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 extern unsigned int ldlm_cancel_unused_locks_before_replay;
  70
  71 #ifdef LPROCFS
  72 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  73                              unsigned long count, void *data)
  74 {
  75         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  76         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  77         RETURN(count);
  78 }
  79
  80 int ldlm_proc_setup(void)
  81 {
  82         int rc;
  83         struct lprocfs_vars list[] = {
  84                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  85                 { "cancel_unused_locks_before_replay",
  86                   lprocfs_rd_uint, lprocfs_wr_uint,
  87                   &ldlm_cancel_unused_locks_before_replay, NULL },
  88                 { NULL }};
  89         ENTRY;
  90         LASSERT(ldlm_ns_proc_dir == NULL);
  91
  92         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
  93                                               proc_lustre_root,
  94                                               NULL, NULL);
  95         if (IS_ERR(ldlm_type_proc_dir)) {
  96                 CERROR("LProcFS failed in ldlm-init\n");
  97                 rc = PTR_ERR(ldlm_type_proc_dir);
  98                 GOTO(err, rc);
  99         }
 100
 101         ldlm_ns_proc_dir = lprocfs_register("namespaces",
 102                                             ldlm_type_proc_dir,
 103                                             NULL, NULL);
 104         if (IS_ERR(ldlm_ns_proc_dir)) {
 105                 CERROR("LProcFS failed in ldlm-init\n");
 106                 rc = PTR_ERR(ldlm_ns_proc_dir);
 107                 GOTO(err_type, rc);
 108         }
 109
 110         ldlm_svc_proc_dir = lprocfs_register("services",
 111                                             ldlm_type_proc_dir,
 112                                             NULL, NULL);
 113         if (IS_ERR(ldlm_svc_proc_dir)) {
 114                 CERROR("LProcFS failed in ldlm-init\n");
 115                 rc = PTR_ERR(ldlm_svc_proc_dir);
 116                 GOTO(err_ns, rc);
 117         }
 118
 119         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 120
 121         RETURN(0);
 122
 123 err_ns:
 124         lprocfs_remove(&ldlm_ns_proc_dir);
 125 err_type:
 126         lprocfs_remove(&ldlm_type_proc_dir);
 127 err:
 128         ldlm_svc_proc_dir = NULL;
 129         RETURN(rc);
 130 }
 131
 132 void ldlm_proc_cleanup(void)
 133 {
 134         if (ldlm_svc_proc_dir)
 135                 lprocfs_remove(&ldlm_svc_proc_dir);
 136
 137         if (ldlm_ns_proc_dir)
 138                 lprocfs_remove(&ldlm_ns_proc_dir);
 139
 140         if (ldlm_type_proc_dir)
 141                 lprocfs_remove(&ldlm_type_proc_dir);
 142 }
 143
 144 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 145                                int count, int *eof, void *data)
 146 {
 147         struct ldlm_namespace *ns = data;
 148         __u32 *nr = &ns->ns_max_unused;
 149
 150         if (ns_connect_lru_resize(ns))
 151                 nr = &ns->ns_nr_unused;
 152         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 153 }
 154
 155 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 156                                unsigned long count, void *data)
 157 {
 158         struct ldlm_namespace *ns = data;
 159         char dummy[MAX_STRING_SIZE + 1] = { '\0' }, *end;
 160         unsigned long tmp;
 161         int lru_resize;
 162
 163         if (count >= sizeof(dummy) || count == 0)
 164                 return -EINVAL;
 165
 166         if (copy_from_user(dummy, buffer, count))
 167                 return -EFAULT;
 168
 169         if (strncmp(dummy, "clear", 5) == 0) {
 170                 CDEBUG(D_DLMTRACE,
 171                        "dropping all unused locks from namespace %s\n",
 172                        ns->ns_name);
 173                 if (ns_connect_lru_resize(ns)) {
 174                         int canceled, unused  = ns->ns_nr_unused;
 175
 176                         /* Try to cancel all @ns_nr_unused locks. */
 177                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
 178                                                    LDLM_CANCEL_PASSED);
 179                         if (canceled < unused) {
 180                                 CERROR("not all requested locks are canceled, "
 181                                        "requested: %d, canceled: %d\n", unused,
 182                                        canceled);
 183                                 return -EINVAL;
 184                         }
 185                 } else {
 186                         tmp = ns->ns_max_unused;
 187                         ns->ns_max_unused = 0;
 188                         ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
 189                         ns->ns_max_unused = tmp;
 190                 }
 191                 return count;
 192         }
 193
 194         tmp = simple_strtoul(dummy, &end, 0);
 195         if (dummy == end) {
 196                 CERROR("invalid value written\n");
 197                 return -EINVAL;
 198         }
 199         lru_resize = (tmp == 0);
 200
 201         if (ns_connect_lru_resize(ns)) {
 202                 if (!lru_resize)
 203                         ns->ns_max_unused = (unsigned int)tmp;
 204
 205                 if (tmp > ns->ns_nr_unused)
 206                         tmp = ns->ns_nr_unused;
 207                 tmp = ns->ns_nr_unused - tmp;
 208
 209                 CDEBUG(D_DLMTRACE,
 210                        "changing namespace %s unused locks from %u to %u\n",
 211                        ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
 212                 ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 213
 214                 if (!lru_resize) {
 215                         CDEBUG(D_DLMTRACE,
 216                                "disable lru_resize for namespace %s\n",
 217                                ns->ns_name);
 218                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 219                 }
 220         } else {
 221                 CDEBUG(D_DLMTRACE,
 222                        "changing namespace %s max_unused from %u to %u\n",
 223                        ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
 224                 ns->ns_max_unused = (unsigned int)tmp;
 225                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 226
 227                 /* Make sure that originally lru resize was supported before
 228                  * turning it on here. */
 229                 if (lru_resize &&
 230                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 231                         CDEBUG(D_DLMTRACE,
 232                                "enable lru_resize for namespace %s\n",
 233                                ns->ns_name);
 234                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 235                 }
 236         }
 237         return count;
 238 }
 239
 240 void ldlm_proc_namespace(struct ldlm_namespace *ns)
 241 {
 242         struct lprocfs_vars lock_vars[2];
 243         char lock_name[MAX_STRING_SIZE + 1];
 244
 245         LASSERT(ns != NULL);
 246         LASSERT(ns->ns_name != NULL);
 247
 248         lock_name[MAX_STRING_SIZE] = '\0';
 249
 250         memset(lock_vars, 0, sizeof(lock_vars));
 251         lock_vars[0].name = lock_name;
 252
 253         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
 254         lock_vars[0].data = &ns->ns_refcount;
 255         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 256         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 257
 258         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
 259         lock_vars[0].data = &ns->ns_locks;
 260         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 261         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 262
 263         if (ns_is_client(ns)) {
 264                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 265                          ns->ns_name);
 266                 lock_vars[0].data = &ns->ns_nr_unused;
 267                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 268                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 269
 270                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 271                          ns->ns_name);
 272                 lock_vars[0].data = ns;
 273                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 274                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 275                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 276
 277                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 278                          ns->ns_name);
 279                 lock_vars[0].data = &ns->ns_max_age;
 280                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 281                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 282                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 283         } else {
 284                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
 285                          ns->ns_name);
 286                 lock_vars[0].data = &ns->ns_timeouts;
 287                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 288                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 289
 290                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 291                          ns->ns_name);
 292                 lock_vars[0].data = &ns->ns_max_nolock_size;
 293                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 294                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 295                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 296
 297                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 298                          ns->ns_name);
 299                 lock_vars[0].data = &ns->ns_contention_time;
 300                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 301                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 302                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 303
 304                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 305                          ns->ns_name);
 306                 lock_vars[0].data = &ns->ns_contended_locks;
 307                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 308                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 309                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 310         }
 311 }
 312 #undef MAX_STRING_SIZE
 313 #else
 314 #define ldlm_proc_namespace(ns) do {} while (0)
 315 #endif /* LPROCFS */
 316
 317 struct ldlm_namespace *
 318 ldlm_namespace_new(struct obd_device *obd, char *name,
 319                    ldlm_side_t client, ldlm_appetite_t apt)
 320 {
 321         struct ldlm_namespace *ns = NULL;
 322         struct list_head *bucket;
 323         int rc, idx, namelen;
 324         ENTRY;
 325
 326         rc = ldlm_get_ref();
 327         if (rc) {
 328                 CERROR("ldlm_get_ref failed: %d\n", rc);
 329                 RETURN(NULL);
 330         }
 331
 332         OBD_ALLOC_PTR(ns);
 333         if (!ns)
 334                 GOTO(out_ref, NULL);
 335
 336         OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 337         if (!ns->ns_hash)
 338                 GOTO(out_ns, NULL);
 339
 340         namelen = strlen(name);
 341         OBD_ALLOC(ns->ns_name, namelen + 1);
 342         if (!ns->ns_name)
 343                 GOTO(out_hash, NULL);
 344
 345         ns->ns_appetite = apt;
 346
 347         LASSERT(obd != NULL);
 348         ns->ns_obd = obd;
 349
 350         strcpy(ns->ns_name, name);
 351
 352         CFS_INIT_LIST_HEAD(&ns->ns_root_list);
 353         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 354         ns->ns_refcount = 0;
 355         ns->ns_client = client;
 356         spin_lock_init(&ns->ns_hash_lock);
 357         atomic_set(&ns->ns_locks, 0);
 358         ns->ns_resources = 0;
 359         cfs_waitq_init(&ns->ns_waitq);
 360         ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
 361         ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
 362         ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
 363
 364         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 365              bucket--)
 366                 CFS_INIT_LIST_HEAD(bucket);
 367
 368         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 369         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 370         ns->ns_nr_unused = 0;
 371         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
 372         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
 373         ns->ns_timeouts = 0;
 374         spin_lock_init(&ns->ns_unused_lock);
 375         ns->ns_orig_connect_flags = 0;
 376         ns->ns_connect_flags = 0;
 377
 378         ldlm_proc_namespace(ns);
 379
 380         idx = atomic_read(ldlm_namespace_nr(client));
 381
 382         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 383         if (rc) {
 384                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 385                 GOTO(out_proc, rc);
 386         }
 387
 388         at_init(&ns->ns_at_estimate, ldlm_enqueue_min, 0);
 389         ldlm_namespace_register(ns, client);
 390         RETURN(ns);
 391 out_proc:
 392         ldlm_namespace_cleanup(ns, 0);
 393         OBD_FREE(ns->ns_name, namelen + 1);
 394 out_hash:
 395         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 396 out_ns:
 397         OBD_FREE_PTR(ns);
 398 out_ref:
 399         ldlm_put_ref();
 400         RETURN(NULL);
 401 }
 402
 403 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 404
 405 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 406  * This is currently only used for recovery, and we make certain assumptions
 407  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 408  *
 409  * Called with the ns_lock held. */
 410 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 411                              int flags)
 412 {
 413         struct list_head *tmp;
 414         int rc = 0, client = ns_is_client(res->lr_namespace);
 415         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 416         ENTRY;
 417
 418         do {
 419                 struct ldlm_lock *lock = NULL;
 420
 421                 /* first, we look for non-cleaned-yet lock
 422                  * all cleaned locks are marked by CLEANED flag */
 423                 lock_res(res);
 424                 list_for_each(tmp, q) {
 425                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 426                         if (lock->l_flags & LDLM_FL_CLEANED) {
 427                                 lock = NULL;
 428                                 continue;
 429                         }
 430                         LDLM_LOCK_GET(lock);
 431                         lock->l_flags |= LDLM_FL_CLEANED;
 432                         break;
 433                 }
 434
 435                 if (lock == NULL) {
 436                         unlock_res(res);
 437                         break;
 438                 }
 439
 440                 /* Set CBPENDING so nothing in the cancellation path
 441                  * can match this lock */
 442                 lock->l_flags |= LDLM_FL_CBPENDING;
 443                 lock->l_flags |= LDLM_FL_FAILED;
 444                 lock->l_flags |= flags;
 445
 446                 /* ... without sending a CANCEL message for local_only. */
 447                 if (local_only)
 448                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 449
 450                 if (local_only && (lock->l_readers || lock->l_writers)) {
 451                         /* This is a little bit gross, but much better than the
 452                          * alternative: pretend that we got a blocking AST from
 453                          * the server, so that when the lock is decref'd, it
 454                          * will go away ... */
 455                         unlock_res(res);
 456                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 457                         if (lock->l_completion_ast)
 458                                 lock->l_completion_ast(lock, 0, NULL);
 459                         LDLM_LOCK_PUT(lock);
 460                         continue;
 461                 }
 462
 463                 if (client) {
 464                         struct lustre_handle lockh;
 465
 466                         unlock_res(res);
 467                         ldlm_lock2handle(lock, &lockh);
 468                         rc = ldlm_cli_cancel(&lockh);
 469                         if (rc)
 470                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 471                 } else {
 472                         ldlm_resource_unlink_lock(lock);
 473                         unlock_res(res);
 474                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 475                                    "client node");
 476                         ldlm_lock_destroy(lock);
 477                 }
 478                 LDLM_LOCK_PUT(lock);
 479         } while (1);
 480
 481         EXIT;
 482 }
 483
 484 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 485 {
 486         struct list_head *tmp;
 487         int i;
 488
 489         if (ns == NULL) {
 490                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 491                 return ELDLM_OK;
 492         }
 493
 494         for (i = 0; i < RES_HASH_SIZE; i++) {
 495                 spin_lock(&ns->ns_hash_lock);
 496                 tmp = ns->ns_hash[i].next;
 497                 while (tmp != &(ns->ns_hash[i])) {
 498                         struct ldlm_resource *res;
 499                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 500                         ldlm_resource_getref(res);
 501                         spin_unlock(&ns->ns_hash_lock);
 502
 503                         cleanup_resource(res, &res->lr_granted, flags);
 504                         cleanup_resource(res, &res->lr_converting, flags);
 505                         cleanup_resource(res, &res->lr_waiting, flags);
 506
 507                         spin_lock(&ns->ns_hash_lock);
 508                         tmp = tmp->next;
 509
 510                         /* XXX: former stuff caused issues in case of race
 511                          * between ldlm_namespace_cleanup() and lockd() when
 512                          * client gets blocking ast when lock gets distracted by
 513                          * server. This is 1_4 branch solution, let's see how
 514                          * will it behave. */
 515                         if (!ldlm_resource_putref_locked(res)) {
 516                                 CERROR("Namespace %s resource refcount nonzero "
 517                                        "(%d) after lock cleanup; forcing cleanup.\n",
 518                                        ns->ns_name, atomic_read(&res->lr_refcount));
 519                                 CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
 520                                        LPU64") (rc: %d)\n", res,
 521                                        res->lr_name.name[0], res->lr_name.name[1],
 522                                        res->lr_name.name[2], res->lr_name.name[3],
 523                                        atomic_read(&res->lr_refcount));
 524                         }
 525                 }
 526                 spin_unlock(&ns->ns_hash_lock);
 527         }
 528
 529         return ELDLM_OK;
 530 }
 531
 532 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 533 {
 534         ENTRY;
 535         /* At shutdown time, don't call the cancellation callback */
 536         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 537
 538         if (ns->ns_refcount > 0) {
 539                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 540                 int rc;
 541                 CDEBUG(D_DLMTRACE,
 542                        "dlm namespace %s free waiting on refcount %d\n",
 543                        ns->ns_name, ns->ns_refcount);
 544 force_wait:
 545                 if (force)
 546                         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 547
 548                 rc = l_wait_event(ns->ns_waitq,
 549                                   ns->ns_refcount == 0, &lwi);
 550
 551                 /* Forced cleanups should be able to reclaim all references,
 552                  * so it's safe to wait forever... we can't leak locks... */
 553                 if (force && rc == -ETIMEDOUT) {
 554                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 555                                        "namespace with %d resources in use, "
 556                                        "(rc=%d)\n", ns->ns_name,
 557                                        ns->ns_refcount, rc);
 558                         GOTO(force_wait, rc);
 559                 }
 560
 561                 if (ns->ns_refcount) {
 562                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 563                                        "with %d resources in use, (rc=%d)\n",
 564                                        ns->ns_name,
 565                                        ns->ns_refcount, rc);
 566                         RETURN(ELDLM_NAMESPACE_EXISTS);
 567                 }
 568                 CDEBUG(D_DLMTRACE,
 569                        "dlm namespace %s free done waiting\n", ns->ns_name);
 570         }
 571
 572         RETURN(ELDLM_OK);
 573 }
 574
 575 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 576                                struct obd_import *imp,
 577                                int force)
 578 {
 579         int rc;
 580         ENTRY;
 581         if (!ns) {
 582                 EXIT;
 583                 return;
 584         }
 585
 586
 587         /* Can fail with -EINTR when force == 0 in which case try harder */
 588         rc = __ldlm_namespace_free(ns, force);
 589         if (rc != ELDLM_OK) {
 590                 if (imp) {
 591                         ptlrpc_disconnect_import(imp, 0);
 592                         ptlrpc_invalidate_import(imp);
 593                 }
 594
 595                 /* With all requests dropped and the import inactive
 596                  * we are gaurenteed all reference will be dropped. */
 597                 rc = __ldlm_namespace_free(ns, 1);
 598                 LASSERT(rc == 0);
 599         }
 600         EXIT;
 601 }
 602
 603 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 604 {
 605         ENTRY;
 606         if (!ns) {
 607                 EXIT;
 608                 return;
 609         }
 610
 611         /* Make sure that nobody can find this ns in its list. */
 612         ldlm_namespace_unregister(ns, ns->ns_client);
 613
 614         /* Fini pool _before_ parent proc dir is removed. This is important
 615          * as ldlm_pool_fini() removes own proc dir which is child to @dir.
 616          * Removing it after @dir may cause oops. */
 617         ldlm_pool_fini(&ns->ns_pool);
 618
 619 #ifdef LPROCFS
 620         {
 621                 struct proc_dir_entry *dir;
 622                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
 623                 if (dir == NULL) {
 624                         CERROR("dlm namespace %s has no procfs dir?\n",
 625                                ns->ns_name);
 626                 } else {
 627                         lprocfs_remove(&dir);
 628                 }
 629         }
 630 #endif
 631         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 632         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 633
 634         /* @ns should be not on list in this time, otherwise this will cause
 635          * issues realted to using freed @ns in pools thread. */
 636         LASSERT(list_empty(&ns->ns_list_chain));
 637         OBD_FREE_PTR(ns);
 638         ldlm_put_ref();
 639         EXIT;
 640 }
 641
 642 /* Cleanup the resource, and free namespace.
 643  * bug 12864:
 644  * Deadlock issue:
 645  * proc1: destroy import
 646  *        class_disconnect_export(grab cl_sem) ->
 647  *              -> ldlm_namespace_free ->
 648  *              -> lprocfs_remove(grab _lprocfs_lock).
 649  * proc2: read proc info
 650  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 651  *              -> osc_rd_active, etc(grab cl_sem).
 652  *
 653  * So that I have to split the ldlm_namespace_free into two parts - the first
 654  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 655  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 656  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 657  * held.
 658  */
 659 void ldlm_namespace_free(struct ldlm_namespace *ns,
 660                          struct obd_import *imp,
 661                          int force)
 662 {
 663         ldlm_namespace_free_prior(ns, imp, force);
 664         ldlm_namespace_free_post(ns);
 665 }
 666
 667 void ldlm_namespace_get_locked(struct ldlm_namespace *ns)
 668 {
 669         ns->ns_refcount++;
 670 }
 671
 672 void ldlm_namespace_get(struct ldlm_namespace *ns)
 673 {
 674         spin_lock(&ns->ns_hash_lock);
 675         ldlm_namespace_get_locked(ns);
 676         spin_unlock(&ns->ns_hash_lock);
 677 }
 678
 679 void ldlm_namespace_put_locked(struct ldlm_namespace *ns, int wakeup)
 680 {
 681         LASSERT(ns->ns_refcount > 0);
 682         ns->ns_refcount--;
 683         if (ns->ns_refcount == 0 && wakeup)
 684                 wake_up(&ns->ns_waitq);
 685 }
 686
 687 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
 688 {
 689         spin_lock(&ns->ns_hash_lock);
 690         ldlm_namespace_put_locked(ns, wakeup);
 691         spin_unlock(&ns->ns_hash_lock);
 692 }
 693
 694 /* Register @ns in the list of namespaces */
 695 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
 696 {
 697         mutex_down(ldlm_namespace_lock(client));
 698         LASSERT(list_empty(&ns->ns_list_chain));
 699         list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 700         atomic_inc(ldlm_namespace_nr(client));
 701         mutex_up(ldlm_namespace_lock(client));
 702 }
 703
 704 /* Unregister @ns from the list of namespaces */
 705 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
 706 {
 707         mutex_down(ldlm_namespace_lock(client));
 708         LASSERT(!list_empty(&ns->ns_list_chain));
 709         /*
 710          * Some asserts and possibly other parts of code still using
 711          * list_empty(&ns->ns_list_chain). This is why it is important
 712          * to use list_del_init() here.
 713          */
 714         list_del_init(&ns->ns_list_chain);
 715         atomic_dec(ldlm_namespace_nr(client));
 716         mutex_up(ldlm_namespace_lock(client));
 717 }
 718
 719 /* Should be called under ldlm_namespace_lock(client) taken */
 720 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
 721 {
 722         LASSERT(!list_empty(&ns->ns_list_chain));
 723         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 724         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 725 }
 726
 727 /* Should be called under ldlm_namespace_lock(client) taken */
 728 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
 729 {
 730         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 731         LASSERT(!list_empty(ldlm_namespace_list(client)));
 732         return container_of(ldlm_namespace_list(client)->next,
 733                 struct ldlm_namespace, ns_list_chain);
 734 }
 735
 736 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, struct ldlm_res_id name)
 737 {
 738         __u32 hash = 0;
 739         int i;
 740
 741         for (i = 0; i < RES_NAME_SIZE; i++)
 742                 hash += name.name[i];
 743
 744         hash += (__u32)((unsigned long)parent >> 4);
 745
 746         return (hash & RES_HASH_MASK);
 747 }
 748
 749 static struct ldlm_resource *ldlm_resource_new(void)
 750 {
 751         struct ldlm_resource *res;
 752         int idx;
 753
 754         OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
 755         if (res == NULL)
 756                 return NULL;
 757
 758         memset(res, 0, sizeof(*res));
 759
 760         CFS_INIT_LIST_HEAD(&res->lr_children);
 761         CFS_INIT_LIST_HEAD(&res->lr_childof);
 762         CFS_INIT_LIST_HEAD(&res->lr_granted);
 763         CFS_INIT_LIST_HEAD(&res->lr_converting);
 764         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 765
 766         /* initialize interval trees for each lock mode*/
 767         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 768                 res->lr_itree[idx].lit_size = 0;
 769                 res->lr_itree[idx].lit_mode = 1 << idx;
 770                 res->lr_itree[idx].lit_root = NULL;
 771         }
 772
 773         atomic_set(&res->lr_refcount, 1);
 774         spin_lock_init(&res->lr_lock);
 775
 776         /* one who creates the resource must unlock
 777          * the semaphore after lvb initialization */
 778         init_MUTEX_LOCKED(&res->lr_lvb_sem);
 779
 780         return res;
 781 }
 782
 783 /* must be called with hash lock held */
 784 static struct ldlm_resource *
 785 ldlm_resource_find(struct ldlm_namespace *ns, struct ldlm_res_id name, __u32 hash)
 786 {
 787         struct list_head *bucket, *tmp;
 788         struct ldlm_resource *res;
 789
 790         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 791         bucket = ns->ns_hash + hash;
 792
 793         list_for_each(tmp, bucket) {
 794                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 795                 if (memcmp(&res->lr_name, &name, sizeof(res->lr_name)) == 0)
 796                         return res;
 797         }
 798
 799         return NULL;
 800 }
 801
 802 /* Args: locked namespace
 803  * Returns: newly-allocated, referenced, unlocked resource */
 804 static struct ldlm_resource *
 805 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 806                   struct ldlm_res_id name, __u32 hash, ldlm_type_t type)
 807 {
 808         struct list_head *bucket;
 809         struct ldlm_resource *res, *old_res;
 810         ENTRY;
 811
 812         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
 813                  "type: %d\n", type);
 814
 815         res = ldlm_resource_new();
 816         if (!res)
 817                 RETURN(NULL);
 818
 819         res->lr_name = name;
 820         res->lr_namespace = ns;
 821         res->lr_type = type;
 822         res->lr_most_restr = LCK_NL;
 823
 824         spin_lock(&ns->ns_hash_lock);
 825         old_res = ldlm_resource_find(ns, name, hash);
 826         if (old_res) {
 827                 /* someone won the race and added the resource before */
 828                 ldlm_resource_getref(old_res);
 829                 spin_unlock(&ns->ns_hash_lock);
 830                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 831                 /* synchronize WRT resource creation */
 832                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 833                         down(&old_res->lr_lvb_sem);
 834                         up(&old_res->lr_lvb_sem);
 835                 }
 836                 RETURN(old_res);
 837         }
 838
 839         /* we won! let's add the resource */
 840         bucket = ns->ns_hash + hash;
 841         list_add(&res->lr_hash, bucket);
 842         ns->ns_resources++;
 843         ldlm_namespace_get_locked(ns);
 844
 845         if (parent == NULL) {
 846                 list_add(&res->lr_childof, &ns->ns_root_list);
 847         } else {
 848                 res->lr_parent = parent;
 849                 list_add(&res->lr_childof, &parent->lr_children);
 850         }
 851         spin_unlock(&ns->ns_hash_lock);
 852
 853         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 854                 int rc;
 855
 856                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
 857                 rc = ns->ns_lvbo->lvbo_init(res);
 858                 if (rc)
 859                         CERROR("lvbo_init failed for resource "
 860                                LPU64": rc %d\n", name.name[0], rc);
 861                 /* we create resource with locked lr_lvb_sem */
 862                 up(&res->lr_lvb_sem);
 863         }
 864
 865         RETURN(res);
 866 }
 867
 868 /* Args: unlocked namespace
 869  * Locks: takes and releases ns->ns_lock and res->lr_lock
 870  * Returns: referenced, unlocked ldlm_resource or NULL */
 871 struct ldlm_resource *
 872 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 873                   struct ldlm_res_id name, ldlm_type_t type, int create)
 874 {
 875         __u32 hash = ldlm_hash_fn(parent, name);
 876         struct ldlm_resource *res = NULL;
 877         ENTRY;
 878
 879         LASSERT(ns != NULL);
 880         LASSERT(ns->ns_hash != NULL);
 881         LASSERT(name.name[0] != 0);
 882
 883         spin_lock(&ns->ns_hash_lock);
 884         res = ldlm_resource_find(ns, name, hash);
 885         if (res) {
 886                 ldlm_resource_getref(res);
 887                 spin_unlock(&ns->ns_hash_lock);
 888                 /* synchronize WRT resource creation */
 889                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 890                         down(&res->lr_lvb_sem);
 891                         up(&res->lr_lvb_sem);
 892                 }
 893                 RETURN(res);
 894         }
 895         spin_unlock(&ns->ns_hash_lock);
 896
 897         if (create == 0)
 898                 RETURN(NULL);
 899
 900         res = ldlm_resource_add(ns, parent, name, hash, type);
 901         RETURN(res);
 902 }
 903
 904 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 905 {
 906         LASSERT(res != NULL);
 907         LASSERT(res != LP_POISON);
 908         atomic_inc(&res->lr_refcount);
 909         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 910                atomic_read(&res->lr_refcount));
 911         return res;
 912 }
 913
 914 void __ldlm_resource_putref_final(struct ldlm_resource *res)
 915 {
 916         struct ldlm_namespace *ns = res->lr_namespace;
 917
 918         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 919
 920         if (!list_empty(&res->lr_granted)) {
 921                 ldlm_resource_dump(D_ERROR, res);
 922                 LBUG();
 923         }
 924
 925         if (!list_empty(&res->lr_converting)) {
 926                 ldlm_resource_dump(D_ERROR, res);
 927                 LBUG();
 928         }
 929
 930         if (!list_empty(&res->lr_waiting)) {
 931                 ldlm_resource_dump(D_ERROR, res);
 932                 LBUG();
 933         }
 934
 935         if (!list_empty(&res->lr_children)) {
 936                 ldlm_resource_dump(D_ERROR, res);
 937                 LBUG();
 938         }
 939
 940         /* Pass 0 as second argument to not wake up ->ns_waitq yet, will do it
 941          * later. */
 942         ldlm_namespace_put_locked(ns, 0);
 943         list_del_init(&res->lr_hash);
 944         list_del_init(&res->lr_childof);
 945
 946         ns->ns_resources--;
 947         if (ns->ns_resources == 0)
 948                 wake_up(&ns->ns_waitq);
 949 }
 950
 951 /* Returns 1 if the resource was freed, 0 if it remains. */
 952 int ldlm_resource_putref(struct ldlm_resource *res)
 953 {
 954         struct ldlm_namespace *ns = res->lr_namespace;
 955         int rc = 0;
 956         ENTRY;
 957
 958         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 959                atomic_read(&res->lr_refcount) - 1);
 960         LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
 961                  atomic_read(&res->lr_refcount));
 962         LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
 963                  atomic_read(&res->lr_refcount));
 964
 965         if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
 966                 __ldlm_resource_putref_final(res);
 967                 spin_unlock(&ns->ns_hash_lock);
 968                 if (res->lr_lvb_data)
 969                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 970                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 971                 rc = 1;
 972         }
 973
 974         RETURN(rc);
 975 }
 976
 977 /* Returns 1 if the resource was freed, 0 if it remains. */
 978 int ldlm_resource_putref_locked(struct ldlm_resource *res)
 979 {
 980         int rc = 0;
 981         ENTRY;
 982
 983         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 984                atomic_read(&res->lr_refcount) - 1);
 985         LASSERT(atomic_read(&res->lr_refcount) > 0);
 986         LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
 987
 988         LASSERT(atomic_read(&res->lr_refcount) >= 0);
 989         if (atomic_dec_and_test(&res->lr_refcount)) {
 990                 __ldlm_resource_putref_final(res);
 991                 if (res->lr_lvb_data)
 992                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 993                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 994                 rc = 1;
 995         }
 996
 997         RETURN(rc);
 998 }
 999
1000 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1001                             struct ldlm_lock *lock)
1002 {
1003         check_res_locked(res);
1004
1005         ldlm_resource_dump(D_INFO, res);
1006         CDEBUG(D_OTHER, "About to add this lock:\n");
1007         ldlm_lock_dump(D_OTHER, lock, 0);
1008
1009         if (lock->l_destroyed) {
1010                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1011                 return;
1012         }
1013
1014         LASSERT(list_empty(&lock->l_res_link));
1015
1016         list_add_tail(&lock->l_res_link, head);
1017 }
1018
1019 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1020                                      struct ldlm_lock *new)
1021 {
1022         struct ldlm_resource *res = original->l_resource;
1023
1024         check_res_locked(res);
1025
1026         ldlm_resource_dump(D_INFO, res);
1027         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1028         ldlm_lock_dump(D_OTHER, new, 0);
1029
1030         if (new->l_destroyed) {
1031                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1032                 goto out;
1033         }
1034
1035         LASSERT(list_empty(&new->l_res_link));
1036
1037         list_add(&new->l_res_link, &original->l_res_link);
1038  out:;
1039 }
1040
1041 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1042 {
1043         int type = lock->l_resource->lr_type;
1044
1045         check_res_locked(lock->l_resource);
1046         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1047                 ldlm_unlink_lock_skiplist(lock);
1048         else if (type == LDLM_EXTENT)
1049                 ldlm_extent_unlink_lock(lock);
1050         list_del_init(&lock->l_res_link);
1051 }
1052
1053 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1054 {
1055         desc->lr_type = res->lr_type;
1056         desc->lr_name = res->lr_name;
1057 }
1058
1059 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1060 {
1061         struct list_head *tmp;
1062
1063         if (!((libcfs_debug | D_ERROR) & level))
1064                 return;
1065
1066         mutex_down(ldlm_namespace_lock(client));
1067
1068         list_for_each(tmp, ldlm_namespace_list(client)) {
1069                 struct ldlm_namespace *ns;
1070                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1071                 ldlm_namespace_dump(level, ns);
1072         }
1073
1074         mutex_up(ldlm_namespace_lock(client));
1075 }
1076
1077 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1078 {
1079         struct list_head *tmp;
1080
1081         if (!((libcfs_debug | D_ERROR) & level))
1082                 return;
1083
1084         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1085                ns->ns_name, ns->ns_refcount,
1086                ns_is_client(ns) ? "client" : "server");
1087
1088         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1089                 return;
1090
1091         spin_lock(&ns->ns_hash_lock);
1092         tmp = ns->ns_root_list.next;
1093         while (tmp != &ns->ns_root_list) {
1094                 struct ldlm_resource *res;
1095                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
1096
1097                 ldlm_resource_getref(res);
1098                 spin_unlock(&ns->ns_hash_lock);
1099
1100                 lock_res(res);
1101                 ldlm_resource_dump(level, res);
1102                 unlock_res(res);
1103
1104                 spin_lock(&ns->ns_hash_lock);
1105                 tmp = tmp->next;
1106                 ldlm_resource_putref_locked(res);
1107         }
1108         ns->ns_next_dump = cfs_time_shift(10);
1109         spin_unlock(&ns->ns_hash_lock);
1110 }
1111
1112 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1113 {
1114         struct list_head *tmp;
1115         int pos;
1116
1117         CLASSERT(RES_NAME_SIZE == 4);
1118
1119         if (!((libcfs_debug | D_ERROR) & level))
1120                 return;
1121
1122         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1123                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1124                res->lr_name.name[2], res->lr_name.name[3],
1125                atomic_read(&res->lr_refcount));
1126
1127         if (!list_empty(&res->lr_granted)) {
1128                 pos = 0;
1129                 CDEBUG(level, "Granted locks:\n");
1130                 list_for_each(tmp, &res->lr_granted) {
1131                         struct ldlm_lock *lock;
1132                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1133                         ldlm_lock_dump(level, lock, ++pos);
1134                 }
1135         }
1136         if (!list_empty(&res->lr_converting)) {
1137                 pos = 0;
1138                 CDEBUG(level, "Converting locks:\n");
1139                 list_for_each(tmp, &res->lr_converting) {
1140                         struct ldlm_lock *lock;
1141                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1142                         ldlm_lock_dump(level, lock, ++pos);
1143                 }
1144         }
1145         if (!list_empty(&res->lr_waiting)) {
1146                 pos = 0;
1147                 CDEBUG(level, "Waiting locks:\n");
1148                 list_for_each(tmp, &res->lr_waiting) {
1149                         struct ldlm_lock *lock;
1150                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1151                         ldlm_lock_dump(level, lock, ++pos);
1152                 }
1153         }
1154 }