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 struct semaphore ldlm_namespace_lock;
  42 struct list_head ldlm_namespace_list = CFS_LIST_HEAD_INIT(ldlm_namespace_list);
  43 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
  44 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
  45 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
  46
  47 #ifdef LPROCFS
  48 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  49                              unsigned long count, void *data)
  50 {
  51         ldlm_dump_all_namespaces(D_DLMTRACE);
  52         RETURN(count);
  53 }
  54
  55 int ldlm_proc_setup(void)
  56 {
  57         int rc;
  58         struct lprocfs_vars list[] = {
  59                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  60                 { NULL }};
  61         ENTRY;
  62         LASSERT(ldlm_ns_proc_dir == NULL);
  63
  64         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
  65                                               proc_lustre_root,
  66                                               NULL, NULL);
  67         if (IS_ERR(ldlm_type_proc_dir)) {
  68                 CERROR("LProcFS failed in ldlm-init\n");
  69                 rc = PTR_ERR(ldlm_type_proc_dir);
  70                 GOTO(err, rc);
  71         }
  72
  73         ldlm_ns_proc_dir = lprocfs_register("namespaces",
  74                                             ldlm_type_proc_dir,
  75                                             NULL, NULL);
  76         if (IS_ERR(ldlm_ns_proc_dir)) {
  77                 CERROR("LProcFS failed in ldlm-init\n");
  78                 rc = PTR_ERR(ldlm_ns_proc_dir);
  79                 GOTO(err_type, rc);
  80         }
  81
  82         ldlm_svc_proc_dir = lprocfs_register("services",
  83                                             ldlm_type_proc_dir,
  84                                             NULL, NULL);
  85         if (IS_ERR(ldlm_svc_proc_dir)) {
  86                 CERROR("LProcFS failed in ldlm-init\n");
  87                 rc = PTR_ERR(ldlm_svc_proc_dir);
  88                 GOTO(err_ns, rc);
  89         }
  90
  91         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
  92
  93         RETURN(0);
  94
  95 err_ns:
  96         lprocfs_remove(&ldlm_ns_proc_dir);
  97 err_type:
  98         lprocfs_remove(&ldlm_type_proc_dir);
  99 err:
 100         ldlm_svc_proc_dir = NULL;
 101         RETURN(rc);
 102 }
 103
 104 void ldlm_proc_cleanup(void)
 105 {
 106         if (ldlm_svc_proc_dir)
 107                 lprocfs_remove(&ldlm_svc_proc_dir);
 108
 109         if (ldlm_ns_proc_dir)
 110                 lprocfs_remove(&ldlm_ns_proc_dir);
 111
 112         if (ldlm_type_proc_dir)
 113                 lprocfs_remove(&ldlm_type_proc_dir);
 114 }
 115
 116 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 117                                int count, int *eof, void *data)
 118 {
 119         struct ldlm_namespace *ns = data;
 120         __u32 *nr = &ns->ns_max_unused;
 121
 122         if (ns_connect_lru_resize(ns))
 123                 nr = &ns->ns_nr_unused;
 124         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 125 }
 126
 127 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 128                                unsigned long count, void *data)
 129 {
 130         struct ldlm_namespace *ns = data;
 131         char dummy[MAX_STRING_SIZE + 1], *end;
 132         unsigned long tmp;
 133
 134         dummy[MAX_STRING_SIZE] = '\0';
 135         if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 136                 return -EFAULT;
 137
 138         if (count == 6 && memcmp(dummy, "clear", 5) == 0) {
 139                 CDEBUG(D_DLMTRACE,
 140                        "dropping all unused locks from namespace %s\n",
 141                        ns->ns_name);
 142                 if (ns_connect_lru_resize(ns)) {
 143                         int canceled, unused  = ns->ns_nr_unused;
 144
 145                         /* Try to cancel all @ns_nr_unused locks. */
 146                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC);
 147                         if (canceled < unused) {
 148                                 CERROR("not all requested locks are canceled, "
 149                                        "requested: %d, canceled: %d\n", unused,
 150                                        canceled);
 151                                 return -EINVAL;
 152                         }
 153                 } else {
 154                         tmp = ns->ns_max_unused;
 155                         ns->ns_max_unused = 0;
 156                         ldlm_cancel_lru(ns, 0, LDLM_SYNC);
 157                         ns->ns_max_unused = tmp;
 158                 }
 159                 return count;
 160         }
 161
 162         tmp = simple_strtoul(dummy, &end, 0);
 163         if (*end) {
 164                 CERROR("invalid value written\n");
 165                 return -EINVAL;
 166         }
 167
 168         if (ns_connect_lru_resize(ns)) {
 169                 if (tmp > ns->ns_nr_unused)
 170                         tmp = ns->ns_nr_unused;
 171                 tmp = ns->ns_nr_unused - tmp;
 172
 173                 CDEBUG(D_DLMTRACE, "changing namespace %s unused locks from %u to %u\n",
 174                        ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
 175                 ldlm_cancel_lru(ns, (unsigned int)tmp, LDLM_ASYNC);
 176         } else {
 177                 CDEBUG(D_DLMTRACE, "changing namespace %s max_unused from %u to %u\n",
 178                        ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
 179                 ns->ns_max_unused = (unsigned int)tmp;
 180                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC);
 181         }
 182         return count;
 183 }
 184
 185 void ldlm_proc_namespace(struct ldlm_namespace *ns)
 186 {
 187         struct lprocfs_vars lock_vars[2];
 188         char lock_name[MAX_STRING_SIZE + 1];
 189
 190         LASSERT(ns != NULL);
 191         LASSERT(ns->ns_name != NULL);
 192
 193         lock_name[MAX_STRING_SIZE] = '\0';
 194
 195         memset(lock_vars, 0, sizeof(lock_vars));
 196         lock_vars[0].name = lock_name;
 197
 198         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
 199         lock_vars[0].data = &ns->ns_refcount;
 200         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 201         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 202
 203         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
 204         lock_vars[0].data = &ns->ns_locks;
 205         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 206         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 207
 208         if (ns->ns_client) {
 209                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 210                          ns->ns_name);
 211                 lock_vars[0].data = &ns->ns_nr_unused;
 212                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 213                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 214
 215                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 216                          ns->ns_name);
 217                 lock_vars[0].data = ns;
 218                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 219                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 220                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 221
 222                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 223                          ns->ns_name);
 224                 lock_vars[0].data = &ns->ns_max_age;
 225                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 226                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 227                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 228         } else {
 229                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 230                          ns->ns_name);
 231                 lock_vars[0].data = &ns->ns_max_nolock_size;
 232                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 233                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 234                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 235
 236                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 237                          ns->ns_name);
 238                 lock_vars[0].data = &ns->ns_contention_time;
 239                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 240                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 241                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 242
 243                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 244                          ns->ns_name);
 245                 lock_vars[0].data = &ns->ns_contended_locks;
 246                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 247                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 248                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 249         }
 250 }
 251 #undef MAX_STRING_SIZE
 252 #else
 253 #define ldlm_proc_namespace(ns) do {} while (0)
 254 #endif /* LPROCFS */
 255
 256 static atomic_t *ldlm_namespace_nr(ldlm_side_t client)
 257 {
 258         return client == LDLM_NAMESPACE_SERVER ?
 259                 &ldlm_srv_namespace_nr : &ldlm_cli_namespace_nr;
 260 }
 261
 262 struct ldlm_namespace *ldlm_namespace_new(char *name, ldlm_side_t client,
 263                                           ldlm_appetite_t apt)
 264 {
 265         struct ldlm_namespace *ns = NULL;
 266         struct list_head *bucket;
 267         int rc, idx;
 268         ENTRY;
 269
 270         rc = ldlm_get_ref(client);
 271         if (rc) {
 272                 CERROR("ldlm_get_ref failed: %d\n", rc);
 273                 RETURN(NULL);
 274         }
 275
 276         OBD_ALLOC_PTR(ns);
 277         if (!ns)
 278                 GOTO(out_ref, NULL);
 279
 280         OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 281         if (!ns->ns_hash)
 282                 GOTO(out_ns, NULL);
 283
 284         OBD_ALLOC(ns->ns_name, strlen(name) + 1);
 285         if (!ns->ns_name)
 286                 GOTO(out_hash, NULL);
 287
 288         ns->ns_appetite = apt;
 289         strcpy(ns->ns_name, name);
 290
 291         CFS_INIT_LIST_HEAD(&ns->ns_root_list);
 292         ns->ns_refcount = 0;
 293         ns->ns_client = client;
 294         spin_lock_init(&ns->ns_hash_lock);
 295         atomic_set(&ns->ns_locks, 0);
 296         ns->ns_resources = 0;
 297         cfs_waitq_init(&ns->ns_waitq);
 298         ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
 299         ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
 300         ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
 301
 302         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 303              bucket--)
 304                 CFS_INIT_LIST_HEAD(bucket);
 305
 306         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 307         ns->ns_nr_unused = 0;
 308         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
 309         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
 310         spin_lock_init(&ns->ns_unused_lock);
 311
 312         ns->ns_connect_flags = 0;
 313         mutex_down(&ldlm_namespace_lock);
 314         list_add(&ns->ns_list_chain, &ldlm_namespace_list);
 315         idx = atomic_read(ldlm_namespace_nr(client));
 316         atomic_inc(ldlm_namespace_nr(client));
 317         mutex_up(&ldlm_namespace_lock);
 318
 319         ldlm_proc_namespace(ns);
 320
 321         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 322         if (rc) {
 323                 CERROR("can't initialize lock pool, rc %d\n", rc);
 324                 GOTO(out_del, rc);
 325         }
 326         RETURN(ns);
 327
 328 out_del:
 329         mutex_down(&ldlm_namespace_lock);
 330         list_del(&ns->ns_list_chain);
 331         atomic_dec(ldlm_namespace_nr(client));
 332         mutex_up(&ldlm_namespace_lock);
 333 out_hash:
 334         POISON(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 335         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 336 out_ns:
 337         OBD_FREE_PTR(ns);
 338 out_ref:
 339         ldlm_put_ref(client, 0);
 340         RETURN(NULL);
 341 }
 342
 343 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 344
 345 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 346  * This is currently only used for recovery, and we make certain assumptions
 347  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 348  *
 349  * Called with the ns_lock held. */
 350 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 351                              int flags)
 352 {
 353         struct list_head *tmp;
 354         int rc = 0, client = res->lr_namespace->ns_client;
 355         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 356         ENTRY;
 357
 358
 359         do {
 360                 struct ldlm_lock *lock = NULL;
 361
 362                 /* first, we look for non-cleaned-yet lock
 363                  * all cleaned locks are marked by CLEANED flag */
 364                 lock_res(res);
 365                 list_for_each(tmp, q) {
 366                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 367                         if (lock->l_flags & LDLM_FL_CLEANED) {
 368                                 lock = NULL;
 369                                 continue;
 370                         }
 371                         LDLM_LOCK_GET(lock);
 372                         lock->l_flags |= LDLM_FL_CLEANED;
 373                         break;
 374                 }
 375
 376                 if (lock == NULL) {
 377                         unlock_res(res);
 378                         break;
 379                 }
 380
 381                 /* Set CBPENDING so nothing in the cancellation path
 382                  * can match this lock */
 383                 lock->l_flags |= LDLM_FL_CBPENDING;
 384                 lock->l_flags |= LDLM_FL_FAILED;
 385                 lock->l_flags |= flags;
 386
 387                 /* ... without sending a CANCEL message for local_only. */
 388                 if (local_only)
 389                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 390
 391                 if (local_only && (lock->l_readers || lock->l_writers)) {
 392                         /* This is a little bit gross, but much better than the
 393                          * alternative: pretend that we got a blocking AST from
 394                          * the server, so that when the lock is decref'd, it
 395                          * will go away ... */
 396                         unlock_res(res);
 397                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 398                         if (lock->l_completion_ast)
 399                                 lock->l_completion_ast(lock, 0, NULL);
 400                         LDLM_LOCK_PUT(lock);
 401                         continue;
 402                 }
 403
 404                 if (client) {
 405                         struct lustre_handle lockh;
 406
 407                         unlock_res(res);
 408                         ldlm_lock2handle(lock, &lockh);
 409                         rc = ldlm_cli_cancel(&lockh);
 410                         if (rc)
 411                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 412                 } else {
 413                         ldlm_resource_unlink_lock(lock);
 414                         unlock_res(res);
 415                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 416                                    "client node");
 417                         ldlm_lock_destroy(lock);
 418                 }
 419                 LDLM_LOCK_PUT(lock);
 420         } while (1);
 421
 422         EXIT;
 423 }
 424
 425 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 426 {
 427         struct list_head *tmp;
 428         int i;
 429
 430         if (ns == NULL) {
 431                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 432                 return ELDLM_OK;
 433         }
 434
 435         for (i = 0; i < RES_HASH_SIZE; i++) {
 436                 spin_lock(&ns->ns_hash_lock);
 437                 tmp = ns->ns_hash[i].next;
 438                 while (tmp != &(ns->ns_hash[i])) {
 439                         struct ldlm_resource *res;
 440                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 441                         ldlm_resource_getref(res);
 442                         spin_unlock(&ns->ns_hash_lock);
 443
 444                         cleanup_resource(res, &res->lr_granted, flags);
 445                         cleanup_resource(res, &res->lr_converting, flags);
 446                         cleanup_resource(res, &res->lr_waiting, flags);
 447
 448                         spin_lock(&ns->ns_hash_lock);
 449                         tmp  = tmp->next;
 450
 451                         /* XXX: former stuff caused issues in case of race
 452                          * between ldlm_namespace_cleanup() and lockd() when
 453                          * client gets blocking ast when lock gets distracted by
 454                          * server. This is 1_4 branch solution, let's see how
 455                          * will it behave. */
 456                         if (!ldlm_resource_putref_locked(res))
 457                                 CDEBUG(D_INFO,
 458                                        "Namespace %s resource refcount nonzero "
 459                                        "(%d) after lock cleanup; forcing cleanup.\n",
 460                                        ns->ns_name, atomic_read(&res->lr_refcount));
 461                 }
 462                 spin_unlock(&ns->ns_hash_lock);
 463         }
 464
 465         return ELDLM_OK;
 466 }
 467
 468 int ldlm_namespace_free_prior(struct ldlm_namespace *ns)
 469 {
 470         ENTRY;
 471         if (!ns)
 472                 RETURN(ELDLM_OK);
 473
 474         mutex_down(&ldlm_namespace_lock);
 475         list_del(&ns->ns_list_chain);
 476         atomic_dec(ldlm_namespace_nr(ns->ns_client));
 477         ldlm_pool_fini(&ns->ns_pool);
 478         mutex_up(&ldlm_namespace_lock);
 479
 480         /* At shutdown time, don't call the cancellation callback */
 481         ldlm_namespace_cleanup(ns, 0);
 482
 483         if (ns->ns_refcount > 0) {
 484                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 485                 int rc;
 486                 CDEBUG(D_DLMTRACE,
 487                        "dlm namespace %s free waiting on refcount %d\n",
 488                        ns->ns_name, ns->ns_refcount);
 489                 rc = l_wait_event(ns->ns_waitq,
 490                                   ns->ns_refcount == 0, &lwi);
 491                 if (ns->ns_refcount)
 492                         LCONSOLE_ERROR_MSG(0x139, "Lock manager: wait for %s "
 493                                            "namespace cleanup aborted with %d "
 494                                            "resources in use. (%d)\nI'm going "
 495                                            "to try to clean up anyway, but I "
 496                                            "might need a reboot of this node.\n",
 497                                             ns->ns_name, (int) ns->ns_refcount,
 498                                             rc);
 499                 CDEBUG(D_DLMTRACE,
 500                        "dlm namespace %s free done waiting\n", ns->ns_name);
 501         }
 502
 503         RETURN(ELDLM_OK);
 504 }
 505
 506 int ldlm_namespace_free_post(struct ldlm_namespace *ns, int force)
 507 {
 508         ldlm_side_t client;
 509         ENTRY;
 510         if (!ns)
 511                 RETURN(ELDLM_OK);
 512
 513 #ifdef LPROCFS
 514         {
 515                 struct proc_dir_entry *dir;
 516                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
 517                 if (dir == NULL) {
 518                         CERROR("dlm namespace %s has no procfs dir?\n",
 519                                ns->ns_name);
 520                 } else {
 521                         lprocfs_remove(&dir);
 522                 }
 523         }
 524 #endif
 525         client = ns->ns_client;
 526         POISON(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 527         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 528         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 529         OBD_FREE_PTR(ns);
 530         ldlm_put_ref(client, force);
 531         RETURN(ELDLM_OK);
 532 }
 533
 534 /* Cleanup the resource, and free namespace.
 535  * bug 12864:
 536  * Deadlock issue:
 537  * proc1: destroy import
 538  *        class_disconnect_export(grab cl_sem) ->
 539  *              -> ldlm_namespace_free ->
 540  *              -> lprocfs_remove(grab _lprocfs_lock).
 541  * proc2: read proc info
 542  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 543  *              -> osc_rd_active, etc(grab cl_sem).
 544  *
 545  * So that I have to split the ldlm_namespace_free into two parts - the first
 546  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 547  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 548  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 549  * held.
 550  */
 551 int ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 552 {
 553         ldlm_namespace_free_prior(ns);
 554         ldlm_namespace_free_post(ns, force);
 555         return ELDLM_OK;
 556 }
 557
 558 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, struct ldlm_res_id name)
 559 {
 560         __u32 hash = 0;
 561         int i;
 562
 563         for (i = 0; i < RES_NAME_SIZE; i++)
 564                 hash += name.name[i];
 565
 566         hash += (__u32)((unsigned long)parent >> 4);
 567
 568         return (hash & RES_HASH_MASK);
 569 }
 570
 571 static struct ldlm_resource *ldlm_resource_new(void)
 572 {
 573         struct ldlm_resource *res;
 574
 575         OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
 576         if (res == NULL)
 577                 return NULL;
 578
 579         memset(res, 0, sizeof(*res));
 580
 581         CFS_INIT_LIST_HEAD(&res->lr_children);
 582         CFS_INIT_LIST_HEAD(&res->lr_childof);
 583         CFS_INIT_LIST_HEAD(&res->lr_granted);
 584         CFS_INIT_LIST_HEAD(&res->lr_converting);
 585         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 586         atomic_set(&res->lr_refcount, 1);
 587         spin_lock_init(&res->lr_lock);
 588
 589         /* one who creates the resource must unlock
 590          * the semaphore after lvb initialization */
 591         init_MUTEX_LOCKED(&res->lr_lvb_sem);
 592
 593         return res;
 594 }
 595
 596 /* must be called with hash lock held */
 597 static struct ldlm_resource *
 598 ldlm_resource_find(struct ldlm_namespace *ns, struct ldlm_res_id name, __u32 hash)
 599 {
 600         struct list_head *bucket, *tmp;
 601         struct ldlm_resource *res;
 602
 603         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 604         bucket = ns->ns_hash + hash;
 605
 606         list_for_each(tmp, bucket) {
 607                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 608                 if (memcmp(&res->lr_name, &name, sizeof(res->lr_name)) == 0)
 609                         return res;
 610         }
 611
 612         return NULL;
 613 }
 614
 615 /* Args: locked namespace
 616  * Returns: newly-allocated, referenced, unlocked resource */
 617 static struct ldlm_resource *
 618 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 619                   struct ldlm_res_id name, __u32 hash, ldlm_type_t type)
 620 {
 621         struct list_head *bucket;
 622         struct ldlm_resource *res, *old_res;
 623         ENTRY;
 624
 625         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
 626                  "type: %d\n", type);
 627
 628         res = ldlm_resource_new();
 629         if (!res)
 630                 RETURN(NULL);
 631
 632         res->lr_name = name;
 633         res->lr_namespace = ns;
 634         res->lr_type = type;
 635         res->lr_most_restr = LCK_NL;
 636
 637         spin_lock(&ns->ns_hash_lock);
 638         old_res = ldlm_resource_find(ns, name, hash);
 639         if (old_res) {
 640                 /* someone won the race and added the resource before */
 641                 ldlm_resource_getref(old_res);
 642                 spin_unlock(&ns->ns_hash_lock);
 643                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 644                 /* synchronize WRT resource creation */
 645                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 646                         down(&old_res->lr_lvb_sem);
 647                         up(&old_res->lr_lvb_sem);
 648                 }
 649                 RETURN(old_res);
 650         }
 651
 652         /* we won! let's add the resource */
 653         bucket = ns->ns_hash + hash;
 654         list_add(&res->lr_hash, bucket);
 655         ns->ns_resources++;
 656         ns->ns_refcount++;
 657
 658         if (parent == NULL) {
 659                 list_add(&res->lr_childof, &ns->ns_root_list);
 660         } else {
 661                 res->lr_parent = parent;
 662                 list_add(&res->lr_childof, &parent->lr_children);
 663         }
 664         spin_unlock(&ns->ns_hash_lock);
 665
 666         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 667                 int rc;
 668
 669                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
 670                 rc = ns->ns_lvbo->lvbo_init(res);
 671                 if (rc)
 672                         CERROR("lvbo_init failed for resource "
 673                                LPU64": rc %d\n", name.name[0], rc);
 674                 /* we create resource with locked lr_lvb_sem */
 675                 up(&res->lr_lvb_sem);
 676         }
 677
 678         RETURN(res);
 679 }
 680
 681 /* Args: unlocked namespace
 682  * Locks: takes and releases ns->ns_lock and res->lr_lock
 683  * Returns: referenced, unlocked ldlm_resource or NULL */
 684 struct ldlm_resource *
 685 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 686                   struct ldlm_res_id name, ldlm_type_t type, int create)
 687 {
 688         __u32 hash = ldlm_hash_fn(parent, name);
 689         struct ldlm_resource *res = NULL;
 690         ENTRY;
 691
 692         LASSERT(ns != NULL);
 693         LASSERT(ns->ns_hash != NULL);
 694         LASSERT(name.name[0] != 0);
 695
 696         spin_lock(&ns->ns_hash_lock);
 697         res = ldlm_resource_find(ns, name, hash);
 698         if (res) {
 699                 ldlm_resource_getref(res);
 700                 spin_unlock(&ns->ns_hash_lock);
 701                 /* synchronize WRT resource creation */
 702                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 703                         down(&res->lr_lvb_sem);
 704                         up(&res->lr_lvb_sem);
 705                 }
 706                 RETURN(res);
 707         }
 708         spin_unlock(&ns->ns_hash_lock);
 709
 710         if (create == 0)
 711                 RETURN(NULL);
 712
 713         res = ldlm_resource_add(ns, parent, name, hash, type);
 714         RETURN(res);
 715 }
 716
 717 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 718 {
 719         LASSERT(res != NULL);
 720         LASSERT(res != LP_POISON);
 721         atomic_inc(&res->lr_refcount);
 722         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 723                atomic_read(&res->lr_refcount));
 724         return res;
 725 }
 726
 727 void __ldlm_resource_putref_final(struct ldlm_resource *res)
 728 {
 729         struct ldlm_namespace *ns = res->lr_namespace;
 730
 731         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 732
 733         if (!list_empty(&res->lr_granted)) {
 734                 ldlm_resource_dump(D_ERROR, res);
 735                 LBUG();
 736         }
 737
 738         if (!list_empty(&res->lr_converting)) {
 739                 ldlm_resource_dump(D_ERROR, res);
 740                 LBUG();
 741         }
 742
 743         if (!list_empty(&res->lr_waiting)) {
 744                 ldlm_resource_dump(D_ERROR, res);
 745                 LBUG();
 746         }
 747
 748         if (!list_empty(&res->lr_children)) {
 749                 ldlm_resource_dump(D_ERROR, res);
 750                 LBUG();
 751         }
 752
 753         ns->ns_refcount--;
 754         list_del_init(&res->lr_hash);
 755         list_del_init(&res->lr_childof);
 756
 757         ns->ns_resources--;
 758         if (ns->ns_resources == 0)
 759                 wake_up(&ns->ns_waitq);
 760 }
 761
 762 /* Returns 1 if the resource was freed, 0 if it remains. */
 763 int ldlm_resource_putref(struct ldlm_resource *res)
 764 {
 765         struct ldlm_namespace *ns = res->lr_namespace;
 766         int rc = 0;
 767         ENTRY;
 768
 769         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 770                atomic_read(&res->lr_refcount) - 1);
 771         LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
 772                  atomic_read(&res->lr_refcount));
 773         LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
 774                  atomic_read(&res->lr_refcount));
 775
 776         if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
 777                 __ldlm_resource_putref_final(res);
 778                 spin_unlock(&ns->ns_hash_lock);
 779                 if (res->lr_lvb_data)
 780                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 781                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 782                 rc = 1;
 783         }
 784
 785         RETURN(rc);
 786 }
 787
 788 /* Returns 1 if the resource was freed, 0 if it remains. */
 789 int ldlm_resource_putref_locked(struct ldlm_resource *res)
 790 {
 791         int rc = 0;
 792         ENTRY;
 793
 794         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 795                atomic_read(&res->lr_refcount) - 1);
 796         LASSERT(atomic_read(&res->lr_refcount) > 0);
 797         LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
 798
 799         LASSERT(atomic_read(&res->lr_refcount) >= 0);
 800         if (atomic_dec_and_test(&res->lr_refcount)) {
 801                 __ldlm_resource_putref_final(res);
 802                 if (res->lr_lvb_data)
 803                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 804                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 805                 rc = 1;
 806         }
 807
 808         RETURN(rc);
 809 }
 810
 811 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 812                             struct ldlm_lock *lock)
 813 {
 814         check_res_locked(res);
 815
 816         ldlm_resource_dump(D_OTHER, res);
 817         CDEBUG(D_OTHER, "About to add this lock:\n");
 818         ldlm_lock_dump(D_OTHER, lock, 0);
 819
 820         if (lock->l_destroyed) {
 821                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 822                 return;
 823         }
 824
 825         LASSERT(list_empty(&lock->l_res_link));
 826
 827         list_add_tail(&lock->l_res_link, head);
 828 }
 829
 830 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
 831                                      struct ldlm_lock *new)
 832 {
 833         struct ldlm_resource *res = original->l_resource;
 834
 835         check_res_locked(res);
 836
 837         ldlm_resource_dump(D_OTHER, res);
 838         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
 839         ldlm_lock_dump(D_OTHER, new, 0);
 840
 841         if (new->l_destroyed) {
 842                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 843                 goto out;
 844         }
 845
 846         LASSERT(list_empty(&new->l_res_link));
 847
 848         list_add(&new->l_res_link, &original->l_res_link);
 849  out:;
 850 }
 851
 852 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 853 {
 854         check_res_locked(lock->l_resource);
 855         ldlm_unlink_lock_skiplist(lock);
 856         list_del_init(&lock->l_res_link);
 857 }
 858
 859 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 860 {
 861         desc->lr_type = res->lr_type;
 862         desc->lr_name = res->lr_name;
 863 }
 864
 865 void ldlm_dump_all_namespaces(int level)
 866 {
 867         struct list_head *tmp;
 868
 869         if (!((libcfs_debug | D_ERROR) & level))
 870                 return;
 871
 872         mutex_down(&ldlm_namespace_lock);
 873
 874         list_for_each(tmp, &ldlm_namespace_list) {
 875                 struct ldlm_namespace *ns;
 876                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 877                 ldlm_namespace_dump(level, ns);
 878         }
 879
 880         mutex_up(&ldlm_namespace_lock);
 881 }
 882
 883 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
 884 {
 885         struct list_head *tmp;
 886
 887         if (!((libcfs_debug | D_ERROR) & level))
 888                 return;
 889
 890         CDEBUG(level, "--- Namespace: %s (rc: %d, client: %d)\n",
 891                ns->ns_name, ns->ns_refcount, ns->ns_client);
 892
 893         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
 894                 return;
 895
 896         spin_lock(&ns->ns_hash_lock);
 897         tmp = ns->ns_root_list.next;
 898         while (tmp != &ns->ns_root_list) {
 899                 struct ldlm_resource *res;
 900                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
 901
 902                 ldlm_resource_getref(res);
 903                 spin_unlock(&ns->ns_hash_lock);
 904
 905                 lock_res(res);
 906                 ldlm_resource_dump(level, res);
 907                 unlock_res(res);
 908
 909                 spin_lock(&ns->ns_hash_lock);
 910                 tmp = tmp->next;
 911                 ldlm_resource_putref_locked(res);
 912         }
 913         ns->ns_next_dump = cfs_time_shift(10);
 914         spin_unlock(&ns->ns_hash_lock);
 915 }
 916
 917 void ldlm_resource_dump(int level, struct ldlm_resource *res)
 918 {
 919         struct list_head *tmp;
 920         int pos;
 921
 922         CLASSERT(RES_NAME_SIZE == 4);
 923
 924         if (!((libcfs_debug | D_ERROR) & level))
 925                 return;
 926
 927         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
 928                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
 929                res->lr_name.name[2], res->lr_name.name[3],
 930                atomic_read(&res->lr_refcount));
 931
 932         if (!list_empty(&res->lr_granted)) {
 933                 pos = 0;
 934                 CDEBUG(level, "Granted locks:\n");
 935                 list_for_each(tmp, &res->lr_granted) {
 936                         struct ldlm_lock *lock;
 937                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 938                         ldlm_lock_dump(level, lock, ++pos);
 939                 }
 940         }
 941         if (!list_empty(&res->lr_converting)) {
 942                 pos = 0;
 943                 CDEBUG(level, "Converting locks:\n");
 944                 list_for_each(tmp, &res->lr_converting) {
 945                         struct ldlm_lock *lock;
 946                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 947                         ldlm_lock_dump(level, lock, ++pos);
 948                 }
 949         }
 950         if (!list_empty(&res->lr_waiting)) {
 951                 pos = 0;
 952                 CDEBUG(level, "Waiting locks:\n");
 953                 list_for_each(tmp, &res->lr_waiting) {
 954                         struct ldlm_lock *lock;
 955                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 956                         ldlm_lock_dump(level, lock, ++pos);
 957                 }
 958         }
 959 }