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 static int lprocfs_uint_rd(char *page, char **start, off_t off,
  37                            int count, int *eof, void *data)
  38 {
  39         unsigned int *temp = (unsigned int *)data;
  40         int len;
  41         len = snprintf(page, count, "%u\n", *temp);
  42         return len;
  43 }
  44
  45 #define MAX_STRING_SIZE 100
  46 void ldlm_proc_namespace(struct ldlm_namespace *ns)
  47 {
  48         struct lprocfs_vars lock_vars[2];
  49         char lock_names[MAX_STRING_SIZE + 1];
  50
  51         memset(lock_vars, 0, sizeof(lock_vars));
  52         snprintf(lock_names, MAX_STRING_SIZE, "%s/resource_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_resources;
  58         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
  59
  60         memset(lock_vars, 0, sizeof(lock_vars));
  61         snprintf(lock_names, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
  62         lock_names[MAX_STRING_SIZE] = '\0';
  63         lock_vars[0].name = lock_names;
  64         lock_vars[0].read_fptr = lprocfs_ll_rd;
  65         lock_vars[0].write_fptr = NULL;
  66         lock_vars[0].data = &ns->ns_locks;
  67         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
  68
  69         memset(lock_vars, 0, sizeof(lock_vars));
  70         snprintf(lock_names, MAX_STRING_SIZE, "%s/lock_unused_count",
  71                  ns->ns_name);
  72         lock_names[MAX_STRING_SIZE] = '\0';
  73         lock_vars[0].name = lock_names;
  74         lock_vars[0].read_fptr = lprocfs_uint_rd;
  75         lock_vars[0].write_fptr = NULL;
  76         lock_vars[0].data = &ns->ns_nr_unused;
  77         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
  78 }
  79 #undef MAX_STRING_SIZE
  80
  81 #define LDLM_MAX_UNUSED 20
  82 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
  83 {
  84         struct ldlm_namespace *ns = NULL;
  85         struct list_head *bucket;
  86         ENTRY;
  87
  88         OBD_ALLOC(ns, sizeof(*ns));
  89         if (!ns)
  90                 RETURN(NULL);
  91
  92         ns->ns_hash = vmalloc(sizeof(*ns->ns_hash) * RES_HASH_SIZE);
  93         if (!ns->ns_hash)
  94                 GOTO(out_ns, NULL);
  95
  96         atomic_add(sizeof(*ns->ns_hash) * RES_HASH_SIZE, &obd_memory);
  97
  98         OBD_ALLOC(ns->ns_name, strlen(name) + 1);
  99         if (!ns->ns_name)
 100                 GOTO(out_hash, NULL);
 101
 102         strcpy(ns->ns_name, name);
 103
 104         INIT_LIST_HEAD(&ns->ns_root_list);
 105         l_lock_init(&ns->ns_lock);
 106         ns->ns_refcount = 0;
 107         ns->ns_client = client;
 108         spin_lock_init(&ns->ns_counter_lock);
 109         ns->ns_locks = 0;
 110         ns->ns_resources = 0;
 111
 112         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 113              bucket--)
 114                 INIT_LIST_HEAD(bucket);
 115
 116         INIT_LIST_HEAD(&ns->ns_unused_list);
 117         ns->ns_nr_unused = 0;
 118         ns->ns_max_unused = LDLM_MAX_UNUSED;
 119
 120         spin_lock(&ldlm_namespace_lock);
 121         list_add(&ns->ns_list_chain, &ldlm_namespace_list);
 122         spin_unlock(&ldlm_namespace_lock);
 123         ldlm_proc_namespace(ns);
 124         RETURN(ns);
 125
 126 out_hash:
 127         memset(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 128         vfree(ns->ns_hash);
 129         atomic_sub(sizeof(*ns->ns_hash) * RES_HASH_SIZE, &obd_memory);
 130 out_ns:
 131         OBD_FREE(ns, sizeof(*ns));
 132         return NULL;
 133 }
 134
 135 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 136
 137 /* If 'local_only' is true, don't try to tell the server, just cleanup. */
 138 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 139                              int local_only)
 140 {
 141         struct list_head *tmp, *pos;
 142         int rc = 0, client = res->lr_namespace->ns_client;
 143         ENTRY;
 144
 145         list_for_each_safe(tmp, pos, q) {
 146                 struct ldlm_lock *lock;
 147                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 148                 LDLM_LOCK_GET(lock);
 149
 150                 /* At shutdown time, don't call the cancellation callback */
 151                 lock->l_flags |= LDLM_FL_CANCEL;
 152
 153                 if (client) {
 154                         struct lustre_handle lockh;
 155                         ldlm_lock2handle(lock, &lockh);
 156                         if (!local_only) {
 157                                 rc = ldlm_cli_cancel(&lockh);
 158                                 if (rc)
 159                                         CERROR("ldlm_cli_cancel: %d\n", rc);
 160                         }
 161                         /* Force local cleanup on errors, too. */
 162                         if (local_only || rc != ELDLM_OK)
 163                                 ldlm_lock_cancel(lock);
 164                 } else {
 165                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 166                                    "client node.\n");
 167
 168                         ldlm_resource_unlink_lock(lock);
 169                         ldlm_lock_destroy(lock);
 170                 }
 171                 LDLM_LOCK_PUT(lock);
 172         }
 173 }
 174
 175 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only)
 176 {
 177         int i;
 178
 179         l_lock(&ns->ns_lock);
 180         for (i = 0; i < RES_HASH_SIZE; i++) {
 181                 struct list_head *tmp, *pos;
 182                 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
 183                         struct ldlm_resource *res;
 184                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 185                         ldlm_resource_getref(res);
 186
 187                         cleanup_resource(res, &res->lr_granted, local_only);
 188                         cleanup_resource(res, &res->lr_converting, local_only);
 189                         cleanup_resource(res, &res->lr_waiting, local_only);
 190
 191                         /* XXX what a mess: don't force cleanup if we're
 192                          * local_only (which is only used by recovery).  In that
 193                          * case, we probably still have outstanding lock refs
 194                          * which reference these resources. -phil */
 195                         if (!ldlm_resource_putref(res) && !local_only) {
 196                                 CERROR("Resource refcount nonzero (%d) after "
 197                                        "lock cleanup; forcing cleanup.\n",
 198                                        atomic_read(&res->lr_refcount));
 199                                 ldlm_resource_dump(res);
 200                                 atomic_set(&res->lr_refcount, 1);
 201                                 ldlm_resource_putref(res);
 202                         }
 203                 }
 204         }
 205         l_unlock(&ns->ns_lock);
 206
 207         return ELDLM_OK;
 208 }
 209
 210 /* Cleanup, but also free, the namespace */
 211 int ldlm_namespace_free(struct ldlm_namespace *ns)
 212 {
 213         if (!ns)
 214                 RETURN(ELDLM_OK);
 215
 216         spin_lock(&ldlm_namespace_lock);
 217         list_del(&ns->ns_list_chain);
 218
 219         spin_unlock(&ldlm_namespace_lock);
 220
 221         ldlm_namespace_cleanup(ns, 0);
 222
 223         memset(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 224         vfree(ns->ns_hash /* , sizeof(*ns->ns_hash) * RES_HASH_SIZE */);
 225         atomic_sub(sizeof(*ns->ns_hash) * RES_HASH_SIZE, &obd_memory);
 226         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 227         OBD_FREE(ns, sizeof(*ns));
 228
 229         return ELDLM_OK;
 230 }
 231
 232 int ldlm_client_free(struct obd_export *exp)
 233 {
 234         struct ldlm_export_data *led = &exp->exp_ldlm_data;
 235         ptlrpc_cleanup_client(&led->led_import);
 236         RETURN(0);
 237 }
 238
 239 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, __u64 *name)
 240 {
 241         __u32 hash = 0;
 242         int i;
 243
 244         for (i = 0; i < RES_NAME_SIZE; i++)
 245                 hash += name[i];
 246
 247         hash += (__u32)((unsigned long)parent >> 4);
 248
 249         return (hash & RES_HASH_MASK);
 250 }
 251
 252 static struct ldlm_resource *ldlm_resource_new(void)
 253 {
 254         struct ldlm_resource *res;
 255
 256         res = kmem_cache_alloc(ldlm_resource_slab, SLAB_KERNEL);
 257         if (res == NULL) {
 258                 LBUG();
 259                 return NULL;
 260         }
 261         memset(res, 0, sizeof(*res));
 262
 263         INIT_LIST_HEAD(&res->lr_children);
 264         INIT_LIST_HEAD(&res->lr_childof);
 265         INIT_LIST_HEAD(&res->lr_granted);
 266         INIT_LIST_HEAD(&res->lr_converting);
 267         INIT_LIST_HEAD(&res->lr_waiting);
 268
 269         atomic_set(&res->lr_refcount, 1);
 270
 271         return res;
 272 }
 273
 274 /* Args: locked namespace
 275  * Returns: newly-allocated, referenced, unlocked resource */
 276 static struct ldlm_resource *ldlm_resource_add(struct ldlm_namespace *ns,
 277                                                struct ldlm_resource *parent,
 278                                                __u64 *name, __u32 type)
 279 {
 280         struct list_head *bucket;
 281         struct ldlm_resource *res;
 282         ENTRY;
 283
 284         if (type < LDLM_MIN_TYPE || type > LDLM_MAX_TYPE) {
 285                 LBUG();
 286                 RETURN(NULL);
 287         }
 288
 289         res = ldlm_resource_new();
 290         if (!res) {
 291                 LBUG();
 292                 RETURN(NULL);
 293         }
 294
 295         spin_lock(&ns->ns_counter_lock);
 296         ns->ns_resources++;
 297         spin_unlock(&ns->ns_counter_lock);
 298
 299         l_lock(&ns->ns_lock);
 300         memcpy(res->lr_name, name, sizeof(res->lr_name));
 301         res->lr_namespace = ns;
 302         ns->ns_refcount++;
 303
 304         res->lr_type = type;
 305         res->lr_most_restr = LCK_NL;
 306
 307         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 308         list_add(&res->lr_hash, bucket);
 309
 310         if (parent == NULL) {
 311                 list_add(&res->lr_childof, &ns->ns_root_list);
 312         } else {
 313                 res->lr_parent = parent;
 314                 list_add(&res->lr_childof, &parent->lr_children);
 315         }
 316         l_unlock(&ns->ns_lock);
 317
 318         RETURN(res);
 319 }
 320
 321 /* Args: unlocked namespace
 322  * Locks: takes and releases ns->ns_lock and res->lr_lock
 323  * Returns: referenced, unlocked ldlm_resource or NULL */
 324 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 325                                         struct ldlm_resource *parent,
 326                                         __u64 *name, __u32 type, int create)
 327 {
 328         struct list_head *bucket, *tmp;
 329         struct ldlm_resource *res = NULL;
 330         ENTRY;
 331
 332         LASSERT(ns != NULL);
 333         LASSERT(ns->ns_hash != NULL);
 334
 335         l_lock(&ns->ns_lock);
 336         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 337
 338         list_for_each(tmp, bucket) {
 339                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 340
 341                 if (memcmp(res->lr_name, name, sizeof(res->lr_name)) == 0) {
 342                         ldlm_resource_getref(res);
 343                         l_unlock(&ns->ns_lock);
 344                         RETURN(res);
 345                 }
 346         }
 347
 348         if (create)
 349                 res = ldlm_resource_add(ns, parent, name, type);
 350         else
 351                 res = NULL;
 352
 353         l_unlock(&ns->ns_lock);
 354
 355         RETURN(res);
 356 }
 357
 358 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 359 {
 360         atomic_inc(&res->lr_refcount);
 361         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 362                atomic_read(&res->lr_refcount));
 363         return res;
 364 }
 365
 366 /* Returns 1 if the resource was freed, 0 if it remains. */
 367 int ldlm_resource_putref(struct ldlm_resource *res)
 368 {
 369         int rc = 0;
 370         ENTRY;
 371
 372         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 373                atomic_read(&res->lr_refcount) - 1);
 374         LASSERT(atomic_read(&res->lr_refcount) > 0);
 375         LASSERT(atomic_read(&res->lr_refcount) < 0x5a5a5a5a);
 376
 377         if (atomic_dec_and_test(&res->lr_refcount)) {
 378                 struct ldlm_namespace *ns = res->lr_namespace;
 379                 ENTRY;
 380
 381                 l_lock(&ns->ns_lock);
 382
 383                 if (atomic_read(&res->lr_refcount) != 0) {
 384                         /* We lost the race. */
 385                         l_unlock(&ns->ns_lock);
 386                         RETURN(rc);
 387                 }
 388
 389                 if (!list_empty(&res->lr_granted)) {
 390                         ldlm_resource_dump(res);
 391                         LBUG();
 392                 }
 393
 394                 if (!list_empty(&res->lr_converting)) {
 395                         ldlm_resource_dump(res);
 396                         LBUG();
 397                 }
 398
 399                 if (!list_empty(&res->lr_waiting)) {
 400                         ldlm_resource_dump(res);
 401                         LBUG();
 402                 }
 403
 404                 if (!list_empty(&res->lr_children)) {
 405                         ldlm_resource_dump(res);
 406                         LBUG();
 407                 }
 408
 409                 ns->ns_refcount--;
 410                 list_del_init(&res->lr_hash);
 411                 list_del_init(&res->lr_childof);
 412
 413                 memset(res, 0x5a, sizeof(*res));
 414                 kmem_cache_free(ldlm_resource_slab, res);
 415                 l_unlock(&ns->ns_lock);
 416
 417                 spin_lock(&ns->ns_counter_lock);
 418                 ns->ns_resources--;
 419                 spin_unlock(&ns->ns_counter_lock);
 420
 421                 rc = 1;
 422                 EXIT;
 423         }
 424
 425         RETURN(rc);
 426 }
 427
 428 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 429                             struct ldlm_lock *lock)
 430 {
 431         l_lock(&res->lr_namespace->ns_lock);
 432
 433         ldlm_resource_dump(res);
 434         ldlm_lock_dump(lock);
 435
 436         LASSERT(list_empty(&lock->l_res_link));
 437
 438         list_add(&lock->l_res_link, head);
 439         l_unlock(&res->lr_namespace->ns_lock);
 440 }
 441
 442 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 443 {
 444         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 445         list_del_init(&lock->l_res_link);
 446         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 447 }
 448
 449 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 450 {
 451         desc->lr_type = res->lr_type;
 452         memcpy(desc->lr_name, res->lr_name, sizeof(desc->lr_name));
 453         memcpy(desc->lr_version, res->lr_version, sizeof(desc->lr_version));
 454 }
 455
 456 void ldlm_dump_all_namespaces(void)
 457 {
 458         struct list_head *tmp;
 459
 460         spin_lock(&ldlm_namespace_lock);
 461
 462         list_for_each(tmp, &ldlm_namespace_list) {
 463                 struct ldlm_namespace *ns;
 464                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 465                 ldlm_namespace_dump(ns);
 466         }
 467
 468         spin_unlock(&ldlm_namespace_lock);
 469 }
 470
 471 void ldlm_namespace_dump(struct ldlm_namespace *ns)
 472 {
 473         struct list_head *tmp;
 474
 475         l_lock(&ns->ns_lock);
 476         CDEBUG(D_OTHER, "--- Namespace: %s (rc: %d, client: %d)\n", ns->ns_name,
 477                ns->ns_refcount, ns->ns_client);
 478
 479         list_for_each(tmp, &ns->ns_root_list) {
 480                 struct ldlm_resource *res;
 481                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
 482
 483                 /* Once we have resources with children, this should really dump
 484                  * them recursively. */
 485                 ldlm_resource_dump(res);
 486         }
 487         l_unlock(&ns->ns_lock);
 488 }
 489
 490 void ldlm_resource_dump(struct ldlm_resource *res)
 491 {
 492         struct list_head *tmp;
 493         char name[256];
 494
 495         if (RES_NAME_SIZE != 3)
 496                 LBUG();
 497
 498         snprintf(name, sizeof(name), "%Lx %Lx %Lx",
 499                  (unsigned long long)res->lr_name[0],
 500                  (unsigned long long)res->lr_name[1],
 501                  (unsigned long long)res->lr_name[2]);
 502
 503         CDEBUG(D_OTHER, "--- Resource: %p (%s) (rc: %d)\n", res, name,
 504                atomic_read(&res->lr_refcount));
 505         CDEBUG(D_OTHER, "Namespace: %p (%s)\n", res->lr_namespace,
 506                res->lr_namespace->ns_name);
 507         CDEBUG(D_OTHER, "Parent: %p, root: %p\n", res->lr_parent, res->lr_root);
 508
 509         CDEBUG(D_OTHER, "Granted locks:\n");
 510         list_for_each(tmp, &res->lr_granted) {
 511                 struct ldlm_lock *lock;
 512                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 513                 ldlm_lock_dump(lock);
 514         }
 515
 516         CDEBUG(D_OTHER, "Converting locks:\n");
 517         list_for_each(tmp, &res->lr_converting) {
 518                 struct ldlm_lock *lock;
 519                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 520                 ldlm_lock_dump(lock);
 521         }
 522
 523         CDEBUG(D_OTHER, "Waiting locks:\n");
 524         list_for_each(tmp, &res->lr_waiting) {
 525                 struct ldlm_lock *lock;
 526                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 527                 ldlm_lock_dump(lock);
 528         }
 529 }