing_policy_table[res->lr_type];
1497         LASSERT(policy);
1498
1499         cfs_list_for_each_safe(tmp, pos, queue) {
1500                 struct ldlm_lock *pending;
1501                 pending = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1502
1503                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1504
1505                 flags = 0;
1506                 rc = policy(pending, &flags, 0, &err, work_list);
1507                 if (rc != LDLM_ITER_CONTINUE)
1508                         break;
1509         }
1510
1511         RETURN(rc);
1512 }
1513 #endif
1514
1515 static int
1516 ldlm_work_bl_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1517 {
1518         struct ldlm_lock_desc d;
1519         struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1520                                                 l_bl_ast);
1521         int rc;
1522         ENTRY;
1523
1524         /* nobody should touch l_bl_ast */
1525         lock_res_and_lock(lock);
1526         cfs_list_del_init(&lock->l_bl_ast);
1527
1528         LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1529         LASSERT(lock->l_bl_ast_run == 0);
1530         LASSERT(lock->l_blocking_lock);
1531         lock->l_bl_ast_run++;
1532         unlock_res_and_lock(lock);
1533
1534         ldlm_lock2desc(lock->l_blocking_lock, &d);
1535
1536         rc = lock->l_blocking_ast(lock, &d, (void *)arg,
1537                                   LDLM_CB_BLOCKING);
1538         LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1539         lock->l_blocking_lock = NULL;
1540         LDLM_LOCK_RELEASE(lock);
1541
1542         RETURN(rc);
1543 }
1544
1545 static int
1546 ldlm_work_cp_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1547 {
1548         struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock, l_cp_ast);
1549         ldlm_completion_callback completion_callback;
1550         int rc = 0;
1551         ENTRY;
1552
1553         /* It's possible to receive a completion AST before we've set
1554          * the l_completion_ast pointer: either because the AST arrived
1555          * before the reply, or simply because there's a small race
1556          * window between receiving the reply and finishing the local
1557          * enqueue. (bug 842)
1558          *
1559          * This can't happen with the blocking_ast, however, because we
1560          * will never call the local blocking_ast until we drop our
1561          * reader/writer reference, which we won't do until we get the
1562          * reply and finish enqueueing. */
1563
1564         /* nobody should touch l_cp_ast */
1565         lock_res_and_lock(lock);
1566         cfs_list_del_init(&lock->l_cp_ast);
1567         LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1568         /* save l_completion_ast since it can be changed by
1569          * mds_intent_policy(), see bug 14225 */
1570         completion_callback = lock->l_completion_ast;
1571         lock->l_flags &= ~LDLM_FL_CP_REQD;
1572         unlock_res_and_lock(lock);
1573
1574         if (completion_callback != NULL)
1575                 rc = completion_callback(lock, 0, (void *)arg);
1576         LDLM_LOCK_RELEASE(lock);
1577
1578         RETURN(rc);
1579 }
1580
1581 static int
1582 ldlm_work_revoke_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1583 {
1584         struct ldlm_lock_desc desc;
1585         struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1586                                                 l_rk_ast);
1587         int rc;
1588         ENTRY;
1589
1590         cfs_list_del_init(&lock->l_rk_ast);
1591
1592         /* the desc just pretend to exclusive */
1593         ldlm_lock2desc(lock, &desc);
1594         desc.l_req_mode = LCK_EX;
1595         desc.l_granted_mode = 0;
1596
1597         rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1598         LDLM_LOCK_RELEASE(lock);
1599
1600         RETURN(rc);
1601 }
1602
1603 int ldlm_run_ast_work(struct ldlm_namespace *ns, cfs_list_t *rpc_list,
1604                       ldlm_desc_ast_t ast_type)
1605 {
1606         struct l_wait_info     lwi = { 0 };
1607         struct ldlm_cb_set_arg *arg;
1608         cfs_list_t *tmp, *pos;
1609         int (*work_ast_lock)(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg);
1610         unsigned int max_ast_count;
1611         int rc;
1612         ENTRY;
1613
1614         if (cfs_list_empty(rpc_list))
1615                 RETURN(0);
1616
1617         OBD_ALLOC_PTR(arg);
1618         if (arg == NULL)
1619                 RETURN(-ENOMEM);
1620
1621         cfs_atomic_set(&arg->restart, 0);
1622         cfs_atomic_set(&arg->rpcs, 0);
1623         cfs_atomic_set(&arg->refcount, 1);
1624         cfs_waitq_init(&arg->waitq);
1625
1626         switch (ast_type) {
1627         case LDLM_WORK_BL_AST:
1628                 arg->type = LDLM_BL_CALLBACK;
1629                 work_ast_lock = ldlm_work_bl_ast_lock;
1630                 break;
1631         case LDLM_WORK_CP_AST:
1632                 arg->type = LDLM_CP_CALLBACK;
1633                 work_ast_lock = ldlm_work_cp_ast_lock;
1634                 break;
1635         case LDLM_WORK_REVOKE_AST:
1636                 arg->type = LDLM_BL_CALLBACK;
1637                 work_ast_lock = ldlm_work_revoke_ast_lock;
1638                 break;
1639         default:
1640                 LBUG();
1641         }
1642
1643         max_ast_count = ns->ns_max_parallel_ast ? : UINT_MAX;
1644         arg->threshold = max_ast_count;
1645
1646         cfs_list_for_each_safe(tmp, pos, rpc_list) {
1647                 (void)work_ast_lock(tmp, arg);
1648                 if (cfs_atomic_read(&arg->rpcs) < max_ast_count)
1649                         continue;
1650
1651                 l_wait_event(arg->waitq,
1652                              cfs_atomic_read(&arg->rpcs) < arg->threshold,
1653                              &lwi);
1654         }
1655
1656         arg->threshold = 1;
1657         l_wait_event(arg->waitq, cfs_atomic_read(&arg->rpcs) == 0, &lwi);
1658
1659         rc = cfs_atomic_read(&arg->restart) ? -ERESTART : 0;
1660         ldlm_csa_put(arg);
1661         RETURN(rc);
1662 }
1663
1664 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1665 {
1666         ldlm_reprocess_all(res);
1667         return LDLM_ITER_CONTINUE;
1668 }
1669
1670 static int ldlm_reprocess_res(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1671                               cfs_hlist_node_t *hnode, void *arg)
1672 {
1673         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1674         int    rc;
1675
1676         rc = reprocess_one_queue(res, arg);
1677
1678         return rc == LDLM_ITER_STOP;
1679 }
1680
1681 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1682 {
1683         ENTRY;
1684
1685         if (ns != NULL) {
1686                 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1687                                          ldlm_reprocess_res, NULL);
1688         }
1689         EXIT;
1690 }
1691
1692 void ldlm_reprocess_all(struct ldlm_resource *res)
1693 {
1694         CFS_LIST_HEAD(rpc_list);
1695
1696 #ifdef HAVE_SERVER_SUPPORT
1697         int rc;
1698         ENTRY;
1699         /* Local lock trees don't get reprocessed. */
1700         if (ns_is_client(ldlm_res_to_ns(res))) {
1701                 EXIT;
1702                 return;
1703         }
1704
1705 restart:
1706         lock_res(res);
1707         rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1708         if (rc == LDLM_ITER_CONTINUE)
1709                 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1710         unlock_res(res);
1711
1712         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
1713                                LDLM_WORK_CP_AST);
1714         if (rc == -ERESTART) {
1715                 LASSERT(cfs_list_empty(&rpc_list));
1716                 goto restart;
1717         }
1718 #else
1719         ENTRY;
1720         if (!ns_is_client(ldlm_res_to_ns(res))) {
1721                 CERROR("This is client-side-only module, cannot handle "
1722                        "LDLM_NAMESPACE_SERVER resource type lock.\n");
1723                 LBUG();
1724         }
1725 #endif
1726         EXIT;
1727 }
1728
1729 void ldlm_cancel_callback(struct ldlm_lock *lock)
1730 {
1731         check_res_locked(lock->l_resource);
1732         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1733                 lock->l_flags |= LDLM_FL_CANCEL;
1734                 if (lock->l_blocking_ast) {
1735                         // l_check_no_ns_lock(ns);
1736                         unlock_res_and_lock(lock);
1737                         lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1738                                              LDLM_CB_CANCELING);
1739                         lock_res_and_lock(lock);
1740                 } else {
1741                         LDLM_DEBUG(lock, "no blocking ast");
1742                 }
1743         }
1744         lock->l_flags |= LDLM_FL_BL_DONE;
1745 }
1746
1747 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1748 {
1749         if (req->l_resource->lr_type != LDLM_PLAIN &&
1750             req->l_resource->lr_type != LDLM_IBITS)
1751                 return;
1752
1753         cfs_list_del_init(&req->l_sl_policy);
1754         cfs_list_del_init(&req->l_sl_mode);
1755 }
1756
1757 void ldlm_lock_cancel(struct ldlm_lock *lock)
1758 {
1759         struct ldlm_resource *res;
1760         struct ldlm_namespace *ns;
1761         ENTRY;
1762
1763         lock_res_and_lock(lock);
1764
1765         res = lock->l_resource;
1766         ns  = ldlm_res_to_ns(res);
1767
1768         /* Please do not, no matter how tempting, remove this LBUG without
1769          * talking to me first. -phik */
1770         if (lock->l_readers || lock->l_writers) {
1771                 LDLM_ERROR(lock, "lock still has references");
1772                 LBUG();
1773         }
1774
1775         ldlm_del_waiting_lock(lock);
1776
1777         /* Releases cancel callback. */
1778         ldlm_cancel_callback(lock);
1779
1780         /* Yes, second time, just in case it was added again while we were
1781            running with no res lock in ldlm_cancel_callback */
1782         ldlm_del_waiting_lock(lock);
1783         ldlm_resource_unlink_lock(lock);
1784         ldlm_lock_destroy_nolock(lock);
1785
1786         if (lock->l_granted_mode == lock->l_req_mode)
1787                 ldlm_pool_del(&ns->ns_pool, lock);
1788
1789         /* Make sure we will not be called again for same lock what is possible
1790          * if not to zero out lock->l_granted_mode */
1791         lock->l_granted_mode = LCK_MINMODE;
1792         unlock_res_and_lock(lock);
1793
1794         EXIT;
1795 }
1796
1797 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1798 {
1799         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1800         int rc = -EINVAL;
1801         ENTRY;
1802
1803         if (lock) {
1804                 if (lock->l_ast_data == NULL)
1805                         lock->l_ast_data = data;
1806                 if (lock->l_ast_data == data)
1807                         rc = 0;
1808                 LDLM_LOCK_PUT(lock);
1809         }
1810         RETURN(rc);
1811 }
1812 EXPORT_SYMBOL(ldlm_lock_set_data);
1813
1814 int ldlm_cancel_locks_for_export_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1815                                     cfs_hlist_node_t *hnode, void *data)
1816
1817 {
1818         struct obd_export    *exp  = data;
1819         struct ldlm_lock     *lock = cfs_hash_object(hs, hnode);
1820         struct ldlm_resource *res;
1821
1822         res = ldlm_resource_getref(lock->l_resource);
1823         LDLM_LOCK_GET(lock);
1824
1825         LDLM_DEBUG(lock, "export %p", exp);
1826         ldlm_res_lvbo_update(res, NULL, 1);
1827         ldlm_lock_cancel(lock);
1828         ldlm_reprocess_all(res);
1829         ldlm_resource_putref(res);
1830         LDLM_LOCK_RELEASE(lock);
1831         return 0;
1832 }
1833
1834 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1835 {
1836         cfs_hash_for_each_empty(exp->exp_lock_hash,
1837                                 ldlm_cancel_locks_for_export_cb, exp);
1838 }
1839
1840 /**
1841  * Downgrade an exclusive lock.
1842  *
1843  * A fast variant of ldlm_lock_convert for convertion of exclusive
1844  * locks. The convertion is always successful.
1845  *
1846  * \param lock A lock to convert
1847  * \param new_mode new lock mode
1848  */
1849 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1850 {
1851         ENTRY;
1852
1853         LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1854         LASSERT(new_mode == LCK_COS);
1855
1856         lock_res_and_lock(lock);
1857         ldlm_resource_unlink_lock(lock);
1858         /*
1859          * Remove the lock from pool as it will be added again in
1860          * ldlm_grant_lock() called below.
1861          */
1862         ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
1863
1864         lock->l_req_mode = new_mode;
1865         ldlm_grant_lock(lock, NULL);
1866         unlock_res_and_lock(lock);
1867         ldlm_reprocess_all(lock->l_resource);
1868
1869         EXIT;
1870 }
1871
1872 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1873                                         __u32 *flags)
1874 {
1875         CFS_LIST_HEAD(rpc_list);
1876         struct ldlm_resource *res;
1877         struct ldlm_namespace *ns;
1878         int granted = 0;
1879         int old_mode;
1880         struct sl_insert_point prev;
1881         struct ldlm_interval *node;
1882         ENTRY;
1883
1884         if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1885                 *flags |= LDLM_FL_BLOCK_GRANTED;
1886                 RETURN(lock->l_resource);
1887         }
1888
1889         /* I can't check the type of lock here because the bitlock of lock
1890          * is not held here, so do the allocation blindly. -jay */
1891         OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1892         if (node == NULL)  /* Actually, this causes EDEADLOCK to be returned */
1893                 RETURN(NULL);
1894
1895         LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1896                  "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1897
1898         lock_res_and_lock(lock);
1899
1900         res = lock->l_resource;
1901         ns  = ldlm_res_to_ns(res);
1902
1903         old_mode = lock->l_req_mode;
1904         lock->l_req_mode = new_mode;
1905         if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1906                 /* remember the lock position where the lock might be
1907                  * added back to the granted list later and also
1908                  * remember the join mode for skiplist fixing. */
1909                 prev.res_link = lock->l_res_link.prev;
1910                 prev.mode_link = lock->l_sl_mode.prev;
1911                 prev.policy_link = lock->l_sl_policy.prev;
1912                 ldlm_resource_unlink_lock(lock);
1913         } else {
1914                 ldlm_resource_unlink_lock(lock);
1915                 if (res->lr_type == LDLM_EXTENT) {
1916                         /* FIXME: ugly code, I have to attach the lock to a
1917                          * interval node again since perhaps it will be granted
1918                          * soon */
1919                         CFS_INIT_LIST_HEAD(&node->li_group);
1920                         ldlm_interval_attach(node, lock);
1921                         node = NULL;
1922                 }
1923         }
1924
1925         /*
1926          * Remove old lock from the pool before adding the lock with new
1927          * mode below in ->policy()
1928          */
1929         ldlm_pool_del(&ns->ns_pool, lock);
1930
1931         /* If this is a local resource, put it on the appropriate list. */
1932         if (ns_is_client(ldlm_res_to_ns(res))) {
1933                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1934                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1935                 } else {
1936                         /* This should never happen, because of the way the
1937                          * server handles conversions. */
1938                         LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1939                                    *flags);
1940                         LBUG();
1941
1942                         ldlm_grant_lock(lock, &rpc_list);
1943                         granted = 1;
1944                         /* FIXME: completion handling not with lr_lock held ! */
1945                         if (lock->l_completion_ast)
1946                                 lock->l_completion_ast(lock, 0, NULL);
1947                 }
1948 #ifdef HAVE_SERVER_SUPPORT
1949         } else {
1950                 int rc;
1951                 ldlm_error_t err;
1952                 int pflags = 0;
1953                 ldlm_processing_policy policy;
1954                 policy = ldlm_processing_policy_table[res->lr_type];
1955                 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1956                 if (rc == LDLM_ITER_STOP) {
1957                         lock->l_req_mode = old_mode;
1958                         if (res->lr_type == LDLM_EXTENT)
1959                                 ldlm_extent_add_lock(res, lock);
1960                         else
1961                                 ldlm_granted_list_add_lock(lock, &prev);
1962
1963                         res = NULL;
1964                 } else {
1965                         *flags |= LDLM_FL_BLOCK_GRANTED;
1966                         granted = 1;
1967                 }
1968         }
1969 #else
1970         } else {
1971                 CERROR("This is client-side-only module, cannot handle "
1972                        "LDLM_NAMESPACE_SERVER resource type lock.\n");
1973                 LBUG();
1974         }
1975 #endif
1976         unlock_res_and_lock(lock);
1977
1978         if (granted)
1979                 ldlm_run_ast_work(ns, &rpc_list, LDLM_WORK_CP_AST);
1980         if (node)
1981                 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1982         RETURN(res);
1983 }
1984
1985 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1986 {
1987         struct obd_device *obd = NULL;
1988
1989         if (!((libcfs_debug | D_ERROR) & level))
1990                 return;
1991
1992         if (!lock) {
1993                 CDEBUG(level, "  NULL LDLM lock\n");
1994                 return;
1995         }
1996
1997         CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1998                lock, lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1999                pos, lock->l_pid);
2000         if (lock->l_conn_export != NULL)
2001                 obd = lock->l_conn_export->exp_obd;
2002         if (lock->l_export && lock->l_export->exp_connection) {
2003                 CDEBUG(level, "  Node: NID %s (rhandle: "LPX64")\n",
2004                      libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
2005                      lock->l_remote_handle.cookie);
2006         } else if (obd == NULL) {
2007                 CDEBUG(level, "  Node: local\n");
2008         } else {
2009                 struct obd_import *imp = obd->u.cli.cl_import;
2010                 CDEBUG(level, "  Node: NID %s (rhandle: "LPX64")\n",
2011                        libcfs_nid2str(imp->imp_connection->c_peer.nid),
2012                        lock->l_remote_handle.cookie);
2013         }
2014         CDEBUG(level, "  Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
2015                   lock->l_resource,
2016                   lock->l_resource->lr_name.name[0],
2017                   lock->l_resource->lr_name.name[1],
2018                   lock->l_resource->lr_name.name[2]);
2019         CDEBUG(level, "  Req mode: %s, grant mode: %s, rc: %u, read: %d, "
2020                "write: %d flags: "LPX64"\n", ldlm_lockname[lock->l_req_mode],
2021                ldlm_lockname[lock->l_granted_mode],
2022                cfs_atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
2023                lock->l_flags);
2024         if (lock->l_resource->lr_type == LDLM_EXTENT)
2025                 CDEBUG(level, "  Extent: "LPU64" -> "LPU64
2026                        " (req "LPU64"-"LPU64")\n",
2027                        lock->l_policy_data.l_extent.start,
2028                        lock->l_policy_data.l_extent.end,
2029                        lock->l_req_extent.start, lock->l_req_extent.end);
2030         else if (lock->l_resource->lr_type == LDLM_FLOCK)
2031                 CDEBUG(level, "  Pid: %d Extent: "LPU64" -> "LPU64"\n",
2032                        lock->l_policy_data.l_flock.pid,
2033                        lock->l_policy_data.l_flock.start,
2034                        lock->l_policy_data.l_flock.end);
2035        else if (lock->l_resource->lr_type == LDLM_IBITS)
2036                 CDEBUG(level, "  Bits: "LPX64"\n",
2037                        lock->l_policy_data.l_inodebits.bits);
2038 }
2039
2040 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
2041 {
2042         struct ldlm_lock *lock;
2043
2044         if (!((libcfs_debug | D_ERROR) & level))
2045                 return;
2046
2047         lock = ldlm_handle2lock(lockh);
2048         if (lock == NULL)
2049                 return;
2050
2051         ldlm_lock_dump(D_OTHER, lock, 0);
2052
2053         LDLM_LOCK_PUT(lock);
2054 }
2055
2056 void _ldlm_lock_debug(struct ldlm_lock *lock,
2057                       struct libcfs_debug_msg_data *msgdata,
2058                       const char *fmt, ...)
2059 {
2060         va_list args;
2061
2062         va_start(args, fmt);
2063
2064         if (lock->l_resource == NULL) {
2065                 libcfs_debug_vmsg2(msgdata, fmt, args,
2066                        " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2067                        "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" remote: "
2068                        LPX64" expref: %d pid: %u timeout: %lu\n", lock,
2069                        lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2070                        lock->l_readers, lock->l_writers,
2071                        ldlm_lockname[lock->l_granted_mode],
2072                        ldlm_lockname[lock->l_req_mode],
2073                        lock->l_flags, lock->l_remote_handle.cookie,
2074                        lock->l_export ?
2075                        cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2076                        lock->l_pid, lock->l_callback_timeout);
2077                 va_end(args);
2078                 return;
2079         }
2080
2081         switch (lock->l_resource->lr_type) {
2082         case LDLM_EXTENT:
2083                 libcfs_debug_vmsg2(msgdata, fmt, args,
2084                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2085                        "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
2086                        "] (req "LPU64"->"LPU64") flags: "LPX64" remote: "LPX64
2087                        " expref: %d pid: %u timeout %lu\n",
2088                        ldlm_lock_to_ns_name(lock), lock,
2089                        lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2090                        lock->l_readers, lock->l_writers,
2091                        ldlm_lockname[lock->l_granted_mode],
2092                        ldlm_lockname[lock->l_req_mode],
2093                        lock->l_resource->lr_name.name[0],
2094                        lock->l_resource->lr_name.name[1],
2095                        cfs_atomic_read(&lock->l_resource->lr_refcount),
2096                        ldlm_typename[lock->l_resource->lr_type],
2097                        lock->l_policy_data.l_extent.start,
2098                        lock->l_policy_data.l_extent.end,
2099                        lock->l_req_extent.start, lock->l_req_extent.end,
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         case LDLM_FLOCK:
2107                 libcfs_debug_vmsg2(msgdata, fmt, args,
2108                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2109                        "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
2110                        "["LPU64"->"LPU64"] flags: "LPX64" remote: "LPX64
2111                        " expref: %d pid: %u timeout: %lu\n",
2112                        ldlm_lock_to_ns_name(lock), lock,
2113                        lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2114                        lock->l_readers, lock->l_writers,
2115                        ldlm_lockname[lock->l_granted_mode],
2116                        ldlm_lockname[lock->l_req_mode],
2117                        lock->l_resource->lr_name.name[0],
2118                        lock->l_resource->lr_name.name[1],
2119                        cfs_atomic_read(&lock->l_resource->lr_refcount),
2120                        ldlm_typename[lock->l_resource->lr_type],
2121                        lock->l_policy_data.l_flock.pid,
2122                        lock->l_policy_data.l_flock.start,
2123                        lock->l_policy_data.l_flock.end,
2124                        lock->l_flags, lock->l_remote_handle.cookie,
2125                        lock->l_export ?
2126                        cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2127                        lock->l_pid, lock->l_callback_timeout);
2128                 break;
2129
2130         case LDLM_IBITS:
2131                 libcfs_debug_vmsg2(msgdata, fmt, args,
2132                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2133                        "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
2134                        "flags: "LPX64" remote: "LPX64" expref: %d "
2135                        "pid: %u timeout: %lu\n",
2136                        ldlm_lock_to_ns_name(lock),
2137                        lock, lock->l_handle.h_cookie,
2138                        cfs_atomic_read (&lock->l_refc),
2139                        lock->l_readers, lock->l_writers,
2140                        ldlm_lockname[lock->l_granted_mode],
2141                        ldlm_lockname[lock->l_req_mode],
2142                        lock->l_resource->lr_name.name[0],
2143                        lock->l_resource->lr_name.name[1],
2144                        lock->l_policy_data.l_inodebits.bits,
2145                        cfs_atomic_read(&lock->l_resource->lr_refcount),
2146                        ldlm_typename[lock->l_resource->lr_type],
2147                        lock->l_flags, lock->l_remote_handle.cookie,
2148                        lock->l_export ?
2149                        cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2150                        lock->l_pid, lock->l_callback_timeout);
2151                 break;
2152
2153         default:
2154                 libcfs_debug_vmsg2(msgdata, fmt, args,
2155                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2156                        "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" "
2157                        "remote: "LPX64" expref: %d pid: %u timeout %lu\n",
2158                        ldlm_lock_to_ns_name(lock),
2159                        lock, lock->l_handle.h_cookie,
2160                        cfs_atomic_read (&lock->l_refc),
2161                        lock->l_readers, lock->l_writers,
2162                        ldlm_lockname[lock->l_granted_mode],
2163                        ldlm_lockname[lock->l_req_mode],
2164                        lock->l_resource->lr_name.name[0],
2165                        lock->l_resource->lr_name.name[1],
2166                        cfs_atomic_read(&lock->l_resource->lr_refcount),
2167                        ldlm_typename[lock->l_resource->lr_type],
2168                        lock->l_flags, lock->l_remote_handle.cookie,
2169                        lock->l_export ?
2170                        cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2171                        lock->l_pid, lock->l_callback_timeout);
2172                 break;
2173         }
2174         va_end(args);
2175 }
2176 EXPORT_SYMBOL(_ldlm_lock_debug);