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 #ifndef LPROC_SNMP
  45 static int lprocfs_ll_rd(char *page, char **start, off_t off,
  46                          int count, int *eof, void *data)
  47 {
  48         int len;
  49         __u64 *temp = (__u64 *)data;
  50
  51         len = snprintf(page, count, "%Lu\n", *temp);
  52
  53         return len;
  54 }
  55 #endif
  56
  57 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
  58 {
  59         struct ldlm_namespace *ns = NULL;
  60         struct list_head *bucket;
  61         struct proc_dir_entry *proc_entry;
  62
  63         OBD_ALLOC(ns, sizeof(*ns));
  64         if (!ns) {
  65                 LBUG();
  66                 GOTO(out, NULL);
  67         }
  68
  69         ns->ns_hash = vmalloc(sizeof(*ns->ns_hash) * RES_HASH_SIZE);
  70         if (!ns->ns_hash) {
  71                 LBUG();
  72                 GOTO(out, ns);
  73         }
  74         obd_memory += sizeof(*ns->ns_hash) * RES_HASH_SIZE;
  75
  76         OBD_ALLOC(ns->ns_name, strlen(name) + 1);
  77         if (!ns->ns_name) {
  78                 LBUG();
  79                 GOTO(out, ns);
  80         }
  81         strcpy(ns->ns_name, name);
  82
  83         INIT_LIST_HEAD(&ns->ns_root_list);
  84         l_lock_init(&ns->ns_lock);
  85         ns->ns_refcount = 0;
  86         ns->ns_client = client;
  87         spin_lock_init(&ns->ns_counter_lock);
  88         ns->ns_locks = 0;
  89         ns->ns_resources = 0;
  90
  91         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
  92              bucket--)
  93                 INIT_LIST_HEAD(bucket);
  94
  95         spin_lock(&ldlm_namespace_lock);
  96         list_add(&ns->ns_list_chain, &ldlm_namespace_list);
  97         spin_unlock(&ldlm_namespace_lock);
  98
  99         ns->ns_proc_dir = proc_mkdir(ns->ns_name, ldlm_ns_proc_dir);
 100         if (ns->ns_proc_dir == NULL)
 101                 CERROR("Unable to create proc directory for namespace.\n");
 102         proc_entry = create_proc_entry("resource_count", 0444, ns->ns_proc_dir);
 103         proc_entry->read_proc = lprocfs_ll_rd;
 104         proc_entry->data = &ns->ns_resources;
 105         proc_entry = create_proc_entry("lock_count", 0444, ns->ns_proc_dir);
 106         proc_entry->read_proc = lprocfs_ll_rd;
 107         proc_entry->data = &ns->ns_locks;
 108
 109         RETURN(ns);
 110
 111  out:
 112         if (ns && ns->ns_hash) {
 113                 vfree(ns->ns_hash);
 114                 obd_memory -= sizeof(*ns->ns_hash) * RES_HASH_SIZE;
 115         }
 116         if (ns && ns->ns_name)
 117                 OBD_FREE(ns->ns_name, strlen(name) + 1);
 118         if (ns)
 119                 OBD_FREE(ns, sizeof(*ns));
 120         return NULL;
 121 }
 122
 123 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 124
 125 /* If 'local_only' is true, don't try to tell the server, just cleanup. */
 126 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 127                              int local_only)
 128 {
 129         struct list_head *tmp, *pos;
 130         int rc = 0, client = res->lr_namespace->ns_client;
 131         ENTRY;
 132
 133         list_for_each_safe(tmp, pos, q) {
 134                 struct ldlm_lock *lock;
 135                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 136                 LDLM_LOCK_GET(lock);
 137
 138                 /* At shutdown time, don't call the cancellation callback */
 139                 lock->l_flags |= LDLM_FL_CANCEL;
 140
 141                 if (client) {
 142                         struct lustre_handle lockh;
 143                         ldlm_lock2handle(lock, &lockh);
 144                         if (!local_only) {
 145                                 rc = ldlm_cli_cancel(&lockh);
 146                                 if (rc)
 147                                         CERROR("ldlm_cli_cancel: %d\n", rc);
 148                         }
 149                         /* Force local cleanup on errors, too. */
 150                         if (local_only || rc != ELDLM_OK)
 151                                 ldlm_lock_cancel(lock);
 152                 } else {
 153                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 154                                    "client node.\n");
 155
 156                         ldlm_resource_unlink_lock(lock);
 157                         ldlm_lock_destroy(lock);
 158                 }
 159                 LDLM_LOCK_PUT(lock);
 160         }
 161 }
 162
 163 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only)
 164 {
 165         int i;
 166
 167         l_lock(&ns->ns_lock);
 168         for (i = 0; i < RES_HASH_SIZE; i++) {
 169                 struct list_head *tmp, *pos;
 170                 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
 171                         struct ldlm_resource *res;
 172                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 173                         ldlm_resource_getref(res);
 174
 175                         cleanup_resource(res, &res->lr_granted, local_only);
 176                         cleanup_resource(res, &res->lr_converting, local_only);
 177                         cleanup_resource(res, &res->lr_waiting, local_only);
 178
 179                         /* XXX this is a bit counter-intuitive and should
 180                          * probably be cleaner: don't force cleanup if we're
 181                          * local_only (which is only used by recovery).  We
 182                          * probably still have outstanding lock refs which
 183                          * reference these resources. -phil */
 184                         if (!ldlm_resource_put(res) && !local_only) {
 185                                 CERROR("Resource refcount nonzero (%d) after "
 186                                        "lock cleanup; forcing cleanup.\n",
 187                                        atomic_read(&res->lr_refcount));
 188                                 ldlm_resource_dump(res);
 189                                 atomic_set(&res->lr_refcount, 1);
 190                                 ldlm_resource_put(res);
 191                         }
 192                 }
 193         }
 194         l_unlock(&ns->ns_lock);
 195
 196         return ELDLM_OK;
 197 }
 198
 199 /* Cleanup, but also free, the namespace */
 200 int ldlm_namespace_free(struct ldlm_namespace *ns)
 201 {
 202         if (!ns)
 203                 RETURN(ELDLM_OK);
 204
 205         spin_lock(&ldlm_namespace_lock);
 206         list_del(&ns->ns_list_chain);
 207         remove_proc_entry("resource_count", ns->ns_proc_dir);
 208         remove_proc_entry("lock_count", ns->ns_proc_dir);
 209         remove_proc_entry(ns->ns_name, ldlm_ns_proc_dir);
 210         spin_unlock(&ldlm_namespace_lock);
 211
 212         ldlm_namespace_cleanup(ns, 0);
 213
 214         vfree(ns->ns_hash /* , sizeof(*ns->ns_hash) * RES_HASH_SIZE */);
 215         obd_memory -= sizeof(*ns->ns_hash) * RES_HASH_SIZE;
 216         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 217         OBD_FREE(ns, sizeof(*ns));
 218
 219         return ELDLM_OK;
 220 }
 221
 222 int ldlm_client_free(struct obd_export *exp)
 223 {
 224         struct ldlm_export_data *led = &exp->exp_ldlm_data;
 225         ptlrpc_cleanup_client(&led->led_import);
 226         RETURN(0);
 227 }
 228
 229 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, __u64 *name)
 230 {
 231         __u32 hash = 0;
 232         int i;
 233
 234         for (i = 0; i < RES_NAME_SIZE; i++)
 235                 hash += name[i];
 236
 237         hash += (__u32)((unsigned long)parent >> 4);
 238
 239         return (hash & RES_HASH_MASK);
 240 }
 241
 242 static struct ldlm_resource *ldlm_resource_new(void)
 243 {
 244         struct ldlm_resource *res;
 245
 246         res = kmem_cache_alloc(ldlm_resource_slab, SLAB_KERNEL);
 247         if (res == NULL) {
 248                 LBUG();
 249                 return NULL;
 250         }
 251         memset(res, 0, sizeof(*res));
 252
 253         INIT_LIST_HEAD(&res->lr_children);
 254         INIT_LIST_HEAD(&res->lr_childof);
 255         INIT_LIST_HEAD(&res->lr_granted);
 256         INIT_LIST_HEAD(&res->lr_converting);
 257         INIT_LIST_HEAD(&res->lr_waiting);
 258
 259         atomic_set(&res->lr_refcount, 1);
 260
 261         return res;
 262 }
 263
 264 /* Args: locked namespace
 265  * Returns: newly-allocated, referenced, unlocked resource */
 266 static struct ldlm_resource *ldlm_resource_add(struct ldlm_namespace *ns,
 267                                                struct ldlm_resource *parent,
 268                                                __u64 *name, __u32 type)
 269 {
 270         struct list_head *bucket;
 271         struct ldlm_resource *res;
 272         ENTRY;
 273
 274         if (type < LDLM_MIN_TYPE || type > LDLM_MAX_TYPE) {
 275                 LBUG();
 276                 RETURN(NULL);
 277         }
 278
 279         res = ldlm_resource_new();
 280         if (!res) {
 281                 LBUG();
 282                 RETURN(NULL);
 283         }
 284
 285         spin_lock(&ns->ns_counter_lock);
 286         ns->ns_resources++;
 287         spin_unlock(&ns->ns_counter_lock);
 288
 289         memcpy(res->lr_name, name, sizeof(res->lr_name));
 290         res->lr_namespace = ns;
 291         ns->ns_refcount++;
 292
 293         res->lr_type = type;
 294         res->lr_most_restr = LCK_NL;
 295
 296         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 297         list_add(&res->lr_hash, bucket);
 298
 299         if (parent == NULL)
 300                 list_add(&res->lr_childof, &ns->ns_root_list);
 301         else {
 302                 res->lr_parent = parent;
 303                 list_add(&res->lr_childof, &parent->lr_children);
 304         }
 305
 306         RETURN(res);
 307 }
 308
 309 /* Args: unlocked namespace
 310  * Locks: takes and releases ns->ns_lock and res->lr_lock
 311  * Returns: referenced, unlocked ldlm_resource or NULL */
 312 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 313                                         struct ldlm_resource *parent,
 314                                         __u64 *name, __u32 type, int create)
 315 {
 316         struct list_head *bucket;
 317         struct list_head *tmp = bucket;
 318         struct ldlm_resource *res = NULL;
 319         ENTRY;
 320
 321         if (ns == NULL || ns->ns_hash == NULL) {
 322                 LBUG();
 323                 RETURN(NULL);
 324         }
 325
 326         l_lock(&ns->ns_lock);
 327         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 328
 329         list_for_each(tmp, bucket) {
 330                 struct ldlm_resource *chk;
 331                 chk = list_entry(tmp, struct ldlm_resource, lr_hash);
 332
 333                 if (memcmp(chk->lr_name, name, sizeof(chk->lr_name)) == 0) {
 334                         res = chk;
 335                         atomic_inc(&res->lr_refcount);
 336                         EXIT;
 337                         break;
 338                 }
 339         }
 340
 341         if (res == NULL && create)
 342                 res = ldlm_resource_add(ns, parent, name, type);
 343         l_unlock(&ns->ns_lock);
 344
 345         RETURN(res);
 346 }
 347
 348 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 349 {
 350         atomic_inc(&res->lr_refcount);
 351         return res;
 352 }
 353
 354 /* Returns 1 if the resource was freed, 0 if it remains. */
 355 int ldlm_resource_put(struct ldlm_resource *res)
 356 {
 357         int rc = 0;
 358
 359         if (atomic_dec_and_test(&res->lr_refcount)) {
 360                 struct ldlm_namespace *ns = res->lr_namespace;
 361                 ENTRY;
 362
 363                 l_lock(&ns->ns_lock);
 364
 365                 if (atomic_read(&res->lr_refcount) != 0) {
 366                         /* We lost the race. */
 367                         l_unlock(&ns->ns_lock);
 368                         goto out;
 369                 }
 370
 371                 if (!list_empty(&res->lr_granted))
 372                         LBUG();
 373
 374                 if (!list_empty(&res->lr_converting))
 375                         LBUG();
 376
 377                 if (!list_empty(&res->lr_waiting))
 378                         LBUG();
 379
 380                 if (!list_empty(&res->lr_children))
 381                         LBUG();
 382
 383                 ns->ns_refcount--;
 384                 list_del(&res->lr_hash);
 385                 list_del(&res->lr_childof);
 386
 387                 kmem_cache_free(ldlm_resource_slab, res);
 388                 l_unlock(&ns->ns_lock);
 389
 390                 spin_lock(&ns->ns_counter_lock);
 391                 ns->ns_resources--;
 392                 spin_unlock(&ns->ns_counter_lock);
 393
 394                 rc = 1;
 395         } else {
 396                 ENTRY;
 397         out:
 398                 if (atomic_read(&res->lr_refcount) < 0)
 399                         LBUG();
 400         }
 401
 402         RETURN(rc);
 403 }
 404
 405 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 406                             struct ldlm_lock *lock)
 407 {
 408         l_lock(&res->lr_namespace->ns_lock);
 409
 410         ldlm_resource_dump(res);
 411         ldlm_lock_dump(lock);
 412
 413         if (!list_empty(&lock->l_res_link))
 414                 LBUG();
 415
 416         list_add(&lock->l_res_link, head);
 417         l_unlock(&res->lr_namespace->ns_lock);
 418 }
 419
 420 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 421 {
 422         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 423         list_del_init(&lock->l_res_link);
 424         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 425 }
 426
 427 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 428 {
 429         desc->lr_type = res->lr_type;
 430         memcpy(desc->lr_name, res->lr_name, sizeof(desc->lr_name));
 431         memcpy(desc->lr_version, res->lr_version, sizeof(desc->lr_version));
 432 }
 433
 434 void ldlm_dump_all_namespaces(void)
 435 {
 436         struct list_head *tmp;
 437
 438         spin_lock(&ldlm_namespace_lock);
 439
 440         list_for_each(tmp, &ldlm_namespace_list) {
 441                 struct ldlm_namespace *ns;
 442                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 443                 ldlm_namespace_dump(ns);
 444         }
 445
 446         spin_unlock(&ldlm_namespace_lock);
 447 }
 448
 449 void ldlm_namespace_dump(struct ldlm_namespace *ns)
 450 {
 451         struct list_head *tmp;
 452
 453         l_lock(&ns->ns_lock);
 454         CDEBUG(D_OTHER, "--- Namespace: %s (rc: %d, client: %d)\n", ns->ns_name,
 455                ns->ns_refcount, ns->ns_client);
 456
 457         list_for_each(tmp, &ns->ns_root_list) {
 458                 struct ldlm_resource *res;
 459                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
 460
 461                 /* Once we have resources with children, this should really dump
 462                  * them recursively. */
 463                 ldlm_resource_dump(res);
 464         }
 465         l_unlock(&ns->ns_lock);
 466 }
 467
 468 void ldlm_resource_dump(struct ldlm_resource *res)
 469 {
 470         struct list_head *tmp;
 471         char name[256];
 472
 473         if (RES_NAME_SIZE != 3)
 474                 LBUG();
 475
 476         snprintf(name, sizeof(name), "%Lx %Lx %Lx",
 477                  (unsigned long long)res->lr_name[0],
 478                  (unsigned long long)res->lr_name[1],
 479                  (unsigned long long)res->lr_name[2]);
 480
 481         CDEBUG(D_OTHER, "--- Resource: %p (%s) (rc: %d)\n", res, name,
 482                atomic_read(&res->lr_refcount));
 483         CDEBUG(D_OTHER, "Namespace: %p (%s)\n", res->lr_namespace,
 484                res->lr_namespace->ns_name);
 485         CDEBUG(D_OTHER, "Parent: %p, root: %p\n", res->lr_parent, res->lr_root);
 486
 487         CDEBUG(D_OTHER, "Granted locks:\n");
 488         list_for_each(tmp, &res->lr_granted) {
 489                 struct ldlm_lock *lock;
 490                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 491                 ldlm_lock_dump(lock);
 492         }
 493
 494         CDEBUG(D_OTHER, "Converting locks:\n");
 495         list_for_each(tmp, &res->lr_converting) {
 496                 struct ldlm_lock *lock;
 497                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 498                 ldlm_lock_dump(lock);
 499         }
 500
 501         CDEBUG(D_OTHER, "Waiting locks:\n");
 502         list_for_each(tmp, &res->lr_waiting) {
 503                 struct ldlm_lock *lock;
 504                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 505                 ldlm_lock_dump(lock);
 506         }
 507 }