lustre/ldlm/ldlm_lockd.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 EXPORT_SYMTAB
  13 #define DEBUG_SUBSYSTEM S_LDLM
  14
  15 #include <linux/module.h>
  16 #include <linux/slab.h>
  17 #include <linux/lustre_dlm.h>
  18
  19 extern kmem_cache_t *ldlm_resource_slab;
  20 extern kmem_cache_t *ldlm_lock_slab;
  21 extern int (*mds_reint_p)(int offset, struct ptlrpc_request *req);
  22 extern int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req);
  23
  24 static int ldlm_handle_enqueue(struct obd_device *obddev, struct ptlrpc_service *svc,
  25                          struct ptlrpc_request *req)
  26 {
  27         struct ldlm_reply *dlm_rep;
  28         struct ldlm_request *dlm_req;
  29         int rc, size = sizeof(*dlm_rep), cookielen = 0;
  30         __u32 flags;
  31         ldlm_error_t err;
  32         struct ldlm_lock *lock = NULL;
  33         void *cookie = NULL;
  34         ENTRY;
  35
  36         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
  37
  38         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
  39         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_MDSINTENT) {
  40                 /* In this case, the reply buffer is allocated deep in
  41                  * local_lock_enqueue by the policy function. */
  42                 cookie = req;
  43                 cookielen = sizeof(*req);
  44         } else {
  45                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
  46                                      &req->rq_repmsg);
  47                 if (rc) {
  48                         CERROR("out of memory\n");
  49                         RETURN(-ENOMEM);
  50                 }
  51                 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) {
  52                         cookie = &dlm_req->lock_desc.l_extent;
  53                         cookielen = sizeof(struct ldlm_extent);
  54                 }
  55         }
  56
  57         lock = ldlm_lock_create(obddev->obd_namespace,
  58                                 &dlm_req->lock_handle2,
  59                                 dlm_req->lock_desc.l_resource.lr_name,
  60                                 dlm_req->lock_desc.l_resource.lr_type,
  61                                 dlm_req->lock_desc.l_req_mode, NULL, 0);
  62         if (!lock)
  63                 GOTO(out, -ENOMEM);
  64
  65         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
  66                sizeof(lock->l_remote_handle));
  67         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
  68
  69         flags = dlm_req->lock_flags;
  70         err = ldlm_lock_enqueue(lock, cookie, cookielen, &flags,
  71                                 ldlm_server_ast, ldlm_server_ast);
  72         if (err != ELDLM_OK)
  73                 GOTO(out, err);
  74
  75         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
  76         dlm_rep->lock_flags = flags;
  77
  78         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
  79         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
  80                 memcpy(&dlm_rep->lock_extent, &lock->l_extent,
  81                        sizeof(lock->l_extent));
  82         if (dlm_rep->lock_flags & LDLM_FL_LOCK_CHANGED)
  83                 memcpy(dlm_rep->lock_resource_name, lock->l_resource->lr_name,
  84                        sizeof(dlm_rep->lock_resource_name));
  85
  86         lock->l_connection = ptlrpc_connection_addref(req->rq_connection);
  87         EXIT;
  88  out:
  89         if (lock)
  90                 ldlm_lock_put(lock);
  91         req->rq_status = err;
  92         CDEBUG(D_INFO, "err = %d\n", err);
  93
  94         if (ptlrpc_reply(svc, req))
  95                 LBUG();
  96
  97         if (err)
  98                 LDLM_DEBUG_NOLOCK("server-side enqueue handler END");
  99         else {
 100                 ldlm_reprocess_all(lock->l_resource);
 101                 LDLM_DEBUG(lock, "server-side enqueue handler END");
 102         }
 103
 104         return 0;
 105 }
 106
 107 static int ldlm_handle_convert(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 108 {
 109         struct ldlm_request *dlm_req;
 110         struct ldlm_reply *dlm_rep;
 111         struct ldlm_lock *lock;
 112         int rc, size = sizeof(*dlm_rep);
 113         ENTRY;
 114
 115         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 116         if (rc) {
 117                 CERROR("out of memory\n");
 118                 RETURN(-ENOMEM);
 119         }
 120         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 121         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
 122         dlm_rep->lock_flags = dlm_req->lock_flags;
 123
 124         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 125         if (!lock) {
 126                 req->rq_status = EINVAL;
 127         } else {
 128                 LDLM_DEBUG(lock, "server-side convert handler START");
 129                 ldlm_convert(lock, dlm_req->lock_desc.l_req_mode,
 130                                   &dlm_rep->lock_flags);
 131                 req->rq_status = 0;
 132         }
 133         if (ptlrpc_reply(svc, req) != 0)
 134                 LBUG();
 135
 136         ldlm_reprocess_all(lock->l_resource);
 137         ldlm_lock_put(lock);
 138         LDLM_DEBUG(lock, "server-side convert handler END");
 139
 140         RETURN(0);
 141 }
 142
 143 static int ldlm_handle_cancel(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 144 {
 145         struct ldlm_request *dlm_req;
 146         struct ldlm_lock *lock;
 147         int rc;
 148         ENTRY;
 149
 150         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 151         if (rc) {
 152                 CERROR("out of memory\n");
 153                 RETURN(-ENOMEM);
 154         }
 155         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 156
 157         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 158         if (!lock) {
 159                 req->rq_status = ESTALE;
 160         } else {
 161                 LDLM_DEBUG(lock, "server-side cancel handler START");
 162                 ldlm_lock_cancel(lock);
 163                 req->rq_status = 0;
 164         }
 165
 166         if (ptlrpc_reply(svc, req) != 0)
 167                 LBUG();
 168
 169         ldlm_reprocess_all(lock->l_resource);
 170         ldlm_lock_put(lock);
 171         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
 172
 173         RETURN(0);
 174 }
 175
 176 static int ldlm_handle_callback(struct ptlrpc_service *svc,
 177                           struct ptlrpc_request *req)
 178 {
 179         struct ldlm_request *dlm_req;
 180         struct ldlm_lock_desc *descp;
 181         struct ldlm_lock *lock;
 182         __u64 is_blocking_ast;
 183         int rc;
 184         ENTRY;
 185
 186         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 187         if (rc) {
 188                 CERROR("out of memory\n");
 189                 RETURN(-ENOMEM);
 190         }
 191         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 192         descp = &dlm_req->lock_desc;
 193
 194         /* We must send the reply first, so that the thread is free to handle
 195          * any requests made in common_callback() */
 196         rc = ptlrpc_reply(svc, req);
 197         if (rc != 0)
 198                 RETURN(rc);
 199
 200         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 201         /* check if this is a blocking AST */
 202         if (!descp->l_req_mode)
 203                 descp = NULL;
 204
 205         if (!lock) {
 206                 CERROR("callback on lock %Lx - lock disappeared\n",
 207                        dlm_req->lock_handle1.addr);
 208                 RETURN(0);
 209         }
 210
 211         LDLM_DEBUG(lock, "client %s callback handler START",
 212                    is_blocking_ast ? "blocked" : "completion");
 213
 214         if (descp) {
 215                 int do_ast;
 216                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 217                 lock->l_flags |= LDLM_FL_CBPENDING;
 218                 do_ast = (!lock->l_readers && !lock->l_writers);
 219                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 220
 221                 if (do_ast) {
 222                         CDEBUG(D_INFO, "Lock already unused, calling "
 223                                "callback (%p).\n", lock->l_blocking_ast);
 224                         if (lock->l_blocking_ast != NULL) {
 225                                 struct lustre_handle lockh;
 226                                 ldlm_lock2handle(lock, &lockh);
 227                                 lock->l_blocking_ast(&lockh, descp,
 228                                                      lock->l_data,
 229                                                      lock->l_data_len);
 230                         }
 231                 } else {
 232                         LDLM_DEBUG(lock, "Lock still has references, will be"
 233                                " cancelled later");
 234                 }
 235                 ldlm_lock_put(lock);
 236         } else {
 237                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 238
 239                 /* If we receive the completion AST before the actual enqueue
 240                  * returned, then we might need to switch resources. */
 241                 if (memcmp(dlm_req->lock_desc.l_resource.lr_name,
 242                            lock->l_resource->lr_name,
 243                            sizeof(__u64) * RES_NAME_SIZE) != 0) {
 244                         ldlm_lock_change_resource(lock, dlm_req->lock_desc.l_resource.lr_name);
 245                         LDLM_DEBUG(lock, "completion AST, new resource");
 246                 }
 247                 ldlm_grant_lock(lock);
 248                 /*  FIXME: we want any completion function, not just wake_up */
 249                 wake_up(&lock->l_waitq);
 250                 ldlm_lock_put(lock);
 251         }
 252
 253         LDLM_DEBUG_NOLOCK("client %s callback handler END (lock: %p)",
 254                           is_blocking_ast ? "blocked" : "completion", lock);
 255         RETURN(0);
 256 }
 257
 258 static int lustre_handle(struct obd_device *dev, struct ptlrpc_service *svc,
 259                        struct ptlrpc_request *req)
 260 {
 261         struct obd_device *req_dev;
 262         int id, rc;
 263         ENTRY;
 264
 265         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 266         if (rc) {
 267                 CERROR("lustre_ldlm: Invalid request\n");
 268                 GOTO(out, rc);
 269         }
 270
 271         if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 272                 CERROR("lustre_ldlm: wrong packet type sent %d\n",
 273                        req->rq_reqmsg->type);
 274                 GOTO(out, rc = -EINVAL);
 275         }
 276
 277         id = req->rq_reqmsg->target_id;
 278         if (id < 0 || id > MAX_OBD_DEVICES)
 279                 GOTO(out, rc = -ENODEV);
 280         req_dev = req->rq_obd = &obd_dev[id];
 281
 282         switch (req->rq_reqmsg->opc) {
 283         case LDLM_ENQUEUE:
 284                 CDEBUG(D_INODE, "enqueue\n");
 285                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
 286                 rc = ldlm_handle_enqueue(req_dev, svc, req);
 287                 break;
 288
 289         case LDLM_CONVERT:
 290                 CDEBUG(D_INODE, "convert\n");
 291                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
 292                 rc = ldlm_handle_convert(svc, req);
 293                 break;
 294
 295         case LDLM_CANCEL:
 296                 CDEBUG(D_INODE, "cancel\n");
 297                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
 298                 rc = ldlm_handle_cancel(svc, req);
 299                 break;
 300
 301         case LDLM_CALLBACK:
 302                 CDEBUG(D_INODE, "callback\n");
 303                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CALLBACK, 0);
 304                 rc = ldlm_handle_callback(svc, req);
 305                 break;
 306
 307         default:
 308                 rc = ptlrpc_error(svc, req);
 309                 RETURN(rc);
 310         }
 311
 312         EXIT;
 313 out:
 314         if (rc)
 315                 RETURN(ptlrpc_error(svc, req));
 316         return 0;
 317 }
 318
 319 static int ldlm_iocontrol(long cmd, struct obd_conn *conn, int len, void *karg,
 320                           void *uarg)
 321 {
 322         struct obd_device *obddev = conn->oc_dev;
 323         struct ptlrpc_connection *connection;
 324         int err;
 325         ENTRY;
 326
 327         if (_IOC_TYPE(cmd) != IOC_LDLM_TYPE || _IOC_NR(cmd) < IOC_LDLM_MIN_NR ||
 328             _IOC_NR(cmd) > IOC_LDLM_MAX_NR) {
 329                 CDEBUG(D_IOCTL, "invalid ioctl (type %ld, nr %ld, size %ld)\n",
 330                        _IOC_TYPE(cmd), _IOC_NR(cmd), _IOC_SIZE(cmd));
 331                 RETURN(-EINVAL);
 332         }
 333
 334         OBD_ALLOC(obddev->u.ldlm.ldlm_client,
 335                   sizeof(*obddev->u.ldlm.ldlm_client));
 336         ptlrpc_init_client(NULL, NULL,
 337                            LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
 338                            obddev->u.ldlm.ldlm_client);
 339         connection = ptlrpc_uuid_to_connection("ldlm");
 340         if (!connection)
 341                 CERROR("No LDLM UUID found: assuming ldlm is local.\n");
 342
 343         switch (cmd) {
 344         case IOC_LDLM_TEST: {
 345                 err = ldlm_test(obddev, connection);
 346                 CERROR("-- done err %d\n", err);
 347                 GOTO(out, err);
 348         }
 349         default:
 350                 GOTO(out, err = -EINVAL);
 351         }
 352
 353  out:
 354         if (connection)
 355                 ptlrpc_put_connection(connection);
 356         OBD_FREE(obddev->u.ldlm.ldlm_client,
 357                  sizeof(*obddev->u.ldlm.ldlm_client));
 358         return err;
 359 }
 360
 361 #define LDLM_NUM_THREADS        8
 362
 363 static int ldlm_setup(struct obd_device *obddev, obd_count len, void *buf)
 364 {
 365         struct ldlm_obd *ldlm = &obddev->u.ldlm;
 366         int rc;
 367         int i;
 368         ENTRY;
 369
 370         MOD_INC_USE_COUNT;
 371         ldlm->ldlm_service =
 372                 ptlrpc_init_svc(64 * 1024, LDLM_REQUEST_PORTAL,
 373                                 LDLM_REPLY_PORTAL, "self", lustre_handle);
 374         if (!ldlm->ldlm_service) {
 375                 LBUG();
 376                 GOTO(out_dec, rc = -ENOMEM);
 377         }
 378
 379         for (i = 0; i < LDLM_NUM_THREADS; i++) {
 380                 rc = ptlrpc_start_thread(obddev, ldlm->ldlm_service,
 381                                          "lustre_dlm");
 382                 /* XXX We could just continue if we had started at least
 383                  *     a few threads here.
 384                  */
 385                 if (rc) {
 386                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
 387                         LBUG();
 388                         GOTO(out_thread, rc);
 389                 }
 390         }
 391
 392         RETURN(0);
 393
 394 out_thread:
 395         ptlrpc_stop_all_threads(ldlm->ldlm_service);
 396         ptlrpc_unregister_service(ldlm->ldlm_service);
 397 out_dec:
 398         MOD_DEC_USE_COUNT;
 399         return rc;
 400 }
 401
 402 static int ldlm_cleanup(struct obd_device *obddev)
 403 {
 404         struct ldlm_obd *ldlm = &obddev->u.ldlm;
 405         ENTRY;
 406
 407         ptlrpc_stop_all_threads(ldlm->ldlm_service);
 408         ptlrpc_unregister_service(ldlm->ldlm_service);
 409
 410         if (mds_reint_p != NULL)
 411                 inter_module_put("mds_reint");
 412         if (mds_getattr_name_p != NULL)
 413                 inter_module_put("mds_getattr_name");
 414
 415         MOD_DEC_USE_COUNT;
 416         RETURN(0);
 417 }
 418
 419 struct obd_ops ldlm_obd_ops = {
 420         o_iocontrol:   ldlm_iocontrol,
 421         o_setup:       ldlm_setup,
 422         o_cleanup:     ldlm_cleanup,
 423         o_connect:     gen_connect,
 424         o_disconnect:  gen_disconnect
 425 };
 426
 427
 428 static int __init ldlm_init(void)
 429 {
 430         int rc = obd_register_type(&ldlm_obd_ops, OBD_LDLM_DEVICENAME);
 431         if (rc != 0)
 432                 return rc;
 433
 434         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
 435                                                sizeof(struct ldlm_resource), 0,
 436                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
 437         if (ldlm_resource_slab == NULL)
 438                 return -ENOMEM;
 439
 440         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
 441                                            sizeof(struct ldlm_lock), 0,
 442                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
 443         if (ldlm_lock_slab == NULL) {
 444                 kmem_cache_destroy(ldlm_resource_slab);
 445                 return -ENOMEM;
 446         }
 447
 448         return 0;
 449 }
 450
 451 static void __exit ldlm_exit(void)
 452 {
 453         obd_unregister_type(OBD_LDLM_DEVICENAME);
 454         if (kmem_cache_destroy(ldlm_resource_slab) != 0)
 455                 CERROR("couldn't free ldlm resource slab\n");
 456         if (kmem_cache_destroy(ldlm_lock_slab) != 0)
 457                 CERROR("couldn't free ldlm lock slab\n");
 458 }
 459
 460 EXPORT_SYMBOL(ldlm_lockname);
 461 EXPORT_SYMBOL(ldlm_typename);
 462 EXPORT_SYMBOL(ldlm_handle2lock);
 463 EXPORT_SYMBOL(ldlm_lock_match);
 464 EXPORT_SYMBOL(ldlm_lock_addref);
 465 EXPORT_SYMBOL(ldlm_lock_decref);
 466 EXPORT_SYMBOL(ldlm_cli_convert);
 467 EXPORT_SYMBOL(ldlm_cli_enqueue);
 468 EXPORT_SYMBOL(ldlm_cli_cancel);
 469 EXPORT_SYMBOL(ldlm_test);
 470 EXPORT_SYMBOL(ldlm_lock_dump);
 471 EXPORT_SYMBOL(ldlm_namespace_new);
 472 EXPORT_SYMBOL(ldlm_namespace_free);
 473
 474 MODULE_AUTHOR("Cluster File Systems, Inc. <braam@clusterfs.com>");
 475 MODULE_DESCRIPTION("Lustre Lock Management Module v0.1");
 476 MODULE_LICENSE("GPL");
 477
 478 module_init(ldlm_init);
 479 module_exit(ldlm_exit);