lustre/ldlm/ldlm_resource.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: Phil Schwan <phil@clusterfs.com>
   6  *   Author: Peter Braam <braam@clusterfs.com>
   7  *
   8  *   This file is part of the Lustre file system, http://www.lustre.org
   9  *   Lustre is a trademark of Cluster File Systems, Inc.
  10  *
  11  *   You may have signed or agreed to another license before downloading
  12  *   this software.  If so, you are bound by the terms and conditions
  13  *   of that agreement, and the following does not apply to you.  See the
  14  *   LICENSE file included with this distribution for more information.
  15  *
  16  *   If you did not agree to a different license, then this copy of Lustre
  17  *   is open source software; you can redistribute it and/or modify it
  18  *   under the terms of version 2 of the GNU General Public License as
  19  *   published by the Free Software Foundation.
  20  *
  21  *   In either case, Lustre is distributed in the hope that it will be
  22  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  23  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  24  *   license text for more details.
  25  */
  26
  27 #define DEBUG_SUBSYSTEM S_LDLM
  28 #ifdef __KERNEL__
  29 # include <lustre_dlm.h>
  30 #else
  31 # include <liblustre.h>
  32 #endif
  33
  34 #include <obd_class.h>
  35 #include "ldlm_internal.h"
  36
  37 cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
  38
  39 atomic_t ldlm_srv_namespace_nr = ATOMIC_INIT(0);
  40 atomic_t ldlm_cli_namespace_nr = ATOMIC_INIT(0);
  41
  42 struct semaphore ldlm_srv_namespace_lock;
  43 struct list_head ldlm_srv_namespace_list =
  44         CFS_LIST_HEAD_INIT(ldlm_srv_namespace_list);
  45
  46 struct semaphore ldlm_cli_namespace_lock;
  47 struct list_head ldlm_cli_namespace_list =
  48         CFS_LIST_HEAD_INIT(ldlm_cli_namespace_list);
  49
  50 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
  51 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
  52 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
  53
  54 #ifdef LPROCFS
  55 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  56                              unsigned long count, void *data)
  57 {
  58         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  59         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  60         RETURN(count);
  61 }
  62
  63 int ldlm_proc_setup(void)
  64 {
  65         int rc;
  66         struct lprocfs_vars list[] = {
  67                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  68                 { NULL }};
  69         ENTRY;
  70         LASSERT(ldlm_ns_proc_dir == NULL);
  71
  72         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
  73                                               proc_lustre_root,
  74                                               NULL, NULL);
  75         if (IS_ERR(ldlm_type_proc_dir)) {
  76                 CERROR("LProcFS failed in ldlm-init\n");
  77                 rc = PTR_ERR(ldlm_type_proc_dir);
  78                 GOTO(err, rc);
  79         }
  80
  81         ldlm_ns_proc_dir = lprocfs_register("namespaces",
  82                                             ldlm_type_proc_dir,
  83                                             NULL, NULL);
  84         if (IS_ERR(ldlm_ns_proc_dir)) {
  85                 CERROR("LProcFS failed in ldlm-init\n");
  86                 rc = PTR_ERR(ldlm_ns_proc_dir);
  87                 GOTO(err_type, rc);
  88         }
  89
  90         ldlm_svc_proc_dir = lprocfs_register("services",
  91                                             ldlm_type_proc_dir,
  92                                             NULL, NULL);
  93         if (IS_ERR(ldlm_svc_proc_dir)) {
  94                 CERROR("LProcFS failed in ldlm-init\n");
  95                 rc = PTR_ERR(ldlm_svc_proc_dir);
  96                 GOTO(err_ns, rc);
  97         }
  98
  99         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 100
 101         RETURN(0);
 102
 103 err_ns:
 104         lprocfs_remove(&ldlm_ns_proc_dir);
 105 err_type:
 106         lprocfs_remove(&ldlm_type_proc_dir);
 107 err:
 108         ldlm_svc_proc_dir = NULL;
 109         RETURN(rc);
 110 }
 111
 112 void ldlm_proc_cleanup(void)
 113 {
 114         if (ldlm_svc_proc_dir)
 115                 lprocfs_remove(&ldlm_svc_proc_dir);
 116
 117         if (ldlm_ns_proc_dir)
 118                 lprocfs_remove(&ldlm_ns_proc_dir);
 119
 120         if (ldlm_type_proc_dir)
 121                 lprocfs_remove(&ldlm_type_proc_dir);
 122 }
 123
 124 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 125                                int count, int *eof, void *data)
 126 {
 127         struct ldlm_namespace *ns = data;
 128         __u32 *nr = &ns->ns_max_unused;
 129
 130         if (ns_connect_lru_resize(ns))
 131                 nr = &ns->ns_nr_unused;
 132         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 133 }
 134
 135 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 136                                unsigned long count, void *data)
 137 {
 138         struct ldlm_namespace *ns = data;
 139         char dummy[MAX_STRING_SIZE + 1], *end;
 140         unsigned long tmp;
 141
 142         dummy[MAX_STRING_SIZE] = '\0';
 143         if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 144                 return -EFAULT;
 145
 146         if (count == 6 && memcmp(dummy, "clear", 5) == 0) {
 147                 CDEBUG(D_DLMTRACE,
 148                        "dropping all unused locks from namespace %s\n",
 149                        ns->ns_name);
 150                 if (ns_connect_lru_resize(ns)) {
 151                         int canceled, unused  = ns->ns_nr_unused;
 152
 153                         /* Try to cancel all @ns_nr_unused locks. */
 154                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC);
 155                         if (canceled < unused) {
 156                                 CERROR("not all requested locks are canceled, "
 157                                        "requested: %d, canceled: %d\n", unused,
 158                                        canceled);
 159                                 return -EINVAL;
 160                         }
 161                 } else {
 162                         tmp = ns->ns_max_unused;
 163                         ns->ns_max_unused = 0;
 164                         ldlm_cancel_lru(ns, 0, LDLM_SYNC);
 165                         ns->ns_max_unused = tmp;
 166                 }
 167                 return count;
 168         }
 169
 170         tmp = simple_strtoul(dummy, &end, 0);
 171         if (dummy == end) {
 172                 CERROR("invalid value written\n");
 173                 return -EINVAL;
 174         }
 175
 176         if (ns_connect_lru_resize(ns)) {
 177                 if (tmp > ns->ns_nr_unused)
 178                         tmp = ns->ns_nr_unused;
 179                 tmp = ns->ns_nr_unused - tmp;
 180
 181                 CDEBUG(D_DLMTRACE, "changing namespace %s unused locks from %u to %u\n",
 182                        ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
 183                 ldlm_cancel_lru(ns, (unsigned int)tmp, LDLM_ASYNC);
 184         } else {
 185                 CDEBUG(D_DLMTRACE, "changing namespace %s max_unused from %u to %u\n",
 186                        ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
 187                 ns->ns_max_unused = (unsigned int)tmp;
 188                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC);
 189         }
 190
 191         return count;
 192 }
 193
 194 void ldlm_proc_namespace(struct ldlm_namespace *ns)
 195 {
 196         struct lprocfs_vars lock_vars[2];
 197         char lock_name[MAX_STRING_SIZE + 1];
 198
 199         LASSERT(ns != NULL);
 200         LASSERT(ns->ns_name != NULL);
 201
 202         lock_name[MAX_STRING_SIZE] = '\0';
 203
 204         memset(lock_vars, 0, sizeof(lock_vars));
 205         lock_vars[0].name = lock_name;
 206
 207         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
 208         lock_vars[0].data = &ns->ns_refcount;
 209         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 210         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 211
 212         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
 213         lock_vars[0].data = &ns->ns_locks;
 214         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 215         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 216
 217         if (ns_is_client(ns)) {
 218                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 219                          ns->ns_name);
 220                 lock_vars[0].data = &ns->ns_nr_unused;
 221                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 222                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 223
 224                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 225                          ns->ns_name);
 226                 lock_vars[0].data = ns;
 227                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 228                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 229                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 230
 231                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 232                          ns->ns_name);
 233                 lock_vars[0].data = &ns->ns_max_age;
 234                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 235                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 236                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 237
 238         }
 239 }
 240 #undef MAX_STRING_SIZE
 241 #else
 242 #define ldlm_proc_namespace(ns) do {} while (0)
 243 #endif /* LPROCFS */
 244
 245 struct ldlm_namespace *ldlm_namespace_new(char *name, ldlm_side_t client,
 246                                           ldlm_appetite_t apt)
 247 {
 248         struct ldlm_namespace *ns = NULL;
 249         struct list_head *bucket;
 250         int rc, idx, namelen;
 251         ENTRY;
 252
 253         rc = ldlm_get_ref();
 254         if (rc) {
 255                 CERROR("ldlm_get_ref failed: %d\n", rc);
 256                 RETURN(NULL);
 257         }
 258
 259         OBD_ALLOC_PTR(ns);
 260         if (!ns)
 261                 GOTO(out_ref, NULL);
 262
 263         OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 264         if (!ns->ns_hash)
 265                 GOTO(out_ns, NULL);
 266
 267         ns->ns_appetite = apt;
 268         namelen = strlen(name);
 269         OBD_ALLOC(ns->ns_name, namelen + 1);
 270         if (!ns->ns_name)
 271                 GOTO(out_hash, NULL);
 272
 273         strcpy(ns->ns_name, name);
 274
 275         CFS_INIT_LIST_HEAD(&ns->ns_root_list);
 276         ns->ns_refcount = 0;
 277         ns->ns_client = client;
 278         spin_lock_init(&ns->ns_hash_lock);
 279         atomic_set(&ns->ns_locks, 0);
 280         ns->ns_resources = 0;
 281         cfs_waitq_init(&ns->ns_waitq);
 282
 283         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 284              bucket--)
 285                 CFS_INIT_LIST_HEAD(bucket);
 286
 287         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 288         ns->ns_nr_unused = 0;
 289         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
 290         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
 291         spin_lock_init(&ns->ns_unused_lock);
 292         ns->ns_connect_flags = 0;
 293         ldlm_proc_namespace(ns);
 294
 295         idx = atomic_read(ldlm_namespace_nr(client));
 296         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 297         if (rc) {
 298                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 299                 GOTO(out_proc, rc);
 300         }
 301
 302         mutex_down(ldlm_namespace_lock(client));
 303         list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 304         atomic_inc(ldlm_namespace_nr(client));
 305         mutex_up(ldlm_namespace_lock(client));
 306
 307         RETURN(ns);
 308 out_proc:
 309         ldlm_namespace_cleanup(ns, 0);
 310         OBD_FREE(ns->ns_name, namelen + 1);
 311 out_hash:
 312         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 313 out_ns:
 314         OBD_FREE_PTR(ns);
 315 out_ref:
 316         ldlm_put_ref(0);
 317         RETURN(NULL);
 318 }
 319
 320 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 321
 322 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 323  * This is currently only used for recovery, and we make certain assumptions
 324  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 325  *
 326  * Called with the ns_lock held. */
 327 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 328                              int flags)
 329 {
 330         struct list_head *tmp;
 331         int rc = 0, client = ns_is_client(res->lr_namespace);
 332         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 333         ENTRY;
 334
 335
 336         do {
 337                 struct ldlm_lock *lock = NULL;
 338
 339                 /* first, we look for non-cleaned-yet lock
 340                  * all cleaned locks are marked by CLEANED flag */
 341                 lock_res(res);
 342                 list_for_each(tmp, q) {
 343                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 344                         if (lock->l_flags & LDLM_FL_CLEANED) {
 345                                 lock = NULL;
 346                                 continue;
 347                         }
 348                         LDLM_LOCK_GET(lock);
 349                         lock->l_flags |= LDLM_FL_CLEANED;
 350                         break;
 351                 }
 352
 353                 if (lock == NULL) {
 354                         unlock_res(res);
 355                         break;
 356                 }
 357
 358                 /* Set CBPENDING so nothing in the cancellation path
 359                  * can match this lock */
 360                 lock->l_flags |= LDLM_FL_CBPENDING;
 361                 lock->l_flags |= LDLM_FL_FAILED;
 362                 lock->l_flags |= flags;
 363
 364                 /* ... without sending a CANCEL message for local_only. */
 365                 if (local_only)
 366                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 367
 368                 if (local_only && (lock->l_readers || lock->l_writers)) {
 369                         /* This is a little bit gross, but much better than the
 370                          * alternative: pretend that we got a blocking AST from
 371                          * the server, so that when the lock is decref'd, it
 372                          * will go away ... */
 373                         unlock_res(res);
 374                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 375                         if (lock->l_completion_ast)
 376                                 lock->l_completion_ast(lock, 0, NULL);
 377                         LDLM_LOCK_PUT(lock);
 378                         continue;
 379                 }
 380
 381                 if (client) {
 382                         struct lustre_handle lockh;
 383
 384                         unlock_res(res);
 385                         ldlm_lock2handle(lock, &lockh);
 386                         rc = ldlm_cli_cancel(&lockh);
 387                         if (rc)
 388                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 389                 } else {
 390                         ldlm_resource_unlink_lock(lock);
 391                         unlock_res(res);
 392                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 393                                    "client node");
 394                         ldlm_lock_destroy(lock);
 395                 }
 396                 LDLM_LOCK_PUT(lock);
 397         } while (1);
 398
 399         EXIT;
 400 }
 401
 402 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 403 {
 404         struct list_head *tmp;
 405         int i;
 406
 407         if (ns == NULL) {
 408                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 409                 return ELDLM_OK;
 410         }
 411
 412         for (i = 0; i < RES_HASH_SIZE; i++) {
 413                 spin_lock(&ns->ns_hash_lock);
 414                 tmp = ns->ns_hash[i].next;
 415                 while (tmp != &(ns->ns_hash[i])) {
 416                         struct ldlm_resource *res;
 417                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 418                         ldlm_resource_getref(res);
 419                         spin_unlock(&ns->ns_hash_lock);
 420
 421                         cleanup_resource(res, &res->lr_granted, flags);
 422                         cleanup_resource(res, &res->lr_converting, flags);
 423                         cleanup_resource(res, &res->lr_waiting, flags);
 424
 425                         spin_lock(&ns->ns_hash_lock);
 426                         tmp  = tmp->next;
 427
 428                         /* XXX: former stuff caused issues in case of race
 429                          * between ldlm_namespace_cleanup() and lockd() when
 430                          * client gets blocking ast when lock gets distracted by
 431                          * server. This is 1_4 branch solution, let's see how
 432                          * will it behave. */
 433                         if (!ldlm_resource_putref_locked(res))
 434                                 CDEBUG(D_INFO,
 435                                        "Namespace %s resource refcount nonzero "
 436                                        "(%d) after lock cleanup; forcing cleanup.\n",
 437                                        ns->ns_name, atomic_read(&res->lr_refcount));
 438                 }
 439                 spin_unlock(&ns->ns_hash_lock);
 440         }
 441
 442         return ELDLM_OK;
 443 }
 444
 445 /* Cleanup, but also free, the namespace */
 446 int ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 447 {
 448         ENTRY;
 449         if (!ns)
 450                 RETURN(ELDLM_OK);
 451
 452         mutex_down(ldlm_namespace_lock(ns->ns_client));
 453         list_del(&ns->ns_list_chain);
 454         atomic_dec(ldlm_namespace_nr(ns->ns_client));
 455         ldlm_pool_fini(&ns->ns_pool);
 456         mutex_up(ldlm_namespace_lock(ns->ns_client));
 457
 458         /* At shutdown time, don't call the cancellation callback */
 459         ldlm_namespace_cleanup(ns, 0);
 460
 461 #ifdef LPROCFS
 462         {
 463                 struct proc_dir_entry *dir;
 464                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
 465                 if (dir == NULL) {
 466                         CERROR("dlm namespace %s has no procfs dir?\n",
 467                                ns->ns_name);
 468                 } else {
 469                         lprocfs_remove(&dir);
 470                 }
 471         }
 472 #endif
 473
 474         if (ns->ns_refcount > 0) {
 475                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 476                 int rc;
 477                 CDEBUG(D_DLMTRACE,
 478                        "dlm namespace %s free waiting on refcount %d\n",
 479                        ns->ns_name, ns->ns_refcount);
 480                 rc = l_wait_event(ns->ns_waitq,
 481                                   ns->ns_refcount == 0, &lwi);
 482                 if (ns->ns_refcount)
 483                         LCONSOLE_ERROR_MSG(0x139, "Lock manager: wait for %s "
 484                                            "namespace cleanup aborted with %d "
 485                                            "resources in use. (%d)\nI'm going "
 486                                            "to try to clean up anyway, but I "
 487                                            "might need a reboot of this node.\n",
 488                                             ns->ns_name, (int) ns->ns_refcount,
 489                                             rc);
 490                 CDEBUG(D_DLMTRACE,
 491                        "dlm namespace %s free done waiting\n", ns->ns_name);
 492         }
 493
 494         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 495         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 496         OBD_FREE_PTR(ns);
 497
 498         ldlm_put_ref(force);
 499
 500         RETURN(ELDLM_OK);
 501 }
 502
 503 void ldlm_namespace_get_nolock(struct ldlm_namespace *ns)
 504 {
 505         LASSERT(ns->ns_refcount >= 0);
 506         ns->ns_refcount++;
 507 }
 508
 509 void ldlm_namespace_get(struct ldlm_namespace *ns)
 510 {
 511         spin_lock(&ns->ns_hash_lock);
 512         ldlm_namespace_get_nolock(ns);
 513         spin_unlock(&ns->ns_hash_lock);
 514 }
 515
 516 void ldlm_namespace_put_nolock(struct ldlm_namespace *ns, int wakeup)
 517 {
 518         LASSERT(ns->ns_refcount > 0);
 519         ns->ns_refcount--;
 520         if (ns->ns_refcount == 0 && wakeup)
 521                 wake_up(&ns->ns_waitq);
 522 }
 523
 524 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
 525 {
 526         spin_lock(&ns->ns_hash_lock);
 527         ldlm_namespace_put_nolock(ns, wakeup);
 528         spin_unlock(&ns->ns_hash_lock);
 529 }
 530
 531 /* Should be called under ldlm_namespace_lock(client) taken */
 532 void ldlm_namespace_move(struct ldlm_namespace *ns, ldlm_side_t client)
 533 {
 534         LASSERT(!list_empty(&ns->ns_list_chain));
 535         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 536         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 537 }
 538
 539 /* Should be called under ldlm_namespace_lock(client) taken */
 540 struct ldlm_namespace *ldlm_namespace_first(ldlm_side_t client)
 541 {
 542         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 543         LASSERT(!list_empty(ldlm_namespace_list(client)));
 544         return container_of(ldlm_namespace_list(client)->next,
 545                 struct ldlm_namespace, ns_list_chain);
 546 }
 547 static __u32 ldlm_hash_fn(struct ldlm_resource *parent,
 548                           const struct ldlm_res_id *name)
 549 {
 550         __u32 hash = 0;
 551         int i;
 552
 553         for (i = 0; i < RES_NAME_SIZE; i++)
 554                 hash += name->name[i];
 555
 556         hash += (__u32)((unsigned long)parent >> 4);
 557
 558         return (hash & RES_HASH_MASK);
 559 }
 560
 561 static struct ldlm_resource *ldlm_resource_new(void)
 562 {
 563         struct ldlm_resource *res;
 564
 565         OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
 566         if (res == NULL)
 567                 return NULL;
 568
 569         memset(res, 0, sizeof(*res));
 570
 571         CFS_INIT_LIST_HEAD(&res->lr_children);
 572         CFS_INIT_LIST_HEAD(&res->lr_childof);
 573         CFS_INIT_LIST_HEAD(&res->lr_granted);
 574         CFS_INIT_LIST_HEAD(&res->lr_converting);
 575         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 576         atomic_set(&res->lr_refcount, 1);
 577         spin_lock_init(&res->lr_lock);
 578
 579         /* one who creates the resource must unlock
 580          * the semaphore after lvb initialization */
 581         init_MUTEX_LOCKED(&res->lr_lvb_sem);
 582
 583         return res;
 584 }
 585
 586 /* must be called with hash lock held */
 587 static struct ldlm_resource *
 588 ldlm_resource_find(struct ldlm_namespace *ns, const struct ldlm_res_id *name,
 589                    __u32 hash)
 590 {
 591         struct list_head *bucket, *tmp;
 592         struct ldlm_resource *res;
 593
 594         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 595         bucket = ns->ns_hash + hash;
 596
 597         list_for_each(tmp, bucket) {
 598                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 599                 if (memcmp(&res->lr_name, name, sizeof(res->lr_name)) == 0)
 600                         return res;
 601         }
 602
 603         return NULL;
 604 }
 605
 606 /* Args: locked namespace
 607  * Returns: newly-allocated, referenced, unlocked resource */
 608 static struct ldlm_resource *
 609 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 610                   const struct ldlm_res_id *name, __u32 hash, ldlm_type_t type)
 611 {
 612         struct list_head *bucket;
 613         struct ldlm_resource *res, *old_res;
 614         ENTRY;
 615
 616         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
 617                  "type: %d\n", type);
 618
 619         res = ldlm_resource_new();
 620         if (!res)
 621                 RETURN(NULL);
 622
 623         res->lr_name = *name;
 624         res->lr_namespace = ns;
 625         res->lr_type = type;
 626         res->lr_most_restr = LCK_NL;
 627
 628         spin_lock(&ns->ns_hash_lock);
 629         old_res = ldlm_resource_find(ns, name, hash);
 630         if (old_res) {
 631                 /* someone won the race and added the resource before */
 632                 ldlm_resource_getref(old_res);
 633                 spin_unlock(&ns->ns_hash_lock);
 634                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 635                 /* synchronize WRT resource creation */
 636                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 637                         down(&old_res->lr_lvb_sem);
 638                         up(&old_res->lr_lvb_sem);
 639                 }
 640                 RETURN(old_res);
 641         }
 642
 643         /* we won! let's add the resource */
 644         bucket = ns->ns_hash + hash;
 645         list_add(&res->lr_hash, bucket);
 646         ns->ns_resources++;
 647         ldlm_namespace_get_nolock(ns);
 648
 649         if (parent == NULL) {
 650                 list_add(&res->lr_childof, &ns->ns_root_list);
 651         } else {
 652                 res->lr_parent = parent;
 653                 list_add(&res->lr_childof, &parent->lr_children);
 654         }
 655         spin_unlock(&ns->ns_hash_lock);
 656
 657         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 658                 int rc;
 659
 660                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
 661                 rc = ns->ns_lvbo->lvbo_init(res);
 662                 if (rc)
 663                         CERROR("lvbo_init failed for resource "
 664                                LPU64": rc %d\n", name->name[0], rc);
 665                 /* we create resource with locked lr_lvb_sem */
 666                 up(&res->lr_lvb_sem);
 667         }
 668
 669         RETURN(res);
 670 }
 671
 672 /* Args: unlocked namespace
 673  * Locks: takes and releases ns->ns_lock and res->lr_lock
 674  * Returns: referenced, unlocked ldlm_resource or NULL */
 675 struct ldlm_resource *
 676 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 677                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
 678 {
 679         __u32 hash = ldlm_hash_fn(parent, name);
 680         struct ldlm_resource *res = NULL;
 681         ENTRY;
 682
 683         LASSERT(ns != NULL);
 684         LASSERT(ns->ns_hash != NULL);
 685         LASSERT(name->name[0] != 0);
 686
 687         spin_lock(&ns->ns_hash_lock);
 688         res = ldlm_resource_find(ns, name, hash);
 689         if (res) {
 690                 ldlm_resource_getref(res);
 691                 spin_unlock(&ns->ns_hash_lock);
 692                 /* synchronize WRT resource creation */
 693                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 694                         down(&res->lr_lvb_sem);
 695                         up(&res->lr_lvb_sem);
 696                 }
 697                 RETURN(res);
 698         }
 699         spin_unlock(&ns->ns_hash_lock);
 700
 701         if (create == 0)
 702                 RETURN(NULL);
 703
 704         res = ldlm_resource_add(ns, parent, name, hash, type);
 705         RETURN(res);
 706 }
 707
 708 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 709 {
 710         LASSERT(res != NULL);
 711         LASSERT(res != LP_POISON);
 712         atomic_inc(&res->lr_refcount);
 713         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 714                atomic_read(&res->lr_refcount));
 715         return res;
 716 }
 717
 718 void __ldlm_resource_putref_final(struct ldlm_resource *res)
 719 {
 720         struct ldlm_namespace *ns = res->lr_namespace;
 721
 722         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 723
 724         if (!list_empty(&res->lr_granted)) {
 725                 ldlm_resource_dump(D_ERROR, res);
 726                 LBUG();
 727         }
 728
 729         if (!list_empty(&res->lr_converting)) {
 730                 ldlm_resource_dump(D_ERROR, res);
 731                 LBUG();
 732         }
 733
 734         if (!list_empty(&res->lr_waiting)) {
 735                 ldlm_resource_dump(D_ERROR, res);
 736                 LBUG();
 737         }
 738
 739         if (!list_empty(&res->lr_children)) {
 740                 ldlm_resource_dump(D_ERROR, res);
 741                 LBUG();
 742         }
 743
 744         /* Pass 0 here to not wake ->ns_waitq up yet, we will do it few
 745          * lines below when all children are freed. */
 746         ldlm_namespace_put_nolock(ns, 0);
 747         list_del_init(&res->lr_hash);
 748         list_del_init(&res->lr_childof);
 749
 750         ns->ns_resources--;
 751         if (ns->ns_resources == 0)
 752                 wake_up(&ns->ns_waitq);
 753 }
 754
 755 /* Returns 1 if the resource was freed, 0 if it remains. */
 756 int ldlm_resource_putref(struct ldlm_resource *res)
 757 {
 758         struct ldlm_namespace *ns = res->lr_namespace;
 759         int rc = 0;
 760         ENTRY;
 761
 762         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 763                atomic_read(&res->lr_refcount) - 1);
 764         LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
 765                  atomic_read(&res->lr_refcount));
 766         LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
 767                  atomic_read(&res->lr_refcount));
 768
 769         if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
 770                 __ldlm_resource_putref_final(res);
 771                 spin_unlock(&ns->ns_hash_lock);
 772                 if (res->lr_lvb_data)
 773                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 774                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 775                 rc = 1;
 776         }
 777
 778         RETURN(rc);
 779 }
 780
 781 /* Returns 1 if the resource was freed, 0 if it remains. */
 782 int ldlm_resource_putref_locked(struct ldlm_resource *res)
 783 {
 784         int rc = 0;
 785         ENTRY;
 786
 787         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 788                atomic_read(&res->lr_refcount) - 1);
 789         LASSERT(atomic_read(&res->lr_refcount) > 0);
 790         LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
 791
 792         LASSERT(atomic_read(&res->lr_refcount) >= 0);
 793         if (atomic_dec_and_test(&res->lr_refcount)) {
 794                 __ldlm_resource_putref_final(res);
 795                 if (res->lr_lvb_data)
 796                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 797                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 798                 rc = 1;
 799         }
 800
 801         RETURN(rc);
 802 }
 803
 804 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 805                             struct ldlm_lock *lock)
 806 {
 807         check_res_locked(res);
 808
 809         ldlm_resource_dump(D_OTHER, res);
 810         CDEBUG(D_OTHER, "About to add this lock:\n");
 811         ldlm_lock_dump(D_OTHER, lock, 0);
 812
 813         if (lock->l_destroyed) {
 814                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 815                 return;
 816         }
 817
 818         LASSERT(list_empty(&lock->l_res_link));
 819
 820         list_add_tail(&lock->l_res_link, head);
 821 }
 822
 823 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
 824                                      struct ldlm_lock *new)
 825 {
 826         struct ldlm_resource *res = original->l_resource;
 827
 828         check_res_locked(res);
 829
 830         ldlm_resource_dump(D_OTHER, res);
 831         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
 832         ldlm_lock_dump(D_OTHER, new, 0);
 833
 834         if (new->l_destroyed) {
 835                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 836                 goto out;
 837         }
 838
 839         LASSERT(list_empty(&new->l_res_link));
 840
 841         list_add(&new->l_res_link, &original->l_res_link);
 842  out:;
 843 }
 844
 845 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 846 {
 847         check_res_locked(lock->l_resource);
 848         ldlm_unlink_lock_skiplist(lock);
 849         list_del_init(&lock->l_res_link);
 850 }
 851
 852 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 853 {
 854         desc->lr_type = res->lr_type;
 855         desc->lr_name = res->lr_name;
 856 }
 857
 858 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
 859 {
 860         struct list_head *tmp;
 861
 862         if (!((libcfs_debug | D_ERROR) & level))
 863                 return;
 864
 865         mutex_down(ldlm_namespace_lock(client));
 866
 867         list_for_each(tmp, ldlm_namespace_list(client)) {
 868                 struct ldlm_namespace *ns;
 869                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 870                 ldlm_namespace_dump(level, ns);
 871         }
 872
 873         mutex_up(ldlm_namespace_lock(client));
 874 }
 875
 876 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
 877 {
 878         struct list_head *tmp;
 879
 880         if (!((libcfs_debug | D_ERROR) & level))
 881                 return;
 882
 883         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
 884                ns->ns_name, ns->ns_refcount,
 885                ns_is_client(ns) ? "client" : "server");
 886
 887         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
 888                 return;
 889
 890         spin_lock(&ns->ns_hash_lock);
 891         tmp = ns->ns_root_list.next;
 892         while (tmp != &ns->ns_root_list) {
 893                 struct ldlm_resource *res;
 894                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
 895
 896                 ldlm_resource_getref(res);
 897                 spin_unlock(&ns->ns_hash_lock);
 898
 899                 lock_res(res);
 900                 ldlm_resource_dump(level, res);
 901                 unlock_res(res);
 902
 903                 spin_lock(&ns->ns_hash_lock);
 904                 tmp = tmp->next;
 905                 ldlm_resource_putref_locked(res);
 906         }
 907         ns->ns_next_dump = cfs_time_shift(10);
 908         spin_unlock(&ns->ns_hash_lock);
 909 }
 910
 911 void ldlm_resource_dump(int level, struct ldlm_resource *res)
 912 {
 913         struct list_head *tmp;
 914         int pos;
 915
 916         CLASSERT(RES_NAME_SIZE == 4);
 917
 918         if (!((libcfs_debug | D_ERROR) & level))
 919                 return;
 920
 921         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
 922                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
 923                res->lr_name.name[2], res->lr_name.name[3],
 924                atomic_read(&res->lr_refcount));
 925
 926         if (!list_empty(&res->lr_granted)) {
 927                 pos = 0;
 928                 CDEBUG(level, "Granted locks:\n");
 929                 list_for_each(tmp, &res->lr_granted) {
 930                         struct ldlm_lock *lock;
 931                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 932                         ldlm_lock_dump(level, lock, ++pos);
 933                 }
 934         }
 935         if (!list_empty(&res->lr_converting)) {
 936                 pos = 0;
 937                 CDEBUG(level, "Converting locks:\n");
 938                 list_for_each(tmp, &res->lr_converting) {
 939                         struct ldlm_lock *lock;
 940                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 941                         ldlm_lock_dump(level, lock, ++pos);
 942                 }
 943         }
 944         if (!list_empty(&res->lr_waiting)) {
 945                 pos = 0;
 946                 CDEBUG(level, "Waiting locks:\n");
 947                 list_for_each(tmp, &res->lr_waiting) {
 948                         struct ldlm_lock *lock;
 949                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 950                         ldlm_lock_dump(level, lock, ++pos);
 951                 }
 952         }
 953 }