lustre/ldlm/ldlm_resource.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Phil Schwan <phil@clusterfs.com>
   6  *   Author: Peter Braam <braam@clusterfs.com>
   7  *
   8  *   This file is part of the Lustre file system, http://www.lustre.org
   9  *   Lustre is a trademark of Cluster File Systems, Inc.
  10  *
  11  *   You may have signed or agreed to another license before downloading
  12  *   this software.  If so, you are bound by the terms and conditions
  13  *   of that agreement, and the following does not apply to you.  See the
  14  *   LICENSE file included with this distribution for more information.
  15  *
  16  *   If you did not agree to a different license, then this copy of Lustre
  17  *   is open source software; you can redistribute it and/or modify it
  18  *   under the terms of version 2 of the GNU General Public License as
  19  *   published by the Free Software Foundation.
  20  *
  21  *   In either case, Lustre is distributed in the hope that it will be
  22  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  23  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  24  *   license text for more details.
  25  */
  26
  27 #define DEBUG_SUBSYSTEM S_LDLM
  28 #ifdef __KERNEL__
  29 # include <lustre_dlm.h>
  30 #else
  31 # include <liblustre.h>
  32 #endif
  33
  34 #include <obd_class.h>
  35 #include "ldlm_internal.h"
  36
  37 cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
  38
  39 atomic_t ldlm_srv_namespace_nr = ATOMIC_INIT(0);
  40 atomic_t ldlm_cli_namespace_nr = ATOMIC_INIT(0);
  41
  42 struct semaphore ldlm_srv_namespace_lock;
  43 struct list_head ldlm_srv_namespace_list =
  44         CFS_LIST_HEAD_INIT(ldlm_srv_namespace_list);
  45
  46 struct semaphore ldlm_cli_namespace_lock;
  47 struct list_head ldlm_cli_namespace_list =
  48         CFS_LIST_HEAD_INIT(ldlm_cli_namespace_list);
  49
  50 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
  51 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
  52 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
  53
  54 #ifdef LPROCFS
  55 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  56                              unsigned long count, void *data)
  57 {
  58         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  59         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  60         RETURN(count);
  61 }
  62
  63 int ldlm_proc_setup(void)
  64 {
  65         int rc;
  66         struct lprocfs_vars list[] = {
  67                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  68                 { NULL }};
  69         ENTRY;
  70         LASSERT(ldlm_ns_proc_dir == NULL);
  71
  72         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
  73                                               proc_lustre_root,
  74                                               NULL, NULL);
  75         if (IS_ERR(ldlm_type_proc_dir)) {
  76                 CERROR("LProcFS failed in ldlm-init\n");
  77                 rc = PTR_ERR(ldlm_type_proc_dir);
  78                 GOTO(err, rc);
  79         }
  80
  81         ldlm_ns_proc_dir = lprocfs_register("namespaces",
  82                                             ldlm_type_proc_dir,
  83                                             NULL, NULL);
  84         if (IS_ERR(ldlm_ns_proc_dir)) {
  85                 CERROR("LProcFS failed in ldlm-init\n");
  86                 rc = PTR_ERR(ldlm_ns_proc_dir);
  87                 GOTO(err_type, rc);
  88         }
  89
  90         ldlm_svc_proc_dir = lprocfs_register("services",
  91                                             ldlm_type_proc_dir,
  92                                             NULL, NULL);
  93         if (IS_ERR(ldlm_svc_proc_dir)) {
  94                 CERROR("LProcFS failed in ldlm-init\n");
  95                 rc = PTR_ERR(ldlm_svc_proc_dir);
  96                 GOTO(err_ns, rc);
  97         }
  98
  99         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 100
 101         RETURN(0);
 102
 103 err_ns:
 104         lprocfs_remove(&ldlm_ns_proc_dir);
 105 err_type:
 106         lprocfs_remove(&ldlm_type_proc_dir);
 107 err:
 108         ldlm_svc_proc_dir = NULL;
 109         RETURN(rc);
 110 }
 111
 112 void ldlm_proc_cleanup(void)
 113 {
 114         if (ldlm_svc_proc_dir)
 115                 lprocfs_remove(&ldlm_svc_proc_dir);
 116
 117         if (ldlm_ns_proc_dir)
 118                 lprocfs_remove(&ldlm_ns_proc_dir);
 119
 120         if (ldlm_type_proc_dir)
 121                 lprocfs_remove(&ldlm_type_proc_dir);
 122 }
 123
 124 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 125                                int count, int *eof, void *data)
 126 {
 127         struct ldlm_namespace *ns = data;
 128         __u32 *nr = &ns->ns_max_unused;
 129
 130         if (ns_connect_lru_resize(ns))
 131                 nr = &ns->ns_nr_unused;
 132         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 133 }
 134
 135 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 136                                unsigned long count, void *data)
 137 {
 138         struct ldlm_namespace *ns = data;
 139         char dummy[MAX_STRING_SIZE + 1], *end;
 140         unsigned long tmp;
 141         int lru_resize;
 142
 143         dummy[MAX_STRING_SIZE] = '\0';
 144         if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 145                 return -EFAULT;
 146
 147         if (count == 6 && memcmp(dummy, "clear", 5) == 0) {
 148                 CDEBUG(D_DLMTRACE,
 149                        "dropping all unused locks from namespace %s\n",
 150                        ns->ns_name);
 151                 if (ns_connect_lru_resize(ns)) {
 152                         int canceled, unused  = ns->ns_nr_unused;
 153
 154                         /* Try to cancel all @ns_nr_unused locks. */
 155                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
 156                                                    LDLM_CANCEL_PASSED);
 157                         if (canceled < unused) {
 158                                 CERROR("not all requested locks are canceled, "
 159                                        "requested: %d, canceled: %d\n", unused,
 160                                        canceled);
 161                                 return -EINVAL;
 162                         }
 163                 } else {
 164                         tmp = ns->ns_max_unused;
 165                         ns->ns_max_unused = 0;
 166                         ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
 167                         ns->ns_max_unused = tmp;
 168                 }
 169                 return count;
 170         }
 171
 172         tmp = simple_strtoul(dummy, &end, 0);
 173         if (dummy == end) {
 174                 CERROR("invalid value written\n");
 175                 return -EINVAL;
 176         }
 177         lru_resize = (tmp == 0);
 178
 179         if (ns_connect_lru_resize(ns)) {
 180                 if (!lru_resize)
 181                         ns->ns_max_unused = (unsigned int)tmp;
 182
 183                 if (tmp > ns->ns_nr_unused)
 184                         tmp = ns->ns_nr_unused;
 185                 tmp = ns->ns_nr_unused - tmp;
 186
 187                 CDEBUG(D_DLMTRACE, "changing namespace %s unused locks from %u to %u\n",
 188                        ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
 189                 ldlm_cancel_lru(ns, (unsigned int)tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 190
 191                 if (!lru_resize) {
 192                         CDEBUG(D_DLMTRACE, "disable lru_resize for namespace %s\n",
 193                                ns->ns_name);
 194                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 195                 }
 196         } else {
 197                 CDEBUG(D_DLMTRACE, "changing namespace %s max_unused from %u to %u\n",
 198                        ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
 199                 ns->ns_max_unused = (unsigned int)tmp;
 200                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 201
 202                 /* Make sure that originally lru resize was supported before
 203                  * turning it on here. */
 204                 if (lru_resize &&
 205                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 206                         CDEBUG(D_DLMTRACE, "enable lru_resize for namespace %s\n",
 207                                ns->ns_name);
 208                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 209                 }
 210         }
 211
 212         return count;
 213 }
 214
 215 void ldlm_proc_namespace(struct ldlm_namespace *ns)
 216 {
 217         struct lprocfs_vars lock_vars[2];
 218         char lock_name[MAX_STRING_SIZE + 1];
 219
 220         LASSERT(ns != NULL);
 221         LASSERT(ns->ns_name != NULL);
 222
 223         lock_name[MAX_STRING_SIZE] = '\0';
 224
 225         memset(lock_vars, 0, sizeof(lock_vars));
 226         lock_vars[0].name = lock_name;
 227
 228         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
 229         lock_vars[0].data = &ns->ns_refcount;
 230         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 231         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 232
 233         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
 234         lock_vars[0].data = &ns->ns_locks;
 235         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 236         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 237
 238         if (ns_is_client(ns)) {
 239                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 240                          ns->ns_name);
 241                 lock_vars[0].data = &ns->ns_nr_unused;
 242                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 243                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 244
 245                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 246                          ns->ns_name);
 247                 lock_vars[0].data = ns;
 248                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 249                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 250                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 251
 252                 snprintf(lock_name, MAX_STRING_SIZE, "%s/shrink_thumb",
 253                          ns->ns_name);
 254                 lock_vars[0].data = ns;
 255                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 256                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 257                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 258
 259                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 260                          ns->ns_name);
 261                 lock_vars[0].data = &ns->ns_max_age;
 262                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 263                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 264                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 265         }
 266 }
 267 #undef MAX_STRING_SIZE
 268 #else
 269 #define ldlm_proc_namespace(ns) do {} while (0)
 270 #endif /* LPROCFS */
 271
 272 struct ldlm_namespace *ldlm_namespace_new(char *name, ldlm_side_t client,
 273                                           ldlm_appetite_t apt)
 274 {
 275         struct ldlm_namespace *ns = NULL;
 276         struct list_head *bucket;
 277         int rc, idx, namelen;
 278         ENTRY;
 279
 280         rc = ldlm_get_ref();
 281         if (rc) {
 282                 CERROR("ldlm_get_ref failed: %d\n", rc);
 283                 RETURN(NULL);
 284         }
 285
 286         OBD_ALLOC_PTR(ns);
 287         if (!ns)
 288                 GOTO(out_ref, NULL);
 289
 290         OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 291         if (!ns->ns_hash)
 292                 GOTO(out_ns, NULL);
 293
 294         ns->ns_shrink_thumb = LDLM_LOCK_SHRINK_THUMB;
 295         ns->ns_appetite = apt;
 296         namelen = strlen(name);
 297         OBD_ALLOC(ns->ns_name, namelen + 1);
 298         if (!ns->ns_name)
 299                 GOTO(out_hash, NULL);
 300
 301         strcpy(ns->ns_name, name);
 302
 303         CFS_INIT_LIST_HEAD(&ns->ns_root_list);
 304         ns->ns_refcount = 0;
 305         ns->ns_client = client;
 306         spin_lock_init(&ns->ns_hash_lock);
 307         atomic_set(&ns->ns_locks, 0);
 308         ns->ns_resources = 0;
 309         cfs_waitq_init(&ns->ns_waitq);
 310
 311         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 312              bucket--)
 313                 CFS_INIT_LIST_HEAD(bucket);
 314
 315         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 316         ns->ns_nr_unused = 0;
 317         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
 318         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
 319         spin_lock_init(&ns->ns_unused_lock);
 320         ns->ns_orig_connect_flags = 0;
 321         ns->ns_connect_flags = 0;
 322         ldlm_proc_namespace(ns);
 323
 324         idx = atomic_read(ldlm_namespace_nr(client));
 325         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 326         if (rc) {
 327                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 328                 GOTO(out_proc, rc);
 329         }
 330
 331         mutex_down(ldlm_namespace_lock(client));
 332         list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 333         atomic_inc(ldlm_namespace_nr(client));
 334         mutex_up(ldlm_namespace_lock(client));
 335
 336         RETURN(ns);
 337 out_proc:
 338         ldlm_namespace_cleanup(ns, 0);
 339         OBD_FREE(ns->ns_name, namelen + 1);
 340 out_hash:
 341         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 342 out_ns:
 343         OBD_FREE_PTR(ns);
 344 out_ref:
 345         ldlm_put_ref(0);
 346         RETURN(NULL);
 347 }
 348
 349 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 350
 351 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 352  * This is currently only used for recovery, and we make certain assumptions
 353  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 354  *
 355  * Called with the ns_lock held. */
 356 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 357                              int flags)
 358 {
 359         struct list_head *tmp;
 360         int rc = 0, client = ns_is_client(res->lr_namespace);
 361         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 362         ENTRY;
 363
 364
 365         do {
 366                 struct ldlm_lock *lock = NULL;
 367
 368                 /* first, we look for non-cleaned-yet lock
 369                  * all cleaned locks are marked by CLEANED flag */
 370                 lock_res(res);
 371                 list_for_each(tmp, q) {
 372                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 373                         if (lock->l_flags & LDLM_FL_CLEANED) {
 374                                 lock = NULL;
 375                                 continue;
 376                         }
 377                         LDLM_LOCK_GET(lock);
 378                         lock->l_flags |= LDLM_FL_CLEANED;
 379                         break;
 380                 }
 381
 382                 if (lock == NULL) {
 383                         unlock_res(res);
 384                         break;
 385                 }
 386
 387                 /* Set CBPENDING so nothing in the cancellation path
 388                  * can match this lock */
 389                 lock->l_flags |= LDLM_FL_CBPENDING;
 390                 lock->l_flags |= LDLM_FL_FAILED;
 391                 lock->l_flags |= flags;
 392
 393                 /* ... without sending a CANCEL message for local_only. */
 394                 if (local_only)
 395                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 396
 397                 if (local_only && (lock->l_readers || lock->l_writers)) {
 398                         /* This is a little bit gross, but much better than the
 399                          * alternative: pretend that we got a blocking AST from
 400                          * the server, so that when the lock is decref'd, it
 401                          * will go away ... */
 402                         unlock_res(res);
 403                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 404                         if (lock->l_completion_ast)
 405                                 lock->l_completion_ast(lock, 0, NULL);
 406                         LDLM_LOCK_PUT(lock);
 407                         continue;
 408                 }
 409
 410                 if (client) {
 411                         struct lustre_handle lockh;
 412
 413                         unlock_res(res);
 414                         ldlm_lock2handle(lock, &lockh);
 415                         rc = ldlm_cli_cancel(&lockh);
 416                         if (rc)
 417                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 418                 } else {
 419                         ldlm_resource_unlink_lock(lock);
 420                         unlock_res(res);
 421                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 422                                    "client node");
 423                         ldlm_lock_destroy(lock);
 424                 }
 425                 LDLM_LOCK_PUT(lock);
 426         } while (1);
 427
 428         EXIT;
 429 }
 430
 431 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 432 {
 433         struct list_head *tmp;
 434         int i;
 435
 436         if (ns == NULL) {
 437                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 438                 return ELDLM_OK;
 439         }
 440
 441         for (i = 0; i < RES_HASH_SIZE; i++) {
 442                 spin_lock(&ns->ns_hash_lock);
 443                 tmp = ns->ns_hash[i].next;
 444                 while (tmp != &(ns->ns_hash[i])) {
 445                         struct ldlm_resource *res;
 446                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 447                         ldlm_resource_getref(res);
 448                         spin_unlock(&ns->ns_hash_lock);
 449
 450                         cleanup_resource(res, &res->lr_granted, flags);
 451                         cleanup_resource(res, &res->lr_converting, flags);
 452                         cleanup_resource(res, &res->lr_waiting, flags);
 453
 454                         spin_lock(&ns->ns_hash_lock);
 455                         tmp  = tmp->next;
 456
 457                         /* XXX: former stuff caused issues in case of race
 458                          * between ldlm_namespace_cleanup() and lockd() when
 459                          * client gets blocking ast when lock gets distracted by
 460                          * server. This is 1_4 branch solution, let's see how
 461                          * will it behave. */
 462                         if (!ldlm_resource_putref_locked(res))
 463                                 CDEBUG(D_INFO,
 464                                        "Namespace %s resource refcount nonzero "
 465                                        "(%d) after lock cleanup; forcing cleanup.\n",
 466                                        ns->ns_name, atomic_read(&res->lr_refcount));
 467                 }
 468                 spin_unlock(&ns->ns_hash_lock);
 469         }
 470
 471         return ELDLM_OK;
 472 }
 473
 474 /* Cleanup, but also free, the namespace */
 475 int ldlm_namespace_free_prior(struct ldlm_namespace *ns)
 476 {
 477         ENTRY;
 478         if (!ns)
 479                 RETURN(ELDLM_OK);
 480
 481         mutex_down(ldlm_namespace_lock(ns->ns_client));
 482         list_del(&ns->ns_list_chain);
 483         atomic_dec(ldlm_namespace_nr(ns->ns_client));
 484         ldlm_pool_fini(&ns->ns_pool);
 485         mutex_up(ldlm_namespace_lock(ns->ns_client));
 486
 487         /* At shutdown time, don't call the cancellation callback */
 488         ldlm_namespace_cleanup(ns, 0);
 489
 490         if (ns->ns_refcount > 0) {
 491                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 492                 int rc;
 493                 CDEBUG(D_DLMTRACE,
 494                        "dlm namespace %s free waiting on refcount %d\n",
 495                        ns->ns_name, ns->ns_refcount);
 496                 rc = l_wait_event(ns->ns_waitq,
 497                                   ns->ns_refcount == 0, &lwi);
 498                 if (ns->ns_refcount)
 499                         LCONSOLE_ERROR_MSG(0x139, "Lock manager: wait for %s "
 500                                            "namespace cleanup aborted with %d "
 501                                            "resources in use. (%d)\nI'm going "
 502                                            "to try to clean up anyway, but I "
 503                                            "might need a reboot of this node.\n",
 504                                             ns->ns_name, (int) ns->ns_refcount,
 505                                             rc);
 506                 CDEBUG(D_DLMTRACE,
 507                        "dlm namespace %s free done waiting\n", ns->ns_name);
 508         }
 509
 510         RETURN(ELDLM_OK);
 511 }
 512
 513 int ldlm_namespace_free_post(struct ldlm_namespace *ns, int force)
 514 {
 515         ENTRY;
 516         if (!ns)
 517                 RETURN(ELDLM_OK);
 518
 519 #ifdef LPROCFS
 520         {
 521                 struct proc_dir_entry *dir;
 522                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
 523                 if (dir == NULL) {
 524                         CERROR("dlm namespace %s has no procfs dir?\n",
 525                                ns->ns_name);
 526                 } else {
 527                         lprocfs_remove(&dir);
 528                 }
 529         }
 530 #endif
 531
 532         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 533         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 534         OBD_FREE_PTR(ns);
 535         ldlm_put_ref(force);
 536         RETURN(ELDLM_OK);
 537 }
 538
 539
 540 /* Cleanup the resource, and free namespace.
 541  * bug 12864:
 542  * Deadlock issue:
 543  * proc1: destroy import
 544  *        class_disconnect_export(grab cl_sem) ->
 545  *              -> ldlm_namespace_free ->
 546  *              -> lprocfs_remove(grab _lprocfs_lock).
 547  * proc2: read proc info
 548  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 549  *              -> osc_rd_active, etc(grab cl_sem).
 550  *
 551  * So that I have to split the ldlm_namespace_free into two parts - the first
 552  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 553  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 554  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 555  * held.
 556  */
 557 int ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 558 {
 559         ldlm_namespace_free_prior(ns);
 560         ldlm_namespace_free_post(ns, force);
 561         return ELDLM_OK;
 562 }
 563
 564
 565 void ldlm_namespace_get_nolock(struct ldlm_namespace *ns)
 566 {
 567         LASSERT(ns->ns_refcount >= 0);
 568         ns->ns_refcount++;
 569 }
 570
 571 void ldlm_namespace_get(struct ldlm_namespace *ns)
 572 {
 573         spin_lock(&ns->ns_hash_lock);
 574         ldlm_namespace_get_nolock(ns);
 575         spin_unlock(&ns->ns_hash_lock);
 576 }
 577
 578 void ldlm_namespace_put_nolock(struct ldlm_namespace *ns, int wakeup)
 579 {
 580         LASSERT(ns->ns_refcount > 0);
 581         ns->ns_refcount--;
 582         if (ns->ns_refcount == 0 && wakeup)
 583                 wake_up(&ns->ns_waitq);
 584 }
 585
 586 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
 587 {
 588         spin_lock(&ns->ns_hash_lock);
 589         ldlm_namespace_put_nolock(ns, wakeup);
 590         spin_unlock(&ns->ns_hash_lock);
 591 }
 592
 593 /* Should be called under ldlm_namespace_lock(client) taken */
 594 void ldlm_namespace_move(struct ldlm_namespace *ns, ldlm_side_t client)
 595 {
 596         LASSERT(!list_empty(&ns->ns_list_chain));
 597         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 598         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 599 }
 600
 601 /* Should be called under ldlm_namespace_lock(client) taken */
 602 struct ldlm_namespace *ldlm_namespace_first(ldlm_side_t client)
 603 {
 604         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 605         LASSERT(!list_empty(ldlm_namespace_list(client)));
 606         return container_of(ldlm_namespace_list(client)->next,
 607                 struct ldlm_namespace, ns_list_chain);
 608 }
 609 static __u32 ldlm_hash_fn(struct ldlm_resource *parent,
 610                           const struct ldlm_res_id *name)
 611 {
 612         __u32 hash = 0;
 613         int i;
 614
 615         for (i = 0; i < RES_NAME_SIZE; i++)
 616                 hash += name->name[i];
 617
 618         hash += (__u32)((unsigned long)parent >> 4);
 619
 620         return (hash & RES_HASH_MASK);
 621 }
 622
 623 static struct ldlm_resource *ldlm_resource_new(void)
 624 {
 625         struct ldlm_resource *res;
 626         int idx;
 627
 628         OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
 629         if (res == NULL)
 630                 return NULL;
 631
 632         memset(res, 0, sizeof(*res));
 633
 634         CFS_INIT_LIST_HEAD(&res->lr_children);
 635         CFS_INIT_LIST_HEAD(&res->lr_childof);
 636         CFS_INIT_LIST_HEAD(&res->lr_granted);
 637         CFS_INIT_LIST_HEAD(&res->lr_converting);
 638         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 639
 640         /* initialize interval trees for each lock mode*/
 641         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 642                 res->lr_itree[idx].lit_size = 0;
 643                 res->lr_itree[idx].lit_mode = 1 << idx;
 644                 res->lr_itree[idx].lit_root = NULL;
 645         }
 646
 647         atomic_set(&res->lr_refcount, 1);
 648         spin_lock_init(&res->lr_lock);
 649
 650         /* one who creates the resource must unlock
 651          * the semaphore after lvb initialization */
 652         init_MUTEX_LOCKED(&res->lr_lvb_sem);
 653
 654         return res;
 655 }
 656
 657 /* must be called with hash lock held */
 658 static struct ldlm_resource *
 659 ldlm_resource_find(struct ldlm_namespace *ns, const struct ldlm_res_id *name,
 660                    __u32 hash)
 661 {
 662         struct list_head *bucket, *tmp;
 663         struct ldlm_resource *res;
 664
 665         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 666         bucket = ns->ns_hash + hash;
 667
 668         list_for_each(tmp, bucket) {
 669                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 670                 if (memcmp(&res->lr_name, name, sizeof(res->lr_name)) == 0)
 671                         return res;
 672         }
 673
 674         return NULL;
 675 }
 676
 677 /* Args: locked namespace
 678  * Returns: newly-allocated, referenced, unlocked resource */
 679 static struct ldlm_resource *
 680 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 681                   const struct ldlm_res_id *name, __u32 hash, ldlm_type_t type)
 682 {
 683         struct list_head *bucket;
 684         struct ldlm_resource *res, *old_res;
 685         ENTRY;
 686
 687         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
 688                  "type: %d\n", type);
 689
 690         res = ldlm_resource_new();
 691         if (!res)
 692                 RETURN(NULL);
 693
 694         res->lr_name = *name;
 695         res->lr_namespace = ns;
 696         res->lr_type = type;
 697         res->lr_most_restr = LCK_NL;
 698
 699         spin_lock(&ns->ns_hash_lock);
 700         old_res = ldlm_resource_find(ns, name, hash);
 701         if (old_res) {
 702                 /* someone won the race and added the resource before */
 703                 ldlm_resource_getref(old_res);
 704                 spin_unlock(&ns->ns_hash_lock);
 705                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 706                 /* synchronize WRT resource creation */
 707                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 708                         down(&old_res->lr_lvb_sem);
 709                         up(&old_res->lr_lvb_sem);
 710                 }
 711                 RETURN(old_res);
 712         }
 713
 714         /* we won! let's add the resource */
 715         bucket = ns->ns_hash + hash;
 716         list_add(&res->lr_hash, bucket);
 717         ns->ns_resources++;
 718         ldlm_namespace_get_nolock(ns);
 719
 720         if (parent == NULL) {
 721                 list_add(&res->lr_childof, &ns->ns_root_list);
 722         } else {
 723                 res->lr_parent = parent;
 724                 list_add(&res->lr_childof, &parent->lr_children);
 725         }
 726         spin_unlock(&ns->ns_hash_lock);
 727
 728         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 729                 int rc;
 730
 731                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
 732                 rc = ns->ns_lvbo->lvbo_init(res);
 733                 if (rc)
 734                         CERROR("lvbo_init failed for resource "
 735                                LPU64": rc %d\n", name->name[0], rc);
 736                 /* we create resource with locked lr_lvb_sem */
 737                 up(&res->lr_lvb_sem);
 738         }
 739
 740         RETURN(res);
 741 }
 742
 743 /* Args: unlocked namespace
 744  * Locks: takes and releases ns->ns_lock and res->lr_lock
 745  * Returns: referenced, unlocked ldlm_resource or NULL */
 746 struct ldlm_resource *
 747 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 748                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
 749 {
 750         __u32 hash = ldlm_hash_fn(parent, name);
 751         struct ldlm_resource *res = NULL;
 752         ENTRY;
 753
 754         LASSERT(ns != NULL);
 755         LASSERT(ns->ns_hash != NULL);
 756         LASSERT(name->name[0] != 0);
 757
 758         spin_lock(&ns->ns_hash_lock);
 759         res = ldlm_resource_find(ns, name, hash);
 760         if (res) {
 761                 ldlm_resource_getref(res);
 762                 spin_unlock(&ns->ns_hash_lock);
 763                 /* synchronize WRT resource creation */
 764                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 765                         down(&res->lr_lvb_sem);
 766                         up(&res->lr_lvb_sem);
 767                 }
 768                 RETURN(res);
 769         }
 770         spin_unlock(&ns->ns_hash_lock);
 771
 772         if (create == 0)
 773                 RETURN(NULL);
 774
 775         res = ldlm_resource_add(ns, parent, name, hash, type);
 776         RETURN(res);
 777 }
 778
 779 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 780 {
 781         LASSERT(res != NULL);
 782         LASSERT(res != LP_POISON);
 783         atomic_inc(&res->lr_refcount);
 784         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 785                atomic_read(&res->lr_refcount));
 786         return res;
 787 }
 788
 789 void __ldlm_resource_putref_final(struct ldlm_resource *res)
 790 {
 791         struct ldlm_namespace *ns = res->lr_namespace;
 792
 793         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 794
 795         if (!list_empty(&res->lr_granted)) {
 796                 ldlm_resource_dump(D_ERROR, res);
 797                 LBUG();
 798         }
 799
 800         if (!list_empty(&res->lr_converting)) {
 801                 ldlm_resource_dump(D_ERROR, res);
 802                 LBUG();
 803         }
 804
 805         if (!list_empty(&res->lr_waiting)) {
 806                 ldlm_resource_dump(D_ERROR, res);
 807                 LBUG();
 808         }
 809
 810         if (!list_empty(&res->lr_children)) {
 811                 ldlm_resource_dump(D_ERROR, res);
 812                 LBUG();
 813         }
 814
 815         /* Pass 0 here to not wake ->ns_waitq up yet, we will do it few
 816          * lines below when all children are freed. */
 817         ldlm_namespace_put_nolock(ns, 0);
 818         list_del_init(&res->lr_hash);
 819         list_del_init(&res->lr_childof);
 820
 821         ns->ns_resources--;
 822         if (ns->ns_resources == 0)
 823                 wake_up(&ns->ns_waitq);
 824 }
 825
 826 /* Returns 1 if the resource was freed, 0 if it remains. */
 827 int ldlm_resource_putref(struct ldlm_resource *res)
 828 {
 829         struct ldlm_namespace *ns = res->lr_namespace;
 830         int rc = 0;
 831         ENTRY;
 832
 833         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 834                atomic_read(&res->lr_refcount) - 1);
 835         LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
 836                  atomic_read(&res->lr_refcount));
 837         LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
 838                  atomic_read(&res->lr_refcount));
 839
 840         if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
 841                 __ldlm_resource_putref_final(res);
 842                 spin_unlock(&ns->ns_hash_lock);
 843                 if (res->lr_lvb_data)
 844                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 845                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 846                 rc = 1;
 847         }
 848
 849         RETURN(rc);
 850 }
 851
 852 /* Returns 1 if the resource was freed, 0 if it remains. */
 853 int ldlm_resource_putref_locked(struct ldlm_resource *res)
 854 {
 855         int rc = 0;
 856         ENTRY;
 857
 858         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 859                atomic_read(&res->lr_refcount) - 1);
 860         LASSERT(atomic_read(&res->lr_refcount) > 0);
 861         LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
 862
 863         LASSERT(atomic_read(&res->lr_refcount) >= 0);
 864         if (atomic_dec_and_test(&res->lr_refcount)) {
 865                 __ldlm_resource_putref_final(res);
 866                 if (res->lr_lvb_data)
 867                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 868                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 869                 rc = 1;
 870         }
 871
 872         RETURN(rc);
 873 }
 874
 875 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 876                             struct ldlm_lock *lock)
 877 {
 878         check_res_locked(res);
 879
 880         ldlm_resource_dump(D_OTHER, res);
 881         CDEBUG(D_OTHER, "About to add this lock:\n");
 882         ldlm_lock_dump(D_OTHER, lock, 0);
 883
 884         if (lock->l_destroyed) {
 885                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 886                 return;
 887         }
 888
 889         LASSERT(list_empty(&lock->l_res_link));
 890
 891         list_add_tail(&lock->l_res_link, head);
 892 }
 893
 894 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
 895                                      struct ldlm_lock *new)
 896 {
 897         struct ldlm_resource *res = original->l_resource;
 898
 899         check_res_locked(res);
 900
 901         ldlm_resource_dump(D_OTHER, res);
 902         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
 903         ldlm_lock_dump(D_OTHER, new, 0);
 904
 905         if (new->l_destroyed) {
 906                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 907                 goto out;
 908         }
 909
 910         LASSERT(list_empty(&new->l_res_link));
 911
 912         list_add(&new->l_res_link, &original->l_res_link);
 913  out:;
 914 }
 915
 916 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 917 {
 918         int type = lock->l_resource->lr_type;
 919
 920         check_res_locked(lock->l_resource);
 921         if (type == LDLM_IBITS || type == LDLM_PLAIN)
 922                 ldlm_unlink_lock_skiplist(lock);
 923         else if (type == LDLM_EXTENT)
 924                 ldlm_extent_unlink_lock(lock);
 925         list_del_init(&lock->l_res_link);
 926 }
 927
 928 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 929 {
 930         desc->lr_type = res->lr_type;
 931         desc->lr_name = res->lr_name;
 932 }
 933
 934 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
 935 {
 936         struct list_head *tmp;
 937
 938         if (!((libcfs_debug | D_ERROR) & level))
 939                 return;
 940
 941         mutex_down(ldlm_namespace_lock(client));
 942
 943         list_for_each(tmp, ldlm_namespace_list(client)) {
 944                 struct ldlm_namespace *ns;
 945                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 946                 ldlm_namespace_dump(level, ns);
 947         }
 948
 949         mutex_up(ldlm_namespace_lock(client));
 950 }
 951
 952 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
 953 {
 954         struct list_head *tmp;
 955
 956         if (!((libcfs_debug | D_ERROR) & level))
 957                 return;
 958
 959         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
 960                ns->ns_name, ns->ns_refcount,
 961                ns_is_client(ns) ? "client" : "server");
 962
 963         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
 964                 return;
 965
 966         spin_lock(&ns->ns_hash_lock);
 967         tmp = ns->ns_root_list.next;
 968         while (tmp != &ns->ns_root_list) {
 969                 struct ldlm_resource *res;
 970                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
 971
 972                 ldlm_resource_getref(res);
 973                 spin_unlock(&ns->ns_hash_lock);
 974
 975                 lock_res(res);
 976                 ldlm_resource_dump(level, res);
 977                 unlock_res(res);
 978
 979                 spin_lock(&ns->ns_hash_lock);
 980                 tmp = tmp->next;
 981                 ldlm_resource_putref_locked(res);
 982         }
 983         ns->ns_next_dump = cfs_time_shift(10);
 984         spin_unlock(&ns->ns_hash_lock);
 985 }
 986
 987 void ldlm_resource_dump(int level, struct ldlm_resource *res)
 988 {
 989         struct list_head *tmp;
 990         int pos;
 991
 992         CLASSERT(RES_NAME_SIZE == 4);
 993
 994         if (!((libcfs_debug | D_ERROR) & level))
 995                 return;
 996
 997         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
 998                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
 999                res->lr_name.name[2], res->lr_name.name[3],
1000                atomic_read(&res->lr_refcount));
1001
1002         if (!list_empty(&res->lr_granted)) {
1003                 pos = 0;
1004                 CDEBUG(level, "Granted locks:\n");
1005                 list_for_each(tmp, &res->lr_granted) {
1006                         struct ldlm_lock *lock;
1007                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1008                         ldlm_lock_dump(level, lock, ++pos);
1009                 }
1010         }
1011         if (!list_empty(&res->lr_converting)) {
1012                 pos = 0;
1013                 CDEBUG(level, "Converting locks:\n");
1014                 list_for_each(tmp, &res->lr_converting) {
1015                         struct ldlm_lock *lock;
1016                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1017                         ldlm_lock_dump(level, lock, ++pos);
1018                 }
1019         }
1020         if (!list_empty(&res->lr_waiting)) {
1021                 pos = 0;
1022                 CDEBUG(level, "Waiting locks:\n");
1023                 list_for_each(tmp, &res->lr_waiting) {
1024                         struct ldlm_lock *lock;
1025                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1026                         ldlm_lock_dump(level, lock, ++pos);
1027                 }
1028         }
1029 }