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 CFS_LIST_HEAD(ldlm_srv_namespace_list);
  44
  45 struct semaphore ldlm_cli_namespace_lock;
  46 CFS_LIST_HEAD(ldlm_cli_namespace_list);
  47
  48 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
  49 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
  50 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
  51
  52 #ifdef LPROCFS
  53 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  54                              unsigned long count, void *data)
  55 {
  56         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  57         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  58         RETURN(count);
  59 }
  60
  61 int ldlm_proc_setup(void)
  62 {
  63         int rc;
  64         struct lprocfs_vars list[] = {
  65                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  66                 { NULL }};
  67         ENTRY;
  68         LASSERT(ldlm_ns_proc_dir == NULL);
  69
  70         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
  71                                               proc_lustre_root,
  72                                               NULL, NULL);
  73         if (IS_ERR(ldlm_type_proc_dir)) {
  74                 CERROR("LProcFS failed in ldlm-init\n");
  75                 rc = PTR_ERR(ldlm_type_proc_dir);
  76                 GOTO(err, rc);
  77         }
  78
  79         ldlm_ns_proc_dir = lprocfs_register("namespaces",
  80                                             ldlm_type_proc_dir,
  81                                             NULL, NULL);
  82         if (IS_ERR(ldlm_ns_proc_dir)) {
  83                 CERROR("LProcFS failed in ldlm-init\n");
  84                 rc = PTR_ERR(ldlm_ns_proc_dir);
  85                 GOTO(err_type, rc);
  86         }
  87
  88         ldlm_svc_proc_dir = lprocfs_register("services",
  89                                             ldlm_type_proc_dir,
  90                                             NULL, NULL);
  91         if (IS_ERR(ldlm_svc_proc_dir)) {
  92                 CERROR("LProcFS failed in ldlm-init\n");
  93                 rc = PTR_ERR(ldlm_svc_proc_dir);
  94                 GOTO(err_ns, rc);
  95         }
  96
  97         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
  98
  99         RETURN(0);
 100
 101 err_ns:
 102         lprocfs_remove(&ldlm_ns_proc_dir);
 103 err_type:
 104         lprocfs_remove(&ldlm_type_proc_dir);
 105 err:
 106         ldlm_svc_proc_dir = NULL;
 107         RETURN(rc);
 108 }
 109
 110 void ldlm_proc_cleanup(void)
 111 {
 112         if (ldlm_svc_proc_dir)
 113                 lprocfs_remove(&ldlm_svc_proc_dir);
 114
 115         if (ldlm_ns_proc_dir)
 116                 lprocfs_remove(&ldlm_ns_proc_dir);
 117
 118         if (ldlm_type_proc_dir)
 119                 lprocfs_remove(&ldlm_type_proc_dir);
 120 }
 121
 122 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 123                                int count, int *eof, void *data)
 124 {
 125         struct ldlm_namespace *ns = data;
 126         __u32 *nr = &ns->ns_max_unused;
 127
 128         if (ns_connect_lru_resize(ns))
 129                 nr = &ns->ns_nr_unused;
 130         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 131 }
 132
 133 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 134                                unsigned long count, void *data)
 135 {
 136         struct ldlm_namespace *ns = data;
 137         char dummy[MAX_STRING_SIZE + 1], *end;
 138         unsigned long tmp;
 139         int lru_resize;
 140
 141         dummy[MAX_STRING_SIZE] = '\0';
 142         if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 143                 return -EFAULT;
 144
 145         if (count == 6 && memcmp(dummy, "clear", 5) == 0) {
 146                 CDEBUG(D_DLMTRACE,
 147                        "dropping all unused locks from namespace %s\n",
 148                        ns->ns_name);
 149                 if (ns_connect_lru_resize(ns)) {
 150                         int canceled, unused  = ns->ns_nr_unused;
 151
 152                         /* Try to cancel all @ns_nr_unused locks. */
 153                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
 154                                                    LDLM_CANCEL_PASSED);
 155                         if (canceled < unused) {
 156                                 CERROR("not all requested locks are canceled, "
 157                                        "requested: %d, canceled: %d\n", unused,
 158                                        canceled);
 159                                 return -EINVAL;
 160                         }
 161                 } else {
 162                         tmp = ns->ns_max_unused;
 163                         ns->ns_max_unused = 0;
 164                         ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
 165                         ns->ns_max_unused = tmp;
 166                 }
 167                 return count;
 168         }
 169
 170         tmp = simple_strtoul(dummy, &end, 0);
 171         if (dummy == end) {
 172                 CERROR("invalid value written\n");
 173                 return -EINVAL;
 174         }
 175         lru_resize = (tmp == 0);
 176
 177         if (ns_connect_lru_resize(ns)) {
 178                 if (!lru_resize)
 179                         ns->ns_max_unused = (unsigned int)tmp;
 180
 181                 if (tmp > ns->ns_nr_unused)
 182                         tmp = ns->ns_nr_unused;
 183                 tmp = ns->ns_nr_unused - tmp;
 184
 185                 CDEBUG(D_DLMTRACE, "changing namespace %s unused locks from %u to %u\n",
 186                        ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
 187                 ldlm_cancel_lru(ns, (unsigned int)tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 188
 189                 if (!lru_resize) {
 190                         CDEBUG(D_DLMTRACE, "disable lru_resize for namespace %s\n",
 191                                ns->ns_name);
 192                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 193                 }
 194         } else {
 195                 CDEBUG(D_DLMTRACE, "changing namespace %s max_unused from %u to %u\n",
 196                        ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
 197                 ns->ns_max_unused = (unsigned int)tmp;
 198                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 199
 200                 /* Make sure that originally lru resize was supported before
 201                  * turning it on here. */
 202                 if (lru_resize &&
 203                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 204                         CDEBUG(D_DLMTRACE, "enable lru_resize for namespace %s\n",
 205                                ns->ns_name);
 206                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 207                 }
 208         }
 209
 210         return count;
 211 }
 212
 213 void ldlm_proc_namespace(struct ldlm_namespace *ns)
 214 {
 215         struct lprocfs_vars lock_vars[2];
 216         char lock_name[MAX_STRING_SIZE + 1];
 217
 218         LASSERT(ns != NULL);
 219         LASSERT(ns->ns_name != NULL);
 220
 221         lock_name[MAX_STRING_SIZE] = '\0';
 222
 223         memset(lock_vars, 0, sizeof(lock_vars));
 224         lock_vars[0].name = lock_name;
 225
 226         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
 227         lock_vars[0].data = &ns->ns_refcount;
 228         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 229         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 230
 231         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
 232         lock_vars[0].data = &ns->ns_locks;
 233         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 234         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 235
 236         if (ns_is_client(ns)) {
 237                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 238                          ns->ns_name);
 239                 lock_vars[0].data = &ns->ns_nr_unused;
 240                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 241                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 242
 243                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 244                          ns->ns_name);
 245                 lock_vars[0].data = ns;
 246                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 247                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 248                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 249
 250                 snprintf(lock_name, MAX_STRING_SIZE, "%s/shrink_thumb",
 251                          ns->ns_name);
 252                 lock_vars[0].data = ns;
 253                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 254                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 255                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 256
 257                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 258                          ns->ns_name);
 259                 lock_vars[0].data = &ns->ns_max_age;
 260                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 261                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 262                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 263         } else {
 264                 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
 265                          ns->ns_name);
 266                 lock_vars[0].data = &ns->ns_ctime_age_limit;
 267                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 268                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 269                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 270
 271                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 272                          ns->ns_name);
 273                 lock_vars[0].data = &ns->ns_max_nolock_size;
 274                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 275                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 276                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 277
 278                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 279                          ns->ns_name);
 280                 lock_vars[0].data = &ns->ns_contention_time;
 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
 285                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 286                          ns->ns_name);
 287                 lock_vars[0].data = &ns->ns_contended_locks;
 288                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 289                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 290                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 291         }
 292 }
 293 #undef MAX_STRING_SIZE
 294 #else
 295 #define ldlm_proc_namespace(ns) do {} while (0)
 296 #endif /* LPROCFS */
 297
 298 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 299                                           ldlm_side_t client, ldlm_appetite_t apt)
 300 {
 301         struct ldlm_namespace *ns = NULL;
 302         struct list_head *bucket;
 303         int rc, idx, namelen;
 304         ENTRY;
 305
 306         rc = ldlm_get_ref();
 307         if (rc) {
 308                 CERROR("ldlm_get_ref failed: %d\n", rc);
 309                 RETURN(NULL);
 310         }
 311
 312         OBD_ALLOC_PTR(ns);
 313         if (!ns)
 314                 GOTO(out_ref, NULL);
 315
 316         OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 317         if (!ns->ns_hash)
 318                 GOTO(out_ns, 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         namelen = strlen(name);
 327         OBD_ALLOC(ns->ns_name, namelen + 1);
 328         if (!ns->ns_name)
 329                 GOTO(out_hash, NULL);
 330
 331         strcpy(ns->ns_name, name);
 332
 333         CFS_INIT_LIST_HEAD(&ns->ns_root_list);
 334         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 335         ns->ns_refcount = 0;
 336         ns->ns_client = client;
 337         spin_lock_init(&ns->ns_hash_lock);
 338         atomic_set(&ns->ns_locks, 0);
 339         ns->ns_resources = 0;
 340         cfs_waitq_init(&ns->ns_waitq);
 341         ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
 342         ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
 343         ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
 344
 345         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 346              bucket--)
 347                 CFS_INIT_LIST_HEAD(bucket);
 348
 349         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 350         ns->ns_nr_unused = 0;
 351         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
 352         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
 353         ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
 354         spin_lock_init(&ns->ns_unused_lock);
 355         ns->ns_orig_connect_flags = 0;
 356         ns->ns_connect_flags = 0;
 357         ldlm_proc_namespace(ns);
 358
 359         idx = atomic_read(ldlm_namespace_nr(client));
 360         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 361         if (rc) {
 362                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 363                 GOTO(out_proc, rc);
 364         }
 365
 366         ldlm_namespace_register(ns, client);
 367         RETURN(ns);
 368 out_proc:
 369         ldlm_namespace_cleanup(ns, 0);
 370         OBD_FREE(ns->ns_name, namelen + 1);
 371 out_hash:
 372         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 373 out_ns:
 374         OBD_FREE_PTR(ns);
 375 out_ref:
 376         ldlm_put_ref();
 377         RETURN(NULL);
 378 }
 379
 380 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 381
 382 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 383  * This is currently only used for recovery, and we make certain assumptions
 384  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 385  *
 386  * Called with the ns_lock held. */
 387 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 388                              int flags)
 389 {
 390         struct list_head *tmp;
 391         int rc = 0, client = ns_is_client(res->lr_namespace);
 392         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 393         ENTRY;
 394
 395
 396         do {
 397                 struct ldlm_lock *lock = NULL;
 398
 399                 /* first, we look for non-cleaned-yet lock
 400                  * all cleaned locks are marked by CLEANED flag */
 401                 lock_res(res);
 402                 list_for_each(tmp, q) {
 403                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 404                         if (lock->l_flags & LDLM_FL_CLEANED) {
 405                                 lock = NULL;
 406                                 continue;
 407                         }
 408                         LDLM_LOCK_GET(lock);
 409                         lock->l_flags |= LDLM_FL_CLEANED;
 410                         break;
 411                 }
 412
 413                 if (lock == NULL) {
 414                         unlock_res(res);
 415                         break;
 416                 }
 417
 418                 /* Set CBPENDING so nothing in the cancellation path
 419                  * can match this lock */
 420                 lock->l_flags |= LDLM_FL_CBPENDING;
 421                 lock->l_flags |= LDLM_FL_FAILED;
 422                 lock->l_flags |= flags;
 423
 424                 /* ... without sending a CANCEL message for local_only. */
 425                 if (local_only)
 426                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 427
 428                 if (local_only && (lock->l_readers || lock->l_writers)) {
 429                         /* This is a little bit gross, but much better than the
 430                          * alternative: pretend that we got a blocking AST from
 431                          * the server, so that when the lock is decref'd, it
 432                          * will go away ... */
 433                         unlock_res(res);
 434                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 435                         if (lock->l_completion_ast)
 436                                 lock->l_completion_ast(lock, 0, NULL);
 437                         LDLM_LOCK_PUT(lock);
 438                         continue;
 439                 }
 440
 441                 if (client) {
 442                         struct lustre_handle lockh;
 443
 444                         unlock_res(res);
 445                         ldlm_lock2handle(lock, &lockh);
 446                         rc = ldlm_cli_cancel(&lockh);
 447                         if (rc)
 448                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 449                 } else {
 450                         ldlm_resource_unlink_lock(lock);
 451                         unlock_res(res);
 452                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 453                                    "client node");
 454                         ldlm_lock_destroy(lock);
 455                 }
 456                 LDLM_LOCK_PUT(lock);
 457         } while (1);
 458
 459         EXIT;
 460 }
 461
 462 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 463 {
 464         struct list_head *tmp;
 465         int i;
 466
 467         if (ns == NULL) {
 468                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 469                 return ELDLM_OK;
 470         }
 471
 472         for (i = 0; i < RES_HASH_SIZE; i++) {
 473                 spin_lock(&ns->ns_hash_lock);
 474                 tmp = ns->ns_hash[i].next;
 475                 while (tmp != &(ns->ns_hash[i])) {
 476                         struct ldlm_resource *res;
 477                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 478                         ldlm_resource_getref(res);
 479                         spin_unlock(&ns->ns_hash_lock);
 480
 481                         cleanup_resource(res, &res->lr_granted, flags);
 482                         cleanup_resource(res, &res->lr_converting, flags);
 483                         cleanup_resource(res, &res->lr_waiting, flags);
 484
 485                         spin_lock(&ns->ns_hash_lock);
 486                         tmp  = tmp->next;
 487
 488                         /* XXX: former stuff caused issues in case of race
 489                          * between ldlm_namespace_cleanup() and lockd() when
 490                          * client gets blocking ast when lock gets distracted by
 491                          * server. This is 1_4 branch solution, let's see how
 492                          * will it behave. */
 493                         if (!ldlm_resource_putref_locked(res))
 494                                 CDEBUG(D_INFO,
 495                                        "Namespace %s resource refcount nonzero "
 496                                        "(%d) after lock cleanup; forcing cleanup.\n",
 497                                        ns->ns_name, atomic_read(&res->lr_refcount));
 498                 }
 499                 spin_unlock(&ns->ns_hash_lock);
 500         }
 501
 502         return ELDLM_OK;
 503 }
 504
 505 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 506 {
 507         ENTRY;
 508
 509         /* At shutdown time, don't call the cancellation callback */
 510         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 511
 512         if (ns->ns_refcount > 0) {
 513                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 514                 int rc;
 515                 CDEBUG(D_DLMTRACE,
 516                        "dlm namespace %s free waiting on refcount %d\n",
 517                        ns->ns_name, ns->ns_refcount);
 518 force_wait:
 519                 if (force)
 520                         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 521
 522                 rc = l_wait_event(ns->ns_waitq,
 523                                   ns->ns_refcount == 0, &lwi);
 524
 525                 /* Forced cleanups should be able to reclaim all references,
 526                  * so it's safe to wait forever... we can't leak locks... */
 527                 if (force && rc == -ETIMEDOUT) {
 528                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 529                                        "namespace with %d resources in use, "
 530                                        "(rc=%d)\n", ns->ns_name,
 531                                        ns->ns_refcount, rc);
 532                         GOTO(force_wait, rc);
 533                 }
 534
 535                 if (ns->ns_refcount) {
 536                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 537                                        "with %d resources in use, (rc=%d)\n",
 538                                        ns->ns_name,
 539                                        ns->ns_refcount, rc);
 540                         RETURN(ELDLM_NAMESPACE_EXISTS);
 541                 }
 542                 CDEBUG(D_DLMTRACE,
 543                        "dlm namespace %s free done waiting\n", ns->ns_name);
 544         }
 545
 546         RETURN(ELDLM_OK);
 547 }
 548
 549 /**
 550  * Performs various cleanups for passed \a ns to make it drop refc and be ready
 551  * for freeing. Waits for refc == 0.
 552  *
 553  * The following is done:
 554  * (0) Unregister \a ns from its list to make inaccessible for potential users
 555  * like pools thread and others;
 556  * (1) Clear all locks in \a ns.
 557  */
 558 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 559                                struct obd_import *imp,
 560                                int force)
 561 {
 562         int rc;
 563         ENTRY;
 564         if (!ns) {
 565                 EXIT;
 566                 return;
 567         }
 568
 569         /*
 570          * Make sure that nobody can find this ns in its list.
 571          */
 572         ldlm_namespace_unregister(ns, ns->ns_client);
 573
 574         /*
 575          * Can fail with -EINTR when force == 0 in which case try harder.
 576          */
 577         rc = __ldlm_namespace_free(ns, force);
 578         if (rc != ELDLM_OK) {
 579                 if (imp) {
 580                         ptlrpc_disconnect_import(imp, 0);
 581                         ptlrpc_invalidate_import(imp);
 582                 }
 583
 584                 /*
 585                  * With all requests dropped and the import inactive
 586                  * we are gaurenteed all reference will be dropped.
 587                  */
 588                 rc = __ldlm_namespace_free(ns, 1);
 589                 LASSERT(rc == 0);
 590         }
 591         EXIT;
 592 }
 593
 594 /**
 595  * Performs freeing memory structures related to \a ns. This is only done when
 596  * ldlm_namespce_free_prior() successfully removed all resources referencing
 597  * \a ns and its refc == 0.
 598  */
 599 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 600 {
 601         ENTRY;
 602         if (!ns) {
 603                 EXIT;
 604                 return;
 605         }
 606
 607         /*
 608          * Fini pool _before_ parent proc dir is removed. This is important as
 609          * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
 610          * it after @dir may cause oops.
 611          */
 612         ldlm_pool_fini(&ns->ns_pool);
 613
 614 #ifdef LPROCFS
 615         {
 616                 struct proc_dir_entry *dir;
 617                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
 618                 if (dir == NULL) {
 619                         CERROR("dlm namespace %s has no procfs dir?\n",
 620                                ns->ns_name);
 621                 } else {
 622                         lprocfs_remove(&dir);
 623                 }
 624         }
 625 #endif
 626
 627         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 628         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 629
 630         /*
 631          * Namespace \a ns should be not on list in this time, otherwise this
 632          * will cause issues realted to using freed \a ns in pools thread.
 633          */
 634         LASSERT(list_empty(&ns->ns_list_chain));
 635         OBD_FREE_PTR(ns);
 636         ldlm_put_ref();
 637         EXIT;
 638 }
 639
 640
 641 /* Cleanup the resource, and free namespace.
 642  * bug 12864:
 643  * Deadlock issue:
 644  * proc1: destroy import
 645  *        class_disconnect_export(grab cl_sem) ->
 646  *              -> ldlm_namespace_free ->
 647  *              -> lprocfs_remove(grab _lprocfs_lock).
 648  * proc2: read proc info
 649  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 650  *              -> osc_rd_active, etc(grab cl_sem).
 651  *
 652  * So that I have to split the ldlm_namespace_free into two parts - the first
 653  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 654  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 655  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 656  * held.
 657  */
 658 void ldlm_namespace_free(struct ldlm_namespace *ns,
 659                          struct obd_import *imp,
 660                          int force)
 661 {
 662         ldlm_namespace_free_prior(ns, imp, force);
 663         ldlm_namespace_free_post(ns);
 664 }
 665
 666
 667 void ldlm_namespace_get_locked(struct ldlm_namespace *ns)
 668 {
 669         LASSERT(ns->ns_refcount >= 0);
 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 static __u32 ldlm_hash_fn(struct ldlm_resource *parent,
 737                           const 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, const struct ldlm_res_id *name,
 787                    __u32 hash)
 788 {
 789         struct list_head *bucket, *tmp;
 790         struct ldlm_resource *res;
 791
 792         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 793         bucket = ns->ns_hash + hash;
 794
 795         list_for_each(tmp, bucket) {
 796                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 797                 if (memcmp(&res->lr_name, name, sizeof(res->lr_name)) == 0)
 798                         return res;
 799         }
 800
 801         return NULL;
 802 }
 803
 804 /* Args: locked namespace
 805  * Returns: newly-allocated, referenced, unlocked resource */
 806 static struct ldlm_resource *
 807 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 808                   const struct ldlm_res_id *name, __u32 hash, ldlm_type_t type)
 809 {
 810         struct list_head *bucket;
 811         struct ldlm_resource *res, *old_res;
 812         ENTRY;
 813
 814         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
 815                  "type: %d\n", type);
 816
 817         res = ldlm_resource_new();
 818         if (!res)
 819                 RETURN(NULL);
 820
 821         res->lr_name = *name;
 822         res->lr_namespace = ns;
 823         res->lr_type = type;
 824         res->lr_most_restr = LCK_NL;
 825
 826         spin_lock(&ns->ns_hash_lock);
 827         old_res = ldlm_resource_find(ns, name, hash);
 828         if (old_res) {
 829                 /* someone won the race and added the resource before */
 830                 ldlm_resource_getref(old_res);
 831                 spin_unlock(&ns->ns_hash_lock);
 832                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 833                 /* synchronize WRT resource creation */
 834                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 835                         down(&old_res->lr_lvb_sem);
 836                         up(&old_res->lr_lvb_sem);
 837                 }
 838                 RETURN(old_res);
 839         }
 840
 841         /* we won! let's add the resource */
 842         bucket = ns->ns_hash + hash;
 843         list_add(&res->lr_hash, bucket);
 844         ns->ns_resources++;
 845         ldlm_namespace_get_locked(ns);
 846
 847         if (parent == NULL) {
 848                 list_add(&res->lr_childof, &ns->ns_root_list);
 849         } else {
 850                 res->lr_parent = parent;
 851                 list_add(&res->lr_childof, &parent->lr_children);
 852         }
 853         spin_unlock(&ns->ns_hash_lock);
 854
 855         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 856                 int rc;
 857
 858                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
 859                 rc = ns->ns_lvbo->lvbo_init(res);
 860                 if (rc)
 861                         CERROR("lvbo_init failed for resource "
 862                                LPU64": rc %d\n", 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                   const 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 here to not wake ->ns_waitq up yet, we will do it few
 943          * lines below when all children are freed. */
 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 /* Returns 1 if the resource was freed, 0 if it remains. */
 954 int ldlm_resource_putref(struct ldlm_resource *res)
 955 {
 956         struct ldlm_namespace *ns = res->lr_namespace;
 957         int rc = 0;
 958         ENTRY;
 959
 960         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 961                atomic_read(&res->lr_refcount) - 1);
 962         LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
 963                  atomic_read(&res->lr_refcount));
 964         LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
 965                  atomic_read(&res->lr_refcount));
 966
 967         if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
 968                 __ldlm_resource_putref_final(res);
 969                 spin_unlock(&ns->ns_hash_lock);
 970                 if (res->lr_lvb_data)
 971                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 972                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 973                 rc = 1;
 974         }
 975
 976         RETURN(rc);
 977 }
 978
 979 /* Returns 1 if the resource was freed, 0 if it remains. */
 980 int ldlm_resource_putref_locked(struct ldlm_resource *res)
 981 {
 982         int rc = 0;
 983         ENTRY;
 984
 985         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 986                atomic_read(&res->lr_refcount) - 1);
 987         LASSERT(atomic_read(&res->lr_refcount) > 0);
 988         LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
 989
 990         LASSERT(atomic_read(&res->lr_refcount) >= 0);
 991         if (atomic_dec_and_test(&res->lr_refcount)) {
 992                 __ldlm_resource_putref_final(res);
 993                 if (res->lr_lvb_data)
 994                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 995                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 996                 rc = 1;
 997         }
 998
 999         RETURN(rc);
1000 }
1001
1002 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1003                             struct ldlm_lock *lock)
1004 {
1005         check_res_locked(res);
1006
1007         ldlm_resource_dump(D_OTHER, res);
1008         CDEBUG(D_OTHER, "About to add this lock:\n");
1009         ldlm_lock_dump(D_OTHER, lock, 0);
1010
1011         if (lock->l_destroyed) {
1012                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1013                 return;
1014         }
1015
1016         LASSERT(list_empty(&lock->l_res_link));
1017
1018         list_add_tail(&lock->l_res_link, head);
1019 }
1020
1021 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1022                                      struct ldlm_lock *new)
1023 {
1024         struct ldlm_resource *res = original->l_resource;
1025
1026         check_res_locked(res);
1027
1028         ldlm_resource_dump(D_OTHER, res);
1029         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1030         ldlm_lock_dump(D_OTHER, new, 0);
1031
1032         if (new->l_destroyed) {
1033                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1034                 goto out;
1035         }
1036
1037         LASSERT(list_empty(&new->l_res_link));
1038
1039         list_add(&new->l_res_link, &original->l_res_link);
1040  out:;
1041 }
1042
1043 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1044 {
1045         int type = lock->l_resource->lr_type;
1046
1047         check_res_locked(lock->l_resource);
1048         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1049                 ldlm_unlink_lock_skiplist(lock);
1050         else if (type == LDLM_EXTENT)
1051                 ldlm_extent_unlink_lock(lock);
1052         list_del_init(&lock->l_res_link);
1053 }
1054
1055 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1056 {
1057         desc->lr_type = res->lr_type;
1058         desc->lr_name = res->lr_name;
1059 }
1060
1061 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1062 {
1063         struct list_head *tmp;
1064
1065         if (!((libcfs_debug | D_ERROR) & level))
1066                 return;
1067
1068         mutex_down(ldlm_namespace_lock(client));
1069
1070         list_for_each(tmp, ldlm_namespace_list(client)) {
1071                 struct ldlm_namespace *ns;
1072                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1073                 ldlm_namespace_dump(level, ns);
1074         }
1075
1076         mutex_up(ldlm_namespace_lock(client));
1077 }
1078
1079 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1080 {
1081         struct list_head *tmp;
1082
1083         if (!((libcfs_debug | D_ERROR) & level))
1084                 return;
1085
1086         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1087                ns->ns_name, ns->ns_refcount,
1088                ns_is_client(ns) ? "client" : "server");
1089
1090         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1091                 return;
1092
1093         spin_lock(&ns->ns_hash_lock);
1094         tmp = ns->ns_root_list.next;
1095         while (tmp != &ns->ns_root_list) {
1096                 struct ldlm_resource *res;
1097                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
1098
1099                 ldlm_resource_getref(res);
1100                 spin_unlock(&ns->ns_hash_lock);
1101
1102                 lock_res(res);
1103                 ldlm_resource_dump(level, res);
1104                 unlock_res(res);
1105
1106                 spin_lock(&ns->ns_hash_lock);
1107                 tmp = tmp->next;
1108                 ldlm_resource_putref_locked(res);
1109         }
1110         ns->ns_next_dump = cfs_time_shift(10);
1111         spin_unlock(&ns->ns_hash_lock);
1112 }
1113
1114 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1115 {
1116         struct list_head *tmp;
1117         int pos;
1118
1119         CLASSERT(RES_NAME_SIZE == 4);
1120
1121         if (!((libcfs_debug | D_ERROR) & level))
1122                 return;
1123
1124         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1125                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1126                res->lr_name.name[2], res->lr_name.name[3],
1127                atomic_read(&res->lr_refcount));
1128
1129         if (!list_empty(&res->lr_granted)) {
1130                 pos = 0;
1131                 CDEBUG(level, "Granted locks:\n");
1132                 list_for_each(tmp, &res->lr_granted) {
1133                         struct ldlm_lock *lock;
1134                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1135                         ldlm_lock_dump(level, lock, ++pos);
1136                 }
1137         }
1138         if (!list_empty(&res->lr_converting)) {
1139                 pos = 0;
1140                 CDEBUG(level, "Converting locks:\n");
1141                 list_for_each(tmp, &res->lr_converting) {
1142                         struct ldlm_lock *lock;
1143                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1144                         ldlm_lock_dump(level, lock, ++pos);
1145                 }
1146         }
1147         if (!list_empty(&res->lr_waiting)) {
1148                 pos = 0;
1149                 CDEBUG(level, "Waiting locks:\n");
1150                 list_for_each(tmp, &res->lr_waiting) {
1151                         struct ldlm_lock *lock;
1152                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1153                         ldlm_lock_dump(level, lock, ++pos);
1154                 }
1155         }
1156 }