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 struct semaphore ldlm_namespace_lock;
  40 struct list_head ldlm_namespace_list = CFS_LIST_HEAD_INIT(ldlm_namespace_list);
  41 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
  42 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
  43 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
  44
  45 #ifdef LPROCFS
  46 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  47                              unsigned long count, void *data)
  48 {
  49         ldlm_dump_all_namespaces(D_DLMTRACE);
  50         RETURN(count);
  51 }
  52
  53 int ldlm_proc_setup(void)
  54 {
  55         int rc;
  56         struct lprocfs_vars list[] = {
  57                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  58                 { NULL }};
  59         ENTRY;
  60         LASSERT(ldlm_ns_proc_dir == NULL);
  61
  62         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
  63                                               proc_lustre_root,
  64                                               NULL, NULL);
  65         if (IS_ERR(ldlm_type_proc_dir)) {
  66                 CERROR("LProcFS failed in ldlm-init\n");
  67                 rc = PTR_ERR(ldlm_type_proc_dir);
  68                 GOTO(err, rc);
  69         }
  70
  71         ldlm_ns_proc_dir = lprocfs_register("namespaces",
  72                                             ldlm_type_proc_dir,
  73                                             NULL, NULL);
  74         if (IS_ERR(ldlm_ns_proc_dir)) {
  75                 CERROR("LProcFS failed in ldlm-init\n");
  76                 rc = PTR_ERR(ldlm_ns_proc_dir);
  77                 GOTO(err_type, rc);
  78         }
  79
  80         ldlm_svc_proc_dir = lprocfs_register("services",
  81                                             ldlm_type_proc_dir,
  82                                             NULL, NULL);
  83         if (IS_ERR(ldlm_svc_proc_dir)) {
  84                 CERROR("LProcFS failed in ldlm-init\n");
  85                 rc = PTR_ERR(ldlm_svc_proc_dir);
  86                 GOTO(err_ns, rc);
  87         }
  88
  89         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
  90
  91         RETURN(0);
  92
  93 err_ns:
  94         lprocfs_remove(&ldlm_ns_proc_dir);
  95 err_type:
  96         lprocfs_remove(&ldlm_type_proc_dir);
  97 err:
  98         ldlm_svc_proc_dir = NULL;
  99         RETURN(rc);
 100 }
 101
 102 void ldlm_proc_cleanup(void)
 103 {
 104         if (ldlm_svc_proc_dir)
 105                 lprocfs_remove(&ldlm_svc_proc_dir);
 106
 107         if (ldlm_ns_proc_dir)
 108                 lprocfs_remove(&ldlm_ns_proc_dir);
 109
 110         if (ldlm_type_proc_dir)
 111                 lprocfs_remove(&ldlm_type_proc_dir);
 112 }
 113
 114 static int lprocfs_uint_rd(char *page, char **start, off_t off,
 115                            int count, int *eof, void *data)
 116 {
 117         unsigned int *temp = (unsigned int *)data;
 118         return snprintf(page, count, "%u\n", *temp);
 119 }
 120
 121 #define MAX_STRING_SIZE 128
 122 static int lprocfs_uint_wr(struct file *file, const char *buffer,
 123                            unsigned long count, void *data)
 124 {
 125         unsigned *p = data;
 126         char dummy[MAX_STRING_SIZE + 1], *end;
 127         unsigned long tmp;
 128
 129         dummy[MAX_STRING_SIZE] = '\0';
 130         if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 131                 return -EFAULT;
 132
 133         tmp = simple_strtoul(dummy, &end, 0);
 134         if (dummy == end)
 135                 return -EINVAL;
 136
 137         *p = (unsigned int)tmp;
 138         return count;
 139 }
 140
 141 static int lprocfs_read_lru_size(char *page, char **start, off_t off,
 142                                  int count, int *eof, void *data)
 143 {
 144         struct ldlm_namespace *ns = data;
 145         return lprocfs_uint_rd(page, start, off, count, eof,
 146                                &ns->ns_max_unused);
 147 }
 148
 149 static int lprocfs_write_lru_size(struct file *file, const char *buffer,
 150                                   unsigned long count, void *data)
 151 {
 152         struct ldlm_namespace *ns = data;
 153         char dummy[MAX_STRING_SIZE + 1], *end;
 154         unsigned long tmp;
 155
 156         dummy[MAX_STRING_SIZE] = '\0';
 157         if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 158                 return -EFAULT;
 159
 160         if (count == 6 && memcmp(dummy, "clear", 5) == 0) {
 161                 CDEBUG(D_DLMTRACE,
 162                        "dropping all unused locks from namespace %s\n",
 163                        ns->ns_name);
 164                 tmp = ns->ns_max_unused;
 165                 ns->ns_max_unused = 0;
 166                 ldlm_cancel_lru(ns, LDLM_SYNC);
 167                 ns->ns_max_unused = tmp;
 168                 return count;
 169         }
 170
 171         tmp = simple_strtoul(dummy, &end, 0);
 172         if (tmp == 0 && *end) {
 173                 CERROR("invalid value written\n");
 174                 return -EINVAL;
 175         }
 176
 177         CDEBUG(D_DLMTRACE, "changing namespace %s max_unused from %u to %u\n",
 178                ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
 179         ns->ns_max_unused = (unsigned int)tmp;
 180
 181         ldlm_cancel_lru(ns, LDLM_ASYNC);
 182
 183         return count;
 184 }
 185
 186 void ldlm_proc_namespace(struct ldlm_namespace *ns)
 187 {
 188         struct lprocfs_vars lock_vars[2];
 189         char lock_name[MAX_STRING_SIZE + 1];
 190
 191         LASSERT(ns != NULL);
 192         LASSERT(ns->ns_name != NULL);
 193
 194         lock_name[MAX_STRING_SIZE] = '\0';
 195
 196         memset(lock_vars, 0, sizeof(lock_vars));
 197         lock_vars[0].name = lock_name;
 198
 199         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
 200         lock_vars[0].data = &ns->ns_refcount;
 201         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 202         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 203
 204         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
 205         lock_vars[0].data = &ns->ns_locks;
 206         lock_vars[0].read_fptr = lprocfs_rd_atomic;
 207         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 208
 209         if (ns->ns_client) {
 210                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 211                          ns->ns_name);
 212                 lock_vars[0].data = &ns->ns_nr_unused;
 213                 lock_vars[0].read_fptr = lprocfs_uint_rd;
 214                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 215
 216                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 217                          ns->ns_name);
 218                 lock_vars[0].data = ns;
 219                 lock_vars[0].read_fptr = lprocfs_read_lru_size;
 220                 lock_vars[0].write_fptr = lprocfs_write_lru_size;
 221                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 222
 223                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 224                          ns->ns_name);
 225                 lock_vars[0].data = &ns->ns_max_age;
 226                 lock_vars[0].read_fptr = lprocfs_uint_rd;
 227                 lock_vars[0].write_fptr = lprocfs_uint_wr;
 228                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 229         }
 230 }
 231 #undef MAX_STRING_SIZE
 232 #else
 233 #define ldlm_proc_namespace(ns) do {} while (0)
 234 #endif /* LPROCFS */
 235
 236 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
 237 {
 238         struct ldlm_namespace *ns = NULL;
 239         struct list_head *bucket;
 240         int rc;
 241         ENTRY;
 242
 243         rc = ldlm_get_ref();
 244         if (rc) {
 245                 CERROR("ldlm_get_ref failed: %d\n", rc);
 246                 RETURN(NULL);
 247         }
 248
 249         OBD_ALLOC(ns, sizeof(*ns));
 250         if (!ns)
 251                 GOTO(out_ref, NULL);
 252
 253         OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 254         if (!ns->ns_hash)
 255                 GOTO(out_ns, NULL);
 256
 257         OBD_ALLOC(ns->ns_name, strlen(name) + 1);
 258         if (!ns->ns_name)
 259                 GOTO(out_hash, NULL);
 260
 261         strcpy(ns->ns_name, name);
 262
 263         CFS_INIT_LIST_HEAD(&ns->ns_root_list);
 264         ns->ns_refcount = 0;
 265         ns->ns_client = client;
 266         spin_lock_init(&ns->ns_hash_lock);
 267         atomic_set(&ns->ns_locks, 0);
 268         ns->ns_resources = 0;
 269         cfs_waitq_init(&ns->ns_waitq);
 270
 271         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 272              bucket--)
 273                 CFS_INIT_LIST_HEAD(bucket);
 274
 275         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 276         ns->ns_nr_unused = 0;
 277         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
 278         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
 279         spin_lock_init(&ns->ns_unused_lock);
 280
 281         mutex_down(&ldlm_namespace_lock);
 282         list_add(&ns->ns_list_chain, &ldlm_namespace_list);
 283         mutex_up(&ldlm_namespace_lock);
 284         ldlm_proc_namespace(ns);
 285         RETURN(ns);
 286
 287 out_hash:
 288         POISON(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 289         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 290 out_ns:
 291         OBD_FREE(ns, sizeof(*ns));
 292 out_ref:
 293         ldlm_put_ref(0);
 294         RETURN(NULL);
 295 }
 296
 297 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 298
 299 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 300  * This is currently only used for recovery, and we make certain assumptions
 301  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 302  *
 303  * Called with the ns_lock held. */
 304 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 305                              int flags)
 306 {
 307         struct list_head *tmp;
 308         int rc = 0, client = res->lr_namespace->ns_client;
 309         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 310         ENTRY;
 311
 312
 313         do {
 314                 struct ldlm_lock *lock = NULL;
 315
 316                 /* first, we look for non-cleaned-yet lock
 317                  * all cleaned locks are marked by CLEANED flag */
 318                 lock_res(res);
 319                 list_for_each(tmp, q) {
 320                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 321                         if (lock->l_flags & LDLM_FL_CLEANED) {
 322                                 lock = NULL;
 323                                 continue;
 324                         }
 325                         LDLM_LOCK_GET(lock);
 326                         lock->l_flags |= LDLM_FL_CLEANED;
 327                         break;
 328                 }
 329
 330                 if (lock == NULL) {
 331                         unlock_res(res);
 332                         break;
 333                 }
 334
 335                 /* Set CBPENDING so nothing in the cancellation path
 336                  * can match this lock */
 337                 lock->l_flags |= LDLM_FL_CBPENDING;
 338                 lock->l_flags |= LDLM_FL_FAILED;
 339                 lock->l_flags |= flags;
 340
 341                 /* ... without sending a CANCEL message for local_only. */
 342                 if (local_only)
 343                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 344
 345                 if (local_only && (lock->l_readers || lock->l_writers)) {
 346                         /* This is a little bit gross, but much better than the
 347                          * alternative: pretend that we got a blocking AST from
 348                          * the server, so that when the lock is decref'd, it
 349                          * will go away ... */
 350                         unlock_res(res);
 351                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 352                         if (lock->l_completion_ast)
 353                                 lock->l_completion_ast(lock, 0, NULL);
 354                         LDLM_LOCK_PUT(lock);
 355                         continue;
 356                 }
 357
 358                 if (client) {
 359                         struct lustre_handle lockh;
 360
 361                         unlock_res(res);
 362                         ldlm_lock2handle(lock, &lockh);
 363                         rc = ldlm_cli_cancel(&lockh);
 364                         if (rc)
 365                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 366                 } else {
 367                         ldlm_resource_unlink_lock(lock);
 368                         unlock_res(res);
 369                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 370                                    "client node");
 371                         ldlm_lock_destroy(lock);
 372                 }
 373                 LDLM_LOCK_PUT(lock);
 374         } while (1);
 375
 376         EXIT;
 377 }
 378
 379 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 380 {
 381         struct list_head *tmp;
 382         int i;
 383
 384         if (ns == NULL) {
 385                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 386                 return ELDLM_OK;
 387         }
 388
 389         for (i = 0; i < RES_HASH_SIZE; i++) {
 390                 spin_lock(&ns->ns_hash_lock);
 391                 tmp = ns->ns_hash[i].next;
 392                 while (tmp != &(ns->ns_hash[i])) {
 393                         struct ldlm_resource *res;
 394                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 395                         ldlm_resource_getref(res);
 396                         spin_unlock(&ns->ns_hash_lock);
 397
 398                         cleanup_resource(res, &res->lr_granted, flags);
 399                         cleanup_resource(res, &res->lr_converting, flags);
 400                         cleanup_resource(res, &res->lr_waiting, flags);
 401
 402                         spin_lock(&ns->ns_hash_lock);
 403                         tmp  = tmp->next;
 404
 405                         /* XXX: former stuff caused issues in case of race
 406                          * between ldlm_namespace_cleanup() and lockd() when
 407                          * client gets blocking ast when lock gets distracted by
 408                          * server. This is 1_4 branch solution, let's see how
 409                          * will it behave. */
 410                         if (!ldlm_resource_putref_locked(res))
 411                                 CDEBUG(D_INFO,
 412                                        "Namespace %s resource refcount nonzero "
 413                                        "(%d) after lock cleanup; forcing cleanup.\n",
 414                                        ns->ns_name, atomic_read(&res->lr_refcount));
 415                 }
 416                 spin_unlock(&ns->ns_hash_lock);
 417         }
 418
 419         return ELDLM_OK;
 420 }
 421
 422 int ldlm_namespace_free_prior(struct ldlm_namespace *ns)
 423 {
 424         ENTRY;
 425         if (!ns)
 426                 RETURN(ELDLM_OK);
 427
 428         mutex_down(&ldlm_namespace_lock);
 429         list_del(&ns->ns_list_chain);
 430         mutex_up(&ldlm_namespace_lock);
 431
 432         /* At shutdown time, don't call the cancellation callback */
 433         ldlm_namespace_cleanup(ns, 0);
 434
 435         if (ns->ns_refcount > 0) {
 436                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 437                 int rc;
 438                 CDEBUG(D_DLMTRACE,
 439                        "dlm namespace %s free waiting on refcount %d\n",
 440                        ns->ns_name, ns->ns_refcount);
 441                 rc = l_wait_event(ns->ns_waitq,
 442                                   ns->ns_refcount == 0, &lwi);
 443                 if (ns->ns_refcount)
 444                         LCONSOLE_ERROR_MSG(0x139, "Lock manager: wait for %s "
 445                                            "namespace cleanup aborted with %d "
 446                                            "resources in use. (%d)\nI'm going "
 447                                            "to try to clean up anyway, but I "
 448                                            "might need a reboot of this node.\n",
 449                                             ns->ns_name, (int) ns->ns_refcount,
 450                                             rc);
 451                 CDEBUG(D_DLMTRACE,
 452                        "dlm namespace %s free done waiting\n", ns->ns_name);
 453         }
 454
 455         RETURN(ELDLM_OK);
 456 }
 457
 458 int ldlm_namespace_free_post(struct ldlm_namespace *ns, int force)
 459 {
 460         ENTRY;
 461         if (!ns)
 462                 RETURN(ELDLM_OK);
 463
 464 #ifdef LPROCFS
 465         {
 466                 struct proc_dir_entry *dir;
 467                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
 468                 if (dir == NULL) {
 469                         CERROR("dlm namespace %s has no procfs dir?\n",
 470                                ns->ns_name);
 471                 } else {
 472                         lprocfs_remove(&dir);
 473                 }
 474         }
 475 #endif
 476
 477         POISON(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 478         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 479         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 480         OBD_FREE(ns, sizeof(*ns));
 481
 482         ldlm_put_ref(force);
 483         RETURN(ELDLM_OK);
 484 }
 485
 486 /* Cleanup the resource, and free namespace.
 487  * bug 12864:
 488  * Deadlock issue:
 489  * proc1: destroy import
 490  *        class_disconnect_export(grab cl_sem) ->
 491  *              -> ldlm_namespace_free ->
 492  *              -> lprocfs_remove(grab _lprocfs_lock).
 493  * proc2: read proc info
 494  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 495  *              -> osc_rd_active, etc(grab cl_sem).
 496  *
 497  * So that I have to split the ldlm_namespace_free into two parts - the first
 498  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 499  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 500  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 501  * held.
 502  */
 503 int ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 504 {
 505         ldlm_namespace_free_prior(ns);
 506         ldlm_namespace_free_post(ns, force);
 507         return ELDLM_OK;
 508 }
 509
 510 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, struct ldlm_res_id name)
 511 {
 512         __u32 hash = 0;
 513         int i;
 514
 515         for (i = 0; i < RES_NAME_SIZE; i++)
 516                 hash += name.name[i];
 517
 518         hash += (__u32)((unsigned long)parent >> 4);
 519
 520         return (hash & RES_HASH_MASK);
 521 }
 522
 523 static struct ldlm_resource *ldlm_resource_new(void)
 524 {
 525         struct ldlm_resource *res;
 526
 527         OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
 528         if (res == NULL)
 529                 return NULL;
 530
 531         memset(res, 0, sizeof(*res));
 532
 533         CFS_INIT_LIST_HEAD(&res->lr_children);
 534         CFS_INIT_LIST_HEAD(&res->lr_childof);
 535         CFS_INIT_LIST_HEAD(&res->lr_granted);
 536         CFS_INIT_LIST_HEAD(&res->lr_converting);
 537         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 538         atomic_set(&res->lr_refcount, 1);
 539         spin_lock_init(&res->lr_lock);
 540
 541         /* one who creates the resource must unlock
 542          * the semaphore after lvb initialization */
 543         init_MUTEX_LOCKED(&res->lr_lvb_sem);
 544
 545         return res;
 546 }
 547
 548 /* must be called with hash lock held */
 549 static struct ldlm_resource *
 550 ldlm_resource_find(struct ldlm_namespace *ns, struct ldlm_res_id name, __u32 hash)
 551 {
 552         struct list_head *bucket, *tmp;
 553         struct ldlm_resource *res;
 554
 555         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 556         bucket = ns->ns_hash + hash;
 557
 558         list_for_each(tmp, bucket) {
 559                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 560                 if (memcmp(&res->lr_name, &name, sizeof(res->lr_name)) == 0)
 561                         return res;
 562         }
 563
 564         return NULL;
 565 }
 566
 567 /* Args: locked namespace
 568  * Returns: newly-allocated, referenced, unlocked resource */
 569 static struct ldlm_resource *
 570 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 571                   struct ldlm_res_id name, __u32 hash, ldlm_type_t type)
 572 {
 573         struct list_head *bucket;
 574         struct ldlm_resource *res, *old_res;
 575         ENTRY;
 576
 577         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
 578                  "type: %d\n", type);
 579
 580         res = ldlm_resource_new();
 581         if (!res)
 582                 RETURN(NULL);
 583
 584         res->lr_name = name;
 585         res->lr_namespace = ns;
 586         res->lr_type = type;
 587         res->lr_most_restr = LCK_NL;
 588
 589         spin_lock(&ns->ns_hash_lock);
 590         old_res = ldlm_resource_find(ns, name, hash);
 591         if (old_res) {
 592                 /* someone won the race and added the resource before */
 593                 ldlm_resource_getref(old_res);
 594                 spin_unlock(&ns->ns_hash_lock);
 595                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 596                 /* synchronize WRT resource creation */
 597                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 598                         down(&old_res->lr_lvb_sem);
 599                         up(&old_res->lr_lvb_sem);
 600                 }
 601                 RETURN(old_res);
 602         }
 603
 604         /* we won! let's add the resource */
 605         bucket = ns->ns_hash + hash;
 606         list_add(&res->lr_hash, bucket);
 607         ns->ns_resources++;
 608         ns->ns_refcount++;
 609
 610         if (parent == NULL) {
 611                 list_add(&res->lr_childof, &ns->ns_root_list);
 612         } else {
 613                 res->lr_parent = parent;
 614                 list_add(&res->lr_childof, &parent->lr_children);
 615         }
 616         spin_unlock(&ns->ns_hash_lock);
 617
 618         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 619                 int rc;
 620
 621                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
 622                 rc = ns->ns_lvbo->lvbo_init(res);
 623                 if (rc)
 624                         CERROR("lvbo_init failed for resource "
 625                                LPU64": rc %d\n", name.name[0], rc);
 626                 /* we create resource with locked lr_lvb_sem */
 627                 up(&res->lr_lvb_sem);
 628         }
 629
 630         RETURN(res);
 631 }
 632
 633 /* Args: unlocked namespace
 634  * Locks: takes and releases ns->ns_lock and res->lr_lock
 635  * Returns: referenced, unlocked ldlm_resource or NULL */
 636 struct ldlm_resource *
 637 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 638                   struct ldlm_res_id name, ldlm_type_t type, int create)
 639 {
 640         __u32 hash = ldlm_hash_fn(parent, name);
 641         struct ldlm_resource *res = NULL;
 642         ENTRY;
 643
 644         LASSERT(ns != NULL);
 645         LASSERT(ns->ns_hash != NULL);
 646         LASSERT(name.name[0] != 0);
 647
 648         spin_lock(&ns->ns_hash_lock);
 649         res = ldlm_resource_find(ns, name, hash);
 650         if (res) {
 651                 ldlm_resource_getref(res);
 652                 spin_unlock(&ns->ns_hash_lock);
 653                 /* synchronize WRT resource creation */
 654                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
 655                         down(&res->lr_lvb_sem);
 656                         up(&res->lr_lvb_sem);
 657                 }
 658                 RETURN(res);
 659         }
 660         spin_unlock(&ns->ns_hash_lock);
 661
 662         if (create == 0)
 663                 RETURN(NULL);
 664
 665         res = ldlm_resource_add(ns, parent, name, hash, type);
 666         RETURN(res);
 667 }
 668
 669 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 670 {
 671         LASSERT(res != NULL);
 672         LASSERT(res != LP_POISON);
 673         atomic_inc(&res->lr_refcount);
 674         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 675                atomic_read(&res->lr_refcount));
 676         return res;
 677 }
 678
 679 void __ldlm_resource_putref_final(struct ldlm_resource *res)
 680 {
 681         struct ldlm_namespace *ns = res->lr_namespace;
 682
 683         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
 684
 685         if (!list_empty(&res->lr_granted)) {
 686                 ldlm_resource_dump(D_ERROR, res);
 687                 LBUG();
 688         }
 689
 690         if (!list_empty(&res->lr_converting)) {
 691                 ldlm_resource_dump(D_ERROR, res);
 692                 LBUG();
 693         }
 694
 695         if (!list_empty(&res->lr_waiting)) {
 696                 ldlm_resource_dump(D_ERROR, res);
 697                 LBUG();
 698         }
 699
 700         if (!list_empty(&res->lr_children)) {
 701                 ldlm_resource_dump(D_ERROR, res);
 702                 LBUG();
 703         }
 704
 705         ns->ns_refcount--;
 706         list_del_init(&res->lr_hash);
 707         list_del_init(&res->lr_childof);
 708
 709         ns->ns_resources--;
 710         if (ns->ns_resources == 0)
 711                 wake_up(&ns->ns_waitq);
 712 }
 713
 714 /* Returns 1 if the resource was freed, 0 if it remains. */
 715 int ldlm_resource_putref(struct ldlm_resource *res)
 716 {
 717         struct ldlm_namespace *ns = res->lr_namespace;
 718         int rc = 0;
 719         ENTRY;
 720
 721         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 722                atomic_read(&res->lr_refcount) - 1);
 723         LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
 724                  atomic_read(&res->lr_refcount));
 725         LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
 726                  atomic_read(&res->lr_refcount));
 727
 728         if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
 729                 __ldlm_resource_putref_final(res);
 730                 spin_unlock(&ns->ns_hash_lock);
 731                 if (res->lr_lvb_data)
 732                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 733                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 734                 rc = 1;
 735         }
 736
 737         RETURN(rc);
 738 }
 739
 740 /* Returns 1 if the resource was freed, 0 if it remains. */
 741 int ldlm_resource_putref_locked(struct ldlm_resource *res)
 742 {
 743         int rc = 0;
 744         ENTRY;
 745
 746         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 747                atomic_read(&res->lr_refcount) - 1);
 748         LASSERT(atomic_read(&res->lr_refcount) > 0);
 749         LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
 750
 751         LASSERT(atomic_read(&res->lr_refcount) >= 0);
 752         if (atomic_dec_and_test(&res->lr_refcount)) {
 753                 __ldlm_resource_putref_final(res);
 754                 if (res->lr_lvb_data)
 755                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
 756                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 757                 rc = 1;
 758         }
 759
 760         RETURN(rc);
 761 }
 762
 763 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 764                             struct ldlm_lock *lock)
 765 {
 766         check_res_locked(res);
 767
 768         ldlm_resource_dump(D_OTHER, res);
 769         CDEBUG(D_OTHER, "About to add this lock:\n");
 770         ldlm_lock_dump(D_OTHER, lock, 0);
 771
 772         if (lock->l_destroyed) {
 773                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 774                 return;
 775         }
 776
 777         LASSERT(list_empty(&lock->l_res_link));
 778
 779         list_add_tail(&lock->l_res_link, head);
 780 }
 781
 782 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
 783                                      struct ldlm_lock *new)
 784 {
 785         struct ldlm_resource *res = original->l_resource;
 786
 787         check_res_locked(res);
 788
 789         ldlm_resource_dump(D_OTHER, res);
 790         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
 791         ldlm_lock_dump(D_OTHER, new, 0);
 792
 793         if (new->l_destroyed) {
 794                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 795                 goto out;
 796         }
 797
 798         LASSERT(list_empty(&new->l_res_link));
 799
 800         list_add(&new->l_res_link, &original->l_res_link);
 801  out:;
 802 }
 803
 804 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 805 {
 806         check_res_locked(lock->l_resource);
 807         ldlm_unlink_lock_skiplist(lock);
 808         list_del_init(&lock->l_res_link);
 809 }
 810
 811 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 812 {
 813         desc->lr_type = res->lr_type;
 814         desc->lr_name = res->lr_name;
 815 }
 816
 817 void ldlm_dump_all_namespaces(int level)
 818 {
 819         struct list_head *tmp;
 820
 821         if (!((libcfs_debug | D_ERROR) & level))
 822                 return;
 823
 824         mutex_down(&ldlm_namespace_lock);
 825
 826         list_for_each(tmp, &ldlm_namespace_list) {
 827                 struct ldlm_namespace *ns;
 828                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 829                 ldlm_namespace_dump(level, ns);
 830         }
 831
 832         mutex_up(&ldlm_namespace_lock);
 833 }
 834
 835 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
 836 {
 837         struct list_head *tmp;
 838
 839         if (!((libcfs_debug | D_ERROR) & level))
 840                 return;
 841
 842         CDEBUG(level, "--- Namespace: %s (rc: %d, client: %d)\n",
 843                ns->ns_name, ns->ns_refcount, ns->ns_client);
 844
 845         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
 846                 return;
 847
 848         spin_lock(&ns->ns_hash_lock);
 849         tmp = ns->ns_root_list.next;
 850         while (tmp != &ns->ns_root_list) {
 851                 struct ldlm_resource *res;
 852                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
 853
 854                 ldlm_resource_getref(res);
 855                 spin_unlock(&ns->ns_hash_lock);
 856
 857                 lock_res(res);
 858                 ldlm_resource_dump(level, res);
 859                 unlock_res(res);
 860
 861                 spin_lock(&ns->ns_hash_lock);
 862                 tmp = tmp->next;
 863                 ldlm_resource_putref_locked(res);
 864         }
 865         ns->ns_next_dump = cfs_time_shift(10);
 866         spin_unlock(&ns->ns_hash_lock);
 867 }
 868
 869 void ldlm_resource_dump(int level, struct ldlm_resource *res)
 870 {
 871         struct list_head *tmp;
 872         int pos;
 873
 874         CLASSERT(RES_NAME_SIZE == 4);
 875
 876         if (!((libcfs_debug | D_ERROR) & level))
 877                 return;
 878
 879         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
 880                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
 881                res->lr_name.name[2], res->lr_name.name[3],
 882                atomic_read(&res->lr_refcount));
 883
 884         if (!list_empty(&res->lr_granted)) {
 885                 pos = 0;
 886                 CDEBUG(level, "Granted locks:\n");
 887                 list_for_each(tmp, &res->lr_granted) {
 888                         struct ldlm_lock *lock;
 889                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 890                         ldlm_lock_dump(level, lock, ++pos);
 891                 }
 892         }
 893         if (!list_empty(&res->lr_converting)) {
 894                 pos = 0;
 895                 CDEBUG(level, "Converting locks:\n");
 896                 list_for_each(tmp, &res->lr_converting) {
 897                         struct ldlm_lock *lock;
 898                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 899                         ldlm_lock_dump(level, lock, ++pos);
 900                 }
 901         }
 902         if (!list_empty(&res->lr_waiting)) {
 903                 pos = 0;
 904                 CDEBUG(level, "Waiting locks:\n");
 905                 list_for_each(tmp, &res->lr_waiting) {
 906                         struct ldlm_lock *lock;
 907                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 908                         ldlm_lock_dump(level, lock, ++pos);
 909                 }
 910         }
 911 }