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