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 extern struct proc_dir_entry proc_root;
  23
  24 int ldlm_proc_setup(struct obd_device *obd)
  25 {
  26         ENTRY;
  27
  28         LASSERT(ldlm_ns_proc_dir == NULL);
  29
  30         ldlm_ns_proc_dir = proc_mkdir("ldlm", &proc_root);
  31         if (ldlm_ns_proc_dir == NULL) {
  32                 CERROR("Couldn't create /proc/ldlm\n");
  33                 RETURN(-EPERM);
  34         }
  35         RETURN(0);
  36 }
  37
  38 void ldlm_proc_cleanup(struct obd_device *obd)
  39 {
  40         remove_proc_entry("ldlm", &proc_root);
  41 }
  42
  43 /* FIXME: This can go away when we start to really use lprocfs */
  44 static int ldlm_proc_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 #define LDLM_MAX_UNUSED 20
  56 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
  57 {
  58         struct ldlm_namespace *ns = NULL;
  59         struct list_head *bucket;
  60         struct proc_dir_entry *proc_entry;
  61
  62         OBD_ALLOC(ns, sizeof(*ns));
  63         if (!ns) {
  64                 LBUG();
  65                 GOTO(out, NULL);
  66         }
  67
  68         ns->ns_hash = vmalloc(sizeof(*ns->ns_hash) * RES_HASH_SIZE);
  69         if (!ns->ns_hash) {
  70                 LBUG();
  71                 GOTO(out, ns);
  72         }
  73         obd_memory += sizeof(*ns->ns_hash) * RES_HASH_SIZE;
  74
  75         OBD_ALLOC(ns->ns_name, strlen(name) + 1);
  76         if (!ns->ns_name) {
  77                 LBUG();
  78                 GOTO(out, ns);
  79         }
  80         strcpy(ns->ns_name, name);
  81
  82         INIT_LIST_HEAD(&ns->ns_root_list);
  83         l_lock_init(&ns->ns_lock);
  84         ns->ns_refcount = 0;
  85         ns->ns_client = client;
  86         spin_lock_init(&ns->ns_counter_lock);
  87         ns->ns_locks = 0;
  88         ns->ns_resources = 0;
  89
  90         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
  91              bucket--)
  92                 INIT_LIST_HEAD(bucket);
  93
  94         INIT_LIST_HEAD(&ns->ns_unused_list);
  95         ns->ns_nr_unused = 0;
  96         ns->ns_max_unused = LDLM_MAX_UNUSED;
  97
  98         spin_lock(&ldlm_namespace_lock);
  99         list_add(&ns->ns_list_chain, &ldlm_namespace_list);
 100         spin_unlock(&ldlm_namespace_lock);
 101
 102         ns->ns_proc_dir = proc_mkdir(ns->ns_name, ldlm_ns_proc_dir);
 103         if (ns->ns_proc_dir == NULL)
 104                 CERROR("Unable to create proc directory for namespace.\n");
 105         proc_entry = create_proc_entry("resource_count", 0444, ns->ns_proc_dir);
 106         proc_entry->read_proc = ldlm_proc_ll_rd;
 107         proc_entry->data = &ns->ns_resources;
 108         proc_entry = create_proc_entry("lock_count", 0444, ns->ns_proc_dir);
 109         proc_entry->read_proc = ldlm_proc_ll_rd;
 110         proc_entry->data = &ns->ns_locks;
 111
 112         RETURN(ns);
 113
 114  out:
 115         if (ns && ns->ns_hash) {
 116                 vfree(ns->ns_hash);
 117                 obd_memory -= sizeof(*ns->ns_hash) * RES_HASH_SIZE;
 118         }
 119         if (ns && ns->ns_name)
 120                 OBD_FREE(ns->ns_name, strlen(name) + 1);
 121         if (ns)
 122                 OBD_FREE(ns, sizeof(*ns));
 123         return NULL;
 124 }
 125
 126 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 127
 128 /* If 'local_only' is true, don't try to tell the server, just cleanup. */
 129 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 130                              int local_only)
 131 {
 132         struct list_head *tmp, *pos;
 133         int rc = 0, client = res->lr_namespace->ns_client;
 134         ENTRY;
 135
 136         list_for_each_safe(tmp, pos, q) {
 137                 struct ldlm_lock *lock;
 138                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 139                 LDLM_LOCK_GET(lock);
 140
 141                 /* At shutdown time, don't call the cancellation callback */
 142                 lock->l_flags |= LDLM_FL_CANCEL;
 143
 144                 if (client) {
 145                         struct lustre_handle lockh;
 146                         ldlm_lock2handle(lock, &lockh);
 147                         if (!local_only) {
 148                                 rc = ldlm_cli_cancel(&lockh);
 149                                 if (rc)
 150                                         CERROR("ldlm_cli_cancel: %d\n", rc);
 151                         }
 152                         /* Force local cleanup on errors, too. */
 153                         if (local_only || rc != ELDLM_OK)
 154                                 ldlm_lock_cancel(lock);
 155                 } else {
 156                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 157                                    "client node.\n");
 158
 159                         ldlm_resource_unlink_lock(lock);
 160                         ldlm_lock_destroy(lock);
 161                 }
 162                 LDLM_LOCK_PUT(lock);
 163         }
 164 }
 165
 166 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only)
 167 {
 168         int i;
 169
 170         l_lock(&ns->ns_lock);
 171         for (i = 0; i < RES_HASH_SIZE; i++) {
 172                 struct list_head *tmp, *pos;
 173                 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
 174                         struct ldlm_resource *res;
 175                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 176                         ldlm_resource_getref(res);
 177
 178                         cleanup_resource(res, &res->lr_granted, local_only);
 179                         cleanup_resource(res, &res->lr_converting, local_only);
 180                         cleanup_resource(res, &res->lr_waiting, local_only);
 181
 182                         /* XXX this is a bit counter-intuitive and should
 183                          * probably be cleaner: don't force cleanup if we're
 184                          * local_only (which is only used by recovery).  We
 185                          * probably still have outstanding lock refs which
 186                          * reference these resources. -phil */
 187                         if (!ldlm_resource_put(res) && !local_only) {
 188                                 CERROR("Resource refcount nonzero (%d) after "
 189                                        "lock cleanup; forcing cleanup.\n",
 190                                        atomic_read(&res->lr_refcount));
 191                                 ldlm_resource_dump(res);
 192                                 atomic_set(&res->lr_refcount, 1);
 193                                 ldlm_resource_put(res);
 194                         }
 195                 }
 196         }
 197         l_unlock(&ns->ns_lock);
 198
 199         return ELDLM_OK;
 200 }
 201
 202 /* Cleanup, but also free, the namespace */
 203 int ldlm_namespace_free(struct ldlm_namespace *ns)
 204 {
 205         if (!ns)
 206                 RETURN(ELDLM_OK);
 207
 208         spin_lock(&ldlm_namespace_lock);
 209         list_del(&ns->ns_list_chain);
 210         remove_proc_entry("resource_count", ns->ns_proc_dir);
 211         remove_proc_entry("lock_count", ns->ns_proc_dir);
 212         remove_proc_entry(ns->ns_name, ldlm_ns_proc_dir);
 213         spin_unlock(&ldlm_namespace_lock);
 214
 215         ldlm_namespace_cleanup(ns, 0);
 216
 217         vfree(ns->ns_hash /* , sizeof(*ns->ns_hash) * RES_HASH_SIZE */);
 218         obd_memory -= sizeof(*ns->ns_hash) * RES_HASH_SIZE;
 219         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 220         OBD_FREE(ns, sizeof(*ns));
 221
 222         return ELDLM_OK;
 223 }
 224
 225 int ldlm_client_free(struct obd_export *exp)
 226 {
 227         struct ldlm_export_data *led = &exp->exp_ldlm_data;
 228         ptlrpc_cleanup_client(&led->led_import);
 229         RETURN(0);
 230 }
 231
 232 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, __u64 *name)
 233 {
 234         __u32 hash = 0;
 235         int i;
 236
 237         for (i = 0; i < RES_NAME_SIZE; i++)
 238                 hash += name[i];
 239
 240         hash += (__u32)((unsigned long)parent >> 4);
 241
 242         return (hash & RES_HASH_MASK);
 243 }
 244
 245 static struct ldlm_resource *ldlm_resource_new(void)
 246 {
 247         struct ldlm_resource *res;
 248
 249         res = kmem_cache_alloc(ldlm_resource_slab, SLAB_KERNEL);
 250         if (res == NULL) {
 251                 LBUG();
 252                 return NULL;
 253         }
 254         memset(res, 0, sizeof(*res));
 255
 256         INIT_LIST_HEAD(&res->lr_children);
 257         INIT_LIST_HEAD(&res->lr_childof);
 258         INIT_LIST_HEAD(&res->lr_granted);
 259         INIT_LIST_HEAD(&res->lr_converting);
 260         INIT_LIST_HEAD(&res->lr_waiting);
 261
 262         atomic_set(&res->lr_refcount, 1);
 263
 264         return res;
 265 }
 266
 267 /* Args: locked namespace
 268  * Returns: newly-allocated, referenced, unlocked resource */
 269 static struct ldlm_resource *ldlm_resource_add(struct ldlm_namespace *ns,
 270                                                struct ldlm_resource *parent,
 271                                                __u64 *name, __u32 type)
 272 {
 273         struct list_head *bucket;
 274         struct ldlm_resource *res;
 275         ENTRY;
 276
 277         if (type < LDLM_MIN_TYPE || type > LDLM_MAX_TYPE) {
 278                 LBUG();
 279                 RETURN(NULL);
 280         }
 281
 282         res = ldlm_resource_new();
 283         if (!res) {
 284                 LBUG();
 285                 RETURN(NULL);
 286         }
 287
 288         spin_lock(&ns->ns_counter_lock);
 289         ns->ns_resources++;
 290         spin_unlock(&ns->ns_counter_lock);
 291
 292         memcpy(res->lr_name, name, sizeof(res->lr_name));
 293         res->lr_namespace = ns;
 294         ns->ns_refcount++;
 295
 296         res->lr_type = type;
 297         res->lr_most_restr = LCK_NL;
 298
 299         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 300         list_add(&res->lr_hash, bucket);
 301
 302         if (parent == NULL)
 303                 list_add(&res->lr_childof, &ns->ns_root_list);
 304         else {
 305                 res->lr_parent = parent;
 306                 list_add(&res->lr_childof, &parent->lr_children);
 307         }
 308
 309         RETURN(res);
 310 }
 311
 312 /* Args: unlocked namespace
 313  * Locks: takes and releases ns->ns_lock and res->lr_lock
 314  * Returns: referenced, unlocked ldlm_resource or NULL */
 315 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 316                                         struct ldlm_resource *parent,
 317                                         __u64 *name, __u32 type, int create)
 318 {
 319         struct list_head *bucket;
 320         struct list_head *tmp = bucket;
 321         struct ldlm_resource *res = NULL;
 322         ENTRY;
 323
 324         if (ns == NULL || ns->ns_hash == NULL) {
 325                 LBUG();
 326                 RETURN(NULL);
 327         }
 328
 329         l_lock(&ns->ns_lock);
 330         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 331
 332         list_for_each(tmp, bucket) {
 333                 struct ldlm_resource *chk;
 334                 chk = list_entry(tmp, struct ldlm_resource, lr_hash);
 335
 336                 if (memcmp(chk->lr_name, name, sizeof(chk->lr_name)) == 0) {
 337                         res = chk;
 338                         atomic_inc(&res->lr_refcount);
 339                         EXIT;
 340                         break;
 341                 }
 342         }
 343
 344         if (res == NULL && create)
 345                 res = ldlm_resource_add(ns, parent, name, type);
 346         l_unlock(&ns->ns_lock);
 347
 348         RETURN(res);
 349 }
 350
 351 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 352 {
 353         atomic_inc(&res->lr_refcount);
 354         return res;
 355 }
 356
 357 /* Returns 1 if the resource was freed, 0 if it remains. */
 358 int ldlm_resource_put(struct ldlm_resource *res)
 359 {
 360         int rc = 0;
 361
 362         if (atomic_dec_and_test(&res->lr_refcount)) {
 363                 struct ldlm_namespace *ns = res->lr_namespace;
 364                 ENTRY;
 365
 366                 l_lock(&ns->ns_lock);
 367
 368                 if (atomic_read(&res->lr_refcount) != 0) {
 369                         /* We lost the race. */
 370                         l_unlock(&ns->ns_lock);
 371                         goto out;
 372                 }
 373
 374                 if (!list_empty(&res->lr_granted))
 375                         LBUG();
 376
 377                 if (!list_empty(&res->lr_converting))
 378                         LBUG();
 379
 380                 if (!list_empty(&res->lr_waiting))
 381                         LBUG();
 382
 383                 if (!list_empty(&res->lr_children))
 384                         LBUG();
 385
 386                 ns->ns_refcount--;
 387                 list_del(&res->lr_hash);
 388                 list_del(&res->lr_childof);
 389
 390                 kmem_cache_free(ldlm_resource_slab, res);
 391                 l_unlock(&ns->ns_lock);
 392
 393                 spin_lock(&ns->ns_counter_lock);
 394                 ns->ns_resources--;
 395                 spin_unlock(&ns->ns_counter_lock);
 396
 397                 rc = 1;
 398         } else {
 399                 ENTRY;
 400         out:
 401                 if (atomic_read(&res->lr_refcount) < 0)
 402                         LBUG();
 403         }
 404
 405         RETURN(rc);
 406 }
 407
 408 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 409                             struct ldlm_lock *lock)
 410 {
 411         l_lock(&res->lr_namespace->ns_lock);
 412
 413         ldlm_resource_dump(res);
 414         ldlm_lock_dump(lock);
 415
 416         if (!list_empty(&lock->l_res_link))
 417                 LBUG();
 418
 419         list_add(&lock->l_res_link, head);
 420         l_unlock(&res->lr_namespace->ns_lock);
 421 }
 422
 423 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 424 {
 425         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 426         list_del_init(&lock->l_res_link);
 427         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 428 }
 429
 430 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 431 {
 432         desc->lr_type = res->lr_type;
 433         memcpy(desc->lr_name, res->lr_name, sizeof(desc->lr_name));
 434         memcpy(desc->lr_version, res->lr_version, sizeof(desc->lr_version));
 435 }
 436
 437 void ldlm_dump_all_namespaces(void)
 438 {
 439         struct list_head *tmp;
 440
 441         spin_lock(&ldlm_namespace_lock);
 442
 443         list_for_each(tmp, &ldlm_namespace_list) {
 444                 struct ldlm_namespace *ns;
 445                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 446                 ldlm_namespace_dump(ns);
 447         }
 448
 449         spin_unlock(&ldlm_namespace_lock);
 450 }
 451
 452 void ldlm_namespace_dump(struct ldlm_namespace *ns)
 453 {
 454         struct list_head *tmp;
 455
 456         l_lock(&ns->ns_lock);
 457         CDEBUG(D_OTHER, "--- Namespace: %s (rc: %d, client: %d)\n", ns->ns_name,
 458                ns->ns_refcount, ns->ns_client);
 459
 460         list_for_each(tmp, &ns->ns_root_list) {
 461                 struct ldlm_resource *res;
 462                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
 463
 464                 /* Once we have resources with children, this should really dump
 465                  * them recursively. */
 466                 ldlm_resource_dump(res);
 467         }
 468         l_unlock(&ns->ns_lock);
 469 }
 470
 471 void ldlm_resource_dump(struct ldlm_resource *res)
 472 {
 473         struct list_head *tmp;
 474         char name[256];
 475
 476         if (RES_NAME_SIZE != 3)
 477                 LBUG();
 478
 479         snprintf(name, sizeof(name), "%Lx %Lx %Lx",
 480                  (unsigned long long)res->lr_name[0],
 481                  (unsigned long long)res->lr_name[1],
 482                  (unsigned long long)res->lr_name[2]);
 483
 484         CDEBUG(D_OTHER, "--- Resource: %p (%s) (rc: %d)\n", res, name,
 485                atomic_read(&res->lr_refcount));
 486         CDEBUG(D_OTHER, "Namespace: %p (%s)\n", res->lr_namespace,
 487                res->lr_namespace->ns_name);
 488         CDEBUG(D_OTHER, "Parent: %p, root: %p\n", res->lr_parent, res->lr_root);
 489
 490         CDEBUG(D_OTHER, "Granted locks:\n");
 491         list_for_each(tmp, &res->lr_granted) {
 492                 struct ldlm_lock *lock;
 493                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 494                 ldlm_lock_dump(lock);
 495         }
 496
 497         CDEBUG(D_OTHER, "Converting locks:\n");
 498         list_for_each(tmp, &res->lr_converting) {
 499                 struct ldlm_lock *lock;
 500                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 501                 ldlm_lock_dump(lock);
 502         }
 503
 504         CDEBUG(D_OTHER, "Waiting locks:\n");
 505         list_for_each(tmp, &res->lr_waiting) {
 506                 struct ldlm_lock *lock;
 507                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 508                 ldlm_lock_dump(lock);
 509         }
 510 }