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 ptlrpc_request *req)
  25 {
  26         struct obd_device *obddev = req->rq_export->export_obd;
  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, err = -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_DEBUG(lock, "server-side enqueue handler, sending reply");
  91                 ldlm_lock_put(lock);
  92         }
  93         req->rq_status = err;
  94         CDEBUG(D_INFO, "err = %d\n", err);
  95
  96         if (ptlrpc_reply(req->rq_svc, req))
  97                 LBUG();
  98
  99         if (!err)
 100                 ldlm_reprocess_all(lock->l_resource);
 101         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p)", lock);
 102
 103         return 0;
 104 }
 105
 106 static int ldlm_handle_convert(struct ptlrpc_request *req)
 107 {
 108         struct ldlm_request *dlm_req;
 109         struct ldlm_reply *dlm_rep;
 110         struct ldlm_lock *lock;
 111         int rc, size = sizeof(*dlm_rep);
 112         ENTRY;
 113
 114         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 115         if (rc) {
 116                 CERROR("out of memory\n");
 117                 RETURN(-ENOMEM);
 118         }
 119         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 120         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
 121         dlm_rep->lock_flags = dlm_req->lock_flags;
 122
 123         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 124         if (!lock) {
 125                 req->rq_status = EINVAL;
 126         } else {
 127                 LDLM_DEBUG(lock, "server-side convert handler START");
 128                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 129                                   &dlm_rep->lock_flags);
 130                 req->rq_status = 0;
 131         }
 132         if (ptlrpc_reply(req->rq_svc, req) != 0)
 133                 LBUG();
 134
 135         if (lock) {
 136                 ldlm_reprocess_all(lock->l_resource);
 137                 ldlm_lock_put(lock);
 138                 LDLM_DEBUG(lock, "server-side convert handler END");
 139         } else
 140                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 141
 142         RETURN(0);
 143 }
 144
 145 static int ldlm_handle_cancel(struct ptlrpc_request *req)
 146 {
 147         struct ldlm_request *dlm_req;
 148         struct ldlm_lock *lock;
 149         int rc;
 150         char *ns_name;
 151         ENTRY;
 152
 153         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 154         if (rc) {
 155                 CERROR("out of memory\n");
 156                 RETURN(-ENOMEM);
 157         }
 158         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 159
 160         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 161         if (!lock) {
 162                 req->rq_status = ESTALE;
 163         } else {
 164                 LDLM_DEBUG(lock, "server-side cancel handler START");
 165                 ns_name = lock->l_resource->lr_namespace->ns_name;
 166                 ldlm_lock_cancel(lock);
 167                 req->rq_status = 0;
 168         }
 169
 170         if (ptlrpc_reply(req->rq_svc, req) != 0)
 171                 LBUG();
 172
 173         if (lock) {
 174                 ldlm_reprocess_all(lock->l_resource);
 175                 ldlm_lock_put(lock);
 176                 LDLM_DEBUG_NOLOCK("server-side cancel handler END (%s: lock "
 177                                   "%p)", ns_name, lock);
 178         } else
 179                 LDLM_DEBUG_NOLOCK("server-side cancel handler END (lock %p)",
 180                                   lock);
 181
 182         RETURN(0);
 183 }
 184
 185 static int ldlm_handle_callback(struct ptlrpc_request *req)
 186 {
 187         struct ldlm_request *dlm_req;
 188         struct ldlm_lock_desc *descp = NULL;
 189         struct ldlm_lock *lock;
 190         __u64 is_blocking_ast = 0;
 191         int rc;
 192         ENTRY;
 193
 194         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 195         if (rc) {
 196                 CERROR("out of memory\n");
 197                 RETURN(-ENOMEM);
 198         }
 199         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 200
 201         /* We must send the reply first, so that the thread is free to handle
 202          * any requests made in common_callback() */
 203         rc = ptlrpc_reply(req->rq_svc, req);
 204         if (rc != 0)
 205                 RETURN(rc);
 206
 207         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 208         if (!lock) {
 209                 CERROR("callback on lock %Lx - lock disappeared\n",
 210                        dlm_req->lock_handle1.addr);
 211                 RETURN(0);
 212         }
 213
 214         /* check if this is a blocking AST */
 215         if (dlm_req->lock_desc.l_req_mode !=
 216             dlm_req->lock_desc.l_granted_mode) {
 217                 descp = &dlm_req->lock_desc;
 218                 is_blocking_ast = 1;
 219         }
 220
 221         LDLM_DEBUG(lock, "client %s callback handler START",
 222                    is_blocking_ast ? "blocked" : "completion");
 223
 224         if (descp) {
 225                 int do_ast;
 226                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 227                 lock->l_flags |= LDLM_FL_CBPENDING;
 228                 do_ast = (!lock->l_readers && !lock->l_writers);
 229                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 230
 231                 if (do_ast) {
 232                         LDLM_DEBUG(lock, "already unused, calling "
 233                                    "callback (%p)", lock->l_blocking_ast);
 234                         if (lock->l_blocking_ast != NULL) {
 235                                 struct lustre_handle lockh;
 236                                 ldlm_lock2handle(lock, &lockh);
 237                                 lock->l_blocking_ast(&lockh, descp,
 238                                                      lock->l_data,
 239                                                      lock->l_data_len);
 240                         }
 241                 } else {
 242                         LDLM_DEBUG(lock, "Lock still has references, will be"
 243                                " cancelled later");
 244                 }
 245                 ldlm_lock_put(lock);
 246         } else {
 247                 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
 248
 249                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 250                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 251
 252                 /* If we receive the completion AST before the actual enqueue
 253                  * returned, then we might need to switch resources. */
 254                 if (memcmp(dlm_req->lock_desc.l_resource.lr_name,
 255                            lock->l_resource->lr_name,
 256                            sizeof(__u64) * RES_NAME_SIZE) != 0) {
 257                         ldlm_lock_change_resource(lock, dlm_req->lock_desc.l_resource.lr_name);
 258                         LDLM_DEBUG(lock, "completion AST, new resource");
 259                 }
 260                 lock->l_resource->lr_tmp = &rpc_list;
 261                 ldlm_resource_unlink_lock(lock);
 262                 ldlm_grant_lock(lock);
 263                 /*  FIXME: we want any completion function, not just wake_up */
 264                 wake_up(&lock->l_waitq);
 265                 lock->l_resource->lr_tmp = NULL;
 266                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 267                 ldlm_lock_put(lock);
 268
 269                 ldlm_run_ast_work(&rpc_list);
 270         }
 271
 272         LDLM_DEBUG_NOLOCK("client %s callback handler END (lock %p)",
 273                           is_blocking_ast ? "blocked" : "completion", lock);
 274         RETURN(0);
 275 }
 276
 277 static int ldlm_handle(struct ptlrpc_request *req)
 278 {
 279         int rc;
 280         ENTRY;
 281
 282         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 283         if (rc) {
 284                 CERROR("lustre_ldlm: Invalid request\n");
 285                 GOTO(out, rc);
 286         }
 287
 288         if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 289                 CERROR("lustre_ldlm: wrong packet type sent %d\n",
 290                        req->rq_reqmsg->type);
 291                 GOTO(out, rc = -EINVAL);
 292         }
 293
 294         if (!req->rq_export && req->rq_reqmsg->opc == LDLM_ENQUEUE) {
 295                 CERROR("No export handle for enqueue request.\n");
 296                 GOTO(out, rc = -ENOTCONN);
 297         }
 298
 299         switch (req->rq_reqmsg->opc) {
 300         case LDLM_ENQUEUE:
 301                 CDEBUG(D_INODE, "enqueue\n");
 302                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
 303                 rc = ldlm_handle_enqueue(req);
 304                 break;
 305
 306         case LDLM_CONVERT:
 307                 CDEBUG(D_INODE, "convert\n");
 308                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
 309                 rc = ldlm_handle_convert(req);
 310                 break;
 311
 312         case LDLM_CANCEL:
 313                 CDEBUG(D_INODE, "cancel\n");
 314                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
 315                 rc = ldlm_handle_cancel(req);
 316                 break;
 317
 318         case LDLM_CALLBACK:
 319                 CDEBUG(D_INODE, "callback\n");
 320                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CALLBACK, 0);
 321                 rc = ldlm_handle_callback(req);
 322                 break;
 323
 324         default:
 325                 rc = ptlrpc_error(req->rq_svc, req);
 326                 RETURN(rc);
 327         }
 328
 329         EXIT;
 330 out:
 331         if (rc)
 332                 RETURN(ptlrpc_error(req->rq_svc, req));
 333         return 0;
 334 }
 335
 336 static int ldlm_iocontrol(long cmd, struct lustre_handle *conn, int len,
 337                           void *karg, void *uarg)
 338 {
 339         struct obd_device *obddev = class_conn2obd(conn);
 340         struct ptlrpc_connection *connection;
 341         int err;
 342         ENTRY;
 343
 344         if (_IOC_TYPE(cmd) != IOC_LDLM_TYPE || _IOC_NR(cmd) < IOC_LDLM_MIN_NR ||
 345             _IOC_NR(cmd) > IOC_LDLM_MAX_NR) {
 346                 CDEBUG(D_IOCTL, "invalid ioctl (type %ld, nr %ld, size %ld)\n",
 347                        _IOC_TYPE(cmd), _IOC_NR(cmd), _IOC_SIZE(cmd));
 348                 RETURN(-EINVAL);
 349         }
 350
 351         OBD_ALLOC(obddev->u.ldlm.ldlm_client,
 352                   sizeof(*obddev->u.ldlm.ldlm_client));
 353         ptlrpc_init_client(NULL, NULL,
 354                            LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
 355                            obddev->u.ldlm.ldlm_client);
 356         connection = ptlrpc_uuid_to_connection("ldlm");
 357         if (!connection)
 358                 CERROR("No LDLM UUID found: assuming ldlm is local.\n");
 359
 360         switch (cmd) {
 361         case IOC_LDLM_TEST: {
 362                 err = ldlm_test(obddev, connection);
 363                 CERROR("-- done err %d\n", err);
 364                 GOTO(out, err);
 365         }
 366         default:
 367                 GOTO(out, err = -EINVAL);
 368         }
 369
 370  out:
 371         if (connection)
 372                 ptlrpc_put_connection(connection);
 373         OBD_FREE(obddev->u.ldlm.ldlm_client,
 374                  sizeof(*obddev->u.ldlm.ldlm_client));
 375         return err;
 376 }
 377
 378 #define LDLM_NUM_THREADS        8
 379
 380 static int ldlm_setup(struct obd_device *obddev, obd_count len, void *buf)
 381 {
 382         struct ldlm_obd *ldlm = &obddev->u.ldlm;
 383         int rc;
 384         int i;
 385         ENTRY;
 386
 387         MOD_INC_USE_COUNT;
 388         ldlm->ldlm_service =
 389                 ptlrpc_init_svc(64 * 1024, LDLM_REQUEST_PORTAL,
 390                                 LDLM_REPLY_PORTAL, "self", ldlm_handle);
 391         if (!ldlm->ldlm_service) {
 392                 LBUG();
 393                 GOTO(out_dec, rc = -ENOMEM);
 394         }
 395
 396         for (i = 0; i < LDLM_NUM_THREADS; i++) {
 397                 rc = ptlrpc_start_thread(obddev, ldlm->ldlm_service,
 398                                          "lustre_dlm");
 399                 if (rc) {
 400                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
 401                         LBUG();
 402                         GOTO(out_thread, rc);
 403                 }
 404         }
 405
 406         RETURN(0);
 407
 408 out_thread:
 409         ptlrpc_stop_all_threads(ldlm->ldlm_service);
 410         ptlrpc_unregister_service(ldlm->ldlm_service);
 411 out_dec:
 412         MOD_DEC_USE_COUNT;
 413         return rc;
 414 }
 415
 416 static int ldlm_cleanup(struct obd_device *obddev)
 417 {
 418         struct ldlm_obd *ldlm = &obddev->u.ldlm;
 419         ENTRY;
 420
 421         ptlrpc_stop_all_threads(ldlm->ldlm_service);
 422         ptlrpc_unregister_service(ldlm->ldlm_service);
 423
 424         if (mds_reint_p != NULL)
 425                 inter_module_put("mds_reint");
 426         if (mds_getattr_name_p != NULL)
 427                 inter_module_put("mds_getattr_name");
 428
 429         MOD_DEC_USE_COUNT;
 430         RETURN(0);
 431 }
 432
 433 struct obd_ops ldlm_obd_ops = {
 434         o_iocontrol:   ldlm_iocontrol,
 435         o_setup:       ldlm_setup,
 436         o_cleanup:     ldlm_cleanup,
 437         o_connect:     class_connect,
 438         o_disconnect:  class_disconnect
 439 };
 440
 441
 442 static int __init ldlm_init(void)
 443 {
 444         int rc = class_register_type(&ldlm_obd_ops, OBD_LDLM_DEVICENAME);
 445         if (rc != 0)
 446                 return rc;
 447
 448         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
 449                                                sizeof(struct ldlm_resource), 0,
 450                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
 451         if (ldlm_resource_slab == NULL)
 452                 return -ENOMEM;
 453
 454         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
 455                                            sizeof(struct ldlm_lock), 0,
 456                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
 457         if (ldlm_lock_slab == NULL) {
 458                 kmem_cache_destroy(ldlm_resource_slab);
 459                 return -ENOMEM;
 460         }
 461
 462         return 0;
 463 }
 464
 465 static void __exit ldlm_exit(void)
 466 {
 467         class_unregister_type(OBD_LDLM_DEVICENAME);
 468         if (kmem_cache_destroy(ldlm_resource_slab) != 0)
 469                 CERROR("couldn't free ldlm resource slab\n");
 470         if (kmem_cache_destroy(ldlm_lock_slab) != 0)
 471                 CERROR("couldn't free ldlm lock slab\n");
 472 }
 473
 474 EXPORT_SYMBOL(ldlm_lockname);
 475 EXPORT_SYMBOL(ldlm_typename);
 476 EXPORT_SYMBOL(ldlm_handle2lock);
 477 EXPORT_SYMBOL(ldlm_lock_match);
 478 EXPORT_SYMBOL(ldlm_lock_addref);
 479 EXPORT_SYMBOL(ldlm_lock_decref);
 480 EXPORT_SYMBOL(ldlm_cli_convert);
 481 EXPORT_SYMBOL(ldlm_cli_enqueue);
 482 EXPORT_SYMBOL(ldlm_cli_cancel);
 483 EXPORT_SYMBOL(ldlm_test);
 484 EXPORT_SYMBOL(ldlm_lock_dump);
 485 EXPORT_SYMBOL(ldlm_namespace_new);
 486 EXPORT_SYMBOL(ldlm_namespace_free);
 487
 488 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
 489 MODULE_DESCRIPTION("Lustre Lock Management Module v0.1");
 490 MODULE_LICENSE("GPL");
 491
 492 module_init(ldlm_init);
 493 module_exit(ldlm_exit);