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 #ifdef __KERNEL__
  26 #include <linux/lustre_dlm.h>
  27 #else
  28 #include <liblustre.h>
  29 #endif
  30
  31 #include <linux/obd_class.h>
  32
  33 kmem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
  34
  35 spinlock_t ldlm_namespace_lock = SPIN_LOCK_UNLOCKED;
  36 struct list_head ldlm_namespace_list = LIST_HEAD_INIT(ldlm_namespace_list);
  37 static struct proc_dir_entry *ldlm_ns_proc_dir = NULL;
  38
  39 int ldlm_proc_setup(struct obd_device *obd)
  40 {
  41         int rc;
  42         ENTRY;
  43         LASSERT(ldlm_ns_proc_dir == NULL);
  44         LASSERT(obd != NULL);
  45         rc = lprocfs_obd_attach(obd, 0);
  46         if (rc) {
  47                 CERROR("LProcFS failed in ldlm-init\n");
  48                 RETURN(rc);
  49         }
  50         ldlm_ns_proc_dir = obd->obd_proc_entry;
  51         RETURN(0);
  52 }
  53
  54 void ldlm_proc_cleanup(struct obd_device *obd)
  55 {
  56         if (ldlm_ns_proc_dir) {
  57                 lprocfs_obd_detach(obd);
  58                 ldlm_ns_proc_dir = NULL;
  59         }
  60 }
  61
  62 #ifdef __KERNEL__
  63 static int lprocfs_uint_rd(char *page, char **start, off_t off,
  64                            int count, int *eof, void *data)
  65 {
  66         unsigned int *temp = (unsigned int *)data;
  67         return snprintf(page, count, "%u\n", *temp);
  68 }
  69
  70
  71 #define MAX_STRING_SIZE 128
  72 void ldlm_proc_namespace(struct ldlm_namespace *ns)
  73 {
  74         struct lprocfs_vars lock_vars[2];
  75         char lock_name[MAX_STRING_SIZE + 1];
  76
  77         LASSERT(ns != NULL);
  78         LASSERT(ns->ns_name != NULL);
  79
  80         lock_name[MAX_STRING_SIZE] = '\0';
  81
  82         memset(lock_vars, 0, sizeof(lock_vars));
  83         lock_vars[0].read_fptr = lprocfs_rd_u64;
  84
  85         lock_vars[0].name = lock_name;
  86
  87         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
  88
  89         lock_vars[0].data = &ns->ns_resources;
  90         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
  91
  92         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
  93
  94         lock_vars[0].data = &ns->ns_locks;
  95         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
  96
  97         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
  98                  ns->ns_name);
  99         lock_vars[0].data = &ns->ns_nr_unused;
 100         lock_vars[0].read_fptr = lprocfs_uint_rd;
 101         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 102 }
 103 #endif
 104 #undef MAX_STRING_SIZE
 105
 106 #define LDLM_MAX_UNUSED 100
 107 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
 108 {
 109         struct ldlm_namespace *ns = NULL;
 110         struct list_head *bucket;
 111         ENTRY;
 112
 113         OBD_ALLOC(ns, sizeof(*ns));
 114         if (!ns)
 115                 RETURN(NULL);
 116
 117         ns->ns_hash = vmalloc(sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 118         if (!ns->ns_hash)
 119                 GOTO(out_ns, NULL);
 120
 121         atomic_add(sizeof(*ns->ns_hash) * RES_HASH_SIZE, &obd_memory);
 122
 123         OBD_ALLOC(ns->ns_name, strlen(name) + 1);
 124         if (!ns->ns_name)
 125                 GOTO(out_hash, NULL);
 126
 127         strcpy(ns->ns_name, name);
 128
 129         INIT_LIST_HEAD(&ns->ns_root_list);
 130         l_lock_init(&ns->ns_lock);
 131         ns->ns_refcount = 0;
 132         ns->ns_client = client;
 133         spin_lock_init(&ns->ns_counter_lock);
 134         ns->ns_locks = 0;
 135         ns->ns_resources = 0;
 136
 137         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
 138              bucket--)
 139                 INIT_LIST_HEAD(bucket);
 140
 141         INIT_LIST_HEAD(&ns->ns_unused_list);
 142         ns->ns_nr_unused = 0;
 143         ns->ns_max_unused = LDLM_MAX_UNUSED;
 144
 145         spin_lock(&ldlm_namespace_lock);
 146         list_add(&ns->ns_list_chain, &ldlm_namespace_list);
 147         spin_unlock(&ldlm_namespace_lock);
 148 #ifdef __KERNEL__
 149         ldlm_proc_namespace(ns);
 150 #endif
 151         RETURN(ns);
 152
 153 out_hash:
 154         POISON(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 155         vfree(ns->ns_hash);
 156         atomic_sub(sizeof(*ns->ns_hash) * RES_HASH_SIZE, &obd_memory);
 157 out_ns:
 158         OBD_FREE(ns, sizeof(*ns));
 159         return NULL;
 160 }
 161
 162 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 163
 164 /* If 'local_only' is true, don't try to tell the server, just cleanup.
 165  * This is currently only used for recovery, and we make certain assumptions
 166  * as a result--notably, that we shouldn't cancel locks with refs. -phil
 167  *
 168  * Called with the ns_lock held. */
 169 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 170                              int local_only)
 171 {
 172         struct list_head *tmp, *pos;
 173         int rc = 0, client = res->lr_namespace->ns_client;
 174         ENTRY;
 175
 176         list_for_each_safe(tmp, pos, q) {
 177                 struct ldlm_lock *lock;
 178                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 179                 LDLM_LOCK_GET(lock);
 180
 181                 if (local_only && (lock->l_readers || lock->l_writers)) {
 182                         /* This is a little bit gross, but much better than the
 183                          * alternative: pretend that we got a blocking AST from
 184                          * the server, so that when the lock is decref'd, it
 185                          * will go away ... */
 186                         lock->l_flags |= LDLM_FL_CBPENDING;
 187                         /* ... without sending a CANCEL message. */
 188                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 189                         /* ... and without calling the cancellation callback */
 190                         lock->l_flags |= LDLM_FL_CANCEL;
 191                         LDLM_LOCK_PUT(lock);
 192                         continue;
 193                 }
 194
 195                 /* At shutdown time, don't call the cancellation callback */
 196                 lock->l_flags |= LDLM_FL_CANCEL;
 197
 198                 if (client) {
 199                         struct lustre_handle lockh;
 200                         ldlm_lock2handle(lock, &lockh);
 201                         if (!local_only) {
 202                                 rc = ldlm_cli_cancel(&lockh);
 203                                 if (rc)
 204                                         CERROR("ldlm_cli_cancel: %d\n", rc);
 205                         }
 206                         /* Force local cleanup on errors, too. */
 207                         if (local_only || rc != ELDLM_OK)
 208                                 ldlm_lock_cancel(lock);
 209                 } else {
 210                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 211                                    "client node");
 212
 213                         ldlm_resource_unlink_lock(lock);
 214                         ldlm_lock_destroy(lock);
 215                 }
 216                 LDLM_LOCK_PUT(lock);
 217         }
 218         EXIT;
 219 }
 220
 221 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only)
 222 {
 223         int i;
 224
 225         if (ns == NULL) {
 226                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 227                 return ELDLM_OK;
 228         }
 229
 230         l_lock(&ns->ns_lock);
 231         for (i = 0; i < RES_HASH_SIZE; i++) {
 232                 struct list_head *tmp, *pos;
 233                 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
 234                         struct ldlm_resource *res;
 235                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 236                         ldlm_resource_getref(res);
 237
 238                         cleanup_resource(res, &res->lr_granted, local_only);
 239                         cleanup_resource(res, &res->lr_converting, local_only);
 240                         cleanup_resource(res, &res->lr_waiting, local_only);
 241
 242                         /* XXX what a mess: don't force cleanup if we're
 243                          * local_only (which is only used by recovery).  In that
 244                          * case, we probably still have outstanding lock refs
 245                          * which reference these resources. -phil */
 246                         if (!ldlm_resource_putref(res) && !local_only) {
 247                                 CERROR("Resource refcount nonzero (%d) after "
 248                                        "lock cleanup; forcing cleanup.\n",
 249                                        atomic_read(&res->lr_refcount));
 250                                 ldlm_resource_dump(res);
 251                                 atomic_set(&res->lr_refcount, 1);
 252                                 ldlm_resource_putref(res);
 253                         }
 254                 }
 255         }
 256         l_unlock(&ns->ns_lock);
 257
 258         return ELDLM_OK;
 259 }
 260
 261 /* Cleanup, but also free, the namespace */
 262 int ldlm_namespace_free(struct ldlm_namespace *ns)
 263 {
 264         if (!ns)
 265                 RETURN(ELDLM_OK);
 266
 267         spin_lock(&ldlm_namespace_lock);
 268         list_del(&ns->ns_list_chain);
 269
 270         spin_unlock(&ldlm_namespace_lock);
 271
 272         ldlm_namespace_cleanup(ns, 0);
 273
 274         POISON(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 275         vfree(ns->ns_hash /* , sizeof(*ns->ns_hash) * RES_HASH_SIZE */);
 276         atomic_sub(sizeof(*ns->ns_hash) * RES_HASH_SIZE, &obd_memory);
 277         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
 278         OBD_FREE(ns, sizeof(*ns));
 279
 280         return ELDLM_OK;
 281 }
 282
 283 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, struct ldlm_res_id name)
 284 {
 285         __u32 hash = 0;
 286         int i;
 287
 288         for (i = 0; i < RES_NAME_SIZE; i++)
 289                 hash += name.name[i];
 290
 291         hash += (__u32)((unsigned long)parent >> 4);
 292
 293         return (hash & RES_HASH_MASK);
 294 }
 295
 296 static struct ldlm_resource *ldlm_resource_new(void)
 297 {
 298         struct ldlm_resource *res;
 299
 300         OBD_SLAB_ALLOC(res, ldlm_resource_slab, SLAB_KERNEL, sizeof *res);
 301         if (res == NULL) {
 302                 LBUG();
 303                 return NULL;
 304         }
 305         memset(res, 0, sizeof(*res));
 306
 307         INIT_LIST_HEAD(&res->lr_children);
 308         INIT_LIST_HEAD(&res->lr_childof);
 309         INIT_LIST_HEAD(&res->lr_granted);
 310         INIT_LIST_HEAD(&res->lr_converting);
 311         INIT_LIST_HEAD(&res->lr_waiting);
 312
 313         atomic_set(&res->lr_refcount, 1);
 314
 315         return res;
 316 }
 317
 318 /* Args: locked namespace
 319  * Returns: newly-allocated, referenced, unlocked resource */
 320 static struct ldlm_resource *
 321 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 322                   struct ldlm_res_id name, __u32 type)
 323 {
 324         struct list_head *bucket;
 325         struct ldlm_resource *res;
 326         ENTRY;
 327
 328         if (type < LDLM_MIN_TYPE || type > LDLM_MAX_TYPE) {
 329                 LBUG();
 330                 RETURN(NULL);
 331         }
 332
 333         res = ldlm_resource_new();
 334         if (!res) {
 335                 LBUG();
 336                 RETURN(NULL);
 337         }
 338
 339         spin_lock(&ns->ns_counter_lock);
 340         ns->ns_resources++;
 341         spin_unlock(&ns->ns_counter_lock);
 342
 343         l_lock(&ns->ns_lock);
 344         memcpy(&res->lr_name, &name, sizeof(res->lr_name));
 345         res->lr_namespace = ns;
 346         ns->ns_refcount++;
 347
 348         res->lr_type = type;
 349         res->lr_most_restr = LCK_NL;
 350
 351         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 352         list_add(&res->lr_hash, bucket);
 353
 354         if (parent == NULL) {
 355                 list_add(&res->lr_childof, &ns->ns_root_list);
 356         } else {
 357                 res->lr_parent = parent;
 358                 list_add(&res->lr_childof, &parent->lr_children);
 359         }
 360         l_unlock(&ns->ns_lock);
 361
 362         RETURN(res);
 363 }
 364
 365 /* Args: unlocked namespace
 366  * Locks: takes and releases ns->ns_lock and res->lr_lock
 367  * Returns: referenced, unlocked ldlm_resource or NULL */
 368 struct ldlm_resource *
 369 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
 370                   struct ldlm_res_id name, __u32 type, int create)
 371 {
 372         struct list_head *bucket, *tmp;
 373         struct ldlm_resource *res = NULL;
 374         ENTRY;
 375
 376         LASSERT(ns != NULL);
 377         LASSERT(ns->ns_hash != NULL);
 378
 379         l_lock(&ns->ns_lock);
 380         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 381
 382         list_for_each(tmp, bucket) {
 383                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
 384
 385                 if (memcmp(&res->lr_name, &name, sizeof(res->lr_name)) == 0) {
 386                         ldlm_resource_getref(res);
 387                         l_unlock(&ns->ns_lock);
 388                         RETURN(res);
 389                 }
 390         }
 391
 392         if (create)
 393                 res = ldlm_resource_add(ns, parent, name, type);
 394         else
 395                 res = NULL;
 396
 397         l_unlock(&ns->ns_lock);
 398
 399         RETURN(res);
 400 }
 401
 402 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
 403 {
 404         atomic_inc(&res->lr_refcount);
 405         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
 406                atomic_read(&res->lr_refcount));
 407         return res;
 408 }
 409
 410 /* Returns 1 if the resource was freed, 0 if it remains. */
 411 int ldlm_resource_putref(struct ldlm_resource *res)
 412 {
 413         int rc = 0;
 414         ENTRY;
 415
 416         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
 417                atomic_read(&res->lr_refcount) - 1);
 418         LASSERT(atomic_read(&res->lr_refcount) > 0);
 419         LASSERT(atomic_read(&res->lr_refcount) < 0x5a5a5a5a);
 420
 421         if (atomic_dec_and_test(&res->lr_refcount)) {
 422                 struct ldlm_namespace *ns = res->lr_namespace;
 423                 ENTRY;
 424
 425                 l_lock(&ns->ns_lock);
 426
 427                 if (atomic_read(&res->lr_refcount) != 0) {
 428                         /* We lost the race. */
 429                         l_unlock(&ns->ns_lock);
 430                         RETURN(rc);
 431                 }
 432
 433                 if (!list_empty(&res->lr_granted)) {
 434                         ldlm_resource_dump(res);
 435                         LBUG();
 436                 }
 437
 438                 if (!list_empty(&res->lr_converting)) {
 439                         ldlm_resource_dump(res);
 440                         LBUG();
 441                 }
 442
 443                 if (!list_empty(&res->lr_waiting)) {
 444                         ldlm_resource_dump(res);
 445                         LBUG();
 446                 }
 447
 448                 if (!list_empty(&res->lr_children)) {
 449                         ldlm_resource_dump(res);
 450                         LBUG();
 451                 }
 452
 453                 ns->ns_refcount--;
 454                 list_del_init(&res->lr_hash);
 455                 list_del_init(&res->lr_childof);
 456
 457                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
 458                 l_unlock(&ns->ns_lock);
 459
 460                 spin_lock(&ns->ns_counter_lock);
 461                 ns->ns_resources--;
 462                 spin_unlock(&ns->ns_counter_lock);
 463
 464                 rc = 1;
 465                 EXIT;
 466         }
 467
 468         RETURN(rc);
 469 }
 470
 471 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 472                             struct ldlm_lock *lock)
 473 {
 474         l_lock(&res->lr_namespace->ns_lock);
 475
 476         ldlm_resource_dump(res);
 477         CDEBUG(D_OTHER, "About to add this lock:\n");
 478         ldlm_lock_dump(D_OTHER, lock);
 479
 480         if (lock->l_destroyed) {
 481                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
 482                 return;
 483         }
 484
 485         LASSERT(list_empty(&lock->l_res_link));
 486
 487         list_add_tail(&lock->l_res_link, head);
 488         l_unlock(&res->lr_namespace->ns_lock);
 489 }
 490
 491 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 492 {
 493         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 494         list_del_init(&lock->l_res_link);
 495         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 496 }
 497
 498 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
 499 {
 500         desc->lr_type = res->lr_type;
 501         memcpy(&desc->lr_name, &res->lr_name, sizeof(desc->lr_name));
 502         memcpy(desc->lr_version, res->lr_version, sizeof(desc->lr_version));
 503 }
 504
 505 void ldlm_dump_all_namespaces(void)
 506 {
 507         struct list_head *tmp;
 508
 509         spin_lock(&ldlm_namespace_lock);
 510
 511         list_for_each(tmp, &ldlm_namespace_list) {
 512                 struct ldlm_namespace *ns;
 513                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
 514                 ldlm_namespace_dump(ns);
 515         }
 516
 517         spin_unlock(&ldlm_namespace_lock);
 518 }
 519
 520 void ldlm_namespace_dump(struct ldlm_namespace *ns)
 521 {
 522         struct list_head *tmp;
 523
 524         l_lock(&ns->ns_lock);
 525         CDEBUG(D_OTHER, "--- Namespace: %s (rc: %d, client: %d)\n", ns->ns_name,
 526                ns->ns_refcount, ns->ns_client);
 527
 528         list_for_each(tmp, &ns->ns_root_list) {
 529                 struct ldlm_resource *res;
 530                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
 531
 532                 /* Once we have resources with children, this should really dump
 533                  * them recursively. */
 534                 ldlm_resource_dump(res);
 535         }
 536         l_unlock(&ns->ns_lock);
 537 }
 538
 539 void ldlm_resource_dump(struct ldlm_resource *res)
 540 {
 541         struct list_head *tmp;
 542         char name[256];
 543
 544         if (RES_NAME_SIZE != 3)
 545                 LBUG();
 546
 547         snprintf(name, sizeof(name), "%Lx %Lx %Lx",
 548                  (unsigned long long)res->lr_name.name[0],
 549                  (unsigned long long)res->lr_name.name[1],
 550                  (unsigned long long)res->lr_name.name[2]);
 551
 552         CDEBUG(D_OTHER, "--- Resource: %p (%s) (rc: %d)\n", res, name,
 553                atomic_read(&res->lr_refcount));
 554         CDEBUG(D_OTHER, "Namespace: %p (%s)\n", res->lr_namespace,
 555                res->lr_namespace->ns_name);
 556         CDEBUG(D_OTHER, "Parent: %p, root: %p\n", res->lr_parent, res->lr_root);
 557
 558         CDEBUG(D_OTHER, "Granted locks:\n");
 559         list_for_each(tmp, &res->lr_granted) {
 560                 struct ldlm_lock *lock;
 561                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 562                 ldlm_lock_dump(D_OTHER, lock);
 563         }
 564
 565         CDEBUG(D_OTHER, "Converting locks:\n");
 566         list_for_each(tmp, &res->lr_converting) {
 567                 struct ldlm_lock *lock;
 568                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 569                 ldlm_lock_dump(D_OTHER, lock);
 570         }
 571
 572         CDEBUG(D_OTHER, "Waiting locks:\n");
 573         list_for_each(tmp, &res->lr_waiting) {
 574                 struct ldlm_lock *lock;
 575                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 576                 ldlm_lock_dump(D_OTHER, lock);
 577         }
 578 }