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
 212         return count;
 213 }
 214
 215 void ldlm_proc_namespace(struct ldlm_namespace *ns)
 216 {
 217         struct lprocfs_vars lock_vars[2];
 218         char lock_name[MAX_STRING_SIZE + 1];
 219
 220         LASSERT(ns != NULL);
 221         LASSERT(ns->ns_name != NULL);
 222
 223         lock_name[MAX_STRING_SIZE] = '\0';
 224
 225         memset(lock_vars, 0, sizeof(lock_vars));
 226         lock_vars[0].name = lock_name;
 227
 228         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
 229         lock_vars[0].data = &ns->ns_refcount;
 230         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 231         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 232
 233         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
 234         lock_vars[0].data = &ns->ns_locks;
 235         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 236         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 237
 238         if (ns_is_client(ns)) {
 239                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 240                          ns->ns_name);
 241                 lock_vars[0].data = &ns->ns_nr_unused;
 242                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 243                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 244
 245                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 246                          ns->ns_name);
 247                 lock_vars[0].data = ns;
 248                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 249                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 250                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 251
 252                 snprintf(lock_name, MAX_STRING_SIZE, "%s/shrink_thumb",
 253                          ns->ns_name);
 254                 lock_vars[0].data = ns;
 255                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 256                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 257                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 258
 259                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 260                          ns->ns_name);
 261                 lock_vars[0].data = &ns->ns_max_age;
 262                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 263                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 264                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 265         } else {
 266                 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
 267                          ns->ns_name);
 268                 lock_vars[0].data = &ns->ns_ctime_age_limit;
 269                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 270                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 271                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 272         }
 273 }
 274 #undef MAX_STRING_SIZE
 275 #else
 276 #define ldlm_proc_namespace(ns) do {} while (0)
 277 #endif /* LPROCFS */
 278
 279 struct ldlm_namespace *ldlm_namespace_new(char *name, ldlm_side_t client,
 280                                           ldlm_appetite_t apt)
 281 {
 282         struct ldlm_namespace *ns = NULL;
 283         struct list_head *bucket;
 284         int rc, idx, namelen;
 285         ENTRY;
 286
 287         rc = ldlm_get_ref();
 288         if (rc) {
 289                 CERROR("ldlm_get_ref failed: %d\n", rc);
 290                 RETURN(NULL);
 291         }
 292
 293         OBD_ALLOC_PTR(ns);
 294         if (!ns)
 295                 GOTO(out_ref, NULL);
 296
 297         OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 298         if (!ns->ns_hash)
 299                 GOTO(out_ns, NULL);
 300
 301         ns->ns_shrink_thumb = LDLM_LOCK_SHRINK_THUMB;
 302         ns->ns_appetite = apt;
 303         namelen = strlen(name);
 304         OBD_ALLOC(ns->ns_name, namelen + 1);
 305         if (!ns->ns_name)
 306                 GOTO(out_hash, NULL);
 307
 308         strcpy(ns->ns_name, name);
 309
 310         CFS_INIT_LIST_HEAD(&ns->ns_root_list);
 311         ns->ns_refcount = 0;
 312         ns->ns_client = client;
 313         spin_lock_init(&ns->ns_hash_lock);
 314         atomic_set(&ns->ns_locks, 0);
 315         ns->ns_resources = 0;
 316         cfs_waitq_init(&ns->ns_waitq);
 317
 318         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 319              bucket--)
 320                 CFS_INIT_LIST_HEAD(bucket);
 321
 322         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 323         ns->ns_nr_unused = 0;
 324         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
 325         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
 326         ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
 327         spin_lock_init(&ns->ns_unused_lock);
 328         ns->ns_orig_connect_flags = 0;
 329         ns->ns_connect_flags = 0;
 330         ldlm_proc_namespace(ns);
 331
 332         idx = atomic_read(ldlm_namespace_nr(client));
 333         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 334         if (rc) {
 335                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 336                 GOTO(out_proc, rc);
 337         }
 338
 339         mutex_down(ldlm_namespace_lock(client));
 340         list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 341         atomic_inc(ldlm_namespace_nr(client));
 342         mutex_up(ldlm_namespace_lock(client));
 343
 344         RETURN(ns);
 345 out_proc:
 346         ldlm_namespace_cleanup(ns, 0);
 347         OBD_FREE(ns->ns_name, namelen + 1);
 348 out_hash:
 349         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 350 out_ns:
 351         OBD_FREE_PTR(ns);
 352 out_ref:
 353         ldlm_put_ref(0);
 354         RETURN(NULL);
 355 }
 356
 357 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 358
 359 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 360  * This is currently only used for recovery, and we make certain assumptions
 361  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 362  *
 363  * Called with the ns_lock held. */
 364 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 365                              int flags)
 366 {
 367         struct list_head *tmp;
 368         int rc = 0, client = ns_is_client(res->lr_namespace);
 369         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 370         ENTRY;
 371
 372
 373         do {
 374                 struct ldlm_lock *lock = NULL;
 375
 376                 /* first, we look for non-cleaned-yet lock
 377                  * all cleaned locks are marked by CLEANED flag */
 378                 lock_res(res);
 379                 list_for_each(tmp, q) {
 380                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 381                         if (lock->l_flags & LDLM_FL_CLEANED) {
 382                                 lock = NULL;
 383                                 continue;
 384                         }
 385                         LDLM_LOCK_GET(lock);
 386                         lock->l_flags |= LDLM_FL_CLEANED;
 387                         break;
 388                 }
 389
 390                 if (lock == NULL) {
 391                         unlock_res(res);
 392                         break;
 393                 }
 394
 395                 /* Set CBPENDING so nothing in the cancellation path
 396                  * can match this lock */
 397                 lock->l_flags |= LDLM_FL_CBPENDING;
 398                 lock->l_flags |= LDLM_FL_FAILED;
 399                 lock->l_flags |= flags;
 400
 401                 /* ... without sending a CANCEL message for local_only. */
 402                 if (local_only)
 403                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 404
 405                 if (local_only && (lock->l_readers || lock->l_writers)) {
 406                         /* This is a little bit gross, but much better than the
 407                          * alternative: pretend that we got a blocking AST from
 408                          * the server, so that when the lock is decref'd, it
 409                          * will go away ... */
 410                         unlock_res(res);
 411                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 412                         if (lock->l_completion_ast)
 413                                 lock->l_completion_ast(lock, 0, NULL);
 414                         LDLM_LOCK_PUT(lock);
 415                         continue;
 416                 }
 417
 418                 if (client) {
 419                         struct lustre_handle lockh;
 420
 421                         unlock_res(res);
 422                         ldlm_lock2handle(lock, &lockh);
 423                         rc = ldlm_cli_cancel(&lockh);
 424                         if (rc)
 425                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 426                 } else {
 427                         ldlm_resource_unlink_lock(lock);
 428                         unlock_res(res);
 429                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 430                                    "client node");
 431                         ldlm_lock_destroy(lock);
 432                 }
 433                 LDLM_LOCK_PUT(lock);
 434         } while (1);
 435
 436         EXIT;
 437 }
 438
 439 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 440 {
 441         struct list_head *tmp;
 442         int i;
 443
 444         if (ns == NULL) {
 445                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 446                 return ELDLM_OK;
 447         }
 448
 449         for (i = 0; i < RES_HASH_SIZE; i++) {
 450                 spin_lock(&ns->ns_hash_lock);
 451                 tmp = ns->ns_hash[i].next;
 452                 while (tmp != &(ns->ns_hash[i])) {
 453                         struct ldlm_resource *res;
 454                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 455                         ldlm_resource_getref(res);
 456                         spin_unlock(&ns->ns_hash_lock);
 457
 458                         cleanup_resource(res, &res->lr_granted, flags);
 459                         cleanup_resource(res, &res->lr_converting, flags);
 460                         cleanup_resource(res, &res->lr_waiting, flags);
 461
 462                         spin_lock(&ns->ns_hash_lock);
 463                         tmp  = tmp->next;
 464
 465                         /* XXX: former stuff caused issues in case of race
 466                          * between ldlm_namespace_cleanup() and lockd() when
 467                          * client gets blocking ast when lock gets distracted by
 468                          * server. This is 1_4 branch solution, let's see how
 469                          * will it behave. */
 470                         if (!ldlm_resource_putref_locked(res))
 471                                 CDEBUG(D_INFO,
 472                                        "Namespace %s resource refcount nonzero "
 473                                        "(%d) after lock cleanup; forcing cleanup.\n",
 474                                        ns->ns_name, atomic_read(&res->lr_refcount));
 475                 }
 476                 spin_unlock(&ns->ns_hash_lock);
 477         }
 478
 479         return ELDLM_OK;
 480 }
 481
 482 /* Cleanup, but also free, the namespace */
 483 int ldlm_namespace_free_prior(struct ldlm_namespace *ns)
 484 {
 485         ENTRY;
 486         if (!ns)
 487                 RETURN(ELDLM_OK);
 488
 489         mutex_down(ldlm_namespace_lock(ns->ns_client));
 490         /*
 491          * Some asserts and possibly other parts of code still using
 492          * list_empty(&ns->ns_list_chain). This is why it is important
 493          * to use list_del_init() here.
 494          */
 495         list_del_init(&ns->ns_list_chain);
 496         atomic_dec(ldlm_namespace_nr(ns->ns_client));
 497         ldlm_pool_fini(&ns->ns_pool);
 498         mutex_up(ldlm_namespace_lock(ns->ns_client));
 499
 500         /* At shutdown time, don't call the cancellation callback */
 501         ldlm_namespace_cleanup(ns, 0);
 502
 503         if (ns->ns_refcount > 0) {
 504                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 505                 int rc;
 506                 CDEBUG(D_DLMTRACE,
 507                        "dlm namespace %s free waiting on refcount %d\n",
 508                        ns->ns_name, ns->ns_refcount);
 509                 rc = l_wait_event(ns->ns_waitq,
 510                                   ns->ns_refcount == 0, &lwi);
 511                 if (ns->ns_refcount)
 512                         LCONSOLE_ERROR_MSG(0x139, "Lock manager: wait for %s "
 513                                            "namespace cleanup aborted with %d "
 514                                            "resources in use. (%d)\nI'm going "
 515                                            "to try to clean up anyway, but I "
 516                                            "might need a reboot of this node.\n",
 517                                             ns->ns_name, (int) ns->ns_refcount,
 518                                             rc);
 519                 CDEBUG(D_DLMTRACE,
 520                        "dlm namespace %s free done waiting\n", ns->ns_name);
 521         }
 522
 523         RETURN(ELDLM_OK);
 524 }
 525
 526 int ldlm_namespace_free_post(struct ldlm_namespace *ns, int force)
 527 {
 528         ENTRY;
 529         if (!ns)
 530                 RETURN(ELDLM_OK);
 531
 532 #ifdef LPROCFS
 533         {
 534                 struct proc_dir_entry *dir;
 535                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
 536                 if (dir == NULL) {
 537                         CERROR("dlm namespace %s has no procfs dir?\n",
 538                                ns->ns_name);
 539                 } else {
 540                         lprocfs_remove(&dir);
 541                 }
 542         }
 543 #endif
 544
 545         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 546         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 547         /*
 548          * @ns should be not on list in this time, otherwise this will cause
 549          * issues realted to using freed @ns in pools thread.
 550          */
 551         LASSERT(list_empty(&ns->ns_list_chain));
 552         OBD_FREE_PTR(ns);
 553         ldlm_put_ref(force);
 554         RETURN(ELDLM_OK);
 555 }
 556
 557
 558 /* Cleanup the resource, and free namespace.
 559  * bug 12864:
 560  * Deadlock issue:
 561  * proc1: destroy import
 562  *        class_disconnect_export(grab cl_sem) ->
 563  *              -> ldlm_namespace_free ->
 564  *              -> lprocfs_remove(grab _lprocfs_lock).
 565  * proc2: read proc info
 566  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 567  *              -> osc_rd_active, etc(grab cl_sem).
 568  *
 569  * So that I have to split the ldlm_namespace_free into two parts - the first
 570  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 571  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 572  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 573  * held.
 574  */
 575 int ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 576 {
 577         ldlm_namespace_free_prior(ns);
 578         ldlm_namespace_free_post(ns, force);
 579         return ELDLM_OK;
 580 }
 581
 582
 583 void ldlm_namespace_get_nolock(struct ldlm_namespace *ns)
 584 {
 585         LASSERT(ns->ns_refcount >= 0);
 586         ns->ns_refcount++;
 587 }
 588
 589 void ldlm_namespace_get(struct ldlm_namespace *ns)
 590 {
 591         spin_lock(&ns->ns_hash_lock);
 592         ldlm_namespace_get_nolock(ns);
 593         spin_unlock(&ns->ns_hash_lock);
 594 }
 595
 596 void ldlm_namespace_put_nolock(struct ldlm_namespace *ns, int wakeup)
 597 {
 598         LASSERT(ns->ns_refcount > 0);
 599         ns->ns_refcount--;
 600         if (ns->ns_refcount == 0 && wakeup)
 601                 wake_up(&ns->ns_waitq);
 602 }
 603
 604 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
 605 {
 606         spin_lock(&ns->ns_hash_lock);
 607         ldlm_namespace_put_nolock(ns, wakeup);
 608         spin_unlock(&ns->ns_hash_lock);
 609 }
 610
 611 /* Should be called under ldlm_namespace_lock(client) taken */
 612 void ldlm_namespace_move(struct ldlm_namespace *ns, ldlm_side_t client)
 613 {
 614         LASSERT(!list_empty(&ns->ns_list_chain));
 615         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 616         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 617 }
 618
 619 /* Should be called under ldlm_namespace_lock(client) taken */
 620 struct ldlm_namespace *ldlm_namespace_first(ldlm_side_t client)
 621 {
 622         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 623         LASSERT(!list_empty(ldlm_namespace_list(client)));
 624         return container_of(ldlm_namespace_list(client)->next,
 625                 struct ldlm_namespace, ns_list_chain);
 626 }
 627 static __u32 ldlm_hash_fn(struct ldlm_resource *parent,
 628                           const struct ldlm_res_id *name)
 629 {
 630         __u32 hash = 0;
 631         int i;
 632
 633         for (i = 0; i < RES_NAME_SIZE; i++)
 634                 hash += name->name[i];
 635
 636         hash += (__u32)((unsigned long)parent >> 4);
 637
 638         return (hash & RES_HASH_MASK);
 639 }
 640
 641 static struct ldlm_resource *ldlm_resource_new(void)
 642 {
 643         struct ldlm_resource *res;
 644         int idx;
 645
 646         OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
 647         if (res == NULL)
 648                 return NULL;
 649
 650         memset(res, 0, sizeof(*res));
 651
 652         CFS_INIT_LIST_HEAD(&res->lr_children);
 653         CFS_INIT_LIST_HEAD(&res->lr_childof);
 654         CFS_INIT_LIST_HEAD(&res->lr_granted);
 655         CFS_INIT_LIST_HEAD(&res->lr_converting);
 656         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 657
 658         /* initialize interval trees for each lock mode*/
 659         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 660                 res->lr_itree[idx].lit_size = 0;
 661                 res->lr_itree[idx].lit_mode = 1 << idx;
 662                 res->lr_itree[idx].lit_root = NULL;
 663         }
 664
 665         atomic_set(&res->lr_refcount, 1);
 666         spin_lock_init(&res->lr_lock);
 667
 668         /* one who creates the resource must unlock
 669          * the semaphore after lvb initialization */
 670         init_MUTEX_LOCKED(&res->lr_lvb_sem);
 671
 672         return res;
 673 }
 674
 675 /* must be called with hash lock held */
 676 static struct ldlm_resource *
 677 ldlm_resource_find(struct ldlm_namespace *ns, const struct ldlm_res_id *name,
 678                    __u32 hash)
 679 {
 680         struct list_head *bucket, *tmp;
 681         struct ldlm_resource *res;
 682
 683         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 684         bucket = ns->ns_hash + hash;
 685
 686         list_for_each(tmp, bucket) {
 687                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 688                 if (memcmp(&res->lr_name, name, sizeof(res->lr_name)) == 0)
 689                         return res;
 690         }
 691
 692         return NULL;
 693 }
 694
 695 /* Args: locked namespace
 696  * Returns: newly-allocated, referenced, unlocked resource */
 697 static struct ldlm_resource *
 698 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 699                   const struct ldlm_res_id *name, __u32 hash, ldlm_type_t type)
 700 {
 701         struct list_head *bucket;
 702         struct ldlm_resource *res, *old_res;
 703         ENTRY;
 704
 705         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
 706                  "type: %d\n", type);
 707
 708         res = ldlm_resource_new();
 709         if (!res)
 710                 RETURN(NULL);
 711
 712         res->lr_name = *name;
 713         res->lr_namespace = ns;
 714         res->lr_type = type;
 715         res->lr_most_restr = LCK_NL;
 716
 717         spin_lock(&ns->ns_hash_lock);
 718         old_res = ldlm_resource_find(ns, name, hash);
 719         if (old_res) {
 720                 /* someone won the race and added the resource before */
 721                 ldlm_resource_getref(old_res);
 722                 spin_unlock(&ns->ns_hash_lock);
 723                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 724                 /* synchronize WRT resource creation */
 725                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 726                         down(&old_res->lr_lvb_sem);
 727                         up(&old_res->lr_lvb_sem);
 728                 }
 729                 RETURN(old_res);
 730         }
 731
 732         /* we won! let's add the resource */
 733         bucket = ns->ns_hash + hash;
 734         list_add(&res->lr_hash, bucket);
 735         ns->ns_resources++;
 736         ldlm_namespace_get_nolock(ns);
 737
 738         if (parent == NULL) {
 739                 list_add(&res->lr_childof, &ns->ns_root_list);
 740         } else {
 741                 res->lr_parent = parent;
 742                 list_add(&res->lr_childof, &parent->lr_children);
 743         }
 744         spin_unlock(&ns->ns_hash_lock);
 745
 746         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 747                 int rc;
 748
 749                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
 750                 rc = ns->ns_lvbo->lvbo_init(res);
 751                 if (rc)
 752                         CERROR("lvbo_init failed for resource "
 753                                LPU64": rc %d\n", name->name[0], rc);
 754                 /* we create resource with locked lr_lvb_sem */
 755                 up(&res->lr_lvb_sem);
 756         }
 757
 758         RETURN(res);
 759 }
 760
 761 /* Args: unlocked namespace
 762  * Locks: takes and releases ns->ns_lock and res->lr_lock
 763  * Returns: referenced, unlocked ldlm_resource or NULL */
 764 struct ldlm_resource *
 765 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 766                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
 767 {
 768         __u32 hash = ldlm_hash_fn(parent, name);
 769         struct ldlm_resource *res = NULL;
 770         ENTRY;
 771
 772         LASSERT(ns != NULL);
 773         LASSERT(ns->ns_hash != NULL);
 774         LASSERT(name->name[0] != 0);
 775
 776         spin_lock(&ns->ns_hash_lock);
 777         res = ldlm_resource_find(ns, name, hash);
 778         if (res) {
 779                 ldlm_resource_getref(res);
 780                 spin_unlock(&ns->ns_hash_lock);
 781                 /* synchronize WRT resource creation */
 782                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 783                         down(&res->lr_lvb_sem);
 784                         up(&res->lr_lvb_sem);
 785                 }
 786                 RETURN(res);
 787         }
 788         spin_unlock(&ns->ns_hash_lock);
 789
 790         if (create == 0)
 791                 RETURN(NULL);
 792
 793         res = ldlm_resource_add(ns, parent, name, hash, type);
 794         RETURN(res);
 795 }
 796
 797 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 798 {
 799         LASSERT(res != NULL);
 800         LASSERT(res != LP_POISON);
 801         atomic_inc(&res->lr_refcount);
 802         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 803                atomic_read(&res->lr_refcount));
 804         return res;
 805 }
 806
 807 void __ldlm_resource_putref_final(struct ldlm_resource *res)
 808 {
 809         struct ldlm_namespace *ns = res->lr_namespace;
 810
 811         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 812
 813         if (!list_empty(&res->lr_granted)) {
 814                 ldlm_resource_dump(D_ERROR, res);
 815                 LBUG();
 816         }
 817
 818         if (!list_empty(&res->lr_converting)) {
 819                 ldlm_resource_dump(D_ERROR, res);
 820                 LBUG();
 821         }
 822
 823         if (!list_empty(&res->lr_waiting)) {
 824                 ldlm_resource_dump(D_ERROR, res);
 825                 LBUG();
 826         }
 827
 828         if (!list_empty(&res->lr_children)) {
 829                 ldlm_resource_dump(D_ERROR, res);
 830                 LBUG();
 831         }
 832
 833         /* Pass 0 here to not wake ->ns_waitq up yet, we will do it few
 834          * lines below when all children are freed. */
 835         ldlm_namespace_put_nolock(ns, 0);
 836         list_del_init(&res->lr_hash);
 837         list_del_init(&res->lr_childof);
 838
 839         ns->ns_resources--;
 840         if (ns->ns_resources == 0)
 841                 wake_up(&ns->ns_waitq);
 842 }
 843
 844 /* Returns 1 if the resource was freed, 0 if it remains. */
 845 int ldlm_resource_putref(struct ldlm_resource *res)
 846 {
 847         struct ldlm_namespace *ns = res->lr_namespace;
 848         int rc = 0;
 849         ENTRY;
 850
 851         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 852                atomic_read(&res->lr_refcount) - 1);
 853         LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
 854                  atomic_read(&res->lr_refcount));
 855         LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
 856                  atomic_read(&res->lr_refcount));
 857
 858         if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
 859                 __ldlm_resource_putref_final(res);
 860                 spin_unlock(&ns->ns_hash_lock);
 861                 if (res->lr_lvb_data)
 862                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 863                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 864                 rc = 1;
 865         }
 866
 867         RETURN(rc);
 868 }
 869
 870 /* Returns 1 if the resource was freed, 0 if it remains. */
 871 int ldlm_resource_putref_locked(struct ldlm_resource *res)
 872 {
 873         int rc = 0;
 874         ENTRY;
 875
 876         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 877                atomic_read(&res->lr_refcount) - 1);
 878         LASSERT(atomic_read(&res->lr_refcount) > 0);
 879         LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
 880
 881         LASSERT(atomic_read(&res->lr_refcount) >= 0);
 882         if (atomic_dec_and_test(&res->lr_refcount)) {
 883                 __ldlm_resource_putref_final(res);
 884                 if (res->lr_lvb_data)
 885                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 886                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 887                 rc = 1;
 888         }
 889
 890         RETURN(rc);
 891 }
 892
 893 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 894                             struct ldlm_lock *lock)
 895 {
 896         check_res_locked(res);
 897
 898         ldlm_resource_dump(D_OTHER, res);
 899         CDEBUG(D_OTHER, "About to add this lock:\n");
 900         ldlm_lock_dump(D_OTHER, lock, 0);
 901
 902         if (lock->l_destroyed) {
 903                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 904                 return;
 905         }
 906
 907         LASSERT(list_empty(&lock->l_res_link));
 908
 909         list_add_tail(&lock->l_res_link, head);
 910 }
 911
 912 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
 913                                      struct ldlm_lock *new)
 914 {
 915         struct ldlm_resource *res = original->l_resource;
 916
 917         check_res_locked(res);
 918
 919         ldlm_resource_dump(D_OTHER, res);
 920         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
 921         ldlm_lock_dump(D_OTHER, new, 0);
 922
 923         if (new->l_destroyed) {
 924                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 925                 goto out;
 926         }
 927
 928         LASSERT(list_empty(&new->l_res_link));
 929
 930         list_add(&new->l_res_link, &original->l_res_link);
 931  out:;
 932 }
 933
 934 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 935 {
 936         int type = lock->l_resource->lr_type;
 937
 938         check_res_locked(lock->l_resource);
 939         if (type == LDLM_IBITS || type == LDLM_PLAIN)
 940                 ldlm_unlink_lock_skiplist(lock);
 941         else if (type == LDLM_EXTENT)
 942                 ldlm_extent_unlink_lock(lock);
 943         list_del_init(&lock->l_res_link);
 944 }
 945
 946 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 947 {
 948         desc->lr_type = res->lr_type;
 949         desc->lr_name = res->lr_name;
 950 }
 951
 952 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
 953 {
 954         struct list_head *tmp;
 955
 956         if (!((libcfs_debug | D_ERROR) & level))
 957                 return;
 958
 959         mutex_down(ldlm_namespace_lock(client));
 960
 961         list_for_each(tmp, ldlm_namespace_list(client)) {
 962                 struct ldlm_namespace *ns;
 963                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 964                 ldlm_namespace_dump(level, ns);
 965         }
 966
 967         mutex_up(ldlm_namespace_lock(client));
 968 }
 969
 970 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
 971 {
 972         struct list_head *tmp;
 973
 974         if (!((libcfs_debug | D_ERROR) & level))
 975                 return;
 976
 977         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
 978                ns->ns_name, ns->ns_refcount,
 979                ns_is_client(ns) ? "client" : "server");
 980
 981         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
 982                 return;
 983
 984         spin_lock(&ns->ns_hash_lock);
 985         tmp = ns->ns_root_list.next;
 986         while (tmp != &ns->ns_root_list) {
 987                 struct ldlm_resource *res;
 988                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
 989
 990                 ldlm_resource_getref(res);
 991                 spin_unlock(&ns->ns_hash_lock);
 992
 993                 lock_res(res);
 994                 ldlm_resource_dump(level, res);
 995                 unlock_res(res);
 996
 997                 spin_lock(&ns->ns_hash_lock);
 998                 tmp = tmp->next;
 999                 ldlm_resource_putref_locked(res);
1000         }
1001         ns->ns_next_dump = cfs_time_shift(10);
1002         spin_unlock(&ns->ns_hash_lock);
1003 }
1004
1005 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1006 {
1007         struct list_head *tmp;
1008         int pos;
1009
1010         CLASSERT(RES_NAME_SIZE == 4);
1011
1012         if (!((libcfs_debug | D_ERROR) & level))
1013                 return;
1014
1015         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1016                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1017                res->lr_name.name[2], res->lr_name.name[3],
1018                atomic_read(&res->lr_refcount));
1019
1020         if (!list_empty(&res->lr_granted)) {
1021                 pos = 0;
1022                 CDEBUG(level, "Granted locks:\n");
1023                 list_for_each(tmp, &res->lr_granted) {
1024                         struct ldlm_lock *lock;
1025                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1026                         ldlm_lock_dump(level, lock, ++pos);
1027                 }
1028         }
1029         if (!list_empty(&res->lr_converting)) {
1030                 pos = 0;
1031                 CDEBUG(level, "Converting locks:\n");
1032                 list_for_each(tmp, &res->lr_converting) {
1033                         struct ldlm_lock *lock;
1034                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1035                         ldlm_lock_dump(level, lock, ++pos);
1036                 }
1037         }
1038         if (!list_empty(&res->lr_waiting)) {
1039                 pos = 0;
1040                 CDEBUG(level, "Waiting locks:\n");
1041                 list_for_each(tmp, &res->lr_waiting) {
1042                         struct ldlm_lock *lock;
1043                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1044                         ldlm_lock_dump(level, lock, ++pos);
1045                 }
1046         }
1047 }