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                 LBUG();
  73                 GOTO(out, NULL);
  74         }
  75
  76         ns->ns_hash = vmalloc(sizeof(*ns->ns_hash) * RES_HASH_SIZE);
  77         if (!ns->ns_hash) {
  78                 LBUG();
  79                 GOTO(out, ns);
  80         }
  81         obd_memory += sizeof(*ns->ns_hash) * RES_HASH_SIZE;
  82
  83         OBD_ALLOC(ns->ns_name, strlen(name) + 1);
  84         if (!ns->ns_name) {
  85                 LBUG();
  86                 GOTO(out, ns);
  87         }
  88         strcpy(ns->ns_name, name);
  89
  90         INIT_LIST_HEAD(&ns->ns_root_list);
  91         l_lock_init(&ns->ns_lock);
  92         ns->ns_refcount = 0;
  93         ns->ns_client = client;
  94         spin_lock_init(&ns->ns_counter_lock);
  95         ns->ns_locks = 0;
  96         ns->ns_resources = 0;
  97
  98         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
  99              bucket--)
 100                 INIT_LIST_HEAD(bucket);
 101
 102         INIT_LIST_HEAD(&ns->ns_unused_list);
 103         ns->ns_nr_unused = 0;
 104         ns->ns_max_unused = LDLM_MAX_UNUSED;
 105
 106         spin_lock(&ldlm_namespace_lock);
 107         list_add(&ns->ns_list_chain, &ldlm_namespace_list);
 108         spin_unlock(&ldlm_namespace_lock);
 109         ldlm_proc_namespace(ns);
 110         RETURN(ns);
 111
 112  out:
 113         if (ns && ns->ns_hash) {
 114                 memset(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 115                 vfree(ns->ns_hash);
 116                 obd_memory -= sizeof(*ns->ns_hash) * RES_HASH_SIZE;
 117         }
 118         if (ns && ns->ns_name)
 119                 OBD_FREE(ns->ns_name, strlen(name) + 1);
 120         if (ns)
 121                 OBD_FREE(ns, sizeof(*ns));
 122         return NULL;
 123 }
 124
 125 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 126
 127 /* If 'local_only' is true, don't try to tell the server, just cleanup. */
 128 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 129                              int local_only)
 130 {
 131         struct list_head *tmp, *pos;
 132         int rc = 0, client = res->lr_namespace->ns_client;
 133         ENTRY;
 134
 135         list_for_each_safe(tmp, pos, q) {
 136                 struct ldlm_lock *lock;
 137                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 138                 LDLM_LOCK_GET(lock);
 139
 140                 /* At shutdown time, don't call the cancellation callback */
 141                 lock->l_flags |= LDLM_FL_CANCEL;
 142
 143                 if (client) {
 144                         struct lustre_handle lockh;
 145                         ldlm_lock2handle(lock, &lockh);
 146                         if (!local_only) {
 147                                 rc = ldlm_cli_cancel(&lockh);
 148                                 if (rc)
 149                                         CERROR("ldlm_cli_cancel: %d\n", rc);
 150                         }
 151                         /* Force local cleanup on errors, too. */
 152                         if (local_only || rc != ELDLM_OK)
 153                                 ldlm_lock_cancel(lock);
 154                 } else {
 155                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 156                                    "client node.\n");
 157
 158                         ldlm_resource_unlink_lock(lock);
 159                         ldlm_lock_destroy(lock);
 160                 }
 161                 LDLM_LOCK_PUT(lock);
 162         }
 163 }
 164
 165 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only)
 166 {
 167         int i;
 168
 169         l_lock(&ns->ns_lock);
 170         for (i = 0; i < RES_HASH_SIZE; i++) {
 171                 struct list_head *tmp, *pos;
 172                 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
 173                         struct ldlm_resource *res;
 174                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 175                         ldlm_resource_getref(res);
 176
 177                         cleanup_resource(res, &res->lr_granted, local_only);
 178                         cleanup_resource(res, &res->lr_converting, local_only);
 179                         cleanup_resource(res, &res->lr_waiting, local_only);
 180
 181                         /* XXX what a mess: don't force cleanup if we're
 182                          * local_only (which is only used by recovery).  In that
 183                          * case, we probably still have outstanding lock refs
 184                          * which reference these resources. -phil */
 185                         if (!ldlm_resource_putref(res) && !local_only) {
 186                                 CERROR("Resource refcount nonzero (%d) after "
 187                                        "lock cleanup; forcing cleanup.\n",
 188                                        atomic_read(&res->lr_refcount));
 189                                 ldlm_resource_dump(res);
 190                                 atomic_set(&res->lr_refcount, 1);
 191                                 ldlm_resource_putref(res);
 192                         }
 193                 }
 194         }
 195         l_unlock(&ns->ns_lock);
 196
 197         return ELDLM_OK;
 198 }
 199
 200 /* Cleanup, but also free, the namespace */
 201 int ldlm_namespace_free(struct ldlm_namespace *ns)
 202 {
 203         if (!ns)
 204                 RETURN(ELDLM_OK);
 205
 206         spin_lock(&ldlm_namespace_lock);
 207         list_del(&ns->ns_list_chain);
 208
 209         spin_unlock(&ldlm_namespace_lock);
 210
 211         ldlm_namespace_cleanup(ns, 0);
 212
 213         memset(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 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         l_lock(&ns->ns_lock);
 290         memcpy(res->lr_name, name, sizeof(res->lr_name));
 291         res->lr_namespace = ns;
 292         ns->ns_refcount++;
 293
 294         res->lr_type = type;
 295         res->lr_most_restr = LCK_NL;
 296
 297         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 298         list_add(&res->lr_hash, bucket);
 299
 300         if (parent == NULL) {
 301                 list_add(&res->lr_childof, &ns->ns_root_list);
 302         } else {
 303                 res->lr_parent = parent;
 304                 list_add(&res->lr_childof, &parent->lr_children);
 305         }
 306         l_unlock(&ns->ns_lock);
 307
 308         RETURN(res);
 309 }
 310
 311 /* Args: unlocked namespace
 312  * Locks: takes and releases ns->ns_lock and res->lr_lock
 313  * Returns: referenced, unlocked ldlm_resource or NULL */
 314 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 315                                         struct ldlm_resource *parent,
 316                                         __u64 *name, __u32 type, int create)
 317 {
 318         struct list_head *bucket;
 319         struct list_head *tmp = bucket;
 320         struct ldlm_resource *res = NULL;
 321         ENTRY;
 322
 323         if (ns == NULL || ns->ns_hash == NULL) {
 324                 LBUG();
 325                 RETURN(NULL);
 326         }
 327
 328         l_lock(&ns->ns_lock);
 329         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 330
 331         list_for_each(tmp, bucket) {
 332                 struct ldlm_resource *chk;
 333                 chk = list_entry(tmp, struct ldlm_resource, lr_hash);
 334
 335                 if (memcmp(chk->lr_name, name, sizeof(chk->lr_name)) == 0) {
 336                         res = chk;
 337                         atomic_inc(&res->lr_refcount);
 338                         l_unlock(&ns->ns_lock);
 339                         RETURN(res);
 340                 }
 341         }
 342
 343         if (create)
 344                 res = ldlm_resource_add(ns, parent, name, type);
 345         l_unlock(&ns->ns_lock);
 346
 347         RETURN(res);
 348 }
 349
 350 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 351 {
 352         atomic_inc(&res->lr_refcount);
 353         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 354                atomic_read(&res->lr_refcount));
 355         return res;
 356 }
 357
 358 /* Returns 1 if the resource was freed, 0 if it remains. */
 359 int ldlm_resource_putref(struct ldlm_resource *res)
 360 {
 361         int rc = 0;
 362
 363         if (atomic_dec_and_test(&res->lr_refcount)) {
 364                 struct ldlm_namespace *ns = res->lr_namespace;
 365                 ENTRY;
 366                 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 367                        atomic_read(&res->lr_refcount));
 368
 369                 l_lock(&ns->ns_lock);
 370
 371                 if (atomic_read(&res->lr_refcount) != 0) {
 372                         /* We lost the race. */
 373                         l_unlock(&ns->ns_lock);
 374                         RETURN(rc);
 375                 }
 376
 377                 if (!list_empty(&res->lr_granted)) {
 378                         ldlm_resource_dump(res);
 379                         LBUG();
 380                 }
 381
 382                 if (!list_empty(&res->lr_converting)) {
 383                         ldlm_resource_dump(res);
 384                         LBUG();
 385                 }
 386
 387                 if (!list_empty(&res->lr_waiting)) {
 388                         ldlm_resource_dump(res);
 389                         LBUG();
 390                 }
 391
 392                 if (!list_empty(&res->lr_children)) {
 393                         ldlm_resource_dump(res);
 394                         LBUG();
 395                 }
 396
 397                 ns->ns_refcount--;
 398                 list_del(&res->lr_hash);
 399                 list_del(&res->lr_childof);
 400
 401                 memset(res, 0x5a, sizeof(*res));
 402                 kmem_cache_free(ldlm_resource_slab, res);
 403                 l_unlock(&ns->ns_lock);
 404
 405                 spin_lock(&ns->ns_counter_lock);
 406                 ns->ns_resources--;
 407                 spin_unlock(&ns->ns_counter_lock);
 408
 409                 rc = 1;
 410         } else {
 411                 ENTRY;
 412                 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 413                        atomic_read(&res->lr_refcount));
 414         out:
 415                 LASSERT(atomic_read(&res->lr_refcount) >= 0);
 416         }
 417
 418         RETURN(rc);
 419 }
 420
 421 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 422                             struct ldlm_lock *lock)
 423 {
 424         l_lock(&res->lr_namespace->ns_lock);
 425
 426         ldlm_resource_dump(res);
 427         ldlm_lock_dump(lock);
 428
 429         LASSERT(list_empty(&lock->l_res_link));
 430
 431         list_add(&lock->l_res_link, head);
 432         l_unlock(&res->lr_namespace->ns_lock);
 433 }
 434
 435 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 436 {
 437         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 438         list_del_init(&lock->l_res_link);
 439         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 440 }
 441
 442 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 443 {
 444         desc->lr_type = res->lr_type;
 445         memcpy(desc->lr_name, res->lr_name, sizeof(desc->lr_name));
 446         memcpy(desc->lr_version, res->lr_version, sizeof(desc->lr_version));
 447 }
 448
 449 void ldlm_dump_all_namespaces(void)
 450 {
 451         struct list_head *tmp;
 452
 453         spin_lock(&ldlm_namespace_lock);
 454
 455         list_for_each(tmp, &ldlm_namespace_list) {
 456                 struct ldlm_namespace *ns;
 457                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 458                 ldlm_namespace_dump(ns);
 459         }
 460
 461         spin_unlock(&ldlm_namespace_lock);
 462 }
 463
 464 void ldlm_namespace_dump(struct ldlm_namespace *ns)
 465 {
 466         struct list_head *tmp;
 467
 468         l_lock(&ns->ns_lock);
 469         CDEBUG(D_OTHER, "--- Namespace: %s (rc: %d, client: %d)\n", ns->ns_name,
 470                ns->ns_refcount, ns->ns_client);
 471
 472         list_for_each(tmp, &ns->ns_root_list) {
 473                 struct ldlm_resource *res;
 474                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
 475
 476                 /* Once we have resources with children, this should really dump
 477                  * them recursively. */
 478                 ldlm_resource_dump(res);
 479         }
 480         l_unlock(&ns->ns_lock);
 481 }
 482
 483 void ldlm_resource_dump(struct ldlm_resource *res)
 484 {
 485         struct list_head *tmp;
 486         char name[256];
 487
 488         if (RES_NAME_SIZE != 3)
 489                 LBUG();
 490
 491         snprintf(name, sizeof(name), "%Lx %Lx %Lx",
 492                  (unsigned long long)res->lr_name[0],
 493                  (unsigned long long)res->lr_name[1],
 494                  (unsigned long long)res->lr_name[2]);
 495
 496         CDEBUG(D_OTHER, "--- Resource: %p (%s) (rc: %d)\n", res, name,
 497                atomic_read(&res->lr_refcount));
 498         CDEBUG(D_OTHER, "Namespace: %p (%s)\n", res->lr_namespace,
 499                res->lr_namespace->ns_name);
 500         CDEBUG(D_OTHER, "Parent: %p, root: %p\n", res->lr_parent, res->lr_root);
 501
 502         CDEBUG(D_OTHER, "Granted locks:\n");
 503         list_for_each(tmp, &res->lr_granted) {
 504                 struct ldlm_lock *lock;
 505                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 506                 ldlm_lock_dump(lock);
 507         }
 508
 509         CDEBUG(D_OTHER, "Converting locks:\n");
 510         list_for_each(tmp, &res->lr_converting) {
 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, "Waiting locks:\n");
 517         list_for_each(tmp, &res->lr_waiting) {
 518                 struct ldlm_lock *lock;
 519                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 520                 ldlm_lock_dump(lock);
 521         }
 522 }