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