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