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];
  41
  42         memset(lock_vars, 0, sizeof(lock_vars));
  43         snprintf(lock_names, MAX_STRING_SIZE, "%s/resource_count",
  44                  ns->ns_name);
  45         lock_names[MAX_STRING_SIZE] = '\0';
  46         lock_vars[0].name = lock_names;
  47         lock_vars[0].read_fptr = lprocfs_ll_rd;
  48         lock_vars[0].write_fptr = NULL;
  49         lock_vars[0].data = &ns->ns_resources;
  50         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
  51
  52         memset(lock_vars, 0, sizeof(lock_vars));
  53         snprintf(lock_names, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
  54         lock_names[MAX_STRING_SIZE] = '\0';
  55         lock_vars[0].name = lock_names;
  56         lock_vars[0].read_fptr = lprocfs_ll_rd;
  57         lock_vars[0].write_fptr = NULL;
  58         lock_vars[0].data = &ns->ns_locks;
  59         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
  60
  61 }
  62 #undef MAX_STRING_SIZE
  63
  64 #define LDLM_MAX_UNUSED 20
  65 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
  66 {
  67         struct ldlm_namespace *ns = NULL;
  68         struct list_head *bucket;
  69
  70
  71
  72         OBD_ALLOC(ns, sizeof(*ns));
  73         if (!ns) {
  74                 LBUG();
  75                 GOTO(out, NULL);
  76         }
  77
  78         ns->ns_hash = vmalloc(sizeof(*ns->ns_hash) * RES_HASH_SIZE);
  79         if (!ns->ns_hash) {
  80                 LBUG();
  81                 GOTO(out, ns);
  82         }
  83         obd_memory += sizeof(*ns->ns_hash) * RES_HASH_SIZE;
  84
  85         OBD_ALLOC(ns->ns_name, strlen(name) + 1);
  86         if (!ns->ns_name) {
  87                 LBUG();
  88                 GOTO(out, ns);
  89         }
  90         strcpy(ns->ns_name, name);
  91
  92         INIT_LIST_HEAD(&ns->ns_root_list);
  93         l_lock_init(&ns->ns_lock);
  94         ns->ns_refcount = 0;
  95         ns->ns_client = client;
  96         spin_lock_init(&ns->ns_counter_lock);
  97         ns->ns_locks = 0;
  98         ns->ns_resources = 0;
  99
 100         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 101              bucket--)
 102                 INIT_LIST_HEAD(bucket);
 103
 104         INIT_LIST_HEAD(&ns->ns_unused_list);
 105         ns->ns_nr_unused = 0;
 106         ns->ns_max_unused = LDLM_MAX_UNUSED;
 107
 108         spin_lock(&ldlm_namespace_lock);
 109         list_add(&ns->ns_list_chain, &ldlm_namespace_list);
 110         spin_unlock(&ldlm_namespace_lock);
 111         ldlm_proc_namespace(ns);
 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
 211         spin_unlock(&ldlm_namespace_lock);
 212
 213         ldlm_namespace_cleanup(ns, 0);
 214
 215         vfree(ns->ns_hash /* , sizeof(*ns->ns_hash) * RES_HASH_SIZE */);
 216         obd_memory -= sizeof(*ns->ns_hash) * RES_HASH_SIZE;
 217         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 218         OBD_FREE(ns, sizeof(*ns));
 219
 220         return ELDLM_OK;
 221 }
 222
 223 int ldlm_client_free(struct obd_export *exp)
 224 {
 225         struct ldlm_export_data *led = &exp->exp_ldlm_data;
 226         ptlrpc_cleanup_client(&led->led_import);
 227         RETURN(0);
 228 }
 229
 230 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, __u64 *name)
 231 {
 232         __u32 hash = 0;
 233         int i;
 234
 235         for (i = 0; i < RES_NAME_SIZE; i++)
 236                 hash += name[i];
 237
 238         hash += (__u32)((unsigned long)parent >> 4);
 239
 240         return (hash & RES_HASH_MASK);
 241 }
 242
 243 static struct ldlm_resource *ldlm_resource_new(void)
 244 {
 245         struct ldlm_resource *res;
 246
 247         res = kmem_cache_alloc(ldlm_resource_slab, SLAB_KERNEL);
 248         if (res == NULL) {
 249                 LBUG();
 250                 return NULL;
 251         }
 252         memset(res, 0, sizeof(*res));
 253
 254         INIT_LIST_HEAD(&res->lr_children);
 255         INIT_LIST_HEAD(&res->lr_childof);
 256         INIT_LIST_HEAD(&res->lr_granted);
 257         INIT_LIST_HEAD(&res->lr_converting);
 258         INIT_LIST_HEAD(&res->lr_waiting);
 259
 260         atomic_set(&res->lr_refcount, 1);
 261
 262         return res;
 263 }
 264
 265 /* Args: locked namespace
 266  * Returns: newly-allocated, referenced, unlocked resource */
 267 static struct ldlm_resource *ldlm_resource_add(struct ldlm_namespace *ns,
 268                                                struct ldlm_resource *parent,
 269                                                __u64 *name, __u32 type)
 270 {
 271         struct list_head *bucket;
 272         struct ldlm_resource *res;
 273         ENTRY;
 274
 275         if (type < LDLM_MIN_TYPE || type > LDLM_MAX_TYPE) {
 276                 LBUG();
 277                 RETURN(NULL);
 278         }
 279
 280         res = ldlm_resource_new();
 281         if (!res) {
 282                 LBUG();
 283                 RETURN(NULL);
 284         }
 285
 286         spin_lock(&ns->ns_counter_lock);
 287         ns->ns_resources++;
 288         spin_unlock(&ns->ns_counter_lock);
 289
 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
 307         RETURN(res);
 308 }
 309
 310 /* Args: unlocked namespace
 311  * Locks: takes and releases ns->ns_lock and res->lr_lock
 312  * Returns: referenced, unlocked ldlm_resource or NULL */
 313 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 314                                         struct ldlm_resource *parent,
 315                                         __u64 *name, __u32 type, int create)
 316 {
 317         struct list_head *bucket;
 318         struct list_head *tmp = bucket;
 319         struct ldlm_resource *res = NULL;
 320         ENTRY;
 321
 322         if (ns == NULL || ns->ns_hash == NULL) {
 323                 LBUG();
 324                 RETURN(NULL);
 325         }
 326
 327         l_lock(&ns->ns_lock);
 328         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 329
 330         list_for_each(tmp, bucket) {
 331                 struct ldlm_resource *chk;
 332                 chk = list_entry(tmp, struct ldlm_resource, lr_hash);
 333
 334                 if (memcmp(chk->lr_name, name, sizeof(chk->lr_name)) == 0) {
 335                         res = chk;
 336                         atomic_inc(&res->lr_refcount);
 337                         EXIT;
 338                         break;
 339                 }
 340         }
 341
 342         if (res == NULL && create)
 343                 res = ldlm_resource_add(ns, parent, name, type);
 344         l_unlock(&ns->ns_lock);
 345
 346         RETURN(res);
 347 }
 348
 349 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 350 {
 351         atomic_inc(&res->lr_refcount);
 352         return res;
 353 }
 354
 355 /* Returns 1 if the resource was freed, 0 if it remains. */
 356 int ldlm_resource_put(struct ldlm_resource *res)
 357 {
 358         int rc = 0;
 359
 360         if (atomic_dec_and_test(&res->lr_refcount)) {
 361                 struct ldlm_namespace *ns = res->lr_namespace;
 362                 ENTRY;
 363
 364                 l_lock(&ns->ns_lock);
 365
 366                 if (atomic_read(&res->lr_refcount) != 0) {
 367                         /* We lost the race. */
 368                         l_unlock(&ns->ns_lock);
 369                         goto out;
 370                 }
 371
 372                 if (!list_empty(&res->lr_granted))
 373                         LBUG();
 374
 375                 if (!list_empty(&res->lr_converting))
 376                         LBUG();
 377
 378                 if (!list_empty(&res->lr_waiting))
 379                         LBUG();
 380
 381                 if (!list_empty(&res->lr_children))
 382                         LBUG();
 383
 384                 ns->ns_refcount--;
 385                 list_del(&res->lr_hash);
 386                 list_del(&res->lr_childof);
 387
 388                 kmem_cache_free(ldlm_resource_slab, res);
 389                 l_unlock(&ns->ns_lock);
 390
 391                 spin_lock(&ns->ns_counter_lock);
 392                 ns->ns_resources--;
 393                 spin_unlock(&ns->ns_counter_lock);
 394
 395                 rc = 1;
 396         } else {
 397                 ENTRY;
 398         out:
 399                 if (atomic_read(&res->lr_refcount) < 0)
 400                         LBUG();
 401         }
 402
 403         RETURN(rc);
 404 }
 405
 406 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 407                             struct ldlm_lock *lock)
 408 {
 409         l_lock(&res->lr_namespace->ns_lock);
 410
 411         ldlm_resource_dump(res);
 412         ldlm_lock_dump(lock);
 413
 414         if (!list_empty(&lock->l_res_link))
 415                 LBUG();
 416
 417         list_add(&lock->l_res_link, head);
 418         l_unlock(&res->lr_namespace->ns_lock);
 419 }
 420
 421 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 422 {
 423         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 424         list_del_init(&lock->l_res_link);
 425         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 426 }
 427
 428 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 429 {
 430         desc->lr_type = res->lr_type;
 431         memcpy(desc->lr_name, res->lr_name, sizeof(desc->lr_name));
 432         memcpy(desc->lr_version, res->lr_version, sizeof(desc->lr_version));
 433 }
 434
 435 void ldlm_dump_all_namespaces(void)
 436 {
 437         struct list_head *tmp;
 438
 439         spin_lock(&ldlm_namespace_lock);
 440
 441         list_for_each(tmp, &ldlm_namespace_list) {
 442                 struct ldlm_namespace *ns;
 443                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 444                 ldlm_namespace_dump(ns);
 445         }
 446
 447         spin_unlock(&ldlm_namespace_lock);
 448 }
 449
 450 void ldlm_namespace_dump(struct ldlm_namespace *ns)
 451 {
 452         struct list_head *tmp;
 453
 454         l_lock(&ns->ns_lock);
 455         CDEBUG(D_OTHER, "--- Namespace: %s (rc: %d, client: %d)\n", ns->ns_name,
 456                ns->ns_refcount, ns->ns_client);
 457
 458         list_for_each(tmp, &ns->ns_root_list) {
 459                 struct ldlm_resource *res;
 460                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
 461
 462                 /* Once we have resources with children, this should really dump
 463                  * them recursively. */
 464                 ldlm_resource_dump(res);
 465         }
 466         l_unlock(&ns->ns_lock);
 467 }
 468
 469 void ldlm_resource_dump(struct ldlm_resource *res)
 470 {
 471         struct list_head *tmp;
 472         char name[256];
 473
 474         if (RES_NAME_SIZE != 3)
 475                 LBUG();
 476
 477         snprintf(name, sizeof(name), "%Lx %Lx %Lx",
 478                  (unsigned long long)res->lr_name[0],
 479                  (unsigned long long)res->lr_name[1],
 480                  (unsigned long long)res->lr_name[2]);
 481
 482         CDEBUG(D_OTHER, "--- Resource: %p (%s) (rc: %d)\n", res, name,
 483                atomic_read(&res->lr_refcount));
 484         CDEBUG(D_OTHER, "Namespace: %p (%s)\n", res->lr_namespace,
 485                res->lr_namespace->ns_name);
 486         CDEBUG(D_OTHER, "Parent: %p, root: %p\n", res->lr_parent, res->lr_root);
 487
 488         CDEBUG(D_OTHER, "Granted locks:\n");
 489         list_for_each(tmp, &res->lr_granted) {
 490                 struct ldlm_lock *lock;
 491                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 492                 ldlm_lock_dump(lock);
 493         }
 494
 495         CDEBUG(D_OTHER, "Converting locks:\n");
 496         list_for_each(tmp, &res->lr_converting) {
 497                 struct ldlm_lock *lock;
 498                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 499                 ldlm_lock_dump(lock);
 500         }
 501
 502         CDEBUG(D_OTHER, "Waiting locks:\n");
 503         list_for_each(tmp, &res->lr_waiting) {
 504                 struct ldlm_lock *lock;
 505                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 506                 ldlm_lock_dump(lock);
 507         }
 508 }