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