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 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 #include <linux/slab.h>
  27 #include <linux/module.h>
  28 #include <linux/random.h>
  29 #include <linux/lustre_dlm.h>
  30 #include <linux/lustre_mds.h>
  31 #include <linux/obd_class.h>
  32
  33 /* this lock protects ldlm_handle2lock's integrity */
  34 //static spinlock_t ldlm_handle_lock = SPIN_LOCK_UNLOCKED;
  35
  36 /* lock types */
  37 char *ldlm_lockname[] = {
  38         [0] "--",
  39         [LCK_EX] "EX",
  40         [LCK_PW] "PW",
  41         [LCK_PR] "PR",
  42         [LCK_CW] "CW",
  43         [LCK_CR] "CR",
  44         [LCK_NL] "NL"
  45 };
  46 char *ldlm_typename[] = {
  47         [LDLM_PLAIN] "PLN",
  48         [LDLM_EXTENT] "EXT",
  49         [LDLM_MDSINTENT] "INT"
  50 };
  51
  52 char *ldlm_it2str(int it)
  53 {
  54         switch (it) {
  55         case IT_OPEN:
  56                 return "open";
  57         case IT_CREAT:
  58                 return "creat";
  59         case (IT_OPEN | IT_CREAT):
  60                 return "open|creat";
  61         case IT_MKDIR:
  62                 return "mkdir";
  63         case IT_LINK:
  64                 return "link";
  65         case IT_LINK2:
  66                 return "link2";
  67         case IT_SYMLINK:
  68                 return "symlink";
  69         case IT_UNLINK:
  70                 return "unlink";
  71         case IT_RMDIR:
  72                 return "rmdir";
  73         case IT_RENAME:
  74                 return "rename";
  75         case IT_RENAME2:
  76                 return "rename2";
  77         case IT_READDIR:
  78                 return "readdir";
  79         case IT_GETATTR:
  80                 return "getattr";
  81         case IT_SETATTR:
  82                 return "setattr";
  83         case IT_READLINK:
  84                 return "readlink";
  85         case IT_MKNOD:
  86                 return "mknod";
  87         case IT_LOOKUP:
  88                 return "lookup";
  89         default:
  90                 CERROR("Unknown intent %d\n", it);
  91                 return "UNKNOWN";
  92         }
  93 }
  94
  95 extern kmem_cache_t *ldlm_lock_slab;
  96
  97 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b);
  98
  99 ldlm_res_compat ldlm_res_compat_table[] = {
 100         [LDLM_PLAIN] ldlm_plain_compat,
 101         [LDLM_EXTENT] ldlm_extent_compat,
 102         [LDLM_MDSINTENT] ldlm_plain_compat
 103 };
 104
 105 ldlm_res_policy ldlm_res_policy_table[] = {
 106         [LDLM_PLAIN] NULL,
 107         [LDLM_EXTENT] ldlm_extent_policy,
 108         [LDLM_MDSINTENT] NULL
 109 };
 110
 111 void ldlm_register_intent(int (*arg) (struct ldlm_lock * lock, void *req_cookie,
 112                                       ldlm_mode_t mode, void *data))
 113 {
 114         ldlm_res_policy_table[LDLM_MDSINTENT] = arg;
 115 }
 116
 117 void ldlm_unregister_intent(void)
 118 {
 119         ldlm_res_policy_table[LDLM_MDSINTENT] = NULL;
 120 }
 121
 122 /*
 123  * REFCOUNTED LOCK OBJECTS
 124  */
 125
 126
 127 /*
 128  * Lock refcounts, during creation:
 129  *   - one special one for allocation, dec'd only once in destroy
 130  *   - one for being a lock that's in-use
 131  *   - one for the addref associated with a new lock
 132  */
 133 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 134 {
 135         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 136         lock->l_refc++;
 137         ldlm_resource_getref(lock->l_resource);
 138         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 139         return lock;
 140 }
 141
 142 void ldlm_lock_put(struct ldlm_lock *lock)
 143 {
 144         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 145         ENTRY;
 146
 147         l_lock(&ns->ns_lock);
 148         lock->l_refc--;
 149         //LDLM_DEBUG(lock, "after refc--");
 150         if (lock->l_refc < 0)
 151                 LBUG();
 152
 153         if (ldlm_resource_put(lock->l_resource))
 154                 lock->l_resource = NULL;
 155         if (lock->l_parent)
 156                 LDLM_LOCK_PUT(lock->l_parent);
 157
 158         if (lock->l_refc == 0 && (lock->l_flags & LDLM_FL_DESTROYED)) {
 159                 l_unlock(&ns->ns_lock);
 160                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
 161
 162                 //spin_lock(&ldlm_handle_lock);
 163                 spin_lock(&ns->ns_counter_lock);
 164                 ns->ns_locks--;
 165                 spin_unlock(&ns->ns_counter_lock);
 166
 167                 lock->l_resource = NULL;
 168                 lock->l_random = DEAD_HANDLE_MAGIC;
 169                 if (lock->l_export && lock->l_export->exp_connection)
 170                         ptlrpc_put_connection(lock->l_export->exp_connection);
 171                 kmem_cache_free(ldlm_lock_slab, lock);
 172                 //spin_unlock(&ldlm_handle_lock);
 173                 CDEBUG(D_MALLOC, "kfreed 'lock': %d at %p (tot 0).\n",
 174                        sizeof(*lock), lock);
 175         } else
 176                 l_unlock(&ns->ns_lock);
 177
 178         EXIT;
 179 }
 180
 181 void ldlm_lock_destroy(struct ldlm_lock *lock)
 182 {
 183         ENTRY;
 184         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 185
 186         if (!list_empty(&lock->l_children)) {
 187                 LDLM_DEBUG(lock, "still has children (%p)!",
 188                            lock->l_children.next);
 189                 ldlm_lock_dump(lock);
 190                 LBUG();
 191         }
 192         if (lock->l_readers || lock->l_writers) {
 193                 LDLM_DEBUG(lock, "lock still has references");
 194                 ldlm_lock_dump(lock);
 195                 LBUG();
 196         }
 197
 198         if (!list_empty(&lock->l_res_link)) {
 199                 ldlm_lock_dump(lock);
 200                 LBUG();
 201         }
 202
 203         if (lock->l_flags & LDLM_FL_DESTROYED) {
 204                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 205                 EXIT;
 206                 return;
 207         }
 208
 209         list_del(&lock->l_export_chain);
 210         lock->l_export = NULL;
 211         lock->l_flags |= LDLM_FL_DESTROYED;
 212
 213         /* Wake anyone waiting for this lock */
 214         if (lock->l_completion_ast)
 215                 lock->l_completion_ast(lock, 0);
 216
 217         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 218         LDLM_LOCK_PUT(lock);
 219         EXIT;
 220 }
 221
 222 /*
 223    usage: pass in a resource on which you have done get
 224           pass in a parent lock on which you have done a get
 225           do not put the resource or the parent
 226    returns: lock with refcount 1
 227 */
 228 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
 229                                        struct ldlm_resource *resource)
 230 {
 231         struct ldlm_lock *lock;
 232         ENTRY;
 233
 234         if (resource == NULL)
 235                 LBUG();
 236
 237         lock = kmem_cache_alloc(ldlm_lock_slab, SLAB_KERNEL);
 238         if (lock == NULL)
 239                 RETURN(NULL);
 240
 241         memset(lock, 0, sizeof(*lock));
 242         get_random_bytes(&lock->l_random, sizeof(__u64));
 243
 244         lock->l_resource = resource;
 245         /* this refcount matches the one of the resource passed
 246            in which is not being put away */
 247         lock->l_refc = 1;
 248         INIT_LIST_HEAD(&lock->l_children);
 249         INIT_LIST_HEAD(&lock->l_res_link);
 250         INIT_LIST_HEAD(&lock->l_export_chain);
 251         INIT_LIST_HEAD(&lock->l_pending_chain);
 252         init_waitqueue_head(&lock->l_waitq);
 253
 254         spin_lock(&resource->lr_namespace->ns_counter_lock);
 255         resource->lr_namespace->ns_locks++;
 256         spin_unlock(&resource->lr_namespace->ns_counter_lock);
 257
 258         if (parent != NULL) {
 259                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
 260                 lock->l_parent = parent;
 261                 list_add(&lock->l_childof, &parent->l_children);
 262                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
 263         }
 264
 265         CDEBUG(D_MALLOC, "kmalloced 'lock': %d at "
 266                "%p (tot %d).\n", sizeof(*lock), lock, 1);
 267         /* this is the extra refcount, to prevent the lock from evaporating */
 268         LDLM_LOCK_GET(lock);
 269         RETURN(lock);
 270 }
 271
 272 int ldlm_lock_change_resource(struct ldlm_lock *lock, __u64 new_resid[3])
 273 {
 274         struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
 275         struct ldlm_resource *oldres = lock->l_resource;
 276         int type, i;
 277         ENTRY;
 278
 279         l_lock(&ns->ns_lock);
 280         if (memcmp(new_resid, lock->l_resource->lr_name,
 281                    sizeof(lock->l_resource->lr_name)) == 0) {
 282                 /* Nothing to do */
 283                 l_unlock(&ns->ns_lock);
 284                 RETURN(0);
 285         }
 286
 287         type = lock->l_resource->lr_type;
 288         if (new_resid[0] == 0)
 289                 LBUG();
 290         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid, type, 1);
 291         if (lock->l_resource == NULL) {
 292                 LBUG();
 293                 RETURN(-ENOMEM);
 294         }
 295
 296         /* move references over */
 297         for (i = 0; i < lock->l_refc; i++) {
 298                 int rc;
 299                 ldlm_resource_getref(lock->l_resource);
 300                 rc = ldlm_resource_put(oldres);
 301                 if (rc == 1 && i != lock->l_refc - 1)
 302                         LBUG();
 303         }
 304         /* compensate for the initial get above.. */
 305         ldlm_resource_put(lock->l_resource);
 306
 307         l_unlock(&ns->ns_lock);
 308         RETURN(0);
 309 }
 310
 311 /*
 312  *  HANDLES
 313  */
 314
 315 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
 316 {
 317         lockh->addr = (__u64) (unsigned long)lock;
 318         lockh->cookie = lock->l_random;
 319 }
 320
 321 struct ldlm_lock *ldlm_handle2lock(struct lustre_handle *handle)
 322 {
 323         struct ldlm_lock *lock = NULL, *retval = NULL;
 324         ENTRY;
 325
 326         if (!handle || !handle->addr)
 327                 RETURN(NULL);
 328
 329         //spin_lock(&ldlm_handle_lock);
 330         lock = (struct ldlm_lock *)(unsigned long)(handle->addr);
 331         if (!kmem_cache_validate(ldlm_lock_slab, (void *)lock)) {
 332                 CERROR("bogus lock %p\n", lock);
 333                 GOTO(out2, retval);
 334         }
 335
 336         if (lock->l_random != handle->cookie) {
 337                 CERROR("bogus cookie: lock %p has "LPX64" vs. handle "LPX64"\n",
 338                        lock, lock->l_random, handle->cookie);
 339                 GOTO(out2, NULL);
 340         }
 341         if (!lock->l_resource) {
 342                 CERROR("trying to lock bogus resource: lock %p\n", lock);
 343                 LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 344                 GOTO(out2, retval);
 345         }
 346         if (!lock->l_resource->lr_namespace) {
 347                 CERROR("trying to lock bogus namespace: lock %p\n", lock);
 348                 LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 349                 GOTO(out2, retval);
 350         }
 351
 352         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 353         if (lock->l_flags & LDLM_FL_DESTROYED) {
 354                 CERROR("lock already destroyed: lock %p\n", lock);
 355                 LDLM_DEBUG(lock, "ldlm_handle2lock(%p)", lock);
 356                 GOTO(out, NULL);
 357         }
 358
 359         retval = LDLM_LOCK_GET(lock);
 360         if (!retval)
 361                 CERROR("lock disappeared below us!!! %p\n", lock);
 362         EXIT;
 363  out:
 364         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 365  out2:
 366         //spin_unlock(&ldlm_handle_lock);
 367         return retval;
 368 }
 369
 370 static int ldlm_plain_compat(struct ldlm_lock *a, struct ldlm_lock *b)
 371 {
 372         return lockmode_compat(a->l_req_mode, b->l_req_mode);
 373 }
 374
 375 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
 376 {
 377         ldlm_res2desc(lock->l_resource, &desc->l_resource);
 378         desc->l_req_mode = lock->l_req_mode;
 379         desc->l_granted_mode = lock->l_granted_mode;
 380         memcpy(&desc->l_extent, &lock->l_extent, sizeof(desc->l_extent));
 381         memcpy(desc->l_version, lock->l_version, sizeof(desc->l_version));
 382 }
 383
 384 static void ldlm_add_ast_work_item(struct ldlm_lock *lock,
 385                                    struct ldlm_lock *new)
 386 {
 387         struct ldlm_ast_work *w;
 388         ENTRY;
 389
 390         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 391         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
 392                 GOTO(out, 0);
 393
 394         OBD_ALLOC(w, sizeof(*w));
 395         if (!w) {
 396                 LBUG();
 397                 GOTO(out, 0);
 398         }
 399
 400         if (new) {
 401                 lock->l_flags |= LDLM_FL_AST_SENT;
 402                 w->w_blocking = 1;
 403                 ldlm_lock2desc(new, &w->w_desc);
 404         }
 405
 406         w->w_lock = LDLM_LOCK_GET(lock);
 407         list_add(&w->w_list, lock->l_resource->lr_tmp);
 408       out:
 409         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 410         return;
 411 }
 412
 413 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
 414 {
 415         struct ldlm_lock *lock;
 416
 417         lock = ldlm_handle2lock(lockh);
 418         ldlm_lock_addref_internal(lock, mode);
 419         LDLM_LOCK_PUT(lock);
 420 }
 421
 422 /* only called for local locks */
 423 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
 424 {
 425         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 426         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 427                 lock->l_readers++;
 428         else
 429                 lock->l_writers++;
 430         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 431         LDLM_LOCK_GET(lock);
 432         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
 433 }
 434
 435 /* Args: unlocked lock */
 436 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
 437 {
 438         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 439         ENTRY;
 440
 441         if (lock == NULL)
 442                 LBUG();
 443
 444         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
 445         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 446         if (mode == LCK_NL || mode == LCK_CR || mode == LCK_PR)
 447                 lock->l_readers--;
 448         else
 449                 lock->l_writers--;
 450
 451         /* If we received a blocked AST and this was the last reference,
 452          * run the callback. */
 453         if (!lock->l_readers && !lock->l_writers &&
 454             (lock->l_flags & LDLM_FL_CBPENDING)) {
 455                 if (!lock->l_resource->lr_namespace->ns_client &&
 456                     lock->l_export)
 457                         CERROR("FL_CBPENDING set on non-local lock--just a "
 458                                "warning\n");
 459
 460                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
 461                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 462
 463                 /* FIXME: need a real 'desc' here */
 464                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 465                                      lock->l_data_len, LDLM_CB_BLOCKING);
 466         } else
 467                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 468
 469         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
 470         LDLM_LOCK_PUT(lock);    /* matches the handle2lock above */
 471
 472         EXIT;
 473 }
 474
 475 static int ldlm_lock_compat_list(struct ldlm_lock *lock, int send_cbs,
 476                                  struct list_head *queue)
 477 {
 478         struct list_head *tmp, *pos;
 479         int rc = 1;
 480
 481         list_for_each_safe(tmp, pos, queue) {
 482                 struct ldlm_lock *child;
 483                 ldlm_res_compat compat;
 484
 485                 child = list_entry(tmp, struct ldlm_lock, l_res_link);
 486                 if (lock == child)
 487                         continue;
 488
 489                 compat = ldlm_res_compat_table[child->l_resource->lr_type];
 490                 if (compat && compat(child, lock)) {
 491                         CDEBUG(D_OTHER, "compat function succeded, next.\n");
 492                         continue;
 493                 }
 494                 if (lockmode_compat(child->l_granted_mode, lock->l_req_mode)) {
 495                         CDEBUG(D_OTHER, "lock modes are compatible, next.\n");
 496                         continue;
 497                 }
 498
 499                 rc = 0;
 500
 501                 if (send_cbs && child->l_blocking_ast != NULL) {
 502                         CDEBUG(D_OTHER, "lock %p incompatible; sending "
 503                                "blocking AST.\n", child);
 504                         ldlm_add_ast_work_item(child, lock);
 505                 }
 506         }
 507
 508         return rc;
 509 }
 510
 511 static int ldlm_lock_compat(struct ldlm_lock *lock, int send_cbs)
 512 {
 513         int rc;
 514         ENTRY;
 515
 516         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 517         rc = ldlm_lock_compat_list(lock, send_cbs,
 518                                    &lock->l_resource->lr_granted);
 519         /* FIXME: should we be sending ASTs to converting? */
 520         if (rc)
 521                 rc = ldlm_lock_compat_list
 522                         (lock, send_cbs, &lock->l_resource->lr_converting);
 523
 524         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 525         RETURN(rc);
 526 }
 527
 528 /* NOTE: called by
 529    - ldlm_handle_enqueuque - resource
 530 */
 531 void ldlm_grant_lock(struct ldlm_lock *lock)
 532 {
 533         struct ldlm_resource *res = lock->l_resource;
 534         ENTRY;
 535
 536         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 537         ldlm_resource_add_lock(res, &res->lr_granted, lock);
 538         lock->l_granted_mode = lock->l_req_mode;
 539
 540         if (lock->l_granted_mode < res->lr_most_restr)
 541                 res->lr_most_restr = lock->l_granted_mode;
 542
 543         if (lock->l_completion_ast) {
 544                 ldlm_add_ast_work_item(lock, NULL);
 545         }
 546         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 547         EXIT;
 548 }
 549
 550 /* returns a referenced lock or NULL */
 551 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
 552                                       struct ldlm_extent *extent)
 553 {
 554         struct ldlm_lock *lock;
 555         struct list_head *tmp;
 556
 557         list_for_each(tmp, queue) {
 558                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 559
 560                 if (lock->l_flags & (LDLM_FL_CBPENDING | LDLM_FL_DESTROYED))
 561                         continue;
 562
 563                 /* lock_convert() takes the resource lock, so we're sure that
 564                  * req_mode, lr_type, and l_cookie won't change beneath us */
 565                 if (lock->l_req_mode != mode)
 566                         continue;
 567
 568                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
 569                     (lock->l_extent.start > extent->start ||
 570                      lock->l_extent.end < extent->end))
 571                         continue;
 572
 573                 ldlm_lock_addref_internal(lock, mode);
 574                 return lock;
 575         }
 576
 577         return NULL;
 578 }
 579
 580 /* Must be called with no resource or lock locks held.
 581  *
 582  * Returns 1 if it finds an already-existing lock that is compatible; in this
 583  * case, lockh is filled in with a addref()ed lock
 584 */
 585 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 * res_id, __u32 type,
 586                     void *cookie, int cookielen, ldlm_mode_t mode,
 587                     struct lustre_handle *lockh)
 588 {
 589         struct ldlm_resource *res;
 590         struct ldlm_lock *lock;
 591         int rc = 0;
 592         ENTRY;
 593
 594         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
 595         if (res == NULL)
 596                 RETURN(0);
 597
 598         ns = res->lr_namespace;
 599         l_lock(&ns->ns_lock);
 600
 601         if ((lock = search_queue(&res->lr_granted, mode, cookie)))
 602                 GOTO(out, rc = 1);
 603         if ((lock = search_queue(&res->lr_converting, mode, cookie)))
 604                 GOTO(out, rc = 1);
 605         if ((lock = search_queue(&res->lr_waiting, mode, cookie)))
 606                 GOTO(out, rc = 1);
 607
 608         EXIT;
 609       out:
 610         ldlm_resource_put(res);
 611         l_unlock(&ns->ns_lock);
 612
 613         if (lock) {
 614                 ldlm_lock2handle(lock, lockh);
 615                 if (lock->l_completion_ast)
 616                         lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 617         }
 618         if (rc)
 619                 LDLM_DEBUG(lock, "matched");
 620         else
 621                 LDLM_DEBUG_NOLOCK("not matched");
 622         return rc;
 623 }
 624
 625 /* Returns a referenced lock */
 626 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
 627                                    struct lustre_handle *parent_lock_handle,
 628                                    __u64 * res_id, __u32 type,
 629                                    ldlm_mode_t mode, void *data, __u32 data_len)
 630 {
 631         struct ldlm_resource *res, *parent_res = NULL;
 632         struct ldlm_lock *lock, *parent_lock;
 633
 634         parent_lock = ldlm_handle2lock(parent_lock_handle);
 635         if (parent_lock)
 636                 parent_res = parent_lock->l_resource;
 637
 638         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
 639         if (res == NULL)
 640                 RETURN(NULL);
 641
 642         lock = ldlm_lock_new(parent_lock, res);
 643         if (lock == NULL) {
 644                 ldlm_resource_put(res);
 645                 RETURN(NULL);
 646         }
 647
 648         lock->l_req_mode = mode;
 649         lock->l_data = data;
 650         lock->l_data_len = data_len;
 651
 652         return lock;
 653 }
 654
 655 /* Must be called with lock->l_lock and lock->l_resource->lr_lock not held */
 656 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock * lock,
 657                                void *cookie, int cookie_len,
 658                                int *flags,
 659                                ldlm_completion_callback completion,
 660                                ldlm_blocking_callback blocking)
 661 {
 662         struct ldlm_resource *res;
 663         int local;
 664         ldlm_res_policy policy;
 665         ENTRY;
 666
 667         res = lock->l_resource;
 668         lock->l_blocking_ast = blocking;
 669
 670         if (res->lr_type == LDLM_EXTENT)
 671                 memcpy(&lock->l_extent, cookie, sizeof(lock->l_extent));
 672
 673         /* policies are not executed on the client */
 674         local = res->lr_namespace->ns_client;
 675         if (!local && (policy = ldlm_res_policy_table[res->lr_type])) {
 676                 int rc;
 677                 rc = policy(lock, cookie, lock->l_req_mode, NULL);
 678
 679                 if (rc == ELDLM_LOCK_CHANGED) {
 680                         res = lock->l_resource;
 681                         *flags |= LDLM_FL_LOCK_CHANGED;
 682                 } else if (rc == ELDLM_LOCK_ABORTED) {
 683                         ldlm_lock_destroy(lock);
 684                         RETURN(rc);
 685                 }
 686         }
 687
 688         lock->l_cookie = cookie;
 689         lock->l_cookie_len = cookie_len;
 690
 691         l_lock(&res->lr_namespace->ns_lock);
 692         if (local && lock->l_req_mode == lock->l_granted_mode) {
 693                 /* The server returned a blocked lock, but it was granted before
 694                  * we got a chance to actually enqueue it.  We don't need to do
 695                  * anything else. */
 696                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
 697                           LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
 698                 GOTO(out, ELDLM_OK);
 699         }
 700
 701         /* This distinction between local lock trees is very important; a client
 702          * namespace only has information about locks taken by that client, and
 703          * thus doesn't have enough information to decide for itself if it can
 704          * be granted (below).  In this case, we do exactly what the server
 705          * tells us to do, as dictated by the 'flags' */
 706         ldlm_resource_unlink_lock(lock);
 707         if (local) {
 708                 if (*flags & LDLM_FL_BLOCK_CONV)
 709                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 710                                                lock);
 711                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
 712                         ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 713                 else
 714                         ldlm_grant_lock(lock);
 715                 GOTO(out, ELDLM_OK);
 716         }
 717
 718         /* FIXME: We may want to optimize by checking lr_most_restr */
 719         if (!list_empty(&res->lr_converting)) {
 720                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 721                 *flags |= LDLM_FL_BLOCK_CONV;
 722                 GOTO(out, ELDLM_OK);
 723         }
 724         if (!list_empty(&res->lr_waiting)) {
 725                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 726                 *flags |= LDLM_FL_BLOCK_WAIT;
 727                 GOTO(out, ELDLM_OK);
 728         }
 729         if (!ldlm_lock_compat(lock, 0)) {
 730                 ldlm_resource_add_lock(res, res->lr_waiting.prev, lock);
 731                 *flags |= LDLM_FL_BLOCK_GRANTED;
 732                 GOTO(out, ELDLM_OK);
 733         }
 734
 735         ldlm_grant_lock(lock);
 736         EXIT;
 737       out:
 738         l_unlock(&res->lr_namespace->ns_lock);
 739         /* Don't set 'completion_ast' until here so that if the lock is granted
 740          * immediately we don't do an unnecessary completion call. */
 741         lock->l_completion_ast = completion;
 742         return ELDLM_OK;
 743 }
 744
 745 /* Must be called with namespace taken: queue is waiting or converting. */
 746 static int ldlm_reprocess_queue(struct ldlm_resource *res,
 747                                 struct list_head *queue)
 748 {
 749         struct list_head *tmp, *pos;
 750         ENTRY;
 751
 752         list_for_each_safe(tmp, pos, queue) {
 753                 struct ldlm_lock *pending;
 754                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 755
 756                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 757
 758                 if (!ldlm_lock_compat(pending, 1))
 759                         RETURN(1);
 760
 761                 list_del_init(&pending->l_res_link);
 762                 ldlm_grant_lock(pending);
 763         }
 764
 765         RETURN(0);
 766 }
 767
 768 void ldlm_run_ast_work(struct list_head *rpc_list)
 769 {
 770         struct list_head *tmp, *pos;
 771         int rc;
 772         ENTRY;
 773
 774         list_for_each_safe(tmp, pos, rpc_list) {
 775                 struct ldlm_ast_work *w =
 776                         list_entry(tmp, struct ldlm_ast_work, w_list);
 777
 778                 if (w->w_blocking)
 779                         rc = w->w_lock->l_blocking_ast
 780                                 (w->w_lock, &w->w_desc, w->w_data,
 781                                  w->w_datalen, LDLM_CB_BLOCKING);
 782                 else
 783                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags);
 784                 if (rc)
 785                         CERROR("Failed AST - should clean & disconnect "
 786                                "client\n");
 787                 LDLM_LOCK_PUT(w->w_lock);
 788                 list_del(&w->w_list);
 789                 OBD_FREE(w, sizeof(*w));
 790         }
 791         EXIT;
 792 }
 793
 794 /* Must be called with resource->lr_lock not taken. */
 795 void ldlm_reprocess_all(struct ldlm_resource *res)
 796 {
 797         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 798         ENTRY;
 799
 800         /* Local lock trees don't get reprocessed. */
 801         if (res->lr_namespace->ns_client) {
 802                 EXIT;
 803                 return;
 804         }
 805
 806         l_lock(&res->lr_namespace->ns_lock);
 807         res->lr_tmp = &rpc_list;
 808
 809         ldlm_reprocess_queue(res, &res->lr_converting);
 810         if (list_empty(&res->lr_converting))
 811                 ldlm_reprocess_queue(res, &res->lr_waiting);
 812
 813         res->lr_tmp = NULL;
 814         l_unlock(&res->lr_namespace->ns_lock);
 815
 816         ldlm_run_ast_work(&rpc_list);
 817         EXIT;
 818 }
 819
 820 void ldlm_cancel_callback(struct ldlm_lock *lock)
 821 {
 822         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 823         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
 824                 lock->l_flags |= LDLM_FL_CANCEL;
 825                 lock->l_blocking_ast(lock, NULL, lock->l_data,
 826                                      lock->l_data_len, LDLM_CB_CANCELING);
 827         }
 828         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 829 }
 830
 831 void ldlm_lock_cancel(struct ldlm_lock *lock)
 832 {
 833         struct ldlm_resource *res;
 834         struct ldlm_namespace *ns;
 835         ENTRY;
 836
 837         res = lock->l_resource;
 838         ns = res->lr_namespace;
 839
 840         l_lock(&ns->ns_lock);
 841         if (lock->l_readers || lock->l_writers)
 842                 CDEBUG(D_INFO, "lock still has references (%d readers, %d "
 843                        "writers)\n", lock->l_readers, lock->l_writers);
 844
 845         ldlm_cancel_callback(lock);
 846
 847         ldlm_del_waiting_lock(lock);
 848         ldlm_resource_unlink_lock(lock);
 849         ldlm_lock_destroy(lock);
 850         l_unlock(&ns->ns_lock);
 851         EXIT;
 852 }
 853
 854 void ldlm_cancel_locks_for_export(struct obd_export *exp)
 855 {
 856         struct list_head *iter, *n; /* MUST BE CALLED "n"! */
 857
 858         list_for_each_safe(iter, n, &exp->exp_ldlm_data.led_held_locks) {
 859                 struct ldlm_lock *lock;
 860                 struct ldlm_resource *res;
 861                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 862                 res = ldlm_resource_getref(lock->l_resource);
 863                 LDLM_DEBUG(lock, "cancelling lock for export %p", exp);
 864                 ldlm_lock_cancel(lock);
 865                 ldlm_reprocess_all(res);
 866                 ldlm_resource_put(res);
 867         }
 868 }
 869
 870 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 871                                         int *flags)
 872 {
 873         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 874         struct ldlm_resource *res;
 875         struct ldlm_namespace *ns;
 876         int granted = 0;
 877         ENTRY;
 878
 879         res = lock->l_resource;
 880         ns = res->lr_namespace;
 881
 882         l_lock(&ns->ns_lock);
 883
 884         lock->l_req_mode = new_mode;
 885         ldlm_resource_unlink_lock(lock);
 886
 887         /* If this is a local resource, put it on the appropriate list. */
 888         if (res->lr_namespace->ns_client) {
 889                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED))
 890                         ldlm_resource_add_lock(res, res->lr_converting.prev,
 891                                                lock);
 892                 else {
 893                         /* This should never happen, because of the way the
 894                          * server handles conversions. */
 895                         LBUG();
 896
 897                         res->lr_tmp = &rpc_list;
 898                         ldlm_grant_lock(lock);
 899                         res->lr_tmp = NULL;
 900                         granted = 1;
 901                         /* FIXME: completion handling not with ns_lock held ! */
 902                         if (lock->l_completion_ast)
 903                                 lock->l_completion_ast(lock, 0);
 904                 }
 905         } else {
 906                 /* FIXME: We should try the conversion right away and possibly
 907                  * return success without the need for an extra AST */
 908                 ldlm_resource_add_lock(res, res->lr_converting.prev, lock);
 909                 *flags |= LDLM_FL_BLOCK_CONV;
 910         }
 911
 912         l_unlock(&ns->ns_lock);
 913
 914         if (granted)
 915                 ldlm_run_ast_work(&rpc_list);
 916         RETURN(res);
 917 }
 918
 919 void ldlm_lock_dump(struct ldlm_lock *lock)
 920 {
 921         char ver[128];
 922
 923         if (!(portal_debug & D_OTHER))
 924                 return;
 925
 926         if (RES_VERSION_SIZE != 4)
 927                 LBUG();
 928
 929         if (!lock) {
 930                 CDEBUG(D_OTHER, "  NULL LDLM lock\n");
 931                 return;
 932         }
 933
 934         snprintf(ver, sizeof(ver), "%x %x %x %x",
 935                  lock->l_version[0], lock->l_version[1],
 936                  lock->l_version[2], lock->l_version[3]);
 937
 938         CDEBUG(D_OTHER, "  -- Lock dump: %p (%s)\n", lock, ver);
 939         if (lock->l_export && lock->l_export->exp_connection)
 940                 CDEBUG(D_OTHER, "  Node: NID %x (rhandle: "LPX64")\n",
 941                        lock->l_export->exp_connection->c_peer.peer_nid,
 942                        lock->l_remote_handle.addr);
 943         else
 944                 CDEBUG(D_OTHER, "  Node: local\n");
 945         CDEBUG(D_OTHER, "  Parent: %p\n", lock->l_parent);
 946         CDEBUG(D_OTHER, "  Resource: %p ("LPD64")\n", lock->l_resource,
 947                lock->l_resource->lr_name[0]);
 948         CDEBUG(D_OTHER, "  Requested mode: %d, granted mode: %d\n",
 949                (int)lock->l_req_mode, (int)lock->l_granted_mode);
 950         CDEBUG(D_OTHER, "  Readers: %u ; Writers; %u\n",
 951                lock->l_readers, lock->l_writers);
 952         if (lock->l_resource->lr_type == LDLM_EXTENT)
 953                 CDEBUG(D_OTHER, "  Extent: %Lu -> %Lu\n",
 954                        (unsigned long long)lock->l_extent.start,
 955                        (unsigned long long)lock->l_extent.end);
 956 }