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 Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #define DEBUG_SUBSYSTEM S_LDLM
  25
  26 #ifdef __KERNEL__
  27 #include <linux/slab.h>
  28 #include <linux/module.h>
  29 #include <linux/lustre_dlm.h>
  30 #include <linux/lustre_mds.h>
  31 #else
  32 #include <liblustre.h>
  33 #include <linux/kp30.h>
  34 #endif
  35
  36 #include <linux/obd_class.h>
  37
  38 //struct lustre_lock ldlm_everything_lock;
  39
  40 /* lock types */
  41 char *ldlm_lockname[] = {
  42         [0] "--",
  43         [LCK_EX] "EX",
  44         [LCK_PW] "PW",
  45         [LCK_PR] "PR",
  46         [LCK_CW] "CW",
  47         [LCK_CR] "CR",
  48         [LCK_NL] "NL"
  49 };
  50 char *ldlm_typename[] = {
  51         [LDLM_PLAIN] "PLN",
  52         [LDLM_EXTENT] "EXT",
  53 };
  54
  55 #ifdef __KERNEL__
  56 char *ldlm_it2str(int it)
  57 {
  58         switch (it) {
  59         case IT_OPEN:
  60                 return "open";
  61         case IT_CREAT:
  62                 return "creat";
  63         case (IT_OPEN | IT_CREAT):
  64                 return "open|creat";
  65         case IT_READDIR:
  66                 return "readdir";
  67         case IT_GETATTR:
  68                 return "getattr";
  69         case IT_LOOKUP:
  70                 return "lookup";
  71         case IT_UNLINK:
  72                 return "unlink";
  73         default:
  74                 CERROR("Unknown intent %d\n", it);
  75                 return "UNKNOWN";
  76         }
  77 }
  78 #endif
  79
  80 extern kmem_cache_t *ldlm_lock_slab;
  81 struct lustre_lock ldlm_handle_lock;
  82
  83 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b);
  84
  85 static ldlm_res_compat ldlm_res_compat_table[] = {
  86         [LDLM_PLAIN] ldlm_plain_compat,
  87         [LDLM_EXTENT] ldlm_extent_compat,
  88 };
  89
  90 static ldlm_res_policy ldlm_intent_policy_func;
  91
  92 static int ldlm_plain_policy(struct ldlm_namespace *ns, struct ldlm_lock **lock,
  93                              void *req_cookie, ldlm_mode_t mode, int flags,
  94                              void *data)
  95 {
  96         if ((flags & LDLM_FL_HAS_INTENT) && ldlm_intent_policy_func) {
  97                 return ldlm_intent_policy_func(ns, lock, req_cookie, mode,
  98                                                flags, data);
  99         }
 100
 101         return ELDLM_OK;
 102 }
 103
 104 static ldlm_res_policy ldlm_res_policy_table[] = {
 105         [LDLM_PLAIN] ldlm_plain_policy,
 106         [LDLM_EXTENT] ldlm_extent_policy,
 107 };
 108
 109 void ldlm_register_intent(ldlm_res_policy arg)
 110 {
 111         ldlm_intent_policy_func = arg;
 112 }
 113
 114 void ldlm_unregister_intent(void)
 115 {
 116         ldlm_intent_policy_func = NULL;
 117 }
 118
 119 /*
 120  * REFCOUNTED LOCK OBJECTS
 121  */
 122
 123
 124 /*
 125  * Lock refcounts, during creation:
 126  *   - one special one for allocation, dec'd only once in destroy
 127  *   - one for being a lock that's in-use
 128  *   - one for the addref associated with a new lock
 129  */
 130 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 131 {
 132         atomic_inc(&lock->l_refc);
 133         return lock;
 134 }
 135
 136 void ldlm_lock_put(struct ldlm_lock *lock)
 137 {
 138         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 139         ENTRY;
 140
 141         if (atomic_dec_and_test(&lock->l_refc)) {
 142                 l_lock(&ns->ns_lock);
 143                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
 144                 LASSERT(lock->l_destroyed);
 145                 LASSERT(list_empty(&lock->l_res_link));
 146
 147                 spin_lock(&ns->ns_counter_lock);
 148                 ns->ns_locks--;
 149                 spin_unlock(&ns->ns_counter_lock);
 150
 151                 ldlm_resource_putref(lock->l_resource);
 152                 lock->l_resource = NULL;
 153
 154                 if (lock->l_parent)
 155                         LDLM_LOCK_PUT(lock->l_parent);
 156
 157                 PORTAL_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
 158                 l_unlock(&ns->ns_lock);
 159         }
 160
 161         EXIT;
 162 }
 163
 164 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
 165 {
 166         ENTRY;
 167         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 168         if (!list_empty(&lock->l_lru)) {
 169                 list_del_init(&lock->l_lru);
 170                 lock->l_resource->lr_namespace->ns_nr_unused--;
 171                 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
 172         }
 173         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 174         EXIT;
 175 }
 176
 177 /* This used to have a 'strict' flact, which recovery would use to mark an
 178  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
 179  * shall explain why it's gone: with the new hash table scheme, once you call
 180  * ldlm_lock_destroy, you can never drop your final references on this lock.
 181  * Because it's not in the hash table anymore.  -phil */
 182 void ldlm_lock_destroy(struct ldlm_lock *lock)
 183 {
 184         ENTRY;
 185         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 186
 187         if (!list_empty(&lock->l_children)) {
 188                 LDLM_ERROR(lock, "still has children (%p)!",
 189                            lock->l_children.next);
 190                 ldlm_lock_dump(D_ERROR, lock);
 191                 LBUG();
 192         }
 193         if (lock->l_readers || lock->l_writers) {
 194                 LDLM_ERROR(lock, "lock still has references");
 195                 ldlm_lock_dump(D_ERROR, lock);
 196                 LBUG();
 197         }
 198
 199         if (!list_empty(&lock->l_res_link)) {
 200                 ldlm_lock_dump(D_ERROR, lock);
 201                 LBUG();
 202         }
 203
 204         if (lock->l_destroyed) {
 205                 LASSERT(list_empty(&lock->l_lru));
 206                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 207                 EXIT;
 208                 return;
 209         }
 210         lock->l_destroyed = 1;
 211
 212         list_del_init(&lock->l_export_chain);
 213         ldlm_lock_remove_from_lru(lock);
 214         class_handle_unhash(&lock->l_handle);
 215
 216 #if 0
 217         /* Wake anyone waiting for this lock */
 218         /* FIXME: I should probably add yet another flag, instead of using
 219          * l_export to only call this on clients */
 220         lock->l_export = NULL;
 221         if (lock->l_export && lock->l_completion_ast)
 222                 lock->l_completion_ast(lock, 0);
 223 #endif
 224
 225         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 226         LDLM_LOCK_PUT(lock);
 227         EXIT;
 228 }
 229
 230 /* this is called by portals_handle2object with the handle lock taken */
 231 static void lock_handle_addref(void *lock)
 232 {
 233         LDLM_LOCK_GET((struct ldlm_lock *)lock);
 234 }
 235
 236 /*
 237  * usage: pass in a resource on which you have done ldlm_resource_get
 238  *        pass in a parent lock on which you have done a ldlm_lock_get
 239  *        after return, ldlm_*_put the resource and parent
 240  * returns: lock with refcount 1
 241  */
 242 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 243                                        struct ldlm_resource *resource)
 244 {
 245         struct ldlm_lock *lock;
 246         ENTRY;
 247
 248         if (resource == NULL)
 249                 LBUG();
 250
 251         PORTAL_SLAB_ALLOC(lock, ldlm_lock_slab, sizeof(*lock));
 252         if (lock == NULL)
 253                 RETURN(NULL);
 254
 255         lock->l_resource = ldlm_resource_getref(resource);
 256
 257         atomic_set(&lock->l_refc, 2);
 258         INIT_LIST_HEAD(&lock->l_children);
 259         INIT_LIST_HEAD(&lock->l_res_link);
 260         INIT_LIST_HEAD(&lock->l_lru);
 261         INIT_LIST_HEAD(&lock->l_export_chain);
 262         INIT_LIST_HEAD(&lock->l_pending_chain);
 263         init_waitqueue_head(&lock->l_waitq);
 264
 265         spin_lock(&resource->lr_namespace->ns_counter_lock);
 266         resource->lr_namespace->ns_locks++;
 267         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 268
 269         if (parent != NULL) {
 270                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 271                 lock->l_parent = LDLM_LOCK_GET(parent);
 272                 list_add(&lock->l_childof, &parent->l_children);
 273                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 274         }
 275
 276         INIT_LIST_HEAD(&lock->l_handle.h_link);
 277         class_handle_hash(&lock->l_handle, lock_handle_addref);
 278
 279         RETURN(lock);
 280 }
 281
 282 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
 283                               struct ldlm_res_id new_resid)
 284 {
 285         struct ldlm_resource *oldres = lock->l_resource;
 286         ENTRY;
 287
 288         l_lock(&ns->ns_lock);
 289         if (memcmp(&new_resid, &lock->l_resource->lr_name,
 290                    sizeof(lock->l_resource->lr_name)) == 0) {
 291                 /* Nothing to do */
 292                 l_unlock(&ns->ns_lock);
 293                 RETURN(0);
 294         }
 295
 296         LASSERT(new_resid.name[0] != 0);
 297
 298         /* This function assumes that the lock isn't on any lists */
 299         LASSERT(list_empty(&lock->l_res_link));
 300
 301         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
 302                                              lock->l_resource->lr_type, 1);
 303         if (lock->l_resource == NULL) {
 304                 LBUG();
 305                 RETURN(-ENOMEM);
 306         }
 307
 308         /* ...and the flowers are still standing! */
 309         ldlm_resource_putref(oldres);
 310
 311         l_unlock(&ns->ns_lock);
 312         RETURN(0);
 313 }
 314
 315 /*
 316  *  HANDLES
 317  */
 318
 319 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 320 {
 321         POISON(&lockh->addr, 0x69, sizeof(lockh->addr));
 322         lockh->cookie = lock->l_handle.h_cookie;
 323 }
 324
 325 /* if flags: atomically get the lock and set the flags.
 326  *           Return NULL if flag already set
 327  */
 328
 329 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
 330 {
 331         struct ldlm_lock *lock = NULL, *retval = NULL;
 332         ENTRY;
 333
 334         LASSERT(handle);
 335
 336         lock = class_handle2object(handle->cookie);
 337         if (lock == NULL)
 338                 RETURN(NULL);
 339
 340         LASSERT(lock->l_resource != NULL);
 341         LASSERT(lock->l_resource->lr_namespace != NULL);
 342
 343         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 344
 345         /* It's unlikely but possible that someone marked the lock as
 346          * destroyed after we did handle2object on it */
 347         if (lock->l_destroyed) {
 348                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
 349                 LDLM_LOCK_PUT(lock);
 350                 GOTO(out, retval);
 351         }
 352
 353         if (flags && (lock->l_flags & flags)) {
 354                 LDLM_LOCK_PUT(lock);
 355                 GOTO(out, retval);
 356         }
 357
 358         if (flags)
 359                 lock->l_flags |= flags;
 360
 361         retval = lock;
 362         EXIT;
 363  out:
 364         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 365         return retval;
 366 }
 367
 368 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
 369                                       struct lustre_handle *handle)
 370 {
 371         struct ldlm_lock *retval = NULL;
 372
 373         l_lock(&ns->ns_lock);
 374         retval = __ldlm_handle2lock(handle, 0);
 375         l_unlock(&ns->ns_lock);
 376
 377         return retval;
 378 }
 379
 380 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 381 {
 382         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 383 }
 384
 385 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 386 {
 387         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 388         desc->l_req_mode = lock->l_req_mode;
 389         desc->l_granted_mode = lock->l_granted_mode;
 390         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 391         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 392 }
 393
 394 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 395                                    struct ldlm_lock *new,
 396                                    void *data, int datalen)
 397 {
 398         struct ldlm_ast_work *w;
 399         ENTRY;
 400
 401         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 402         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 403                 GOTO(out, 0);
 404
 405         OBD_ALLOC(w, sizeof(*w));
 406         if (!w) {
 407                 LBUG();
 408                 GOTO(out, 0);
 409         }
 410
 411         w->w_data = data;
 412         w->w_datalen = datalen;
 413         if (new) {
 414                 lock->l_flags |= LDLM_FL_AST_SENT;
 415                 w->w_blocking = 1;
 416                 ldlm_lock2desc(new, &w->w_desc);
 417         }
 418
 419         w->w_lock = LDLM_LOCK_GET(lock);
 420         list_add(&w->w_list, lock->l_resource->lr_tmp);
 421         EXIT;
 422  out:
 423         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 424         return;
 425 }
 426
 427 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 428 {
 429         struct ldlm_lock *lock;
 430
 431         lock = ldlm_handle2lock(lockh);
 432         ldlm_lock_addref_internal(lock, mode);
 433         LDLM_LOCK_PUT(lock);
 434 }
 435
 436 /* only called for local locks */
 437 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 438 {
 439         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 440         ldlm_lock_remove_from_lru(lock);
 441         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 442                 lock->l_readers++;
 443         else
 444                 lock->l_writers++;
 445         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 446         LDLM_LOCK_GET(lock);
 447         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 448 }
 449
 450 /* Args: unlocked lock */
 451 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 452                                     struct ldlm_res_id, int flags);
 453
 454 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
 455 {
 456         struct ldlm_namespace *ns;
 457         ENTRY;
 458
 459         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 460         ns = lock->l_resource->lr_namespace;
 461         l_lock(&ns->ns_lock);
 462         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR) {
 463                 LASSERT(lock->l_readers > 0);
 464                 lock->l_readers--;
 465         } else {
 466                 LASSERT(lock->l_writers > 0);
 467                 lock->l_writers--;
 468         }
 469
 470         if (lock->l_flags & LDLM_FL_LOCAL &&
 471             !lock->l_readers && !lock->l_writers) {
 472                 /* If this is a local lock on a server namespace and this was
 473                  * the last reference, cancel the lock. */
 474                 CDEBUG(D_INFO, "forcing cancel of local lock\n");
 475                 lock->l_flags |= LDLM_FL_CBPENDING;
 476         }
 477
 478         if (!lock->l_readers && !lock->l_writers &&
 479             (lock->l_flags & LDLM_FL_CBPENDING)) {
 480                 /* If we received a blocked AST and this was the last reference,
 481                  * run the callback. */
 482                 if (!ns->ns_client && lock->l_export)
 483                         CERROR("FL_CBPENDING set on non-local lock--just a "
 484                                "warning\n");
 485
 486                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 487
 488                 if (lock->l_blocking_ast == NULL) {
 489                         /* The lock wasn't even fully formed; just destroy it */
 490                         ldlm_lock_destroy(lock);
 491                 }
 492                 l_unlock(&ns->ns_lock);
 493
 494                 /* FIXME: need a real 'desc' here */
 495                 if (lock->l_blocking_ast != NULL)
 496                         lock->l_blocking_ast(lock, NULL, lock->l_data,
 497                                              LDLM_CB_BLOCKING);
 498         } else if (ns->ns_client && !lock->l_readers && !lock->l_writers) {
 499                 /* If this is a client-side namespace and this was the last
 500                  * reference, put it on the LRU. */
 501                 LASSERT(list_empty(&lock->l_lru));
 502                 LASSERT(ns->ns_nr_unused >= 0);
 503                 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 504                 ns->ns_nr_unused++;
 505                 l_unlock(&ns->ns_lock);
 506                 ldlm_cancel_lru(ns);
 507         } else {
 508                 l_unlock(&ns->ns_lock);
 509         }
 510
 511         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 512
 513         EXIT;
 514 }
 515
 516 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 517 {
 518         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 519         LASSERT(lock != NULL);
 520         ldlm_lock_decref_internal(lock, mode);
 521         LDLM_LOCK_PUT(lock);
 522 }
 523
 524 /* This will drop a lock reference and mark it for destruction, but will not
 525  * necessarily cancel the lock before returning. */
 526 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
 527 {
 528         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 529         ENTRY;
 530
 531         LASSERT(lock != NULL);
 532
 533         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 534         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 535         lock->l_flags |= LDLM_FL_CBPENDING;
 536         ldlm_lock_decref_internal(lock, mode);
 537         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 538         LDLM_LOCK_PUT(lock);
 539 }
 540
 541 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 542                                  struct list_head *queue)
 543 {
 544         struct list_head *tmp, *pos;
 545         int rc = 1;
 546
 547         list_for_each_safe(tmp, pos, queue) {
 548                 struct ldlm_lock *child;
 549                 ldlm_res_compat compat;
 550
 551                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 552                 if (lock == child)
 553                         continue;
 554
 555                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 556                 if (compat && compat(child, lock)) {
 557                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 558                         continue;
 559                 }
 560                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 561                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 562                         continue;
 563                 }
 564
 565                 rc = 0;
 566
 567                 if (send_cbs && child->l_blocking_ast != NULL) {
 568                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 569                                "blocking AST.\n", child);
 570                         ldlm_add_ast_work_item(child, lock, NULL, 0);
 571                 }
 572         }
 573
 574         return rc;
 575 }
 576
 577 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 578 {
 579         int rc;
 580         ENTRY;
 581
 582         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 583         rc = ldlm_lock_compat_list(lock, send_cbs,
 584                                    &lock->l_resource->lr_granted);
 585         /* FIXME: should we be sending ASTs to converting? */
 586         if (rc)
 587                 rc = ldlm_lock_compat_list
 588                         (lock, send_cbs, &lock->l_resource->lr_converting);
 589
 590         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 591         RETURN(rc);
 592 }
 593
 594 /* NOTE: called by
 595  *  - ldlm_lock_enqueue
 596  *  - ldlm_reprocess_queue
 597  *  - ldlm_lock_convert
 598  */
 599 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen)
 600 {
 601         struct ldlm_resource *res = lock->l_resource;
 602         ENTRY;
 603
 604         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 605         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 606         lock->l_granted_mode = lock->l_req_mode;
 607
 608         if (lock->l_granted_mode < res->lr_most_restr)
 609                 res->lr_most_restr = lock->l_granted_mode;
 610
 611         if (lock->l_completion_ast != NULL)
 612                 ldlm_add_ast_work_item(lock, NULL, data, datalen);
 613
 614         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 615         EXIT;
 616 }
 617
 618 /* returns a referenced lock or NULL.  See the flag descriptions below, in the
 619  * comment above ldlm_lock_match */
 620 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 621                                       struct ldlm_extent *extent,
 622                                       struct ldlm_lock *old_lock, int flags)
 623 {
 624         struct ldlm_lock *lock;
 625         struct list_head *tmp;
 626
 627         list_for_each(tmp, queue) {
 628                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 629
 630                 if (lock == old_lock)
 631                         break;
 632
 633                 if (lock->l_flags & LDLM_FL_CBPENDING)
 634                         continue;
 635
 636                 if (lock->l_req_mode != mode)
 637                         continue;
 638
 639                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 640                     (lock->l_extent.start > extent->start ||
 641                      lock->l_extent.end < extent->end))
 642                         continue;
 643
 644                 if (lock->l_destroyed)
 645                         continue;
 646
 647                 if ((flags & LDLM_FL_LOCAL_ONLY) &&
 648                     !(lock->l_flags & LDLM_FL_LOCAL))
 649                         continue;
 650
 651                 ldlm_lock_addref_internal(lock, mode);
 652                 return lock;
 653         }
 654
 655         return NULL;
 656 }
 657
 658 /* Can be called in two ways:
 659  *
 660  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
 661  * for a duplicate of.
 662  *
 663  * Otherwise, all of the fields must be filled in, to match against.
 664  *
 665  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
 666  *     server (ie, connh is NULL)
 667  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
 668  *     list will be considered
 669  *
 670  * Returns 1 if it finds an already-existing lock that is compatible; in this
 671  * case, lockh is filled in with a addref()ed lock
 672  */
 673 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
 674                     struct ldlm_res_id *res_id, __u32 type, void *cookie,
 675                     int cookielen, ldlm_mode_t mode,struct lustre_handle *lockh)
 676 {
 677         struct ldlm_resource *res;
 678         struct ldlm_lock *lock, *old_lock = NULL;
 679         int rc = 0;
 680         ENTRY;
 681
 682         if (ns == NULL) {
 683                 old_lock = ldlm_handle2lock(lockh);
 684                 LASSERT(old_lock);
 685
 686                 ns = old_lock->l_resource->lr_namespace;
 687                 res_id = &old_lock->l_resource->lr_name;
 688                 type = old_lock->l_resource->lr_type;
 689                 mode = old_lock->l_req_mode;
 690         }
 691
 692         res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
 693         if (res == NULL) {
 694                 LASSERT(old_lock == NULL);
 695                 RETURN(0);
 696         }
 697
 698         l_lock(&ns->ns_lock);
 699
 700         lock = search_queue(&res->lr_granted, mode, cookie, old_lock, flags);
 701         if (lock != NULL)
 702                 GOTO(out, rc = 1);
 703         if (flags & LDLM_FL_BLOCK_GRANTED)
 704                 GOTO(out, rc = 0);
 705         lock = search_queue(&res->lr_converting, mode, cookie, old_lock, flags);
 706         if (lock != NULL)
 707                 GOTO(out, rc = 1);
 708         lock = search_queue(&res->lr_waiting, mode, cookie, old_lock, flags);
 709         if (lock != NULL)
 710                 GOTO(out, rc = 1);
 711
 712         EXIT;
 713        out:
 714         ldlm_resource_putref(res);
 715         l_unlock(&ns->ns_lock);
 716
 717         if (lock) {
 718                 ldlm_lock2handle(lock, lockh);
 719                 if (lock->l_completion_ast)
 720                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC, NULL);
 721         }
 722         if (rc)
 723                 LDLM_DEBUG(lock, "matched");
 724         else
 725                 LDLM_DEBUG_NOLOCK("not matched");
 726
 727         if (old_lock)
 728                 LDLM_LOCK_PUT(old_lock);
 729
 730         return rc;
 731 }
 732
 733 /* Returns a referenced lock */
 734 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 735                                    struct lustre_handle *parent_lock_handle,
 736                                    struct ldlm_res_id res_id, __u32 type,
 737                                    ldlm_mode_t mode, void *data, void *cp_data)
 738 {
 739         struct ldlm_resource *res, *parent_res = NULL;
 740         struct ldlm_lock *lock, *parent_lock = NULL;
 741         ENTRY;
 742
 743         if (parent_lock_handle) {
 744                 parent_lock = ldlm_handle2lock(parent_lock_handle);
 745                 if (parent_lock)
 746                         parent_res = parent_lock->l_resource;
 747         }
 748
 749         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 750         if (res == NULL)
 751                 RETURN(NULL);
 752
 753         lock = ldlm_lock_new(parent_lock, res);
 754         ldlm_resource_putref(res);
 755         if (parent_lock != NULL)
 756                 LDLM_LOCK_PUT(parent_lock);
 757
 758         if (lock == NULL)
 759                 RETURN(NULL);
 760
 761         lock->l_req_mode = mode;
 762         lock->l_data = data;
 763         lock->l_cp_data = cp_data;
 764
 765         RETURN(lock);
 766 }
 767
 768 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
 769                                struct ldlm_lock **lockp,
 770                                void *cookie, int cookie_len,
 771                                int *flags,
 772                                ldlm_completion_callback completion,
 773                                ldlm_blocking_callback blocking)
 774 {
 775         struct ldlm_resource *res;
 776         struct ldlm_lock *lock = *lockp;
 777         int local;
 778         ldlm_res_policy policy;
 779         ENTRY;
 780
 781         res = lock->l_resource;
 782         lock->l_blocking_ast = blocking;
 783
 784         if (res->lr_type == LDLM_EXTENT)
 785                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 786
 787         /* policies are not executed on the client or during replay */
 788         local = res->lr_namespace->ns_client;
 789         if (!local && !(*flags & LDLM_FL_REPLAY) &&
 790             (policy = ldlm_res_policy_table[res->lr_type])) {
 791                 int rc;
 792                 rc = policy(ns, lockp, cookie, lock->l_req_mode, *flags, NULL);
 793                 if (rc == ELDLM_LOCK_CHANGED) {
 794                         res = lock->l_resource;
 795                         *flags |= LDLM_FL_LOCK_CHANGED;
 796                 } else if (rc == ELDLM_LOCK_REPLACED) {
 797                         /* The lock that was returned has already been granted,
 798                          * and placed into lockp.  Destroy the old one and our
 799                          * work here is done. */
 800                         ldlm_lock_destroy(lock);
 801                         LDLM_LOCK_PUT(lock);
 802                         *flags |= LDLM_FL_LOCK_CHANGED;
 803                         RETURN(0);
 804                 } else if (rc == ELDLM_LOCK_ABORTED) {
 805                         ldlm_lock_destroy(lock);
 806                         RETURN(rc);
 807                 }
 808         }
 809
 810         l_lock(&ns->ns_lock);
 811         if (local && lock->l_req_mode == lock->l_granted_mode) {
 812                 /* The server returned a blocked lock, but it was granted before
 813                  * we got a chance to actually enqueue it.  We don't need to do
 814                  * anything else. */
 815                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 816                             LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 817                 GOTO(out, ELDLM_OK);
 818         }
 819
 820         /* This distinction between local lock trees is very important; a client
 821          * namespace only has information about locks taken by that client, and
 822          * thus doesn't have enough information to decide for itself if it can
 823          * be granted (below).  In this case, we do exactly what the server
 824          * tells us to do, as dictated by the 'flags'.
 825          *
 826          * We do exactly the same thing during recovery, when the server is
 827          * more or less trusting the clients not to lie.
 828          *
 829          * FIXME (bug 268): Detect obvious lies by checking compatibility in
 830          * granted/converting queues. */
 831         ldlm_resource_unlink_lock(lock);
 832         if (local) {
 833                 if (*flags & LDLM_FL_BLOCK_CONV)
 834                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 835                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 836                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 837                 else
 838                         ldlm_grant_lock(lock, NULL, 0);
 839                 GOTO(out, ELDLM_OK);
 840         } else if (*flags & LDLM_FL_REPLAY) {
 841                 if (*flags & LDLM_FL_BLOCK_CONV) {
 842                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 843                         GOTO(out, ELDLM_OK);
 844                 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
 845                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 846                         GOTO(out, ELDLM_OK);
 847                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
 848                         ldlm_grant_lock(lock, NULL, 0);
 849                         GOTO(out, ELDLM_OK);
 850                 }
 851                 /* If no flags, fall through to normal enqueue path. */
 852         }
 853
 854         /* FIXME: We may want to optimize by checking lr_most_restr */
 855         if (!list_empty(&res->lr_converting)) {
 856                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 857                 *flags |= LDLM_FL_BLOCK_CONV;
 858                 GOTO(out, ELDLM_OK);
 859         }
 860         if (!list_empty(&res->lr_waiting)) {
 861                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 862                 *flags |= LDLM_FL_BLOCK_WAIT;
 863                 GOTO(out, ELDLM_OK);
 864         }
 865         if (!ldlm_lock_compat(lock, 0)) {
 866                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 867                 *flags |= LDLM_FL_BLOCK_GRANTED;
 868                 GOTO(out, ELDLM_OK);
 869         }
 870
 871         if (lock->l_granted_cb != NULL && lock->l_data != NULL) {
 872                 /* We just -know- */
 873                 struct ptlrpc_request *req = lock->l_data;
 874                 lock->l_granted_cb(lock, req->rq_repmsg, 0);
 875         }
 876         ldlm_grant_lock(lock, NULL, 0);
 877         EXIT;
 878       out:
 879         /* Don't set 'completion_ast' until here so that if the lock is granted
 880          * immediately we don't do an unnecessary completion call. */
 881         lock->l_completion_ast = completion;
 882         l_unlock(&ns->ns_lock);
 883         return ELDLM_OK;
 884 }
 885
 886 /* Must be called with namespace taken: queue is waiting or converting. */
 887 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 888                                 struct list_head *queue)
 889 {
 890         struct list_head *tmp, *pos;
 891         ENTRY;
 892
 893         list_for_each_safe(tmp, pos, queue) {
 894                 struct ldlm_lock *pending;
 895                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 896
 897                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 898
 899                 if (!ldlm_lock_compat(pending, 1))
 900                         RETURN(1);
 901
 902                 list_del_init(&pending->l_res_link);
 903                 ldlm_grant_lock(pending, NULL, 0);
 904         }
 905
 906         RETURN(0);
 907 }
 908
 909 int ldlm_run_ast_work(struct list_head *rpc_list)
 910 {
 911         struct list_head *tmp, *pos;
 912         int rc, retval = 0;
 913         ENTRY;
 914
 915         list_for_each_safe(tmp, pos, rpc_list) {
 916                 struct ldlm_ast_work *w =
 917                         list_entry(tmp, struct ldlm_ast_work, w_list);
 918
 919                 /* It's possible to receive a completion AST before we've set
 920                  * the l_completion_ast pointer: either because the AST arrived
 921                  * before the reply, or simply because there's a small race
 922                  * window between receiving the reply and finishing the local
 923                  * enqueue. (bug 842)
 924                  *
 925                  * This can't happen with the blocking_ast, however, because we
 926                  * will never call the local blocking_ast until we drop our
 927                  * reader/writer reference, which we won't do until we get the
 928                  * reply and finish enqueueing. */
 929                 if (w->w_blocking) {
 930                         LASSERT(w->w_lock->l_blocking_ast != NULL);
 931                         rc = w->w_lock->l_blocking_ast
 932                                 (w->w_lock, &w->w_desc, w->w_data,
 933                                  LDLM_CB_BLOCKING);
 934                 } else if (w->w_lock->l_completion_ast != NULL) {
 935                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags,
 936                                                          w->w_data);
 937                 } else {
 938                         rc = 0;
 939                 }
 940                 if (rc == -ERESTART)
 941                         retval = rc;
 942                 else if (rc)
 943                         CERROR("Failed AST - should clean & disconnect "
 944                                "client\n");
 945                 LDLM_LOCK_PUT(w->w_lock);
 946                 list_del(&w->w_list);
 947                 OBD_FREE(w, sizeof(*w));
 948         }
 949         RETURN(retval);
 950 }
 951
 952 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
 953 {
 954         ldlm_reprocess_all(res);
 955         return LDLM_ITER_CONTINUE;
 956 }
 957
 958 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
 959 {
 960         (void)ldlm_namespace_foreach_res(ns, reprocess_one_queue, NULL);
 961 }
 962
 963 void ldlm_reprocess_all(struct ldlm_resource *res)
 964 {
 965         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 966         int rc;
 967         ENTRY;
 968
 969         /* Local lock trees don't get reprocessed. */
 970         if (res->lr_namespace->ns_client) {
 971                 EXIT;
 972                 return;
 973         }
 974
 975  restart:
 976         l_lock(&res->lr_namespace->ns_lock);
 977         res->lr_tmp = &rpc_list;
 978
 979         ldlm_reprocess_queue(res, &res->lr_converting);
 980         if (list_empty(&res->lr_converting))
 981                 ldlm_reprocess_queue(res, &res->lr_waiting);
 982
 983         res->lr_tmp = NULL;
 984         l_unlock(&res->lr_namespace->ns_lock);
 985
 986         rc = ldlm_run_ast_work(&rpc_list);
 987         if (rc == -ERESTART)
 988                 goto restart;
 989         EXIT;
 990 }
 991
 992 void ldlm_cancel_callback(struct ldlm_lock *lock)
 993 {
 994         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 995         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
 996                 lock->l_flags |= LDLM_FL_CANCEL;
 997                 if (lock->l_blocking_ast)
 998                         lock->l_blocking_ast(lock, NULL, lock->l_data,
 999                                              LDLM_CB_CANCELING);
1000                 else
1001                         LDLM_DEBUG(lock, "no blocking ast");
1002         }
1003         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1004 }
1005
1006 void ldlm_lock_cancel(struct ldlm_lock *lock)
1007 {
1008         struct ldlm_resource *res;
1009         struct ldlm_namespace *ns;
1010         ENTRY;
1011
1012         ldlm_del_waiting_lock(lock);
1013
1014         res = lock->l_resource;
1015         ns = res->lr_namespace;
1016
1017         l_lock(&ns->ns_lock);
1018         /* Please do not, no matter how tempting, remove this LBUG without
1019          * talking to me first. -phik */
1020         if (lock->l_readers || lock->l_writers) {
1021                 LDLM_DEBUG(lock, "lock still has references");
1022                 ldlm_lock_dump(D_OTHER, lock);
1023                 LBUG();
1024         }
1025
1026         ldlm_cancel_callback(lock);
1027
1028         ldlm_resource_unlink_lock(lock);
1029         ldlm_lock_destroy(lock);
1030         l_unlock(&ns->ns_lock);
1031         EXIT;
1032 }
1033
1034 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data, void *cp_data)
1035 {
1036         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1037         ENTRY;
1038
1039         if (lock == NULL)
1040                 RETURN(-EINVAL);
1041
1042         lock->l_data = data;
1043         lock->l_cp_data = cp_data;
1044
1045         LDLM_LOCK_PUT(lock);
1046
1047         RETURN(0);
1048 }
1049
1050 /* This function is only called from one thread (per export); no locking around
1051  * the list ops needed */
1052 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1053 {
1054         struct list_head *iter, *n;
1055
1056         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
1057                 struct ldlm_lock *lock;
1058                 struct ldlm_resource *res;
1059                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1060                 res = ldlm_resource_getref(lock->l_resource);
1061                 LDLM_DEBUG(lock, "export %p", exp);
1062                 ldlm_lock_cancel(lock);
1063                 ldlm_reprocess_all(res);
1064                 ldlm_resource_putref(res);
1065         }
1066 }
1067
1068 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1069                                         int *flags)
1070 {
1071         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1072         struct ldlm_resource *res;
1073         struct ldlm_namespace *ns;
1074         int granted = 0;
1075         ENTRY;
1076
1077         LBUG();
1078
1079         res = lock->l_resource;
1080         ns = res->lr_namespace;
1081
1082         l_lock(&ns->ns_lock);
1083
1084         lock->l_req_mode = new_mode;
1085         ldlm_resource_unlink_lock(lock);
1086
1087         /* If this is a local resource, put it on the appropriate list. */
1088         if (res->lr_namespace->ns_client) {
1089                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1090                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1091                 } else {
1092                         /* This should never happen, because of the way the
1093                          * server handles conversions. */
1094                         LBUG();
1095
1096                         res->lr_tmp = &rpc_list;
1097                         ldlm_grant_lock(lock, NULL, 0);
1098                         res->lr_tmp = NULL;
1099                         granted = 1;
1100                         /* FIXME: completion handling not with ns_lock held ! */
1101                         if (lock->l_completion_ast)
1102                                 lock->l_completion_ast(lock, 0, NULL);
1103                 }
1104         } else {
1105                 /* FIXME: We should try the conversion right away and possibly
1106                  * return success without the need for an extra AST */
1107                 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1108                 *flags |= LDLM_FL_BLOCK_CONV;
1109         }
1110
1111         l_unlock(&ns->ns_lock);
1112
1113         if (granted)
1114                 ldlm_run_ast_work(&rpc_list);
1115         RETURN(res);
1116 }
1117
1118 void ldlm_lock_dump(int level, struct ldlm_lock *lock)
1119 {
1120         char ver[128];
1121
1122         if (!((portal_debug | D_ERROR) & level))
1123                 return;
1124
1125         if (RES_VERSION_SIZE != 4)
1126                 LBUG();
1127
1128         if (!lock) {
1129                 CDEBUG(level, "  NULL LDLM lock\n");
1130                 return;
1131         }
1132
1133         snprintf(ver, sizeof(ver), "%x %x %x %x",
1134                  lock->l_version[0], lock->l_version[1],
1135                  lock->l_version[2], lock->l_version[3]);
1136
1137         CDEBUG(level, "  -- Lock dump: %p (%s) (rc: %d)\n", lock, ver,
1138                atomic_read(&lock->l_refc));
1139         if (lock->l_export && lock->l_export->exp_connection)
1140                 CDEBUG(level, "  Node: NID "LPX64" on %s (rhandle: "LPX64")\n",
1141                        lock->l_export->exp_connection->c_peer.peer_nid,
1142                        lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1143                        lock->l_remote_handle.cookie);
1144         else
1145                 CDEBUG(level, "  Node: local\n");
1146         CDEBUG(level, "  Parent: %p\n", lock->l_parent);
1147         CDEBUG(level, "  Resource: %p ("LPD64")\n", lock->l_resource,
1148                lock->l_resource->lr_name.name[0]);
1149         CDEBUG(level, "  Requested mode: %d, granted mode: %d\n",
1150                (int)lock->l_req_mode, (int)lock->l_granted_mode);
1151         CDEBUG(level, "  Readers: %u ; Writers; %u\n",
1152                lock->l_readers, lock->l_writers);
1153         if (lock->l_resource->lr_type == LDLM_EXTENT)
1154                 CDEBUG(level, "  Extent: "LPU64" -> "LPU64"\n",
1155                        lock->l_extent.start, lock->l_extent.end);
1156 }
1157
1158 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1159 {
1160         struct ldlm_lock *lock;
1161
1162         lock = ldlm_handle2lock(lockh);
1163         if (lock == NULL)
1164                 return;
1165
1166         ldlm_lock_dump(D_OTHER, lock);
1167
1168         LDLM_LOCK_PUT(lock);
1169 }