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         int lru_resize;
 142
 143         dummy[MAX_STRING_SIZE] = '\0';
 144         if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 145                 return -EFAULT;
 146
 147         if (count == 6 && memcmp(dummy, "clear", 5) == 0) {
 148                 CDEBUG(D_DLMTRACE,
 149                        "dropping all unused locks from namespace %s\n",
 150                        ns->ns_name);
 151                 if (ns_connect_lru_resize(ns)) {
 152                         int canceled, unused  = ns->ns_nr_unused;
 153
 154                         /* Try to cancel all @ns_nr_unused locks. */
 155                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
 156                                                    LDLM_CANCEL_PASSED);
 157                         if (canceled < unused) {
 158                                 CERROR("not all requested locks are canceled, "
 159                                        "requested: %d, canceled: %d\n", unused,
 160                                        canceled);
 161                                 return -EINVAL;
 162                         }
 163                 } else {
 164                         tmp = ns->ns_max_unused;
 165                         ns->ns_max_unused = 0;
 166                         ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
 167                         ns->ns_max_unused = tmp;
 168                 }
 169                 return count;
 170         }
 171
 172         tmp = simple_strtoul(dummy, &end, 0);
 173         if (dummy == end) {
 174                 CERROR("invalid value written\n");
 175                 return -EINVAL;
 176         }
 177         lru_resize = (tmp == 0);
 178
 179         if (ns_connect_lru_resize(ns)) {
 180                 if (!lru_resize)
 181                         ns->ns_max_unused = (unsigned int)tmp;
 182
 183                 if (tmp > ns->ns_nr_unused)
 184                         tmp = ns->ns_nr_unused;
 185                 tmp = ns->ns_nr_unused - tmp;
 186
 187                 CDEBUG(D_DLMTRACE, "changing namespace %s unused locks from %u to %u\n",
 188                        ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
 189                 ldlm_cancel_lru(ns, (unsigned int)tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 190
 191                 if (!lru_resize) {
 192                         CDEBUG(D_DLMTRACE, "disable lru_resize for namespace %s\n",
 193                                ns->ns_name);
 194                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 195                 }
 196         } else {
 197                 CDEBUG(D_DLMTRACE, "changing namespace %s max_unused from %u to %u\n",
 198                        ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
 199                 ns->ns_max_unused = (unsigned int)tmp;
 200                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 201
 202                 /* Make sure that originally lru resize was supported before
 203                  * turning it on here. */
 204                 if (lru_resize &&
 205                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 206                         CDEBUG(D_DLMTRACE, "enable lru_resize for namespace %s\n",
 207                                ns->ns_name);
 208                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 209                 }
 210         }
 211
 212         return count;
 213 }
 214
 215 void ldlm_proc_namespace(struct ldlm_namespace *ns)
 216 {
 217         struct lprocfs_vars lock_vars[2];
 218         char lock_name[MAX_STRING_SIZE + 1];
 219
 220         LASSERT(ns != NULL);
 221         LASSERT(ns->ns_name != NULL);
 222
 223         lock_name[MAX_STRING_SIZE] = '\0';
 224
 225         memset(lock_vars, 0, sizeof(lock_vars));
 226         lock_vars[0].name = lock_name;
 227
 228         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
 229         lock_vars[0].data = &ns->ns_refcount;
 230         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 231         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 232
 233         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
 234         lock_vars[0].data = &ns->ns_locks;
 235         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 236         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 237
 238         if (ns_is_client(ns)) {
 239                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 240                          ns->ns_name);
 241                 lock_vars[0].data = &ns->ns_nr_unused;
 242                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 243                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 244
 245                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 246                          ns->ns_name);
 247                 lock_vars[0].data = ns;
 248                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 249                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 250                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 251
 252                 snprintf(lock_name, MAX_STRING_SIZE, "%s/shrink_thumb",
 253                          ns->ns_name);
 254                 lock_vars[0].data = ns;
 255                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 256                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 257                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 258
 259                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 260                          ns->ns_name);
 261                 lock_vars[0].data = &ns->ns_max_age;
 262                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 263                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 264                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 265         } else {
 266                 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
 267                          ns->ns_name);
 268                 lock_vars[0].data = &ns->ns_ctime_age_limit;
 269                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 270                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 271                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 272
 273                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 274                          ns->ns_name);
 275                 lock_vars[0].data = &ns->ns_max_nolock_size;
 276                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 277                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 278                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 279
 280                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 281                          ns->ns_name);
 282                 lock_vars[0].data = &ns->ns_contention_time;
 283                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 284                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 285                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 286
 287                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 288                          ns->ns_name);
 289                 lock_vars[0].data = &ns->ns_contended_locks;
 290                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 291                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 292                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 293         }
 294 }
 295 #undef MAX_STRING_SIZE
 296 #else
 297 #define ldlm_proc_namespace(ns) do {} while (0)
 298 #endif /* LPROCFS */
 299
 300 struct ldlm_namespace *ldlm_namespace_new(char *name, ldlm_side_t client,
 301                                           ldlm_appetite_t apt)
 302 {
 303         struct ldlm_namespace *ns = NULL;
 304         struct list_head *bucket;
 305         int rc, idx, namelen;
 306         ENTRY;
 307
 308         rc = ldlm_get_ref();
 309         if (rc) {
 310                 CERROR("ldlm_get_ref failed: %d\n", rc);
 311                 RETURN(NULL);
 312         }
 313
 314         OBD_ALLOC_PTR(ns);
 315         if (!ns)
 316                 GOTO(out_ref, NULL);
 317
 318         OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 319         if (!ns->ns_hash)
 320                 GOTO(out_ns, NULL);
 321
 322         ns->ns_shrink_thumb = LDLM_LOCK_SHRINK_THUMB;
 323         ns->ns_appetite = apt;
 324         namelen = strlen(name);
 325         OBD_ALLOC(ns->ns_name, namelen + 1);
 326         if (!ns->ns_name)
 327                 GOTO(out_hash, NULL);
 328
 329         strcpy(ns->ns_name, name);
 330
 331         CFS_INIT_LIST_HEAD(&ns->ns_root_list);
 332         ns->ns_refcount = 0;
 333         ns->ns_client = client;
 334         spin_lock_init(&ns->ns_hash_lock);
 335         atomic_set(&ns->ns_locks, 0);
 336         ns->ns_resources = 0;
 337         cfs_waitq_init(&ns->ns_waitq);
 338         ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
 339         ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
 340         ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
 341
 342         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 343              bucket--)
 344                 CFS_INIT_LIST_HEAD(bucket);
 345
 346         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 347         ns->ns_nr_unused = 0;
 348         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
 349         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
 350         ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
 351         spin_lock_init(&ns->ns_unused_lock);
 352         ns->ns_orig_connect_flags = 0;
 353         ns->ns_connect_flags = 0;
 354         ldlm_proc_namespace(ns);
 355
 356         idx = atomic_read(ldlm_namespace_nr(client));
 357         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 358         if (rc) {
 359                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 360                 GOTO(out_proc, rc);
 361         }
 362
 363         mutex_down(ldlm_namespace_lock(client));
 364         list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 365         atomic_inc(ldlm_namespace_nr(client));
 366         mutex_up(ldlm_namespace_lock(client));
 367
 368         RETURN(ns);
 369 out_proc:
 370         ldlm_namespace_cleanup(ns, 0);
 371         OBD_FREE(ns->ns_name, namelen + 1);
 372 out_hash:
 373         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 374 out_ns:
 375         OBD_FREE_PTR(ns);
 376 out_ref:
 377         ldlm_put_ref(0);
 378         RETURN(NULL);
 379 }
 380
 381 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 382
 383 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 384  * This is currently only used for recovery, and we make certain assumptions
 385  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 386  *
 387  * Called with the ns_lock held. */
 388 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 389                              int flags)
 390 {
 391         struct list_head *tmp;
 392         int rc = 0, client = ns_is_client(res->lr_namespace);
 393         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 394         ENTRY;
 395
 396
 397         do {
 398                 struct ldlm_lock *lock = NULL;
 399
 400                 /* first, we look for non-cleaned-yet lock
 401                  * all cleaned locks are marked by CLEANED flag */
 402                 lock_res(res);
 403                 list_for_each(tmp, q) {
 404                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 405                         if (lock->l_flags & LDLM_FL_CLEANED) {
 406                                 lock = NULL;
 407                                 continue;
 408                         }
 409                         LDLM_LOCK_GET(lock);
 410                         lock->l_flags |= LDLM_FL_CLEANED;
 411                         break;
 412                 }
 413
 414                 if (lock == NULL) {
 415                         unlock_res(res);
 416                         break;
 417                 }
 418
 419                 /* Set CBPENDING so nothing in the cancellation path
 420                  * can match this lock */
 421                 lock->l_flags |= LDLM_FL_CBPENDING;
 422                 lock->l_flags |= LDLM_FL_FAILED;
 423                 lock->l_flags |= flags;
 424
 425                 /* ... without sending a CANCEL message for local_only. */
 426                 if (local_only)
 427                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 428
 429                 if (local_only && (lock->l_readers || lock->l_writers)) {
 430                         /* This is a little bit gross, but much better than the
 431                          * alternative: pretend that we got a blocking AST from
 432                          * the server, so that when the lock is decref'd, it
 433                          * will go away ... */
 434                         unlock_res(res);
 435                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 436                         if (lock->l_completion_ast)
 437                                 lock->l_completion_ast(lock, 0, NULL);
 438                         LDLM_LOCK_PUT(lock);
 439                         continue;
 440                 }
 441
 442                 if (client) {
 443                         struct lustre_handle lockh;
 444
 445                         unlock_res(res);
 446                         ldlm_lock2handle(lock, &lockh);
 447                         rc = ldlm_cli_cancel(&lockh);
 448                         if (rc)
 449                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 450                 } else {
 451                         ldlm_resource_unlink_lock(lock);
 452                         unlock_res(res);
 453                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 454                                    "client node");
 455                         ldlm_lock_destroy(lock);
 456                 }
 457                 LDLM_LOCK_PUT(lock);
 458         } while (1);
 459
 460         EXIT;
 461 }
 462
 463 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 464 {
 465         struct list_head *tmp;
 466         int i;
 467
 468         if (ns == NULL) {
 469                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 470                 return ELDLM_OK;
 471         }
 472
 473         for (i = 0; i < RES_HASH_SIZE; i++) {
 474                 spin_lock(&ns->ns_hash_lock);
 475                 tmp = ns->ns_hash[i].next;
 476                 while (tmp != &(ns->ns_hash[i])) {
 477                         struct ldlm_resource *res;
 478                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 479                         ldlm_resource_getref(res);
 480                         spin_unlock(&ns->ns_hash_lock);
 481
 482                         cleanup_resource(res, &res->lr_granted, flags);
 483                         cleanup_resource(res, &res->lr_converting, flags);
 484                         cleanup_resource(res, &res->lr_waiting, flags);
 485
 486                         spin_lock(&ns->ns_hash_lock);
 487                         tmp  = tmp->next;
 488
 489                         /* XXX: former stuff caused issues in case of race
 490                          * between ldlm_namespace_cleanup() and lockd() when
 491                          * client gets blocking ast when lock gets distracted by
 492                          * server. This is 1_4 branch solution, let's see how
 493                          * will it behave. */
 494                         if (!ldlm_resource_putref_locked(res))
 495                                 CDEBUG(D_INFO,
 496                                        "Namespace %s resource refcount nonzero "
 497                                        "(%d) after lock cleanup; forcing cleanup.\n",
 498                                        ns->ns_name, atomic_read(&res->lr_refcount));
 499                 }
 500                 spin_unlock(&ns->ns_hash_lock);
 501         }
 502
 503         return ELDLM_OK;
 504 }
 505
 506 /* Cleanup, but also free, the namespace */
 507 int ldlm_namespace_free_prior(struct ldlm_namespace *ns)
 508 {
 509         ENTRY;
 510         if (!ns)
 511                 RETURN(ELDLM_OK);
 512
 513         mutex_down(ldlm_namespace_lock(ns->ns_client));
 514         /*
 515          * Some asserts and possibly other parts of code still using
 516          * list_empty(&ns->ns_list_chain). This is why it is important
 517          * to use list_del_init() here.
 518          */
 519         list_del_init(&ns->ns_list_chain);
 520         atomic_dec(ldlm_namespace_nr(ns->ns_client));
 521         ldlm_pool_fini(&ns->ns_pool);
 522         mutex_up(ldlm_namespace_lock(ns->ns_client));
 523
 524         /* At shutdown time, don't call the cancellation callback */
 525         ldlm_namespace_cleanup(ns, 0);
 526
 527         if (ns->ns_refcount > 0) {
 528                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 529                 int rc;
 530                 CDEBUG(D_DLMTRACE,
 531                        "dlm namespace %s free waiting on refcount %d\n",
 532                        ns->ns_name, ns->ns_refcount);
 533                 rc = l_wait_event(ns->ns_waitq,
 534                                   ns->ns_refcount == 0, &lwi);
 535                 if (ns->ns_refcount)
 536                         LCONSOLE_ERROR_MSG(0x139, "Lock manager: wait for %s "
 537                                            "namespace cleanup aborted with %d "
 538                                            "resources in use. (%d)\nI'm going "
 539                                            "to try to clean up anyway, but I "
 540                                            "might need a reboot of this node.\n",
 541                                             ns->ns_name, (int) ns->ns_refcount,
 542                                             rc);
 543                 CDEBUG(D_DLMTRACE,
 544                        "dlm namespace %s free done waiting\n", ns->ns_name);
 545         }
 546
 547         RETURN(ELDLM_OK);
 548 }
 549
 550 int ldlm_namespace_free_post(struct ldlm_namespace *ns, int force)
 551 {
 552         ENTRY;
 553         if (!ns)
 554                 RETURN(ELDLM_OK);
 555
 556 #ifdef LPROCFS
 557         {
 558                 struct proc_dir_entry *dir;
 559                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
 560                 if (dir == NULL) {
 561                         CERROR("dlm namespace %s has no procfs dir?\n",
 562                                ns->ns_name);
 563                 } else {
 564                         lprocfs_remove(&dir);
 565                 }
 566         }
 567 #endif
 568
 569         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 570         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 571         /*
 572          * @ns should be not on list in this time, otherwise this will cause
 573          * issues realted to using freed @ns in pools thread.
 574          */
 575         LASSERT(list_empty(&ns->ns_list_chain));
 576         OBD_FREE_PTR(ns);
 577         ldlm_put_ref(force);
 578         RETURN(ELDLM_OK);
 579 }
 580
 581
 582 /* Cleanup the resource, and free namespace.
 583  * bug 12864:
 584  * Deadlock issue:
 585  * proc1: destroy import
 586  *        class_disconnect_export(grab cl_sem) ->
 587  *              -> ldlm_namespace_free ->
 588  *              -> lprocfs_remove(grab _lprocfs_lock).
 589  * proc2: read proc info
 590  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 591  *              -> osc_rd_active, etc(grab cl_sem).
 592  *
 593  * So that I have to split the ldlm_namespace_free into two parts - the first
 594  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 595  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 596  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 597  * held.
 598  */
 599 int ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 600 {
 601         ldlm_namespace_free_prior(ns);
 602         ldlm_namespace_free_post(ns, force);
 603         return ELDLM_OK;
 604 }
 605
 606
 607 void ldlm_namespace_get_nolock(struct ldlm_namespace *ns)
 608 {
 609         LASSERT(ns->ns_refcount >= 0);
 610         ns->ns_refcount++;
 611 }
 612
 613 void ldlm_namespace_get(struct ldlm_namespace *ns)
 614 {
 615         spin_lock(&ns->ns_hash_lock);
 616         ldlm_namespace_get_nolock(ns);
 617         spin_unlock(&ns->ns_hash_lock);
 618 }
 619
 620 void ldlm_namespace_put_nolock(struct ldlm_namespace *ns, int wakeup)
 621 {
 622         LASSERT(ns->ns_refcount > 0);
 623         ns->ns_refcount--;
 624         if (ns->ns_refcount == 0 && wakeup)
 625                 wake_up(&ns->ns_waitq);
 626 }
 627
 628 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
 629 {
 630         spin_lock(&ns->ns_hash_lock);
 631         ldlm_namespace_put_nolock(ns, wakeup);
 632         spin_unlock(&ns->ns_hash_lock);
 633 }
 634
 635 /* Should be called under ldlm_namespace_lock(client) taken */
 636 void ldlm_namespace_move(struct ldlm_namespace *ns, ldlm_side_t client)
 637 {
 638         LASSERT(!list_empty(&ns->ns_list_chain));
 639         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 640         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 641 }
 642
 643 /* Should be called under ldlm_namespace_lock(client) taken */
 644 struct ldlm_namespace *ldlm_namespace_first(ldlm_side_t client)
 645 {
 646         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 647         LASSERT(!list_empty(ldlm_namespace_list(client)));
 648         return container_of(ldlm_namespace_list(client)->next,
 649                 struct ldlm_namespace, ns_list_chain);
 650 }
 651 static __u32 ldlm_hash_fn(struct ldlm_resource *parent,
 652                           const struct ldlm_res_id *name)
 653 {
 654         __u32 hash = 0;
 655         int i;
 656
 657         for (i = 0; i < RES_NAME_SIZE; i++)
 658                 hash += name->name[i];
 659
 660         hash += (__u32)((unsigned long)parent >> 4);
 661
 662         return (hash & RES_HASH_MASK);
 663 }
 664
 665 static struct ldlm_resource *ldlm_resource_new(void)
 666 {
 667         struct ldlm_resource *res;
 668         int idx;
 669
 670         OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
 671         if (res == NULL)
 672                 return NULL;
 673
 674         memset(res, 0, sizeof(*res));
 675
 676         CFS_INIT_LIST_HEAD(&res->lr_children);
 677         CFS_INIT_LIST_HEAD(&res->lr_childof);
 678         CFS_INIT_LIST_HEAD(&res->lr_granted);
 679         CFS_INIT_LIST_HEAD(&res->lr_converting);
 680         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 681
 682         /* initialize interval trees for each lock mode*/
 683         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 684                 res->lr_itree[idx].lit_size = 0;
 685                 res->lr_itree[idx].lit_mode = 1 << idx;
 686                 res->lr_itree[idx].lit_root = NULL;
 687         }
 688
 689         atomic_set(&res->lr_refcount, 1);
 690         spin_lock_init(&res->lr_lock);
 691
 692         /* one who creates the resource must unlock
 693          * the semaphore after lvb initialization */
 694         init_MUTEX_LOCKED(&res->lr_lvb_sem);
 695
 696         return res;
 697 }
 698
 699 /* must be called with hash lock held */
 700 static struct ldlm_resource *
 701 ldlm_resource_find(struct ldlm_namespace *ns, const struct ldlm_res_id *name,
 702                    __u32 hash)
 703 {
 704         struct list_head *bucket, *tmp;
 705         struct ldlm_resource *res;
 706
 707         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 708         bucket = ns->ns_hash + hash;
 709
 710         list_for_each(tmp, bucket) {
 711                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 712                 if (memcmp(&res->lr_name, name, sizeof(res->lr_name)) == 0)
 713                         return res;
 714         }
 715
 716         return NULL;
 717 }
 718
 719 /* Args: locked namespace
 720  * Returns: newly-allocated, referenced, unlocked resource */
 721 static struct ldlm_resource *
 722 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 723                   const struct ldlm_res_id *name, __u32 hash, ldlm_type_t type)
 724 {
 725         struct list_head *bucket;
 726         struct ldlm_resource *res, *old_res;
 727         ENTRY;
 728
 729         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
 730                  "type: %d\n", type);
 731
 732         res = ldlm_resource_new();
 733         if (!res)
 734                 RETURN(NULL);
 735
 736         res->lr_name = *name;
 737         res->lr_namespace = ns;
 738         res->lr_type = type;
 739         res->lr_most_restr = LCK_NL;
 740
 741         spin_lock(&ns->ns_hash_lock);
 742         old_res = ldlm_resource_find(ns, name, hash);
 743         if (old_res) {
 744                 /* someone won the race and added the resource before */
 745                 ldlm_resource_getref(old_res);
 746                 spin_unlock(&ns->ns_hash_lock);
 747                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 748                 /* synchronize WRT resource creation */
 749                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 750                         down(&old_res->lr_lvb_sem);
 751                         up(&old_res->lr_lvb_sem);
 752                 }
 753                 RETURN(old_res);
 754         }
 755
 756         /* we won! let's add the resource */
 757         bucket = ns->ns_hash + hash;
 758         list_add(&res->lr_hash, bucket);
 759         ns->ns_resources++;
 760         ldlm_namespace_get_nolock(ns);
 761
 762         if (parent == NULL) {
 763                 list_add(&res->lr_childof, &ns->ns_root_list);
 764         } else {
 765                 res->lr_parent = parent;
 766                 list_add(&res->lr_childof, &parent->lr_children);
 767         }
 768         spin_unlock(&ns->ns_hash_lock);
 769
 770         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 771                 int rc;
 772
 773                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
 774                 rc = ns->ns_lvbo->lvbo_init(res);
 775                 if (rc)
 776                         CERROR("lvbo_init failed for resource "
 777                                LPU64": rc %d\n", name->name[0], rc);
 778                 /* we create resource with locked lr_lvb_sem */
 779                 up(&res->lr_lvb_sem);
 780         }
 781
 782         RETURN(res);
 783 }
 784
 785 /* Args: unlocked namespace
 786  * Locks: takes and releases ns->ns_lock and res->lr_lock
 787  * Returns: referenced, unlocked ldlm_resource or NULL */
 788 struct ldlm_resource *
 789 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 790                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
 791 {
 792         __u32 hash = ldlm_hash_fn(parent, name);
 793         struct ldlm_resource *res = NULL;
 794         ENTRY;
 795
 796         LASSERT(ns != NULL);
 797         LASSERT(ns->ns_hash != NULL);
 798         LASSERT(name->name[0] != 0);
 799
 800         spin_lock(&ns->ns_hash_lock);
 801         res = ldlm_resource_find(ns, name, hash);
 802         if (res) {
 803                 ldlm_resource_getref(res);
 804                 spin_unlock(&ns->ns_hash_lock);
 805                 /* synchronize WRT resource creation */
 806                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 807                         down(&res->lr_lvb_sem);
 808                         up(&res->lr_lvb_sem);
 809                 }
 810                 RETURN(res);
 811         }
 812         spin_unlock(&ns->ns_hash_lock);
 813
 814         if (create == 0)
 815                 RETURN(NULL);
 816
 817         res = ldlm_resource_add(ns, parent, name, hash, type);
 818         RETURN(res);
 819 }
 820
 821 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 822 {
 823         LASSERT(res != NULL);
 824         LASSERT(res != LP_POISON);
 825         atomic_inc(&res->lr_refcount);
 826         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 827                atomic_read(&res->lr_refcount));
 828         return res;
 829 }
 830
 831 void __ldlm_resource_putref_final(struct ldlm_resource *res)
 832 {
 833         struct ldlm_namespace *ns = res->lr_namespace;
 834
 835         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 836
 837         if (!list_empty(&res->lr_granted)) {
 838                 ldlm_resource_dump(D_ERROR, res);
 839                 LBUG();
 840         }
 841
 842         if (!list_empty(&res->lr_converting)) {
 843                 ldlm_resource_dump(D_ERROR, res);
 844                 LBUG();
 845         }
 846
 847         if (!list_empty(&res->lr_waiting)) {
 848                 ldlm_resource_dump(D_ERROR, res);
 849                 LBUG();
 850         }
 851
 852         if (!list_empty(&res->lr_children)) {
 853                 ldlm_resource_dump(D_ERROR, res);
 854                 LBUG();
 855         }
 856
 857         /* Pass 0 here to not wake ->ns_waitq up yet, we will do it few
 858          * lines below when all children are freed. */
 859         ldlm_namespace_put_nolock(ns, 0);
 860         list_del_init(&res->lr_hash);
 861         list_del_init(&res->lr_childof);
 862
 863         ns->ns_resources--;
 864         if (ns->ns_resources == 0)
 865                 wake_up(&ns->ns_waitq);
 866 }
 867
 868 /* Returns 1 if the resource was freed, 0 if it remains. */
 869 int ldlm_resource_putref(struct ldlm_resource *res)
 870 {
 871         struct ldlm_namespace *ns = res->lr_namespace;
 872         int rc = 0;
 873         ENTRY;
 874
 875         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 876                atomic_read(&res->lr_refcount) - 1);
 877         LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
 878                  atomic_read(&res->lr_refcount));
 879         LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
 880                  atomic_read(&res->lr_refcount));
 881
 882         if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
 883                 __ldlm_resource_putref_final(res);
 884                 spin_unlock(&ns->ns_hash_lock);
 885                 if (res->lr_lvb_data)
 886                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 887                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 888                 rc = 1;
 889         }
 890
 891         RETURN(rc);
 892 }
 893
 894 /* Returns 1 if the resource was freed, 0 if it remains. */
 895 int ldlm_resource_putref_locked(struct ldlm_resource *res)
 896 {
 897         int rc = 0;
 898         ENTRY;
 899
 900         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 901                atomic_read(&res->lr_refcount) - 1);
 902         LASSERT(atomic_read(&res->lr_refcount) > 0);
 903         LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
 904
 905         LASSERT(atomic_read(&res->lr_refcount) >= 0);
 906         if (atomic_dec_and_test(&res->lr_refcount)) {
 907                 __ldlm_resource_putref_final(res);
 908                 if (res->lr_lvb_data)
 909                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 910                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 911                 rc = 1;
 912         }
 913
 914         RETURN(rc);
 915 }
 916
 917 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 918                             struct ldlm_lock *lock)
 919 {
 920         check_res_locked(res);
 921
 922         ldlm_resource_dump(D_OTHER, res);
 923         CDEBUG(D_OTHER, "About to add this lock:\n");
 924         ldlm_lock_dump(D_OTHER, lock, 0);
 925
 926         if (lock->l_destroyed) {
 927                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 928                 return;
 929         }
 930
 931         LASSERT(list_empty(&lock->l_res_link));
 932
 933         list_add_tail(&lock->l_res_link, head);
 934 }
 935
 936 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
 937                                      struct ldlm_lock *new)
 938 {
 939         struct ldlm_resource *res = original->l_resource;
 940
 941         check_res_locked(res);
 942
 943         ldlm_resource_dump(D_OTHER, res);
 944         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
 945         ldlm_lock_dump(D_OTHER, new, 0);
 946
 947         if (new->l_destroyed) {
 948                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 949                 goto out;
 950         }
 951
 952         LASSERT(list_empty(&new->l_res_link));
 953
 954         list_add(&new->l_res_link, &original->l_res_link);
 955  out:;
 956 }
 957
 958 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 959 {
 960         int type = lock->l_resource->lr_type;
 961
 962         check_res_locked(lock->l_resource);
 963         if (type == LDLM_IBITS || type == LDLM_PLAIN)
 964                 ldlm_unlink_lock_skiplist(lock);
 965         else if (type == LDLM_EXTENT)
 966                 ldlm_extent_unlink_lock(lock);
 967         list_del_init(&lock->l_res_link);
 968 }
 969
 970 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 971 {
 972         desc->lr_type = res->lr_type;
 973         desc->lr_name = res->lr_name;
 974 }
 975
 976 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
 977 {
 978         struct list_head *tmp;
 979
 980         if (!((libcfs_debug | D_ERROR) & level))
 981                 return;
 982
 983         mutex_down(ldlm_namespace_lock(client));
 984
 985         list_for_each(tmp, ldlm_namespace_list(client)) {
 986                 struct ldlm_namespace *ns;
 987                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 988                 ldlm_namespace_dump(level, ns);
 989         }
 990
 991         mutex_up(ldlm_namespace_lock(client));
 992 }
 993
 994 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
 995 {
 996         struct list_head *tmp;
 997
 998         if (!((libcfs_debug | D_ERROR) & level))
 999                 return;
1000
1001         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1002                ns->ns_name, ns->ns_refcount,
1003                ns_is_client(ns) ? "client" : "server");
1004
1005         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1006                 return;
1007
1008         spin_lock(&ns->ns_hash_lock);
1009         tmp = ns->ns_root_list.next;
1010         while (tmp != &ns->ns_root_list) {
1011                 struct ldlm_resource *res;
1012                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
1013
1014                 ldlm_resource_getref(res);
1015                 spin_unlock(&ns->ns_hash_lock);
1016
1017                 lock_res(res);
1018                 ldlm_resource_dump(level, res);
1019                 unlock_res(res);
1020
1021                 spin_lock(&ns->ns_hash_lock);
1022                 tmp = tmp->next;
1023                 ldlm_resource_putref_locked(res);
1024         }
1025         ns->ns_next_dump = cfs_time_shift(10);
1026         spin_unlock(&ns->ns_hash_lock);
1027 }
1028
1029 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1030 {
1031         struct list_head *tmp;
1032         int pos;
1033
1034         CLASSERT(RES_NAME_SIZE == 4);
1035
1036         if (!((libcfs_debug | D_ERROR) & level))
1037                 return;
1038
1039         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1040                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1041                res->lr_name.name[2], res->lr_name.name[3],
1042                atomic_read(&res->lr_refcount));
1043
1044         if (!list_empty(&res->lr_granted)) {
1045                 pos = 0;
1046                 CDEBUG(level, "Granted locks:\n");
1047                 list_for_each(tmp, &res->lr_granted) {
1048                         struct ldlm_lock *lock;
1049                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1050                         ldlm_lock_dump(level, lock, ++pos);
1051                 }
1052         }
1053         if (!list_empty(&res->lr_converting)) {
1054                 pos = 0;
1055                 CDEBUG(level, "Converting locks:\n");
1056                 list_for_each(tmp, &res->lr_converting) {
1057                         struct ldlm_lock *lock;
1058                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1059                         ldlm_lock_dump(level, lock, ++pos);
1060                 }
1061         }
1062         if (!list_empty(&res->lr_waiting)) {
1063                 pos = 0;
1064                 CDEBUG(level, "Waiting locks:\n");
1065                 list_for_each(tmp, &res->lr_waiting) {
1066                         struct ldlm_lock *lock;
1067                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1068                         ldlm_lock_dump(level, lock, ++pos);
1069                 }
1070         }
1071 }