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(char *name, ldlm_side_t client,
 299                                           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         namelen = strlen(name);
 323         OBD_ALLOC(ns->ns_name, namelen + 1);
 324         if (!ns->ns_name)
 325                 GOTO(out_hash, NULL);
 326
 327         strcpy(ns->ns_name, name);
 328
 329         CFS_INIT_LIST_HEAD(&ns->ns_root_list);
 330         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 331         ns->ns_refcount = 0;
 332         ns->ns_client = client;
 333         spin_lock_init(&ns->ns_hash_lock);
 334         atomic_set(&ns->ns_locks, 0);
 335         ns->ns_resources = 0;
 336         cfs_waitq_init(&ns->ns_waitq);
 337         ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
 338         ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
 339         ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
 340
 341         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 342              bucket--)
 343                 CFS_INIT_LIST_HEAD(bucket);
 344
 345         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 346         ns->ns_nr_unused = 0;
 347         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
 348         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
 349         ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
 350         spin_lock_init(&ns->ns_unused_lock);
 351         ns->ns_orig_connect_flags = 0;
 352         ns->ns_connect_flags = 0;
 353         ldlm_proc_namespace(ns);
 354
 355         idx = atomic_read(ldlm_namespace_nr(client));
 356         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 357         if (rc) {
 358                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 359                 GOTO(out_proc, rc);
 360         }
 361
 362         ldlm_namespace_register(ns, client);
 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();
 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 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 502 {
 503         ENTRY;
 504
 505         /* At shutdown time, don't call the cancellation callback */
 506         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 507
 508         if (ns->ns_refcount > 0) {
 509                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 510                 int rc;
 511                 CDEBUG(D_DLMTRACE,
 512                        "dlm namespace %s free waiting on refcount %d\n",
 513                        ns->ns_name, ns->ns_refcount);
 514 force_wait:
 515                 if (force)
 516                         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 517
 518                 rc = l_wait_event(ns->ns_waitq,
 519                                   ns->ns_refcount == 0, &lwi);
 520
 521                 /* Forced cleanups should be able to reclaim all references,
 522                  * so it's safe to wait forever... we can't leak locks... */
 523                 if (force && rc == -ETIMEDOUT) {
 524                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 525                                        "namespace with %d resources in use, "
 526                                        "(rc=%d)\n", ns->ns_name,
 527                                        ns->ns_refcount, rc);
 528                         GOTO(force_wait, rc);
 529                 }
 530
 531                 if (ns->ns_refcount) {
 532                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 533                                        "with %d resources in use, (rc=%d)\n",
 534                                        ns->ns_name,
 535                                        ns->ns_refcount, rc);
 536                         RETURN(ELDLM_NAMESPACE_EXISTS);
 537                 }
 538                 CDEBUG(D_DLMTRACE,
 539                        "dlm namespace %s free done waiting\n", ns->ns_name);
 540         }
 541
 542         RETURN(ELDLM_OK);
 543 }
 544
 545 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 546                                struct obd_import *imp,
 547                                int force)
 548 {
 549         int rc;
 550         ENTRY;
 551         if (!ns) {
 552                 EXIT;
 553                 return;
 554         }
 555
 556         /* Remove @ns from list. */
 557         ldlm_namespace_unregister(ns, ns->ns_client);
 558
 559         /* Can fail with -EINTR when force == 0 in which case try harder */
 560         rc = __ldlm_namespace_free(ns, force);
 561         if (rc != ELDLM_OK) {
 562                 if (imp) {
 563                         ptlrpc_disconnect_import(imp, 0);
 564                         ptlrpc_invalidate_import(imp);
 565                 }
 566
 567                 /* With all requests dropped and the import inactive
 568                  * we are gaurenteed all reference will be dropped. */
 569                 rc = __ldlm_namespace_free(ns, 1);
 570                 LASSERT(rc == 0);
 571         }
 572         EXIT;
 573 }
 574
 575 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 576 {
 577         ENTRY;
 578         if (!ns) {
 579                 EXIT;
 580                 return;
 581         }
 582
 583         /*
 584          * Fini pool _before_ parent proc dir is removed. This is important as
 585          * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
 586          * it after @dir may cause oops.
 587          */
 588         ldlm_pool_fini(&ns->ns_pool);
 589 #ifdef LPROCFS
 590         {
 591                 struct proc_dir_entry *dir;
 592                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
 593                 if (dir == NULL) {
 594                         CERROR("dlm namespace %s has no procfs dir?\n",
 595                                ns->ns_name);
 596                 } else {
 597                         lprocfs_remove(&dir);
 598                 }
 599         }
 600 #endif
 601
 602         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 603         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 604         /*
 605          * @ns should be not on list in this time, otherwise this will cause
 606          * issues realted to using freed @ns in pools thread.
 607          */
 608         LASSERT(list_empty(&ns->ns_list_chain));
 609         OBD_FREE_PTR(ns);
 610         ldlm_put_ref();
 611         EXIT;
 612 }
 613
 614
 615 /* Cleanup the resource, and free namespace.
 616  * bug 12864:
 617  * Deadlock issue:
 618  * proc1: destroy import
 619  *        class_disconnect_export(grab cl_sem) ->
 620  *              -> ldlm_namespace_free ->
 621  *              -> lprocfs_remove(grab _lprocfs_lock).
 622  * proc2: read proc info
 623  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 624  *              -> osc_rd_active, etc(grab cl_sem).
 625  *
 626  * So that I have to split the ldlm_namespace_free into two parts - the first
 627  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 628  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 629  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 630  * held.
 631  */
 632 void ldlm_namespace_free(struct ldlm_namespace *ns,
 633                          struct obd_import *imp,
 634                          int force)
 635 {
 636         ldlm_namespace_free_prior(ns, imp, force);
 637         ldlm_namespace_free_post(ns);
 638 }
 639
 640
 641 void ldlm_namespace_get_locked(struct ldlm_namespace *ns)
 642 {
 643         LASSERT(ns->ns_refcount >= 0);
 644         ns->ns_refcount++;
 645 }
 646
 647 void ldlm_namespace_get(struct ldlm_namespace *ns)
 648 {
 649         spin_lock(&ns->ns_hash_lock);
 650         ldlm_namespace_get_locked(ns);
 651         spin_unlock(&ns->ns_hash_lock);
 652 }
 653
 654 void ldlm_namespace_put_locked(struct ldlm_namespace *ns, int wakeup)
 655 {
 656         LASSERT(ns->ns_refcount > 0);
 657         ns->ns_refcount--;
 658         if (ns->ns_refcount == 0 && wakeup)
 659                 wake_up(&ns->ns_waitq);
 660 }
 661
 662 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
 663 {
 664         spin_lock(&ns->ns_hash_lock);
 665         ldlm_namespace_put_locked(ns, wakeup);
 666         spin_unlock(&ns->ns_hash_lock);
 667 }
 668
 669 /* Register @ns in the list of namespaces */
 670 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
 671 {
 672         mutex_down(ldlm_namespace_lock(client));
 673         LASSERT(list_empty(&ns->ns_list_chain));
 674         list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 675         atomic_inc(ldlm_namespace_nr(client));
 676         mutex_up(ldlm_namespace_lock(client));
 677 }
 678
 679 /* Unregister @ns from the list of namespaces */
 680 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
 681 {
 682         mutex_down(ldlm_namespace_lock(client));
 683         LASSERT(!list_empty(&ns->ns_list_chain));
 684         /*
 685          * Some asserts and possibly other parts of code still using
 686          * list_empty(&ns->ns_list_chain). This is why it is important
 687          * to use list_del_init() here.
 688          */
 689         list_del_init(&ns->ns_list_chain);
 690         atomic_dec(ldlm_namespace_nr(client));
 691         mutex_up(ldlm_namespace_lock(client));
 692 }
 693
 694 /* Should be called under ldlm_namespace_lock(client) taken */
 695 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
 696 {
 697         LASSERT(!list_empty(&ns->ns_list_chain));
 698         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 699         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 700 }
 701
 702 /* Should be called under ldlm_namespace_lock(client) taken */
 703 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
 704 {
 705         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 706         LASSERT(!list_empty(ldlm_namespace_list(client)));
 707         return container_of(ldlm_namespace_list(client)->next,
 708                 struct ldlm_namespace, ns_list_chain);
 709 }
 710 static __u32 ldlm_hash_fn(struct ldlm_resource *parent,
 711                           const struct ldlm_res_id *name)
 712 {
 713         __u32 hash = 0;
 714         int i;
 715
 716         for (i = 0; i < RES_NAME_SIZE; i++)
 717                 hash += name->name[i];
 718
 719         hash += (__u32)((unsigned long)parent >> 4);
 720
 721         return (hash & RES_HASH_MASK);
 722 }
 723
 724 static struct ldlm_resource *ldlm_resource_new(void)
 725 {
 726         struct ldlm_resource *res;
 727         int idx;
 728
 729         OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
 730         if (res == NULL)
 731                 return NULL;
 732
 733         memset(res, 0, sizeof(*res));
 734
 735         CFS_INIT_LIST_HEAD(&res->lr_children);
 736         CFS_INIT_LIST_HEAD(&res->lr_childof);
 737         CFS_INIT_LIST_HEAD(&res->lr_granted);
 738         CFS_INIT_LIST_HEAD(&res->lr_converting);
 739         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 740
 741         /* initialize interval trees for each lock mode*/
 742         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 743                 res->lr_itree[idx].lit_size = 0;
 744                 res->lr_itree[idx].lit_mode = 1 << idx;
 745                 res->lr_itree[idx].lit_root = NULL;
 746         }
 747
 748         atomic_set(&res->lr_refcount, 1);
 749         spin_lock_init(&res->lr_lock);
 750
 751         /* one who creates the resource must unlock
 752          * the semaphore after lvb initialization */
 753         init_MUTEX_LOCKED(&res->lr_lvb_sem);
 754
 755         return res;
 756 }
 757
 758 /* must be called with hash lock held */
 759 static struct ldlm_resource *
 760 ldlm_resource_find(struct ldlm_namespace *ns, const struct ldlm_res_id *name,
 761                    __u32 hash)
 762 {
 763         struct list_head *bucket, *tmp;
 764         struct ldlm_resource *res;
 765
 766         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 767         bucket = ns->ns_hash + hash;
 768
 769         list_for_each(tmp, bucket) {
 770                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 771                 if (memcmp(&res->lr_name, name, sizeof(res->lr_name)) == 0)
 772                         return res;
 773         }
 774
 775         return NULL;
 776 }
 777
 778 /* Args: locked namespace
 779  * Returns: newly-allocated, referenced, unlocked resource */
 780 static struct ldlm_resource *
 781 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 782                   const struct ldlm_res_id *name, __u32 hash, ldlm_type_t type)
 783 {
 784         struct list_head *bucket;
 785         struct ldlm_resource *res, *old_res;
 786         ENTRY;
 787
 788         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
 789                  "type: %d\n", type);
 790
 791         res = ldlm_resource_new();
 792         if (!res)
 793                 RETURN(NULL);
 794
 795         res->lr_name = *name;
 796         res->lr_namespace = ns;
 797         res->lr_type = type;
 798         res->lr_most_restr = LCK_NL;
 799
 800         spin_lock(&ns->ns_hash_lock);
 801         old_res = ldlm_resource_find(ns, name, hash);
 802         if (old_res) {
 803                 /* someone won the race and added the resource before */
 804                 ldlm_resource_getref(old_res);
 805                 spin_unlock(&ns->ns_hash_lock);
 806                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 807                 /* synchronize WRT resource creation */
 808                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 809                         down(&old_res->lr_lvb_sem);
 810                         up(&old_res->lr_lvb_sem);
 811                 }
 812                 RETURN(old_res);
 813         }
 814
 815         /* we won! let's add the resource */
 816         bucket = ns->ns_hash + hash;
 817         list_add(&res->lr_hash, bucket);
 818         ns->ns_resources++;
 819         ldlm_namespace_get_locked(ns);
 820
 821         if (parent == NULL) {
 822                 list_add(&res->lr_childof, &ns->ns_root_list);
 823         } else {
 824                 res->lr_parent = parent;
 825                 list_add(&res->lr_childof, &parent->lr_children);
 826         }
 827         spin_unlock(&ns->ns_hash_lock);
 828
 829         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 830                 int rc;
 831
 832                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
 833                 rc = ns->ns_lvbo->lvbo_init(res);
 834                 if (rc)
 835                         CERROR("lvbo_init failed for resource "
 836                                LPU64": rc %d\n", name->name[0], rc);
 837                 /* we create resource with locked lr_lvb_sem */
 838                 up(&res->lr_lvb_sem);
 839         }
 840
 841         RETURN(res);
 842 }
 843
 844 /* Args: unlocked namespace
 845  * Locks: takes and releases ns->ns_lock and res->lr_lock
 846  * Returns: referenced, unlocked ldlm_resource or NULL */
 847 struct ldlm_resource *
 848 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 849                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
 850 {
 851         __u32 hash = ldlm_hash_fn(parent, name);
 852         struct ldlm_resource *res = NULL;
 853         ENTRY;
 854
 855         LASSERT(ns != NULL);
 856         LASSERT(ns->ns_hash != NULL);
 857         LASSERT(name->name[0] != 0);
 858
 859         spin_lock(&ns->ns_hash_lock);
 860         res = ldlm_resource_find(ns, name, hash);
 861         if (res) {
 862                 ldlm_resource_getref(res);
 863                 spin_unlock(&ns->ns_hash_lock);
 864                 /* synchronize WRT resource creation */
 865                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 866                         down(&res->lr_lvb_sem);
 867                         up(&res->lr_lvb_sem);
 868                 }
 869                 RETURN(res);
 870         }
 871         spin_unlock(&ns->ns_hash_lock);
 872
 873         if (create == 0)
 874                 RETURN(NULL);
 875
 876         res = ldlm_resource_add(ns, parent, name, hash, type);
 877         RETURN(res);
 878 }
 879
 880 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 881 {
 882         LASSERT(res != NULL);
 883         LASSERT(res != LP_POISON);
 884         atomic_inc(&res->lr_refcount);
 885         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 886                atomic_read(&res->lr_refcount));
 887         return res;
 888 }
 889
 890 void __ldlm_resource_putref_final(struct ldlm_resource *res)
 891 {
 892         struct ldlm_namespace *ns = res->lr_namespace;
 893
 894         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 895
 896         if (!list_empty(&res->lr_granted)) {
 897                 ldlm_resource_dump(D_ERROR, res);
 898                 LBUG();
 899         }
 900
 901         if (!list_empty(&res->lr_converting)) {
 902                 ldlm_resource_dump(D_ERROR, res);
 903                 LBUG();
 904         }
 905
 906         if (!list_empty(&res->lr_waiting)) {
 907                 ldlm_resource_dump(D_ERROR, res);
 908                 LBUG();
 909         }
 910
 911         if (!list_empty(&res->lr_children)) {
 912                 ldlm_resource_dump(D_ERROR, res);
 913                 LBUG();
 914         }
 915
 916         /* Pass 0 here to not wake ->ns_waitq up yet, we will do it few
 917          * lines below when all children are freed. */
 918         ldlm_namespace_put_locked(ns, 0);
 919         list_del_init(&res->lr_hash);
 920         list_del_init(&res->lr_childof);
 921
 922         ns->ns_resources--;
 923         if (ns->ns_resources == 0)
 924                 wake_up(&ns->ns_waitq);
 925 }
 926
 927 /* Returns 1 if the resource was freed, 0 if it remains. */
 928 int ldlm_resource_putref(struct ldlm_resource *res)
 929 {
 930         struct ldlm_namespace *ns = res->lr_namespace;
 931         int rc = 0;
 932         ENTRY;
 933
 934         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 935                atomic_read(&res->lr_refcount) - 1);
 936         LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
 937                  atomic_read(&res->lr_refcount));
 938         LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
 939                  atomic_read(&res->lr_refcount));
 940
 941         if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
 942                 __ldlm_resource_putref_final(res);
 943                 spin_unlock(&ns->ns_hash_lock);
 944                 if (res->lr_lvb_data)
 945                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 946                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 947                 rc = 1;
 948         }
 949
 950         RETURN(rc);
 951 }
 952
 953 /* Returns 1 if the resource was freed, 0 if it remains. */
 954 int ldlm_resource_putref_locked(struct ldlm_resource *res)
 955 {
 956         int rc = 0;
 957         ENTRY;
 958
 959         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 960                atomic_read(&res->lr_refcount) - 1);
 961         LASSERT(atomic_read(&res->lr_refcount) > 0);
 962         LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
 963
 964         LASSERT(atomic_read(&res->lr_refcount) >= 0);
 965         if (atomic_dec_and_test(&res->lr_refcount)) {
 966                 __ldlm_resource_putref_final(res);
 967                 if (res->lr_lvb_data)
 968                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 969                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 970                 rc = 1;
 971         }
 972
 973         RETURN(rc);
 974 }
 975
 976 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 977                             struct ldlm_lock *lock)
 978 {
 979         check_res_locked(res);
 980
 981         ldlm_resource_dump(D_OTHER, res);
 982         CDEBUG(D_OTHER, "About to add this lock:\n");
 983         ldlm_lock_dump(D_OTHER, lock, 0);
 984
 985         if (lock->l_destroyed) {
 986                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 987                 return;
 988         }
 989
 990         LASSERT(list_empty(&lock->l_res_link));
 991
 992         list_add_tail(&lock->l_res_link, head);
 993 }
 994
 995 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
 996                                      struct ldlm_lock *new)
 997 {
 998         struct ldlm_resource *res = original->l_resource;
 999
1000         check_res_locked(res);
1001
1002         ldlm_resource_dump(D_OTHER, res);
1003         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1004         ldlm_lock_dump(D_OTHER, new, 0);
1005
1006         if (new->l_destroyed) {
1007                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1008                 goto out;
1009         }
1010
1011         LASSERT(list_empty(&new->l_res_link));
1012
1013         list_add(&new->l_res_link, &original->l_res_link);
1014  out:;
1015 }
1016
1017 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1018 {
1019         int type = lock->l_resource->lr_type;
1020
1021         check_res_locked(lock->l_resource);
1022         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1023                 ldlm_unlink_lock_skiplist(lock);
1024         else if (type == LDLM_EXTENT)
1025                 ldlm_extent_unlink_lock(lock);
1026         list_del_init(&lock->l_res_link);
1027 }
1028
1029 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1030 {
1031         desc->lr_type = res->lr_type;
1032         desc->lr_name = res->lr_name;
1033 }
1034
1035 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1036 {
1037         struct list_head *tmp;
1038
1039         if (!((libcfs_debug | D_ERROR) & level))
1040                 return;
1041
1042         mutex_down(ldlm_namespace_lock(client));
1043
1044         list_for_each(tmp, ldlm_namespace_list(client)) {
1045                 struct ldlm_namespace *ns;
1046                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1047                 ldlm_namespace_dump(level, ns);
1048         }
1049
1050         mutex_up(ldlm_namespace_lock(client));
1051 }
1052
1053 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1054 {
1055         struct list_head *tmp;
1056
1057         if (!((libcfs_debug | D_ERROR) & level))
1058                 return;
1059
1060         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1061                ns->ns_name, ns->ns_refcount,
1062                ns_is_client(ns) ? "client" : "server");
1063
1064         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1065                 return;
1066
1067         spin_lock(&ns->ns_hash_lock);
1068         tmp = ns->ns_root_list.next;
1069         while (tmp != &ns->ns_root_list) {
1070                 struct ldlm_resource *res;
1071                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
1072
1073                 ldlm_resource_getref(res);
1074                 spin_unlock(&ns->ns_hash_lock);
1075
1076                 lock_res(res);
1077                 ldlm_resource_dump(level, res);
1078                 unlock_res(res);
1079
1080                 spin_lock(&ns->ns_hash_lock);
1081                 tmp = tmp->next;
1082                 ldlm_resource_putref_locked(res);
1083         }
1084         ns->ns_next_dump = cfs_time_shift(10);
1085         spin_unlock(&ns->ns_hash_lock);
1086 }
1087
1088 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1089 {
1090         struct list_head *tmp;
1091         int pos;
1092
1093         CLASSERT(RES_NAME_SIZE == 4);
1094
1095         if (!((libcfs_debug | D_ERROR) & level))
1096                 return;
1097
1098         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1099                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1100                res->lr_name.name[2], res->lr_name.name[3],
1101                atomic_read(&res->lr_refcount));
1102
1103         if (!list_empty(&res->lr_granted)) {
1104                 pos = 0;
1105                 CDEBUG(level, "Granted locks:\n");
1106                 list_for_each(tmp, &res->lr_granted) {
1107                         struct ldlm_lock *lock;
1108                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1109                         ldlm_lock_dump(level, lock, ++pos);
1110                 }
1111         }
1112         if (!list_empty(&res->lr_converting)) {
1113                 pos = 0;
1114                 CDEBUG(level, "Converting locks:\n");
1115                 list_for_each(tmp, &res->lr_converting) {
1116                         struct ldlm_lock *lock;
1117                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1118                         ldlm_lock_dump(level, lock, ++pos);
1119                 }
1120         }
1121         if (!list_empty(&res->lr_waiting)) {
1122                 pos = 0;
1123                 CDEBUG(level, "Waiting locks:\n");
1124                 list_for_each(tmp, &res->lr_waiting) {
1125                         struct ldlm_lock *lock;
1126                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1127                         ldlm_lock_dump(level, lock, ++pos);
1128                 }
1129         }
1130 }