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 (strncmp(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,
 186                        "changing namespace %s unused locks from %u to %u\n",
 187                        ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
 188                 ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 189
 190                 if (!lru_resize) {
 191                         CDEBUG(D_DLMTRACE,
 192                                "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,
 198                        "changing namespace %s max_unused from %u to %u\n",
 199                        ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
 200                 ns->ns_max_unused = (unsigned int)tmp;
 201                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 202
 203                 /* Make sure that originally lru resize was supported before
 204                  * turning it on here. */
 205                 if (lru_resize &&
 206                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 207                         CDEBUG(D_DLMTRACE,
 208                                "enable lru_resize for namespace %s\n",
 209                                ns->ns_name);
 210                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 211                 }
 212         }
 213
 214         return count;
 215 }
 216
 217 void ldlm_proc_namespace(struct ldlm_namespace *ns)
 218 {
 219         struct lprocfs_vars lock_vars[2];
 220         char lock_name[MAX_STRING_SIZE + 1];
 221
 222         LASSERT(ns != NULL);
 223         LASSERT(ns->ns_name != NULL);
 224
 225         lock_name[MAX_STRING_SIZE] = '\0';
 226
 227         memset(lock_vars, 0, sizeof(lock_vars));
 228         lock_vars[0].name = lock_name;
 229
 230         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
 231         lock_vars[0].data = &ns->ns_refcount;
 232         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 233         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 234
 235         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
 236         lock_vars[0].data = &ns->ns_locks;
 237         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 238         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 239
 240         if (ns_is_client(ns)) {
 241                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 242                          ns->ns_name);
 243                 lock_vars[0].data = &ns->ns_nr_unused;
 244                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 245                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 246
 247                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 248                          ns->ns_name);
 249                 lock_vars[0].data = ns;
 250                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 251                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 252                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 253
 254                 snprintf(lock_name, MAX_STRING_SIZE, "%s/shrink_thumb",
 255                          ns->ns_name);
 256                 lock_vars[0].data = ns;
 257                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 258                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 259                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 260
 261                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 262                          ns->ns_name);
 263                 lock_vars[0].data = &ns->ns_max_age;
 264                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 265                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 266                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 267         } else {
 268                 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
 269                          ns->ns_name);
 270                 lock_vars[0].data = &ns->ns_ctime_age_limit;
 271                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 272                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 273                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 274
 275                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 276                          ns->ns_name);
 277                 lock_vars[0].data = &ns->ns_max_nolock_size;
 278                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 279                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 280                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 281
 282                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 283                          ns->ns_name);
 284                 lock_vars[0].data = &ns->ns_contention_time;
 285                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 286                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 287                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 288
 289                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 290                          ns->ns_name);
 291                 lock_vars[0].data = &ns->ns_contended_locks;
 292                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 293                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 294                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 295         }
 296 }
 297 #undef MAX_STRING_SIZE
 298 #else
 299 #define ldlm_proc_namespace(ns) do {} while (0)
 300 #endif /* LPROCFS */
 301
 302 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 303                                           ldlm_side_t client, ldlm_appetite_t apt)
 304 {
 305         struct ldlm_namespace *ns = NULL;
 306         struct list_head *bucket;
 307         int rc, idx, namelen;
 308         ENTRY;
 309
 310         rc = ldlm_get_ref();
 311         if (rc) {
 312                 CERROR("ldlm_get_ref failed: %d\n", rc);
 313                 RETURN(NULL);
 314         }
 315
 316         OBD_ALLOC_PTR(ns);
 317         if (!ns)
 318                 GOTO(out_ref, NULL);
 319
 320         OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 321         if (!ns->ns_hash)
 322                 GOTO(out_ns, NULL);
 323
 324         ns->ns_shrink_thumb = LDLM_LOCK_SHRINK_THUMB;
 325         ns->ns_appetite = apt;
 326
 327         LASSERT(obd != NULL);
 328         ns->ns_obd = obd;
 329
 330         namelen = strlen(name);
 331         OBD_ALLOC(ns->ns_name, namelen + 1);
 332         if (!ns->ns_name)
 333                 GOTO(out_hash, NULL);
 334
 335         strcpy(ns->ns_name, name);
 336
 337         CFS_INIT_LIST_HEAD(&ns->ns_root_list);
 338         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 339         ns->ns_refcount = 0;
 340         ns->ns_client = client;
 341         spin_lock_init(&ns->ns_hash_lock);
 342         atomic_set(&ns->ns_locks, 0);
 343         ns->ns_resources = 0;
 344         cfs_waitq_init(&ns->ns_waitq);
 345         ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
 346         ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
 347         ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
 348
 349         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 350              bucket--)
 351                 CFS_INIT_LIST_HEAD(bucket);
 352
 353         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 354         ns->ns_nr_unused = 0;
 355         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
 356         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
 357         ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
 358         spin_lock_init(&ns->ns_unused_lock);
 359         ns->ns_orig_connect_flags = 0;
 360         ns->ns_connect_flags = 0;
 361         ldlm_proc_namespace(ns);
 362
 363         idx = atomic_read(ldlm_namespace_nr(client));
 364         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 365         if (rc) {
 366                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 367                 GOTO(out_proc, rc);
 368         }
 369
 370         ldlm_namespace_register(ns, client);
 371         RETURN(ns);
 372 out_proc:
 373         ldlm_namespace_cleanup(ns, 0);
 374         OBD_FREE(ns->ns_name, namelen + 1);
 375 out_hash:
 376         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 377 out_ns:
 378         OBD_FREE_PTR(ns);
 379 out_ref:
 380         ldlm_put_ref();
 381         RETURN(NULL);
 382 }
 383
 384 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 385
 386 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 387  * This is currently only used for recovery, and we make certain assumptions
 388  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 389  *
 390  * Called with the ns_lock held. */
 391 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 392                              int flags)
 393 {
 394         struct list_head *tmp;
 395         int rc = 0, client = ns_is_client(res->lr_namespace);
 396         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 397         ENTRY;
 398
 399
 400         do {
 401                 struct ldlm_lock *lock = NULL;
 402
 403                 /* first, we look for non-cleaned-yet lock
 404                  * all cleaned locks are marked by CLEANED flag */
 405                 lock_res(res);
 406                 list_for_each(tmp, q) {
 407                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 408                         if (lock->l_flags & LDLM_FL_CLEANED) {
 409                                 lock = NULL;
 410                                 continue;
 411                         }
 412                         LDLM_LOCK_GET(lock);
 413                         lock->l_flags |= LDLM_FL_CLEANED;
 414                         break;
 415                 }
 416
 417                 if (lock == NULL) {
 418                         unlock_res(res);
 419                         break;
 420                 }
 421
 422                 /* Set CBPENDING so nothing in the cancellation path
 423                  * can match this lock */
 424                 lock->l_flags |= LDLM_FL_CBPENDING;
 425                 lock->l_flags |= LDLM_FL_FAILED;
 426                 lock->l_flags |= flags;
 427
 428                 /* ... without sending a CANCEL message for local_only. */
 429                 if (local_only)
 430                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 431
 432                 if (local_only && (lock->l_readers || lock->l_writers)) {
 433                         /* This is a little bit gross, but much better than the
 434                          * alternative: pretend that we got a blocking AST from
 435                          * the server, so that when the lock is decref'd, it
 436                          * will go away ... */
 437                         unlock_res(res);
 438                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 439                         if (lock->l_completion_ast)
 440                                 lock->l_completion_ast(lock, 0, NULL);
 441                         LDLM_LOCK_PUT(lock);
 442                         continue;
 443                 }
 444
 445                 if (client) {
 446                         struct lustre_handle lockh;
 447
 448                         unlock_res(res);
 449                         ldlm_lock2handle(lock, &lockh);
 450                         rc = ldlm_cli_cancel(&lockh);
 451                         if (rc)
 452                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 453                 } else {
 454                         ldlm_resource_unlink_lock(lock);
 455                         unlock_res(res);
 456                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 457                                    "client node");
 458                         ldlm_lock_destroy(lock);
 459                 }
 460                 LDLM_LOCK_PUT(lock);
 461         } while (1);
 462
 463         EXIT;
 464 }
 465
 466 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 467 {
 468         struct list_head *tmp;
 469         int i;
 470
 471         if (ns == NULL) {
 472                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 473                 return ELDLM_OK;
 474         }
 475
 476         for (i = 0; i < RES_HASH_SIZE; i++) {
 477                 spin_lock(&ns->ns_hash_lock);
 478                 tmp = ns->ns_hash[i].next;
 479                 while (tmp != &(ns->ns_hash[i])) {
 480                         struct ldlm_resource *res;
 481                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 482                         ldlm_resource_getref(res);
 483                         spin_unlock(&ns->ns_hash_lock);
 484
 485                         cleanup_resource(res, &res->lr_granted, flags);
 486                         cleanup_resource(res, &res->lr_converting, flags);
 487                         cleanup_resource(res, &res->lr_waiting, flags);
 488
 489                         spin_lock(&ns->ns_hash_lock);
 490                         tmp  = tmp->next;
 491
 492                         /* XXX: former stuff caused issues in case of race
 493                          * between ldlm_namespace_cleanup() and lockd() when
 494                          * client gets blocking ast when lock gets distracted by
 495                          * server. This is 1_4 branch solution, let's see how
 496                          * will it behave. */
 497                         if (!ldlm_resource_putref_locked(res))
 498                                 CDEBUG(D_INFO,
 499                                        "Namespace %s resource refcount nonzero "
 500                                        "(%d) after lock cleanup; forcing cleanup.\n",
 501                                        ns->ns_name, atomic_read(&res->lr_refcount));
 502                 }
 503                 spin_unlock(&ns->ns_hash_lock);
 504         }
 505
 506         return ELDLM_OK;
 507 }
 508
 509 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 510 {
 511         ENTRY;
 512
 513         /* At shutdown time, don't call the cancellation callback */
 514         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 515
 516         if (ns->ns_refcount > 0) {
 517                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 518                 int rc;
 519                 CDEBUG(D_DLMTRACE,
 520                        "dlm namespace %s free waiting on refcount %d\n",
 521                        ns->ns_name, ns->ns_refcount);
 522 force_wait:
 523                 if (force)
 524                         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 525
 526                 rc = l_wait_event(ns->ns_waitq,
 527                                   ns->ns_refcount == 0, &lwi);
 528
 529                 /* Forced cleanups should be able to reclaim all references,
 530                  * so it's safe to wait forever... we can't leak locks... */
 531                 if (force && rc == -ETIMEDOUT) {
 532                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 533                                        "namespace with %d resources in use, "
 534                                        "(rc=%d)\n", ns->ns_name,
 535                                        ns->ns_refcount, rc);
 536                         GOTO(force_wait, rc);
 537                 }
 538
 539                 if (ns->ns_refcount) {
 540                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 541                                        "with %d resources in use, (rc=%d)\n",
 542                                        ns->ns_name,
 543                                        ns->ns_refcount, rc);
 544                         RETURN(ELDLM_NAMESPACE_EXISTS);
 545                 }
 546                 CDEBUG(D_DLMTRACE,
 547                        "dlm namespace %s free done waiting\n", ns->ns_name);
 548         }
 549
 550         RETURN(ELDLM_OK);
 551 }
 552
 553 /**
 554  * Performs various cleanups for passed \a ns to make it drop refc and be ready
 555  * for freeing. Waits for refc == 0.
 556  *
 557  * The following is done:
 558  * (0) Unregister \a ns from its list to make inaccessible for potential users
 559  * like pools thread and others;
 560  * (1) Clear all locks in \a ns.
 561  */
 562 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 563                                struct obd_import *imp,
 564                                int force)
 565 {
 566         int rc;
 567         ENTRY;
 568         if (!ns) {
 569                 EXIT;
 570                 return;
 571         }
 572
 573         /*
 574          * Make sure that nobody can find this ns in its list.
 575          */
 576         ldlm_namespace_unregister(ns, ns->ns_client);
 577
 578         /*
 579          * Can fail with -EINTR when force == 0 in which case try harder.
 580          */
 581         rc = __ldlm_namespace_free(ns, force);
 582         if (rc != ELDLM_OK) {
 583                 if (imp) {
 584                         ptlrpc_disconnect_import(imp, 0);
 585                         ptlrpc_invalidate_import(imp);
 586                 }
 587
 588                 /*
 589                  * With all requests dropped and the import inactive
 590                  * we are gaurenteed all reference will be dropped.
 591                  */
 592                 rc = __ldlm_namespace_free(ns, 1);
 593                 LASSERT(rc == 0);
 594         }
 595         EXIT;
 596 }
 597
 598 /**
 599  * Performs freeing memory structures related to \a ns. This is only done when
 600  * ldlm_namespce_free_prior() successfully removed all resources referencing
 601  * \a ns and its refc == 0.
 602  */
 603 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 604 {
 605         ENTRY;
 606         if (!ns) {
 607                 EXIT;
 608                 return;
 609         }
 610
 611         /*
 612          * Fini pool _before_ parent proc dir is removed. This is important as
 613          * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
 614          * it after @dir may cause oops.
 615          */
 616         ldlm_pool_fini(&ns->ns_pool);
 617
 618 #ifdef LPROCFS
 619         {
 620                 struct proc_dir_entry *dir;
 621                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
 622                 if (dir == NULL) {
 623                         CERROR("dlm namespace %s has no procfs dir?\n",
 624                                ns->ns_name);
 625                 } else {
 626                         lprocfs_remove(&dir);
 627                 }
 628         }
 629 #endif
 630
 631         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 632         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 633
 634         /*
 635          * Namespace \a ns should be not on list in this time, otherwise this
 636          * will cause issues realted to using freed \a ns in pools thread.
 637          */
 638         LASSERT(list_empty(&ns->ns_list_chain));
 639         OBD_FREE_PTR(ns);
 640         ldlm_put_ref();
 641         EXIT;
 642 }
 643
 644
 645 /* Cleanup the resource, and free namespace.
 646  * bug 12864:
 647  * Deadlock issue:
 648  * proc1: destroy import
 649  *        class_disconnect_export(grab cl_sem) ->
 650  *              -> ldlm_namespace_free ->
 651  *              -> lprocfs_remove(grab _lprocfs_lock).
 652  * proc2: read proc info
 653  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 654  *              -> osc_rd_active, etc(grab cl_sem).
 655  *
 656  * So that I have to split the ldlm_namespace_free into two parts - the first
 657  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 658  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 659  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 660  * held.
 661  */
 662 void ldlm_namespace_free(struct ldlm_namespace *ns,
 663                          struct obd_import *imp,
 664                          int force)
 665 {
 666         ldlm_namespace_free_prior(ns, imp, force);
 667         ldlm_namespace_free_post(ns);
 668 }
 669
 670
 671 void ldlm_namespace_get_locked(struct ldlm_namespace *ns)
 672 {
 673         LASSERT(ns->ns_refcount >= 0);
 674         ns->ns_refcount++;
 675 }
 676
 677 void ldlm_namespace_get(struct ldlm_namespace *ns)
 678 {
 679         spin_lock(&ns->ns_hash_lock);
 680         ldlm_namespace_get_locked(ns);
 681         spin_unlock(&ns->ns_hash_lock);
 682 }
 683
 684 void ldlm_namespace_put_locked(struct ldlm_namespace *ns, int wakeup)
 685 {
 686         LASSERT(ns->ns_refcount > 0);
 687         ns->ns_refcount--;
 688         if (ns->ns_refcount == 0 && wakeup)
 689                 wake_up(&ns->ns_waitq);
 690 }
 691
 692 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
 693 {
 694         spin_lock(&ns->ns_hash_lock);
 695         ldlm_namespace_put_locked(ns, wakeup);
 696         spin_unlock(&ns->ns_hash_lock);
 697 }
 698
 699 /* Register @ns in the list of namespaces */
 700 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
 701 {
 702         mutex_down(ldlm_namespace_lock(client));
 703         LASSERT(list_empty(&ns->ns_list_chain));
 704         list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 705         atomic_inc(ldlm_namespace_nr(client));
 706         mutex_up(ldlm_namespace_lock(client));
 707 }
 708
 709 /* Unregister @ns from the list of namespaces */
 710 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
 711 {
 712         mutex_down(ldlm_namespace_lock(client));
 713         LASSERT(!list_empty(&ns->ns_list_chain));
 714         /*
 715          * Some asserts and possibly other parts of code still using
 716          * list_empty(&ns->ns_list_chain). This is why it is important
 717          * to use list_del_init() here.
 718          */
 719         list_del_init(&ns->ns_list_chain);
 720         atomic_dec(ldlm_namespace_nr(client));
 721         mutex_up(ldlm_namespace_lock(client));
 722 }
 723
 724 /* Should be called under ldlm_namespace_lock(client) taken */
 725 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
 726 {
 727         LASSERT(!list_empty(&ns->ns_list_chain));
 728         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 729         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 730 }
 731
 732 /* Should be called under ldlm_namespace_lock(client) taken */
 733 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
 734 {
 735         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 736         LASSERT(!list_empty(ldlm_namespace_list(client)));
 737         return container_of(ldlm_namespace_list(client)->next,
 738                 struct ldlm_namespace, ns_list_chain);
 739 }
 740 static __u32 ldlm_hash_fn(struct ldlm_resource *parent,
 741                           const struct ldlm_res_id *name)
 742 {
 743         __u32 hash = 0;
 744         int i;
 745
 746         for (i = 0; i < RES_NAME_SIZE; i++)
 747                 hash += name->name[i];
 748
 749         hash += (__u32)((unsigned long)parent >> 4);
 750
 751         return (hash & RES_HASH_MASK);
 752 }
 753
 754 static struct ldlm_resource *ldlm_resource_new(void)
 755 {
 756         struct ldlm_resource *res;
 757         int idx;
 758
 759         OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
 760         if (res == NULL)
 761                 return NULL;
 762
 763         memset(res, 0, sizeof(*res));
 764
 765         CFS_INIT_LIST_HEAD(&res->lr_children);
 766         CFS_INIT_LIST_HEAD(&res->lr_childof);
 767         CFS_INIT_LIST_HEAD(&res->lr_granted);
 768         CFS_INIT_LIST_HEAD(&res->lr_converting);
 769         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 770
 771         /* initialize interval trees for each lock mode*/
 772         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 773                 res->lr_itree[idx].lit_size = 0;
 774                 res->lr_itree[idx].lit_mode = 1 << idx;
 775                 res->lr_itree[idx].lit_root = NULL;
 776         }
 777
 778         atomic_set(&res->lr_refcount, 1);
 779         spin_lock_init(&res->lr_lock);
 780
 781         /* one who creates the resource must unlock
 782          * the semaphore after lvb initialization */
 783         init_MUTEX_LOCKED(&res->lr_lvb_sem);
 784
 785         return res;
 786 }
 787
 788 /* must be called with hash lock held */
 789 static struct ldlm_resource *
 790 ldlm_resource_find(struct ldlm_namespace *ns, const struct ldlm_res_id *name,
 791                    __u32 hash)
 792 {
 793         struct list_head *bucket, *tmp;
 794         struct ldlm_resource *res;
 795
 796         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 797         bucket = ns->ns_hash + hash;
 798
 799         list_for_each(tmp, bucket) {
 800                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 801                 if (memcmp(&res->lr_name, name, sizeof(res->lr_name)) == 0)
 802                         return res;
 803         }
 804
 805         return NULL;
 806 }
 807
 808 /* Args: locked namespace
 809  * Returns: newly-allocated, referenced, unlocked resource */
 810 static struct ldlm_resource *
 811 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 812                   const struct ldlm_res_id *name, __u32 hash, ldlm_type_t type)
 813 {
 814         struct list_head *bucket;
 815         struct ldlm_resource *res, *old_res;
 816         ENTRY;
 817
 818         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
 819                  "type: %d\n", type);
 820
 821         res = ldlm_resource_new();
 822         if (!res)
 823                 RETURN(NULL);
 824
 825         res->lr_name = *name;
 826         res->lr_namespace = ns;
 827         res->lr_type = type;
 828         res->lr_most_restr = LCK_NL;
 829
 830         spin_lock(&ns->ns_hash_lock);
 831         old_res = ldlm_resource_find(ns, name, hash);
 832         if (old_res) {
 833                 /* someone won the race and added the resource before */
 834                 ldlm_resource_getref(old_res);
 835                 spin_unlock(&ns->ns_hash_lock);
 836                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 837                 /* synchronize WRT resource creation */
 838                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 839                         down(&old_res->lr_lvb_sem);
 840                         up(&old_res->lr_lvb_sem);
 841                 }
 842                 RETURN(old_res);
 843         }
 844
 845         /* we won! let's add the resource */
 846         bucket = ns->ns_hash + hash;
 847         list_add(&res->lr_hash, bucket);
 848         ns->ns_resources++;
 849         ldlm_namespace_get_locked(ns);
 850
 851         if (parent == NULL) {
 852                 list_add(&res->lr_childof, &ns->ns_root_list);
 853         } else {
 854                 res->lr_parent = parent;
 855                 list_add(&res->lr_childof, &parent->lr_children);
 856         }
 857         spin_unlock(&ns->ns_hash_lock);
 858
 859         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 860                 int rc;
 861
 862                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
 863                 rc = ns->ns_lvbo->lvbo_init(res);
 864                 if (rc)
 865                         CERROR("lvbo_init failed for resource "
 866                                LPU64": rc %d\n", name->name[0], rc);
 867                 /* we create resource with locked lr_lvb_sem */
 868                 up(&res->lr_lvb_sem);
 869         }
 870
 871         RETURN(res);
 872 }
 873
 874 /* Args: unlocked namespace
 875  * Locks: takes and releases ns->ns_lock and res->lr_lock
 876  * Returns: referenced, unlocked ldlm_resource or NULL */
 877 struct ldlm_resource *
 878 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 879                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
 880 {
 881         __u32 hash = ldlm_hash_fn(parent, name);
 882         struct ldlm_resource *res = NULL;
 883         ENTRY;
 884
 885         LASSERT(ns != NULL);
 886         LASSERT(ns->ns_hash != NULL);
 887         LASSERT(name->name[0] != 0);
 888
 889         spin_lock(&ns->ns_hash_lock);
 890         res = ldlm_resource_find(ns, name, hash);
 891         if (res) {
 892                 ldlm_resource_getref(res);
 893                 spin_unlock(&ns->ns_hash_lock);
 894                 /* synchronize WRT resource creation */
 895                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 896                         down(&res->lr_lvb_sem);
 897                         up(&res->lr_lvb_sem);
 898                 }
 899                 RETURN(res);
 900         }
 901         spin_unlock(&ns->ns_hash_lock);
 902
 903         if (create == 0)
 904                 RETURN(NULL);
 905
 906         res = ldlm_resource_add(ns, parent, name, hash, type);
 907         RETURN(res);
 908 }
 909
 910 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 911 {
 912         LASSERT(res != NULL);
 913         LASSERT(res != LP_POISON);
 914         atomic_inc(&res->lr_refcount);
 915         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 916                atomic_read(&res->lr_refcount));
 917         return res;
 918 }
 919
 920 void __ldlm_resource_putref_final(struct ldlm_resource *res)
 921 {
 922         struct ldlm_namespace *ns = res->lr_namespace;
 923
 924         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 925
 926         if (!list_empty(&res->lr_granted)) {
 927                 ldlm_resource_dump(D_ERROR, res);
 928                 LBUG();
 929         }
 930
 931         if (!list_empty(&res->lr_converting)) {
 932                 ldlm_resource_dump(D_ERROR, res);
 933                 LBUG();
 934         }
 935
 936         if (!list_empty(&res->lr_waiting)) {
 937                 ldlm_resource_dump(D_ERROR, res);
 938                 LBUG();
 939         }
 940
 941         if (!list_empty(&res->lr_children)) {
 942                 ldlm_resource_dump(D_ERROR, res);
 943                 LBUG();
 944         }
 945
 946         /* Pass 0 here to not wake ->ns_waitq up yet, we will do it few
 947          * lines below when all children are freed. */
 948         ldlm_namespace_put_locked(ns, 0);
 949         list_del_init(&res->lr_hash);
 950         list_del_init(&res->lr_childof);
 951
 952         ns->ns_resources--;
 953         if (ns->ns_resources == 0)
 954                 wake_up(&ns->ns_waitq);
 955 }
 956
 957 /* Returns 1 if the resource was freed, 0 if it remains. */
 958 int ldlm_resource_putref(struct ldlm_resource *res)
 959 {
 960         struct ldlm_namespace *ns = res->lr_namespace;
 961         int rc = 0;
 962         ENTRY;
 963
 964         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 965                atomic_read(&res->lr_refcount) - 1);
 966         LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
 967                  atomic_read(&res->lr_refcount));
 968         LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
 969                  atomic_read(&res->lr_refcount));
 970
 971         if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
 972                 __ldlm_resource_putref_final(res);
 973                 spin_unlock(&ns->ns_hash_lock);
 974                 if (res->lr_lvb_data)
 975                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 976                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 977                 rc = 1;
 978         }
 979
 980         RETURN(rc);
 981 }
 982
 983 /* Returns 1 if the resource was freed, 0 if it remains. */
 984 int ldlm_resource_putref_locked(struct ldlm_resource *res)
 985 {
 986         int rc = 0;
 987         ENTRY;
 988
 989         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 990                atomic_read(&res->lr_refcount) - 1);
 991         LASSERT(atomic_read(&res->lr_refcount) > 0);
 992         LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
 993
 994         LASSERT(atomic_read(&res->lr_refcount) >= 0);
 995         if (atomic_dec_and_test(&res->lr_refcount)) {
 996                 __ldlm_resource_putref_final(res);
 997                 if (res->lr_lvb_data)
 998                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 999                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1000                 rc = 1;
1001         }
1002
1003         RETURN(rc);
1004 }
1005
1006 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1007                             struct ldlm_lock *lock)
1008 {
1009         check_res_locked(res);
1010
1011         ldlm_resource_dump(D_INFO, res);
1012         CDEBUG(D_INFO, "About to add this lock:\n");
1013         ldlm_lock_dump(D_INFO, lock, 0);
1014
1015         if (lock->l_destroyed) {
1016                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1017                 return;
1018         }
1019
1020         LASSERT(list_empty(&lock->l_res_link));
1021
1022         list_add_tail(&lock->l_res_link, head);
1023 }
1024
1025 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1026                                      struct ldlm_lock *new)
1027 {
1028         struct ldlm_resource *res = original->l_resource;
1029
1030         check_res_locked(res);
1031
1032         ldlm_resource_dump(D_OTHER, res);
1033         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1034         ldlm_lock_dump(D_OTHER, new, 0);
1035
1036         if (new->l_destroyed) {
1037                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1038                 goto out;
1039         }
1040
1041         LASSERT(list_empty(&new->l_res_link));
1042
1043         list_add(&new->l_res_link, &original->l_res_link);
1044  out:;
1045 }
1046
1047 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1048 {
1049         int type = lock->l_resource->lr_type;
1050
1051         check_res_locked(lock->l_resource);
1052         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1053                 ldlm_unlink_lock_skiplist(lock);
1054         else if (type == LDLM_EXTENT)
1055                 ldlm_extent_unlink_lock(lock);
1056         list_del_init(&lock->l_res_link);
1057 }
1058
1059 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1060 {
1061         desc->lr_type = res->lr_type;
1062         desc->lr_name = res->lr_name;
1063 }
1064
1065 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1066 {
1067         struct list_head *tmp;
1068
1069         if (!((libcfs_debug | D_ERROR) & level))
1070                 return;
1071
1072         mutex_down(ldlm_namespace_lock(client));
1073
1074         list_for_each(tmp, ldlm_namespace_list(client)) {
1075                 struct ldlm_namespace *ns;
1076                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1077                 ldlm_namespace_dump(level, ns);
1078         }
1079
1080         mutex_up(ldlm_namespace_lock(client));
1081 }
1082
1083 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1084 {
1085         struct list_head *tmp;
1086
1087         if (!((libcfs_debug | D_ERROR) & level))
1088                 return;
1089
1090         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1091                ns->ns_name, ns->ns_refcount,
1092                ns_is_client(ns) ? "client" : "server");
1093
1094         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1095                 return;
1096
1097         spin_lock(&ns->ns_hash_lock);
1098         tmp = ns->ns_root_list.next;
1099         while (tmp != &ns->ns_root_list) {
1100                 struct ldlm_resource *res;
1101                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
1102
1103                 ldlm_resource_getref(res);
1104                 spin_unlock(&ns->ns_hash_lock);
1105
1106                 lock_res(res);
1107                 ldlm_resource_dump(level, res);
1108                 unlock_res(res);
1109
1110                 spin_lock(&ns->ns_hash_lock);
1111                 tmp = tmp->next;
1112                 ldlm_resource_putref_locked(res);
1113         }
1114         ns->ns_next_dump = cfs_time_shift(10);
1115         spin_unlock(&ns->ns_hash_lock);
1116 }
1117
1118 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1119 {
1120         struct list_head *tmp;
1121         int pos;
1122
1123         CLASSERT(RES_NAME_SIZE == 4);
1124
1125         if (!((libcfs_debug | D_ERROR) & level))
1126                 return;
1127
1128         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1129                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1130                res->lr_name.name[2], res->lr_name.name[3],
1131                atomic_read(&res->lr_refcount));
1132
1133         if (!list_empty(&res->lr_granted)) {
1134                 pos = 0;
1135                 CDEBUG(level, "Granted locks:\n");
1136                 list_for_each(tmp, &res->lr_granted) {
1137                         struct ldlm_lock *lock;
1138                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1139                         ldlm_lock_dump(level, lock, ++pos);
1140                 }
1141         }
1142         if (!list_empty(&res->lr_converting)) {
1143                 pos = 0;
1144                 CDEBUG(level, "Converting locks:\n");
1145                 list_for_each(tmp, &res->lr_converting) {
1146                         struct ldlm_lock *lock;
1147                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1148                         ldlm_lock_dump(level, lock, ++pos);
1149                 }
1150         }
1151         if (!list_empty(&res->lr_waiting)) {
1152                 pos = 0;
1153                 CDEBUG(level, "Waiting locks:\n");
1154                 list_for_each(tmp, &res->lr_waiting) {
1155                         struct ldlm_lock *lock;
1156                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1157                         ldlm_lock_dump(level, lock, ++pos);
1158                 }
1159         }
1160 }