lustre/ldlm/ldlm_lib.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #ifndef EXPORT_SYMTAB
  23 # define EXPORT_SYMTAB
  24 #endif
  25 #define DEBUG_SUBSYSTEM S_LDLM
  26
  27 #ifdef __KERNEL__
  28 # include <linux/module.h>
  29 #else
  30 # include <liblustre.h>
  31 #endif
  32 #include <linux/obd.h>
  33 #include <linux/obd_ost.h> /* for LUSTRE_OSC_NAME */
  34 #include <linux/lustre_mds.h> /* for LUSTRE_MDC_NAME */
  35 #include <linux/lustre_mgmt.h>
  36 #include <linux/lustre_dlm.h>
  37 #include <linux/lustre_net.h>
  38
  39 int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
  40 {
  41         struct ptlrpc_connection *conn;
  42         struct lustre_cfg* lcfg = buf;
  43         struct client_obd *cli = &obddev->u.cli;
  44         struct obd_import *imp;
  45         struct obd_uuid server_uuid;
  46         int rq_portal, rp_portal, connect_op;
  47         char *name = obddev->obd_type->typ_name;
  48         char *mgmt_name = NULL;
  49         int rc;
  50         struct obd_device *mgmt_obd;
  51         mgmtcli_register_for_events_t register_f;
  52         ENTRY;
  53
  54         /* In a more perfect world, we would hang a ptlrpc_client off of
  55          * obd_type and just use the values from there. */
  56         if (!strcmp(name, LUSTRE_OSC_NAME)) {
  57                 rq_portal = OST_REQUEST_PORTAL;
  58                 rp_portal = OSC_REPLY_PORTAL;
  59                 connect_op = OST_CONNECT;
  60         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
  61                 rq_portal = MDS_REQUEST_PORTAL;
  62                 rp_portal = MDC_REPLY_PORTAL;
  63                 connect_op = MDS_CONNECT;
  64         } else if (!strcmp(name, LUSTRE_MGMTCLI_NAME)) {
  65                 rq_portal = MGMT_REQUEST_PORTAL;
  66                 rp_portal = MGMT_REPLY_PORTAL;
  67                 connect_op = MGMT_CONNECT;
  68         } else {
  69                 CERROR("unknown client OBD type \"%s\", can't setup\n",
  70                        name);
  71                 RETURN(-EINVAL);
  72         }
  73
  74         if (lcfg->lcfg_inllen1 < 1) {
  75                 CERROR("requires a TARGET UUID\n");
  76                 RETURN(-EINVAL);
  77         }
  78
  79         if (lcfg->lcfg_inllen1 > 37) {
  80                 CERROR("client UUID must be less than 38 characters\n");
  81                 RETURN(-EINVAL);
  82         }
  83
  84         if (lcfg->lcfg_inllen2 < 1) {
  85                 CERROR("setup requires a SERVER UUID\n");
  86                 RETURN(-EINVAL);
  87         }
  88
  89         if (lcfg->lcfg_inllen2 > 37) {
  90                 CERROR("target UUID must be less than 38 characters\n");
  91                 RETURN(-EINVAL);
  92         }
  93
  94         sema_init(&cli->cl_sem, 1);
  95         cli->cl_conn_count = 0;
  96         memcpy(server_uuid.uuid, lcfg->lcfg_inlbuf2,
  97                min_t(unsigned int, lcfg->lcfg_inllen2, sizeof(server_uuid)));
  98
  99         cli->cl_dirty = 0;
 100         cli->cl_avail_grant = 0;
 101         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 102         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 103         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 104         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 105         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 106         spin_lock_init(&cli->cl_loi_list_lock);
 107         cli->cl_brw_in_flight = 0;
 108         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 109         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 110         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 111         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 112         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 113         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 114
 115         rc = ldlm_get_ref();
 116         if (rc) {
 117                 CERROR("ldlm_get_ref failed: %d\n", rc);
 118                 GOTO(err, rc);
 119         }
 120
 121         conn = ptlrpc_uuid_to_connection(&server_uuid);
 122         if (conn == NULL)
 123                 GOTO(err_ldlm, rc = -ENOENT);
 124
 125         ptlrpc_init_client(rq_portal, rp_portal, name,
 126                            &obddev->obd_ldlm_client);
 127
 128         imp = class_new_import();
 129         if (imp == NULL) {
 130                 ptlrpc_put_connection(conn);
 131                 GOTO(err_ldlm, rc = -ENOENT);
 132         }
 133         imp->imp_connection = conn;
 134         imp->imp_client = &obddev->obd_ldlm_client;
 135         imp->imp_obd = obddev;
 136         imp->imp_connect_op = connect_op;
 137         imp->imp_generation = 0;
 138         imp->imp_initial_recov = 1;
 139         INIT_LIST_HEAD(&imp->imp_pinger_chain);
 140         memcpy(imp->imp_target_uuid.uuid, lcfg->lcfg_inlbuf1,
 141               lcfg->lcfg_inllen1);
 142         class_import_put(imp);
 143
 144         cli->cl_import = imp;
 145         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 146         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 147         cli->cl_sandev = to_kdev_t(0);
 148
 149         if (lcfg->lcfg_inllen3 != 0) {
 150                 if (!strcmp(lcfg->lcfg_inlbuf3, "inactive")) {
 151                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 152                                name, obddev->obd_name,
 153                                imp->imp_target_uuid.uuid);
 154                         imp->imp_invalid = 1;
 155
 156                         if (lcfg->lcfg_inllen4 != 0)
 157                                 mgmt_name = lcfg->lcfg_inlbuf4;
 158                 } else {
 159                         mgmt_name = lcfg->lcfg_inlbuf3;
 160                 }
 161         }
 162
 163         if (mgmt_name != NULL) {
 164                 /* Register with management client if we need to. */
 165                 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
 166                        obddev->obd_name, mgmt_name, server_uuid.uuid);
 167
 168                 mgmt_obd = class_name2obd(mgmt_name);
 169                 if (!mgmt_obd) {
 170                         CERROR("can't find mgmtcli %s to register\n",
 171                                mgmt_name);
 172                         GOTO(err_import, rc = -ENOSYS);
 173                 }
 174
 175                 register_f = inter_module_get("mgmtcli_register_for_events");
 176                 if (!register_f) {
 177                         CERROR("can't i_m_g mgmtcli_register_for_events\n");
 178                         GOTO(err_import, rc = -ENOSYS);
 179                 }
 180
 181                 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
 182                 inter_module_put("mgmtcli_register_for_events");
 183
 184                 if (!rc)
 185                         cli->cl_mgmtcli_obd = mgmt_obd;
 186         }
 187
 188         RETURN(rc);
 189
 190 err_import:
 191         class_destroy_import(imp);
 192 err_ldlm:
 193         ldlm_put_ref(0);
 194 err:
 195         RETURN(rc);
 196
 197 }
 198
 199 int client_obd_cleanup(struct obd_device *obddev, int flags)
 200 {
 201         struct client_obd *cli = &obddev->u.cli;
 202
 203         if (!cli->cl_import)
 204                 RETURN(-EINVAL);
 205         if (cli->cl_mgmtcli_obd) {
 206                 mgmtcli_deregister_for_events_t dereg_f;
 207
 208                 dereg_f = inter_module_get("mgmtcli_deregister_for_events");
 209                 dereg_f(cli->cl_mgmtcli_obd, obddev);
 210                 inter_module_put("mgmtcli_deregister_for_events");
 211         }
 212         class_destroy_import(cli->cl_import);
 213         cli->cl_import = NULL;
 214
 215         ldlm_put_ref(flags & OBD_OPT_FORCE);
 216
 217         RETURN(0);
 218 }
 219
 220 int client_connect_import(struct lustre_handle *dlm_handle,
 221                           struct obd_device *obd,
 222                           struct obd_uuid *cluuid)
 223 {
 224         struct client_obd *cli = &obd->u.cli;
 225         struct obd_import *imp = cli->cl_import;
 226         struct obd_export *exp;
 227         int rc;
 228         ENTRY;
 229
 230         down(&cli->cl_sem);
 231         rc = class_connect(dlm_handle, obd, cluuid);
 232         if (rc)
 233                 GOTO(out_sem, rc);
 234
 235         cli->cl_conn_count++;
 236         if (cli->cl_conn_count > 1)
 237                 GOTO(out_sem, rc);
 238         exp = class_conn2export(dlm_handle);
 239
 240         if (obd->obd_namespace != NULL)
 241                 CERROR("already have namespace!\n");
 242         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 243                                                 LDLM_NAMESPACE_CLIENT);
 244         if (obd->obd_namespace == NULL)
 245                 GOTO(out_disco, rc = -ENOMEM);
 246
 247         imp->imp_dlm_handle = *dlm_handle;
 248         rc = ptlrpc_init_import(imp);
 249         if (rc != 0)
 250                 GOTO(out_ldlm, rc);
 251
 252         exp->exp_connection = ptlrpc_connection_addref(imp->imp_connection);
 253         rc = ptlrpc_connect_import(imp, NULL);
 254         if (rc != 0) {
 255                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 256                 GOTO(out_ldlm, rc);
 257         }
 258
 259         ptlrpc_pinger_add_import(imp);
 260         EXIT;
 261
 262         if (rc) {
 263 out_ldlm:
 264                 ldlm_namespace_free(obd->obd_namespace, 0);
 265                 obd->obd_namespace = NULL;
 266 out_disco:
 267                 cli->cl_conn_count--;
 268                 class_disconnect(exp, 0);
 269         } else {
 270                 class_export_put(exp);
 271         }
 272 out_sem:
 273         up(&cli->cl_sem);
 274         return rc;
 275 }
 276
 277 int client_disconnect_export(struct obd_export *exp, int failover)
 278 {
 279         struct obd_device *obd = class_exp2obd(exp);
 280         struct client_obd *cli = &obd->u.cli;
 281         struct obd_import *imp = cli->cl_import;
 282         int rc = 0, err;
 283         ENTRY;
 284
 285         if (!obd) {
 286                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 287                        exp, exp ? exp->exp_handle.h_cookie : -1);
 288                 RETURN(-EINVAL);
 289         }
 290
 291         down(&cli->cl_sem);
 292         if (!cli->cl_conn_count) {
 293                 CERROR("disconnecting disconnected device (%s)\n",
 294                        obd->obd_name);
 295                 GOTO(out_sem, rc = -EINVAL);
 296         }
 297
 298         cli->cl_conn_count--;
 299         if (cli->cl_conn_count)
 300                 GOTO(out_no_disconnect, rc = 0);
 301
 302         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 303          * delete it regardless.  (It's safe to delete an import that was
 304          * never added.) */
 305         (void)ptlrpc_pinger_del_import(imp);
 306
 307         if (obd->obd_namespace != NULL) {
 308                 /* obd_no_recov == local only */
 309                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 310                                        obd->obd_no_recov, NULL);
 311                 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
 312                 obd->obd_namespace = NULL;
 313         }
 314
 315         /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
 316         if (obd->obd_no_recov)
 317                 ptlrpc_invalidate_import(imp, 0);
 318         else
 319                 rc = ptlrpc_disconnect_import(imp);
 320
 321         EXIT;
 322  out_no_disconnect:
 323         err = class_disconnect(exp, 0);
 324         if (!rc && err)
 325                 rc = err;
 326  out_sem:
 327         up(&cli->cl_sem);
 328         RETURN(rc);
 329 }
 330
 331 /* --------------------------------------------------------------------------
 332  * from old lib/target.c
 333  * -------------------------------------------------------------------------- */
 334
 335 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 336                             struct obd_uuid *cluuid, int initial_conn)
 337 {
 338         if (exp->exp_connection && !initial_conn) {
 339                 struct lustre_handle *hdl;
 340                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 341                 /* Might be a re-connect after a partition. */
 342                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 343                         CERROR("%s reconnecting\n", cluuid->uuid);
 344                         conn->cookie = exp->exp_handle.h_cookie;
 345                         RETURN(EALREADY);
 346                 } else {
 347                         CERROR("%s reconnecting from %s, "
 348                                "handle mismatch (ours "LPX64", theirs "
 349                                LPX64")\n", cluuid->uuid,
 350                                exp->exp_connection->c_remote_uuid.uuid,
 351                                hdl->cookie, conn->cookie);
 352                         memset(conn, 0, sizeof *conn);
 353                         RETURN(-EALREADY);
 354                 }
 355         }
 356
 357         conn->cookie = exp->exp_handle.h_cookie;
 358         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 359                cluuid->uuid, exp);
 360         CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
 361         RETURN(0);
 362 }
 363
 364 int target_handle_connect(struct ptlrpc_request *req)
 365 {
 366         struct obd_device *target;
 367         struct obd_export *export = NULL;
 368         struct obd_import *revimp;
 369         struct lustre_handle conn;
 370         struct obd_uuid tgtuuid;
 371         struct obd_uuid cluuid;
 372         struct obd_uuid remote_uuid;
 373         struct list_head *p;
 374         char *str, *tmp;
 375         int rc = 0;
 376         unsigned long flags;
 377         int initial_conn = 0;
 378         char peer_str[PTL_NALFMT_SIZE];
 379         ENTRY;
 380
 381         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 382
 383         LASSERT_REQSWAB (req, 0);
 384         str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
 385         if (str == NULL) {
 386                 CERROR("bad target UUID for connect\n");
 387                 GOTO(out, rc = -EINVAL);
 388         }
 389
 390         obd_str2uuid (&tgtuuid, str);
 391         target = class_uuid2obd(&tgtuuid);
 392         if (!target) {
 393                 target = class_name2obd(str);
 394         }
 395
 396         if (!target || target->obd_stopping || !target->obd_set_up) {
 397                 CERROR("UUID '%s' is not available for connect\n", str);
 398
 399                 GOTO(out, rc = -ENODEV);
 400         }
 401
 402         LASSERT_REQSWAB (req, 1);
 403         str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
 404         if (str == NULL) {
 405                 CERROR("bad client UUID for connect\n");
 406                 GOTO(out, rc = -EINVAL);
 407         }
 408
 409         obd_str2uuid (&cluuid, str);
 410
 411         /* XXX extract a nettype and format accordingly */
 412         switch (sizeof(ptl_nid_t)) {
 413                 /* NB the casts only avoid compiler warnings */
 414         case 8:
 415                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 416                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.peer_nid);
 417                 break;
 418         case 4:
 419                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 420                          "NET_%x_UUID", (__u32)req->rq_peer.peer_nid);
 421                 break;
 422         default:
 423                 LBUG();
 424         }
 425
 426         tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
 427         if (tmp == NULL)
 428                 GOTO(out, rc = -EPROTO);
 429
 430         memcpy(&conn, tmp, sizeof conn);
 431
 432         rc = lustre_pack_reply(req, 0, NULL, NULL);
 433         if (rc)
 434                 GOTO(out, rc);
 435
 436         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
 437                 initial_conn = 1;
 438
 439         /* lctl gets a backstage, all-access pass. */
 440         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 441                 goto dont_check_exports;
 442
 443         spin_lock(&target->obd_dev_lock);
 444         list_for_each(p, &target->obd_exports) {
 445                 export = list_entry(p, struct obd_export, exp_obd_chain);
 446                 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
 447                         spin_unlock(&target->obd_dev_lock);
 448                         LASSERT(export->exp_obd == target);
 449
 450                         rc = target_handle_reconnect(&conn, export, &cluuid,
 451                                                      initial_conn);
 452                         break;
 453                 }
 454                 export = NULL;
 455         }
 456         /* If we found an export, we already unlocked. */
 457         if (!export) {
 458                 spin_unlock(&target->obd_dev_lock);
 459         } else if (req->rq_export == NULL &&
 460                    atomic_read(&export->exp_rpc_count) > 0) {
 461                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 462                       target->obd_name, cluuid.uuid,
 463                       ptlrpc_peernid2str(&req->rq_peer, peer_str),
 464                       export, atomic_read(&export->exp_refcount));
 465                 GOTO(out, rc = -EBUSY);
 466         } else if (req->rq_export != NULL &&
 467                    atomic_read(&export->exp_rpc_count) > 1) {
 468                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 469                       target->obd_name, cluuid.uuid,
 470                       ptlrpc_peernid2str(&req->rq_peer, peer_str),
 471                       export, atomic_read(&export->exp_rpc_count));
 472                 GOTO(out, rc = -EBUSY);
 473         }
 474         else if (req->rq_reqmsg->conn_cnt == 1 && !initial_conn) {
 475                 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
 476                        cluuid.uuid);
 477                 GOTO(out, rc = -EALREADY);
 478         }
 479
 480         /* Tell the client if we're in recovery. */
 481         /* If this is the first client, start the recovery timer */
 482         CWARN("%s: connection from %s@%s %s\n", target->obd_name, cluuid.uuid,
 483               ptlrpc_peernid2str(&req->rq_peer, peer_str),
 484               target->obd_recovering ? "(recovering)" : "");
 485         if (target->obd_recovering) {
 486                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 487                 target_start_recovery_timer(target);
 488         }
 489 #if 0
 490         /* Tell the client if we support replayable requests */
 491         if (target->obd_replayable)
 492                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 493 #endif
 494         if (export == NULL) {
 495                 if (target->obd_recovering) {
 496                         CERROR("denying connection for new client %s@%s: "
 497                                "%d clients in recovery for %lds\n", cluuid.uuid,
 498                                ptlrpc_peernid2str(&req->rq_peer, peer_str),
 499                                target->obd_recoverable_clients,
 500                                (target->obd_recovery_timer.expires-jiffies)/HZ);
 501                         rc = -EBUSY;
 502                 } else {
 503  dont_check_exports:
 504                         rc = obd_connect(&conn, target, &cluuid);
 505                 }
 506         }
 507         /* Tell the client if we support replayable requests */
 508         if (target->obd_replayable)
 509                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 510
 511         /* If all else goes well, this is our RPC return code. */
 512         req->rq_status = 0;
 513
 514         if (rc && rc != EALREADY)
 515                 GOTO(out, rc);
 516
 517         req->rq_repmsg->handle = conn;
 518
 519         /* If the client and the server are the same node, we will already
 520          * have an export that really points to the client's DLM export,
 521          * because we have a shared handles table.
 522          *
 523          * XXX this will go away when shaver stops sending the "connect" handle
 524          * in the real "remote handle" field of the request --phik 24 Apr 2003
 525          */
 526         if (req->rq_export != NULL)
 527                 class_export_put(req->rq_export);
 528
 529         /* ownership of this export ref transfers to the request */
 530         export = req->rq_export = class_conn2export(&conn);
 531         LASSERT(export != NULL);
 532
 533         spin_lock_irqsave(&export->exp_lock, flags);
 534         if (initial_conn) {
 535                 req->rq_repmsg->conn_cnt = export->exp_conn_cnt + 1;
 536         } else if (export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
 537                 CERROR("%s@%s: already connected at a higher conn_cnt: %d > %d\n",
 538                        cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str),
 539                        export->exp_conn_cnt,
 540                        req->rq_reqmsg->conn_cnt);
 541                 spin_unlock_irqrestore(&export->exp_lock, flags);
 542                 GOTO(out, rc = -EALREADY);
 543         }
 544         export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
 545         spin_unlock_irqrestore(&export->exp_lock, flags);
 546
 547         /* request from liblustre? */
 548         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
 549                 export->exp_libclient = 1;
 550
 551         if (export->exp_connection != NULL)
 552                 ptlrpc_put_connection(export->exp_connection);
 553         export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
 554                                                        &remote_uuid);
 555
 556         if (rc == EALREADY) {
 557                 /* We indicate the reconnection in a flag, not an error code. */
 558                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 559                 GOTO(out, rc = 0);
 560         }
 561
 562         if (target->obd_recovering) {
 563                 target->obd_connected_clients++;
 564         }
 565
 566         memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
 567                sizeof conn);
 568
 569         if (export->exp_imp_reverse != NULL)
 570                 class_destroy_import(export->exp_imp_reverse);
 571         revimp = export->exp_imp_reverse = class_new_import();
 572         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 573         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 574         revimp->imp_remote_handle = conn;
 575         revimp->imp_obd = target;
 576         revimp->imp_dlm_fake = 1;
 577         revimp->imp_state = LUSTRE_IMP_FULL;
 578         class_import_put(revimp);
 579 out:
 580         if (rc)
 581                 req->rq_status = rc;
 582         RETURN(rc);
 583 }
 584
 585 int target_handle_disconnect(struct ptlrpc_request *req)
 586 {
 587         struct obd_export *exp;
 588         int rc;
 589         ENTRY;
 590
 591         rc = lustre_pack_reply(req, 0, NULL, NULL);
 592         if (rc)
 593                 RETURN(rc);
 594
 595         /* keep the rq_export around so we can send the reply */
 596         exp = class_export_get(req->rq_export);
 597         req->rq_status = obd_disconnect(exp, 0);
 598         RETURN(0);
 599 }
 600
 601 void target_destroy_export(struct obd_export *exp)
 602 {
 603         /* exports created from last_rcvd data, and "fake"
 604            exports created by lctl don't have an import */
 605         if (exp->exp_imp_reverse != NULL)
 606                 class_destroy_import(exp->exp_imp_reverse);
 607
 608         /* We cancel locks at disconnect time, but this will catch any locks
 609          * granted in a race with recovery-induced disconnect. */
 610         if (exp->exp_obd->obd_namespace != NULL)
 611                 ldlm_cancel_locks_for_export(exp);
 612 }
 613
 614 /*
 615  * Recovery functions
 616  */
 617
 618 struct ptlrpc_request *
 619 ptlrpc_clone_req( struct ptlrpc_request *orig_req)
 620 {
 621         struct ptlrpc_request *copy_req;
 622         struct lustre_msg *copy_reqmsg;
 623
 624         OBD_ALLOC(copy_req, sizeof *copy_req);
 625         if (!copy_req)
 626                 return NULL;
 627         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
 628         if (!copy_reqmsg){
 629                 OBD_FREE(copy_req, sizeof *copy_req);
 630                 return NULL;
 631         }
 632
 633         memcpy(copy_req, orig_req, sizeof *copy_req);
 634         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
 635         /* the copied req takes over the reply state */
 636         orig_req->rq_reply_state = NULL;
 637
 638         copy_req->rq_reqmsg = copy_reqmsg;
 639         class_export_get(copy_req->rq_export);
 640         INIT_LIST_HEAD(&copy_req->rq_list);
 641
 642         return copy_req;
 643 }
 644 void ptlrpc_free_clone( struct ptlrpc_request *req)
 645 {
 646         class_export_put(req->rq_export);
 647         list_del(&req->rq_list);
 648         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 649         OBD_FREE(req, sizeof *req);
 650 }
 651
 652 static void abort_delayed_replies(struct obd_device *obd)
 653 {
 654         struct ptlrpc_request *req;
 655         struct list_head *tmp, *n;
 656         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 657                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 658                 DEBUG_REQ(D_ERROR, req, "aborted:");
 659                 req->rq_status = -ENOTCONN;
 660                 req->rq_type = PTL_RPC_MSG_ERR;
 661                 ptlrpc_reply(req);
 662                 class_export_put(req->rq_export);
 663                 list_del(&req->rq_list);
 664                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 665                 OBD_FREE(req, sizeof *req);
 666         }
 667 }
 668
 669 static void abort_recovery_queue(struct obd_device *obd)
 670 {
 671         struct ptlrpc_request *req;
 672         struct list_head *tmp, *n;
 673         int rc;
 674
 675         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
 676                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 677                 DEBUG_REQ(D_ERROR, req, "aborted:");
 678                 req->rq_status = -ENOTCONN;
 679                 req->rq_type = PTL_RPC_MSG_ERR;
 680                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 681                 if (rc == 0) {
 682                         ptlrpc_reply(req);
 683                 } else {
 684                         DEBUG_REQ(D_ERROR, req,
 685                                   "packing failed for abort-reply; skipping");
 686                 }
 687                 list_del(&req->rq_list);
 688                 class_export_put(req->rq_export);
 689                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 690                 OBD_FREE(req, sizeof *req);
 691         }
 692 }
 693
 694 static void target_abort_recovery(void *data)
 695 {
 696         struct obd_device *obd = data;
 697         int rc;
 698
 699         CERROR("disconnecting clients and aborting recovery\n");
 700         LASSERT(!obd->obd_recovering);
 701
 702         class_disconnect_exports(obd, 0);
 703
 704         /* when recovery was aborted, cleanup orphans on mds and ost */
 705         if (OBT(obd) && OBP(obd, postrecov)) {
 706                 rc = OBP(obd, postrecov)(obd);
 707                 if (rc >= 0)
 708                         CWARN("Cleanup %d orphans after recovery was aborted\n",
 709                               rc);
 710                 else
 711                         CERROR("postrecov failed %d\n", rc);
 712         }
 713
 714         abort_delayed_replies(obd);
 715         abort_recovery_queue(obd);
 716         ptlrpc_run_recovery_over_upcall(obd);
 717 }
 718
 719 static void target_recovery_expired(unsigned long castmeharder)
 720 {
 721         struct obd_device *obd = (struct obd_device *)castmeharder;
 722         CERROR("recovery timed out, aborting\n");
 723         spin_lock_bh(&obd->obd_processing_task_lock);
 724         obd->obd_abort_recovery = 1;
 725         wake_up(&obd->obd_next_transno_waitq);
 726         spin_unlock_bh(&obd->obd_processing_task_lock);
 727 }
 728
 729
 730 /* obd_processing_task_lock should be held */
 731 void target_cancel_recovery_timer(struct obd_device *obd)
 732 {
 733         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
 734         del_timer(&obd->obd_recovery_timer);
 735 }
 736
 737 static void reset_recovery_timer(struct obd_device *obd)
 738 {
 739         spin_lock_bh(&obd->obd_processing_task_lock);
 740         if (!obd->obd_recovering) {
 741                 spin_unlock_bh(&obd->obd_processing_task_lock);
 742                 return;
 743         }
 744         CDEBUG(D_HA, "timer will expire in %u seconds\n",
 745                OBD_RECOVERY_TIMEOUT / HZ);
 746         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
 747         spin_unlock_bh(&obd->obd_processing_task_lock);
 748 }
 749
 750
 751 /* Only start it the first time called */
 752 void target_start_recovery_timer(struct obd_device *obd)
 753 {
 754         spin_lock_bh(&obd->obd_processing_task_lock);
 755         if (!obd->obd_recovering || timer_pending(&obd->obd_recovery_timer)) {
 756                 spin_unlock_bh(&obd->obd_processing_task_lock);
 757                 return;
 758         }
 759         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
 760                OBD_RECOVERY_TIMEOUT / HZ);
 761         obd->obd_recovery_timer.function = target_recovery_expired;
 762         obd->obd_recovery_timer.data = (unsigned long)obd;
 763         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
 764         spin_unlock_bh(&obd->obd_processing_task_lock);
 765 }
 766
 767 static void target_finish_recovery(struct obd_device *obd)
 768 {
 769         struct list_head *tmp, *n;
 770         int rc2;
 771
 772         ldlm_reprocess_all_ns(obd->obd_namespace);
 773
 774         CWARN("%s: all clients recovered, calling postrecov\n",
 775               obd->obd_name);
 776         /* when recovery finished, cleanup orphans on mds and ost */
 777         if (OBT(obd) && OBP(obd, postrecov)) {
 778                 rc2 = OBP(obd, postrecov)(obd);
 779                 if (rc2 >= 0)
 780                         CWARN("%s: all clients recovered, %d MDS "
 781                               "orphans deleted\n", obd->obd_name, rc2);
 782                 else
 783                         CERROR("postrecov failed %d\n", rc2);
 784         }
 785
 786         CWARN("%s: recovery over, sending delayed replies\n",
 787               obd->obd_name);
 788         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 789                 struct ptlrpc_request *req;
 790                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 791                 DEBUG_REQ(D_ERROR, req, "delayed:");
 792                 ptlrpc_reply(req);
 793                 ptlrpc_free_clone(req);
 794         }
 795         ptlrpc_run_recovery_over_upcall(obd);
 796 }
 797
 798 static int check_for_next_transno(struct obd_device *obd)
 799 {
 800         struct ptlrpc_request *req;
 801         int wake_up = 0, connected, completed, queue_len, max;
 802         __u64 next_transno, req_transno;
 803
 804         spin_lock_bh(&obd->obd_processing_task_lock);
 805         if (!list_empty(&obd->obd_recovery_queue)) {
 806                 req = list_entry(obd->obd_recovery_queue.next,
 807                                  struct ptlrpc_request, rq_list);
 808                 req_transno = req->rq_reqmsg->transno;
 809         } else {
 810                 req_transno = 0;
 811         }
 812
 813         max = obd->obd_max_recoverable_clients;
 814         connected = obd->obd_connected_clients;
 815         completed = max - obd->obd_recoverable_clients;
 816         queue_len = obd->obd_requests_queued_for_recovery;
 817         next_transno = obd->obd_next_recovery_transno;
 818
 819         CDEBUG(D_HA,"max: %d, connected: %d, completed: %d, queue_len: %d, "
 820                "req_transno: "LPU64", next_transno: "LPU64"\n",
 821                max, connected, completed, queue_len, req_transno, next_transno);
 822         if (obd->obd_abort_recovery) {
 823                 CDEBUG(D_HA, "waking for aborted recovery\n");
 824                 wake_up = 1;
 825         } else if (max == completed) {
 826                 CDEBUG(D_HA, "waking for completed recovery\n");
 827                 wake_up = 1;
 828         } else if (req_transno == next_transno) {
 829                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
 830                 wake_up = 1;
 831         } else if (queue_len + completed == max) {
 832                 LASSERT(req->rq_reqmsg->transno >= next_transno);
 833                 CDEBUG(D_ERROR,
 834                        "waking for skipped transno (skip: "LPD64
 835                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
 836                        next_transno, queue_len, completed, max, req_transno);
 837                 obd->obd_next_recovery_transno = req_transno;
 838                 wake_up = 1;
 839         }
 840         spin_unlock_bh(&obd->obd_processing_task_lock);
 841
 842         return wake_up;
 843 }
 844
 845 static struct ptlrpc_request *
 846 target_next_replay_req(struct obd_device *obd)
 847 {
 848         struct l_wait_info lwi = { 0 };
 849         struct ptlrpc_request *req;
 850
 851         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
 852                obd->obd_next_recovery_transno);
 853         l_wait_event(obd->obd_next_transno_waitq,
 854                      check_for_next_transno(obd), &lwi);
 855
 856         spin_lock_bh(&obd->obd_processing_task_lock);
 857         if (obd->obd_abort_recovery) {
 858                 req = NULL;
 859         } else if (!list_empty(&obd->obd_recovery_queue)) {
 860                 req = list_entry(obd->obd_recovery_queue.next,
 861                                  struct ptlrpc_request, rq_list);
 862                 list_del_init(&req->rq_list);
 863                 obd->obd_requests_queued_for_recovery--;
 864         } else {
 865                 req = NULL;
 866         }
 867         spin_unlock_bh(&obd->obd_processing_task_lock);
 868         return req;
 869 }
 870
 871
 872 static int target_recovery_thread(void *arg)
 873 {
 874         struct obd_device *obd = arg;
 875         struct ptlrpc_request *req;
 876         struct target_recovery_data *trd = &obd->obd_recovery_data;
 877         unsigned long flags;
 878         ENTRY;
 879
 880         kportal_daemonize("tgt-recov");
 881
 882         SIGNAL_MASK_LOCK(current, flags);
 883         sigfillset(&current->blocked);
 884         RECALC_SIGPENDING;
 885         SIGNAL_MASK_UNLOCK(current, flags);
 886
 887         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
 888                current->pid);
 889         trd->trd_processing_task = current->pid;
 890
 891         obd->obd_recovering = 1;
 892         complete(&trd->trd_starting);
 893
 894         while (obd->obd_recovering) {
 895                 LASSERT(trd->trd_processing_task == current->pid);
 896                 req = target_next_replay_req(obd);
 897                 if (req != NULL) {
 898                         char peer_str[PTL_NALFMT_SIZE];
 899                         DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s: ",
 900                                   req->rq_reqmsg->transno,
 901                                   ptlrpc_peernid2str(&req->rq_peer, peer_str));
 902                         (void)trd->trd_recovery_handler(req);
 903                         obd->obd_replayed_requests++;
 904                         reset_recovery_timer(obd);
 905                         /* bug 1580: decide how to properly sync() in recovery*/
 906                         //mds_fsync_super(mds->mds_sb);
 907                         ptlrpc_free_clone(req);
 908                         spin_lock_bh(&obd->obd_processing_task_lock);
 909                         obd->obd_next_recovery_transno++;
 910                         spin_unlock_bh(&obd->obd_processing_task_lock);
 911                 } else {
 912                         /* recovery is over */
 913                         spin_lock_bh(&obd->obd_processing_task_lock);
 914                         obd->obd_recovering = 0;
 915                         target_cancel_recovery_timer(obd);
 916                         if (obd->obd_abort_recovery) {
 917                                 obd->obd_abort_recovery = 0;
 918                                 spin_unlock_bh(&obd->obd_processing_task_lock);
 919                                 target_abort_recovery(obd);
 920                         } else {
 921                                 LASSERT(obd->obd_recoverable_clients == 0);
 922                                 spin_unlock_bh(&obd->obd_processing_task_lock);
 923                                 target_finish_recovery(obd);
 924                         }
 925                 }
 926         }
 927
 928         trd->trd_processing_task = 0;
 929         complete(&trd->trd_finishing);
 930         return 0;
 931 }
 932
 933 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
 934 {
 935         int rc = 0;
 936         struct target_recovery_data *trd = &obd->obd_recovery_data;
 937
 938         memset(trd, 0, sizeof(*trd));
 939         init_completion(&trd->trd_starting);
 940         init_completion(&trd->trd_finishing);
 941         trd->trd_recovery_handler = handler;
 942
 943         if (kernel_thread(target_recovery_thread, obd, 0) == 0)
 944                 wait_for_completion(&trd->trd_starting);
 945         else
 946                 rc = -ECHILD;
 947
 948         return rc;
 949 }
 950
 951 void target_stop_recovery_thread(struct obd_device *obd)
 952 {
 953         spin_lock_bh(&obd->obd_processing_task_lock);
 954         if (obd->obd_recovery_data.trd_processing_task > 0) {
 955                 struct target_recovery_data *trd = &obd->obd_recovery_data;
 956                 CERROR("%s: aborting recovery\n", obd->obd_name);
 957                 obd->obd_abort_recovery = 1;
 958                 wake_up(&obd->obd_next_transno_waitq);
 959                 spin_unlock_bh(&obd->obd_processing_task_lock);
 960                 wait_for_completion(&trd->trd_finishing);
 961         } else {
 962                 spin_unlock_bh(&obd->obd_processing_task_lock);
 963         }
 964 }
 965
 966 int target_queue_recovery_request(struct ptlrpc_request *req,
 967                                   struct obd_device *obd)
 968 {
 969         struct list_head *tmp;
 970         int inserted = 0;
 971         __u64 transno = req->rq_reqmsg->transno;
 972
 973         /* CAVEAT EMPTOR: The incoming request message has been swabbed
 974          * (i.e. buflens etc are in my own byte order), but type-dependent
 975          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
 976
 977         if (!transno) {
 978                 INIT_LIST_HEAD(&req->rq_list);
 979                 DEBUG_REQ(D_HA, req, "not queueing");
 980                 return 1;
 981         }
 982
 983
 984         /* If we're processing the queue, we want don't want to queue this
 985          * message.
 986          *
 987          * Also, if this request has a transno less than the one we're waiting
 988          * for, we should process it now.  It could (and currently always will)
 989          * be an open request for a descriptor that was opened some time ago.
 990          *
 991          * Also, a resent, replayed request that has already been
 992          * handled will pass through here and be processed immediately.
 993          */
 994         spin_lock_bh(&obd->obd_processing_task_lock);
 995         if (obd->obd_recovery_data.trd_processing_task == current->pid ||
 996             transno < obd->obd_next_recovery_transno) {
 997                 /* Processing the queue right now, don't re-add. */
 998                 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
 999                 LASSERT(list_empty(&req->rq_list));
1000                 spin_unlock_bh(&obd->obd_processing_task_lock);
1001                 return 1;
1002         }
1003         spin_unlock_bh(&obd->obd_processing_task_lock);
1004
1005         /* A resent, replayed request that is still on the queue; just drop it.
1006            The queued request will handle this. */
1007         if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY))
1008             == (MSG_RESENT | MSG_REPLAY)) {
1009                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1010                 return 0;
1011         }
1012
1013         req = ptlrpc_clone_req(req);
1014         if (req == NULL)
1015                 return -ENOMEM;
1016
1017         spin_lock_bh(&obd->obd_processing_task_lock);
1018
1019         /* XXX O(n^2) */
1020         list_for_each(tmp, &obd->obd_recovery_queue) {
1021                 struct ptlrpc_request *reqiter =
1022                         list_entry(tmp, struct ptlrpc_request, rq_list);
1023
1024                 if (reqiter->rq_reqmsg->transno > transno) {
1025                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1026                         inserted = 1;
1027                         break;
1028                 }
1029         }
1030
1031         if (!inserted) {
1032                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
1033         }
1034
1035         obd->obd_requests_queued_for_recovery++;
1036         wake_up(&obd->obd_next_transno_waitq);
1037         spin_unlock_bh(&obd->obd_processing_task_lock);
1038
1039         return 0;
1040 }
1041
1042 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1043 {
1044         return req->rq_export->exp_obd;
1045 }
1046
1047 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
1048 {
1049         struct obd_device *obd = target_req2obd(req);
1050
1051         LASSERT ((rc == 0) == (req->rq_reply_state != NULL));
1052
1053         if (rc) {
1054                 /* Just like ptlrpc_error, but without the sending. */
1055                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1056                 LASSERT(rc == 0); /* XXX handle this */
1057                 req->rq_type = PTL_RPC_MSG_ERR;
1058         }
1059
1060         LASSERT (!req->rq_reply_state->rs_difficult);
1061         LASSERT(list_empty(&req->rq_list));
1062
1063         req = ptlrpc_clone_req(req);
1064
1065         spin_lock_bh(&obd->obd_processing_task_lock);
1066
1067         list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
1068
1069         /* only count the first "replay over" request from each
1070            export */
1071         if (req->rq_export->exp_replay_needed) {
1072                 --obd->obd_recoverable_clients;
1073                 req->rq_export->exp_replay_needed = 0;
1074                 CWARN("%s: %d recoverable clients remain\n",
1075                       obd->obd_name, obd->obd_recoverable_clients);
1076         }
1077         wake_up(&obd->obd_next_transno_waitq);
1078         spin_unlock_bh(&obd->obd_processing_task_lock);
1079         return 1;
1080 }
1081
1082 int
1083 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
1084 {
1085         if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1086                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1087                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1088                 /* NB this does _not_ send with ACK disabled, to simulate
1089                  * sending OK, but timing out for the ACK */
1090                 if (req->rq_reply_state != NULL) {
1091                         if (!req->rq_reply_state->rs_difficult) {
1092                                 lustre_free_reply_state (req->rq_reply_state);
1093                                 req->rq_reply_state = NULL;
1094                         } else {
1095                                 struct ptlrpc_service *svc =
1096                                         req->rq_rqbd->rqbd_srv_ni->sni_service;
1097                                 atomic_inc(&svc->srv_outstanding_replies);
1098                         }
1099                 }
1100                 return (-ECOMM);
1101         }
1102
1103         if (rc) {
1104                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1105                 if (req->rq_reply_state == NULL) {
1106                         rc = lustre_pack_reply (req, 0, NULL, NULL);
1107                         if (rc != 0) {
1108                                 CERROR ("can't allocate reply\n");
1109                                 return (rc);
1110                         }
1111                 }
1112                 req->rq_type = PTL_RPC_MSG_ERR;
1113         } else {
1114                 DEBUG_REQ(D_NET, req, "sending reply");
1115         }
1116
1117         return (ptlrpc_send_reply(req, 1));
1118 }
1119
1120 void
1121 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1122 {
1123         int                        netrc;
1124         unsigned long              flags;
1125         struct ptlrpc_reply_state *rs;
1126         struct obd_device         *obd;
1127         struct obd_export         *exp;
1128         struct ptlrpc_srv_ni      *sni;
1129         struct ptlrpc_service     *svc;
1130
1131         sni = req->rq_rqbd->rqbd_srv_ni;
1132         svc = sni->sni_service;
1133
1134         rs = req->rq_reply_state;
1135         if (rs == NULL || !rs->rs_difficult) {
1136                 /* The easy case; no notifiers and reply_out_callback()
1137                  * cleans up (i.e. we can't look inside rs after a
1138                  * successful send) */
1139                 netrc = target_send_reply_msg (req, rc, fail_id);
1140
1141                 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1142                 return;
1143         }
1144
1145         /* must be an export if locks saved */
1146         LASSERT (req->rq_export != NULL);
1147         /* req/reply consistent */
1148         LASSERT (rs->rs_srv_ni == sni);
1149
1150         /* "fresh" reply */
1151         LASSERT (!rs->rs_scheduled);
1152         LASSERT (!rs->rs_scheduled_ever);
1153         LASSERT (!rs->rs_handled);
1154         LASSERT (!rs->rs_on_net);
1155         LASSERT (rs->rs_export == NULL);
1156         LASSERT (list_empty(&rs->rs_obd_list));
1157         LASSERT (list_empty(&rs->rs_exp_list));
1158
1159         exp = class_export_get (req->rq_export);
1160         obd = exp->exp_obd;
1161
1162         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1163         rs->rs_scheduled = 1;
1164         rs->rs_on_net    = 1;
1165         rs->rs_xid       = req->rq_xid;
1166         rs->rs_transno   = req->rq_transno;
1167         rs->rs_export    = exp;
1168
1169         spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1170
1171         if (rs->rs_transno > obd->obd_last_committed) {
1172                 /* not committed already */
1173                 list_add_tail (&rs->rs_obd_list,
1174                                &obd->obd_uncommitted_replies);
1175         }
1176
1177         spin_unlock (&obd->obd_uncommitted_replies_lock);
1178         spin_lock (&exp->exp_lock);
1179
1180         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1181
1182         spin_unlock_irqrestore (&exp->exp_lock, flags);
1183
1184         netrc = target_send_reply_msg (req, rc, fail_id);
1185
1186         spin_lock_irqsave (&svc->srv_lock, flags);
1187
1188         svc->srv_n_difficult_replies++;
1189
1190         if (netrc != 0) /* error sending: reply is off the net */
1191                 rs->rs_on_net = 0;
1192
1193         if (!rs->rs_on_net ||                   /* some notifier */
1194             list_empty(&rs->rs_exp_list) ||     /* completed already */
1195             list_empty(&rs->rs_obd_list)) {
1196                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1197                 wake_up (&svc->srv_waitq);
1198         } else {
1199                 list_add (&rs->rs_list, &sni->sni_active_replies);
1200                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1201         }
1202
1203         spin_unlock_irqrestore (&svc->srv_lock, flags);
1204 }
1205
1206 int target_handle_ping(struct ptlrpc_request *req)
1207 {
1208         return lustre_pack_reply(req, 0, NULL, NULL);
1209 }