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 Cluster File Systems, Inc.
   5  *
   6  * This code is issued under the GNU General Public License.
   7  * See the file COPYING in this distribution
   8  *
   9  * by Cluster File Systems, Inc.
  10  */
  11
  12 #define DEBUG_SUBSYSTEM S_LDLM
  13
  14 #include <linux/lustre_dlm.h>
  15 #include <linux/obd_class.h>
  16
  17 kmem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
  18
  19 spinlock_t ldlm_namespace_lock = SPIN_LOCK_UNLOCKED;
  20 struct list_head ldlm_namespace_list = LIST_HEAD_INIT(ldlm_namespace_list);
  21 static struct proc_dir_entry *ldlm_ns_proc_dir = NULL;
  22
  23 int ldlm_proc_setup(struct obd_device *obd)
  24 {
  25         ENTRY;
  26         LASSERT(ldlm_ns_proc_dir == NULL);
  27         ldlm_ns_proc_dir=obd->obd_type->typ_procroot;
  28         RETURN(0);
  29 }
  30
  31 void ldlm_proc_cleanup(struct obd_device *obd)
  32 {
  33         ldlm_ns_proc_dir = NULL;
  34 }
  35
  36 #define MAX_STRING_SIZE 100
  37 void ldlm_proc_namespace(struct ldlm_namespace *ns)
  38 {
  39         struct lprocfs_vars lock_vars[2];
  40         char lock_names[MAX_STRING_SIZE+1];
  41
  42         memset(lock_vars, 0, sizeof(lock_vars));
  43         snprintf(lock_names, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
  44         lock_names[MAX_STRING_SIZE] = '\0';
  45         lock_vars[0].name = lock_names;
  46         lock_vars[0].read_fptr = lprocfs_ll_rd;
  47         lock_vars[0].write_fptr = NULL;
  48         lock_vars[0].data = &ns->ns_resources;
  49         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
  50
  51         memset(lock_vars, 0, sizeof(lock_vars));
  52         snprintf(lock_names, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
  53         lock_names[MAX_STRING_SIZE] = '\0';
  54         lock_vars[0].name = lock_names;
  55         lock_vars[0].read_fptr = lprocfs_ll_rd;
  56         lock_vars[0].write_fptr = NULL;
  57         lock_vars[0].data = &ns->ns_locks;
  58         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
  59
  60 }
  61 #undef MAX_STRING_SIZE
  62
  63 #define LDLM_MAX_UNUSED 20
  64 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
  65 {
  66         struct ldlm_namespace *ns = NULL;
  67         struct list_head *bucket;
  68         ENTRY;
  69
  70         OBD_ALLOC(ns, sizeof(*ns));
  71         if (!ns)
  72                 RETURN(NULL);
  73
  74         ns->ns_hash = vmalloc(sizeof(*ns->ns_hash) * RES_HASH_SIZE);
  75         if (!ns->ns_hash)
  76                 GOTO(out_ns, NULL);
  77
  78         atomic_add(sizeof(*ns->ns_hash) * RES_HASH_SIZE, &obd_memory);
  79
  80         OBD_ALLOC(ns->ns_name, strlen(name) + 1);
  81         if (!ns->ns_name)
  82                 GOTO(out_hash, NULL);
  83
  84         strcpy(ns->ns_name, name);
  85
  86         INIT_LIST_HEAD(&ns->ns_root_list);
  87         l_lock_init(&ns->ns_lock);
  88         ns->ns_refcount = 0;
  89         ns->ns_client = client;
  90         spin_lock_init(&ns->ns_counter_lock);
  91         ns->ns_locks = 0;
  92         ns->ns_resources = 0;
  93
  94         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
  95              bucket--)
  96                 INIT_LIST_HEAD(bucket);
  97
  98         INIT_LIST_HEAD(&ns->ns_unused_list);
  99         ns->ns_nr_unused = 0;
 100         ns->ns_max_unused = LDLM_MAX_UNUSED;
 101
 102         spin_lock(&ldlm_namespace_lock);
 103         list_add(&ns->ns_list_chain, &ldlm_namespace_list);
 104         spin_unlock(&ldlm_namespace_lock);
 105         ldlm_proc_namespace(ns);
 106         RETURN(ns);
 107
 108 out_hash:
 109         memset(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 110         vfree(ns->ns_hash);
 111         atomic_sub(sizeof(*ns->ns_hash) * RES_HASH_SIZE, &obd_memory);
 112 out_ns:
 113         OBD_FREE(ns, sizeof(*ns));
 114         return NULL;
 115 }
 116
 117 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 118
 119 /* If 'local_only' is true, don't try to tell the server, just cleanup. */
 120 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 121                              int local_only)
 122 {
 123         struct list_head *tmp, *pos;
 124         int rc = 0, client = res->lr_namespace->ns_client;
 125         ENTRY;
 126
 127         list_for_each_safe(tmp, pos, q) {
 128                 struct ldlm_lock *lock;
 129                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 130                 LDLM_LOCK_GET(lock);
 131
 132                 /* At shutdown time, don't call the cancellation callback */
 133                 lock->l_flags |= LDLM_FL_CANCEL;
 134
 135                 if (client) {
 136                         struct lustre_handle lockh;
 137                         ldlm_lock2handle(lock, &lockh);
 138                         if (!local_only) {
 139                                 rc = ldlm_cli_cancel(&lockh);
 140                                 if (rc)
 141                                         CERROR("ldlm_cli_cancel: %d\n", rc);
 142                         }
 143                         /* Force local cleanup on errors, too. */
 144                         if (local_only || rc != ELDLM_OK)
 145                                 ldlm_lock_cancel(lock);
 146                 } else {
 147                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 148                                    "client node.\n");
 149
 150                         ldlm_resource_unlink_lock(lock);
 151                         ldlm_lock_destroy(lock);
 152                 }
 153                 LDLM_LOCK_PUT(lock);
 154         }
 155 }
 156
 157 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only)
 158 {
 159         int i;
 160
 161         l_lock(&ns->ns_lock);
 162         for (i = 0; i < RES_HASH_SIZE; i++) {
 163                 struct list_head *tmp, *pos;
 164                 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
 165                         struct ldlm_resource *res;
 166                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 167                         ldlm_resource_getref(res);
 168
 169                         cleanup_resource(res, &res->lr_granted, local_only);
 170                         cleanup_resource(res, &res->lr_converting, local_only);
 171                         cleanup_resource(res, &res->lr_waiting, local_only);
 172
 173                         /* XXX what a mess: don't force cleanup if we're
 174                          * local_only (which is only used by recovery).  In that
 175                          * case, we probably still have outstanding lock refs
 176                          * which reference these resources. -phil */
 177                         if (!ldlm_resource_putref(res) && !local_only) {
 178                                 CERROR("Resource refcount nonzero (%d) after "
 179                                        "lock cleanup; forcing cleanup.\n",
 180                                        atomic_read(&res->lr_refcount));
 181                                 ldlm_resource_dump(res);
 182                                 atomic_set(&res->lr_refcount, 1);
 183                                 ldlm_resource_putref(res);
 184                         }
 185                 }
 186         }
 187         l_unlock(&ns->ns_lock);
 188
 189         return ELDLM_OK;
 190 }
 191
 192 /* Cleanup, but also free, the namespace */
 193 int ldlm_namespace_free(struct ldlm_namespace *ns)
 194 {
 195         if (!ns)
 196                 RETURN(ELDLM_OK);
 197
 198         spin_lock(&ldlm_namespace_lock);
 199         list_del(&ns->ns_list_chain);
 200
 201         spin_unlock(&ldlm_namespace_lock);
 202
 203         ldlm_namespace_cleanup(ns, 0);
 204
 205         memset(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 206         vfree(ns->ns_hash /* , sizeof(*ns->ns_hash) * RES_HASH_SIZE */);
 207         atomic_sub(sizeof(*ns->ns_hash) * RES_HASH_SIZE, &obd_memory);
 208         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 209         OBD_FREE(ns, sizeof(*ns));
 210
 211         return ELDLM_OK;
 212 }
 213
 214 int ldlm_client_free(struct obd_export *exp)
 215 {
 216         struct ldlm_export_data *led = &exp->exp_ldlm_data;
 217         ptlrpc_cleanup_client(&led->led_import);
 218         RETURN(0);
 219 }
 220
 221 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, __u64 *name)
 222 {
 223         __u32 hash = 0;
 224         int i;
 225
 226         for (i = 0; i < RES_NAME_SIZE; i++)
 227                 hash += name[i];
 228
 229         hash += (__u32)((unsigned long)parent >> 4);
 230
 231         return (hash & RES_HASH_MASK);
 232 }
 233
 234 static struct ldlm_resource *ldlm_resource_new(void)
 235 {
 236         struct ldlm_resource *res;
 237
 238         res = kmem_cache_alloc(ldlm_resource_slab, SLAB_KERNEL);
 239         if (res == NULL) {
 240                 LBUG();
 241                 return NULL;
 242         }
 243         memset(res, 0, sizeof(*res));
 244
 245         INIT_LIST_HEAD(&res->lr_children);
 246         INIT_LIST_HEAD(&res->lr_childof);
 247         INIT_LIST_HEAD(&res->lr_granted);
 248         INIT_LIST_HEAD(&res->lr_converting);
 249         INIT_LIST_HEAD(&res->lr_waiting);
 250
 251         atomic_set(&res->lr_refcount, 1);
 252
 253         return res;
 254 }
 255
 256 /* Args: locked namespace
 257  * Returns: newly-allocated, referenced, unlocked resource */
 258 static struct ldlm_resource *ldlm_resource_add(struct ldlm_namespace *ns,
 259                                                struct ldlm_resource *parent,
 260                                                __u64 *name, __u32 type)
 261 {
 262         struct list_head *bucket;
 263         struct ldlm_resource *res;
 264         ENTRY;
 265
 266         if (type < LDLM_MIN_TYPE || type > LDLM_MAX_TYPE) {
 267                 LBUG();
 268                 RETURN(NULL);
 269         }
 270
 271         res = ldlm_resource_new();
 272         if (!res) {
 273                 LBUG();
 274                 RETURN(NULL);
 275         }
 276
 277         spin_lock(&ns->ns_counter_lock);
 278         ns->ns_resources++;
 279         spin_unlock(&ns->ns_counter_lock);
 280
 281         l_lock(&ns->ns_lock);
 282         memcpy(res->lr_name, name, sizeof(res->lr_name));
 283         res->lr_namespace = ns;
 284         ns->ns_refcount++;
 285
 286         res->lr_type = type;
 287         res->lr_most_restr = LCK_NL;
 288
 289         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 290         list_add(&res->lr_hash, bucket);
 291
 292         if (parent == NULL) {
 293                 list_add(&res->lr_childof, &ns->ns_root_list);
 294         } else {
 295                 res->lr_parent = parent;
 296                 list_add(&res->lr_childof, &parent->lr_children);
 297         }
 298         l_unlock(&ns->ns_lock);
 299
 300         RETURN(res);
 301 }
 302
 303 /* Args: unlocked namespace
 304  * Locks: takes and releases ns->ns_lock and res->lr_lock
 305  * Returns: referenced, unlocked ldlm_resource or NULL */
 306 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 307                                         struct ldlm_resource *parent,
 308                                         __u64 *name, __u32 type, int create)
 309 {
 310         struct list_head *bucket;
 311         struct list_head *tmp = bucket;
 312         struct ldlm_resource *res = NULL;
 313         ENTRY;
 314
 315         if (ns == NULL || ns->ns_hash == NULL) {
 316                 LBUG();
 317                 RETURN(NULL);
 318         }
 319
 320         l_lock(&ns->ns_lock);
 321         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 322
 323         list_for_each(tmp, bucket) {
 324                 struct ldlm_resource *chk;
 325                 chk = list_entry(tmp, struct ldlm_resource, lr_hash);
 326
 327                 if (memcmp(chk->lr_name, name, sizeof(chk->lr_name)) == 0) {
 328                         res = chk;
 329                         atomic_inc(&res->lr_refcount);
 330                         l_unlock(&ns->ns_lock);
 331                         RETURN(res);
 332                 }
 333         }
 334
 335         if (create)
 336                 res = ldlm_resource_add(ns, parent, name, type);
 337         l_unlock(&ns->ns_lock);
 338
 339         RETURN(res);
 340 }
 341
 342 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 343 {
 344         atomic_inc(&res->lr_refcount);
 345         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 346                atomic_read(&res->lr_refcount));
 347         return res;
 348 }
 349
 350 /* Returns 1 if the resource was freed, 0 if it remains. */
 351 int ldlm_resource_putref(struct ldlm_resource *res)
 352 {
 353         int rc = 0;
 354
 355         if (atomic_dec_and_test(&res->lr_refcount)) {
 356                 struct ldlm_namespace *ns = res->lr_namespace;
 357                 ENTRY;
 358                 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 359                        atomic_read(&res->lr_refcount));
 360
 361                 l_lock(&ns->ns_lock);
 362
 363                 if (atomic_read(&res->lr_refcount) != 0) {
 364                         /* We lost the race. */
 365                         l_unlock(&ns->ns_lock);
 366                         RETURN(rc);
 367                 }
 368
 369                 if (!list_empty(&res->lr_granted)) {
 370                         ldlm_resource_dump(res);
 371                         LBUG();
 372                 }
 373
 374                 if (!list_empty(&res->lr_converting)) {
 375                         ldlm_resource_dump(res);
 376                         LBUG();
 377                 }
 378
 379                 if (!list_empty(&res->lr_waiting)) {
 380                         ldlm_resource_dump(res);
 381                         LBUG();
 382                 }
 383
 384                 if (!list_empty(&res->lr_children)) {
 385                         ldlm_resource_dump(res);
 386                         LBUG();
 387                 }
 388
 389                 ns->ns_refcount--;
 390                 list_del(&res->lr_hash);
 391                 list_del(&res->lr_childof);
 392
 393                 memset(res, 0x5a, sizeof(*res));
 394                 kmem_cache_free(ldlm_resource_slab, res);
 395                 l_unlock(&ns->ns_lock);
 396
 397                 spin_lock(&ns->ns_counter_lock);
 398                 ns->ns_resources--;
 399                 spin_unlock(&ns->ns_counter_lock);
 400
 401                 rc = 1;
 402         } else {
 403                 ENTRY;
 404                 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 405                        atomic_read(&res->lr_refcount));
 406                 LASSERT(atomic_read(&res->lr_refcount) >= 0);
 407         }
 408
 409         RETURN(rc);
 410 }
 411
 412 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 413                             struct ldlm_lock *lock)
 414 {
 415         l_lock(&res->lr_namespace->ns_lock);
 416
 417         ldlm_resource_dump(res);
 418         ldlm_lock_dump(lock);
 419
 420         LASSERT(list_empty(&lock->l_res_link));
 421
 422         list_add(&lock->l_res_link, head);
 423         l_unlock(&res->lr_namespace->ns_lock);
 424 }
 425
 426 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 427 {
 428         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 429         list_del_init(&lock->l_res_link);
 430         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 431 }
 432
 433 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 434 {
 435         desc->lr_type = res->lr_type;
 436         memcpy(desc->lr_name, res->lr_name, sizeof(desc->lr_name));
 437         memcpy(desc->lr_version, res->lr_version, sizeof(desc->lr_version));
 438 }
 439
 440 void ldlm_dump_all_namespaces(void)
 441 {
 442         struct list_head *tmp;
 443
 444         spin_lock(&ldlm_namespace_lock);
 445
 446         list_for_each(tmp, &ldlm_namespace_list) {
 447                 struct ldlm_namespace *ns;
 448                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 449                 ldlm_namespace_dump(ns);
 450         }
 451
 452         spin_unlock(&ldlm_namespace_lock);
 453 }
 454
 455 void ldlm_namespace_dump(struct ldlm_namespace *ns)
 456 {
 457         struct list_head *tmp;
 458
 459         l_lock(&ns->ns_lock);
 460         CDEBUG(D_OTHER, "--- Namespace: %s (rc: %d, client: %d)\n", ns->ns_name,
 461                ns->ns_refcount, ns->ns_client);
 462
 463         list_for_each(tmp, &ns->ns_root_list) {
 464                 struct ldlm_resource *res;
 465                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
 466
 467                 /* Once we have resources with children, this should really dump
 468                  * them recursively. */
 469                 ldlm_resource_dump(res);
 470         }
 471         l_unlock(&ns->ns_lock);
 472 }
 473
 474 void ldlm_resource_dump(struct ldlm_resource *res)
 475 {
 476         struct list_head *tmp;
 477         char name[256];
 478
 479         if (RES_NAME_SIZE != 3)
 480                 LBUG();
 481
 482         snprintf(name, sizeof(name), "%Lx %Lx %Lx",
 483                  (unsigned long long)res->lr_name[0],
 484                  (unsigned long long)res->lr_name[1],
 485                  (unsigned long long)res->lr_name[2]);
 486
 487         CDEBUG(D_OTHER, "--- Resource: %p (%s) (rc: %d)\n", res, name,
 488                atomic_read(&res->lr_refcount));
 489         CDEBUG(D_OTHER, "Namespace: %p (%s)\n", res->lr_namespace,
 490                res->lr_namespace->ns_name);
 491         CDEBUG(D_OTHER, "Parent: %p, root: %p\n", res->lr_parent, res->lr_root);
 492
 493         CDEBUG(D_OTHER, "Granted locks:\n");
 494         list_for_each(tmp, &res->lr_granted) {
 495                 struct ldlm_lock *lock;
 496                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 497                 ldlm_lock_dump(lock);
 498         }
 499
 500         CDEBUG(D_OTHER, "Converting locks:\n");
 501         list_for_each(tmp, &res->lr_converting) {
 502                 struct ldlm_lock *lock;
 503                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 504                 ldlm_lock_dump(lock);
 505         }
 506
 507         CDEBUG(D_OTHER, "Waiting locks:\n");
 508         list_for_each(tmp, &res->lr_waiting) {
 509                 struct ldlm_lock *lock;
 510                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 511                 ldlm_lock_dump(lock);
 512         }
 513 }