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  *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of the Lustre file system, http://www.lustre.org
   9  *   Lustre is a trademark of Cluster File Systems, Inc.
  10  *
  11  *   You may have signed or agreed to another license before downloading
  12  *   this software.  If so, you are bound by the terms and conditions
  13  *   of that agreement, and the following does not apply to you.  See the
  14  *   LICENSE file included with this distribution for more information.
  15  *
  16  *   If you did not agree to a different license, then this copy of Lustre
  17  *   is open source software; you can redistribute it and/or modify it
  18  *   under the terms of version 2 of the GNU General Public License as
  19  *   published by the Free Software Foundation.
  20  *
  21  *   In either case, Lustre is distributed in the hope that it will be
  22  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  23  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  24  *   license text for more details.
  25  */
  26
  27 #define DEBUG_SUBSYSTEM S_LDLM
  28
  29 #ifdef __KERNEL__
  30 # include <libcfs/libcfs.h>
  31 # include <linux/lustre_intent.h>
  32 #else
  33 # include <liblustre.h>
  34 # include <libcfs/kp30.h>
  35 #endif
  36
  37 #include <obd_class.h>
  38 #include "ldlm_internal.h"
  39
  40 //struct lustre_lock ldlm_everything_lock;
  41
  42 /* lock's skip list pointers fix mode */
  43 #define LDLM_JOIN_NONE          0
  44 #define LDLM_MODE_JOIN_RIGHT    1
  45 #define LDLM_MODE_JOIN_LEFT     (1 << 1)
  46 #define LDLM_POLICY_JOIN_RIGHT  (1 << 2)
  47 #define LDLM_POLICY_JOIN_LEFT   (1 << 3)
  48
  49 /* lock types */
  50 char *ldlm_lockname[] = {
  51         [0] "--",
  52         [LCK_EX] "EX",
  53         [LCK_PW] "PW",
  54         [LCK_PR] "PR",
  55         [LCK_CW] "CW",
  56         [LCK_CR] "CR",
  57         [LCK_NL] "NL",
  58         [LCK_GROUP] "GROUP"
  59 };
  60
  61 char *ldlm_typename[] = {
  62         [LDLM_PLAIN] "PLN",
  63         [LDLM_EXTENT] "EXT",
  64         [LDLM_FLOCK] "FLK",
  65         [LDLM_IBITS] "IBT",
  66 };
  67
  68 char *ldlm_it2str(int it)
  69 {
  70         switch (it) {
  71         case IT_OPEN:
  72                 return "open";
  73         case IT_CREAT:
  74                 return "creat";
  75         case (IT_OPEN | IT_CREAT):
  76                 return "open|creat";
  77         case IT_READDIR:
  78                 return "readdir";
  79         case IT_GETATTR:
  80                 return "getattr";
  81         case IT_LOOKUP:
  82                 return "lookup";
  83         case IT_UNLINK:
  84                 return "unlink";
  85         case IT_GETXATTR:
  86                 return "getxattr";
  87         default:
  88                 CERROR("Unknown intent %d\n", it);
  89                 return "UNKNOWN";
  90         }
  91 }
  92
  93 extern cfs_mem_cache_t *ldlm_lock_slab;
  94
  95 static ldlm_processing_policy ldlm_processing_policy_table[] = {
  96         [LDLM_PLAIN] ldlm_process_plain_lock,
  97         [LDLM_EXTENT] ldlm_process_extent_lock,
  98 #ifdef __KERNEL__
  99         [LDLM_FLOCK] ldlm_process_flock_lock,
 100 #endif
 101         [LDLM_IBITS] ldlm_process_inodebits_lock,
 102 };
 103
 104 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
 105 {
 106         return ldlm_processing_policy_table[res->lr_type];
 107 }
 108
 109 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
 110 {
 111         ns->ns_policy = arg;
 112 }
 113
 114 /*
 115  * REFCOUNTED LOCK OBJECTS
 116  */
 117
 118
 119 /*
 120  * Lock refcounts, during creation:
 121  *   - one special one for allocation, dec'd only once in destroy
 122  *   - one for being a lock that's in-use
 123  *   - one for the addref associated with a new lock
 124  */
 125 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 126 {
 127         atomic_inc(&lock->l_refc);
 128         return lock;
 129 }
 130
 131 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
 132 {
 133         LASSERT(size == sizeof(*lock));
 134         OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
 135 }
 136
 137 void ldlm_lock_put(struct ldlm_lock *lock)
 138 {
 139         ENTRY;
 140
 141         LASSERT(lock->l_resource != LP_POISON);
 142         LASSERT(atomic_read(&lock->l_refc) > 0);
 143         if (atomic_dec_and_test(&lock->l_refc)) {
 144                 struct ldlm_resource *res;
 145
 146                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing it.");
 147
 148                 res = lock->l_resource;
 149                 LASSERT(lock->l_destroyed);
 150                 LASSERT(list_empty(&lock->l_res_link));
 151                 LASSERT(list_empty(&lock->l_pending_chain));
 152
 153                 atomic_dec(&res->lr_namespace->ns_locks);
 154                 ldlm_resource_putref(res);
 155                 lock->l_resource = NULL;
 156                 if (lock->l_export) {
 157                         class_export_put(lock->l_export);
 158                         lock->l_export = NULL;
 159                 }
 160
 161                 if (lock->l_lvb_data != NULL)
 162                         OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
 163
 164                 ldlm_interval_free(ldlm_interval_detach(lock));
 165                 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
 166                                 ldlm_lock_free);
 167         }
 168
 169         EXIT;
 170 }
 171
 172 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
 173 {
 174         int rc = 0;
 175         if (!list_empty(&lock->l_lru)) {
 176                 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 177                 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
 178                 list_del_init(&lock->l_lru);
 179                 ns->ns_nr_unused--;
 180                 LASSERT(ns->ns_nr_unused >= 0);
 181                 rc = 1;
 182         }
 183         return rc;
 184 }
 185
 186 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
 187 {
 188         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 189         int rc;
 190         ENTRY;
 191         spin_lock(&ns->ns_unused_lock);
 192         rc = ldlm_lock_remove_from_lru_nolock(lock);
 193         spin_unlock(&ns->ns_unused_lock);
 194         EXIT;
 195         return rc;
 196 }
 197
 198 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
 199 {
 200         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 201         lock->l_last_used = cfs_time_current();
 202         LASSERT(list_empty(&lock->l_lru));
 203         list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 204         LASSERT(ns->ns_nr_unused >= 0);
 205         ns->ns_nr_unused++;
 206 }
 207
 208 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
 209 {
 210         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 211         ENTRY;
 212         spin_lock(&ns->ns_unused_lock);
 213         ldlm_lock_add_to_lru_nolock(lock);
 214         spin_unlock(&ns->ns_unused_lock);
 215         EXIT;
 216 }
 217
 218 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
 219 {
 220         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 221         ENTRY;
 222         spin_lock(&ns->ns_unused_lock);
 223         if (!list_empty(&lock->l_lru)) {
 224                 ldlm_lock_remove_from_lru_nolock(lock);
 225                 ldlm_lock_add_to_lru_nolock(lock);
 226         }
 227         spin_unlock(&ns->ns_unused_lock);
 228         EXIT;
 229 }
 230
 231 /* This used to have a 'strict' flag, which recovery would use to mark an
 232  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
 233  * shall explain why it's gone: with the new hash table scheme, once you call
 234  * ldlm_lock_destroy, you can never drop your final references on this lock.
 235  * Because it's not in the hash table anymore.  -phil */
 236 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
 237 {
 238         ENTRY;
 239
 240         if (lock->l_readers || lock->l_writers) {
 241                 LDLM_ERROR(lock, "lock still has references");
 242                 ldlm_lock_dump(D_ERROR, lock, 0);
 243                 LBUG();
 244         }
 245
 246         if (!list_empty(&lock->l_res_link)) {
 247                 LDLM_ERROR(lock, "lock still on resource");
 248                 ldlm_lock_dump(D_ERROR, lock, 0);
 249                 LBUG();
 250         }
 251
 252         if (lock->l_destroyed) {
 253                 LASSERT(list_empty(&lock->l_lru));
 254                 EXIT;
 255                 return 0;
 256         }
 257         lock->l_destroyed = 1;
 258
 259         if (lock->l_export)
 260                 spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
 261         list_del_init(&lock->l_export_chain);
 262         if (lock->l_export)
 263                 spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
 264
 265         ldlm_lock_remove_from_lru(lock);
 266         class_handle_unhash(&lock->l_handle);
 267
 268 #if 0
 269         /* Wake anyone waiting for this lock */
 270         /* FIXME: I should probably add yet another flag, instead of using
 271          * l_export to only call this on clients */
 272         if (lock->l_export)
 273                 class_export_put(lock->l_export);
 274         lock->l_export = NULL;
 275         if (lock->l_export && lock->l_completion_ast)
 276                 lock->l_completion_ast(lock, 0);
 277 #endif
 278         EXIT;
 279         return 1;
 280 }
 281
 282 void ldlm_lock_destroy(struct ldlm_lock *lock)
 283 {
 284         int first;
 285         ENTRY;
 286         lock_res_and_lock(lock);
 287         first = ldlm_lock_destroy_internal(lock);
 288         unlock_res_and_lock(lock);
 289
 290         /* drop reference from hashtable only for first destroy */
 291         if (first)
 292                 LDLM_LOCK_PUT(lock);
 293         EXIT;
 294 }
 295
 296 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
 297 {
 298         int first;
 299         ENTRY;
 300         first = ldlm_lock_destroy_internal(lock);
 301         /* drop reference from hashtable only for first destroy */
 302         if (first)
 303                 LDLM_LOCK_PUT(lock);
 304         EXIT;
 305 }
 306
 307 /* this is called by portals_handle2object with the handle lock taken */
 308 static void lock_handle_addref(void *lock)
 309 {
 310         LDLM_LOCK_GET((struct ldlm_lock *)lock);
 311 }
 312
 313 /*
 314  * usage: pass in a resource on which you have done ldlm_resource_get
 315  *        pass in a parent lock on which you have done a ldlm_lock_get
 316  *        after return, ldlm_*_put the resource and parent
 317  * returns: lock with refcount 2 - one for current caller and one for remote
 318  */
 319 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
 320 {
 321         struct ldlm_lock *lock;
 322         ENTRY;
 323
 324         if (resource == NULL)
 325                 LBUG();
 326
 327         OBD_SLAB_ALLOC(lock, ldlm_lock_slab, CFS_ALLOC_IO, sizeof(*lock));
 328         if (lock == NULL)
 329                 RETURN(NULL);
 330
 331         spin_lock_init(&lock->l_lock);
 332         lock->l_resource = ldlm_resource_getref(resource);
 333
 334         atomic_set(&lock->l_refc, 2);
 335         CFS_INIT_LIST_HEAD(&lock->l_res_link);
 336         CFS_INIT_LIST_HEAD(&lock->l_lru);
 337         CFS_INIT_LIST_HEAD(&lock->l_export_chain);
 338         CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
 339         CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
 340         CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
 341         cfs_waitq_init(&lock->l_waitq);
 342         lock->l_blocking_lock = NULL;
 343         lock->l_sl_mode.prev = NULL;
 344         lock->l_sl_mode.next = NULL;
 345         lock->l_sl_policy.prev = NULL;
 346         lock->l_sl_policy.next = NULL;
 347
 348         atomic_inc(&resource->lr_namespace->ns_locks);
 349         CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
 350         class_handle_hash(&lock->l_handle, lock_handle_addref);
 351
 352         CFS_INIT_LIST_HEAD(&lock->l_extents_list);
 353         spin_lock_init(&lock->l_extents_list_lock);
 354         CFS_INIT_LIST_HEAD(&lock->l_cache_locks_list);
 355
 356         RETURN(lock);
 357 }
 358
 359 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
 360                               struct ldlm_res_id new_resid)
 361 {
 362         struct ldlm_resource *oldres = lock->l_resource;
 363         struct ldlm_resource *newres;
 364         int type;
 365         ENTRY;
 366
 367         LASSERT(ns_is_client(ns));
 368
 369         lock_res_and_lock(lock);
 370         if (memcmp(&new_resid, &lock->l_resource->lr_name,
 371                    sizeof(lock->l_resource->lr_name)) == 0) {
 372                 /* Nothing to do */
 373                 unlock_res_and_lock(lock);
 374                 RETURN(0);
 375         }
 376
 377         LASSERT(new_resid.name[0] != 0);
 378
 379         /* This function assumes that the lock isn't on any lists */
 380         LASSERT(list_empty(&lock->l_res_link));
 381
 382         type = oldres->lr_type;
 383         unlock_res_and_lock(lock);
 384
 385         newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
 386         if (newres == NULL) {
 387                 LBUG();
 388                 RETURN(-ENOMEM);
 389         }
 390
 391         lock_res_and_lock(lock);
 392         LASSERT(memcmp(&new_resid, &lock->l_resource->lr_name,
 393                        sizeof(lock->l_resource->lr_name)) != 0);
 394         lock_res(newres);
 395         lock->l_resource = newres;
 396         unlock_res(oldres);
 397         unlock_res_and_lock(lock);
 398
 399         /* ...and the flowers are still standing! */
 400         ldlm_resource_putref(oldres);
 401
 402         RETURN(0);
 403 }
 404
 405 /*
 406  *  HANDLES
 407  */
 408
 409 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 410 {
 411         lockh->cookie = lock->l_handle.h_cookie;
 412 }
 413
 414 /* if flags: atomically get the lock and set the flags.
 415  *           Return NULL if flag already set
 416  */
 417
 418 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
 419 {
 420         struct ldlm_namespace *ns;
 421         struct ldlm_lock *lock = NULL, *retval = NULL;
 422         ENTRY;
 423
 424         LASSERT(handle);
 425
 426         lock = class_handle2object(handle->cookie);
 427         if (lock == NULL)
 428                 RETURN(NULL);
 429
 430         LASSERT(lock->l_resource != NULL);
 431         ns = lock->l_resource->lr_namespace;
 432         LASSERT(ns != NULL);
 433
 434         lock_res_and_lock(lock);
 435
 436         /* It's unlikely but possible that someone marked the lock as
 437          * destroyed after we did handle2object on it */
 438         if (lock->l_destroyed) {
 439                 unlock_res_and_lock(lock);
 440                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
 441                 LDLM_LOCK_PUT(lock);
 442                 GOTO(out, retval);
 443         }
 444
 445         if (flags && (lock->l_flags & flags)) {
 446                 unlock_res_and_lock(lock);
 447                 LDLM_LOCK_PUT(lock);
 448                 GOTO(out, retval);
 449         }
 450
 451         if (flags)
 452                 lock->l_flags |= flags;
 453
 454         unlock_res_and_lock(lock);
 455         retval = lock;
 456         EXIT;
 457  out:
 458         return retval;
 459 }
 460
 461 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
 462                                       struct lustre_handle *handle)
 463 {
 464         struct ldlm_lock *retval = NULL;
 465         retval = __ldlm_handle2lock(handle, 0);
 466         return retval;
 467 }
 468
 469 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 470 {
 471         struct obd_export *exp = lock->l_export?:lock->l_conn_export;
 472         /* INODEBITS_INTEROP: If the other side does not support
 473          * inodebits, reply with a plain lock descriptor.
 474          */
 475         if ((lock->l_resource->lr_type == LDLM_IBITS) &&
 476             (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
 477                 struct ldlm_resource res = *lock->l_resource;
 478
 479                 /* Make sure all the right bits are set in this lock we
 480                    are going to pass to client */
 481                 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
 482                          (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
 483                          "Inappropriate inode lock bits during "
 484                          "conversion " LPU64 "\n",
 485                          lock->l_policy_data.l_inodebits.bits);
 486                 res.lr_type = LDLM_PLAIN;
 487                 ldlm_res2desc(&res, &desc->l_resource);
 488                 /* Convert "new" lock mode to something old client can
 489                    understand */
 490                 if ((lock->l_req_mode == LCK_CR) ||
 491                     (lock->l_req_mode == LCK_CW))
 492                         desc->l_req_mode = LCK_PR;
 493                 else
 494                         desc->l_req_mode = lock->l_req_mode;
 495                 if ((lock->l_granted_mode == LCK_CR) ||
 496                     (lock->l_granted_mode == LCK_CW)) {
 497                         desc->l_granted_mode = LCK_PR;
 498                 } else {
 499                         /* We never grant PW/EX locks to clients */
 500                         LASSERT((lock->l_granted_mode != LCK_PW) &&
 501                                 (lock->l_granted_mode != LCK_EX));
 502                         desc->l_granted_mode = lock->l_granted_mode;
 503                 }
 504
 505                 /* We do not copy policy here, because there is no
 506                    policy for plain locks */
 507         } else {
 508                 ldlm_res2desc(lock->l_resource, &desc->l_resource);
 509                 desc->l_req_mode = lock->l_req_mode;
 510                 desc->l_granted_mode = lock->l_granted_mode;
 511                 desc->l_policy_data = lock->l_policy_data;
 512         }
 513 }
 514
 515 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
 516                            struct list_head *work_list)
 517 {
 518         if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
 519                 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
 520                 lock->l_flags |= LDLM_FL_AST_SENT;
 521                 /* If the enqueuing client said so, tell the AST recipient to
 522                  * discard dirty data, rather than writing back. */
 523                 if (new->l_flags & LDLM_AST_DISCARD_DATA)
 524                         lock->l_flags |= LDLM_FL_DISCARD_DATA;
 525                 LASSERT(list_empty(&lock->l_bl_ast));
 526                 list_add(&lock->l_bl_ast, work_list);
 527                 LDLM_LOCK_GET(lock);
 528                 LASSERT(lock->l_blocking_lock == NULL);
 529                 lock->l_blocking_lock = LDLM_LOCK_GET(new);
 530         }
 531 }
 532
 533 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
 534 {
 535         if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
 536                 lock->l_flags |= LDLM_FL_CP_REQD;
 537                 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
 538                 LASSERT(list_empty(&lock->l_cp_ast));
 539                 list_add(&lock->l_cp_ast, work_list);
 540                 LDLM_LOCK_GET(lock);
 541         }
 542 }
 543
 544 /* must be called with lr_lock held */
 545 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
 546                                 struct list_head *work_list)
 547 {
 548         ENTRY;
 549         check_res_locked(lock->l_resource);
 550         if (new)
 551                 ldlm_add_bl_work_item(lock, new, work_list);
 552         else
 553                 ldlm_add_cp_work_item(lock, work_list);
 554         EXIT;
 555 }
 556
 557 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 558 {
 559         struct ldlm_lock *lock;
 560
 561         lock = ldlm_handle2lock(lockh);
 562         LASSERT(lock != NULL);
 563         ldlm_lock_addref_internal(lock, mode);
 564         LDLM_LOCK_PUT(lock);
 565 }
 566
 567 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
 568 {
 569         ldlm_lock_remove_from_lru(lock);
 570         if (mode & (LCK_NL | LCK_CR | LCK_PR))
 571                 lock->l_readers++;
 572         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
 573                 lock->l_writers++;
 574         LDLM_LOCK_GET(lock);
 575         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 576 }
 577
 578 /* only called for local locks */
 579 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 580 {
 581         lock_res_and_lock(lock);
 582         ldlm_lock_addref_internal_nolock(lock, mode);
 583         unlock_res_and_lock(lock);
 584 }
 585
 586 /* only called in ldlm_flock_destroy and for local locks.
 587  * for LDLM_FLOCK type locks, l_blocking_ast is null, and
 588  * ldlm_lock_remove_from_lru() does nothing, it is safe
 589  * for ldlm_flock_destroy usage by dropping some code */
 590 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
 591 {
 592         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 593         if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
 594                 LASSERT(lock->l_readers > 0);
 595                 lock->l_readers--;
 596         }
 597         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
 598                 LASSERT(lock->l_writers > 0);
 599                 lock->l_writers--;
 600         }
 601
 602         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 603 }
 604
 605 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
 606 {
 607         struct ldlm_namespace *ns;
 608         ENTRY;
 609
 610         lock_res_and_lock(lock);
 611
 612         ns = lock->l_resource->lr_namespace;
 613
 614         ldlm_lock_decref_internal_nolock(lock, mode);
 615
 616         if (lock->l_flags & LDLM_FL_LOCAL &&
 617             !lock->l_readers && !lock->l_writers) {
 618                 /* If this is a local lock on a server namespace and this was
 619                  * the last reference, cancel the lock. */
 620                 CDEBUG(D_INFO, "forcing cancel of local lock\n");
 621                 lock->l_flags |= LDLM_FL_CBPENDING;
 622         }
 623
 624         if (!lock->l_readers && !lock->l_writers &&
 625             (lock->l_flags & LDLM_FL_CBPENDING)) {
 626                 /* If we received a blocked AST and this was the last reference,
 627                  * run the callback. */
 628                 if (ns_is_server(ns) && lock->l_export)
 629                         CERROR("FL_CBPENDING set on non-local lock--just a "
 630                                "warning\n");
 631
 632                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 633
 634                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
 635                 ldlm_lock_remove_from_lru(lock);
 636                 unlock_res_and_lock(lock);
 637                 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
 638                     ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
 639                         ldlm_handle_bl_callback(ns, NULL, lock);
 640         } else if (ns_is_client(ns) &&
 641                    !lock->l_readers && !lock->l_writers &&
 642                    !(lock->l_flags & LDLM_FL_NO_LRU)) {
 643                 /* If this is a client-side namespace and this was the last
 644                  * reference, put it on the LRU. */
 645                 ldlm_lock_add_to_lru(lock);
 646                 unlock_res_and_lock(lock);
 647                 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
 648                  * are not supported by the server, otherwise, it is done on
 649                  * enqueue. */
 650                 if (!exp_connect_cancelset(lock->l_conn_export) &&
 651                     !ns_connect_lru_resize(ns))
 652                         ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
 653         } else {
 654                 unlock_res_and_lock(lock);
 655         }
 656
 657         EXIT;
 658 }
 659
 660 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 661 {
 662         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 663         LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
 664         ldlm_lock_decref_internal(lock, mode);
 665         LDLM_LOCK_PUT(lock);
 666 }
 667
 668 /* This will drop a lock reference and mark it for destruction, but will not
 669  * necessarily cancel the lock before returning. */
 670 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
 671 {
 672         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 673         ENTRY;
 674
 675         LASSERT(lock != NULL);
 676
 677         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 678         lock_res_and_lock(lock);
 679         lock->l_flags |= LDLM_FL_CBPENDING;
 680         unlock_res_and_lock(lock);
 681         ldlm_lock_decref_internal(lock, mode);
 682         LDLM_LOCK_PUT(lock);
 683 }
 684
 685 /*
 686  * search_granted_lock
 687  *
 688  * Description:
 689  *      Finds a position to insert the new lock.
 690  * Parameters:
 691  *      queue [input]:  the granted list where search acts on;
 692  *      req [input]:    the lock whose position to be located;
 693  *      lockp [output]: the position where the lock should be inserted before, or
 694  *                      NULL indicating @req should be appended to @queue.
 695  * Return Values:
 696  *      Bit-masks combination of following values indicating in which way the
 697  *      lock need to be inserted.
 698  *      - LDLM_JOIN_NONE:       noting about skip list needs to be fixed;
 699  *      - LDLM_MODE_JOIN_RIGHT: @req needs join right becoming the head of a
 700  *                              mode group;
 701  *      - LDLM_POLICY_JOIN_RIGHT: @req needs join right becoming the head of
 702  *                                a policy group.
 703  * NOTE: called by
 704  *  - ldlm_grant_lock_with_skiplist
 705  */
 706 static int search_granted_lock(struct list_head *queue,
 707                         struct ldlm_lock *req,
 708                         struct ldlm_lock **lockp)
 709 {
 710         struct list_head *tmp, *tmp_tail;
 711         struct ldlm_lock *lock, *mode_head_lock;
 712         int rc = LDLM_JOIN_NONE;
 713         ENTRY;
 714
 715         list_for_each(tmp, queue) {
 716                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 717
 718                 if (lock->l_req_mode != req->l_req_mode) {
 719                         if (LDLM_SL_HEAD(&lock->l_sl_mode))
 720                                 tmp = &list_entry(lock->l_sl_mode.next,
 721                                                   struct ldlm_lock,
 722                                                   l_sl_mode)->l_res_link;
 723                         continue;
 724                 }
 725
 726                 /* found the same mode group */
 727                 if (lock->l_resource->lr_type == LDLM_PLAIN) {
 728                         *lockp = lock;
 729                         rc = LDLM_MODE_JOIN_RIGHT;
 730                         GOTO(out, rc);
 731                 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
 732                         tmp_tail = tmp;
 733                         if (LDLM_SL_HEAD(&lock->l_sl_mode))
 734                                 tmp_tail = &list_entry(lock->l_sl_mode.next,
 735                                                        struct ldlm_lock,
 736                                                        l_sl_mode)->l_res_link;
 737                         mode_head_lock = lock;
 738                         for (;;) {
 739                                 if (lock->l_policy_data.l_inodebits.bits ==
 740                                     req->l_policy_data.l_inodebits.bits) {
 741                                         /* matched policy lock is found */
 742                                         *lockp = lock;
 743                                         rc |= LDLM_POLICY_JOIN_RIGHT;
 744
 745                                         /* if the policy group head is also a
 746                                          * mode group head or a single mode
 747                                          * group lock */
 748                                         if (LDLM_SL_HEAD(&lock->l_sl_mode) ||
 749                                             (tmp == tmp_tail &&
 750                                              LDLM_SL_EMPTY(&lock->l_sl_mode)))
 751                                                 rc |= LDLM_MODE_JOIN_RIGHT;
 752                                         GOTO(out, rc);
 753                                 }
 754
 755                                 if (LDLM_SL_HEAD(&lock->l_sl_policy))
 756                                         tmp = &list_entry(lock->l_sl_policy.next,
 757                                                           struct ldlm_lock,
 758                                                           l_sl_policy)->l_res_link;
 759
 760                                 if (tmp == tmp_tail)
 761                                         break;
 762                                 else
 763                                         tmp = tmp->next;
 764                                 lock = list_entry(tmp, struct ldlm_lock,
 765                                                   l_res_link);
 766                         }  /* for all locks in the matched mode group */
 767
 768                         /* no matched policy group is found, insert before
 769                          * the mode group head lock */
 770                         *lockp = mode_head_lock;
 771                         rc = LDLM_MODE_JOIN_RIGHT;
 772                         GOTO(out, rc);
 773                 } else {
 774                         LDLM_ERROR(lock, "is not LDLM_PLAIN or LDLM_IBITS lock");
 775                         LBUG();
 776                 }
 777         }
 778
 779         /* no matched mode group is found, append to the end */
 780         *lockp = NULL;
 781         rc = LDLM_JOIN_NONE;
 782         EXIT;
 783 out:
 784         return rc;
 785 }
 786
 787 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
 788                                        struct ldlm_lock *lockp,
 789                                        int join)
 790 {
 791         struct ldlm_resource *res = lock->l_resource;
 792         ENTRY;
 793
 794         LASSERT(lockp || join == LDLM_JOIN_NONE);
 795
 796         check_res_locked(res);
 797
 798         ldlm_resource_dump(D_OTHER, res);
 799         CDEBUG(D_OTHER, "About to add this lock:\n");
 800         ldlm_lock_dump(D_OTHER, lock, 0);
 801
 802         if (lock->l_destroyed) {
 803                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 804                 return;
 805         }
 806
 807         LASSERT(list_empty(&lock->l_res_link));
 808
 809         if (!lockp)
 810                 list_add_tail(&lock->l_res_link, &lock->l_resource->lr_granted);
 811         else if ((join & LDLM_MODE_JOIN_LEFT) || (join & LDLM_POLICY_JOIN_LEFT))
 812                 list_add(&lock->l_res_link, &lockp->l_res_link);
 813         else
 814                 list_add_tail(&lock->l_res_link, &lockp->l_res_link);
 815
 816         /* fix skip lists */
 817         if (join & LDLM_MODE_JOIN_RIGHT) {
 818                 LASSERT(! LDLM_SL_TAIL(&lockp->l_sl_mode));
 819                 if (LDLM_SL_EMPTY(&lockp->l_sl_mode)) {
 820                         lock->l_sl_mode.next = &lockp->l_sl_mode;
 821                         lockp->l_sl_mode.prev = &lock->l_sl_mode;
 822                 } else if (LDLM_SL_HEAD(&lockp->l_sl_mode)) {
 823                         lock->l_sl_mode.next = lockp->l_sl_mode.next;
 824                         lockp->l_sl_mode.next = NULL;
 825                         lock->l_sl_mode.next->prev = &lock->l_sl_mode;
 826                 }
 827         } else if (join & LDLM_MODE_JOIN_LEFT) {
 828                 LASSERT(! LDLM_SL_HEAD(&lockp->l_sl_mode));
 829                 if (LDLM_SL_EMPTY(&lockp->l_sl_mode)) {
 830                         lock->l_sl_mode.prev = &lockp->l_sl_mode;
 831                         lockp->l_sl_mode.next = &lock->l_sl_mode;
 832                 } else if (LDLM_SL_TAIL(&lockp->l_sl_mode)) {
 833                         lock->l_sl_mode.prev = lockp->l_sl_mode.prev;
 834                         lockp->l_sl_mode.prev = NULL;
 835                         lock->l_sl_mode.prev->next = &lock->l_sl_mode;
 836                 }
 837         }
 838
 839         if (join & LDLM_POLICY_JOIN_RIGHT) {
 840                 LASSERT(! LDLM_SL_TAIL(&lockp->l_sl_policy));
 841                 if (LDLM_SL_EMPTY(&lockp->l_sl_policy)) {
 842                         lock->l_sl_policy.next = &lockp->l_sl_policy;
 843                         lockp->l_sl_policy.prev = &lock->l_sl_policy;
 844                 } else if (LDLM_SL_HEAD(&lockp->l_sl_policy)) {
 845                         lock->l_sl_policy.next = lockp->l_sl_policy.next;
 846                         lockp->l_sl_policy.next = NULL;
 847                         lock->l_sl_policy.next->prev = &lock->l_sl_policy;
 848                 }
 849         } else if (join & LDLM_POLICY_JOIN_LEFT) {
 850                 LASSERT(! LDLM_SL_HEAD(&lockp->l_sl_policy));
 851                 if (LDLM_SL_EMPTY(&lockp->l_sl_policy)) {
 852                         lock->l_sl_policy.prev = &lockp->l_sl_policy;
 853                         lockp->l_sl_policy.next = &lock->l_sl_policy;
 854                 } else if (LDLM_SL_TAIL(&lockp->l_sl_policy)) {
 855                         lock->l_sl_policy.prev = lockp->l_sl_policy.prev;
 856                         lockp->l_sl_policy.prev = NULL;
 857                         lock->l_sl_policy.prev->next = &lock->l_sl_policy;
 858                 }
 859         }
 860
 861         EXIT;
 862 }
 863
 864 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
 865 {
 866         int join = LDLM_JOIN_NONE;
 867         struct ldlm_lock *lockp = NULL;
 868         ENTRY;
 869
 870         LASSERT(lock->l_req_mode == lock->l_granted_mode);
 871
 872         join = search_granted_lock(&lock->l_resource->lr_granted, lock, &lockp);
 873         ldlm_granted_list_add_lock(lock, lockp, join);
 874         EXIT;
 875 }
 876
 877 /* NOTE: called by
 878  *  - ldlm_lock_enqueue
 879  *  - ldlm_reprocess_queue
 880  *  - ldlm_lock_convert
 881  *
 882  * must be called with lr_lock held
 883  */
 884 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
 885 {
 886         struct ldlm_resource *res = lock->l_resource;
 887         ENTRY;
 888
 889         check_res_locked(res);
 890
 891         lock->l_granted_mode = lock->l_req_mode;
 892         if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
 893                 ldlm_grant_lock_with_skiplist(lock);
 894         else if (res->lr_type == LDLM_EXTENT)
 895                 ldlm_extent_add_lock(res, lock);
 896         else
 897                 ldlm_resource_add_lock(res, &res->lr_granted, lock);
 898
 899         if (lock->l_granted_mode < res->lr_most_restr)
 900                 res->lr_most_restr = lock->l_granted_mode;
 901
 902         if (work_list && lock->l_completion_ast != NULL)
 903                 ldlm_add_ast_work_item(lock, NULL, work_list);
 904
 905         ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
 906         EXIT;
 907 }
 908
 909 /* returns a referenced lock or NULL.  See the flag descriptions below, in the
 910  * comment above ldlm_lock_match */
 911 static struct ldlm_lock *search_queue(struct list_head *queue,
 912                                       ldlm_mode_t *mode,
 913                                       ldlm_policy_data_t *policy,
 914                                       struct ldlm_lock *old_lock, int flags)
 915 {
 916         struct ldlm_lock *lock;
 917         struct list_head *tmp;
 918
 919         list_for_each(tmp, queue) {
 920                 ldlm_mode_t match;
 921
 922                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 923
 924                 if (lock == old_lock)
 925                         continue;
 926
 927                 /* llite sometimes wants to match locks that will be
 928                  * canceled when their users drop, but we allow it to match
 929                  * if it passes in CBPENDING and the lock still has users.
 930                  * this is generally only going to be used by children
 931                  * whose parents already hold a lock so forward progress
 932                  * can still happen. */
 933                 if (lock->l_flags & LDLM_FL_CBPENDING &&
 934                     !(flags & LDLM_FL_CBPENDING))
 935                         continue;
 936                 if (lock->l_flags & LDLM_FL_CBPENDING &&
 937                     lock->l_readers == 0 && lock->l_writers == 0)
 938                         continue;
 939
 940                 if (!(lock->l_req_mode & *mode))
 941                         continue;
 942
 943                 match = lock->l_req_mode;
 944                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 945                     (lock->l_policy_data.l_extent.start >
 946                      policy->l_extent.start ||
 947                      lock->l_policy_data.l_extent.end < policy->l_extent.end))
 948                         continue;
 949
 950                 if (unlikely(match == LCK_GROUP) &&
 951                     lock->l_resource->lr_type == LDLM_EXTENT &&
 952                     lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
 953                         continue;
 954
 955                 /* We match if we have existing lock with same or wider set
 956                    of bits. */
 957                 if (lock->l_resource->lr_type == LDLM_IBITS &&
 958                      ((lock->l_policy_data.l_inodebits.bits &
 959                       policy->l_inodebits.bits) !=
 960                       policy->l_inodebits.bits))
 961                         continue;
 962
 963                 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
 964                         continue;
 965
 966                 if ((flags & LDLM_FL_LOCAL_ONLY) &&
 967                     !(lock->l_flags & LDLM_FL_LOCAL))
 968                         continue;
 969
 970                 if (flags & LDLM_FL_TEST_LOCK) {
 971                         LDLM_LOCK_GET(lock);
 972                         ldlm_lock_touch_in_lru(lock);
 973                 } else {
 974                         ldlm_lock_addref_internal_nolock(lock, match);
 975                 }
 976                 *mode = match;
 977                 return lock;
 978         }
 979
 980         return NULL;
 981 }
 982
 983 void ldlm_lock_allow_match(struct ldlm_lock *lock)
 984 {
 985         lock_res_and_lock(lock);
 986         lock->l_flags |= LDLM_FL_LVB_READY;
 987         cfs_waitq_signal(&lock->l_waitq);
 988         unlock_res_and_lock(lock);
 989 }
 990
 991 /* Can be called in two ways:
 992  *
 993  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
 994  * for a duplicate of.
 995  *
 996  * Otherwise, all of the fields must be filled in, to match against.
 997  *
 998  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
 999  *     server (ie, connh is NULL)
1000  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1001  *     list will be considered
1002  * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1003  *     to be canceled can still be matched as long as they still have reader
1004  *     or writer refernces
1005  * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1006  *     just tell us if we would have matched.
1007  *
1008  * Returns 1 if it finds an already-existing lock that is compatible; in this
1009  * case, lockh is filled in with a addref()ed lock
1010  */
1011 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1012                             struct ldlm_res_id *res_id, ldlm_type_t type,
1013                             ldlm_policy_data_t *policy, ldlm_mode_t mode,
1014                             struct lustre_handle *lockh)
1015 {
1016         struct ldlm_resource *res;
1017         struct ldlm_lock *lock, *old_lock = NULL;
1018         int rc = 0;
1019         ENTRY;
1020
1021         if (ns == NULL) {
1022                 old_lock = ldlm_handle2lock(lockh);
1023                 LASSERT(old_lock);
1024
1025                 ns = old_lock->l_resource->lr_namespace;
1026                 res_id = &old_lock->l_resource->lr_name;
1027                 type = old_lock->l_resource->lr_type;
1028                 mode = old_lock->l_req_mode;
1029         }
1030
1031         res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
1032         if (res == NULL) {
1033                 LASSERT(old_lock == NULL);
1034                 RETURN(0);
1035         }
1036
1037         lock_res(res);
1038
1039         lock = search_queue(&res->lr_granted, &mode, policy, old_lock, flags);
1040         if (lock != NULL)
1041                 GOTO(out, rc = 1);
1042         if (flags & LDLM_FL_BLOCK_GRANTED)
1043                 GOTO(out, rc = 0);
1044         lock = search_queue(&res->lr_converting, &mode, policy, old_lock, flags);
1045         if (lock != NULL)
1046                 GOTO(out, rc = 1);
1047         lock = search_queue(&res->lr_waiting, &mode, policy, old_lock, flags);
1048         if (lock != NULL)
1049                 GOTO(out, rc = 1);
1050
1051         EXIT;
1052  out:
1053         unlock_res(res);
1054         ldlm_resource_putref(res);
1055
1056         if (lock) {
1057                 ldlm_lock2handle(lock, lockh);
1058                 if ((flags & LDLM_FL_LVB_READY) && (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1059                         struct l_wait_info lwi;
1060                         if (lock->l_completion_ast) {
1061                                 int err = lock->l_completion_ast(lock,
1062                                                           LDLM_FL_WAIT_NOREPROC,
1063                                                                  NULL);
1064                                 if (err) {
1065                                         if (flags & LDLM_FL_TEST_LOCK)
1066                                                 LDLM_LOCK_PUT(lock);
1067                                         else
1068                                                 ldlm_lock_decref_internal(lock, mode);
1069                                         rc = 0;
1070                                         goto out2;
1071                                 }
1072                         }
1073
1074                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), NULL,
1075                                                LWI_ON_SIGNAL_NOOP, NULL);
1076
1077                         /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1078                         l_wait_event(lock->l_waitq,
1079                                      (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1080                 }
1081         }
1082  out2:
1083         if (rc) {
1084                 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1085                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1086                                 res_id->name[2] : policy->l_extent.start,
1087                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1088                                 res_id->name[3] : policy->l_extent.end);
1089         } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1090                 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1091                                   LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1092                                   type, mode, res_id->name[0], res_id->name[1],
1093                                   (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1094                                         res_id->name[2] :policy->l_extent.start,
1095                                 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1096                                         res_id->name[3] : policy->l_extent.end);
1097         }
1098         if (old_lock)
1099                 LDLM_LOCK_PUT(old_lock);
1100         if (flags & LDLM_FL_TEST_LOCK && rc)
1101                 LDLM_LOCK_PUT(lock);
1102
1103         return rc ? mode : 0;
1104 }
1105
1106 /* Returns a referenced lock */
1107 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1108                                    struct ldlm_res_id res_id, ldlm_type_t type,
1109                                    ldlm_mode_t mode,
1110                                    ldlm_blocking_callback blocking,
1111                                    ldlm_completion_callback completion,
1112                                    ldlm_glimpse_callback glimpse,
1113                                    void *data, __u32 lvb_len)
1114 {
1115         struct ldlm_lock *lock;
1116         struct ldlm_resource *res;
1117         ENTRY;
1118
1119         res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1120         if (res == NULL)
1121                 RETURN(NULL);
1122
1123         lock = ldlm_lock_new(res);
1124         ldlm_resource_putref(res);
1125
1126         if (lock == NULL)
1127                 RETURN(NULL);
1128
1129         lock->l_req_mode = mode;
1130         lock->l_ast_data = data;
1131         lock->l_blocking_ast = blocking;
1132         lock->l_completion_ast = completion;
1133         lock->l_glimpse_ast = glimpse;
1134         lock->l_pid = cfs_curproc_pid();
1135
1136         lock->l_tree_node = NULL;
1137         /* if this is the extent lock, allocate the interval tree node */
1138         if (type == LDLM_EXTENT) {
1139                 if (ldlm_interval_alloc(lock) == NULL)
1140                         GOTO(out, 0);
1141         }
1142
1143         if (lvb_len) {
1144                 lock->l_lvb_len = lvb_len;
1145                 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1146                 if (lock->l_lvb_data == NULL)
1147                         GOTO(out, 0);
1148         }
1149
1150         RETURN(lock);
1151
1152 out:
1153         if (lock->l_lvb_data)
1154                 OBD_FREE(lock->l_lvb_data, lvb_len);
1155         ldlm_interval_free(ldlm_interval_detach(lock));
1156         OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1157         return NULL;
1158 }
1159
1160 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1161                                struct ldlm_lock **lockp,
1162                                void *cookie, int *flags)
1163 {
1164         struct ldlm_lock *lock = *lockp;
1165         struct ldlm_resource *res = lock->l_resource;
1166         int local = ns_is_client(res->lr_namespace);
1167         ldlm_processing_policy policy;
1168         ldlm_error_t rc = ELDLM_OK;
1169         ENTRY;
1170
1171         do_gettimeofday(&lock->l_enqueued_time);
1172         /* policies are not executed on the client or during replay */
1173         if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1174             && !local && ns->ns_policy) {
1175                 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1176                                    NULL);
1177                 if (rc == ELDLM_LOCK_REPLACED) {
1178                         /* The lock that was returned has already been granted,
1179                          * and placed into lockp.  If it's not the same as the
1180                          * one we passed in, then destroy the old one and our
1181                          * work here is done. */
1182                         if (lock != *lockp) {
1183                                 ldlm_lock_destroy(lock);
1184                                 LDLM_LOCK_PUT(lock);
1185                         }
1186                         *flags |= LDLM_FL_LOCK_CHANGED;
1187                         RETURN(0);
1188                 } else if (rc != ELDLM_OK ||
1189                            (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1190                         ldlm_lock_destroy(lock);
1191                         RETURN(rc);
1192                 }
1193         }
1194
1195         lock_res_and_lock(lock);
1196         if (local && lock->l_req_mode == lock->l_granted_mode) {
1197                 /* The server returned a blocked lock, but it was granted before
1198                  * we got a chance to actually enqueue it.  We don't need to do
1199                  * anything else. */
1200                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1201                             LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1202                 GOTO(out, ELDLM_OK);
1203         }
1204
1205         /* Some flags from the enqueue want to make it into the AST, via the
1206          * lock's l_flags. */
1207         lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1208
1209         /* This distinction between local lock trees is very important; a client
1210          * namespace only has information about locks taken by that client, and
1211          * thus doesn't have enough information to decide for itself if it can
1212          * be granted (below).  In this case, we do exactly what the server
1213          * tells us to do, as dictated by the 'flags'.
1214          *
1215          * We do exactly the same thing during recovery, when the server is
1216          * more or less trusting the clients not to lie.
1217          *
1218          * FIXME (bug 268): Detect obvious lies by checking compatibility in
1219          * granted/converting queues. */
1220         ldlm_resource_unlink_lock(lock);
1221         if (local) {
1222                 if (*flags & LDLM_FL_BLOCK_CONV)
1223                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1224                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1225                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1226                 else
1227                         ldlm_grant_lock(lock, NULL);
1228                 GOTO(out, ELDLM_OK);
1229         } else if (*flags & LDLM_FL_REPLAY) {
1230                 if (*flags & LDLM_FL_BLOCK_CONV) {
1231                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1232                         GOTO(out, ELDLM_OK);
1233                 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1234                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1235                         GOTO(out, ELDLM_OK);
1236                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1237                         ldlm_grant_lock(lock, NULL);
1238                         GOTO(out, ELDLM_OK);
1239                 }
1240                 /* If no flags, fall through to normal enqueue path. */
1241         }
1242
1243         policy = ldlm_processing_policy_table[res->lr_type];
1244         policy(lock, flags, 1, &rc, NULL);
1245         GOTO(out, rc);
1246 out:
1247         unlock_res_and_lock(lock);
1248         return rc;
1249 }
1250
1251 /* Must be called with namespace taken: queue is waiting or converting. */
1252 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1253                          struct list_head *work_list)
1254 {
1255         struct list_head *tmp, *pos;
1256         ldlm_processing_policy policy;
1257         int flags;
1258         int rc = LDLM_ITER_CONTINUE;
1259         ldlm_error_t err;
1260         ENTRY;
1261
1262         check_res_locked(res);
1263
1264         policy = ldlm_processing_policy_table[res->lr_type];
1265         LASSERT(policy);
1266
1267         list_for_each_safe(tmp, pos, queue) {
1268                 struct ldlm_lock *pending;
1269                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1270
1271                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1272
1273                 flags = 0;
1274                 rc = policy(pending, &flags, 0, &err, work_list);
1275                 if (rc != LDLM_ITER_CONTINUE)
1276                         break;
1277         }
1278
1279         RETURN(rc);
1280 }
1281
1282 /* Helper function for pair ldlm_run_{bl,cp}_ast_work().
1283  *
1284  * Send an existing rpc set specified by @arg->set and then
1285  * destroy it. Create new one if @do_create flag is set. */
1286 static void
1287 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1288 {
1289         int rc;
1290
1291         rc = ptlrpc_set_wait(arg->set);
1292         if (arg->type == LDLM_BL_CALLBACK)
1293                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1294         ptlrpc_set_destroy(arg->set);
1295
1296         if (do_create)
1297                 arg->set = ptlrpc_prep_set();
1298 }
1299
1300 int ldlm_run_bl_ast_work(struct list_head *rpc_list)
1301 {
1302         struct ldlm_cb_set_arg arg;
1303         struct list_head *tmp, *pos;
1304         struct ldlm_lock_desc d;
1305         int ast_count;
1306         int rc = 0;
1307         ENTRY;
1308
1309         arg.set = ptlrpc_prep_set();
1310         atomic_set(&arg.restart, 0);
1311         arg.type = LDLM_BL_CALLBACK;
1312
1313         ast_count = 0;
1314         list_for_each_safe(tmp, pos, rpc_list) {
1315                 struct ldlm_lock *lock =
1316                         list_entry(tmp, struct ldlm_lock, l_bl_ast);
1317
1318                 /* nobody should touch l_bl_ast */
1319                 lock_res_and_lock(lock);
1320                 list_del_init(&lock->l_bl_ast);
1321
1322                 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1323                 LASSERT(lock->l_bl_ast_run == 0);
1324                 LASSERT(lock->l_blocking_lock);
1325                 lock->l_bl_ast_run++;
1326                 unlock_res_and_lock(lock);
1327
1328                 ldlm_lock2desc(lock->l_blocking_lock, &d);
1329
1330                 LDLM_LOCK_PUT(lock->l_blocking_lock);
1331                 lock->l_blocking_lock = NULL;
1332                 rc = lock->l_blocking_ast(lock, &d, (void *)&arg,
1333                                           LDLM_CB_BLOCKING);
1334                 LDLM_LOCK_PUT(lock);
1335                 ast_count++;
1336
1337                 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1338                  * and create a new set for requests that remained in
1339                  * @rpc_list */
1340                 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1341                         ldlm_send_and_maybe_create_set(&arg, 1);
1342                         ast_count = 0;
1343                 }
1344         }
1345
1346         if (ast_count > 0)
1347                 ldlm_send_and_maybe_create_set(&arg, 0);
1348         else
1349                 /* In case when number of ASTs is multiply of
1350                  * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1351                  * @arg.set must be destroyed here, otherwise we get
1352                  * write memory leaking. */
1353                 ptlrpc_set_destroy(arg.set);
1354
1355         RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1356 }
1357
1358 int ldlm_run_cp_ast_work(struct list_head *rpc_list)
1359 {
1360         struct ldlm_cb_set_arg arg;
1361         struct list_head *tmp, *pos;
1362         int ast_count;
1363         int rc = 0;
1364         ENTRY;
1365
1366         arg.set = ptlrpc_prep_set();
1367         atomic_set(&arg.restart, 0);
1368         arg.type = LDLM_CP_CALLBACK;
1369
1370         /* It's possible to receive a completion AST before we've set
1371          * the l_completion_ast pointer: either because the AST arrived
1372          * before the reply, or simply because there's a small race
1373          * window between receiving the reply and finishing the local
1374          * enqueue. (bug 842)
1375          *
1376          * This can't happen with the blocking_ast, however, because we
1377          * will never call the local blocking_ast until we drop our
1378          * reader/writer reference, which we won't do until we get the
1379          * reply and finish enqueueing. */
1380
1381         ast_count = 0;
1382         list_for_each_safe(tmp, pos, rpc_list) {
1383                 struct ldlm_lock *lock =
1384                         list_entry(tmp, struct ldlm_lock, l_cp_ast);
1385                 ldlm_completion_callback completion_callback;
1386
1387                 /* nobody should touch l_cp_ast */
1388                 lock_res_and_lock(lock);
1389                 list_del_init(&lock->l_cp_ast);
1390                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1391                 /* save l_completion_ast since it can be changed by
1392                  * mds_intent_policy(), see bug 14225 */
1393                 completion_callback = lock->l_completion_ast;
1394                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1395                 unlock_res_and_lock(lock);
1396
1397                 if (completion_callback != NULL) {
1398                         rc = completion_callback(lock, 0, (void *)&arg);
1399                         ast_count++;
1400                 }
1401                 LDLM_LOCK_PUT(lock);
1402
1403                 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1404                  * and create a new set for requests that remained in
1405                  * @rpc_list */
1406                 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1407                         ldlm_send_and_maybe_create_set(&arg, 1);
1408                         ast_count = 0;
1409                 }
1410         }
1411
1412         if (ast_count > 0)
1413                 ldlm_send_and_maybe_create_set(&arg, 0);
1414         else
1415                 /* In case when number of ASTs is multiply of
1416                  * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1417                  * @arg.set must be destroyed here, otherwise we get
1418                  * write memory leaking. */
1419                 ptlrpc_set_destroy(arg.set);
1420
1421         RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1422 }
1423
1424 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1425 {
1426         ldlm_reprocess_all(res);
1427         return LDLM_ITER_CONTINUE;
1428 }
1429
1430 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1431 {
1432         struct list_head *tmp;
1433         int i, rc;
1434
1435         ENTRY;
1436         spin_lock(&ns->ns_hash_lock);
1437         for (i = 0; i < RES_HASH_SIZE; i++) {
1438                 tmp = ns->ns_hash[i].next;
1439                 while (tmp != &(ns->ns_hash[i])) {
1440                         struct ldlm_resource *res =
1441                                 list_entry(tmp, struct ldlm_resource, lr_hash);
1442
1443                         ldlm_resource_getref(res);
1444                         spin_unlock(&ns->ns_hash_lock);
1445
1446                         rc = reprocess_one_queue(res, NULL);
1447
1448                         spin_lock(&ns->ns_hash_lock);
1449                         tmp = tmp->next;
1450                         ldlm_resource_putref_locked(res);
1451
1452                         if (rc == LDLM_ITER_STOP)
1453                                 GOTO(out, rc);
1454                 }
1455         }
1456  out:
1457         spin_unlock(&ns->ns_hash_lock);
1458         EXIT;
1459 }
1460
1461 void ldlm_reprocess_all(struct ldlm_resource *res)
1462 {
1463         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1464         int rc;
1465         ENTRY;
1466
1467         /* Local lock trees don't get reprocessed. */
1468         if (ns_is_client(res->lr_namespace)) {
1469                 EXIT;
1470                 return;
1471         }
1472
1473  restart:
1474         lock_res(res);
1475         rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1476         if (rc == LDLM_ITER_CONTINUE)
1477                 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1478         unlock_res(res);
1479
1480         rc = ldlm_run_cp_ast_work(&rpc_list);
1481         if (rc == -ERESTART) {
1482                 LASSERT(list_empty(&rpc_list));
1483                 goto restart;
1484         }
1485         EXIT;
1486 }
1487
1488 void ldlm_cancel_callback(struct ldlm_lock *lock)
1489 {
1490         check_res_locked(lock->l_resource);
1491         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1492                 lock->l_flags |= LDLM_FL_CANCEL;
1493                 if (lock->l_blocking_ast) {
1494                         // l_check_no_ns_lock(ns);
1495                         unlock_res_and_lock(lock);
1496                         lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1497                                              LDLM_CB_CANCELING);
1498                         lock_res_and_lock(lock);
1499                 } else {
1500                         LDLM_DEBUG(lock, "no blocking ast");
1501                 }
1502         }
1503         lock->l_flags |= LDLM_FL_BL_DONE;
1504 }
1505
1506 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1507 {
1508         struct ldlm_lock *lock;
1509
1510         if (req->l_resource->lr_type != LDLM_PLAIN &&
1511             req->l_resource->lr_type != LDLM_IBITS)
1512                 return;
1513
1514         if (LDLM_SL_HEAD(&req->l_sl_mode)) {
1515                 lock = list_entry(req->l_res_link.next, struct ldlm_lock,
1516                                   l_res_link);
1517                 if (req->l_sl_mode.next == &lock->l_sl_mode) {
1518                         lock->l_sl_mode.prev = NULL;
1519                 } else {
1520                         lock->l_sl_mode.next = req->l_sl_mode.next;
1521                         lock->l_sl_mode.next->prev = &lock->l_sl_mode;
1522                 }
1523                 req->l_sl_mode.next = NULL;
1524         } else if (LDLM_SL_TAIL(&req->l_sl_mode)) {
1525                 lock = list_entry(req->l_res_link.prev, struct ldlm_lock,
1526                                   l_res_link);
1527                 if (req->l_sl_mode.prev == &lock->l_sl_mode) {
1528                         lock->l_sl_mode.next = NULL;
1529                 } else {
1530                         lock->l_sl_mode.prev = req->l_sl_mode.prev;
1531                         lock->l_sl_mode.prev->next = &lock->l_sl_mode;
1532                 }
1533                 req->l_sl_mode.prev = NULL;
1534         }
1535
1536         if (LDLM_SL_HEAD(&req->l_sl_policy)) {
1537                 lock = list_entry(req->l_res_link.next, struct ldlm_lock,
1538                                   l_res_link);
1539                 if (req->l_sl_policy.next == &lock->l_sl_policy) {
1540                         lock->l_sl_policy.prev = NULL;
1541                 } else {
1542                         lock->l_sl_policy.next = req->l_sl_policy.next;
1543                         lock->l_sl_policy.next->prev = &lock->l_sl_policy;
1544                 }
1545                 req->l_sl_policy.next = NULL;
1546         } else if (LDLM_SL_TAIL(&req->l_sl_policy)) {
1547                 lock = list_entry(req->l_res_link.prev, struct ldlm_lock,
1548                                   l_res_link);
1549                 if (req->l_sl_policy.prev == &lock->l_sl_policy) {
1550                         lock->l_sl_policy.next = NULL;
1551                 } else {
1552                         lock->l_sl_policy.prev = req->l_sl_policy.prev;
1553                         lock->l_sl_policy.prev->next = &lock->l_sl_policy;
1554                 }
1555                 req->l_sl_policy.prev = NULL;
1556         }
1557 }
1558
1559 void ldlm_lock_cancel(struct ldlm_lock *lock)
1560 {
1561         struct ldlm_resource *res;
1562         struct ldlm_namespace *ns;
1563         ENTRY;
1564
1565         lock_res_and_lock(lock);
1566
1567         res = lock->l_resource;
1568         ns = res->lr_namespace;
1569
1570         /* Please do not, no matter how tempting, remove this LBUG without
1571          * talking to me first. -phik */
1572         if (lock->l_readers || lock->l_writers) {
1573                 LDLM_ERROR(lock, "lock still has references");
1574                 LBUG();
1575         }
1576
1577         ldlm_del_waiting_lock(lock);
1578
1579         /* Releases res lock */
1580         ldlm_cancel_callback(lock);
1581
1582         /* Yes, second time, just in case it was added again while we were
1583            running with no res lock in ldlm_cancel_callback */
1584         ldlm_del_waiting_lock(lock);
1585         ldlm_resource_unlink_lock(lock);
1586         ldlm_lock_destroy_nolock(lock);
1587
1588         if (lock->l_granted_mode == lock->l_req_mode)
1589                 ldlm_pool_del(&ns->ns_pool, lock);
1590
1591         /* Make sure we will not be called again for same lock what is possible
1592          * if not to zero out lock->l_granted_mode */
1593         lock->l_granted_mode = 0;
1594         unlock_res_and_lock(lock);
1595
1596         EXIT;
1597 }
1598
1599 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1600 {
1601         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1602         ENTRY;
1603
1604         if (lock == NULL)
1605                 RETURN(-EINVAL);
1606
1607         lock->l_ast_data = data;
1608         LDLM_LOCK_PUT(lock);
1609         RETURN(0);
1610 }
1611
1612 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1613 {
1614         struct ldlm_lock *lock;
1615         struct ldlm_resource *res;
1616
1617         spin_lock(&exp->exp_ldlm_data.led_lock);
1618         while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1619                 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1620                                   struct ldlm_lock, l_export_chain);
1621                 res = ldlm_resource_getref(lock->l_resource);
1622                 LDLM_LOCK_GET(lock);
1623                 spin_unlock(&exp->exp_ldlm_data.led_lock);
1624
1625                 LDLM_DEBUG(lock, "export %p", exp);
1626                 ldlm_res_lvbo_update(res, NULL, 0, 1);
1627
1628                 ldlm_lock_cancel(lock);
1629                 ldlm_reprocess_all(res);
1630
1631                 ldlm_resource_putref(res);
1632                 LDLM_LOCK_PUT(lock);
1633                 spin_lock(&exp->exp_ldlm_data.led_lock);
1634         }
1635         spin_unlock(&exp->exp_ldlm_data.led_lock);
1636 }
1637
1638 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1639                                         __u32 *flags)
1640 {
1641         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1642         struct ldlm_resource *res;
1643         struct ldlm_namespace *ns;
1644         int granted = 0;
1645         int old_mode, rc;
1646         struct ldlm_lock *mark_lock = NULL;
1647         int join= LDLM_JOIN_NONE;
1648         ldlm_error_t err;
1649         struct ldlm_interval *node;
1650         ENTRY;
1651
1652         if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1653                 *flags |= LDLM_FL_BLOCK_GRANTED;
1654                 RETURN(lock->l_resource);
1655         }
1656
1657         /* I can't check the type of lock here because the bitlock of lock
1658          * is not held here, so do the allocation blindly. -jay */
1659         OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO, sizeof(*node));
1660         if (node == NULL)  /* Actually, this causes EDEADLOCK to be returned */
1661                 RETURN(NULL);
1662
1663         LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1664                  "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1665
1666         lock_res_and_lock(lock);
1667
1668         res = lock->l_resource;
1669         ns = res->lr_namespace;
1670
1671         old_mode = lock->l_req_mode;
1672         lock->l_req_mode = new_mode;
1673         if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1674                 /* remember the lock position where the lock might be
1675                  * added back to the granted list later and also
1676                  * remember the join mode for skiplist fixing. */
1677                 if (LDLM_SL_HEAD(&lock->l_sl_mode))
1678                         join = LDLM_MODE_JOIN_RIGHT;
1679                 else if (LDLM_SL_TAIL(&lock->l_sl_mode))
1680                         join = LDLM_MODE_JOIN_LEFT;
1681                 if (LDLM_SL_HEAD(&lock->l_sl_policy))
1682                         join |= LDLM_POLICY_JOIN_RIGHT;
1683                 else if (LDLM_SL_TAIL(&lock->l_sl_policy))
1684                         join |= LDLM_POLICY_JOIN_LEFT;
1685
1686                 LASSERT(!((join & LDLM_MODE_JOIN_RIGHT) &&
1687                           (join & LDLM_POLICY_JOIN_LEFT)));
1688                 LASSERT(!((join & LDLM_MODE_JOIN_LEFT) &&
1689                           (join & LDLM_POLICY_JOIN_RIGHT)));
1690
1691                 if ((join & LDLM_MODE_JOIN_LEFT) ||
1692                     (join & LDLM_POLICY_JOIN_LEFT))
1693                         mark_lock = list_entry(lock->l_res_link.prev,
1694                                                struct ldlm_lock, l_res_link);
1695                 else if (lock->l_res_link.next != &res->lr_granted)
1696                         mark_lock = list_entry(lock->l_res_link.next,
1697                                                struct ldlm_lock, l_res_link);
1698         } else {
1699                 ldlm_resource_unlink_lock(lock);
1700                 if (res->lr_type == LDLM_EXTENT) {
1701                         /* FIXME: ugly code, I have to attach the lock to a
1702                          * interval node again since perhaps it will be granted
1703                          * soon */
1704                         CFS_INIT_LIST_HEAD(&node->li_group);
1705                         ldlm_interval_attach(node, lock);
1706                         node = NULL;
1707                 }
1708         }
1709
1710         /* If this is a local resource, put it on the appropriate list. */
1711         if (ns_is_client(res->lr_namespace)) {
1712                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1713                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1714                 } else {
1715                         /* This should never happen, because of the way the
1716                          * server handles conversions. */
1717                         LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1718                                    *flags);
1719                         LBUG();
1720
1721                         ldlm_grant_lock(lock, &rpc_list);
1722                         granted = 1;
1723                         /* FIXME: completion handling not with ns_lock held ! */
1724                         if (lock->l_completion_ast)
1725                                 lock->l_completion_ast(lock, 0, NULL);
1726                 }
1727         } else {
1728                 int pflags = 0;
1729                 ldlm_processing_policy policy;
1730                 policy = ldlm_processing_policy_table[res->lr_type];
1731                 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1732                 if (rc == LDLM_ITER_STOP) {
1733                         lock->l_req_mode = old_mode;
1734                         if (res->lr_type == LDLM_EXTENT)
1735                                 ldlm_extent_add_lock(res, lock);
1736                         else
1737                                 ldlm_granted_list_add_lock(lock, mark_lock,
1738                                                            join);
1739                         res = NULL;
1740                 } else {
1741                         *flags |= LDLM_FL_BLOCK_GRANTED;
1742                         granted = 1;
1743                 }
1744         }
1745         unlock_res_and_lock(lock);
1746
1747         if (granted)
1748                 ldlm_run_cp_ast_work(&rpc_list);
1749         if (node)
1750                 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1751         RETURN(res);
1752 }
1753
1754 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1755 {
1756         struct obd_device *obd = NULL;
1757
1758         if (!((libcfs_debug | D_ERROR) & level))
1759                 return;
1760
1761         if (!lock) {
1762                 CDEBUG(level, "  NULL LDLM lock\n");
1763                 return;
1764         }
1765
1766         CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1767                lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1768                pos, lock->l_pid);
1769         if (lock->l_conn_export != NULL)
1770                 obd = lock->l_conn_export->exp_obd;
1771         if (lock->l_export && lock->l_export->exp_connection) {
1772                 CDEBUG(level, "  Node: NID %s (rhandle: "LPX64")\n",
1773                      libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1774                      lock->l_remote_handle.cookie);
1775         } else if (obd == NULL) {
1776                 CDEBUG(level, "  Node: local\n");
1777         } else {
1778                 struct obd_import *imp = obd->u.cli.cl_import;
1779                 CDEBUG(level, "  Node: NID %s (rhandle: "LPX64")\n",
1780                        libcfs_nid2str(imp->imp_connection->c_peer.nid),
1781                        lock->l_remote_handle.cookie);
1782         }
1783         CDEBUG(level, "  Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource,
1784                lock->l_resource->lr_name.name[0],
1785                lock->l_resource->lr_name.name[1]);
1786         CDEBUG(level, "  Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1787                "write: %d flags: %#x\n", ldlm_lockname[lock->l_req_mode],
1788                ldlm_lockname[lock->l_granted_mode],
1789                atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1790                lock->l_flags);
1791         if (lock->l_resource->lr_type == LDLM_EXTENT)
1792                 CDEBUG(level, "  Extent: "LPU64" -> "LPU64
1793                        " (req "LPU64"-"LPU64")\n",
1794                        lock->l_policy_data.l_extent.start,
1795                        lock->l_policy_data.l_extent.end,
1796                        lock->l_req_extent.start, lock->l_req_extent.end);
1797         else if (lock->l_resource->lr_type == LDLM_FLOCK)
1798                 CDEBUG(level, "  Pid: %d Extent: "LPU64" -> "LPU64"\n",
1799                        lock->l_policy_data.l_flock.pid,
1800                        lock->l_policy_data.l_flock.start,
1801                        lock->l_policy_data.l_flock.end);
1802        else if (lock->l_resource->lr_type == LDLM_IBITS)
1803                 CDEBUG(level, "  Bits: "LPX64"\n",
1804                        lock->l_policy_data.l_inodebits.bits);
1805 }
1806
1807 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1808 {
1809         struct ldlm_lock *lock;
1810
1811         if (!((libcfs_debug | D_ERROR) & level))
1812                 return;
1813
1814         lock = ldlm_handle2lock(lockh);
1815         if (lock == NULL)
1816                 return;
1817
1818         ldlm_lock_dump(D_OTHER, lock, 0);
1819
1820         LDLM_LOCK_PUT(lock);
1821 }
1822
1823 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1824                       struct libcfs_debug_msg_data *data, const char *fmt,
1825                       ...)
1826 {
1827         va_list args;
1828         cfs_debug_limit_state_t *cdls = data->msg_cdls;
1829
1830         va_start(args, fmt);
1831         if (lock->l_resource == NULL) {
1832                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1833                                    data->msg_fn, data->msg_line, fmt, args,
1834                                    " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1835                                    "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "
1836                                    LPX64" expref: %d pid: %u\n", lock,
1837                                    lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1838                                    lock->l_readers, lock->l_writers,
1839                                    ldlm_lockname[lock->l_granted_mode],
1840                                    ldlm_lockname[lock->l_req_mode],
1841                                    lock->l_flags, lock->l_remote_handle.cookie,
1842                                    lock->l_export ?
1843                                         atomic_read(&lock->l_export->exp_refcount) : -99,
1844                                    lock->l_pid);
1845                 va_end(args);
1846                 return;
1847         }
1848
1849         switch (lock->l_resource->lr_type) {
1850         case LDLM_EXTENT:
1851                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1852                                    data->msg_fn, data->msg_line, fmt, args,
1853                                    " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1854                                    "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1855                                    "] (req "LPU64"->"LPU64") flags: %x remote: "LPX64
1856                                     " expref: %d pid: %u\n",
1857                                     lock->l_resource->lr_namespace->ns_name, lock,
1858                                     lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1859                                     lock->l_readers, lock->l_writers,
1860                                     ldlm_lockname[lock->l_granted_mode],
1861                                     ldlm_lockname[lock->l_req_mode],
1862                                     lock->l_resource->lr_name.name[0],
1863                                     lock->l_resource->lr_name.name[1],
1864                                     atomic_read(&lock->l_resource->lr_refcount),
1865                                     ldlm_typename[lock->l_resource->lr_type],
1866                                     lock->l_policy_data.l_extent.start,
1867                                     lock->l_policy_data.l_extent.end,
1868                                     lock->l_req_extent.start, lock->l_req_extent.end,
1869                                     lock->l_flags, lock->l_remote_handle.cookie,
1870                                     lock->l_export ?
1871                                         atomic_read(&lock->l_export->exp_refcount) : -99,
1872                                     lock->l_pid);
1873                 break;
1874         case LDLM_FLOCK:
1875                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1876                                    data->msg_fn, data->msg_line, fmt, args,
1877                                    " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1878                                    "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1879                                    "["LPU64"->"LPU64"] flags: %x remote: "LPX64
1880                                    " expref: %d pid: %u\n",
1881                                    lock->l_resource->lr_namespace->ns_name, lock,
1882                                    lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1883                                    lock->l_readers, lock->l_writers,
1884                                    ldlm_lockname[lock->l_granted_mode],
1885                                    ldlm_lockname[lock->l_req_mode],
1886                                    lock->l_resource->lr_name.name[0],
1887                                    lock->l_resource->lr_name.name[1],
1888                                    atomic_read(&lock->l_resource->lr_refcount),
1889                                    ldlm_typename[lock->l_resource->lr_type],
1890                                    lock->l_policy_data.l_flock.pid,
1891                                    lock->l_policy_data.l_flock.start,
1892                                    lock->l_policy_data.l_flock.end,
1893                                    lock->l_flags, lock->l_remote_handle.cookie,
1894                                    lock->l_export ?
1895                                         atomic_read(&lock->l_export->exp_refcount) : -99,
1896                                    lock->l_pid);
1897                 break;
1898         case LDLM_IBITS:
1899                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1900                                    data->msg_fn, data->msg_line, fmt, args,
1901                                    " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1902                                    "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1903                                    "flags: %x remote: "LPX64" expref: %d "
1904                                    "pid %u\n",
1905                                    lock->l_resource->lr_namespace->ns_name,
1906                                    lock, lock->l_handle.h_cookie,
1907                                    atomic_read (&lock->l_refc),
1908                                    lock->l_readers, lock->l_writers,
1909                                    ldlm_lockname[lock->l_granted_mode],
1910                                    ldlm_lockname[lock->l_req_mode],
1911                                    lock->l_resource->lr_name.name[0],
1912                                    lock->l_resource->lr_name.name[1],
1913                                    lock->l_policy_data.l_inodebits.bits,
1914                                    atomic_read(&lock->l_resource->lr_refcount),
1915                                    ldlm_typename[lock->l_resource->lr_type],
1916                                    lock->l_flags, lock->l_remote_handle.cookie,
1917                                    lock->l_export ?
1918                                         atomic_read(&lock->l_export->exp_refcount) : -99,
1919                                    lock->l_pid);
1920                 break;
1921         default:
1922                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1923                                    data->msg_fn, data->msg_line, fmt, args,
1924                                    " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1925                                    "res: "LPU64"/"LPU64" rrc: %d type: %s flags: %x "
1926                                    "remote: "LPX64" expref: %d pid: %u\n",
1927                                    lock->l_resource->lr_namespace->ns_name,
1928                                    lock, lock->l_handle.h_cookie,
1929                                    atomic_read (&lock->l_refc),
1930                                    lock->l_readers, lock->l_writers,
1931                                    ldlm_lockname[lock->l_granted_mode],
1932                                    ldlm_lockname[lock->l_req_mode],
1933                                    lock->l_resource->lr_name.name[0],
1934                                    lock->l_resource->lr_name.name[1],
1935                                    atomic_read(&lock->l_resource->lr_refcount),
1936                                    ldlm_typename[lock->l_resource->lr_type],
1937                                    lock->l_flags, lock->l_remote_handle.cookie,
1938                                    lock->l_export ?
1939                                          atomic_read(&lock->l_export->exp_refcount) : -99,
1940                                    lock->l_pid);
1941                 break;
1942         }
1943         va_end(args);
1944 }
1945 EXPORT_SYMBOL(_ldlm_lock_debug);