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 types */
  43 char *ldlm_lockname[] = {
  44         [0] "--",
  45         [LCK_EX] "EX",
  46         [LCK_PW] "PW",
  47         [LCK_PR] "PR",
  48         [LCK_CW] "CW",
  49         [LCK_CR] "CR",
  50         [LCK_NL] "NL",
  51         [LCK_GROUP] "GROUP"
  52 };
  53
  54 char *ldlm_typename[] = {
  55         [LDLM_PLAIN] "PLN",
  56         [LDLM_EXTENT] "EXT",
  57         [LDLM_FLOCK] "FLK",
  58         [LDLM_IBITS] "IBT",
  59 };
  60
  61 char *ldlm_it2str(int it)
  62 {
  63         switch (it) {
  64         case IT_OPEN:
  65                 return "open";
  66         case IT_CREAT:
  67                 return "creat";
  68         case (IT_OPEN | IT_CREAT):
  69                 return "open|creat";
  70         case IT_READDIR:
  71                 return "readdir";
  72         case IT_GETATTR:
  73                 return "getattr";
  74         case IT_LOOKUP:
  75                 return "lookup";
  76         case IT_UNLINK:
  77                 return "unlink";
  78         case IT_GETXATTR:
  79                 return "getxattr";
  80         default:
  81                 CERROR("Unknown intent %d\n", it);
  82                 return "UNKNOWN";
  83         }
  84 }
  85
  86 extern cfs_mem_cache_t *ldlm_lock_slab;
  87
  88 static ldlm_processing_policy ldlm_processing_policy_table[] = {
  89         [LDLM_PLAIN] ldlm_process_plain_lock,
  90         [LDLM_EXTENT] ldlm_process_extent_lock,
  91 #ifdef __KERNEL__
  92         [LDLM_FLOCK] ldlm_process_flock_lock,
  93 #endif
  94         [LDLM_IBITS] ldlm_process_inodebits_lock,
  95 };
  96
  97 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
  98 {
  99         return ldlm_processing_policy_table[res->lr_type];
 100 }
 101
 102 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
 103 {
 104         ns->ns_policy = arg;
 105 }
 106
 107 /*
 108  * REFCOUNTED LOCK OBJECTS
 109  */
 110
 111
 112 /*
 113  * Lock refcounts, during creation:
 114  *   - one special one for allocation, dec'd only once in destroy
 115  *   - one for being a lock that's in-use
 116  *   - one for the addref associated with a new lock
 117  */
 118 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 119 {
 120         atomic_inc(&lock->l_refc);
 121         return lock;
 122 }
 123
 124 void ldlm_lock_put(struct ldlm_lock *lock)
 125 {
 126         ENTRY;
 127
 128         LASSERT(lock->l_resource != LP_POISON);
 129         LASSERT(atomic_read(&lock->l_refc) > 0);
 130         if (atomic_dec_and_test(&lock->l_refc)) {
 131                 struct ldlm_resource *res;
 132
 133                 LDLM_DEBUG(lock,
 134                            "final lock_put on destroyed lock, freeing it.\n");
 135
 136                 lock_res_and_lock(lock);
 137                 res = lock->l_resource;
 138                 LASSERT(lock->l_destroyed);
 139                 LASSERT(list_empty(&lock->l_res_link));
 140
 141                 if (lock->l_parent)
 142                         LDLM_LOCK_PUT(lock->l_parent);
 143                 unlock_res_and_lock(lock);
 144
 145                 atomic_dec(&res->lr_namespace->ns_locks);
 146                 ldlm_resource_putref(res);
 147                 lock->l_resource = NULL;
 148                 if (lock->l_export)
 149                         class_export_put(lock->l_export);
 150
 151                 if (lock->l_lvb_data != NULL)
 152                         OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
 153
 154                 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
 155         }
 156
 157         EXIT;
 158 }
 159
 160 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
 161 {
 162         int rc = 0;
 163         if (!list_empty(&lock->l_lru)) {
 164                 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
 165                 list_del_init(&lock->l_lru);
 166                 lock->l_resource->lr_namespace->ns_nr_unused--;
 167                 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
 168                 rc = 1;
 169         }
 170         return rc;
 171 }
 172
 173 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
 174 {
 175         int rc;
 176         ENTRY;
 177         spin_lock(&lock->l_resource->lr_namespace->ns_unused_lock);
 178         rc = ldlm_lock_remove_from_lru_nolock(lock);
 179         spin_unlock(&lock->l_resource->lr_namespace->ns_unused_lock);
 180         EXIT;
 181         return rc;
 182 }
 183
 184 /* This used to have a 'strict' flag, which recovery would use to mark an
 185  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
 186  * shall explain why it's gone: with the new hash table scheme, once you call
 187  * ldlm_lock_destroy, you can never drop your final references on this lock.
 188  * Because it's not in the hash table anymore.  -phil */
 189 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
 190 {
 191         ENTRY;
 192
 193         if (!list_empty(&lock->l_children)) {
 194                 LDLM_ERROR(lock, "still has children (%p)!",
 195                            lock->l_children.next);
 196                 ldlm_lock_dump(D_ERROR, lock, 0);
 197                 LBUG();
 198         }
 199         if (lock->l_readers || lock->l_writers) {
 200                 LDLM_ERROR(lock, "lock still has references");
 201                 ldlm_lock_dump(D_ERROR, lock, 0);
 202                 LBUG();
 203         }
 204
 205         if (!list_empty(&lock->l_res_link)) {
 206                 LDLM_ERROR(lock, "lock still on resource");
 207                 ldlm_lock_dump(D_ERROR, lock, 0);
 208                 LBUG();
 209         }
 210
 211         if (lock->l_destroyed) {
 212                 LASSERT(list_empty(&lock->l_lru));
 213                 EXIT;
 214                 return 0;
 215         }
 216         lock->l_destroyed = 1;
 217
 218         if (lock->l_export)
 219                 spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
 220         list_del_init(&lock->l_export_chain);
 221         if (lock->l_export)
 222                 spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
 223
 224         ldlm_lock_remove_from_lru(lock);
 225         class_handle_unhash(&lock->l_handle);
 226
 227 #if 0
 228         /* Wake anyone waiting for this lock */
 229         /* FIXME: I should probably add yet another flag, instead of using
 230          * l_export to only call this on clients */
 231         if (lock->l_export)
 232                 class_export_put(lock->l_export);
 233         lock->l_export = NULL;
 234         if (lock->l_export && lock->l_completion_ast)
 235                 lock->l_completion_ast(lock, 0);
 236 #endif
 237         EXIT;
 238         return 1;
 239 }
 240
 241 void ldlm_lock_destroy(struct ldlm_lock *lock)
 242 {
 243         int first;
 244         ENTRY;
 245         lock_res_and_lock(lock);
 246         first = ldlm_lock_destroy_internal(lock);
 247         unlock_res_and_lock(lock);
 248
 249         /* drop reference from hashtable only for first destroy */
 250         if (first)
 251                 LDLM_LOCK_PUT(lock);
 252         EXIT;
 253 }
 254
 255 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
 256 {
 257         int first;
 258         ENTRY;
 259         first = ldlm_lock_destroy_internal(lock);
 260         /* drop reference from hashtable only for first destroy */
 261         if (first)
 262                 LDLM_LOCK_PUT(lock);
 263         EXIT;
 264 }
 265
 266 /* this is called by portals_handle2object with the handle lock taken */
 267 static void lock_handle_addref(void *lock)
 268 {
 269         LDLM_LOCK_GET((struct ldlm_lock *)lock);
 270 }
 271
 272 /*
 273  * usage: pass in a resource on which you have done ldlm_resource_get
 274  *        pass in a parent lock on which you have done a ldlm_lock_get
 275  *        after return, ldlm_*_put the resource and parent
 276  * returns: lock with refcount 2 - one for current caller and one for remote
 277  */
 278 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 279                                        struct ldlm_resource *resource)
 280 {
 281         struct ldlm_lock *lock;
 282         ENTRY;
 283
 284         if (resource == NULL)
 285                 LBUG();
 286
 287         OBD_SLAB_ALLOC(lock, ldlm_lock_slab, CFS_ALLOC_IO, sizeof(*lock));
 288         if (lock == NULL)
 289                 RETURN(NULL);
 290
 291         lock->l_resource = ldlm_resource_getref(resource);
 292
 293         atomic_set(&lock->l_refc, 2);
 294         CFS_INIT_LIST_HEAD(&lock->l_children);
 295         CFS_INIT_LIST_HEAD(&lock->l_res_link);
 296         CFS_INIT_LIST_HEAD(&lock->l_lru);
 297         CFS_INIT_LIST_HEAD(&lock->l_export_chain);
 298         CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
 299         CFS_INIT_LIST_HEAD(&lock->l_tmp);
 300         CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
 301         CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
 302         cfs_waitq_init(&lock->l_waitq);
 303         lock->l_blocking_lock = NULL;
 304         lock->l_pidb = 0;
 305
 306         atomic_inc(&resource->lr_namespace->ns_locks);
 307
 308         if (parent != NULL) {
 309                 spin_lock(&resource->lr_namespace->ns_hash_lock);
 310                 lock->l_parent = LDLM_LOCK_GET(parent);
 311                 list_add(&lock->l_childof, &parent->l_children);
 312                 spin_unlock(&resource->lr_namespace->ns_hash_lock);
 313         }
 314
 315         CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
 316         class_handle_hash(&lock->l_handle, lock_handle_addref);
 317
 318         RETURN(lock);
 319 }
 320
 321 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
 322                               struct ldlm_res_id new_resid)
 323 {
 324         struct ldlm_resource *oldres = lock->l_resource;
 325         struct ldlm_resource *newres;
 326         int type;
 327         ENTRY;
 328
 329         LASSERT(ns->ns_client != 0);
 330
 331         lock_res_and_lock(lock);
 332         if (memcmp(&new_resid, &lock->l_resource->lr_name,
 333                    sizeof(lock->l_resource->lr_name)) == 0) {
 334                 /* Nothing to do */
 335                 unlock_res_and_lock(lock);
 336                 RETURN(0);
 337         }
 338
 339         LASSERT(new_resid.name[0] != 0);
 340
 341         /* This function assumes that the lock isn't on any lists */
 342         LASSERT(list_empty(&lock->l_res_link));
 343
 344         type = oldres->lr_type;
 345         unlock_res_and_lock(lock);
 346
 347         newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
 348         if (newres == NULL) {
 349                 LBUG();
 350                 RETURN(-ENOMEM);
 351         }
 352
 353         lock_res_and_lock(lock);
 354         LASSERT(memcmp(&new_resid, &lock->l_resource->lr_name,
 355                        sizeof(lock->l_resource->lr_name)) != 0);
 356         lock_res(newres);
 357         lock->l_resource = newres;
 358         unlock_res(newres);
 359         unlock_res(oldres);
 360         unlock_bitlock(lock);
 361
 362         /* ...and the flowers are still standing! */
 363         ldlm_resource_putref(oldres);
 364
 365         RETURN(0);
 366 }
 367
 368 /*
 369  *  HANDLES
 370  */
 371
 372 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
 373 {
 374         lockh->cookie = lock->l_handle.h_cookie;
 375 }
 376
 377 /* if flags: atomically get the lock and set the flags.
 378  *           Return NULL if flag already set
 379  */
 380
 381 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
 382                                      int flags)
 383 {
 384         struct ldlm_namespace *ns;
 385         struct ldlm_lock *lock = NULL, *retval = NULL;
 386         ENTRY;
 387
 388         LASSERT(handle);
 389
 390         lock = class_handle2object(handle->cookie);
 391         if (lock == NULL)
 392                 RETURN(NULL);
 393
 394         LASSERT(lock->l_resource != NULL);
 395         ns = lock->l_resource->lr_namespace;
 396         LASSERT(ns != NULL);
 397
 398         lock_res_and_lock(lock);
 399
 400         /* It's unlikely but possible that someone marked the lock as
 401          * destroyed after we did handle2object on it */
 402         if (lock->l_destroyed) {
 403                 unlock_res_and_lock(lock);
 404                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
 405                 LDLM_LOCK_PUT(lock);
 406                 GOTO(out, retval);
 407         }
 408
 409         if (flags && (lock->l_flags & flags)) {
 410                 unlock_res_and_lock(lock);
 411                 LDLM_LOCK_PUT(lock);
 412                 GOTO(out, retval);
 413         }
 414
 415         if (flags)
 416                 lock->l_flags |= flags;
 417
 418         unlock_res_and_lock(lock);
 419         retval = lock;
 420         EXIT;
 421  out:
 422         return retval;
 423 }
 424
 425 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
 426                                       const struct lustre_handle *handle)
 427 {
 428         struct ldlm_lock *retval = NULL;
 429         retval = __ldlm_handle2lock(handle, 0);
 430         return retval;
 431 }
 432
 433 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 434 {
 435         struct obd_export *exp = lock->l_export?:lock->l_conn_export;
 436         /* INODEBITS_INTEROP: If the other side does not support
 437          * inodebits, reply with a plain lock descriptor.
 438          */
 439         if ((lock->l_resource->lr_type == LDLM_IBITS) &&
 440             (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
 441                 struct ldlm_resource res = *lock->l_resource;
 442
 443                 /* Make sure all the right bits are set in this lock we
 444                    are going to pass to client */
 445                 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
 446                          (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
 447                          "Inappropriate inode lock bits during "
 448                          "conversion " LPU64 "\n",
 449                          lock->l_policy_data.l_inodebits.bits);
 450                 res.lr_type = LDLM_PLAIN;
 451                 ldlm_res2desc(&res, &desc->l_resource);
 452                 /* Convert "new" lock mode to something old client can
 453                    understand */
 454                 if ((lock->l_req_mode == LCK_CR) ||
 455                     (lock->l_req_mode == LCK_CW))
 456                         desc->l_req_mode = LCK_PR;
 457                 else
 458                         desc->l_req_mode = lock->l_req_mode;
 459                 if ((lock->l_granted_mode == LCK_CR) ||
 460                     (lock->l_granted_mode == LCK_CW)) {
 461                         desc->l_granted_mode = LCK_PR;
 462                 } else {
 463                         /* We never grant PW/EX locks to clients */
 464                         LASSERT((lock->l_granted_mode != LCK_PW) &&
 465                                 (lock->l_granted_mode != LCK_EX));
 466                         desc->l_granted_mode = lock->l_granted_mode;
 467                 }
 468
 469                 /* We do not copy policy here, because there is no
 470                    policy for plain locks */
 471         } else {
 472                 ldlm_res2desc(lock->l_resource, &desc->l_resource);
 473                 desc->l_req_mode = lock->l_req_mode;
 474                 desc->l_granted_mode = lock->l_granted_mode;
 475                 desc->l_policy_data = lock->l_policy_data;
 476         }
 477 }
 478
 479 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
 480                            struct list_head *work_list)
 481 {
 482         if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
 483                 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
 484                 lock->l_flags |= LDLM_FL_AST_SENT;
 485                 /* If the enqueuing client said so, tell the AST recipient to
 486                  * discard dirty data, rather than writing back. */
 487                 if (new->l_flags & LDLM_AST_DISCARD_DATA)
 488                         lock->l_flags |= LDLM_FL_DISCARD_DATA;
 489                 LASSERT(list_empty(&lock->l_bl_ast));
 490                 list_add(&lock->l_bl_ast, work_list);
 491                 LDLM_LOCK_GET(lock);
 492                 LASSERT(lock->l_blocking_lock == NULL);
 493                 lock->l_blocking_lock = LDLM_LOCK_GET(new);
 494         }
 495 }
 496
 497 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
 498 {
 499         if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
 500                 lock->l_flags |= LDLM_FL_CP_REQD;
 501                 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
 502                 LASSERT(list_empty(&lock->l_cp_ast));
 503                 list_add(&lock->l_cp_ast, work_list);
 504                 LDLM_LOCK_GET(lock);
 505         }
 506 }
 507
 508 /* must be called with lr_lock held */
 509 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
 510                                 struct list_head *work_list)
 511 {
 512         ENTRY;
 513         check_res_locked(lock->l_resource);
 514         if (new)
 515                 ldlm_add_bl_work_item(lock, new, work_list);
 516         else
 517                 ldlm_add_cp_work_item(lock, work_list);
 518         EXIT;
 519 }
 520
 521 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 522 {
 523         struct ldlm_lock *lock;
 524
 525         lock = ldlm_handle2lock(lockh);
 526         LASSERT(lock != NULL);
 527         ldlm_lock_addref_internal(lock, mode);
 528         LDLM_LOCK_PUT(lock);
 529 }
 530
 531 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
 532 {
 533         ldlm_lock_remove_from_lru(lock);
 534         if (mode & (LCK_NL | LCK_CR | LCK_PR))
 535                 lock->l_readers++;
 536         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
 537                 lock->l_writers++;
 538         lock->l_last_used = cfs_time_current();
 539         LDLM_LOCK_GET(lock);
 540         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 541 }
 542
 543 /* only called for local locks */
 544 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 545 {
 546         lock_res_and_lock(lock);
 547         ldlm_lock_addref_internal_nolock(lock, mode);
 548         unlock_res_and_lock(lock);
 549 }
 550
 551 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
 552 {
 553         struct ldlm_namespace *ns;
 554         ENTRY;
 555
 556         lock_res_and_lock(lock);
 557
 558         ns = lock->l_resource->lr_namespace;
 559
 560         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 561         if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
 562                 LASSERT(lock->l_readers > 0);
 563                 lock->l_readers--;
 564         }
 565         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
 566                 LASSERT(lock->l_writers > 0);
 567                 lock->l_writers--;
 568         }
 569
 570         if (lock->l_flags & LDLM_FL_LOCAL &&
 571             !lock->l_readers && !lock->l_writers) {
 572                 /* If this is a local lock on a server namespace and this was
 573                  * the last reference, cancel the lock. */
 574                 CDEBUG(D_INFO, "forcing cancel of local lock\n");
 575                 lock->l_flags |= LDLM_FL_CBPENDING;
 576         }
 577
 578         if (!lock->l_readers && !lock->l_writers &&
 579             (lock->l_flags & LDLM_FL_CBPENDING)) {
 580                 /* If we received a blocked AST and this was the last reference,
 581                  * run the callback. */
 582                 if (ns->ns_client == LDLM_NAMESPACE_SERVER && lock->l_export)
 583                         CERROR("FL_CBPENDING set on non-local lock--just a "
 584                                "warning\n");
 585
 586                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 587
 588                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
 589                 ldlm_lock_remove_from_lru(lock);
 590                 unlock_res_and_lock(lock);
 591                 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
 592                                 ldlm_bl_to_thread(ns, NULL, lock) != 0)
 593                         ldlm_handle_bl_callback(ns, NULL, lock);
 594         } else if (ns->ns_client == LDLM_NAMESPACE_CLIENT &&
 595                    !lock->l_readers && !lock->l_writers &&
 596                    !(lock->l_flags & LDLM_FL_NO_LRU)) {
 597                 /* If this is a client-side namespace and this was the last
 598                  * reference, put it on the LRU. */
 599                 LASSERT(list_empty(&lock->l_lru));
 600                 LASSERT(ns->ns_nr_unused >= 0);
 601                 spin_lock(&ns->ns_unused_lock);
 602                 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 603                 ns->ns_nr_unused++;
 604                 spin_unlock(&ns->ns_unused_lock);
 605                 unlock_res_and_lock(lock);
 606                 ldlm_cancel_lru(ns, LDLM_ASYNC);
 607         } else {
 608                 unlock_res_and_lock(lock);
 609         }
 610
 611         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 612
 613         EXIT;
 614 }
 615
 616 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 617 {
 618         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 619         LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
 620         ldlm_lock_decref_internal(lock, mode);
 621         LDLM_LOCK_PUT(lock);
 622 }
 623
 624 /* This will drop a lock reference and mark it for destruction, but will not
 625  * necessarily cancel the lock before returning. */
 626 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
 627 {
 628         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 629         ENTRY;
 630
 631         LASSERT(lock != NULL);
 632
 633         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 634         lock_res_and_lock(lock);
 635         lock->l_flags |= LDLM_FL_CBPENDING;
 636         unlock_res_and_lock(lock);
 637         ldlm_lock_decref_internal(lock, mode);
 638         LDLM_LOCK_PUT(lock);
 639 }
 640
 641 /* NOTE: called by
 642  *  - ldlm_lock_enqueue
 643  *  - ldlm_reprocess_queue
 644  *  - ldlm_lock_convert
 645  *
 646  * must be called with lr_lock held
 647  */
 648 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
 649 {
 650         struct ldlm_resource *res = lock->l_resource;
 651         ENTRY;
 652
 653         check_res_locked(res);
 654
 655         lock->l_granted_mode = lock->l_req_mode;
 656         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 657
 658         if (lock->l_granted_mode < res->lr_most_restr)
 659                 res->lr_most_restr = lock->l_granted_mode;
 660
 661         if (work_list && lock->l_completion_ast != NULL)
 662                 ldlm_add_ast_work_item(lock, NULL, work_list);
 663
 664         EXIT;
 665 }
 666
 667 /* returns a referenced lock or NULL.  See the flag descriptions below, in the
 668  * comment above ldlm_lock_match */
 669 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 670                                       ldlm_policy_data_t *policy,
 671                                       struct ldlm_lock *old_lock, int flags)
 672 {
 673         struct ldlm_lock *lock;
 674         struct list_head *tmp;
 675
 676         list_for_each(tmp, queue) {
 677                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 678
 679                 if (lock == old_lock)
 680                         break;
 681
 682                 /* llite sometimes wants to match locks that will be
 683                  * canceled when their users drop, but we allow it to match
 684                  * if it passes in CBPENDING and the lock still has users.
 685                  * this is generally only going to be used by children
 686                  * whose parents already hold a lock so forward progress
 687                  * can still happen. */
 688                 if (lock->l_flags & LDLM_FL_CBPENDING &&
 689                     !(flags & LDLM_FL_CBPENDING))
 690                         continue;
 691                 if (lock->l_flags & LDLM_FL_CBPENDING &&
 692                     lock->l_readers == 0 && lock->l_writers == 0)
 693                         continue;
 694
 695                 if (!(lock->l_req_mode & mode))
 696                         continue;
 697
 698                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 699                     (lock->l_policy_data.l_extent.start >
 700                      policy->l_extent.start ||
 701                      lock->l_policy_data.l_extent.end < policy->l_extent.end))
 702                         continue;
 703
 704                 if (unlikely(mode == LCK_GROUP) &&
 705                     lock->l_resource->lr_type == LDLM_EXTENT &&
 706                     lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
 707                         continue;
 708
 709                 /* We match if we have existing lock with same or wider set
 710                    of bits. */
 711                 if (lock->l_resource->lr_type == LDLM_IBITS &&
 712                      ((lock->l_policy_data.l_inodebits.bits &
 713                       policy->l_inodebits.bits) !=
 714                       policy->l_inodebits.bits))
 715                         continue;
 716
 717                 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
 718                         continue;
 719
 720                 if ((flags & LDLM_FL_LOCAL_ONLY) &&
 721                     !(lock->l_flags & LDLM_FL_LOCAL))
 722                         continue;
 723
 724                 if (flags & LDLM_FL_TEST_LOCK)
 725                         LDLM_LOCK_GET(lock);
 726                 else
 727                         ldlm_lock_addref_internal_nolock(lock, mode);
 728                 return lock;
 729         }
 730
 731         return NULL;
 732 }
 733
 734 void ldlm_lock_allow_match(struct ldlm_lock *lock)
 735 {
 736         lock_res_and_lock(lock);
 737         lock->l_flags |= LDLM_FL_CAN_MATCH;
 738         cfs_waitq_signal(&lock->l_waitq);
 739         unlock_res_and_lock(lock);
 740 }
 741
 742 /* Can be called in two ways:
 743  *
 744  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
 745  * for a duplicate of.
 746  *
 747  * Otherwise, all of the fields must be filled in, to match against.
 748  *
 749  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
 750  *     server (ie, connh is NULL)
 751  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
 752  *     list will be considered
 753  * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
 754  *     to be canceled can still be matched as long as they still have reader
 755  *     or writer refernces
 756  * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
 757  *     just tell us if we would have matched.
 758  *
 759  * Returns 1 if it finds an already-existing lock that is compatible; in this
 760  * case, lockh is filled in with a addref()ed lock
 761  *
 762  * we also check security context, if that failed we simply return 0 (to keep
 763  * caller code unchanged), the context failure will be discovered by caller
 764  * sometime later.
 765  */
 766 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
 767                     struct ldlm_res_id *res_id, ldlm_type_t type,
 768                     ldlm_policy_data_t *policy, ldlm_mode_t mode,
 769                     struct lustre_handle *lockh)
 770 {
 771         struct ldlm_resource *res;
 772         struct ldlm_lock *lock, *old_lock = NULL;
 773         int rc = 0;
 774         ENTRY;
 775
 776         if (ns == NULL) {
 777                 old_lock = ldlm_handle2lock(lockh);
 778                 LASSERT(old_lock);
 779
 780                 ns = old_lock->l_resource->lr_namespace;
 781                 res_id = &old_lock->l_resource->lr_name;
 782                 type = old_lock->l_resource->lr_type;
 783                 mode = old_lock->l_req_mode;
 784         }
 785
 786         res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
 787         if (res == NULL) {
 788                 LASSERT(old_lock == NULL);
 789                 RETURN(0);
 790         }
 791
 792         lock_res(res);
 793
 794         lock = search_queue(&res->lr_granted, mode, policy, old_lock, flags);
 795         if (lock != NULL)
 796                 GOTO(out, rc = 1);
 797         if (flags & LDLM_FL_BLOCK_GRANTED)
 798                 GOTO(out, rc = 0);
 799         lock = search_queue(&res->lr_converting, mode, policy, old_lock, flags);
 800         if (lock != NULL)
 801                 GOTO(out, rc = 1);
 802         lock = search_queue(&res->lr_waiting, mode, policy, old_lock, flags);
 803         if (lock != NULL)
 804                 GOTO(out, rc = 1);
 805
 806         EXIT;
 807  out:
 808         unlock_res(res);
 809         ldlm_resource_putref(res);
 810
 811         if (lock) {
 812                 ldlm_lock2handle(lock, lockh);
 813                 if (!(lock->l_flags & LDLM_FL_CAN_MATCH)) {
 814                         struct l_wait_info lwi;
 815                         if (lock->l_completion_ast) {
 816                                 int err = lock->l_completion_ast(lock,
 817                                                           LDLM_FL_WAIT_NOREPROC,
 818                                                                  NULL);
 819                                 if (err) {
 820                                         if (flags & LDLM_FL_TEST_LOCK)
 821                                                 LDLM_LOCK_PUT(lock);
 822                                         else
 823                                                 ldlm_lock_decref_internal(lock, mode);
 824                                         rc = 0;
 825                                         goto out2;
 826                                 }
 827                         }
 828
 829                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), NULL,
 830                                                LWI_ON_SIGNAL_NOOP, NULL);
 831
 832                         /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
 833                         l_wait_event(lock->l_waitq,
 834                                      (lock->l_flags & LDLM_FL_CAN_MATCH), &lwi);
 835                 }
 836         }
 837  out2:
 838         if (rc) {
 839                 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
 840                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
 841                                 res_id->name[2] : policy->l_extent.start,
 842                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
 843                                 res_id->name[3] : policy->l_extent.end);
 844
 845                 /* check user's security context */
 846                 if (lock->l_conn_export &&
 847                     sptlrpc_import_check_ctx(
 848                                 class_exp2cliimp(lock->l_conn_export))) {
 849                         if (!(flags & LDLM_FL_TEST_LOCK))
 850                                 ldlm_lock_decref_internal(lock, mode);
 851                         rc = 0;
 852                 }
 853
 854                 if (flags & LDLM_FL_TEST_LOCK)
 855                         LDLM_LOCK_PUT(lock);
 856         } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
 857                 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
 858                                   LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
 859                                   type, mode, res_id->name[0], res_id->name[1],
 860                                   (type == LDLM_PLAIN || type == LDLM_IBITS) ?
 861                                         res_id->name[2] :policy->l_extent.start,
 862                                 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
 863                                         res_id->name[3] : policy->l_extent.end);
 864         }
 865         if (old_lock)
 866                 LDLM_LOCK_PUT(old_lock);
 867
 868         return rc;
 869 }
 870
 871 /* Returns a referenced lock */
 872 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 873                                    const struct lustre_handle *parent_lock_handle,
 874                                    const struct ldlm_res_id res_id,
 875                                    ldlm_type_t type,
 876                                    ldlm_mode_t mode,
 877                                    ldlm_blocking_callback blocking,
 878                                    ldlm_completion_callback completion,
 879                                    ldlm_glimpse_callback glimpse,
 880                                    void *data, __u32 lvb_len)
 881 {
 882         struct ldlm_resource *res, *parent_res = NULL;
 883         struct ldlm_lock *lock, *parent_lock = NULL;
 884         ENTRY;
 885
 886         if (parent_lock_handle) {
 887                 parent_lock = ldlm_handle2lock(parent_lock_handle);
 888                 if (parent_lock)
 889                         parent_res = parent_lock->l_resource;
 890         }
 891
 892         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 893         if (res == NULL)
 894                 RETURN(NULL);
 895
 896         lock = ldlm_lock_new(parent_lock, res);
 897         ldlm_resource_putref(res);
 898         if (parent_lock != NULL)
 899                 LDLM_LOCK_PUT(parent_lock);
 900
 901         if (lock == NULL)
 902                 RETURN(NULL);
 903
 904         lock->l_req_mode = mode;
 905         lock->l_ast_data = data;
 906         lock->l_blocking_ast = blocking;
 907         lock->l_completion_ast = completion;
 908         lock->l_glimpse_ast = glimpse;
 909         lock->l_pid = cfs_curproc_pid();
 910
 911         if (lvb_len) {
 912                 lock->l_lvb_len = lvb_len;
 913                 OBD_ALLOC(lock->l_lvb_data, lvb_len);
 914                 if (lock->l_lvb_data == NULL) {
 915                         OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
 916                         RETURN(NULL);
 917                 }
 918         }
 919
 920         RETURN(lock);
 921 }
 922
 923 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
 924                                struct ldlm_lock **lockp,
 925                                void *cookie, int *flags)
 926 {
 927         struct ldlm_lock *lock = *lockp;
 928         struct ldlm_resource *res = lock->l_resource;
 929         int local = res->lr_namespace->ns_client;
 930         ldlm_processing_policy policy;
 931         ldlm_error_t rc = ELDLM_OK;
 932         ENTRY;
 933
 934         do_gettimeofday(&lock->l_enqueued_time);
 935         /* policies are not executed on the client or during replay */
 936         if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
 937             && !local && ns->ns_policy) {
 938                 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
 939                                    NULL);
 940                 if (rc == ELDLM_LOCK_REPLACED) {
 941                         /* The lock that was returned has already been granted,
 942                          * and placed into lockp.  If it's not the same as the
 943                          * one we passed in, then destroy the old one and our
 944                          * work here is done. */
 945                         if (lock != *lockp) {
 946                                 ldlm_lock_destroy(lock);
 947                                 LDLM_LOCK_PUT(lock);
 948                         }
 949                         *flags |= LDLM_FL_LOCK_CHANGED;
 950                         RETURN(0);
 951                 } else if (rc != ELDLM_OK ||
 952                            (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
 953                         ldlm_lock_destroy(lock);
 954                         RETURN(rc);
 955                 }
 956         }
 957
 958         lock_res_and_lock(lock);
 959         if (local && lock->l_req_mode == lock->l_granted_mode) {
 960                 /* The server returned a blocked lock, but it was granted before
 961                  * we got a chance to actually enqueue it.  We don't need to do
 962                  * anything else. */
 963                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 964                             LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 965                 GOTO(out, ELDLM_OK);
 966         }
 967
 968         /* Some flags from the enqueue want to make it into the AST, via the
 969          * lock's l_flags. */
 970         lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
 971
 972         /* This distinction between local lock trees is very important; a client
 973          * namespace only has information about locks taken by that client, and
 974          * thus doesn't have enough information to decide for itself if it can
 975          * be granted (below).  In this case, we do exactly what the server
 976          * tells us to do, as dictated by the 'flags'.
 977          *
 978          * We do exactly the same thing during recovery, when the server is
 979          * more or less trusting the clients not to lie.
 980          *
 981          * FIXME (bug 268): Detect obvious lies by checking compatibility in
 982          * granted/converting queues. */
 983         ldlm_resource_unlink_lock(lock);
 984         if (local) {
 985                 if (*flags & LDLM_FL_BLOCK_CONV)
 986                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 987                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 988                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 989                 else
 990                         ldlm_grant_lock(lock, NULL);
 991                 GOTO(out, ELDLM_OK);
 992         } else if (*flags & LDLM_FL_REPLAY) {
 993                 if (*flags & LDLM_FL_BLOCK_CONV) {
 994                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 995                         GOTO(out, ELDLM_OK);
 996                 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
 997                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 998                         GOTO(out, ELDLM_OK);
 999                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1000                         ldlm_grant_lock(lock, NULL);
1001                         GOTO(out, ELDLM_OK);
1002                 }
1003                 /* If no flags, fall through to normal enqueue path. */
1004         }
1005
1006         policy = ldlm_processing_policy_table[res->lr_type];
1007         policy(lock, flags, 1, &rc, NULL);
1008         GOTO(out, rc);
1009 out:
1010         unlock_res_and_lock(lock);
1011         return rc;
1012 }
1013
1014 /* Must be called with namespace taken: queue is waiting or converting. */
1015 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1016                          struct list_head *work_list)
1017 {
1018         struct list_head *tmp, *pos;
1019         ldlm_processing_policy policy;
1020         int flags;
1021         int rc = LDLM_ITER_CONTINUE;
1022         ldlm_error_t err;
1023         ENTRY;
1024
1025         check_res_locked(res);
1026
1027         policy = ldlm_processing_policy_table[res->lr_type];
1028         LASSERT(policy);
1029
1030         list_for_each_safe(tmp, pos, queue) {
1031                 struct ldlm_lock *pending;
1032                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1033
1034                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1035
1036                 flags = 0;
1037                 rc = policy(pending, &flags, 0, &err, work_list);
1038                 if (rc != LDLM_ITER_CONTINUE)
1039                         break;
1040         }
1041
1042         RETURN(rc);
1043 }
1044
1045 int ldlm_run_bl_ast_work(struct list_head *rpc_list)
1046 {
1047         struct list_head *tmp, *pos;
1048         struct ldlm_lock_desc d;
1049         int rc = 0, retval = 0;
1050         ENTRY;
1051
1052         list_for_each_safe(tmp, pos, rpc_list) {
1053                 struct ldlm_lock *lock =
1054                         list_entry(tmp, struct ldlm_lock, l_bl_ast);
1055
1056                 /* nobody should touch l_bl_ast */
1057                 lock_res_and_lock(lock);
1058                 list_del_init(&lock->l_bl_ast);
1059
1060                 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1061                 LASSERT(lock->l_bl_ast_run == 0);
1062                 LASSERT(lock->l_blocking_lock);
1063                 lock->l_bl_ast_run++;
1064                 unlock_res_and_lock(lock);
1065
1066                 ldlm_lock2desc(lock->l_blocking_lock, &d);
1067
1068                 LDLM_LOCK_PUT(lock->l_blocking_lock);
1069                 lock->l_blocking_lock = NULL;
1070                 rc = lock->l_blocking_ast(lock, &d, NULL, LDLM_CB_BLOCKING);
1071
1072                 if (rc == -ERESTART)
1073                         retval = rc;
1074                 else if (rc)
1075                         CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
1076                                "disconnect client\n");
1077                 LDLM_LOCK_PUT(lock);
1078         }
1079         RETURN(retval);
1080 }
1081
1082 int ldlm_run_cp_ast_work(struct list_head *rpc_list)
1083 {
1084         struct list_head *tmp, *pos;
1085         int rc = 0, retval = 0;
1086         ENTRY;
1087
1088         /* It's possible to receive a completion AST before we've set
1089          * the l_completion_ast pointer: either because the AST arrived
1090          * before the reply, or simply because there's a small race
1091          * window between receiving the reply and finishing the local
1092          * enqueue. (bug 842)
1093          *
1094          * This can't happen with the blocking_ast, however, because we
1095          * will never call the local blocking_ast until we drop our
1096          * reader/writer reference, which we won't do until we get the
1097          * reply and finish enqueueing. */
1098
1099         list_for_each_safe(tmp, pos, rpc_list) {
1100                 struct ldlm_lock *lock =
1101                         list_entry(tmp, struct ldlm_lock, l_cp_ast);
1102
1103                 /* nobody should touch l_cp_ast */
1104                 lock_res_and_lock(lock);
1105                 list_del_init(&lock->l_cp_ast);
1106                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1107                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1108                 unlock_res_and_lock(lock);
1109
1110                 if (lock->l_completion_ast != NULL)
1111                         rc = lock->l_completion_ast(lock, 0, 0);
1112                 if (rc == -ERESTART)
1113                         retval = rc;
1114                 else if (rc)
1115                         CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
1116                                "disconnect client\n");
1117                 LDLM_LOCK_PUT(lock);
1118         }
1119         RETURN(retval);
1120 }
1121
1122 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1123 {
1124         ldlm_reprocess_all(res);
1125         return LDLM_ITER_CONTINUE;
1126 }
1127
1128 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1129 {
1130         struct list_head *tmp;
1131         int i, rc;
1132
1133         if (ns == NULL)
1134                 return;
1135
1136         ENTRY;
1137         spin_lock(&ns->ns_hash_lock);
1138         for (i = 0; i < RES_HASH_SIZE; i++) {
1139                 tmp = ns->ns_hash[i].next;
1140                 while (tmp != &(ns->ns_hash[i])) {
1141                         struct ldlm_resource *res =
1142                                 list_entry(tmp, struct ldlm_resource, lr_hash);
1143
1144                         ldlm_resource_getref(res);
1145                         spin_unlock(&ns->ns_hash_lock);
1146
1147                         rc = reprocess_one_queue(res, NULL);
1148
1149                         spin_lock(&ns->ns_hash_lock);
1150                         tmp = tmp->next;
1151                         ldlm_resource_putref_locked(res);
1152
1153                         if (rc == LDLM_ITER_STOP)
1154                                 GOTO(out, rc);
1155                 }
1156         }
1157  out:
1158         spin_unlock(&ns->ns_hash_lock);
1159         EXIT;
1160 }
1161
1162 void ldlm_reprocess_all(struct ldlm_resource *res)
1163 {
1164         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1165         int rc;
1166         ENTRY;
1167
1168         /* Local lock trees don't get reprocessed. */
1169         if (res->lr_namespace->ns_client) {
1170                 EXIT;
1171                 return;
1172         }
1173
1174  restart:
1175         lock_res(res);
1176         rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1177         if (rc == LDLM_ITER_CONTINUE)
1178                 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1179         unlock_res(res);
1180
1181         rc = ldlm_run_cp_ast_work(&rpc_list);
1182         if (rc == -ERESTART) {
1183                 LASSERT(list_empty(&rpc_list));
1184                 goto restart;
1185         }
1186         EXIT;
1187 }
1188
1189 void ldlm_cancel_callback(struct ldlm_lock *lock)
1190 {
1191         check_res_locked(lock->l_resource);
1192         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1193                 lock->l_flags |= LDLM_FL_CANCEL;
1194                 if (lock->l_blocking_ast) {
1195                         // l_check_no_ns_lock(ns);
1196                         unlock_res_and_lock(lock);
1197                         lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1198                                              LDLM_CB_CANCELING);
1199                         lock_res_and_lock(lock);
1200                 } else {
1201                         LDLM_DEBUG(lock, "no blocking ast");
1202                 }
1203         }
1204 }
1205
1206 void ldlm_lock_cancel(struct ldlm_lock *lock)
1207 {
1208         struct ldlm_resource *res;
1209         struct ldlm_namespace *ns;
1210         ENTRY;
1211
1212         ldlm_del_waiting_lock(lock);
1213         lock_res_and_lock(lock);
1214
1215         res = lock->l_resource;
1216         ns = res->lr_namespace;
1217
1218         /* Please do not, no matter how tempting, remove this LBUG without
1219          * talking to me first. -phik */
1220         if (lock->l_readers || lock->l_writers) {
1221                 LDLM_ERROR(lock, "lock still has references");
1222                 LBUG();
1223         }
1224
1225         ldlm_cancel_callback(lock);
1226
1227         ldlm_resource_unlink_lock(lock);
1228         unlock_res_and_lock(lock);
1229
1230         ldlm_lock_destroy(lock);
1231
1232         EXIT;
1233 }
1234
1235 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1236 {
1237         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1238         ENTRY;
1239
1240         if (lock == NULL)
1241                 RETURN(-EINVAL);
1242
1243         lock->l_ast_data = data;
1244         LDLM_LOCK_PUT(lock);
1245         RETURN(0);
1246 }
1247
1248 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1249 {
1250         struct ldlm_lock *lock;
1251         struct ldlm_resource *res;
1252
1253         spin_lock(&exp->exp_ldlm_data.led_lock);
1254         while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1255                 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1256                                   struct ldlm_lock, l_export_chain);
1257                 res = ldlm_resource_getref(lock->l_resource);
1258                 LDLM_LOCK_GET(lock);
1259                 spin_unlock(&exp->exp_ldlm_data.led_lock);
1260
1261                 LDLM_DEBUG(lock, "export %p", exp);
1262                 ldlm_lock_cancel(lock);
1263                 ldlm_reprocess_all(res);
1264
1265                 ldlm_resource_putref(res);
1266                 LDLM_LOCK_PUT(lock);
1267                 spin_lock(&exp->exp_ldlm_data.led_lock);
1268         }
1269         spin_unlock(&exp->exp_ldlm_data.led_lock);
1270 }
1271
1272 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1273                                         int *flags)
1274 {
1275         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1276         struct ldlm_resource *res;
1277         struct ldlm_namespace *ns;
1278         int granted = 0;
1279         int old_mode, rc;
1280         ldlm_error_t err;
1281         ENTRY;
1282
1283         if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1284                 *flags |= LDLM_FL_BLOCK_GRANTED;
1285                 RETURN(lock->l_resource);
1286         }
1287
1288         LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1289                  "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1290
1291         lock_res_and_lock(lock);
1292
1293         res = lock->l_resource;
1294         ns = res->lr_namespace;
1295
1296         old_mode = lock->l_req_mode;
1297         lock->l_req_mode = new_mode;
1298         ldlm_resource_unlink_lock(lock);
1299
1300         /* If this is a local resource, put it on the appropriate list. */
1301         if (res->lr_namespace->ns_client) {
1302                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1303                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1304                 } else {
1305                         /* This should never happen, because of the way the
1306                          * server handles conversions. */
1307                         LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1308                                    *flags);
1309                         LBUG();
1310
1311                         ldlm_grant_lock(lock, &rpc_list);
1312                         granted = 1;
1313                         /* FIXME: completion handling not with ns_lock held ! */
1314                         if (lock->l_completion_ast)
1315                                 lock->l_completion_ast(lock, 0, NULL);
1316                 }
1317         } else {
1318                 int pflags = 0;
1319                 ldlm_processing_policy policy;
1320                 policy = ldlm_processing_policy_table[res->lr_type];
1321                 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1322                 if (rc == LDLM_ITER_STOP) {
1323                         lock->l_req_mode = old_mode;
1324                         ldlm_resource_add_lock(res, &res->lr_granted, lock);
1325                         res = NULL;
1326                 } else {
1327                         *flags |= LDLM_FL_BLOCK_GRANTED;
1328                         granted = 1;
1329                 }
1330         }
1331         unlock_res_and_lock(lock);
1332
1333         if (granted)
1334                 ldlm_run_cp_ast_work(&rpc_list);
1335         RETURN(res);
1336 }
1337
1338 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1339 {
1340         struct obd_device *obd = NULL;
1341
1342         if (!((libcfs_debug | D_ERROR) & level))
1343                 return;
1344
1345         if (!lock) {
1346                 CDEBUG(level, "  NULL LDLM lock\n");
1347                 return;
1348         }
1349
1350         CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1351                lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1352                pos, lock->l_pid);
1353         if (lock->l_conn_export != NULL)
1354                 obd = lock->l_conn_export->exp_obd;
1355         if (lock->l_export && lock->l_export->exp_connection) {
1356                 CDEBUG(level, "  Node: NID %s (rhandle: "LPX64")\n",
1357                      libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1358                      lock->l_remote_handle.cookie);
1359         } else if (obd == NULL) {
1360                 CDEBUG(level, "  Node: local\n");
1361         } else {
1362                 struct obd_import *imp = obd->u.cli.cl_import;
1363                 CDEBUG(level, "  Node: NID %s (rhandle: "LPX64")\n",
1364                        libcfs_nid2str(imp->imp_connection->c_peer.nid),
1365                        lock->l_remote_handle.cookie);
1366         }
1367         CDEBUG(level, "  Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1368                   lock->l_resource,
1369                   lock->l_resource->lr_name.name[0],
1370                   lock->l_resource->lr_name.name[1],
1371                   lock->l_resource->lr_name.name[2]);
1372         CDEBUG(level, "  Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1373                "write: %d flags: %#x\n", ldlm_lockname[lock->l_req_mode],
1374                ldlm_lockname[lock->l_granted_mode],
1375                atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1376                lock->l_flags);
1377         if (lock->l_resource->lr_type == LDLM_EXTENT)
1378                 CDEBUG(level, "  Extent: "LPU64" -> "LPU64
1379                        " (req "LPU64"-"LPU64")\n",
1380                        lock->l_policy_data.l_extent.start,
1381                        lock->l_policy_data.l_extent.end,
1382                        lock->l_req_extent.start, lock->l_req_extent.end);
1383         else if (lock->l_resource->lr_type == LDLM_FLOCK)
1384                 CDEBUG(level, "  Pid: %d Extent: "LPU64" -> "LPU64"\n",
1385                        lock->l_policy_data.l_flock.pid,
1386                        lock->l_policy_data.l_flock.start,
1387                        lock->l_policy_data.l_flock.end);
1388        else if (lock->l_resource->lr_type == LDLM_IBITS)
1389                 CDEBUG(level, "  Bits: "LPX64"\n",
1390                        lock->l_policy_data.l_inodebits.bits);
1391 }
1392
1393 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1394 {
1395         struct ldlm_lock *lock;
1396
1397         lock = ldlm_handle2lock(lockh);
1398         if (lock == NULL)
1399                 return;
1400
1401         ldlm_lock_dump(D_OTHER, lock, 0);
1402
1403         LDLM_LOCK_PUT(lock);
1404 }
1405
1406 void cdebug_va(cfs_debug_limit_state_t *cdls, __u32 mask,
1407                const char *file, const char *func, const int line,
1408                const char *fmt, va_list args);
1409 void cdebug(cfs_debug_limit_state_t *cdls, __u32 mask,
1410             const char *file, const char *func, const int line,
1411             const char *fmt, ...);
1412
1413 void
1414 ldlm_lock_debug(cfs_debug_limit_state_t *cdls,
1415                 __u32 level, struct ldlm_lock *lock,
1416                 const char *file, const char *func, const int line,
1417                 char *fmt, ...)
1418 {
1419         va_list args;
1420
1421         va_start(args, fmt);
1422         cdebug_va(cdls, level, file, func, line, fmt, args);
1423         va_end(args);
1424
1425         if (lock->l_resource == NULL) {
1426                 cdebug(cdls, level, file, func, line,
1427                        " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1428                        "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "
1429                        LPX64" expref: %d pid: %u\n", lock,
1430                        lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1431                        lock->l_readers, lock->l_writers,
1432                        ldlm_lockname[lock->l_granted_mode],
1433                        ldlm_lockname[lock->l_req_mode],
1434                        lock->l_flags, lock->l_remote_handle.cookie,
1435                        lock->l_export ?
1436                        atomic_read(&lock->l_export->exp_refcount) : -99,
1437                        lock->l_pid);
1438                 return;
1439         }
1440
1441         switch (lock->l_resource->lr_type) {
1442         case LDLM_EXTENT:
1443                 cdebug(cdls, level, file, func, line,
1444                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1445                        "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1446                        "] (req "LPU64"->"LPU64") flags: %x remote: "LPX64
1447                        " expref: %d pid: %u\n",
1448                        lock->l_resource->lr_namespace->ns_name, lock,
1449                        lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1450                        lock->l_readers, lock->l_writers,
1451                        ldlm_lockname[lock->l_granted_mode],
1452                        ldlm_lockname[lock->l_req_mode],
1453                        lock->l_resource->lr_name.name[0],
1454                        lock->l_resource->lr_name.name[1],
1455                        atomic_read(&lock->l_resource->lr_refcount),
1456                        ldlm_typename[lock->l_resource->lr_type],
1457                        lock->l_policy_data.l_extent.start,
1458                        lock->l_policy_data.l_extent.end,
1459                        lock->l_req_extent.start, lock->l_req_extent.end,
1460                        lock->l_flags, lock->l_remote_handle.cookie,
1461                        lock->l_export ?
1462                        atomic_read(&lock->l_export->exp_refcount) : -99,
1463                        lock->l_pid);
1464                 break;
1465
1466         case LDLM_FLOCK:
1467                 cdebug(cdls, level, file, func, line,
1468                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1469                        "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1470                        "["LPU64"->"LPU64"] flags: %x remote: "LPX64
1471                        " expref: %d pid: %u\n",
1472                        lock->l_resource->lr_namespace->ns_name, lock,
1473                        lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1474                        lock->l_readers, lock->l_writers,
1475                        ldlm_lockname[lock->l_granted_mode],
1476                        ldlm_lockname[lock->l_req_mode],
1477                        lock->l_resource->lr_name.name[0],
1478                        lock->l_resource->lr_name.name[1],
1479                        atomic_read(&lock->l_resource->lr_refcount),
1480                        ldlm_typename[lock->l_resource->lr_type],
1481                        lock->l_policy_data.l_flock.pid,
1482                        lock->l_policy_data.l_flock.start,
1483                        lock->l_policy_data.l_flock.end,
1484                        lock->l_flags, lock->l_remote_handle.cookie,
1485                        lock->l_export ?
1486                        atomic_read(&lock->l_export->exp_refcount) : -99,
1487                        lock->l_pid);
1488                 break;
1489
1490         case LDLM_IBITS:
1491                 cdebug(cdls, level, file, func, line,
1492                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1493                        "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1494                        "flags: %x remote: "LPX64" expref: %d "
1495                        "pid %u\n",
1496                        lock->l_resource->lr_namespace->ns_name,
1497                        lock, lock->l_handle.h_cookie,
1498                        atomic_read (&lock->l_refc),
1499                        lock->l_readers, lock->l_writers,
1500                        ldlm_lockname[lock->l_granted_mode],
1501                        ldlm_lockname[lock->l_req_mode],
1502                        lock->l_resource->lr_name.name[0],
1503                        lock->l_resource->lr_name.name[1],
1504                        lock->l_policy_data.l_inodebits.bits,
1505                        atomic_read(&lock->l_resource->lr_refcount),
1506                        ldlm_typename[lock->l_resource->lr_type],
1507                        lock->l_flags, lock->l_remote_handle.cookie,
1508                        lock->l_export ?
1509                        atomic_read(&lock->l_export->exp_refcount) : -99,
1510                        lock->l_pid);
1511                 break;
1512
1513         default:
1514                 cdebug(cdls, level, file, func, line,
1515                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1516                        "res: "LPU64"/"LPU64" rrc: %d type: %s flags: %x "
1517                        "remote: "LPX64" expref: %d pid: %u\n",
1518                        lock->l_resource->lr_namespace->ns_name,
1519                        lock, lock->l_handle.h_cookie,
1520                        atomic_read (&lock->l_refc),
1521                        lock->l_readers, lock->l_writers,
1522                        ldlm_lockname[lock->l_granted_mode],
1523                        ldlm_lockname[lock->l_req_mode],
1524                        lock->l_resource->lr_name.name[0],
1525                        lock->l_resource->lr_name.name[1],
1526                        atomic_read(&lock->l_resource->lr_refcount),
1527                        ldlm_typename[lock->l_resource->lr_type],
1528                        lock->l_flags, lock->l_remote_handle.cookie,
1529                        lock->l_export ?
1530                        atomic_read(&lock->l_export->exp_refcount) : -99,
1531                        lock->l_pid);
1532                 break;
1533         }
1534 }
1535 EXPORT_SYMBOL(ldlm_lock_debug);