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