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, 2003 Cluster File Systems, Inc.
   5  *   Author: Phil Schwan <phil@clusterfs.com>
   6  *   Author: Peter Braam <braam@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #define DEBUG_SUBSYSTEM S_LDLM
  25
  26 #include <linux/lustre_dlm.h>
  27 #include <linux/obd_class.h>
  28
  29 kmem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
  30
  31 spinlock_t ldlm_namespace_lock = SPIN_LOCK_UNLOCKED;
  32 struct list_head ldlm_namespace_list = LIST_HEAD_INIT(ldlm_namespace_list);
  33 static struct proc_dir_entry *ldlm_ns_proc_dir = NULL;
  34
  35 int ldlm_proc_setup(struct obd_device *obd)
  36 {
  37         int rc;
  38         ENTRY;
  39         LASSERT(ldlm_ns_proc_dir == NULL);
  40         rc = lprocfs_obd_attach(obd, 0);
  41         if (rc) {
  42                 CERROR("LProcFS failed in ldlm-init\n");
  43                 RETURN(rc);
  44         }
  45         ldlm_ns_proc_dir = obd->obd_proc_entry;
  46         RETURN(0);
  47 }
  48
  49 void ldlm_proc_cleanup(struct obd_device *obd)
  50 {
  51         if (ldlm_ns_proc_dir) {
  52                 lprocfs_obd_detach(obd);
  53                 ldlm_ns_proc_dir = NULL;
  54         }
  55 }
  56
  57 static int lprocfs_uint_rd(char *page, char **start, off_t off,
  58                            int count, int *eof, void *data)
  59 {
  60         unsigned int *temp = (unsigned int *)data;
  61         return snprintf(page, count, "%u\n", *temp);
  62 }
  63
  64 #define MAX_STRING_SIZE 128
  65 void ldlm_proc_namespace(struct ldlm_namespace *ns)
  66 {
  67         struct lprocfs_vars lock_vars[2];
  68         char lock_name[MAX_STRING_SIZE + 1];
  69
  70         lock_name[MAX_STRING_SIZE] = '\0';
  71
  72         memset(lock_vars, 0, sizeof(lock_vars));
  73         lock_vars[0].read_fptr = lprocfs_rd_u64;
  74
  75         lock_vars[0].name = lock_name;
  76
  77         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
  78
  79         lock_vars[0].data = &ns->ns_resources;
  80         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
  81
  82         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
  83         lock_vars[0].data = &ns->ns_locks;
  84         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
  85
  86         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
  87                  ns->ns_name);
  88         lock_vars[0].data = &ns->ns_nr_unused;
  89         lock_vars[0].read_fptr = lprocfs_uint_rd;
  90         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
  91 }
  92 #undef MAX_STRING_SIZE
  93
  94 #define LDLM_MAX_UNUSED 20
  95 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
  96 {
  97         struct ldlm_namespace *ns = NULL;
  98         struct list_head *bucket;
  99         ENTRY;
 100
 101         OBD_ALLOC(ns, sizeof(*ns));
 102         if (!ns)
 103                 RETURN(NULL);
 104
 105         ns->ns_hash = vmalloc(sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 106         if (!ns->ns_hash)
 107                 GOTO(out_ns, NULL);
 108
 109         atomic_add(sizeof(*ns->ns_hash) * RES_HASH_SIZE, &obd_memory);
 110
 111         OBD_ALLOC(ns->ns_name, strlen(name) + 1);
 112         if (!ns->ns_name)
 113                 GOTO(out_hash, NULL);
 114
 115         strcpy(ns->ns_name, name);
 116
 117         INIT_LIST_HEAD(&ns->ns_root_list);
 118         l_lock_init(&ns->ns_lock);
 119         ns->ns_refcount = 0;
 120         ns->ns_client = client;
 121         spin_lock_init(&ns->ns_counter_lock);
 122         ns->ns_locks = 0;
 123         ns->ns_resources = 0;
 124
 125         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 126              bucket--)
 127                 INIT_LIST_HEAD(bucket);
 128
 129         INIT_LIST_HEAD(&ns->ns_unused_list);
 130         ns->ns_nr_unused = 0;
 131         ns->ns_max_unused = LDLM_MAX_UNUSED;
 132
 133         spin_lock(&ldlm_namespace_lock);
 134         list_add(&ns->ns_list_chain, &ldlm_namespace_list);
 135         spin_unlock(&ldlm_namespace_lock);
 136         ldlm_proc_namespace(ns);
 137         RETURN(ns);
 138
 139 out_hash:
 140         memset(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 141         vfree(ns->ns_hash);
 142         atomic_sub(sizeof(*ns->ns_hash) * RES_HASH_SIZE, &obd_memory);
 143 out_ns:
 144         OBD_FREE(ns, sizeof(*ns));
 145         return NULL;
 146 }
 147
 148 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 149
 150 /* If 'local_only' is true, don't try to tell the server, just cleanup.
 151  * This is currently only used for recovery, and we make certain assumptions
 152  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 153  *
 154  * Called with the ns_lock held. */
 155 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 156                              int local_only)
 157 {
 158         struct list_head *tmp, *pos;
 159         int rc = 0, client = res->lr_namespace->ns_client;
 160         ENTRY;
 161
 162         list_for_each_safe(tmp, pos, q) {
 163                 struct ldlm_lock *lock;
 164                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 165                 LDLM_LOCK_GET(lock);
 166
 167                 if (local_only && (lock->l_readers || lock->l_writers)) {
 168                         /* This is a little bit gross, but much better than the
 169                          * alternative: pretend that we got a blocking AST from
 170                          * the server, so that when the lock is decref'd, it
 171                          * will go away ... */
 172                         lock->l_flags |= LDLM_FL_CBPENDING;
 173                         /* ... without sending a CANCEL message. */
 174                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 175                         /* ... and without calling the cancellation callback */
 176                         lock->l_flags |= LDLM_FL_CANCEL;
 177                         LDLM_LOCK_PUT(lock);
 178                         continue;
 179                 }
 180
 181                 /* At shutdown time, don't call the cancellation callback */
 182                 lock->l_flags |= LDLM_FL_CANCEL;
 183
 184                 if (client) {
 185                         struct lustre_handle lockh;
 186                         ldlm_lock2handle(lock, &lockh);
 187                         if (!local_only) {
 188                                 rc = ldlm_cli_cancel(&lockh);
 189                                 if (rc)
 190                                         CERROR("ldlm_cli_cancel: %d\n", rc);
 191                         }
 192                         /* Force local cleanup on errors, too. */
 193                         if (local_only || rc != ELDLM_OK)
 194                                 ldlm_lock_cancel(lock);
 195                 } else {
 196                         LDLM_DEBUG0(lock, "Freeing a lock still held by a "
 197                                     "client node");
 198
 199                         ldlm_resource_unlink_lock(lock);
 200                         ldlm_lock_destroy(lock);
 201                 }
 202                 LDLM_LOCK_PUT(lock);
 203         }
 204         EXIT;
 205 }
 206
 207 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only)
 208 {
 209         int i;
 210
 211         if (ns == NULL) {
 212                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 213                 return ELDLM_OK;
 214         }
 215
 216         l_lock(&ns->ns_lock);
 217         for (i = 0; i < RES_HASH_SIZE; i++) {
 218                 struct list_head *tmp, *pos;
 219                 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
 220                         struct ldlm_resource *res;
 221                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 222                         ldlm_resource_getref(res);
 223
 224                         cleanup_resource(res, &res->lr_granted, local_only);
 225                         cleanup_resource(res, &res->lr_converting, local_only);
 226                         cleanup_resource(res, &res->lr_waiting, local_only);
 227
 228                         /* XXX what a mess: don't force cleanup if we're
 229                          * local_only (which is only used by recovery).  In that
 230                          * case, we probably still have outstanding lock refs
 231                          * which reference these resources. -phil */
 232                         if (!ldlm_resource_putref(res) && !local_only) {
 233                                 CERROR("Resource refcount nonzero (%d) after "
 234                                        "lock cleanup; forcing cleanup.\n",
 235                                        atomic_read(&res->lr_refcount));
 236                                 ldlm_resource_dump(res);
 237                                 atomic_set(&res->lr_refcount, 1);
 238                                 ldlm_resource_putref(res);
 239                         }
 240                 }
 241         }
 242         l_unlock(&ns->ns_lock);
 243
 244         return ELDLM_OK;
 245 }
 246
 247 /* Cleanup, but also free, the namespace */
 248 int ldlm_namespace_free(struct ldlm_namespace *ns)
 249 {
 250         if (!ns)
 251                 RETURN(ELDLM_OK);
 252
 253         spin_lock(&ldlm_namespace_lock);
 254         list_del(&ns->ns_list_chain);
 255
 256         spin_unlock(&ldlm_namespace_lock);
 257
 258         ldlm_namespace_cleanup(ns, 0);
 259
 260         memset(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 261         vfree(ns->ns_hash /* , sizeof(*ns->ns_hash) * RES_HASH_SIZE */);
 262         atomic_sub(sizeof(*ns->ns_hash) * RES_HASH_SIZE, &obd_memory);
 263         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 264         OBD_FREE(ns, sizeof(*ns));
 265
 266         return ELDLM_OK;
 267 }
 268
 269 int ldlm_client_free(struct obd_export *exp)
 270 {
 271         struct ldlm_export_data *led = &exp->exp_ldlm_data;
 272         ptlrpc_cleanup_client(&led->led_import);
 273         RETURN(0);
 274 }
 275
 276 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, struct ldlm_res_id name)
 277 {
 278         __u32 hash = 0;
 279         int i;
 280
 281         for (i = 0; i < RES_NAME_SIZE; i++)
 282                 hash += name.name[i];
 283
 284         hash += (__u32)((unsigned long)parent >> 4);
 285
 286         return (hash & RES_HASH_MASK);
 287 }
 288
 289 static struct ldlm_resource *ldlm_resource_new(void)
 290 {
 291         struct ldlm_resource *res;
 292
 293         res = kmem_cache_alloc(ldlm_resource_slab, SLAB_KERNEL);
 294         if (res == NULL) {
 295                 LBUG();
 296                 return NULL;
 297         }
 298         memset(res, 0, sizeof(*res));
 299
 300         INIT_LIST_HEAD(&res->lr_children);
 301         INIT_LIST_HEAD(&res->lr_childof);
 302         INIT_LIST_HEAD(&res->lr_granted);
 303         INIT_LIST_HEAD(&res->lr_converting);
 304         INIT_LIST_HEAD(&res->lr_waiting);
 305
 306         atomic_set(&res->lr_refcount, 1);
 307
 308         return res;
 309 }
 310
 311 /* Args: locked namespace
 312  * Returns: newly-allocated, referenced, unlocked resource */
 313 static struct ldlm_resource *
 314 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 315                   struct ldlm_res_id name, __u32 type)
 316 {
 317         struct list_head *bucket;
 318         struct ldlm_resource *res;
 319         ENTRY;
 320
 321         if (type < LDLM_MIN_TYPE || type > LDLM_MAX_TYPE) {
 322                 LBUG();
 323                 RETURN(NULL);
 324         }
 325
 326         res = ldlm_resource_new();
 327         if (!res) {
 328                 LBUG();
 329                 RETURN(NULL);
 330         }
 331
 332         spin_lock(&ns->ns_counter_lock);
 333         ns->ns_resources++;
 334         spin_unlock(&ns->ns_counter_lock);
 335
 336         l_lock(&ns->ns_lock);
 337         memcpy(&res->lr_name, &name, sizeof(res->lr_name));
 338         res->lr_namespace = ns;
 339         ns->ns_refcount++;
 340
 341         res->lr_type = type;
 342         res->lr_most_restr = LCK_NL;
 343
 344         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 345         list_add(&res->lr_hash, bucket);
 346
 347         if (parent == NULL) {
 348                 list_add(&res->lr_childof, &ns->ns_root_list);
 349         } else {
 350                 res->lr_parent = parent;
 351                 list_add(&res->lr_childof, &parent->lr_children);
 352         }
 353         l_unlock(&ns->ns_lock);
 354
 355         RETURN(res);
 356 }
 357
 358 /* Args: unlocked namespace
 359  * Locks: takes and releases ns->ns_lock and res->lr_lock
 360  * Returns: referenced, unlocked ldlm_resource or NULL */
 361 struct ldlm_resource *
 362 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 363                   struct ldlm_res_id name, __u32 type, int create)
 364 {
 365         struct list_head *bucket, *tmp;
 366         struct ldlm_resource *res = NULL;
 367         ENTRY;
 368
 369         LASSERT(ns != NULL);
 370         LASSERT(ns->ns_hash != NULL);
 371
 372         l_lock(&ns->ns_lock);
 373         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 374
 375         list_for_each(tmp, bucket) {
 376                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 377
 378                 if (memcmp(&res->lr_name, &name, sizeof(res->lr_name)) == 0) {
 379                         ldlm_resource_getref(res);
 380                         l_unlock(&ns->ns_lock);
 381                         RETURN(res);
 382                 }
 383         }
 384
 385         if (create)
 386                 res = ldlm_resource_add(ns, parent, name, type);
 387         else
 388                 res = NULL;
 389
 390         l_unlock(&ns->ns_lock);
 391
 392         RETURN(res);
 393 }
 394
 395 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 396 {
 397         atomic_inc(&res->lr_refcount);
 398         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 399                atomic_read(&res->lr_refcount));
 400         return res;
 401 }
 402
 403 /* Returns 1 if the resource was freed, 0 if it remains. */
 404 int ldlm_resource_putref(struct ldlm_resource *res)
 405 {
 406         int rc = 0;
 407         ENTRY;
 408
 409         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 410                atomic_read(&res->lr_refcount) - 1);
 411         LASSERT(atomic_read(&res->lr_refcount) > 0);
 412         LASSERT(atomic_read(&res->lr_refcount) < 0x5a5a5a5a);
 413
 414         if (atomic_dec_and_test(&res->lr_refcount)) {
 415                 struct ldlm_namespace *ns = res->lr_namespace;
 416                 ENTRY;
 417
 418                 l_lock(&ns->ns_lock);
 419
 420                 if (atomic_read(&res->lr_refcount) != 0) {
 421                         /* We lost the race. */
 422                         l_unlock(&ns->ns_lock);
 423                         RETURN(rc);
 424                 }
 425
 426                 if (!list_empty(&res->lr_granted)) {
 427                         ldlm_resource_dump(res);
 428                         LBUG();
 429                 }
 430
 431                 if (!list_empty(&res->lr_converting)) {
 432                         ldlm_resource_dump(res);
 433                         LBUG();
 434                 }
 435
 436                 if (!list_empty(&res->lr_waiting)) {
 437                         ldlm_resource_dump(res);
 438                         LBUG();
 439                 }
 440
 441                 if (!list_empty(&res->lr_children)) {
 442                         ldlm_resource_dump(res);
 443                         LBUG();
 444                 }
 445
 446                 ns->ns_refcount--;
 447                 list_del_init(&res->lr_hash);
 448                 list_del_init(&res->lr_childof);
 449
 450                 memset(res, 0x5a, sizeof(*res));
 451                 kmem_cache_free(ldlm_resource_slab, res);
 452                 l_unlock(&ns->ns_lock);
 453
 454                 spin_lock(&ns->ns_counter_lock);
 455                 ns->ns_resources--;
 456                 spin_unlock(&ns->ns_counter_lock);
 457
 458                 rc = 1;
 459                 EXIT;
 460         }
 461
 462         RETURN(rc);
 463 }
 464
 465 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 466                             struct ldlm_lock *lock)
 467 {
 468         l_lock(&res->lr_namespace->ns_lock);
 469
 470         ldlm_resource_dump(res);
 471         CDEBUG(D_OTHER, "About to add this lock:\n");
 472         ldlm_lock_dump(D_OTHER, lock);
 473
 474         if (lock->l_destroyed) {
 475                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 476                 return;
 477         }
 478
 479         LASSERT(list_empty(&lock->l_res_link));
 480
 481         list_add_tail(&lock->l_res_link, head);
 482         l_unlock(&res->lr_namespace->ns_lock);
 483 }
 484
 485 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 486 {
 487         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 488         list_del_init(&lock->l_res_link);
 489         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 490 }
 491
 492 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 493 {
 494         desc->lr_type = res->lr_type;
 495         memcpy(&desc->lr_name, &res->lr_name, sizeof(desc->lr_name));
 496         memcpy(desc->lr_version, res->lr_version, sizeof(desc->lr_version));
 497 }
 498
 499 void ldlm_dump_all_namespaces(void)
 500 {
 501         struct list_head *tmp;
 502
 503         spin_lock(&ldlm_namespace_lock);
 504
 505         list_for_each(tmp, &ldlm_namespace_list) {
 506                 struct ldlm_namespace *ns;
 507                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 508                 ldlm_namespace_dump(ns);
 509         }
 510
 511         spin_unlock(&ldlm_namespace_lock);
 512 }
 513
 514 void ldlm_namespace_dump(struct ldlm_namespace *ns)
 515 {
 516         struct list_head *tmp;
 517
 518         l_lock(&ns->ns_lock);
 519         CDEBUG(D_OTHER, "--- Namespace: %s (rc: %d, client: %d)\n", ns->ns_name,
 520                ns->ns_refcount, ns->ns_client);
 521
 522         list_for_each(tmp, &ns->ns_root_list) {
 523                 struct ldlm_resource *res;
 524                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
 525
 526                 /* Once we have resources with children, this should really dump
 527                  * them recursively. */
 528                 ldlm_resource_dump(res);
 529         }
 530         l_unlock(&ns->ns_lock);
 531 }
 532
 533 void ldlm_resource_dump(struct ldlm_resource *res)
 534 {
 535         struct list_head *tmp;
 536         char name[256];
 537
 538         if (RES_NAME_SIZE != 3)
 539                 LBUG();
 540
 541         snprintf(name, sizeof(name), "%Lx %Lx %Lx",
 542                  (unsigned long long)res->lr_name.name[0],
 543                  (unsigned long long)res->lr_name.name[1],
 544                  (unsigned long long)res->lr_name.name[2]);
 545
 546         CDEBUG(D_OTHER, "--- Resource: %p (%s) (rc: %d)\n", res, name,
 547                atomic_read(&res->lr_refcount));
 548         CDEBUG(D_OTHER, "Namespace: %p (%s)\n", res->lr_namespace,
 549                res->lr_namespace->ns_name);
 550         CDEBUG(D_OTHER, "Parent: %p, root: %p\n", res->lr_parent, res->lr_root);
 551
 552         CDEBUG(D_OTHER, "Granted locks:\n");
 553         list_for_each(tmp, &res->lr_granted) {
 554                 struct ldlm_lock *lock;
 555                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 556                 ldlm_lock_dump(D_OTHER, lock);
 557         }
 558
 559         CDEBUG(D_OTHER, "Converting locks:\n");
 560         list_for_each(tmp, &res->lr_converting) {
 561                 struct ldlm_lock *lock;
 562                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 563                 ldlm_lock_dump(D_OTHER, lock);
 564         }
 565
 566         CDEBUG(D_OTHER, "Waiting locks:\n");
 567         list_for_each(tmp, &res->lr_waiting) {
 568                 struct ldlm_lock *lock;
 569                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 570                 ldlm_lock_dump(D_OTHER, lock);
 571         }
 572 }