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         ns->ns_refcount = 0;
 331         ns->ns_client = client;
 332         spin_lock_init(&ns->ns_hash_lock);
 333         atomic_set(&ns->ns_locks, 0);
 334         ns->ns_resources = 0;
 335         cfs_waitq_init(&ns->ns_waitq);
 336         ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
 337         ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
 338         ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
 339
 340         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 341              bucket--)
 342                 CFS_INIT_LIST_HEAD(bucket);
 343
 344         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 345         ns->ns_nr_unused = 0;
 346         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
 347         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
 348         ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
 349         spin_lock_init(&ns->ns_unused_lock);
 350         ns->ns_orig_connect_flags = 0;
 351         ns->ns_connect_flags = 0;
 352         ldlm_proc_namespace(ns);
 353
 354         idx = atomic_read(ldlm_namespace_nr(client));
 355         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 356         if (rc) {
 357                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 358                 GOTO(out_proc, rc);
 359         }
 360
 361         mutex_down(ldlm_namespace_lock(client));
 362         list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 363         atomic_inc(ldlm_namespace_nr(client));
 364         mutex_up(ldlm_namespace_lock(client));
 365
 366         RETURN(ns);
 367 out_proc:
 368         ldlm_namespace_cleanup(ns, 0);
 369         OBD_FREE(ns->ns_name, namelen + 1);
 370 out_hash:
 371         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 372 out_ns:
 373         OBD_FREE_PTR(ns);
 374 out_ref:
 375         ldlm_put_ref(0);
 376         RETURN(NULL);
 377 }
 378
 379 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 380
 381 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 382  * This is currently only used for recovery, and we make certain assumptions
 383  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 384  *
 385  * Called with the ns_lock held. */
 386 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 387                              int flags)
 388 {
 389         struct list_head *tmp;
 390         int rc = 0, client = ns_is_client(res->lr_namespace);
 391         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 392         ENTRY;
 393
 394
 395         do {
 396                 struct ldlm_lock *lock = NULL;
 397
 398                 /* first, we look for non-cleaned-yet lock
 399                  * all cleaned locks are marked by CLEANED flag */
 400                 lock_res(res);
 401                 list_for_each(tmp, q) {
 402                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 403                         if (lock->l_flags & LDLM_FL_CLEANED) {
 404                                 lock = NULL;
 405                                 continue;
 406                         }
 407                         LDLM_LOCK_GET(lock);
 408                         lock->l_flags |= LDLM_FL_CLEANED;
 409                         break;
 410                 }
 411
 412                 if (lock == NULL) {
 413                         unlock_res(res);
 414                         break;
 415                 }
 416
 417                 /* Set CBPENDING so nothing in the cancellation path
 418                  * can match this lock */
 419                 lock->l_flags |= LDLM_FL_CBPENDING;
 420                 lock->l_flags |= LDLM_FL_FAILED;
 421                 lock->l_flags |= flags;
 422
 423                 /* ... without sending a CANCEL message for local_only. */
 424                 if (local_only)
 425                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 426
 427                 if (local_only && (lock->l_readers || lock->l_writers)) {
 428                         /* This is a little bit gross, but much better than the
 429                          * alternative: pretend that we got a blocking AST from
 430                          * the server, so that when the lock is decref'd, it
 431                          * will go away ... */
 432                         unlock_res(res);
 433                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 434                         if (lock->l_completion_ast)
 435                                 lock->l_completion_ast(lock, 0, NULL);
 436                         LDLM_LOCK_PUT(lock);
 437                         continue;
 438                 }
 439
 440                 if (client) {
 441                         struct lustre_handle lockh;
 442
 443                         unlock_res(res);
 444                         ldlm_lock2handle(lock, &lockh);
 445                         rc = ldlm_cli_cancel(&lockh);
 446                         if (rc)
 447                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 448                 } else {
 449                         ldlm_resource_unlink_lock(lock);
 450                         unlock_res(res);
 451                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 452                                    "client node");
 453                         ldlm_lock_destroy(lock);
 454                 }
 455                 LDLM_LOCK_PUT(lock);
 456         } while (1);
 457
 458         EXIT;
 459 }
 460
 461 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 462 {
 463         struct list_head *tmp;
 464         int i;
 465
 466         if (ns == NULL) {
 467                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 468                 return ELDLM_OK;
 469         }
 470
 471         for (i = 0; i < RES_HASH_SIZE; i++) {
 472                 spin_lock(&ns->ns_hash_lock);
 473                 tmp = ns->ns_hash[i].next;
 474                 while (tmp != &(ns->ns_hash[i])) {
 475                         struct ldlm_resource *res;
 476                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 477                         ldlm_resource_getref(res);
 478                         spin_unlock(&ns->ns_hash_lock);
 479
 480                         cleanup_resource(res, &res->lr_granted, flags);
 481                         cleanup_resource(res, &res->lr_converting, flags);
 482                         cleanup_resource(res, &res->lr_waiting, flags);
 483
 484                         spin_lock(&ns->ns_hash_lock);
 485                         tmp  = tmp->next;
 486
 487                         /* XXX: former stuff caused issues in case of race
 488                          * between ldlm_namespace_cleanup() and lockd() when
 489                          * client gets blocking ast when lock gets distracted by
 490                          * server. This is 1_4 branch solution, let's see how
 491                          * will it behave. */
 492                         if (!ldlm_resource_putref_locked(res))
 493                                 CDEBUG(D_INFO,
 494                                        "Namespace %s resource refcount nonzero "
 495                                        "(%d) after lock cleanup; forcing cleanup.\n",
 496                                        ns->ns_name, atomic_read(&res->lr_refcount));
 497                 }
 498                 spin_unlock(&ns->ns_hash_lock);
 499         }
 500
 501         return ELDLM_OK;
 502 }
 503
 504 /* Cleanup, but also free, the namespace */
 505 int ldlm_namespace_free_prior(struct ldlm_namespace *ns)
 506 {
 507         ENTRY;
 508         if (!ns)
 509                 RETURN(ELDLM_OK);
 510
 511         mutex_down(ldlm_namespace_lock(ns->ns_client));
 512         /*
 513          * Some asserts and possibly other parts of code still using
 514          * list_empty(&ns->ns_list_chain). This is why it is important
 515          * to use list_del_init() here.
 516          */
 517         list_del_init(&ns->ns_list_chain);
 518         atomic_dec(ldlm_namespace_nr(ns->ns_client));
 519         ldlm_pool_fini(&ns->ns_pool);
 520         mutex_up(ldlm_namespace_lock(ns->ns_client));
 521
 522         /* At shutdown time, don't call the cancellation callback */
 523         ldlm_namespace_cleanup(ns, 0);
 524
 525         if (ns->ns_refcount > 0) {
 526                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 527                 int rc;
 528                 CDEBUG(D_DLMTRACE,
 529                        "dlm namespace %s free waiting on refcount %d\n",
 530                        ns->ns_name, ns->ns_refcount);
 531                 rc = l_wait_event(ns->ns_waitq,
 532                                   ns->ns_refcount == 0, &lwi);
 533                 if (ns->ns_refcount)
 534                         LCONSOLE_ERROR_MSG(0x139, "Lock manager: wait for %s "
 535                                            "namespace cleanup aborted with %d "
 536                                            "resources in use. (%d)\nI'm going "
 537                                            "to try to clean up anyway, but I "
 538                                            "might need a reboot of this node.\n",
 539                                             ns->ns_name, (int) ns->ns_refcount,
 540                                             rc);
 541                 CDEBUG(D_DLMTRACE,
 542                        "dlm namespace %s free done waiting\n", ns->ns_name);
 543         }
 544
 545         RETURN(ELDLM_OK);
 546 }
 547
 548 int ldlm_namespace_free_post(struct ldlm_namespace *ns, int force)
 549 {
 550         ENTRY;
 551         if (!ns)
 552                 RETURN(ELDLM_OK);
 553
 554 #ifdef LPROCFS
 555         {
 556                 struct proc_dir_entry *dir;
 557                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
 558                 if (dir == NULL) {
 559                         CERROR("dlm namespace %s has no procfs dir?\n",
 560                                ns->ns_name);
 561                 } else {
 562                         lprocfs_remove(&dir);
 563                 }
 564         }
 565 #endif
 566
 567         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 568         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 569         /*
 570          * @ns should be not on list in this time, otherwise this will cause
 571          * issues realted to using freed @ns in pools thread.
 572          */
 573         LASSERT(list_empty(&ns->ns_list_chain));
 574         OBD_FREE_PTR(ns);
 575         ldlm_put_ref(force);
 576         RETURN(ELDLM_OK);
 577 }
 578
 579
 580 /* Cleanup the resource, and free namespace.
 581  * bug 12864:
 582  * Deadlock issue:
 583  * proc1: destroy import
 584  *        class_disconnect_export(grab cl_sem) ->
 585  *              -> ldlm_namespace_free ->
 586  *              -> lprocfs_remove(grab _lprocfs_lock).
 587  * proc2: read proc info
 588  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 589  *              -> osc_rd_active, etc(grab cl_sem).
 590  *
 591  * So that I have to split the ldlm_namespace_free into two parts - the first
 592  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 593  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 594  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 595  * held.
 596  */
 597 int ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 598 {
 599         ldlm_namespace_free_prior(ns);
 600         ldlm_namespace_free_post(ns, force);
 601         return ELDLM_OK;
 602 }
 603
 604
 605 void ldlm_namespace_get_nolock(struct ldlm_namespace *ns)
 606 {
 607         LASSERT(ns->ns_refcount >= 0);
 608         ns->ns_refcount++;
 609 }
 610
 611 void ldlm_namespace_get(struct ldlm_namespace *ns)
 612 {
 613         spin_lock(&ns->ns_hash_lock);
 614         ldlm_namespace_get_nolock(ns);
 615         spin_unlock(&ns->ns_hash_lock);
 616 }
 617
 618 void ldlm_namespace_put_nolock(struct ldlm_namespace *ns, int wakeup)
 619 {
 620         LASSERT(ns->ns_refcount > 0);
 621         ns->ns_refcount--;
 622         if (ns->ns_refcount == 0 && wakeup)
 623                 wake_up(&ns->ns_waitq);
 624 }
 625
 626 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
 627 {
 628         spin_lock(&ns->ns_hash_lock);
 629         ldlm_namespace_put_nolock(ns, wakeup);
 630         spin_unlock(&ns->ns_hash_lock);
 631 }
 632
 633 /* Should be called under ldlm_namespace_lock(client) taken */
 634 void ldlm_namespace_move(struct ldlm_namespace *ns, ldlm_side_t client)
 635 {
 636         LASSERT(!list_empty(&ns->ns_list_chain));
 637         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 638         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 639 }
 640
 641 /* Should be called under ldlm_namespace_lock(client) taken */
 642 struct ldlm_namespace *ldlm_namespace_first(ldlm_side_t client)
 643 {
 644         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 645         LASSERT(!list_empty(ldlm_namespace_list(client)));
 646         return container_of(ldlm_namespace_list(client)->next,
 647                 struct ldlm_namespace, ns_list_chain);
 648 }
 649 static __u32 ldlm_hash_fn(struct ldlm_resource *parent,
 650                           const struct ldlm_res_id *name)
 651 {
 652         __u32 hash = 0;
 653         int i;
 654
 655         for (i = 0; i < RES_NAME_SIZE; i++)
 656                 hash += name->name[i];
 657
 658         hash += (__u32)((unsigned long)parent >> 4);
 659
 660         return (hash & RES_HASH_MASK);
 661 }
 662
 663 static struct ldlm_resource *ldlm_resource_new(void)
 664 {
 665         struct ldlm_resource *res;
 666         int idx;
 667
 668         OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
 669         if (res == NULL)
 670                 return NULL;
 671
 672         memset(res, 0, sizeof(*res));
 673
 674         CFS_INIT_LIST_HEAD(&res->lr_children);
 675         CFS_INIT_LIST_HEAD(&res->lr_childof);
 676         CFS_INIT_LIST_HEAD(&res->lr_granted);
 677         CFS_INIT_LIST_HEAD(&res->lr_converting);
 678         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 679
 680         /* initialize interval trees for each lock mode*/
 681         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 682                 res->lr_itree[idx].lit_size = 0;
 683                 res->lr_itree[idx].lit_mode = 1 << idx;
 684                 res->lr_itree[idx].lit_root = NULL;
 685         }
 686
 687         atomic_set(&res->lr_refcount, 1);
 688         spin_lock_init(&res->lr_lock);
 689
 690         /* one who creates the resource must unlock
 691          * the semaphore after lvb initialization */
 692         init_MUTEX_LOCKED(&res->lr_lvb_sem);
 693
 694         return res;
 695 }
 696
 697 /* must be called with hash lock held */
 698 static struct ldlm_resource *
 699 ldlm_resource_find(struct ldlm_namespace *ns, const struct ldlm_res_id *name,
 700                    __u32 hash)
 701 {
 702         struct list_head *bucket, *tmp;
 703         struct ldlm_resource *res;
 704
 705         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 706         bucket = ns->ns_hash + hash;
 707
 708         list_for_each(tmp, bucket) {
 709                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 710                 if (memcmp(&res->lr_name, name, sizeof(res->lr_name)) == 0)
 711                         return res;
 712         }
 713
 714         return NULL;
 715 }
 716
 717 /* Args: locked namespace
 718  * Returns: newly-allocated, referenced, unlocked resource */
 719 static struct ldlm_resource *
 720 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 721                   const struct ldlm_res_id *name, __u32 hash, ldlm_type_t type)
 722 {
 723         struct list_head *bucket;
 724         struct ldlm_resource *res, *old_res;
 725         ENTRY;
 726
 727         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
 728                  "type: %d\n", type);
 729
 730         res = ldlm_resource_new();
 731         if (!res)
 732                 RETURN(NULL);
 733
 734         res->lr_name = *name;
 735         res->lr_namespace = ns;
 736         res->lr_type = type;
 737         res->lr_most_restr = LCK_NL;
 738
 739         spin_lock(&ns->ns_hash_lock);
 740         old_res = ldlm_resource_find(ns, name, hash);
 741         if (old_res) {
 742                 /* someone won the race and added the resource before */
 743                 ldlm_resource_getref(old_res);
 744                 spin_unlock(&ns->ns_hash_lock);
 745                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 746                 /* synchronize WRT resource creation */
 747                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 748                         down(&old_res->lr_lvb_sem);
 749                         up(&old_res->lr_lvb_sem);
 750                 }
 751                 RETURN(old_res);
 752         }
 753
 754         /* we won! let's add the resource */
 755         bucket = ns->ns_hash + hash;
 756         list_add(&res->lr_hash, bucket);
 757         ns->ns_resources++;
 758         ldlm_namespace_get_nolock(ns);
 759
 760         if (parent == NULL) {
 761                 list_add(&res->lr_childof, &ns->ns_root_list);
 762         } else {
 763                 res->lr_parent = parent;
 764                 list_add(&res->lr_childof, &parent->lr_children);
 765         }
 766         spin_unlock(&ns->ns_hash_lock);
 767
 768         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 769                 int rc;
 770
 771                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
 772                 rc = ns->ns_lvbo->lvbo_init(res);
 773                 if (rc)
 774                         CERROR("lvbo_init failed for resource "
 775                                LPU64": rc %d\n", name->name[0], rc);
 776                 /* we create resource with locked lr_lvb_sem */
 777                 up(&res->lr_lvb_sem);
 778         }
 779
 780         RETURN(res);
 781 }
 782
 783 /* Args: unlocked namespace
 784  * Locks: takes and releases ns->ns_lock and res->lr_lock
 785  * Returns: referenced, unlocked ldlm_resource or NULL */
 786 struct ldlm_resource *
 787 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 788                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
 789 {
 790         __u32 hash = ldlm_hash_fn(parent, name);
 791         struct ldlm_resource *res = NULL;
 792         ENTRY;
 793
 794         LASSERT(ns != NULL);
 795         LASSERT(ns->ns_hash != NULL);
 796         LASSERT(name->name[0] != 0);
 797
 798         spin_lock(&ns->ns_hash_lock);
 799         res = ldlm_resource_find(ns, name, hash);
 800         if (res) {
 801                 ldlm_resource_getref(res);
 802                 spin_unlock(&ns->ns_hash_lock);
 803                 /* synchronize WRT resource creation */
 804                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 805                         down(&res->lr_lvb_sem);
 806                         up(&res->lr_lvb_sem);
 807                 }
 808                 RETURN(res);
 809         }
 810         spin_unlock(&ns->ns_hash_lock);
 811
 812         if (create == 0)
 813                 RETURN(NULL);
 814
 815         res = ldlm_resource_add(ns, parent, name, hash, type);
 816         RETURN(res);
 817 }
 818
 819 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 820 {
 821         LASSERT(res != NULL);
 822         LASSERT(res != LP_POISON);
 823         atomic_inc(&res->lr_refcount);
 824         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 825                atomic_read(&res->lr_refcount));
 826         return res;
 827 }
 828
 829 void __ldlm_resource_putref_final(struct ldlm_resource *res)
 830 {
 831         struct ldlm_namespace *ns = res->lr_namespace;
 832
 833         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 834
 835         if (!list_empty(&res->lr_granted)) {
 836                 ldlm_resource_dump(D_ERROR, res);
 837                 LBUG();
 838         }
 839
 840         if (!list_empty(&res->lr_converting)) {
 841                 ldlm_resource_dump(D_ERROR, res);
 842                 LBUG();
 843         }
 844
 845         if (!list_empty(&res->lr_waiting)) {
 846                 ldlm_resource_dump(D_ERROR, res);
 847                 LBUG();
 848         }
 849
 850         if (!list_empty(&res->lr_children)) {
 851                 ldlm_resource_dump(D_ERROR, res);
 852                 LBUG();
 853         }
 854
 855         /* Pass 0 here to not wake ->ns_waitq up yet, we will do it few
 856          * lines below when all children are freed. */
 857         ldlm_namespace_put_nolock(ns, 0);
 858         list_del_init(&res->lr_hash);
 859         list_del_init(&res->lr_childof);
 860
 861         ns->ns_resources--;
 862         if (ns->ns_resources == 0)
 863                 wake_up(&ns->ns_waitq);
 864 }
 865
 866 /* Returns 1 if the resource was freed, 0 if it remains. */
 867 int ldlm_resource_putref(struct ldlm_resource *res)
 868 {
 869         struct ldlm_namespace *ns = res->lr_namespace;
 870         int rc = 0;
 871         ENTRY;
 872
 873         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 874                atomic_read(&res->lr_refcount) - 1);
 875         LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
 876                  atomic_read(&res->lr_refcount));
 877         LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
 878                  atomic_read(&res->lr_refcount));
 879
 880         if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
 881                 __ldlm_resource_putref_final(res);
 882                 spin_unlock(&ns->ns_hash_lock);
 883                 if (res->lr_lvb_data)
 884                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 885                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 886                 rc = 1;
 887         }
 888
 889         RETURN(rc);
 890 }
 891
 892 /* Returns 1 if the resource was freed, 0 if it remains. */
 893 int ldlm_resource_putref_locked(struct ldlm_resource *res)
 894 {
 895         int rc = 0;
 896         ENTRY;
 897
 898         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 899                atomic_read(&res->lr_refcount) - 1);
 900         LASSERT(atomic_read(&res->lr_refcount) > 0);
 901         LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
 902
 903         LASSERT(atomic_read(&res->lr_refcount) >= 0);
 904         if (atomic_dec_and_test(&res->lr_refcount)) {
 905                 __ldlm_resource_putref_final(res);
 906                 if (res->lr_lvb_data)
 907                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 908                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 909                 rc = 1;
 910         }
 911
 912         RETURN(rc);
 913 }
 914
 915 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 916                             struct ldlm_lock *lock)
 917 {
 918         check_res_locked(res);
 919
 920         ldlm_resource_dump(D_OTHER, res);
 921         CDEBUG(D_OTHER, "About to add this lock:\n");
 922         ldlm_lock_dump(D_OTHER, lock, 0);
 923
 924         if (lock->l_destroyed) {
 925                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 926                 return;
 927         }
 928
 929         LASSERT(list_empty(&lock->l_res_link));
 930
 931         list_add_tail(&lock->l_res_link, head);
 932 }
 933
 934 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
 935                                      struct ldlm_lock *new)
 936 {
 937         struct ldlm_resource *res = original->l_resource;
 938
 939         check_res_locked(res);
 940
 941         ldlm_resource_dump(D_OTHER, res);
 942         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
 943         ldlm_lock_dump(D_OTHER, new, 0);
 944
 945         if (new->l_destroyed) {
 946                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 947                 goto out;
 948         }
 949
 950         LASSERT(list_empty(&new->l_res_link));
 951
 952         list_add(&new->l_res_link, &original->l_res_link);
 953  out:;
 954 }
 955
 956 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 957 {
 958         int type = lock->l_resource->lr_type;
 959
 960         check_res_locked(lock->l_resource);
 961         if (type == LDLM_IBITS || type == LDLM_PLAIN)
 962                 ldlm_unlink_lock_skiplist(lock);
 963         else if (type == LDLM_EXTENT)
 964                 ldlm_extent_unlink_lock(lock);
 965         list_del_init(&lock->l_res_link);
 966 }
 967
 968 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 969 {
 970         desc->lr_type = res->lr_type;
 971         desc->lr_name = res->lr_name;
 972 }
 973
 974 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
 975 {
 976         struct list_head *tmp;
 977
 978         if (!((libcfs_debug | D_ERROR) & level))
 979                 return;
 980
 981         mutex_down(ldlm_namespace_lock(client));
 982
 983         list_for_each(tmp, ldlm_namespace_list(client)) {
 984                 struct ldlm_namespace *ns;
 985                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 986                 ldlm_namespace_dump(level, ns);
 987         }
 988
 989         mutex_up(ldlm_namespace_lock(client));
 990 }
 991
 992 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
 993 {
 994         struct list_head *tmp;
 995
 996         if (!((libcfs_debug | D_ERROR) & level))
 997                 return;
 998
 999         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1000                ns->ns_name, ns->ns_refcount,
1001                ns_is_client(ns) ? "client" : "server");
1002
1003         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1004                 return;
1005
1006         spin_lock(&ns->ns_hash_lock);
1007         tmp = ns->ns_root_list.next;
1008         while (tmp != &ns->ns_root_list) {
1009                 struct ldlm_resource *res;
1010                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
1011
1012                 ldlm_resource_getref(res);
1013                 spin_unlock(&ns->ns_hash_lock);
1014
1015                 lock_res(res);
1016                 ldlm_resource_dump(level, res);
1017                 unlock_res(res);
1018
1019                 spin_lock(&ns->ns_hash_lock);
1020                 tmp = tmp->next;
1021                 ldlm_resource_putref_locked(res);
1022         }
1023         ns->ns_next_dump = cfs_time_shift(10);
1024         spin_unlock(&ns->ns_hash_lock);
1025 }
1026
1027 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1028 {
1029         struct list_head *tmp;
1030         int pos;
1031
1032         CLASSERT(RES_NAME_SIZE == 4);
1033
1034         if (!((libcfs_debug | D_ERROR) & level))
1035                 return;
1036
1037         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1038                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1039                res->lr_name.name[2], res->lr_name.name[3],
1040                atomic_read(&res->lr_refcount));
1041
1042         if (!list_empty(&res->lr_granted)) {
1043                 pos = 0;
1044                 CDEBUG(level, "Granted locks:\n");
1045                 list_for_each(tmp, &res->lr_granted) {
1046                         struct ldlm_lock *lock;
1047                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1048                         ldlm_lock_dump(level, lock, ++pos);
1049                 }
1050         }
1051         if (!list_empty(&res->lr_converting)) {
1052                 pos = 0;
1053                 CDEBUG(level, "Converting locks:\n");
1054                 list_for_each(tmp, &res->lr_converting) {
1055                         struct ldlm_lock *lock;
1056                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1057                         ldlm_lock_dump(level, lock, ++pos);
1058                 }
1059         }
1060         if (!list_empty(&res->lr_waiting)) {
1061                 pos = 0;
1062                 CDEBUG(level, "Waiting locks:\n");
1063                 list_for_each(tmp, &res->lr_waiting) {
1064                         struct ldlm_lock *lock;
1065                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1066                         ldlm_lock_dump(level, lock, ++pos);
1067                 }
1068         }
1069 }