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  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ldlm/ldlm_resource.c
  37  *
  38  * Author: Phil Schwan <phil@clusterfs.com>
  39  * Author: Peter Braam <braam@clusterfs.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LDLM
  43 #ifdef __KERNEL__
  44 # include <lustre_dlm.h>
  45 #else
  46 # include <liblustre.h>
  47 #endif
  48
  49 #include <obd_class.h>
  50 #include "ldlm_internal.h"
  51
  52 cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
  53
  54 atomic_t ldlm_srv_namespace_nr = ATOMIC_INIT(0);
  55 atomic_t ldlm_cli_namespace_nr = ATOMIC_INIT(0);
  56
  57 struct semaphore ldlm_srv_namespace_lock;
  58 struct list_head ldlm_srv_namespace_list =
  59         CFS_LIST_HEAD_INIT(ldlm_srv_namespace_list);
  60
  61 struct semaphore ldlm_cli_namespace_lock;
  62 struct list_head ldlm_cli_namespace_list =
  63         CFS_LIST_HEAD_INIT(ldlm_cli_namespace_list);
  64
  65 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
  66 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
  67 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
  68
  69 #ifdef LPROCFS
  70 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  71                              unsigned long count, void *data)
  72 {
  73         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  74         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  75         RETURN(count);
  76 }
  77
  78 int ldlm_proc_setup(void)
  79 {
  80         int rc;
  81         struct lprocfs_vars list[] = {
  82                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  83                 { NULL }};
  84         ENTRY;
  85         LASSERT(ldlm_ns_proc_dir == NULL);
  86
  87         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
  88                                               proc_lustre_root,
  89                                               NULL, NULL);
  90         if (IS_ERR(ldlm_type_proc_dir)) {
  91                 CERROR("LProcFS failed in ldlm-init\n");
  92                 rc = PTR_ERR(ldlm_type_proc_dir);
  93                 GOTO(err, rc);
  94         }
  95
  96         ldlm_ns_proc_dir = lprocfs_register("namespaces",
  97                                             ldlm_type_proc_dir,
  98                                             NULL, NULL);
  99         if (IS_ERR(ldlm_ns_proc_dir)) {
 100                 CERROR("LProcFS failed in ldlm-init\n");
 101                 rc = PTR_ERR(ldlm_ns_proc_dir);
 102                 GOTO(err_type, rc);
 103         }
 104
 105         ldlm_svc_proc_dir = lprocfs_register("services",
 106                                             ldlm_type_proc_dir,
 107                                             NULL, NULL);
 108         if (IS_ERR(ldlm_svc_proc_dir)) {
 109                 CERROR("LProcFS failed in ldlm-init\n");
 110                 rc = PTR_ERR(ldlm_svc_proc_dir);
 111                 GOTO(err_ns, rc);
 112         }
 113
 114         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 115
 116         RETURN(0);
 117
 118 err_ns:
 119         lprocfs_remove(&ldlm_ns_proc_dir);
 120 err_type:
 121         lprocfs_remove(&ldlm_type_proc_dir);
 122 err:
 123         ldlm_svc_proc_dir = NULL;
 124         RETURN(rc);
 125 }
 126
 127 void ldlm_proc_cleanup(void)
 128 {
 129         if (ldlm_svc_proc_dir)
 130                 lprocfs_remove(&ldlm_svc_proc_dir);
 131
 132         if (ldlm_ns_proc_dir)
 133                 lprocfs_remove(&ldlm_ns_proc_dir);
 134
 135         if (ldlm_type_proc_dir)
 136                 lprocfs_remove(&ldlm_type_proc_dir);
 137 }
 138
 139 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 140                                int count, int *eof, void *data)
 141 {
 142         struct ldlm_namespace *ns = data;
 143         __u32 *nr = &ns->ns_max_unused;
 144
 145         if (ns_connect_lru_resize(ns))
 146                 nr = &ns->ns_nr_unused;
 147         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 148 }
 149
 150 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 151                                unsigned long count, void *data)
 152 {
 153         struct ldlm_namespace *ns = data;
 154         char dummy[MAX_STRING_SIZE + 1] = { '\0' }, *end;
 155         unsigned long tmp;
 156         int lru_resize;
 157
 158         if (count >= sizeof(dummy) || count == 0)
 159                 return -EINVAL;
 160
 161         if (copy_from_user(dummy, buffer, count))
 162                 return -EFAULT;
 163
 164         if (strncmp(dummy, "clear", 5) == 0) {
 165                 CDEBUG(D_DLMTRACE,
 166                        "dropping all unused locks from namespace %s\n",
 167                        ns->ns_name);
 168                 if (ns_connect_lru_resize(ns)) {
 169                         int canceled, unused  = ns->ns_nr_unused;
 170
 171                         /* Try to cancel all @ns_nr_unused locks. */
 172                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
 173                                                    LDLM_CANCEL_PASSED);
 174                         if (canceled < unused) {
 175                                 CERROR("not all requested locks are canceled, "
 176                                        "requested: %d, canceled: %d\n", unused,
 177                                        canceled);
 178                                 return -EINVAL;
 179                         }
 180                 } else {
 181                         tmp = ns->ns_max_unused;
 182                         ns->ns_max_unused = 0;
 183                         ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
 184                         ns->ns_max_unused = tmp;
 185                 }
 186                 return count;
 187         }
 188
 189         tmp = simple_strtoul(dummy, &end, 0);
 190         if (dummy == end) {
 191                 CERROR("invalid value written\n");
 192                 return -EINVAL;
 193         }
 194         lru_resize = (tmp == 0);
 195
 196         if (ns_connect_lru_resize(ns)) {
 197                 if (!lru_resize)
 198                         ns->ns_max_unused = (unsigned int)tmp;
 199
 200                 if (tmp > ns->ns_nr_unused)
 201                         tmp = ns->ns_nr_unused;
 202                 tmp = ns->ns_nr_unused - tmp;
 203
 204                 CDEBUG(D_DLMTRACE,
 205                        "changing namespace %s unused locks from %u to %u\n",
 206                        ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
 207                 ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 208
 209                 if (!lru_resize) {
 210                         CDEBUG(D_DLMTRACE,
 211                                "disable lru_resize for namespace %s\n",
 212                                ns->ns_name);
 213                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 214                 }
 215         } else {
 216                 CDEBUG(D_DLMTRACE,
 217                        "changing namespace %s max_unused from %u to %u\n",
 218                        ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
 219                 ns->ns_max_unused = (unsigned int)tmp;
 220                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 221
 222                 /* Make sure that originally lru resize was supported before
 223                  * turning it on here. */
 224                 if (lru_resize &&
 225                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 226                         CDEBUG(D_DLMTRACE,
 227                                "enable lru_resize for namespace %s\n",
 228                                ns->ns_name);
 229                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 230                 }
 231         }
 232         return count;
 233 }
 234
 235 void ldlm_proc_namespace(struct ldlm_namespace *ns)
 236 {
 237         struct lprocfs_vars lock_vars[2];
 238         char lock_name[MAX_STRING_SIZE + 1];
 239
 240         LASSERT(ns != NULL);
 241         LASSERT(ns->ns_name != NULL);
 242
 243         lock_name[MAX_STRING_SIZE] = '\0';
 244
 245         memset(lock_vars, 0, sizeof(lock_vars));
 246         lock_vars[0].name = lock_name;
 247
 248         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
 249         lock_vars[0].data = &ns->ns_refcount;
 250         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 251         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 252
 253         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
 254         lock_vars[0].data = &ns->ns_locks;
 255         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 256         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 257
 258         if (ns_is_client(ns)) {
 259                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 260                          ns->ns_name);
 261                 lock_vars[0].data = &ns->ns_nr_unused;
 262                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 263                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 264
 265                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 266                          ns->ns_name);
 267                 lock_vars[0].data = ns;
 268                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 269                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 270                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 271
 272                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 273                          ns->ns_name);
 274                 lock_vars[0].data = &ns->ns_max_age;
 275                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 276                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 277                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 278         } else {
 279                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
 280                          ns->ns_name);
 281                 lock_vars[0].data = &ns->ns_timeouts;
 282                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 283                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 284
 285                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 286                          ns->ns_name);
 287                 lock_vars[0].data = &ns->ns_max_nolock_size;
 288                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 289                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 290                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 291
 292                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 293                          ns->ns_name);
 294                 lock_vars[0].data = &ns->ns_contention_time;
 295                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 296                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 297                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 298
 299                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 300                          ns->ns_name);
 301                 lock_vars[0].data = &ns->ns_contended_locks;
 302                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 303                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 304                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 305         }
 306 }
 307 #undef MAX_STRING_SIZE
 308 #else
 309 #define ldlm_proc_namespace(ns) do {} while (0)
 310 #endif /* LPROCFS */
 311
 312 struct ldlm_namespace *
 313 ldlm_namespace_new(struct obd_device *obd, char *name,
 314                    ldlm_side_t client, ldlm_appetite_t apt)
 315 {
 316         struct ldlm_namespace *ns = NULL;
 317         struct list_head *bucket;
 318         int rc, idx, namelen;
 319         ENTRY;
 320
 321         rc = ldlm_get_ref();
 322         if (rc) {
 323                 CERROR("ldlm_get_ref failed: %d\n", rc);
 324                 RETURN(NULL);
 325         }
 326
 327         OBD_ALLOC_PTR(ns);
 328         if (!ns)
 329                 GOTO(out_ref, NULL);
 330
 331         OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 332         if (!ns->ns_hash)
 333                 GOTO(out_ns, NULL);
 334
 335         namelen = strlen(name);
 336         OBD_ALLOC(ns->ns_name, namelen + 1);
 337         if (!ns->ns_name)
 338                 GOTO(out_hash, NULL);
 339
 340         ns->ns_appetite = apt;
 341
 342         LASSERT(obd != NULL);
 343         ns->ns_obd = obd;
 344
 345         strcpy(ns->ns_name, name);
 346
 347         CFS_INIT_LIST_HEAD(&ns->ns_root_list);
 348         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 349         ns->ns_refcount = 0;
 350         ns->ns_client = client;
 351         spin_lock_init(&ns->ns_hash_lock);
 352         atomic_set(&ns->ns_locks, 0);
 353         ns->ns_resources = 0;
 354         cfs_waitq_init(&ns->ns_waitq);
 355         ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
 356         ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
 357         ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
 358
 359         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 360              bucket--)
 361                 CFS_INIT_LIST_HEAD(bucket);
 362
 363         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 364         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 365         ns->ns_nr_unused = 0;
 366         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
 367         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
 368         ns->ns_timeouts = 0;
 369         spin_lock_init(&ns->ns_unused_lock);
 370         ns->ns_orig_connect_flags = 0;
 371         ns->ns_connect_flags = 0;
 372
 373         ldlm_proc_namespace(ns);
 374
 375         idx = atomic_read(ldlm_namespace_nr(client));
 376
 377         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 378         if (rc) {
 379                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 380                 GOTO(out_proc, rc);
 381         }
 382
 383         at_init(&ns->ns_at_estimate, ldlm_enqueue_min, 0);
 384         ldlm_namespace_register(ns, client);
 385         RETURN(ns);
 386 out_proc:
 387         ldlm_namespace_cleanup(ns, 0);
 388         OBD_FREE(ns->ns_name, namelen + 1);
 389 out_hash:
 390         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 391 out_ns:
 392         OBD_FREE_PTR(ns);
 393 out_ref:
 394         ldlm_put_ref();
 395         RETURN(NULL);
 396 }
 397
 398 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 399
 400 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 401  * This is currently only used for recovery, and we make certain assumptions
 402  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 403  *
 404  * Called with the ns_lock held. */
 405 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 406                              int flags)
 407 {
 408         struct list_head *tmp;
 409         int rc = 0, client = ns_is_client(res->lr_namespace);
 410         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 411         ENTRY;
 412
 413         do {
 414                 struct ldlm_lock *lock = NULL;
 415
 416                 /* first, we look for non-cleaned-yet lock
 417                  * all cleaned locks are marked by CLEANED flag */
 418                 lock_res(res);
 419                 list_for_each(tmp, q) {
 420                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 421                         if (lock->l_flags & LDLM_FL_CLEANED) {
 422                                 lock = NULL;
 423                                 continue;
 424                         }
 425                         LDLM_LOCK_GET(lock);
 426                         lock->l_flags |= LDLM_FL_CLEANED;
 427                         break;
 428                 }
 429
 430                 if (lock == NULL) {
 431                         unlock_res(res);
 432                         break;
 433                 }
 434
 435                 /* Set CBPENDING so nothing in the cancellation path
 436                  * can match this lock */
 437                 lock->l_flags |= LDLM_FL_CBPENDING;
 438                 lock->l_flags |= LDLM_FL_FAILED;
 439                 lock->l_flags |= flags;
 440
 441                 /* ... without sending a CANCEL message for local_only. */
 442                 if (local_only)
 443                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 444
 445                 if (local_only && (lock->l_readers || lock->l_writers)) {
 446                         /* This is a little bit gross, but much better than the
 447                          * alternative: pretend that we got a blocking AST from
 448                          * the server, so that when the lock is decref'd, it
 449                          * will go away ... */
 450                         unlock_res(res);
 451                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 452                         if (lock->l_completion_ast)
 453                                 lock->l_completion_ast(lock, 0, NULL);
 454                         LDLM_LOCK_PUT(lock);
 455                         continue;
 456                 }
 457
 458                 if (client) {
 459                         struct lustre_handle lockh;
 460
 461                         unlock_res(res);
 462                         ldlm_lock2handle(lock, &lockh);
 463                         rc = ldlm_cli_cancel(&lockh);
 464                         if (rc)
 465                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 466                 } else {
 467                         ldlm_resource_unlink_lock(lock);
 468                         unlock_res(res);
 469                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 470                                    "client node");
 471                         ldlm_lock_destroy(lock);
 472                 }
 473                 LDLM_LOCK_PUT(lock);
 474         } while (1);
 475
 476         EXIT;
 477 }
 478
 479 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 480 {
 481         struct list_head *tmp;
 482         int i;
 483
 484         if (ns == NULL) {
 485                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 486                 return ELDLM_OK;
 487         }
 488
 489         for (i = 0; i < RES_HASH_SIZE; i++) {
 490                 spin_lock(&ns->ns_hash_lock);
 491                 tmp = ns->ns_hash[i].next;
 492                 while (tmp != &(ns->ns_hash[i])) {
 493                         struct ldlm_resource *res;
 494                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 495                         ldlm_resource_getref(res);
 496                         spin_unlock(&ns->ns_hash_lock);
 497
 498                         cleanup_resource(res, &res->lr_granted, flags);
 499                         cleanup_resource(res, &res->lr_converting, flags);
 500                         cleanup_resource(res, &res->lr_waiting, flags);
 501
 502                         spin_lock(&ns->ns_hash_lock);
 503                         tmp  = tmp->next;
 504
 505                         /* XXX: former stuff caused issues in case of race
 506                          * between ldlm_namespace_cleanup() and lockd() when
 507                          * client gets blocking ast when lock gets distracted by
 508                          * server. This is 1_4 branch solution, let's see how
 509                          * will it behave. */
 510                         if (!ldlm_resource_putref_locked(res))
 511                                 CDEBUG(D_INFO,
 512                                        "Namespace %s resource refcount nonzero "
 513                                        "(%d) after lock cleanup; forcing cleanup.\n",
 514                                        ns->ns_name, atomic_read(&res->lr_refcount));
 515                 }
 516                 spin_unlock(&ns->ns_hash_lock);
 517         }
 518
 519         return ELDLM_OK;
 520 }
 521
 522 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 523 {
 524         ENTRY;
 525         /* At shutdown time, don't call the cancellation callback */
 526         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 527
 528         if (ns->ns_refcount > 0) {
 529                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 530                 int rc;
 531                 CDEBUG(D_DLMTRACE,
 532                        "dlm namespace %s free waiting on refcount %d\n",
 533                        ns->ns_name, ns->ns_refcount);
 534 force_wait:
 535                 if (force)
 536                         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 537
 538                 rc = l_wait_event(ns->ns_waitq,
 539                                   ns->ns_refcount == 0, &lwi);
 540
 541                 /* Forced cleanups should be able to reclaim all references,
 542                  * so it's safe to wait forever... we can't leak locks... */
 543                 if (force && rc == -ETIMEDOUT) {
 544                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 545                                        "namespace with %d resources in use, "
 546                                        "(rc=%d)\n", ns->ns_name,
 547                                        ns->ns_refcount, rc);
 548                         GOTO(force_wait, rc);
 549                 }
 550
 551                 if (ns->ns_refcount) {
 552                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 553                                        "with %d resources in use, (rc=%d)\n",
 554                                        ns->ns_name,
 555                                        ns->ns_refcount, rc);
 556                         RETURN(ELDLM_NAMESPACE_EXISTS);
 557                 }
 558                 CDEBUG(D_DLMTRACE,
 559                        "dlm namespace %s free done waiting\n", ns->ns_name);
 560         }
 561
 562         RETURN(ELDLM_OK);
 563 }
 564
 565 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 566                                struct obd_import *imp,
 567                                int force)
 568 {
 569         int rc;
 570         ENTRY;
 571         if (!ns) {
 572                 EXIT;
 573                 return;
 574         }
 575
 576         /* Make sure that nobody can find this ns in its list. */
 577         ldlm_namespace_unregister(ns, ns->ns_client);
 578
 579         /* Can fail with -EINTR when force == 0 in which case try harder */
 580         rc = __ldlm_namespace_free(ns, force);
 581         if (rc != ELDLM_OK) {
 582                 if (imp) {
 583                         ptlrpc_disconnect_import(imp, 0);
 584                         ptlrpc_invalidate_import(imp);
 585                 }
 586
 587                 /* With all requests dropped and the import inactive
 588                  * we are gaurenteed all reference will be dropped. */
 589                 rc = __ldlm_namespace_free(ns, 1);
 590                 LASSERT(rc == 0);
 591         }
 592         EXIT;
 593 }
 594
 595 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 596 {
 597         ENTRY;
 598         if (!ns) {
 599                 EXIT;
 600                 return;
 601         }
 602
 603         /* Fini pool _before_ parent proc dir is removed. This is important
 604          * as ldlm_pool_fini() removes own proc dir which is child to @dir.
 605          * Removing it after @dir may cause oops. */
 606         ldlm_pool_fini(&ns->ns_pool);
 607
 608 #ifdef LPROCFS
 609         {
 610                 struct proc_dir_entry *dir;
 611                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
 612                 if (dir == NULL) {
 613                         CERROR("dlm namespace %s has no procfs dir?\n",
 614                                ns->ns_name);
 615                 } else {
 616                         lprocfs_remove(&dir);
 617                 }
 618         }
 619 #endif
 620         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 621         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 622
 623         /* @ns should be not on list in this time, otherwise this will cause
 624          * issues realted to using freed @ns in pools thread. */
 625         LASSERT(list_empty(&ns->ns_list_chain));
 626         OBD_FREE_PTR(ns);
 627         ldlm_put_ref();
 628         EXIT;
 629 }
 630
 631 /* Cleanup the resource, and free namespace.
 632  * bug 12864:
 633  * Deadlock issue:
 634  * proc1: destroy import
 635  *        class_disconnect_export(grab cl_sem) ->
 636  *              -> ldlm_namespace_free ->
 637  *              -> lprocfs_remove(grab _lprocfs_lock).
 638  * proc2: read proc info
 639  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 640  *              -> osc_rd_active, etc(grab cl_sem).
 641  *
 642  * So that I have to split the ldlm_namespace_free into two parts - the first
 643  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 644  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 645  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 646  * held.
 647  */
 648 void ldlm_namespace_free(struct ldlm_namespace *ns,
 649                          struct obd_import *imp,
 650                          int force)
 651 {
 652         ldlm_namespace_free_prior(ns, imp, force);
 653         ldlm_namespace_free_post(ns);
 654 }
 655
 656 void ldlm_namespace_get_locked(struct ldlm_namespace *ns)
 657 {
 658         ns->ns_refcount++;
 659 }
 660
 661 void ldlm_namespace_get(struct ldlm_namespace *ns)
 662 {
 663         spin_lock(&ns->ns_hash_lock);
 664         ldlm_namespace_get_locked(ns);
 665         spin_unlock(&ns->ns_hash_lock);
 666 }
 667
 668 void ldlm_namespace_put_locked(struct ldlm_namespace *ns, int wakeup)
 669 {
 670         LASSERT(ns->ns_refcount > 0);
 671         ns->ns_refcount--;
 672         if (ns->ns_refcount == 0 && wakeup)
 673                 wake_up(&ns->ns_waitq);
 674 }
 675
 676 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
 677 {
 678         spin_lock(&ns->ns_hash_lock);
 679         ldlm_namespace_put_locked(ns, wakeup);
 680         spin_unlock(&ns->ns_hash_lock);
 681 }
 682
 683 /* Register @ns in the list of namespaces */
 684 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
 685 {
 686         mutex_down(ldlm_namespace_lock(client));
 687         LASSERT(list_empty(&ns->ns_list_chain));
 688         list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 689         atomic_inc(ldlm_namespace_nr(client));
 690         mutex_up(ldlm_namespace_lock(client));
 691 }
 692
 693 /* Unregister @ns from the list of namespaces */
 694 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
 695 {
 696         mutex_down(ldlm_namespace_lock(client));
 697         LASSERT(!list_empty(&ns->ns_list_chain));
 698         /*
 699          * Some asserts and possibly other parts of code still using
 700          * list_empty(&ns->ns_list_chain). This is why it is important
 701          * to use list_del_init() here.
 702          */
 703         list_del_init(&ns->ns_list_chain);
 704         atomic_dec(ldlm_namespace_nr(client));
 705         mutex_up(ldlm_namespace_lock(client));
 706 }
 707
 708 /* Should be called under ldlm_namespace_lock(client) taken */
 709 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
 710 {
 711         LASSERT(!list_empty(&ns->ns_list_chain));
 712         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 713         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 714 }
 715
 716 /* Should be called under ldlm_namespace_lock(client) taken */
 717 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
 718 {
 719         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 720         LASSERT(!list_empty(ldlm_namespace_list(client)));
 721         return container_of(ldlm_namespace_list(client)->next,
 722                 struct ldlm_namespace, ns_list_chain);
 723 }
 724
 725 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, struct ldlm_res_id name)
 726 {
 727         __u32 hash = 0;
 728         int i;
 729
 730         for (i = 0; i < RES_NAME_SIZE; i++)
 731                 hash += name.name[i];
 732
 733         hash += (__u32)((unsigned long)parent >> 4);
 734
 735         return (hash & RES_HASH_MASK);
 736 }
 737
 738 static struct ldlm_resource *ldlm_resource_new(void)
 739 {
 740         struct ldlm_resource *res;
 741         int idx;
 742
 743         OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
 744         if (res == NULL)
 745                 return NULL;
 746
 747         memset(res, 0, sizeof(*res));
 748
 749         CFS_INIT_LIST_HEAD(&res->lr_children);
 750         CFS_INIT_LIST_HEAD(&res->lr_childof);
 751         CFS_INIT_LIST_HEAD(&res->lr_granted);
 752         CFS_INIT_LIST_HEAD(&res->lr_converting);
 753         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 754
 755         /* initialize interval trees for each lock mode*/
 756         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 757                 res->lr_itree[idx].lit_size = 0;
 758                 res->lr_itree[idx].lit_mode = 1 << idx;
 759                 res->lr_itree[idx].lit_root = NULL;
 760         }
 761
 762         atomic_set(&res->lr_refcount, 1);
 763         spin_lock_init(&res->lr_lock);
 764
 765         /* one who creates the resource must unlock
 766          * the semaphore after lvb initialization */
 767         init_MUTEX_LOCKED(&res->lr_lvb_sem);
 768
 769         return res;
 770 }
 771
 772 /* must be called with hash lock held */
 773 static struct ldlm_resource *
 774 ldlm_resource_find(struct ldlm_namespace *ns, struct ldlm_res_id name, __u32 hash)
 775 {
 776         struct list_head *bucket, *tmp;
 777         struct ldlm_resource *res;
 778
 779         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 780         bucket = ns->ns_hash + hash;
 781
 782         list_for_each(tmp, bucket) {
 783                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 784                 if (memcmp(&res->lr_name, &name, sizeof(res->lr_name)) == 0)
 785                         return res;
 786         }
 787
 788         return NULL;
 789 }
 790
 791 /* Args: locked namespace
 792  * Returns: newly-allocated, referenced, unlocked resource */
 793 static struct ldlm_resource *
 794 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 795                   struct ldlm_res_id name, __u32 hash, ldlm_type_t type)
 796 {
 797         struct list_head *bucket;
 798         struct ldlm_resource *res, *old_res;
 799         ENTRY;
 800
 801         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
 802                  "type: %d\n", type);
 803
 804         res = ldlm_resource_new();
 805         if (!res)
 806                 RETURN(NULL);
 807
 808         res->lr_name = name;
 809         res->lr_namespace = ns;
 810         res->lr_type = type;
 811         res->lr_most_restr = LCK_NL;
 812
 813         spin_lock(&ns->ns_hash_lock);
 814         old_res = ldlm_resource_find(ns, name, hash);
 815         if (old_res) {
 816                 /* someone won the race and added the resource before */
 817                 ldlm_resource_getref(old_res);
 818                 spin_unlock(&ns->ns_hash_lock);
 819                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 820                 /* synchronize WRT resource creation */
 821                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 822                         down(&old_res->lr_lvb_sem);
 823                         up(&old_res->lr_lvb_sem);
 824                 }
 825                 RETURN(old_res);
 826         }
 827
 828         /* we won! let's add the resource */
 829         bucket = ns->ns_hash + hash;
 830         list_add(&res->lr_hash, bucket);
 831         ns->ns_resources++;
 832         ldlm_namespace_get_locked(ns);
 833
 834         if (parent == NULL) {
 835                 list_add(&res->lr_childof, &ns->ns_root_list);
 836         } else {
 837                 res->lr_parent = parent;
 838                 list_add(&res->lr_childof, &parent->lr_children);
 839         }
 840         spin_unlock(&ns->ns_hash_lock);
 841
 842         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 843                 int rc;
 844
 845                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
 846                 rc = ns->ns_lvbo->lvbo_init(res);
 847                 if (rc)
 848                         CERROR("lvbo_init failed for resource "
 849                                LPU64": rc %d\n", name.name[0], rc);
 850                 /* we create resource with locked lr_lvb_sem */
 851                 up(&res->lr_lvb_sem);
 852         }
 853
 854         RETURN(res);
 855 }
 856
 857 /* Args: unlocked namespace
 858  * Locks: takes and releases ns->ns_lock and res->lr_lock
 859  * Returns: referenced, unlocked ldlm_resource or NULL */
 860 struct ldlm_resource *
 861 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 862                   struct ldlm_res_id name, ldlm_type_t type, int create)
 863 {
 864         __u32 hash = ldlm_hash_fn(parent, name);
 865         struct ldlm_resource *res = NULL;
 866         ENTRY;
 867
 868         LASSERT(ns != NULL);
 869         LASSERT(ns->ns_hash != NULL);
 870         LASSERT(name.name[0] != 0);
 871
 872         spin_lock(&ns->ns_hash_lock);
 873         res = ldlm_resource_find(ns, name, hash);
 874         if (res) {
 875                 ldlm_resource_getref(res);
 876                 spin_unlock(&ns->ns_hash_lock);
 877                 /* synchronize WRT resource creation */
 878                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 879                         down(&res->lr_lvb_sem);
 880                         up(&res->lr_lvb_sem);
 881                 }
 882                 RETURN(res);
 883         }
 884         spin_unlock(&ns->ns_hash_lock);
 885
 886         if (create == 0)
 887                 RETURN(NULL);
 888
 889         res = ldlm_resource_add(ns, parent, name, hash, type);
 890         RETURN(res);
 891 }
 892
 893 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 894 {
 895         LASSERT(res != NULL);
 896         LASSERT(res != LP_POISON);
 897         atomic_inc(&res->lr_refcount);
 898         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 899                atomic_read(&res->lr_refcount));
 900         return res;
 901 }
 902
 903 void __ldlm_resource_putref_final(struct ldlm_resource *res)
 904 {
 905         struct ldlm_namespace *ns = res->lr_namespace;
 906
 907         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 908
 909         if (!list_empty(&res->lr_granted)) {
 910                 ldlm_resource_dump(D_ERROR, res);
 911                 LBUG();
 912         }
 913
 914         if (!list_empty(&res->lr_converting)) {
 915                 ldlm_resource_dump(D_ERROR, res);
 916                 LBUG();
 917         }
 918
 919         if (!list_empty(&res->lr_waiting)) {
 920                 ldlm_resource_dump(D_ERROR, res);
 921                 LBUG();
 922         }
 923
 924         if (!list_empty(&res->lr_children)) {
 925                 ldlm_resource_dump(D_ERROR, res);
 926                 LBUG();
 927         }
 928
 929         /* Pass 0 as second argument to not wake up ->ns_waitq yet, will do it
 930          * later. */
 931         ldlm_namespace_put_locked(ns, 0);
 932         list_del_init(&res->lr_hash);
 933         list_del_init(&res->lr_childof);
 934
 935         ns->ns_resources--;
 936         if (ns->ns_resources == 0)
 937                 wake_up(&ns->ns_waitq);
 938 }
 939
 940 /* Returns 1 if the resource was freed, 0 if it remains. */
 941 int ldlm_resource_putref(struct ldlm_resource *res)
 942 {
 943         struct ldlm_namespace *ns = res->lr_namespace;
 944         int rc = 0;
 945         ENTRY;
 946
 947         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 948                atomic_read(&res->lr_refcount) - 1);
 949         LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
 950                  atomic_read(&res->lr_refcount));
 951         LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
 952                  atomic_read(&res->lr_refcount));
 953
 954         if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
 955                 __ldlm_resource_putref_final(res);
 956                 spin_unlock(&ns->ns_hash_lock);
 957                 if (res->lr_lvb_data)
 958                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 959                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 960                 rc = 1;
 961         }
 962
 963         RETURN(rc);
 964 }
 965
 966 /* Returns 1 if the resource was freed, 0 if it remains. */
 967 int ldlm_resource_putref_locked(struct ldlm_resource *res)
 968 {
 969         int rc = 0;
 970         ENTRY;
 971
 972         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 973                atomic_read(&res->lr_refcount) - 1);
 974         LASSERT(atomic_read(&res->lr_refcount) > 0);
 975         LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
 976
 977         LASSERT(atomic_read(&res->lr_refcount) >= 0);
 978         if (atomic_dec_and_test(&res->lr_refcount)) {
 979                 __ldlm_resource_putref_final(res);
 980                 if (res->lr_lvb_data)
 981                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 982                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 983                 rc = 1;
 984         }
 985
 986         RETURN(rc);
 987 }
 988
 989 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 990                             struct ldlm_lock *lock)
 991 {
 992         check_res_locked(res);
 993
 994         ldlm_resource_dump(D_INFO, res);
 995         CDEBUG(D_OTHER, "About to add this lock:\n");
 996         ldlm_lock_dump(D_OTHER, lock, 0);
 997
 998         if (lock->l_destroyed) {
 999                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1000                 return;
1001         }
1002
1003         LASSERT(list_empty(&lock->l_res_link));
1004
1005         list_add_tail(&lock->l_res_link, head);
1006 }
1007
1008 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1009                                      struct ldlm_lock *new)
1010 {
1011         struct ldlm_resource *res = original->l_resource;
1012
1013         check_res_locked(res);
1014
1015         ldlm_resource_dump(D_INFO, res);
1016         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1017         ldlm_lock_dump(D_OTHER, new, 0);
1018
1019         if (new->l_destroyed) {
1020                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1021                 goto out;
1022         }
1023
1024         LASSERT(list_empty(&new->l_res_link));
1025
1026         list_add(&new->l_res_link, &original->l_res_link);
1027  out:;
1028 }
1029
1030 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1031 {
1032         int type = lock->l_resource->lr_type;
1033
1034         check_res_locked(lock->l_resource);
1035         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1036                 ldlm_unlink_lock_skiplist(lock);
1037         else if (type == LDLM_EXTENT)
1038                 ldlm_extent_unlink_lock(lock);
1039         list_del_init(&lock->l_res_link);
1040 }
1041
1042 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1043 {
1044         desc->lr_type = res->lr_type;
1045         desc->lr_name = res->lr_name;
1046 }
1047
1048 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1049 {
1050         struct list_head *tmp;
1051
1052         if (!((libcfs_debug | D_ERROR) & level))
1053                 return;
1054
1055         mutex_down(ldlm_namespace_lock(client));
1056
1057         list_for_each(tmp, ldlm_namespace_list(client)) {
1058                 struct ldlm_namespace *ns;
1059                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1060                 ldlm_namespace_dump(level, ns);
1061         }
1062
1063         mutex_up(ldlm_namespace_lock(client));
1064 }
1065
1066 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1067 {
1068         struct list_head *tmp;
1069
1070         if (!((libcfs_debug | D_ERROR) & level))
1071                 return;
1072
1073         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1074                ns->ns_name, ns->ns_refcount,
1075                ns_is_client(ns) ? "client" : "server");
1076
1077         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1078                 return;
1079
1080         spin_lock(&ns->ns_hash_lock);
1081         tmp = ns->ns_root_list.next;
1082         while (tmp != &ns->ns_root_list) {
1083                 struct ldlm_resource *res;
1084                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
1085
1086                 ldlm_resource_getref(res);
1087                 spin_unlock(&ns->ns_hash_lock);
1088
1089                 lock_res(res);
1090                 ldlm_resource_dump(level, res);
1091                 unlock_res(res);
1092
1093                 spin_lock(&ns->ns_hash_lock);
1094                 tmp = tmp->next;
1095                 ldlm_resource_putref_locked(res);
1096         }
1097         ns->ns_next_dump = cfs_time_shift(10);
1098         spin_unlock(&ns->ns_hash_lock);
1099 }
1100
1101 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1102 {
1103         struct list_head *tmp;
1104         int pos;
1105
1106         CLASSERT(RES_NAME_SIZE == 4);
1107
1108         if (!((libcfs_debug | D_ERROR) & level))
1109                 return;
1110
1111         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1112                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1113                res->lr_name.name[2], res->lr_name.name[3],
1114                atomic_read(&res->lr_refcount));
1115
1116         if (!list_empty(&res->lr_granted)) {
1117                 pos = 0;
1118                 CDEBUG(level, "Granted locks:\n");
1119                 list_for_each(tmp, &res->lr_granted) {
1120                         struct ldlm_lock *lock;
1121                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1122                         ldlm_lock_dump(level, lock, ++pos);
1123                 }
1124         }
1125         if (!list_empty(&res->lr_converting)) {
1126                 pos = 0;
1127                 CDEBUG(level, "Converting locks:\n");
1128                 list_for_each(tmp, &res->lr_converting) {
1129                         struct ldlm_lock *lock;
1130                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1131                         ldlm_lock_dump(level, lock, ++pos);
1132                 }
1133         }
1134         if (!list_empty(&res->lr_waiting)) {
1135                 pos = 0;
1136                 CDEBUG(level, "Waiting locks:\n");
1137                 list_for_each(tmp, &res->lr_waiting) {
1138                         struct ldlm_lock *lock;
1139                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1140                         ldlm_lock_dump(level, lock, ++pos);
1141                 }
1142         }
1143 }