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 (c) 2002, 2010, Oracle and/or its affiliates. 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                                 CDEBUG(D_DLMTRACE,
 181                                        "not all requested locks are canceled, "
 182                                        "requested: %d, canceled: %d\n", unused,
 183                                        canceled);
 184                                 return -EINVAL;
 185                         }
 186                 } else {
 187                         tmp = ns->ns_max_unused;
 188                         ns->ns_max_unused = 0;
 189                         ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
 190                         ns->ns_max_unused = tmp;
 191                 }
 192                 return count;
 193         }
 194
 195         tmp = simple_strtoul(dummy, &end, 0);
 196         if (dummy == end) {
 197                 CERROR("invalid value written\n");
 198                 return -EINVAL;
 199         }
 200         lru_resize = (tmp == 0);
 201
 202         if (ns_connect_lru_resize(ns)) {
 203                 if (!lru_resize)
 204                         ns->ns_max_unused = (unsigned int)tmp;
 205
 206                 if (tmp > ns->ns_nr_unused)
 207                         tmp = ns->ns_nr_unused;
 208                 tmp = ns->ns_nr_unused - tmp;
 209
 210                 CDEBUG(D_DLMTRACE,
 211                        "changing namespace %s unused locks from %u to %u\n",
 212                        ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
 213                 ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 214
 215                 if (!lru_resize) {
 216                         CDEBUG(D_DLMTRACE,
 217                                "disable lru_resize for namespace %s\n",
 218                                ns->ns_name);
 219                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 220                 }
 221         } else {
 222                 CDEBUG(D_DLMTRACE,
 223                        "changing namespace %s max_unused from %u to %u\n",
 224                        ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
 225                 ns->ns_max_unused = (unsigned int)tmp;
 226                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 227
 228                 /* Make sure that originally lru resize was supported before
 229                  * turning it on here. */
 230                 if (lru_resize &&
 231                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 232                         CDEBUG(D_DLMTRACE,
 233                                "enable lru_resize for namespace %s\n",
 234                                ns->ns_name);
 235                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 236                 }
 237         }
 238         return count;
 239 }
 240
 241 void ldlm_proc_namespace(struct ldlm_namespace *ns)
 242 {
 243         struct lprocfs_vars lock_vars[2];
 244         char lock_name[MAX_STRING_SIZE + 1];
 245
 246         LASSERT(ns != NULL);
 247         LASSERT(ns->ns_name != NULL);
 248
 249         lock_name[MAX_STRING_SIZE] = '\0';
 250
 251         memset(lock_vars, 0, sizeof(lock_vars));
 252         lock_vars[0].name = lock_name;
 253
 254         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
 255         lock_vars[0].data = &ns->ns_refcount;
 256         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 257         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 258
 259         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
 260         lock_vars[0].data = &ns->ns_locks;
 261         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 262         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 263
 264         if (ns_is_client(ns)) {
 265                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 266                          ns->ns_name);
 267                 lock_vars[0].data = &ns->ns_nr_unused;
 268                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 269                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 270
 271                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 272                          ns->ns_name);
 273                 lock_vars[0].data = ns;
 274                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 275                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 276                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 277
 278                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 279                          ns->ns_name);
 280                 lock_vars[0].data = &ns->ns_max_age;
 281                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 282                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 283                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 284         } else {
 285                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
 286                          ns->ns_name);
 287                 lock_vars[0].data = &ns->ns_timeouts;
 288                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 289                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 290
 291                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 292                          ns->ns_name);
 293                 lock_vars[0].data = &ns->ns_max_nolock_size;
 294                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 295                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 296                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 297
 298                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 299                          ns->ns_name);
 300                 lock_vars[0].data = &ns->ns_contention_time;
 301                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 302                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 303                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 304
 305                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 306                          ns->ns_name);
 307                 lock_vars[0].data = &ns->ns_contended_locks;
 308                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 309                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 310                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 311         }
 312 }
 313 #undef MAX_STRING_SIZE
 314 #else
 315 #define ldlm_proc_namespace(ns) do {} while (0)
 316 #endif /* LPROCFS */
 317
 318 struct ldlm_namespace *
 319 ldlm_namespace_new(struct obd_device *obd, char *name,
 320                    ldlm_side_t client, ldlm_appetite_t apt)
 321 {
 322         struct ldlm_namespace *ns = NULL;
 323         struct list_head *bucket;
 324         int rc, idx, namelen;
 325         ENTRY;
 326
 327         rc = ldlm_get_ref();
 328         if (rc) {
 329                 CERROR("ldlm_get_ref failed: %d\n", rc);
 330                 RETURN(NULL);
 331         }
 332
 333         OBD_ALLOC_PTR(ns);
 334         if (!ns)
 335                 GOTO(out_ref, NULL);
 336
 337         OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 338         if (!ns->ns_hash)
 339                 GOTO(out_ns, NULL);
 340
 341         namelen = strlen(name);
 342         OBD_ALLOC(ns->ns_name, namelen + 1);
 343         if (!ns->ns_name)
 344                 GOTO(out_hash, NULL);
 345
 346         ns->ns_appetite = apt;
 347
 348         LASSERT(obd != NULL);
 349         ns->ns_obd = obd;
 350
 351         strcpy(ns->ns_name, name);
 352
 353         CFS_INIT_LIST_HEAD(&ns->ns_root_list);
 354         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 355         ns->ns_refcount = 0;
 356         ns->ns_client = client;
 357         spin_lock_init(&ns->ns_hash_lock);
 358         atomic_set(&ns->ns_locks, 0);
 359         ns->ns_resources = 0;
 360         cfs_waitq_init(&ns->ns_waitq);
 361         ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
 362         ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
 363         ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
 364
 365         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 366              bucket--)
 367                 CFS_INIT_LIST_HEAD(bucket);
 368
 369         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 370         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 371         ns->ns_nr_unused = 0;
 372         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
 373         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
 374         ns->ns_timeouts = 0;
 375         spin_lock_init(&ns->ns_unused_lock);
 376         ns->ns_orig_connect_flags = 0;
 377         ns->ns_connect_flags = 0;
 378
 379         ldlm_proc_namespace(ns);
 380
 381         idx = atomic_read(ldlm_namespace_nr(client));
 382
 383         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 384         if (rc) {
 385                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 386                 GOTO(out_proc, rc);
 387         }
 388
 389         at_init(&ns->ns_at_estimate, ldlm_enqueue_min, 0);
 390         ldlm_namespace_register(ns, client);
 391         RETURN(ns);
 392 out_proc:
 393         ldlm_namespace_cleanup(ns, 0);
 394         OBD_FREE(ns->ns_name, namelen + 1);
 395 out_hash:
 396         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 397 out_ns:
 398         OBD_FREE_PTR(ns);
 399 out_ref:
 400         ldlm_put_ref();
 401         RETURN(NULL);
 402 }
 403
 404 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 405
 406 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 407  * This is currently only used for recovery, and we make certain assumptions
 408  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 409  *
 410  * Called with the ns_lock held. */
 411 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 412                              int flags)
 413 {
 414         struct list_head *tmp;
 415         int rc = 0, client = ns_is_client(res->lr_namespace);
 416         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 417         ENTRY;
 418
 419         do {
 420                 struct ldlm_lock *lock = NULL;
 421
 422                 /* first, we look for non-cleaned-yet lock
 423                  * all cleaned locks are marked by CLEANED flag */
 424                 lock_res(res);
 425                 list_for_each(tmp, q) {
 426                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 427                         if (lock->l_flags & LDLM_FL_CLEANED) {
 428                                 lock = NULL;
 429                                 continue;
 430                         }
 431                         LDLM_LOCK_GET(lock);
 432                         lock->l_flags |= LDLM_FL_CLEANED;
 433                         break;
 434                 }
 435
 436                 if (lock == NULL) {
 437                         unlock_res(res);
 438                         break;
 439                 }
 440
 441                 /* Set CBPENDING so nothing in the cancellation path
 442                  * can match this lock */
 443                 lock->l_flags |= LDLM_FL_CBPENDING;
 444                 lock->l_flags |= LDLM_FL_FAILED;
 445                 lock->l_flags |= flags;
 446
 447                 /* ... without sending a CANCEL message for local_only. */
 448                 if (local_only)
 449                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 450
 451                 if (local_only && (lock->l_readers || lock->l_writers)) {
 452                         /* This is a little bit gross, but much better than the
 453                          * alternative: pretend that we got a blocking AST from
 454                          * the server, so that when the lock is decref'd, it
 455                          * will go away ... */
 456                         unlock_res(res);
 457                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 458                         if (lock->l_completion_ast)
 459                                 lock->l_completion_ast(lock, 0, NULL);
 460                         LDLM_LOCK_PUT(lock);
 461                         continue;
 462                 }
 463
 464                 if (client) {
 465                         struct lustre_handle lockh;
 466
 467                         unlock_res(res);
 468                         ldlm_lock2handle(lock, &lockh);
 469                         rc = ldlm_cli_cancel(&lockh);
 470                         if (rc)
 471                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 472                 } else {
 473                         ldlm_resource_unlink_lock(lock);
 474                         unlock_res(res);
 475                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 476                                    "client node");
 477                         ldlm_lock_destroy(lock);
 478                 }
 479                 LDLM_LOCK_PUT(lock);
 480         } while (1);
 481
 482         EXIT;
 483 }
 484
 485 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 486 {
 487         struct list_head *tmp;
 488         int i;
 489
 490         if (ns == NULL) {
 491                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 492                 return ELDLM_OK;
 493         }
 494
 495         for (i = 0; i < RES_HASH_SIZE; i++) {
 496                 spin_lock(&ns->ns_hash_lock);
 497                 tmp = ns->ns_hash[i].next;
 498                 while (tmp != &(ns->ns_hash[i])) {
 499                         struct ldlm_resource *res;
 500                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 501                         ldlm_resource_getref(res);
 502                         spin_unlock(&ns->ns_hash_lock);
 503
 504                         cleanup_resource(res, &res->lr_granted, flags);
 505                         cleanup_resource(res, &res->lr_converting, flags);
 506                         cleanup_resource(res, &res->lr_waiting, flags);
 507
 508                         spin_lock(&ns->ns_hash_lock);
 509                         tmp = tmp->next;
 510
 511                         /* XXX: former stuff caused issues in case of race
 512                          * between ldlm_namespace_cleanup() and lockd() when
 513                          * client gets blocking ast when lock gets distracted by
 514                          * server. This is 1_4 branch solution, let's see how
 515                          * will it behave. */
 516                         if (!ldlm_resource_putref_locked(res)) {
 517                                 CERROR("Namespace %s resource refcount nonzero "
 518                                        "(%d) after lock cleanup; forcing cleanup.\n",
 519                                        ns->ns_name, atomic_read(&res->lr_refcount));
 520                                 CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
 521                                        LPU64") (rc: %d)\n", res,
 522                                        res->lr_name.name[0], res->lr_name.name[1],
 523                                        res->lr_name.name[2], res->lr_name.name[3],
 524                                        atomic_read(&res->lr_refcount));
 525                         }
 526                 }
 527                 spin_unlock(&ns->ns_hash_lock);
 528         }
 529
 530         return ELDLM_OK;
 531 }
 532
 533 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 534 {
 535         ENTRY;
 536         /* At shutdown time, don't call the cancellation callback */
 537         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 538
 539         if (ns->ns_refcount > 0) {
 540                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 541                 int rc;
 542                 CDEBUG(D_DLMTRACE,
 543                        "dlm namespace %s free waiting on refcount %d\n",
 544                        ns->ns_name, ns->ns_refcount);
 545 force_wait:
 546                 if (force)
 547                         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 548
 549                 rc = l_wait_event(ns->ns_waitq,
 550                                   ns->ns_refcount == 0, &lwi);
 551
 552                 /* Forced cleanups should be able to reclaim all references,
 553                  * so it's safe to wait forever... we can't leak locks... */
 554                 if (force && rc == -ETIMEDOUT) {
 555                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 556                                        "namespace with %d resources in use, "
 557                                        "(rc=%d)\n", ns->ns_name,
 558                                        ns->ns_refcount, rc);
 559                         GOTO(force_wait, rc);
 560                 }
 561
 562                 if (ns->ns_refcount) {
 563                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 564                                        "with %d resources in use, (rc=%d)\n",
 565                                        ns->ns_name,
 566                                        ns->ns_refcount, rc);
 567                         RETURN(ELDLM_NAMESPACE_EXISTS);
 568                 }
 569                 CDEBUG(D_DLMTRACE,
 570                        "dlm namespace %s free done waiting\n", ns->ns_name);
 571         }
 572
 573         RETURN(ELDLM_OK);
 574 }
 575
 576 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 577                                struct obd_import *imp,
 578                                int force)
 579 {
 580         int rc;
 581         ENTRY;
 582         if (!ns) {
 583                 EXIT;
 584                 return;
 585         }
 586
 587
 588         /* Can fail with -EINTR when force == 0 in which case try harder */
 589         rc = __ldlm_namespace_free(ns, force);
 590         if (rc != ELDLM_OK) {
 591                 if (imp) {
 592                         ptlrpc_disconnect_import(imp, 0);
 593                         ptlrpc_invalidate_import(imp);
 594                 }
 595
 596                 /* With all requests dropped and the import inactive
 597                  * we are gaurenteed all reference will be dropped. */
 598                 rc = __ldlm_namespace_free(ns, 1);
 599                 LASSERT(rc == 0);
 600         }
 601         EXIT;
 602 }
 603
 604 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 605 {
 606         ENTRY;
 607         if (!ns) {
 608                 EXIT;
 609                 return;
 610         }
 611
 612         /* Make sure that nobody can find this ns in its list. */
 613         ldlm_namespace_unregister(ns, ns->ns_client);
 614
 615         /* Fini pool _before_ parent proc dir is removed. This is important
 616          * as ldlm_pool_fini() removes own proc dir which is child to @dir.
 617          * Removing it after @dir may cause oops. */
 618         ldlm_pool_fini(&ns->ns_pool);
 619
 620 #ifdef LPROCFS
 621         {
 622                 struct proc_dir_entry *dir;
 623                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
 624                 if (dir == NULL) {
 625                         CERROR("dlm namespace %s has no procfs dir?\n",
 626                                ns->ns_name);
 627                 } else {
 628                         lprocfs_remove(&dir);
 629                 }
 630         }
 631 #endif
 632         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 633         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 634
 635         /* @ns should be not on list in this time, otherwise this will cause
 636          * issues realted to using freed @ns in pools thread. */
 637         LASSERT(list_empty(&ns->ns_list_chain));
 638         OBD_FREE_PTR(ns);
 639         ldlm_put_ref();
 640         EXIT;
 641 }
 642
 643 /* Cleanup the resource, and free namespace.
 644  * bug 12864:
 645  * Deadlock issue:
 646  * proc1: destroy import
 647  *        class_disconnect_export(grab cl_sem) ->
 648  *              -> ldlm_namespace_free ->
 649  *              -> lprocfs_remove(grab _lprocfs_lock).
 650  * proc2: read proc info
 651  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 652  *              -> osc_rd_active, etc(grab cl_sem).
 653  *
 654  * So that I have to split the ldlm_namespace_free into two parts - the first
 655  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 656  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 657  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 658  * held.
 659  */
 660 void ldlm_namespace_free(struct ldlm_namespace *ns,
 661                          struct obd_import *imp,
 662                          int force)
 663 {
 664         ldlm_namespace_free_prior(ns, imp, force);
 665         ldlm_namespace_free_post(ns);
 666 }
 667
 668 void ldlm_namespace_get_locked(struct ldlm_namespace *ns)
 669 {
 670         ns->ns_refcount++;
 671 }
 672
 673 void ldlm_namespace_get(struct ldlm_namespace *ns)
 674 {
 675         spin_lock(&ns->ns_hash_lock);
 676         ldlm_namespace_get_locked(ns);
 677         spin_unlock(&ns->ns_hash_lock);
 678 }
 679
 680 void ldlm_namespace_put_locked(struct ldlm_namespace *ns, int wakeup)
 681 {
 682         LASSERT(ns->ns_refcount > 0);
 683         ns->ns_refcount--;
 684         if (ns->ns_refcount == 0 && wakeup)
 685                 wake_up(&ns->ns_waitq);
 686 }
 687
 688 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
 689 {
 690         spin_lock(&ns->ns_hash_lock);
 691         ldlm_namespace_put_locked(ns, wakeup);
 692         spin_unlock(&ns->ns_hash_lock);
 693 }
 694
 695 /* Register @ns in the list of namespaces */
 696 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
 697 {
 698         mutex_down(ldlm_namespace_lock(client));
 699         LASSERT(list_empty(&ns->ns_list_chain));
 700         list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 701         atomic_inc(ldlm_namespace_nr(client));
 702         mutex_up(ldlm_namespace_lock(client));
 703 }
 704
 705 /* Unregister @ns from the list of namespaces */
 706 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
 707 {
 708         mutex_down(ldlm_namespace_lock(client));
 709         LASSERT(!list_empty(&ns->ns_list_chain));
 710         /*
 711          * Some asserts and possibly other parts of code still using
 712          * list_empty(&ns->ns_list_chain). This is why it is important
 713          * to use list_del_init() here.
 714          */
 715         list_del_init(&ns->ns_list_chain);
 716         atomic_dec(ldlm_namespace_nr(client));
 717         mutex_up(ldlm_namespace_lock(client));
 718 }
 719
 720 /* Should be called under ldlm_namespace_lock(client) taken */
 721 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
 722 {
 723         LASSERT(!list_empty(&ns->ns_list_chain));
 724         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 725         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 726 }
 727
 728 /* Should be called under ldlm_namespace_lock(client) taken */
 729 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
 730 {
 731         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 732         LASSERT(!list_empty(ldlm_namespace_list(client)));
 733         return container_of(ldlm_namespace_list(client)->next,
 734                 struct ldlm_namespace, ns_list_chain);
 735 }
 736
 737 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, struct ldlm_res_id name)
 738 {
 739         __u32 hash = 0;
 740         int i;
 741
 742         for (i = 0; i < RES_NAME_SIZE; i++)
 743                 hash += name.name[i];
 744
 745         hash += (__u32)((unsigned long)parent >> 4);
 746
 747         return (hash & RES_HASH_MASK);
 748 }
 749
 750 static struct ldlm_resource *ldlm_resource_new(void)
 751 {
 752         struct ldlm_resource *res;
 753         int idx;
 754
 755         OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
 756         if (res == NULL)
 757                 return NULL;
 758
 759         memset(res, 0, sizeof(*res));
 760
 761         CFS_INIT_LIST_HEAD(&res->lr_children);
 762         CFS_INIT_LIST_HEAD(&res->lr_childof);
 763         CFS_INIT_LIST_HEAD(&res->lr_granted);
 764         CFS_INIT_LIST_HEAD(&res->lr_converting);
 765         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 766
 767         /* initialize interval trees for each lock mode*/
 768         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 769                 res->lr_itree[idx].lit_size = 0;
 770                 res->lr_itree[idx].lit_mode = 1 << idx;
 771                 res->lr_itree[idx].lit_root = NULL;
 772         }
 773
 774         atomic_set(&res->lr_refcount, 1);
 775         spin_lock_init(&res->lr_lock);
 776
 777         /* one who creates the resource must unlock
 778          * the semaphore after lvb initialization */
 779         init_MUTEX_LOCKED(&res->lr_lvb_sem);
 780
 781         return res;
 782 }
 783
 784 /* must be called with hash lock held */
 785 static struct ldlm_resource *
 786 ldlm_resource_find(struct ldlm_namespace *ns, struct ldlm_res_id name, __u32 hash)
 787 {
 788         struct list_head *bucket, *tmp;
 789         struct ldlm_resource *res;
 790
 791         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 792         bucket = ns->ns_hash + hash;
 793
 794         list_for_each(tmp, bucket) {
 795                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 796                 if (memcmp(&res->lr_name, &name, sizeof(res->lr_name)) == 0)
 797                         return res;
 798         }
 799
 800         return NULL;
 801 }
 802
 803 /* Args: locked namespace
 804  * Returns: newly-allocated, referenced, unlocked resource */
 805 static struct ldlm_resource *
 806 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 807                   struct ldlm_res_id name, __u32 hash, ldlm_type_t type)
 808 {
 809         struct list_head *bucket;
 810         struct ldlm_resource *res, *old_res;
 811         ENTRY;
 812
 813         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
 814                  "type: %d\n", type);
 815
 816         res = ldlm_resource_new();
 817         if (!res)
 818                 RETURN(NULL);
 819
 820         res->lr_name = name;
 821         res->lr_namespace = ns;
 822         res->lr_type = type;
 823         res->lr_most_restr = LCK_NL;
 824
 825         spin_lock(&ns->ns_hash_lock);
 826         old_res = ldlm_resource_find(ns, name, hash);
 827         if (old_res) {
 828                 /* someone won the race and added the resource before */
 829                 ldlm_resource_getref(old_res);
 830                 spin_unlock(&ns->ns_hash_lock);
 831                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 832                 /* synchronize WRT resource creation */
 833                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 834                         down(&old_res->lr_lvb_sem);
 835                         up(&old_res->lr_lvb_sem);
 836                 }
 837                 RETURN(old_res);
 838         }
 839
 840         /* we won! let's add the resource */
 841         bucket = ns->ns_hash + hash;
 842         list_add(&res->lr_hash, bucket);
 843         ns->ns_resources++;
 844         ldlm_namespace_get_locked(ns);
 845
 846         if (parent == NULL) {
 847                 list_add(&res->lr_childof, &ns->ns_root_list);
 848         } else {
 849                 res->lr_parent = parent;
 850                 list_add(&res->lr_childof, &parent->lr_children);
 851         }
 852         spin_unlock(&ns->ns_hash_lock);
 853
 854         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 855                 int rc;
 856
 857                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
 858                 rc = ns->ns_lvbo->lvbo_init(res);
 859                 if (rc)
 860                         CERROR("%s: lvbo_init failed for resource "
 861                                LPU64": rc %d\n", ns->ns_name,
 862                                name.name[0], rc);
 863                 /* we create resource with locked lr_lvb_sem */
 864                 up(&res->lr_lvb_sem);
 865         }
 866
 867         RETURN(res);
 868 }
 869
 870 /* Args: unlocked namespace
 871  * Locks: takes and releases ns->ns_lock and res->lr_lock
 872  * Returns: referenced, unlocked ldlm_resource or NULL */
 873 struct ldlm_resource *
 874 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 875                   struct ldlm_res_id name, ldlm_type_t type, int create)
 876 {
 877         __u32 hash = ldlm_hash_fn(parent, name);
 878         struct ldlm_resource *res = NULL;
 879         ENTRY;
 880
 881         LASSERT(ns != NULL);
 882         LASSERT(ns->ns_hash != NULL);
 883         LASSERT(name.name[0] != 0);
 884
 885         spin_lock(&ns->ns_hash_lock);
 886         res = ldlm_resource_find(ns, name, hash);
 887         if (res) {
 888                 ldlm_resource_getref(res);
 889                 spin_unlock(&ns->ns_hash_lock);
 890                 /* synchronize WRT resource creation */
 891                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 892                         down(&res->lr_lvb_sem);
 893                         up(&res->lr_lvb_sem);
 894                 }
 895                 RETURN(res);
 896         }
 897         spin_unlock(&ns->ns_hash_lock);
 898
 899         if (create == 0)
 900                 RETURN(NULL);
 901
 902         res = ldlm_resource_add(ns, parent, name, hash, type);
 903         RETURN(res);
 904 }
 905
 906 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 907 {
 908         LASSERT(res != NULL);
 909         LASSERT(res != LP_POISON);
 910         atomic_inc(&res->lr_refcount);
 911         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 912                atomic_read(&res->lr_refcount));
 913         return res;
 914 }
 915
 916 void __ldlm_resource_putref_final(struct ldlm_resource *res)
 917 {
 918         struct ldlm_namespace *ns = res->lr_namespace;
 919
 920         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 921
 922         if (!list_empty(&res->lr_granted)) {
 923                 ldlm_resource_dump(D_ERROR, res);
 924                 LBUG();
 925         }
 926
 927         if (!list_empty(&res->lr_converting)) {
 928                 ldlm_resource_dump(D_ERROR, res);
 929                 LBUG();
 930         }
 931
 932         if (!list_empty(&res->lr_waiting)) {
 933                 ldlm_resource_dump(D_ERROR, res);
 934                 LBUG();
 935         }
 936
 937         if (!list_empty(&res->lr_children)) {
 938                 ldlm_resource_dump(D_ERROR, res);
 939                 LBUG();
 940         }
 941
 942         /* Pass 0 as second argument to not wake up ->ns_waitq yet, will do it
 943          * later. */
 944         ldlm_namespace_put_locked(ns, 0);
 945         list_del_init(&res->lr_hash);
 946         list_del_init(&res->lr_childof);
 947
 948         ns->ns_resources--;
 949         if (ns->ns_resources == 0)
 950                 wake_up(&ns->ns_waitq);
 951 }
 952
 953 int ldlm_resource_putref_internal(struct ldlm_resource *res, int locked)
 954 {
 955         struct ldlm_namespace *ns = res->lr_namespace;
 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 (locked && !atomic_dec_and_test(&res->lr_refcount))
 966                 RETURN(0);
 967         if (!locked && !atomic_dec_and_lock(&res->lr_refcount,
 968                                             &ns->ns_hash_lock))
 969                 RETURN(0);
 970
 971         __ldlm_resource_putref_final(res);
 972
 973         if (!locked)
 974                 spin_unlock(&ns->ns_hash_lock);
 975
 976         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
 977                 ns->ns_lvbo->lvbo_free(res);
 978
 979         LASSERT(res->lr_lvb_inode == NULL);
 980
 981         OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 982
 983         RETURN(1);
 984 }
 985
 986 int ldlm_resource_putref(struct ldlm_resource *res)
 987 {
 988         return ldlm_resource_putref_internal(res, 0);
 989 }
 990
 991 int ldlm_resource_putref_locked(struct ldlm_resource *res)
 992 {
 993         return ldlm_resource_putref_internal(res, 1);
 994 }
 995
 996 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 997                             struct ldlm_lock *lock)
 998 {
 999         check_res_locked(res);
1000
1001         ldlm_resource_dump(D_INFO, res);
1002         CDEBUG(D_OTHER, "About to add this lock:\n");
1003         ldlm_lock_dump(D_OTHER, lock, 0);
1004
1005         if (lock->l_destroyed) {
1006                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1007                 return;
1008         }
1009
1010         LASSERT(list_empty(&lock->l_res_link));
1011
1012         list_add_tail(&lock->l_res_link, head);
1013 }
1014
1015 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1016                                      struct ldlm_lock *new)
1017 {
1018         struct ldlm_resource *res = original->l_resource;
1019
1020         check_res_locked(res);
1021
1022         ldlm_resource_dump(D_INFO, res);
1023         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1024         ldlm_lock_dump(D_OTHER, new, 0);
1025
1026         if (new->l_destroyed) {
1027                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1028                 goto out;
1029         }
1030
1031         LASSERT(list_empty(&new->l_res_link));
1032
1033         list_add(&new->l_res_link, &original->l_res_link);
1034  out:;
1035 }
1036
1037 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1038 {
1039         int type = lock->l_resource->lr_type;
1040
1041         check_res_locked(lock->l_resource);
1042         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1043                 ldlm_unlink_lock_skiplist(lock);
1044         else if (type == LDLM_EXTENT)
1045                 ldlm_extent_unlink_lock(lock);
1046         list_del_init(&lock->l_res_link);
1047 }
1048
1049 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1050 {
1051         desc->lr_type = res->lr_type;
1052         desc->lr_name = res->lr_name;
1053 }
1054
1055 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1056 {
1057         struct list_head *tmp;
1058
1059         if (!((libcfs_debug | D_ERROR) & level))
1060                 return;
1061
1062         mutex_down(ldlm_namespace_lock(client));
1063
1064         list_for_each(tmp, ldlm_namespace_list(client)) {
1065                 struct ldlm_namespace *ns;
1066                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1067                 ldlm_namespace_dump(level, ns);
1068         }
1069
1070         mutex_up(ldlm_namespace_lock(client));
1071 }
1072
1073 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1074 {
1075         struct list_head *tmp;
1076
1077         if (!((libcfs_debug | D_ERROR) & level))
1078                 return;
1079
1080         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1081                ns->ns_name, ns->ns_refcount,
1082                ns_is_client(ns) ? "client" : "server");
1083
1084         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1085                 return;
1086
1087         spin_lock(&ns->ns_hash_lock);
1088         tmp = ns->ns_root_list.next;
1089         while (tmp != &ns->ns_root_list) {
1090                 struct ldlm_resource *res;
1091                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
1092
1093                 ldlm_resource_getref(res);
1094                 spin_unlock(&ns->ns_hash_lock);
1095
1096                 lock_res(res);
1097                 ldlm_resource_dump(level, res);
1098                 unlock_res(res);
1099
1100                 spin_lock(&ns->ns_hash_lock);
1101                 tmp = tmp->next;
1102                 ldlm_resource_putref_locked(res);
1103         }
1104         ns->ns_next_dump = cfs_time_shift(10);
1105         spin_unlock(&ns->ns_hash_lock);
1106 }
1107
1108 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1109 {
1110         struct list_head *tmp;
1111         int pos;
1112
1113         CLASSERT(RES_NAME_SIZE == 4);
1114
1115         if (!((libcfs_debug | D_ERROR) & level))
1116                 return;
1117
1118         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1119                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1120                res->lr_name.name[2], res->lr_name.name[3],
1121                atomic_read(&res->lr_refcount));
1122
1123         if (!list_empty(&res->lr_granted)) {
1124                 pos = 0;
1125                 CDEBUG(level, "Granted locks:\n");
1126                 list_for_each(tmp, &res->lr_granted) {
1127                         struct ldlm_lock *lock;
1128                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1129                         ldlm_lock_dump(level, lock, ++pos);
1130                 }
1131         }
1132         if (!list_empty(&res->lr_converting)) {
1133                 pos = 0;
1134                 CDEBUG(level, "Converting locks:\n");
1135                 list_for_each(tmp, &res->lr_converting) {
1136                         struct ldlm_lock *lock;
1137                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1138                         ldlm_lock_dump(level, lock, ++pos);
1139                 }
1140         }
1141         if (!list_empty(&res->lr_waiting)) {
1142                 pos = 0;
1143                 CDEBUG(level, "Waiting locks:\n");
1144                 list_for_each(tmp, &res->lr_waiting) {
1145                         struct ldlm_lock *lock;
1146                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1147                         ldlm_lock_dump(level, lock, ++pos);
1148                 }
1149         }
1150 }