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 #else
  31 # include <liblustre.h>
  32 # include <linux/kp30.h>
  33 #endif
  34
  35 #include <linux/obd_class.h>
  36 #include "ldlm_internal.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         case IT_GETXATTR:
  74                 return "getxattr";
  75         default:
  76                 CERROR("Unknown intent %d\n", it);
  77                 return "UNKNOWN";
  78         }
  79 }
  80 #endif
  81
  82 extern kmem_cache_t *ldlm_lock_slab;
  83 struct lustre_lock ldlm_handle_lock;
  84
  85 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b);
  86
  87 static ldlm_res_compat ldlm_res_compat_table[] = {
  88         [LDLM_PLAIN] ldlm_plain_compat,
  89         [LDLM_EXTENT] ldlm_extent_compat,
  90 };
  91
  92 static ldlm_res_policy ldlm_intent_policy_func;
  93
  94 static int ldlm_plain_policy(struct ldlm_namespace *ns, struct ldlm_lock **lock,
  95                              void *req_cookie, ldlm_mode_t mode, int flags,
  96                              void *data)
  97 {
  98         if ((flags & LDLM_FL_HAS_INTENT) && ldlm_intent_policy_func) {
  99                 return ldlm_intent_policy_func(ns, lock, req_cookie, mode,
 100                                                flags, data);
 101         }
 102
 103         return ELDLM_OK;
 104 }
 105
 106 static ldlm_res_policy ldlm_res_policy_table[] = {
 107         [LDLM_PLAIN] ldlm_plain_policy,
 108         [LDLM_EXTENT] ldlm_extent_policy,
 109 };
 110
 111 void ldlm_register_intent(ldlm_res_policy arg)
 112 {
 113         ldlm_intent_policy_func = arg;
 114 }
 115
 116 void ldlm_unregister_intent(void)
 117 {
 118         ldlm_intent_policy_func = NULL;
 119 }
 120
 121 /*
 122  * REFCOUNTED LOCK OBJECTS
 123  */
 124
 125
 126 /*
 127  * Lock refcounts, during creation:
 128  *   - one special one for allocation, dec'd only once in destroy
 129  *   - one for being a lock that's in-use
 130  *   - one for the addref associated with a new lock
 131  */
 132 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 133 {
 134         atomic_inc(&lock->l_refc);
 135         return lock;
 136 }
 137
 138 void ldlm_lock_put(struct ldlm_lock *lock)
 139 {
 140         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 141         ENTRY;
 142
 143         if (atomic_dec_and_test(&lock->l_refc)) {
 144                 l_lock(&ns->ns_lock);
 145                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
 146                 LASSERT(lock->l_destroyed);
 147                 LASSERT(list_empty(&lock->l_res_link));
 148
 149                 spin_lock(&ns->ns_counter_lock);
 150                 ns->ns_locks--;
 151                 spin_unlock(&ns->ns_counter_lock);
 152
 153                 ldlm_resource_putref(lock->l_resource);
 154                 lock->l_resource = NULL;
 155                 if (lock->l_export)
 156                         class_export_put(lock->l_export);
 157
 158                 if (lock->l_parent)
 159                         LDLM_LOCK_PUT(lock->l_parent);
 160
 161                 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
 162                 l_unlock(&ns->ns_lock);
 163         }
 164
 165         EXIT;
 166 }
 167
 168 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
 169 {
 170         ENTRY;
 171         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 172         if (!list_empty(&lock->l_lru)) {
 173                 list_del_init(&lock->l_lru);
 174                 lock->l_resource->lr_namespace->ns_nr_unused--;
 175                 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
 176         }
 177         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 178         EXIT;
 179 }
 180
 181 /* This used to have a 'strict' flact, which recovery would use to mark an
 182  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
 183  * shall explain why it's gone: with the new hash table scheme, once you call
 184  * ldlm_lock_destroy, you can never drop your final references on this lock.
 185  * Because it's not in the hash table anymore.  -phil */
 186 void ldlm_lock_destroy(struct ldlm_lock *lock)
 187 {
 188         ENTRY;
 189         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 190
 191         if (!list_empty(&lock->l_children)) {
 192                 LDLM_ERROR(lock, "still has children (%p)!",
 193                            lock->l_children.next);
 194                 ldlm_lock_dump(D_ERROR, lock, 0);
 195                 LBUG();
 196         }
 197         if (lock->l_readers || lock->l_writers) {
 198                 LDLM_ERROR(lock, "lock still has references");
 199                 ldlm_lock_dump(D_ERROR, lock, 0);
 200                 LBUG();
 201         }
 202
 203         if (!list_empty(&lock->l_res_link)) {
 204                 ldlm_lock_dump(D_ERROR, lock, 0);
 205                 LBUG();
 206         }
 207
 208         if (lock->l_destroyed) {
 209                 LASSERT(list_empty(&lock->l_lru));
 210                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 211                 EXIT;
 212                 return;
 213         }
 214         lock->l_destroyed = 1;
 215
 216         list_del_init(&lock->l_export_chain);
 217         ldlm_lock_remove_from_lru(lock);
 218         class_handle_unhash(&lock->l_handle);
 219
 220 #if 0
 221         /* Wake anyone waiting for this lock */
 222         /* FIXME: I should probably add yet another flag, instead of using
 223          * l_export to only call this on clients */
 224         if (lock->l_export)
 225                 class_export_put(lock->l_export);
 226         lock->l_export = NULL;
 227         if (lock->l_export && lock->l_completion_ast)
 228                 lock->l_completion_ast(lock, 0);
 229 #endif
 230
 231         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 232         LDLM_LOCK_PUT(lock);
 233         EXIT;
 234 }
 235
 236 /* this is called by portals_handle2object with the handle lock taken */
 237 static void lock_handle_addref(void *lock)
 238 {
 239         LDLM_LOCK_GET((struct ldlm_lock *)lock);
 240 }
 241
 242 /*
 243  * usage: pass in a resource on which you have done ldlm_resource_get
 244  *        pass in a parent lock on which you have done a ldlm_lock_get
 245  *        after return, ldlm_*_put the resource and parent
 246  * returns: lock with refcount 1
 247  */
 248 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 249                                        struct ldlm_resource *resource)
 250 {
 251         struct ldlm_lock *lock;
 252         ENTRY;
 253
 254         if (resource == NULL)
 255                 LBUG();
 256
 257         OBD_SLAB_ALLOC(lock, ldlm_lock_slab, SLAB_KERNEL, sizeof(*lock));
 258         if (lock == NULL)
 259                 RETURN(NULL);
 260
 261         lock->l_resource = ldlm_resource_getref(resource);
 262
 263         atomic_set(&lock->l_refc, 2);
 264         INIT_LIST_HEAD(&lock->l_children);
 265         INIT_LIST_HEAD(&lock->l_res_link);
 266         INIT_LIST_HEAD(&lock->l_lru);
 267         INIT_LIST_HEAD(&lock->l_export_chain);
 268         INIT_LIST_HEAD(&lock->l_pending_chain);
 269         init_waitqueue_head(&lock->l_waitq);
 270
 271         spin_lock(&resource->lr_namespace->ns_counter_lock);
 272         resource->lr_namespace->ns_locks++;
 273         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 274
 275         if (parent != NULL) {
 276                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 277                 lock->l_parent = LDLM_LOCK_GET(parent);
 278                 list_add(&lock->l_childof, &parent->l_children);
 279                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 280         }
 281
 282         INIT_LIST_HEAD(&lock->l_handle.h_link);
 283         class_handle_hash(&lock->l_handle, lock_handle_addref);
 284
 285         RETURN(lock);
 286 }
 287
 288 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
 289                               struct ldlm_res_id new_resid)
 290 {
 291         struct ldlm_resource *oldres = lock->l_resource;
 292         ENTRY;
 293
 294         l_lock(&ns->ns_lock);
 295         if (memcmp(&new_resid, &lock->l_resource->lr_name,
 296                    sizeof(lock->l_resource->lr_name)) == 0) {
 297                 /* Nothing to do */
 298                 l_unlock(&ns->ns_lock);
 299                 RETURN(0);
 300         }
 301
 302         LASSERT(new_resid.name[0] != 0);
 303
 304         /* This function assumes that the lock isn't on any lists */
 305         LASSERT(list_empty(&lock->l_res_link));
 306
 307         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
 308                                              lock->l_resource->lr_type, 1);
 309         if (lock->l_resource == NULL) {
 310                 LBUG();
 311                 RETURN(-ENOMEM);
 312         }
 313
 314         /* ...and the flowers are still standing! */
 315         ldlm_resource_putref(oldres);
 316
 317         l_unlock(&ns->ns_lock);
 318         RETURN(0);
 319 }
 320
 321 /*
 322  *  HANDLES
 323  */
 324
 325 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 326 {
 327         lockh->cookie = lock->l_handle.h_cookie;
 328 }
 329
 330 /* if flags: atomically get the lock and set the flags.
 331  *           Return NULL if flag already set
 332  */
 333
 334 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
 335 {
 336         struct ldlm_namespace *ns;
 337         struct ldlm_lock *lock = NULL, *retval = NULL;
 338         ENTRY;
 339
 340         LASSERT(handle);
 341
 342         lock = class_handle2object(handle->cookie);
 343         if (lock == NULL)
 344                 RETURN(NULL);
 345
 346         LASSERT(lock->l_resource != NULL);
 347         ns = lock->l_resource->lr_namespace;
 348         LASSERT(ns != NULL);
 349
 350         l_lock(&ns->ns_lock);
 351
 352         /* It's unlikely but possible that someone marked the lock as
 353          * destroyed after we did handle2object on it */
 354         if (lock->l_destroyed) {
 355                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
 356                 LDLM_LOCK_PUT(lock);
 357                 GOTO(out, retval);
 358         }
 359
 360         if (flags && (lock->l_flags & flags)) {
 361                 LDLM_LOCK_PUT(lock);
 362                 GOTO(out, retval);
 363         }
 364
 365         if (flags)
 366                 lock->l_flags |= flags;
 367
 368         retval = lock;
 369         EXIT;
 370  out:
 371         l_unlock(&ns->ns_lock);
 372         return retval;
 373 }
 374
 375 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
 376                                       struct lustre_handle *handle)
 377 {
 378         struct ldlm_lock *retval = NULL;
 379
 380         l_lock(&ns->ns_lock);
 381         retval = __ldlm_handle2lock(handle, 0);
 382         l_unlock(&ns->ns_lock);
 383
 384         return retval;
 385 }
 386
 387 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 388 {
 389         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 390 }
 391
 392 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 393 {
 394         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 395         desc->l_req_mode = lock->l_req_mode;
 396         desc->l_granted_mode = lock->l_granted_mode;
 397         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 398         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 399 }
 400
 401 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 402                                    struct ldlm_lock *new,
 403                                    void *data, int datalen)
 404 {
 405         struct ldlm_ast_work *w;
 406         ENTRY;
 407
 408         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 409         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 410                 GOTO(out, 0);
 411
 412         CDEBUG(D_OTHER, "lock %p incompatible; sending blocking AST.\n", lock);
 413
 414         OBD_ALLOC(w, sizeof(*w));
 415         if (!w) {
 416                 LBUG();
 417                 GOTO(out, 0);
 418         }
 419
 420         w->w_data = data;
 421         w->w_datalen = datalen;
 422         if (new) {
 423                 lock->l_flags |= LDLM_FL_AST_SENT;
 424                 w->w_blocking = 1;
 425                 ldlm_lock2desc(new, &w->w_desc);
 426         }
 427
 428         w->w_lock = LDLM_LOCK_GET(lock);
 429         list_add(&w->w_list, lock->l_resource->lr_tmp);
 430         EXIT;
 431  out:
 432         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 433         return;
 434 }
 435
 436 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 437 {
 438         struct ldlm_lock *lock;
 439
 440         lock = ldlm_handle2lock(lockh);
 441         ldlm_lock_addref_internal(lock, mode);
 442         LDLM_LOCK_PUT(lock);
 443 }
 444
 445 /* only called for local locks */
 446 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 447 {
 448         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 449         ldlm_lock_remove_from_lru(lock);
 450         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 451                 lock->l_readers++;
 452         else
 453                 lock->l_writers++;
 454         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 455         LDLM_LOCK_GET(lock);
 456         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 457 }
 458
 459 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
 460 {
 461         struct ldlm_namespace *ns;
 462         ENTRY;
 463
 464         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 465         ns = lock->l_resource->lr_namespace;
 466         l_lock(&ns->ns_lock);
 467         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR) {
 468                 LASSERT(lock->l_readers > 0);
 469                 lock->l_readers--;
 470         } else {
 471                 LASSERT(lock->l_writers > 0);
 472                 lock->l_writers--;
 473         }
 474
 475         if (lock->l_flags & LDLM_FL_LOCAL &&
 476             !lock->l_readers && !lock->l_writers) {
 477                 /* If this is a local lock on a server namespace and this was
 478                  * the last reference, cancel the lock. */
 479                 CDEBUG(D_INFO, "forcing cancel of local lock\n");
 480                 lock->l_flags |= LDLM_FL_CBPENDING;
 481         }
 482
 483         if (!lock->l_readers && !lock->l_writers &&
 484             (lock->l_flags & LDLM_FL_CBPENDING)) {
 485                 /* If we received a blocked AST and this was the last reference,
 486                  * run the callback. */
 487                 if (!ns->ns_client && lock->l_export)
 488                         CERROR("FL_CBPENDING set on non-local lock--just a "
 489                                "warning\n");
 490
 491                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 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
 499                         LDLM_DEBUG(lock, "No blocking AST?");
 500         } else if (ns->ns_client && !lock->l_readers && !lock->l_writers) {
 501                 /* If this is a client-side namespace and this was the last
 502                  * reference, put it on the LRU. */
 503                 LASSERT(list_empty(&lock->l_lru));
 504                 LASSERT(ns->ns_nr_unused >= 0);
 505                 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 506                 ns->ns_nr_unused++;
 507                 l_unlock(&ns->ns_lock);
 508                 ldlm_cancel_lru(ns);
 509         } else {
 510                 l_unlock(&ns->ns_lock);
 511         }
 512
 513         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 514
 515         EXIT;
 516 }
 517
 518 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 519 {
 520         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 521         LASSERT(lock != NULL);
 522         ldlm_lock_decref_internal(lock, mode);
 523         LDLM_LOCK_PUT(lock);
 524 }
 525
 526 /* This will drop a lock reference and mark it for destruction, but will not
 527  * necessarily cancel the lock before returning. */
 528 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
 529 {
 530         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
 531         ENTRY;
 532
 533         LASSERT(lock != NULL);
 534
 535         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 536         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 537         lock->l_flags |= LDLM_FL_CBPENDING;
 538         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 539         ldlm_lock_decref_internal(lock, mode);
 540         LDLM_LOCK_PUT(lock);
 541 }
 542
 543 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 544                                  struct list_head *queue)
 545 {
 546         struct list_head *tmp, *pos;
 547         int rc = 1;
 548
 549         list_for_each_safe(tmp, pos, queue) {
 550                 struct ldlm_lock *child;
 551                 ldlm_res_compat compat;
 552
 553                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 554                 if (lock == child)
 555                         continue;
 556
 557                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 558                 if (compat && compat(child, lock)) {
 559                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 560                         continue;
 561                 }
 562                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 563                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 564                         continue;
 565                 }
 566
 567                 rc = 0;
 568
 569                 if (send_cbs && child->l_blocking_ast != NULL)
 570                         ldlm_add_ast_work_item(child, lock, NULL, 0);
 571         }
 572
 573         return rc;
 574 }
 575
 576 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 577 {
 578         int rc;
 579         ENTRY;
 580
 581         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 582         rc = ldlm_lock_compat_list(lock, send_cbs,
 583                                    &lock->l_resource->lr_granted);
 584         /* FIXME: should we be sending ASTs to converting? */
 585         if (rc)
 586                 rc = ldlm_lock_compat_list
 587                         (lock, send_cbs, &lock->l_resource->lr_converting);
 588
 589         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 590         RETURN(rc);
 591 }
 592
 593 /* NOTE: called by
 594  *  - ldlm_lock_enqueue
 595  *  - ldlm_reprocess_queue
 596  *  - ldlm_lock_convert
 597  */
 598 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen,
 599                      int run_ast)
 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 (run_ast && 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, void *data,
 623                                       int flags)
 624 {
 625         struct ldlm_lock *lock;
 626         struct list_head *tmp;
 627
 628         list_for_each(tmp, queue) {
 629                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 630
 631                 if (lock == old_lock)
 632                         break;
 633
 634                 /* llite sometimes wants to match locks that will be
 635                  * canceled when their users drop, but we allow it to match
 636                  * if it passes in CBPENDING and the lock still has users.
 637                  * this is generally only going to be used by children
 638                  * whose parents already hold a lock so forward progress
 639                  * can still happen. */
 640                 if (lock->l_flags & LDLM_FL_CBPENDING &&
 641                     !(flags & LDLM_FL_CBPENDING))
 642                         continue;
 643                 if (lock->l_flags & LDLM_FL_CBPENDING &&
 644                     lock->l_readers == 0 && lock->l_writers == 0)
 645                         continue;
 646
 647                 if (lock->l_req_mode != mode)
 648                         continue;
 649
 650                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 651                     (lock->l_extent.start > extent->start ||
 652                      lock->l_extent.end < extent->end))
 653                         continue;
 654
 655                 if (lock->l_destroyed)
 656                         continue;
 657
 658                 if ((flags & LDLM_FL_LOCAL_ONLY) &&
 659                     !(lock->l_flags & LDLM_FL_LOCAL))
 660                         continue;
 661
 662                 if ((flags & LDLM_FL_MATCH_DATA) && lock->l_data != data) {
 663                         LDLM_DEBUG(lock, "data mismatch: have %p, want %p",
 664                                    lock->l_data, data);
 665                         continue;
 666                 }
 667
 668                 ldlm_lock_addref_internal(lock, mode);
 669                 return lock;
 670         }
 671
 672         return NULL;
 673 }
 674
 675 /* Can be called in two ways:
 676  *
 677  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
 678  * for a duplicate of.
 679  *
 680  * Otherwise, all of the fields must be filled in, to match against.
 681  *
 682  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
 683  *     server (ie, connh is NULL)
 684  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
 685  *     list will be considered
 686  * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
 687  *     to be canceled can still be matched as long as they still have reader
 688  *     or writer refernces
 689  * If 'flags' contains LDLM_FL_MATCH_DATA, then only match a lock if the opaque
 690  *     data is the same.
 691  *
 692  * Returns 1 if it finds an already-existing lock that is compatible; in this
 693  * case, lockh is filled in with a addref()ed lock
 694  */
 695 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
 696                     struct ldlm_res_id *res_id, __u32 type, void *cookie,
 697                     int cookielen, ldlm_mode_t mode, void *data,
 698                     struct lustre_handle *lockh)
 699 {
 700         struct ldlm_resource *res;
 701         struct ldlm_lock *lock, *old_lock = NULL;
 702         int rc = 0;
 703         ENTRY;
 704
 705         if (ns == NULL) {
 706                 old_lock = ldlm_handle2lock(lockh);
 707                 LASSERT(old_lock);
 708
 709                 ns = old_lock->l_resource->lr_namespace;
 710                 res_id = &old_lock->l_resource->lr_name;
 711                 type = old_lock->l_resource->lr_type;
 712                 mode = old_lock->l_req_mode;
 713         }
 714
 715         res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
 716         if (res == NULL) {
 717                 LASSERT(old_lock == NULL);
 718                 RETURN(0);
 719         }
 720
 721         l_lock(&ns->ns_lock);
 722
 723         lock = search_queue(&res->lr_granted, mode, cookie, old_lock, data,
 724                             flags);
 725         if (lock != NULL)
 726                 GOTO(out, rc = 1);
 727         if (flags & LDLM_FL_BLOCK_GRANTED)
 728                 GOTO(out, rc = 0);
 729         lock = search_queue(&res->lr_converting, mode, cookie, old_lock, data,
 730                             flags);
 731         if (lock != NULL)
 732                 GOTO(out, rc = 1);
 733         lock = search_queue(&res->lr_waiting, mode, cookie, old_lock, data,
 734                             flags);
 735         if (lock != NULL)
 736                 GOTO(out, rc = 1);
 737
 738         EXIT;
 739  out:
 740         ldlm_resource_putref(res);
 741         l_unlock(&ns->ns_lock);
 742
 743         if (lock) {
 744                 ldlm_lock2handle(lock, lockh);
 745                 if (lock->l_completion_ast)
 746                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 747                                                NULL);
 748         }
 749         if (rc)
 750                 LDLM_DEBUG(lock, "matched");
 751         else
 752                 LDLM_DEBUG_NOLOCK("not matched");
 753
 754         if (old_lock)
 755                 LDLM_LOCK_PUT(old_lock);
 756
 757         return rc;
 758 }
 759
 760 /* Returns a referenced lock */
 761 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 762                                    struct lustre_handle *parent_lock_handle,
 763                                    struct ldlm_res_id res_id, __u32 type,
 764                                    ldlm_mode_t mode,
 765                                    ldlm_blocking_callback blocking,
 766                                    ldlm_completion_callback completion,
 767                                    void *data)
 768 {
 769         struct ldlm_resource *res, *parent_res = NULL;
 770         struct ldlm_lock *lock, *parent_lock = NULL;
 771         ENTRY;
 772
 773         if (parent_lock_handle) {
 774                 parent_lock = ldlm_handle2lock(parent_lock_handle);
 775                 if (parent_lock)
 776                         parent_res = parent_lock->l_resource;
 777         }
 778
 779         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 780         if (res == NULL)
 781                 RETURN(NULL);
 782
 783         lock = ldlm_lock_new(parent_lock, res);
 784         ldlm_resource_putref(res);
 785         if (parent_lock != NULL)
 786                 LDLM_LOCK_PUT(parent_lock);
 787
 788         if (lock == NULL)
 789                 RETURN(NULL);
 790
 791         lock->l_req_mode = mode;
 792         lock->l_data = data;
 793         lock->l_blocking_ast = blocking;
 794         lock->l_completion_ast = completion;
 795
 796         RETURN(lock);
 797 }
 798
 799 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
 800                                struct ldlm_lock **lockp,
 801                                void *cookie, int cookie_len, int *flags)
 802 {
 803         struct ldlm_resource *res;
 804         struct ldlm_lock *lock = *lockp;
 805         int local;
 806         ldlm_res_policy policy;
 807         ENTRY;
 808
 809         res = lock->l_resource;
 810
 811         if (res->lr_type == LDLM_EXTENT)
 812                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 813
 814         /* policies are not executed on the client or during replay */
 815         local = res->lr_namespace->ns_client;
 816         if (!local && !(*flags & LDLM_FL_REPLAY) &&
 817             (policy = ldlm_res_policy_table[res->lr_type])) {
 818                 int rc;
 819                 rc = policy(ns, lockp, cookie, lock->l_req_mode, *flags, NULL);
 820                 if (rc == ELDLM_LOCK_CHANGED) {
 821                         res = lock->l_resource;
 822                         *flags |= LDLM_FL_LOCK_CHANGED;
 823                 } else if (rc == ELDLM_LOCK_REPLACED) {
 824                         /* The lock that was returned has already been granted,
 825                          * and placed into lockp.  Destroy the old one and our
 826                          * work here is done. */
 827                         ldlm_lock_destroy(lock);
 828                         LDLM_LOCK_PUT(lock);
 829                         *flags |= LDLM_FL_LOCK_CHANGED;
 830                         RETURN(0);
 831                 } else if (rc == ELDLM_LOCK_ABORTED) {
 832                         ldlm_lock_destroy(lock);
 833                         RETURN(rc);
 834                 }
 835         }
 836
 837         l_lock(&ns->ns_lock);
 838         if (local && lock->l_req_mode == lock->l_granted_mode) {
 839                 /* The server returned a blocked lock, but it was granted before
 840                  * we got a chance to actually enqueue it.  We don't need to do
 841                  * anything else. */
 842                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 843                             LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 844                 GOTO(out, ELDLM_OK);
 845         }
 846
 847         /* This distinction between local lock trees is very important; a client
 848          * namespace only has information about locks taken by that client, and
 849          * thus doesn't have enough information to decide for itself if it can
 850          * be granted (below).  In this case, we do exactly what the server
 851          * tells us to do, as dictated by the 'flags'.
 852          *
 853          * We do exactly the same thing during recovery, when the server is
 854          * more or less trusting the clients not to lie.
 855          *
 856          * FIXME (bug 268): Detect obvious lies by checking compatibility in
 857          * granted/converting queues. */
 858         ldlm_resource_unlink_lock(lock);
 859         if (local) {
 860                 if (*flags & LDLM_FL_BLOCK_CONV)
 861                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 862                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 863                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 864                 else
 865                         ldlm_grant_lock(lock, NULL, 0, 0);
 866                 GOTO(out, ELDLM_OK);
 867         } else if (*flags & LDLM_FL_REPLAY) {
 868                 if (*flags & LDLM_FL_BLOCK_CONV) {
 869                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
 870                         GOTO(out, ELDLM_OK);
 871                 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
 872                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 873                         GOTO(out, ELDLM_OK);
 874                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
 875                         ldlm_grant_lock(lock, NULL, 0, 0);
 876                         GOTO(out, ELDLM_OK);
 877                 }
 878                 /* If no flags, fall through to normal enqueue path. */
 879         }
 880
 881         /* FIXME: We may want to optimize by checking lr_most_restr */
 882         if (!list_empty(&res->lr_converting)) {
 883                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 884                 *flags |= LDLM_FL_BLOCK_CONV;
 885                 GOTO(out, ELDLM_OK);
 886         }
 887         if (!list_empty(&res->lr_waiting)) {
 888                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 889                 *flags |= LDLM_FL_BLOCK_WAIT;
 890                 GOTO(out, ELDLM_OK);
 891         }
 892         if (!ldlm_lock_compat(lock, 0)) {
 893                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
 894                 *flags |= LDLM_FL_BLOCK_GRANTED;
 895                 GOTO(out, ELDLM_OK);
 896         }
 897         ldlm_grant_lock(lock, NULL, 0, 0);
 898         EXIT;
 899       out:
 900         l_unlock(&ns->ns_lock);
 901         return ELDLM_OK;
 902 }
 903
 904 /* Must be called with namespace taken: queue is waiting or converting. */
 905 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 906                                 struct list_head *queue)
 907 {
 908         struct list_head *tmp, *pos;
 909         ENTRY;
 910
 911         list_for_each_safe(tmp, pos, queue) {
 912                 struct ldlm_lock *pending;
 913                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 914
 915                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 916
 917                 if (!ldlm_lock_compat(pending, 1))
 918                         RETURN(1);
 919
 920                 list_del_init(&pending->l_res_link);
 921                 ldlm_grant_lock(pending, NULL, 0, 1);
 922         }
 923
 924         RETURN(0);
 925 }
 926
 927 int ldlm_run_ast_work(struct list_head *rpc_list)
 928 {
 929         struct list_head *tmp, *pos;
 930         int rc, retval = 0;
 931         ENTRY;
 932
 933         list_for_each_safe(tmp, pos, rpc_list) {
 934                 struct ldlm_ast_work *w =
 935                         list_entry(tmp, struct ldlm_ast_work, w_list);
 936
 937                 /* It's possible to receive a completion AST before we've set
 938                  * the l_completion_ast pointer: either because the AST arrived
 939                  * before the reply, or simply because there's a small race
 940                  * window between receiving the reply and finishing the local
 941                  * enqueue. (bug 842)
 942                  *
 943                  * This can't happen with the blocking_ast, however, because we
 944                  * will never call the local blocking_ast until we drop our
 945                  * reader/writer reference, which we won't do until we get the
 946                  * reply and finish enqueueing. */
 947                 if (w->w_blocking) {
 948                         LASSERT(w->w_lock->l_blocking_ast != NULL);
 949                         rc = w->w_lock->l_blocking_ast
 950                                 (w->w_lock, &w->w_desc, w->w_data,
 951                                  LDLM_CB_BLOCKING);
 952                 } else if (w->w_lock->l_completion_ast != NULL) {
 953                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags,
 954                                                          w->w_data);
 955                 } else {
 956                         rc = 0;
 957                 }
 958                 if (rc == -ERESTART)
 959                         retval = rc;
 960                 else if (rc)
 961                         CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
 962                                "disconnect client\n");
 963                 LDLM_LOCK_PUT(w->w_lock);
 964                 list_del(&w->w_list);
 965                 OBD_FREE(w, sizeof(*w));
 966         }
 967         RETURN(retval);
 968 }
 969
 970 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
 971 {
 972         ldlm_reprocess_all(res);
 973         return LDLM_ITER_CONTINUE;
 974 }
 975
 976 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
 977 {
 978         (void)ldlm_namespace_foreach_res(ns, reprocess_one_queue, NULL);
 979 }
 980
 981 void ldlm_reprocess_all(struct ldlm_resource *res)
 982 {
 983         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 984         int rc;
 985         ENTRY;
 986
 987         /* Local lock trees don't get reprocessed. */
 988         if (res->lr_namespace->ns_client) {
 989                 EXIT;
 990                 return;
 991         }
 992
 993  restart:
 994         l_lock(&res->lr_namespace->ns_lock);
 995         res->lr_tmp = &rpc_list;
 996
 997         ldlm_reprocess_queue(res, &res->lr_converting);
 998         if (list_empty(&res->lr_converting))
 999                 ldlm_reprocess_queue(res, &res->lr_waiting);
1000
1001         res->lr_tmp = NULL;
1002         l_unlock(&res->lr_namespace->ns_lock);
1003
1004         rc = ldlm_run_ast_work(&rpc_list);
1005         if (rc == -ERESTART)
1006                 goto restart;
1007         EXIT;
1008 }
1009
1010 void ldlm_cancel_callback(struct ldlm_lock *lock)
1011 {
1012         l_lock(&lock->l_resource->lr_namespace->ns_lock);
1013         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1014                 lock->l_flags |= LDLM_FL_CANCEL;
1015                 if (lock->l_blocking_ast) {
1016                         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1017                         lock->l_blocking_ast(lock, NULL, lock->l_data,
1018                                              LDLM_CB_CANCELING);
1019                         return;
1020                 } else {
1021                         LDLM_DEBUG(lock, "no blocking ast");
1022                 }
1023         }
1024         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1025 }
1026
1027 void ldlm_lock_cancel(struct ldlm_lock *lock)
1028 {
1029         struct ldlm_resource *res;
1030         struct ldlm_namespace *ns;
1031         ENTRY;
1032
1033         /* There's no race between calling this and taking the ns lock below;
1034          * a lock can only be put on the waiting list once, because it can only
1035          * issue a blocking AST once. */
1036         ldlm_del_waiting_lock(lock);
1037
1038         res = lock->l_resource;
1039         ns = res->lr_namespace;
1040
1041         l_lock(&ns->ns_lock);
1042         /* Please do not, no matter how tempting, remove this LBUG without
1043          * talking to me first. -phik */
1044         if (lock->l_readers || lock->l_writers) {
1045                 LDLM_DEBUG(lock, "lock still has references");
1046                 ldlm_lock_dump(D_OTHER, lock, 0);
1047                 LBUG();
1048         }
1049
1050         ldlm_cancel_callback(lock); /* XXX FIXME bug 1030 */
1051
1052         ldlm_resource_unlink_lock(lock);
1053         ldlm_lock_destroy(lock);
1054         l_unlock(&ns->ns_lock);
1055         EXIT;
1056 }
1057
1058 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1059 {
1060         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1061         ENTRY;
1062
1063         if (lock == NULL)
1064                 RETURN(-EINVAL);
1065
1066         lock->l_data = data;
1067
1068         LDLM_LOCK_PUT(lock);
1069
1070         RETURN(0);
1071 }
1072
1073 /* This function is only called from one thread (per export); no locking around
1074  * the list ops needed */
1075 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1076 {
1077         struct list_head *iter, *n;
1078
1079         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
1080                 struct ldlm_lock *lock;
1081                 struct ldlm_resource *res;
1082                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1083                 res = ldlm_resource_getref(lock->l_resource);
1084                 LDLM_DEBUG(lock, "export %p", exp);
1085                 ldlm_lock_cancel(lock);
1086                 ldlm_reprocess_all(res);
1087                 ldlm_resource_putref(res);
1088         }
1089 }
1090
1091 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1092                                         int *flags)
1093 {
1094         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1095         struct ldlm_resource *res;
1096         struct ldlm_namespace *ns;
1097         int granted = 0;
1098         ENTRY;
1099
1100         LBUG();
1101
1102         res = lock->l_resource;
1103         ns = res->lr_namespace;
1104
1105         l_lock(&ns->ns_lock);
1106
1107         lock->l_req_mode = new_mode;
1108         ldlm_resource_unlink_lock(lock);
1109
1110         /* If this is a local resource, put it on the appropriate list. */
1111         if (res->lr_namespace->ns_client) {
1112                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1113                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1114                 } else {
1115                         /* This should never happen, because of the way the
1116                          * server handles conversions. */
1117                         LBUG();
1118
1119                         res->lr_tmp = &rpc_list;
1120                         ldlm_grant_lock(lock, NULL, 0, 0);
1121                         res->lr_tmp = NULL;
1122                         granted = 1;
1123                         /* FIXME: completion handling not with ns_lock held ! */
1124                         if (lock->l_completion_ast)
1125                                 lock->l_completion_ast(lock, 0, NULL);
1126                 }
1127         } else {
1128                 /* FIXME: We should try the conversion right away and possibly
1129                  * return success without the need for an extra AST */
1130                 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1131                 *flags |= LDLM_FL_BLOCK_CONV;
1132         }
1133
1134         l_unlock(&ns->ns_lock);
1135
1136         if (granted)
1137                 ldlm_run_ast_work(&rpc_list);
1138         RETURN(res);
1139 }
1140
1141 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1142 {
1143         char ver[128];
1144         struct obd_device *obd = NULL;
1145
1146         if (!((portal_debug | D_ERROR) & level))
1147                 return;
1148
1149         if (RES_VERSION_SIZE != 4)
1150                 LBUG();
1151
1152         if (!lock) {
1153                 CDEBUG(level, "  NULL LDLM lock\n");
1154                 return;
1155         }
1156
1157         snprintf(ver, sizeof(ver), "%x %x %x %x",
1158                  lock->l_version[0], lock->l_version[1],
1159                  lock->l_version[2], lock->l_version[3]);
1160
1161         CDEBUG(level, "  -- Lock dump: %p (%s) (rc: %d) (pos: %d)\n", lock, ver,
1162                atomic_read(&lock->l_refc), pos);
1163         if (lock->l_connh != NULL)
1164                 obd = class_conn2obd(lock->l_connh);
1165         if (lock->l_export && lock->l_export->exp_connection) {
1166                 CDEBUG(level, "  Node: NID "LPX64" on %s (rhandle: "LPX64")\n",
1167                        lock->l_export->exp_connection->c_peer.peer_nid,
1168                        lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1169                        lock->l_remote_handle.cookie);
1170         } else if (obd == NULL) {
1171                 CDEBUG(level, "  Node: local\n");
1172         } else {
1173                 struct obd_import *imp = obd->u.cli.cl_import;
1174                 CDEBUG(level, "  Node: NID "LPX64" on %s (rhandle: "LPX64")\n",
1175                        imp->imp_connection->c_peer.peer_nid,
1176                        imp->imp_connection->c_peer.peer_ni->pni_name,
1177                        lock->l_remote_handle.cookie);
1178         }
1179         CDEBUG(level, "  Resource: %p ("LPD64")\n", lock->l_resource,
1180                lock->l_resource->lr_name.name[0]);
1181         CDEBUG(level, "  Req mode: %d, grant mode: %d, readers: %d, writers: "
1182                "%d\n", (int)lock->l_req_mode, (int)lock->l_granted_mode,
1183                lock->l_readers, lock->l_writers);
1184         if (lock->l_resource->lr_type == LDLM_EXTENT)
1185                 CDEBUG(level, "  Extent: "LPU64" -> "LPU64"\n",
1186                        lock->l_extent.start, lock->l_extent.end);
1187 }
1188
1189 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1190 {
1191         struct ldlm_lock *lock;
1192
1193         lock = ldlm_handle2lock(lockh);
1194         if (lock == NULL)
1195                 return;
1196
1197         ldlm_lock_dump(D_OTHER, lock, 0);
1198
1199         LDLM_LOCK_PUT(lock);
1200 }