lustre/ldlm/ldlm_resource.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Phil Schwan <phil@clusterfs.com>
   6  *   Author: Peter Braam <braam@clusterfs.com>
   7  *
   8  *   This file is part of the Lustre file system, http://www.lustre.org
   9  *   Lustre is a trademark of Cluster File Systems, Inc.
  10  *
  11  *   You may have signed or agreed to another license before downloading
  12  *   this software.  If so, you are bound by the terms and conditions
  13  *   of that agreement, and the following does not apply to you.  See the
  14  *   LICENSE file included with this distribution for more information.
  15  *
  16  *   If you did not agree to a different license, then this copy of Lustre
  17  *   is open source software; you can redistribute it and/or modify it
  18  *   under the terms of version 2 of the GNU General Public License as
  19  *   published by the Free Software Foundation.
  20  *
  21  *   In either case, Lustre is distributed in the hope that it will be
  22  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  23  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  24  *   license text for more details.
  25  */
  26
  27 #define DEBUG_SUBSYSTEM S_LDLM
  28 #ifdef __KERNEL__
  29 # include <lustre_dlm.h>
  30 #else
  31 # include <liblustre.h>
  32 #endif
  33
  34 #include <obd_class.h>
  35 #include "ldlm_internal.h"
  36
  37 cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
  38
  39 atomic_t ldlm_srv_namespace_nr = ATOMIC_INIT(0);
  40 atomic_t ldlm_cli_namespace_nr = ATOMIC_INIT(0);
  41
  42 struct semaphore ldlm_srv_namespace_lock;
  43 struct list_head ldlm_srv_namespace_list =
  44         CFS_LIST_HEAD_INIT(ldlm_srv_namespace_list);
  45
  46 struct semaphore ldlm_cli_namespace_lock;
  47 struct list_head ldlm_cli_namespace_list =
  48         CFS_LIST_HEAD_INIT(ldlm_cli_namespace_list);
  49
  50 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
  51 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
  52 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
  53
  54 #ifdef LPROCFS
  55 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  56                              unsigned long count, void *data)
  57 {
  58         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  59         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  60         RETURN(count);
  61 }
  62
  63 int ldlm_proc_setup(void)
  64 {
  65         int rc;
  66         struct lprocfs_vars list[] = {
  67                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  68                 { NULL }};
  69         ENTRY;
  70         LASSERT(ldlm_ns_proc_dir == NULL);
  71
  72         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
  73                                               proc_lustre_root,
  74                                               NULL, NULL);
  75         if (IS_ERR(ldlm_type_proc_dir)) {
  76                 CERROR("LProcFS failed in ldlm-init\n");
  77                 rc = PTR_ERR(ldlm_type_proc_dir);
  78                 GOTO(err, rc);
  79         }
  80
  81         ldlm_ns_proc_dir = lprocfs_register("namespaces",
  82                                             ldlm_type_proc_dir,
  83                                             NULL, NULL);
  84         if (IS_ERR(ldlm_ns_proc_dir)) {
  85                 CERROR("LProcFS failed in ldlm-init\n");
  86                 rc = PTR_ERR(ldlm_ns_proc_dir);
  87                 GOTO(err_type, rc);
  88         }
  89
  90         ldlm_svc_proc_dir = lprocfs_register("services",
  91                                             ldlm_type_proc_dir,
  92                                             NULL, NULL);
  93         if (IS_ERR(ldlm_svc_proc_dir)) {
  94                 CERROR("LProcFS failed in ldlm-init\n");
  95                 rc = PTR_ERR(ldlm_svc_proc_dir);
  96                 GOTO(err_ns, rc);
  97         }
  98
  99         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 100
 101         RETURN(0);
 102
 103 err_ns:
 104         lprocfs_remove(&ldlm_ns_proc_dir);
 105 err_type:
 106         lprocfs_remove(&ldlm_type_proc_dir);
 107 err:
 108         ldlm_svc_proc_dir = NULL;
 109         RETURN(rc);
 110 }
 111
 112 void ldlm_proc_cleanup(void)
 113 {
 114         if (ldlm_svc_proc_dir)
 115                 lprocfs_remove(&ldlm_svc_proc_dir);
 116
 117         if (ldlm_ns_proc_dir)
 118                 lprocfs_remove(&ldlm_ns_proc_dir);
 119
 120         if (ldlm_type_proc_dir)
 121                 lprocfs_remove(&ldlm_type_proc_dir);
 122 }
 123
 124 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 125                                int count, int *eof, void *data)
 126 {
 127         struct ldlm_namespace *ns = data;
 128         __u32 *nr = &ns->ns_max_unused;
 129
 130         if (ns_connect_lru_resize(ns))
 131                 nr = &ns->ns_nr_unused;
 132         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 133 }
 134
 135 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 136                                unsigned long count, void *data)
 137 {
 138         struct ldlm_namespace *ns = data;
 139         char dummy[MAX_STRING_SIZE + 1], *end;
 140         unsigned long tmp;
 141         int lru_resize;
 142
 143         dummy[MAX_STRING_SIZE] = '\0';
 144         if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 145                 return -EFAULT;
 146
 147         if (count == 6 && memcmp(dummy, "clear", 5) == 0) {
 148                 CDEBUG(D_DLMTRACE,
 149                        "dropping all unused locks from namespace %s\n",
 150                        ns->ns_name);
 151                 if (ns_connect_lru_resize(ns)) {
 152                         int canceled, unused  = ns->ns_nr_unused;
 153
 154                         /* Try to cancel all @ns_nr_unused locks. */
 155                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
 156                                                    LDLM_CANCEL_PASSED);
 157                         if (canceled < unused) {
 158                                 CERROR("not all requested locks are canceled, "
 159                                        "requested: %d, canceled: %d\n", unused,
 160                                        canceled);
 161                                 return -EINVAL;
 162                         }
 163                 } else {
 164                         tmp = ns->ns_max_unused;
 165                         ns->ns_max_unused = 0;
 166                         ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
 167                         ns->ns_max_unused = tmp;
 168                 }
 169                 return count;
 170         }
 171
 172         tmp = simple_strtoul(dummy, &end, 0);
 173         if (dummy == end) {
 174                 CERROR("invalid value written\n");
 175                 return -EINVAL;
 176         }
 177         lru_resize = (tmp == 0);
 178
 179         if (ns_connect_lru_resize(ns)) {
 180                 if (!lru_resize)
 181                         ns->ns_max_unused = (unsigned int)tmp;
 182
 183                 if (tmp > ns->ns_nr_unused)
 184                         tmp = ns->ns_nr_unused;
 185                 tmp = ns->ns_nr_unused - tmp;
 186
 187                 CDEBUG(D_DLMTRACE, "changing namespace %s unused locks from %u to %u\n",
 188                        ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
 189                 ldlm_cancel_lru(ns, (unsigned int)tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 190
 191                 if (!lru_resize) {
 192                         CDEBUG(D_DLMTRACE, "disable lru_resize for namespace %s\n",
 193                                ns->ns_name);
 194                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 195                 }
 196         } else {
 197                 CDEBUG(D_DLMTRACE, "changing namespace %s max_unused from %u to %u\n",
 198                        ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
 199                 ns->ns_max_unused = (unsigned int)tmp;
 200                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 201
 202                 /* Make sure that originally lru resize was supported before
 203                  * turning it on here. */
 204                 if (lru_resize &&
 205                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 206                         CDEBUG(D_DLMTRACE, "enable lru_resize for namespace %s\n",
 207                                ns->ns_name);
 208                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 209                 }
 210         }
 211         return count;
 212 }
 213
 214 void ldlm_proc_namespace(struct ldlm_namespace *ns)
 215 {
 216         struct lprocfs_vars lock_vars[2];
 217         char lock_name[MAX_STRING_SIZE + 1];
 218
 219         LASSERT(ns != NULL);
 220         LASSERT(ns->ns_name != NULL);
 221
 222         lock_name[MAX_STRING_SIZE] = '\0';
 223
 224         memset(lock_vars, 0, sizeof(lock_vars));
 225         lock_vars[0].name = lock_name;
 226
 227         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
 228         lock_vars[0].data = &ns->ns_refcount;
 229         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 230         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 231
 232         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
 233         lock_vars[0].data = &ns->ns_locks;
 234         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 235         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 236
 237         if (ns_is_client(ns)) {
 238                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 239                          ns->ns_name);
 240                 lock_vars[0].data = &ns->ns_nr_unused;
 241                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 242                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 243
 244                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 245                          ns->ns_name);
 246                 lock_vars[0].data = ns;
 247                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 248                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 249                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 250
 251                 snprintf(lock_name, MAX_STRING_SIZE, "%s/shrink_thumb",
 252                          ns->ns_name);
 253                 lock_vars[0].data = ns;
 254                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 255                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 256                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 257
 258                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 259                          ns->ns_name);
 260                 lock_vars[0].data = &ns->ns_max_age;
 261                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 262                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 263                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 264         } else {
 265                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 266                          ns->ns_name);
 267                 lock_vars[0].data = &ns->ns_max_nolock_size;
 268                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 269                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 270                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 271
 272                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 273                          ns->ns_name);
 274                 lock_vars[0].data = &ns->ns_contention_time;
 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
 279                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 280                          ns->ns_name);
 281                 lock_vars[0].data = &ns->ns_contended_locks;
 282                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 283                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 284                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 285         }
 286 }
 287 #undef MAX_STRING_SIZE
 288 #else
 289 #define ldlm_proc_namespace(ns) do {} while (0)
 290 #endif /* LPROCFS */
 291
 292 struct ldlm_namespace *
 293 ldlm_namespace_new(struct obd_device *obd, char *name,
 294                    ldlm_side_t client, ldlm_appetite_t apt)
 295 {
 296         struct ldlm_namespace *ns = NULL;
 297         struct list_head *bucket;
 298         int rc, idx, namelen;
 299         ENTRY;
 300
 301         rc = ldlm_get_ref();
 302         if (rc) {
 303                 CERROR("ldlm_get_ref failed: %d\n", rc);
 304                 RETURN(NULL);
 305         }
 306
 307         OBD_ALLOC_PTR(ns);
 308         if (!ns)
 309                 GOTO(out_ref, NULL);
 310
 311         OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 312         if (!ns->ns_hash)
 313                 GOTO(out_ns, NULL);
 314
 315         namelen = strlen(name);
 316         OBD_ALLOC(ns->ns_name, namelen + 1);
 317         if (!ns->ns_name)
 318                 GOTO(out_hash, NULL);
 319
 320         ns->ns_shrink_thumb = LDLM_LOCK_SHRINK_THUMB;
 321         ns->ns_appetite = apt;
 322
 323         LASSERT(obd != NULL);
 324         ns->ns_obd = obd;
 325
 326         strcpy(ns->ns_name, name);
 327
 328         CFS_INIT_LIST_HEAD(&ns->ns_root_list);
 329         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 330         ns->ns_refcount = 0;
 331         ns->ns_client = client;
 332         spin_lock_init(&ns->ns_hash_lock);
 333         atomic_set(&ns->ns_locks, 0);
 334         ns->ns_resources = 0;
 335         cfs_waitq_init(&ns->ns_waitq);
 336         ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
 337         ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
 338         ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
 339
 340         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 341              bucket--)
 342                 CFS_INIT_LIST_HEAD(bucket);
 343
 344         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 345         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 346         ns->ns_nr_unused = 0;
 347         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
 348         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
 349         spin_lock_init(&ns->ns_unused_lock);
 350         ns->ns_orig_connect_flags = 0;
 351         ns->ns_connect_flags = 0;
 352
 353         ldlm_proc_namespace(ns);
 354
 355         idx = atomic_read(ldlm_namespace_nr(client));
 356
 357         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 358         if (rc) {
 359                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 360                 GOTO(out_proc, rc);
 361         }
 362
 363         at_init(&ns->ns_at_estimate, ldlm_enqueue_min, 0);
 364
 365         ldlm_namespace_register(ns, client);
 366         RETURN(ns);
 367 out_proc:
 368         ldlm_namespace_cleanup(ns, 0);
 369         OBD_FREE(ns->ns_name, namelen + 1);
 370 out_hash:
 371         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 372 out_ns:
 373         OBD_FREE_PTR(ns);
 374 out_ref:
 375         ldlm_put_ref();
 376         RETURN(NULL);
 377 }
 378
 379 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 380
 381 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 382  * This is currently only used for recovery, and we make certain assumptions
 383  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 384  *
 385  * Called with the ns_lock held. */
 386 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 387                              int flags)
 388 {
 389         struct list_head *tmp;
 390         int rc = 0, client = ns_is_client(res->lr_namespace);
 391         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 392         ENTRY;
 393
 394
 395         do {
 396                 struct ldlm_lock *lock = NULL;
 397
 398                 /* first, we look for non-cleaned-yet lock
 399                  * all cleaned locks are marked by CLEANED flag */
 400                 lock_res(res);
 401                 list_for_each(tmp, q) {
 402                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 403                         if (lock->l_flags & LDLM_FL_CLEANED) {
 404                                 lock = NULL;
 405                                 continue;
 406                         }
 407                         LDLM_LOCK_GET(lock);
 408                         lock->l_flags |= LDLM_FL_CLEANED;
 409                         break;
 410                 }
 411
 412                 if (lock == NULL) {
 413                         unlock_res(res);
 414                         break;
 415                 }
 416
 417                 /* Set CBPENDING so nothing in the cancellation path
 418                  * can match this lock */
 419                 lock->l_flags |= LDLM_FL_CBPENDING;
 420                 lock->l_flags |= LDLM_FL_FAILED;
 421                 lock->l_flags |= flags;
 422
 423                 /* ... without sending a CANCEL message for local_only. */
 424                 if (local_only)
 425                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 426
 427                 if (local_only && (lock->l_readers || lock->l_writers)) {
 428                         /* This is a little bit gross, but much better than the
 429                          * alternative: pretend that we got a blocking AST from
 430                          * the server, so that when the lock is decref'd, it
 431                          * will go away ... */
 432                         unlock_res(res);
 433                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 434                         if (lock->l_completion_ast)
 435                                 lock->l_completion_ast(lock, 0, NULL);
 436                         LDLM_LOCK_PUT(lock);
 437                         continue;
 438                 }
 439
 440                 if (client) {
 441                         struct lustre_handle lockh;
 442
 443                         unlock_res(res);
 444                         ldlm_lock2handle(lock, &lockh);
 445                         rc = ldlm_cli_cancel(&lockh);
 446                         if (rc)
 447                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 448                 } else {
 449                         ldlm_resource_unlink_lock(lock);
 450                         unlock_res(res);
 451                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 452                                    "client node");
 453                         ldlm_lock_destroy(lock);
 454                 }
 455                 LDLM_LOCK_PUT(lock);
 456         } while (1);
 457
 458         EXIT;
 459 }
 460
 461 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 462 {
 463         struct list_head *tmp;
 464         int i;
 465
 466         if (ns == NULL) {
 467                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 468                 return ELDLM_OK;
 469         }
 470
 471         for (i = 0; i < RES_HASH_SIZE; i++) {
 472                 spin_lock(&ns->ns_hash_lock);
 473                 tmp = ns->ns_hash[i].next;
 474                 while (tmp != &(ns->ns_hash[i])) {
 475                         struct ldlm_resource *res;
 476                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 477                         ldlm_resource_getref(res);
 478                         spin_unlock(&ns->ns_hash_lock);
 479
 480                         cleanup_resource(res, &res->lr_granted, flags);
 481                         cleanup_resource(res, &res->lr_converting, flags);
 482                         cleanup_resource(res, &res->lr_waiting, flags);
 483
 484                         spin_lock(&ns->ns_hash_lock);
 485                         tmp  = tmp->next;
 486
 487                         /* XXX: former stuff caused issues in case of race
 488                          * between ldlm_namespace_cleanup() and lockd() when
 489                          * client gets blocking ast when lock gets distracted by
 490                          * server. This is 1_4 branch solution, let's see how
 491                          * will it behave. */
 492                         if (!ldlm_resource_putref_locked(res))
 493                                 CDEBUG(D_INFO,
 494                                        "Namespace %s resource refcount nonzero "
 495                                        "(%d) after lock cleanup; forcing cleanup.\n",
 496                                        ns->ns_name, atomic_read(&res->lr_refcount));
 497                 }
 498                 spin_unlock(&ns->ns_hash_lock);
 499         }
 500
 501         return ELDLM_OK;
 502 }
 503
 504 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 505 {
 506         ENTRY;
 507         /* At shutdown time, don't call the cancellation callback */
 508         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 509
 510         if (ns->ns_refcount > 0) {
 511                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 512                 int rc;
 513                 CDEBUG(D_DLMTRACE,
 514                        "dlm namespace %s free waiting on refcount %d\n",
 515                        ns->ns_name, ns->ns_refcount);
 516 force_wait:
 517                 if (force)
 518                         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 519
 520                 rc = l_wait_event(ns->ns_waitq,
 521                                   ns->ns_refcount == 0, &lwi);
 522
 523                 /* Forced cleanups should be able to reclaim all references,
 524                  * so it's safe to wait forever... we can't leak locks... */
 525                 if (force && rc == -ETIMEDOUT) {
 526                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 527                                        "namespace with %d resources in use, "
 528                                        "(rc=%d)\n", ns->ns_name,
 529                                        ns->ns_refcount, rc);
 530                         GOTO(force_wait, rc);
 531                 }
 532
 533                 if (ns->ns_refcount) {
 534                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 535                                        "with %d resources in use, (rc=%d)\n",
 536                                        ns->ns_name,
 537                                        ns->ns_refcount, rc);
 538                         RETURN(ELDLM_NAMESPACE_EXISTS);
 539                 }
 540                 CDEBUG(D_DLMTRACE,
 541                        "dlm namespace %s free done waiting\n", ns->ns_name);
 542         }
 543
 544         RETURN(ELDLM_OK);
 545 }
 546
 547 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 548                                struct obd_import *imp,
 549                                int force)
 550 {
 551         int rc;
 552         ENTRY;
 553         if (!ns) {
 554                 EXIT;
 555                 return;
 556         }
 557
 558         /* Make sure that nobody can find this ns in its list. */
 559         ldlm_namespace_unregister(ns, ns->ns_client);
 560
 561         /* Can fail with -EINTR when force == 0 in which case try harder */
 562         rc = __ldlm_namespace_free(ns, force);
 563         if (rc != ELDLM_OK) {
 564                 if (imp) {
 565                         ptlrpc_disconnect_import(imp, 0);
 566                         ptlrpc_invalidate_import(imp);
 567                 }
 568
 569                 /* With all requests dropped and the import inactive
 570                  * we are gaurenteed all reference will be dropped. */
 571                 rc = __ldlm_namespace_free(ns, 1);
 572                 LASSERT(rc == 0);
 573         }
 574         EXIT;
 575 }
 576
 577 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 578 {
 579         ENTRY;
 580         if (!ns) {
 581                 EXIT;
 582                 return;
 583         }
 584
 585         /* Fini pool _before_ parent proc dir is removed. This is important as
 586          * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
 587          * it after @dir may cause oops. */
 588         ldlm_pool_fini(&ns->ns_pool);
 589
 590 #ifdef LPROCFS
 591         {
 592                 struct proc_dir_entry *dir;
 593                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
 594                 if (dir == NULL) {
 595                         CERROR("dlm namespace %s has no procfs dir?\n",
 596                                ns->ns_name);
 597                 } else {
 598                         lprocfs_remove(&dir);
 599                 }
 600         }
 601 #endif
 602         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 603         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 604
 605         /* @ns should be not on list in this time, otherwise this will cause
 606          * issues realted to using freed @ns in pools thread. */
 607         LASSERT(list_empty(&ns->ns_list_chain));
 608         OBD_FREE_PTR(ns);
 609         ldlm_put_ref();
 610         EXIT;
 611 }
 612
 613 /* Cleanup the resource, and free namespace.
 614  * bug 12864:
 615  * Deadlock issue:
 616  * proc1: destroy import
 617  *        class_disconnect_export(grab cl_sem) ->
 618  *              -> ldlm_namespace_free ->
 619  *              -> lprocfs_remove(grab _lprocfs_lock).
 620  * proc2: read proc info
 621  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 622  *              -> osc_rd_active, etc(grab cl_sem).
 623  *
 624  * So that I have to split the ldlm_namespace_free into two parts - the first
 625  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 626  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 627  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 628  * held.
 629  */
 630 void ldlm_namespace_free(struct ldlm_namespace *ns,
 631                          struct obd_import *imp,
 632                          int force)
 633 {
 634         ldlm_namespace_free_prior(ns, imp, force);
 635         ldlm_namespace_free_post(ns);
 636 }
 637
 638 void ldlm_namespace_get_locked(struct ldlm_namespace *ns)
 639 {
 640         LASSERT(ns->ns_refcount >= 0);
 641         ns->ns_refcount++;
 642 }
 643
 644 void ldlm_namespace_get(struct ldlm_namespace *ns)
 645 {
 646         spin_lock(&ns->ns_hash_lock);
 647         ldlm_namespace_get_locked(ns);
 648         spin_unlock(&ns->ns_hash_lock);
 649 }
 650
 651 void ldlm_namespace_put_locked(struct ldlm_namespace *ns, int wakeup)
 652 {
 653         LASSERT(ns->ns_refcount > 0);
 654         ns->ns_refcount--;
 655         if (ns->ns_refcount == 0 && wakeup)
 656                 wake_up(&ns->ns_waitq);
 657 }
 658
 659 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
 660 {
 661         spin_lock(&ns->ns_hash_lock);
 662         ldlm_namespace_put_locked(ns, wakeup);
 663         spin_unlock(&ns->ns_hash_lock);
 664 }
 665
 666 /* Register @ns in the list of namespaces */
 667 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
 668 {
 669         mutex_down(ldlm_namespace_lock(client));
 670         LASSERT(list_empty(&ns->ns_list_chain));
 671         list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 672         atomic_inc(ldlm_namespace_nr(client));
 673         mutex_up(ldlm_namespace_lock(client));
 674 }
 675
 676 /* Unregister @ns from the list of namespaces */
 677 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
 678 {
 679         mutex_down(ldlm_namespace_lock(client));
 680         LASSERT(!list_empty(&ns->ns_list_chain));
 681         /*
 682          * Some asserts and possibly other parts of code still using
 683          * list_empty(&ns->ns_list_chain). This is why it is important
 684          * to use list_del_init() here.
 685          */
 686         list_del_init(&ns->ns_list_chain);
 687         atomic_dec(ldlm_namespace_nr(client));
 688         mutex_up(ldlm_namespace_lock(client));
 689 }
 690
 691 /* Should be called under ldlm_namespace_lock(client) taken */
 692 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
 693 {
 694         LASSERT(!list_empty(&ns->ns_list_chain));
 695         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 696         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 697 }
 698
 699 /* Should be called under ldlm_namespace_lock(client) taken */
 700 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
 701 {
 702         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 703         LASSERT(!list_empty(ldlm_namespace_list(client)));
 704         return container_of(ldlm_namespace_list(client)->next,
 705                 struct ldlm_namespace, ns_list_chain);
 706 }
 707
 708 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, struct ldlm_res_id name)
 709 {
 710         __u32 hash = 0;
 711         int i;
 712
 713         for (i = 0; i < RES_NAME_SIZE; i++)
 714                 hash += name.name[i];
 715
 716         hash += (__u32)((unsigned long)parent >> 4);
 717
 718         return (hash & RES_HASH_MASK);
 719 }
 720
 721 static struct ldlm_resource *ldlm_resource_new(void)
 722 {
 723         struct ldlm_resource *res;
 724         int idx;
 725
 726         OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
 727         if (res == NULL)
 728                 return NULL;
 729
 730         memset(res, 0, sizeof(*res));
 731
 732         CFS_INIT_LIST_HEAD(&res->lr_children);
 733         CFS_INIT_LIST_HEAD(&res->lr_childof);
 734         CFS_INIT_LIST_HEAD(&res->lr_granted);
 735         CFS_INIT_LIST_HEAD(&res->lr_converting);
 736         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 737
 738         /* initialize interval trees for each lock mode*/
 739         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 740                 res->lr_itree[idx].lit_size = 0;
 741                 res->lr_itree[idx].lit_mode = 1 << idx;
 742                 res->lr_itree[idx].lit_root = NULL;
 743         }
 744
 745         atomic_set(&res->lr_refcount, 1);
 746         spin_lock_init(&res->lr_lock);
 747
 748         /* one who creates the resource must unlock
 749          * the semaphore after lvb initialization */
 750         init_MUTEX_LOCKED(&res->lr_lvb_sem);
 751
 752         return res;
 753 }
 754
 755 /* must be called with hash lock held */
 756 static struct ldlm_resource *
 757 ldlm_resource_find(struct ldlm_namespace *ns, struct ldlm_res_id name, __u32 hash)
 758 {
 759         struct list_head *bucket, *tmp;
 760         struct ldlm_resource *res;
 761
 762         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 763         bucket = ns->ns_hash + hash;
 764
 765         list_for_each(tmp, bucket) {
 766                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 767                 if (memcmp(&res->lr_name, &name, sizeof(res->lr_name)) == 0)
 768                         return res;
 769         }
 770
 771         return NULL;
 772 }
 773
 774 /* Args: locked namespace
 775  * Returns: newly-allocated, referenced, unlocked resource */
 776 static struct ldlm_resource *
 777 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 778                   struct ldlm_res_id name, __u32 hash, ldlm_type_t type)
 779 {
 780         struct list_head *bucket;
 781         struct ldlm_resource *res, *old_res;
 782         ENTRY;
 783
 784         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
 785                  "type: %d\n", type);
 786
 787         res = ldlm_resource_new();
 788         if (!res)
 789                 RETURN(NULL);
 790
 791         res->lr_name = name;
 792         res->lr_namespace = ns;
 793         res->lr_type = type;
 794         res->lr_most_restr = LCK_NL;
 795
 796         spin_lock(&ns->ns_hash_lock);
 797         old_res = ldlm_resource_find(ns, name, hash);
 798         if (old_res) {
 799                 /* someone won the race and added the resource before */
 800                 ldlm_resource_getref(old_res);
 801                 spin_unlock(&ns->ns_hash_lock);
 802                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 803                 /* synchronize WRT resource creation */
 804                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 805                         down(&old_res->lr_lvb_sem);
 806                         up(&old_res->lr_lvb_sem);
 807                 }
 808                 RETURN(old_res);
 809         }
 810
 811         /* we won! let's add the resource */
 812         bucket = ns->ns_hash + hash;
 813         list_add(&res->lr_hash, bucket);
 814         ns->ns_resources++;
 815         ldlm_namespace_get_locked(ns);
 816
 817         if (parent == NULL) {
 818                 list_add(&res->lr_childof, &ns->ns_root_list);
 819         } else {
 820                 res->lr_parent = parent;
 821                 list_add(&res->lr_childof, &parent->lr_children);
 822         }
 823         spin_unlock(&ns->ns_hash_lock);
 824
 825         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 826                 int rc;
 827
 828                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
 829                 rc = ns->ns_lvbo->lvbo_init(res);
 830                 if (rc)
 831                         CERROR("lvbo_init failed for resource "
 832                                LPU64": rc %d\n", name.name[0], rc);
 833                 /* we create resource with locked lr_lvb_sem */
 834                 up(&res->lr_lvb_sem);
 835         }
 836
 837         RETURN(res);
 838 }
 839
 840 /* Args: unlocked namespace
 841  * Locks: takes and releases ns->ns_lock and res->lr_lock
 842  * Returns: referenced, unlocked ldlm_resource or NULL */
 843 struct ldlm_resource *
 844 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 845                   struct ldlm_res_id name, ldlm_type_t type, int create)
 846 {
 847         __u32 hash = ldlm_hash_fn(parent, name);
 848         struct ldlm_resource *res = NULL;
 849         ENTRY;
 850
 851         LASSERT(ns != NULL);
 852         LASSERT(ns->ns_hash != NULL);
 853         LASSERT(name.name[0] != 0);
 854
 855         spin_lock(&ns->ns_hash_lock);
 856         res = ldlm_resource_find(ns, name, hash);
 857         if (res) {
 858                 ldlm_resource_getref(res);
 859                 spin_unlock(&ns->ns_hash_lock);
 860                 /* synchronize WRT resource creation */
 861                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 862                         down(&res->lr_lvb_sem);
 863                         up(&res->lr_lvb_sem);
 864                 }
 865                 RETURN(res);
 866         }
 867         spin_unlock(&ns->ns_hash_lock);
 868
 869         if (create == 0)
 870                 RETURN(NULL);
 871
 872         res = ldlm_resource_add(ns, parent, name, hash, type);
 873         RETURN(res);
 874 }
 875
 876 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 877 {
 878         LASSERT(res != NULL);
 879         LASSERT(res != LP_POISON);
 880         atomic_inc(&res->lr_refcount);
 881         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 882                atomic_read(&res->lr_refcount));
 883         return res;
 884 }
 885
 886 void __ldlm_resource_putref_final(struct ldlm_resource *res)
 887 {
 888         struct ldlm_namespace *ns = res->lr_namespace;
 889
 890         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 891
 892         if (!list_empty(&res->lr_granted)) {
 893                 ldlm_resource_dump(D_ERROR, res);
 894                 LBUG();
 895         }
 896
 897         if (!list_empty(&res->lr_converting)) {
 898                 ldlm_resource_dump(D_ERROR, res);
 899                 LBUG();
 900         }
 901
 902         if (!list_empty(&res->lr_waiting)) {
 903                 ldlm_resource_dump(D_ERROR, res);
 904                 LBUG();
 905         }
 906
 907         if (!list_empty(&res->lr_children)) {
 908                 ldlm_resource_dump(D_ERROR, res);
 909                 LBUG();
 910         }
 911
 912         /* Pass 0 as second argument to not wake up ->ns_waitq yet, will do it
 913          * later. */
 914         ldlm_namespace_put_locked(ns, 0);
 915         list_del_init(&res->lr_hash);
 916         list_del_init(&res->lr_childof);
 917
 918         ns->ns_resources--;
 919         if (ns->ns_resources == 0)
 920                 wake_up(&ns->ns_waitq);
 921 }
 922
 923 /* Returns 1 if the resource was freed, 0 if it remains. */
 924 int ldlm_resource_putref(struct ldlm_resource *res)
 925 {
 926         struct ldlm_namespace *ns = res->lr_namespace;
 927         int rc = 0;
 928         ENTRY;
 929
 930         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 931                atomic_read(&res->lr_refcount) - 1);
 932         LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
 933                  atomic_read(&res->lr_refcount));
 934         LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
 935                  atomic_read(&res->lr_refcount));
 936
 937         if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
 938                 __ldlm_resource_putref_final(res);
 939                 spin_unlock(&ns->ns_hash_lock);
 940                 if (res->lr_lvb_data)
 941                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 942                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 943                 rc = 1;
 944         }
 945
 946         RETURN(rc);
 947 }
 948
 949 /* Returns 1 if the resource was freed, 0 if it remains. */
 950 int ldlm_resource_putref_locked(struct ldlm_resource *res)
 951 {
 952         int rc = 0;
 953         ENTRY;
 954
 955         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 956                atomic_read(&res->lr_refcount) - 1);
 957         LASSERT(atomic_read(&res->lr_refcount) > 0);
 958         LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
 959
 960         LASSERT(atomic_read(&res->lr_refcount) >= 0);
 961         if (atomic_dec_and_test(&res->lr_refcount)) {
 962                 __ldlm_resource_putref_final(res);
 963                 if (res->lr_lvb_data)
 964                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 965                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 966                 rc = 1;
 967         }
 968
 969         RETURN(rc);
 970 }
 971
 972 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 973                             struct ldlm_lock *lock)
 974 {
 975         check_res_locked(res);
 976
 977         ldlm_resource_dump(D_OTHER, res);
 978         CDEBUG(D_OTHER, "About to add this lock:\n");
 979         ldlm_lock_dump(D_OTHER, lock, 0);
 980
 981         if (lock->l_destroyed) {
 982                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 983                 return;
 984         }
 985
 986         LASSERT(list_empty(&lock->l_res_link));
 987
 988         list_add_tail(&lock->l_res_link, head);
 989 }
 990
 991 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
 992                                      struct ldlm_lock *new)
 993 {
 994         struct ldlm_resource *res = original->l_resource;
 995
 996         check_res_locked(res);
 997
 998         ldlm_resource_dump(D_OTHER, res);
 999         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1000         ldlm_lock_dump(D_OTHER, new, 0);
1001
1002         if (new->l_destroyed) {
1003                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1004                 goto out;
1005         }
1006
1007         LASSERT(list_empty(&new->l_res_link));
1008
1009         list_add(&new->l_res_link, &original->l_res_link);
1010  out:;
1011 }
1012
1013 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1014 {
1015         int type = lock->l_resource->lr_type;
1016
1017         check_res_locked(lock->l_resource);
1018         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1019                 ldlm_unlink_lock_skiplist(lock);
1020         else if (type == LDLM_EXTENT)
1021                 ldlm_extent_unlink_lock(lock);
1022         list_del_init(&lock->l_res_link);
1023 }
1024
1025 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1026 {
1027         desc->lr_type = res->lr_type;
1028         desc->lr_name = res->lr_name;
1029 }
1030
1031 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1032 {
1033         struct list_head *tmp;
1034
1035         if (!((libcfs_debug | D_ERROR) & level))
1036                 return;
1037
1038         mutex_down(ldlm_namespace_lock(client));
1039
1040         list_for_each(tmp, ldlm_namespace_list(client)) {
1041                 struct ldlm_namespace *ns;
1042                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1043                 ldlm_namespace_dump(level, ns);
1044         }
1045
1046         mutex_up(ldlm_namespace_lock(client));
1047 }
1048
1049 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1050 {
1051         struct list_head *tmp;
1052
1053         if (!((libcfs_debug | D_ERROR) & level))
1054                 return;
1055
1056         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1057                ns->ns_name, ns->ns_refcount,
1058                ns_is_client(ns) ? "client" : "server");
1059
1060         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1061                 return;
1062
1063         spin_lock(&ns->ns_hash_lock);
1064         tmp = ns->ns_root_list.next;
1065         while (tmp != &ns->ns_root_list) {
1066                 struct ldlm_resource *res;
1067                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
1068
1069                 ldlm_resource_getref(res);
1070                 spin_unlock(&ns->ns_hash_lock);
1071
1072                 lock_res(res);
1073                 ldlm_resource_dump(level, res);
1074                 unlock_res(res);
1075
1076                 spin_lock(&ns->ns_hash_lock);
1077                 tmp = tmp->next;
1078                 ldlm_resource_putref_locked(res);
1079         }
1080         ns->ns_next_dump = cfs_time_shift(10);
1081         spin_unlock(&ns->ns_hash_lock);
1082 }
1083
1084 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1085 {
1086         struct list_head *tmp;
1087         int pos;
1088
1089         CLASSERT(RES_NAME_SIZE == 4);
1090
1091         if (!((libcfs_debug | D_ERROR) & level))
1092                 return;
1093
1094         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1095                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1096                res->lr_name.name[2], res->lr_name.name[3],
1097                atomic_read(&res->lr_refcount));
1098
1099         if (!list_empty(&res->lr_granted)) {
1100                 pos = 0;
1101                 CDEBUG(level, "Granted locks:\n");
1102                 list_for_each(tmp, &res->lr_granted) {
1103                         struct ldlm_lock *lock;
1104                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1105                         ldlm_lock_dump(level, lock, ++pos);
1106                 }
1107         }
1108         if (!list_empty(&res->lr_converting)) {
1109                 pos = 0;
1110                 CDEBUG(level, "Converting locks:\n");
1111                 list_for_each(tmp, &res->lr_converting) {
1112                         struct ldlm_lock *lock;
1113                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1114                         ldlm_lock_dump(level, lock, ++pos);
1115                 }
1116         }
1117         if (!list_empty(&res->lr_waiting)) {
1118                 pos = 0;
1119                 CDEBUG(level, "Waiting locks:\n");
1120                 list_for_each(tmp, &res->lr_waiting) {
1121                         struct ldlm_lock *lock;
1122                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1123                         ldlm_lock_dump(level, lock, ++pos);
1124                 }
1125         }
1126 }