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
  27         if (obd->obd_proc_entry == NULL)
  28                 RETURN(-EINVAL);
  29
  30         ldlm_ns_proc_dir = proc_mkdir("namespaces", obd->obd_proc_entry);
  31         if (ldlm_ns_proc_dir == NULL) {
  32                 CERROR("Couldn't create /proc/lustre/ldlm/namespaces\n");
  33                 RETURN(-EPERM);
  34         }
  35         RETURN(0);
  36 }
  37
  38 void ldlm_proc_cleanup(struct obd_device *obd)
  39 {
  40         proc_lustre_remove_obd_entry("namespaces", obd);
  41 }
  42
  43 /* FIXME: This can go away when we start to really use lprocfs */
  44 static int lprocfs_ll_rd(char *page, char **start, off_t off,
  45                          int count, int *eof, void *data)
  46 {
  47         int len;
  48         __u64 *temp = (__u64 *)data;
  49
  50         len = snprintf(page, count, "%Lu\n", *temp);
  51
  52         return len;
  53 }
  54
  55 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
  56 {
  57         struct ldlm_namespace *ns = NULL;
  58         struct list_head *bucket;
  59         struct proc_dir_entry *proc_entry;
  60
  61         OBD_ALLOC(ns, sizeof(*ns));
  62         if (!ns) {
  63                 LBUG();
  64                 GOTO(out, NULL);
  65         }
  66
  67         ns->ns_hash = vmalloc(sizeof(*ns->ns_hash) * RES_HASH_SIZE);
  68         if (!ns->ns_hash) {
  69                 LBUG();
  70                 GOTO(out, ns);
  71         }
  72         obd_memory += sizeof(*ns->ns_hash) * RES_HASH_SIZE;
  73
  74         OBD_ALLOC(ns->ns_name, strlen(name) + 1);
  75         if (!ns->ns_name) {
  76                 LBUG();
  77                 GOTO(out, ns);
  78         }
  79         strcpy(ns->ns_name, name);
  80
  81         INIT_LIST_HEAD(&ns->ns_root_list);
  82         l_lock_init(&ns->ns_lock);
  83         ns->ns_refcount = 0;
  84         ns->ns_client = client;
  85         spin_lock_init(&ns->ns_counter_lock);
  86         ns->ns_locks = 0;
  87         ns->ns_resources = 0;
  88
  89         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
  90              bucket--)
  91                 INIT_LIST_HEAD(bucket);
  92
  93         spin_lock(&ldlm_namespace_lock);
  94         list_add(&ns->ns_list_chain, &ldlm_namespace_list);
  95         spin_unlock(&ldlm_namespace_lock);
  96
  97         ns->ns_proc_dir = proc_mkdir(ns->ns_name, ldlm_ns_proc_dir);
  98         if (ns->ns_proc_dir == NULL)
  99                 CERROR("Unable to create proc directory for namespace.\n");
 100         proc_entry = create_proc_entry("resource_count", 0444, ns->ns_proc_dir);
 101         proc_entry->read_proc = lprocfs_ll_rd;
 102         proc_entry->data = &ns->ns_resources;
 103         proc_entry = create_proc_entry("lock_count", 0444, ns->ns_proc_dir);
 104         proc_entry->read_proc = lprocfs_ll_rd;
 105         proc_entry->data = &ns->ns_locks;
 106
 107         RETURN(ns);
 108
 109  out:
 110         if (ns && ns->ns_hash) {
 111                 vfree(ns->ns_hash);
 112                 obd_memory -= sizeof(*ns->ns_hash) * RES_HASH_SIZE;
 113         }
 114         if (ns && ns->ns_name)
 115                 OBD_FREE(ns->ns_name, strlen(name) + 1);
 116         if (ns)
 117                 OBD_FREE(ns, sizeof(*ns));
 118         return NULL;
 119 }
 120
 121 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 122
 123 /* If 'local_only' is true, don't try to tell the server, just cleanup. */
 124 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 125                              int local_only)
 126 {
 127         struct list_head *tmp, *pos;
 128         int rc = 0, client = res->lr_namespace->ns_client;
 129         ENTRY;
 130
 131         list_for_each_safe(tmp, pos, q) {
 132                 struct ldlm_lock *lock;
 133                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 134                 LDLM_LOCK_GET(lock);
 135
 136                 if (client) {
 137                         struct lustre_handle lockh;
 138                         ldlm_lock2handle(lock, &lockh);
 139                         if (!local_only) {
 140                                 rc = ldlm_cli_cancel(&lockh);
 141                                 if (rc)
 142                                         CERROR("ldlm_cli_cancel: %d\n", rc);
 143                         }
 144                         /* Force local cleanup on errors, too. */
 145                         if (local_only || rc != ELDLM_OK)
 146                                 ldlm_lock_cancel(lock);
 147                 } else {
 148                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 149                                    "client node.\n");
 150
 151                         ldlm_resource_unlink_lock(lock);
 152                         ldlm_lock_destroy(lock);
 153                 }
 154                 LDLM_LOCK_PUT(lock);
 155         }
 156 }
 157
 158 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only)
 159 {
 160         int i;
 161
 162         l_lock(&ns->ns_lock);
 163         for (i = 0; i < RES_HASH_SIZE; i++) {
 164                 struct list_head *tmp, *pos;
 165                 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
 166                         struct ldlm_resource *res;
 167                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 168                         ldlm_resource_getref(res);
 169
 170                         cleanup_resource(res, &res->lr_granted, local_only);
 171                         cleanup_resource(res, &res->lr_converting, local_only);
 172                         cleanup_resource(res, &res->lr_waiting, local_only);
 173
 174                         if (!ldlm_resource_put(res) && !local_only) {
 175                                 CERROR("Resource refcount nonzero (%d) after "
 176                                        "lock cleanup; forcing cleanup.\n",
 177                                        atomic_read(&res->lr_refcount));
 178                                 ldlm_resource_dump(res);
 179                                 atomic_set(&res->lr_refcount, 1);
 180                                 ldlm_resource_put(res);
 181                         }
 182                 }
 183         }
 184         l_unlock(&ns->ns_lock);
 185
 186         return ELDLM_OK;
 187 }
 188
 189 /* Cleanup, but also free, the namespace */
 190 int ldlm_namespace_free(struct ldlm_namespace *ns)
 191 {
 192         if (!ns)
 193                 RETURN(ELDLM_OK);
 194
 195         spin_lock(&ldlm_namespace_lock);
 196         list_del(&ns->ns_list_chain);
 197         remove_proc_entry("resource_count", ns->ns_proc_dir);
 198         remove_proc_entry("lock_count", ns->ns_proc_dir);
 199         remove_proc_entry(ns->ns_name, ldlm_ns_proc_dir);
 200         spin_unlock(&ldlm_namespace_lock);
 201
 202         ldlm_namespace_cleanup(ns, 0);
 203
 204         vfree(ns->ns_hash /* , sizeof(*ns->ns_hash) * RES_HASH_SIZE */);
 205         obd_memory -= sizeof(*ns->ns_hash) * RES_HASH_SIZE;
 206         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 207         OBD_FREE(ns, sizeof(*ns));
 208
 209         return ELDLM_OK;
 210 }
 211
 212 int ldlm_client_free(struct obd_export *exp)
 213 {
 214         struct ldlm_export_data *led = &exp->exp_ldlm_data;
 215         ptlrpc_cleanup_client(&led->led_import);
 216         RETURN(0);
 217 }
 218
 219 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, __u64 *name)
 220 {
 221         __u32 hash = 0;
 222         int i;
 223
 224         for (i = 0; i < RES_NAME_SIZE; i++)
 225                 hash += name[i];
 226
 227         hash += (__u32)((unsigned long)parent >> 4);
 228
 229         return (hash & RES_HASH_MASK);
 230 }
 231
 232 static struct ldlm_resource *ldlm_resource_new(void)
 233 {
 234         struct ldlm_resource *res;
 235
 236         res = kmem_cache_alloc(ldlm_resource_slab, SLAB_KERNEL);
 237         if (res == NULL) {
 238                 LBUG();
 239                 return NULL;
 240         }
 241         memset(res, 0, sizeof(*res));
 242
 243         INIT_LIST_HEAD(&res->lr_children);
 244         INIT_LIST_HEAD(&res->lr_childof);
 245         INIT_LIST_HEAD(&res->lr_granted);
 246         INIT_LIST_HEAD(&res->lr_converting);
 247         INIT_LIST_HEAD(&res->lr_waiting);
 248
 249         atomic_set(&res->lr_refcount, 1);
 250
 251         return res;
 252 }
 253
 254 /* Args: locked namespace
 255  * Returns: newly-allocated, referenced, unlocked resource */
 256 static struct ldlm_resource *ldlm_resource_add(struct ldlm_namespace *ns,
 257                                                struct ldlm_resource *parent,
 258                                                __u64 *name, __u32 type)
 259 {
 260         struct list_head *bucket;
 261         struct ldlm_resource *res;
 262         ENTRY;
 263
 264         if (type < LDLM_MIN_TYPE || type > LDLM_MAX_TYPE) {
 265                 LBUG();
 266                 RETURN(NULL);
 267         }
 268
 269         res = ldlm_resource_new();
 270         if (!res) {
 271                 LBUG();
 272                 RETURN(NULL);
 273         }
 274
 275         spin_lock(&ns->ns_counter_lock);
 276         ns->ns_resources++;
 277         spin_unlock(&ns->ns_counter_lock);
 278
 279         memcpy(res->lr_name, name, sizeof(res->lr_name));
 280         res->lr_namespace = ns;
 281         ns->ns_refcount++;
 282
 283         res->lr_type = type;
 284         res->lr_most_restr = LCK_NL;
 285
 286         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 287         list_add(&res->lr_hash, bucket);
 288
 289         if (parent == NULL)
 290                 list_add(&res->lr_childof, &ns->ns_root_list);
 291         else {
 292                 res->lr_parent = parent;
 293                 list_add(&res->lr_childof, &parent->lr_children);
 294         }
 295
 296         RETURN(res);
 297 }
 298
 299 /* Args: unlocked namespace
 300  * Locks: takes and releases ns->ns_lock and res->lr_lock
 301  * Returns: referenced, unlocked ldlm_resource or NULL */
 302 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 303                                         struct ldlm_resource *parent,
 304                                         __u64 *name, __u32 type, int create)
 305 {
 306         struct list_head *bucket;
 307         struct list_head *tmp = bucket;
 308         struct ldlm_resource *res = NULL;
 309         ENTRY;
 310
 311         if (ns == NULL || ns->ns_hash == NULL) {
 312                 LBUG();
 313                 RETURN(NULL);
 314         }
 315
 316         l_lock(&ns->ns_lock);
 317         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 318
 319         list_for_each(tmp, bucket) {
 320                 struct ldlm_resource *chk;
 321                 chk = list_entry(tmp, struct ldlm_resource, lr_hash);
 322
 323                 if (memcmp(chk->lr_name, name, sizeof(chk->lr_name)) == 0) {
 324                         res = chk;
 325                         atomic_inc(&res->lr_refcount);
 326                         EXIT;
 327                         break;
 328                 }
 329         }
 330
 331         if (res == NULL && create)
 332                 res = ldlm_resource_add(ns, parent, name, type);
 333         l_unlock(&ns->ns_lock);
 334
 335         RETURN(res);
 336 }
 337
 338 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 339 {
 340         atomic_inc(&res->lr_refcount);
 341         return res;
 342 }
 343
 344 /* Returns 1 if the resource was freed, 0 if it remains. */
 345 int ldlm_resource_put(struct ldlm_resource *res)
 346 {
 347         int rc = 0;
 348
 349         if (atomic_dec_and_test(&res->lr_refcount)) {
 350                 struct ldlm_namespace *ns = res->lr_namespace;
 351                 ENTRY;
 352
 353                 l_lock(&ns->ns_lock);
 354
 355                 if (atomic_read(&res->lr_refcount) != 0) {
 356                         /* We lost the race. */
 357                         l_unlock(&ns->ns_lock);
 358                         goto out;
 359                 }
 360
 361                 if (!list_empty(&res->lr_granted))
 362                         LBUG();
 363
 364                 if (!list_empty(&res->lr_converting))
 365                         LBUG();
 366
 367                 if (!list_empty(&res->lr_waiting))
 368                         LBUG();
 369
 370                 if (!list_empty(&res->lr_children))
 371                         LBUG();
 372
 373                 ns->ns_refcount--;
 374                 list_del(&res->lr_hash);
 375                 list_del(&res->lr_childof);
 376
 377                 kmem_cache_free(ldlm_resource_slab, res);
 378                 l_unlock(&ns->ns_lock);
 379
 380                 spin_lock(&ns->ns_counter_lock);
 381                 ns->ns_resources--;
 382                 spin_unlock(&ns->ns_counter_lock);
 383
 384                 rc = 1;
 385         } else {
 386                 ENTRY;
 387         out:
 388                 if (atomic_read(&res->lr_refcount) < 0)
 389                         LBUG();
 390         }
 391
 392         RETURN(rc);
 393 }
 394
 395 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 396                             struct ldlm_lock *lock)
 397 {
 398         l_lock(&res->lr_namespace->ns_lock);
 399
 400         ldlm_resource_dump(res);
 401         ldlm_lock_dump(lock);
 402
 403         if (!list_empty(&lock->l_res_link))
 404                 LBUG();
 405
 406         list_add(&lock->l_res_link, head);
 407         l_unlock(&res->lr_namespace->ns_lock);
 408 }
 409
 410 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 411 {
 412         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 413         list_del_init(&lock->l_res_link);
 414         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 415 }
 416
 417 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 418 {
 419         desc->lr_type = res->lr_type;
 420         memcpy(desc->lr_name, res->lr_name, sizeof(desc->lr_name));
 421         memcpy(desc->lr_version, res->lr_version, sizeof(desc->lr_version));
 422 }
 423
 424 void ldlm_dump_all_namespaces(void)
 425 {
 426         struct list_head *tmp;
 427
 428         spin_lock(&ldlm_namespace_lock);
 429
 430         list_for_each(tmp, &ldlm_namespace_list) {
 431                 struct ldlm_namespace *ns;
 432                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 433                 ldlm_namespace_dump(ns);
 434         }
 435
 436         spin_unlock(&ldlm_namespace_lock);
 437 }
 438
 439 void ldlm_namespace_dump(struct ldlm_namespace *ns)
 440 {
 441         struct list_head *tmp;
 442
 443         l_lock(&ns->ns_lock);
 444         CDEBUG(D_OTHER, "--- Namespace: %s (rc: %d, client: %d)\n", ns->ns_name,
 445                ns->ns_refcount, ns->ns_client);
 446
 447         list_for_each(tmp, &ns->ns_root_list) {
 448                 struct ldlm_resource *res;
 449                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
 450
 451                 /* Once we have resources with children, this should really dump
 452                  * them recursively. */
 453                 ldlm_resource_dump(res);
 454         }
 455         l_unlock(&ns->ns_lock);
 456 }
 457
 458 void ldlm_resource_dump(struct ldlm_resource *res)
 459 {
 460         struct list_head *tmp;
 461         char name[256];
 462
 463         if (RES_NAME_SIZE != 3)
 464                 LBUG();
 465
 466         snprintf(name, sizeof(name), "%Lx %Lx %Lx",
 467                  (unsigned long long)res->lr_name[0],
 468                  (unsigned long long)res->lr_name[1],
 469                  (unsigned long long)res->lr_name[2]);
 470
 471         CDEBUG(D_OTHER, "--- Resource: %p (%s) (rc: %d)\n", res, name,
 472                atomic_read(&res->lr_refcount));
 473         CDEBUG(D_OTHER, "Namespace: %p (%s)\n", res->lr_namespace,
 474                res->lr_namespace->ns_name);
 475         CDEBUG(D_OTHER, "Parent: %p, root: %p\n", res->lr_parent, res->lr_root);
 476
 477         CDEBUG(D_OTHER, "Granted locks:\n");
 478         list_for_each(tmp, &res->lr_granted) {
 479                 struct ldlm_lock *lock;
 480                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 481                 ldlm_lock_dump(lock);
 482         }
 483
 484         CDEBUG(D_OTHER, "Converting locks:\n");
 485         list_for_each(tmp, &res->lr_converting) {
 486                 struct ldlm_lock *lock;
 487                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 488                 ldlm_lock_dump(lock);
 489         }
 490
 491         CDEBUG(D_OTHER, "Waiting locks:\n");
 492         list_for_each(tmp, &res->lr_waiting) {
 493                 struct ldlm_lock *lock;
 494                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 495                 ldlm_lock_dump(lock);
 496         }
 497 }