lustre/ptlrpc/import.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Mike Shaver <shaver@clusterfs.com>
   6  *
   7  *   This file is part of the Lustre file system, http://www.lustre.org
   8  *   Lustre is a trademark of Cluster File Systems, Inc.
   9  *
  10  *   You may have signed or agreed to another license before downloading
  11  *   this software.  If so, you are bound by the terms and conditions
  12  *   of that agreement, and the following does not apply to you.  See the
  13  *   LICENSE file included with this distribution for more information.
  14  *
  15  *   If you did not agree to a different license, then this copy of Lustre
  16  *   is open source software; you can redistribute it and/or modify it
  17  *   under the terms of version 2 of the GNU General Public License as
  18  *   published by the Free Software Foundation.
  19  *
  20  *   In either case, Lustre is distributed in the hope that it will be
  21  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  22  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  23  *   license text for more details.
  24  */
  25
  26 #define DEBUG_SUBSYSTEM S_RPC
  27 #ifndef __KERNEL__
  28 # include <liblustre.h>
  29 #endif
  30
  31 #include <obd_support.h>
  32 #include <lustre_ha.h>
  33 #include <lustre_net.h>
  34 #include <lustre_import.h>
  35 #include <lustre_export.h>
  36 #include <obd.h>
  37 #include <obd_class.h>
  38
  39 #include "ptlrpc_internal.h"
  40
  41 struct ptlrpc_connect_async_args {
  42          __u64 pcaa_peer_committed;
  43         int pcaa_initial_connect;
  44 };
  45
  46 /* A CLOSED import should remain so. */
  47 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  48 do {                                                                           \
  49         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  50                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  51                       imp, obd2cli_tgt(imp->imp_obd),                          \
  52                       ptlrpc_import_state_name(imp->imp_state),                \
  53                       ptlrpc_import_state_name(state));                        \
  54                imp->imp_state = state;                                         \
  55         }                                                                      \
  56 } while(0)
  57
  58 #define IMPORT_SET_STATE(imp, state)            \
  59 do {                                            \
  60         spin_lock(&imp->imp_lock);              \
  61         IMPORT_SET_STATE_NOLOCK(imp, state);    \
  62         spin_unlock(&imp->imp_lock);            \
  63 } while(0)
  64
  65
  66 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
  67                                     void * data, int rc);
  68 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
  69
  70 /* Only this function is allowed to change the import state when it is
  71  * CLOSED. I would rather refcount the import and free it after
  72  * disconnection like we do with exports. To do that, the client_obd
  73  * will need to save the peer info somewhere other than in the import,
  74  * though. */
  75 int ptlrpc_init_import(struct obd_import *imp)
  76 {
  77         spin_lock(&imp->imp_lock);
  78
  79         imp->imp_generation++;
  80         imp->imp_state =  LUSTRE_IMP_NEW;
  81
  82         spin_unlock(&imp->imp_lock);
  83
  84         return 0;
  85 }
  86 EXPORT_SYMBOL(ptlrpc_init_import);
  87
  88 #define UUID_STR "_UUID"
  89 static void deuuidify(char *uuid, const char *prefix, char **uuid_start,
  90                       int *uuid_len)
  91 {
  92         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
  93                 ? uuid : uuid + strlen(prefix);
  94
  95         *uuid_len = strlen(*uuid_start);
  96
  97         if (*uuid_len < strlen(UUID_STR))
  98                 return;
  99
 100         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 101                     UUID_STR, strlen(UUID_STR)))
 102                 *uuid_len -= strlen(UUID_STR);
 103 }
 104
 105 /* Returns true if import was FULL, false if import was already not
 106  * connected.
 107  * @imp - import to be disconnected
 108  * @conn_cnt - connection count (epoch) of the request that timed out
 109  *             and caused the disconnection.  In some cases, multiple
 110  *             inflight requests can fail to a single target (e.g. OST
 111  *             bulk requests) and if one has already caused a reconnection
 112  *             (increasing the import->conn_cnt) the older failure should
 113  *             not also cause a reconnection.  If zero it forces a reconnect.
 114  */
 115 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 116 {
 117         int rc = 0;
 118
 119         spin_lock(&imp->imp_lock);
 120
 121         if (imp->imp_state == LUSTRE_IMP_FULL &&
 122             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 123                 char *target_start;
 124                 int   target_len;
 125
 126                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 127                           &target_start, &target_len);
 128
 129                 if (imp->imp_replayable) {
 130                         LCONSOLE_WARN("%s: Connection to service %.*s via nid "
 131                                "%s was lost; in progress operations using this "
 132                                "service will wait for recovery to complete.\n",
 133                                imp->imp_obd->obd_name, target_len, target_start,
 134                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 135                 } else {
 136                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to service "
 137                                            "%.*s via nid %s was lost; in progress"
 138                                            "operations using this service will"
 139                                            "fail.\n",
 140                                            imp->imp_obd->obd_name,
 141                                            target_len, target_start,
 142                                  libcfs_nid2str(imp->imp_connection->c_peer.nid));
 143                 }
 144                 ptlrpc_deactivate_timeouts(imp);
 145                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 146                 spin_unlock(&imp->imp_lock);
 147
 148                 if (obd_dump_on_timeout)
 149                         libcfs_debug_dumplog();
 150
 151                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 152                 rc = 1;
 153         } else {
 154                 spin_unlock(&imp->imp_lock);
 155                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 156                        imp->imp_client->cli_name, imp,
 157                        (imp->imp_state == LUSTRE_IMP_FULL &&
 158                         imp->imp_conn_cnt > conn_cnt) ?
 159                        "reconnected" : "not connected", imp->imp_conn_cnt,
 160                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 161         }
 162
 163         return rc;
 164 }
 165
 166 /* Must be called with imp_lock held! */
 167 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 168 {
 169         ENTRY;
 170         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 171
 172         if (imp->imp_invalid) {
 173                 spin_unlock(&imp->imp_lock);
 174                 EXIT;
 175                 return;
 176         }
 177
 178         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 179         imp->imp_invalid = 1;
 180         imp->imp_generation++;
 181         spin_unlock(&imp->imp_lock);
 182
 183         ptlrpc_abort_inflight(imp);
 184         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 185
 186         EXIT;
 187 }
 188
 189 /*
 190  * This acts as a barrier; all existing requests are rejected, and
 191  * no new requests will be accepted until the import is valid again.
 192  */
 193 void ptlrpc_deactivate_import(struct obd_import *imp)
 194 {
 195         spin_lock(&imp->imp_lock);
 196         ptlrpc_deactivate_and_unlock_import(imp);
 197 }
 198
 199 /*
 200  * This function will invalidate the import, if necessary, then block
 201  * for all the RPC completions, and finally notify the obd to
 202  * invalidate its state (ie cancel locks, clear pending requests,
 203  * etc).
 204  */
 205 void ptlrpc_invalidate_import(struct obd_import *imp)
 206 {
 207         struct l_wait_info lwi;
 208         int rc;
 209
 210         atomic_inc(&imp->imp_inval_count);
 211
 212         ptlrpc_deactivate_import(imp);
 213
 214         LASSERT(imp->imp_invalid);
 215
 216         /* wait for all requests to error out and call completion callbacks */
 217         lwi = LWI_TIMEOUT_INTERVAL(cfs_timeout_cap(cfs_time_seconds(obd_timeout)),
 218                                    HZ, NULL, NULL);
 219         rc = l_wait_event(imp->imp_recovery_waitq,
 220                           (atomic_read(&imp->imp_inflight) == 0), &lwi);
 221
 222         if (rc)
 223                 CDEBUG(D_HA, "%s: rc = %d waiting for callback (%d != 0)\n",
 224                        obd2cli_tgt(imp->imp_obd), rc,
 225                        atomic_read(&imp->imp_inflight));
 226
 227         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 228         sptlrpc_import_flush_all_ctx(imp);
 229
 230         atomic_dec(&imp->imp_inval_count);
 231         cfs_waitq_signal(&imp->imp_recovery_waitq);
 232 }
 233
 234 /* unset imp_invalid */
 235 void ptlrpc_activate_import(struct obd_import *imp)
 236 {
 237         struct obd_device *obd = imp->imp_obd;
 238
 239         spin_lock(&imp->imp_lock);
 240         imp->imp_invalid = 0;
 241         ptlrpc_activate_timeouts(imp);
 242         spin_unlock(&imp->imp_lock);
 243         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 244 }
 245
 246 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 247 {
 248         ENTRY;
 249
 250         LASSERT(!imp->imp_dlm_fake);
 251
 252         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 253                 if (!imp->imp_replayable) {
 254                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 255                                "auto-deactivating\n",
 256                                obd2cli_tgt(imp->imp_obd),
 257                                imp->imp_connection->c_remote_uuid.uuid,
 258                                imp->imp_obd->obd_name);
 259                         ptlrpc_deactivate_import(imp);
 260                 }
 261
 262                 CDEBUG(D_HA, "%s: waking up pinger\n",
 263                        obd2cli_tgt(imp->imp_obd));
 264
 265                 spin_lock(&imp->imp_lock);
 266                 imp->imp_force_verify = 1;
 267                 spin_unlock(&imp->imp_lock);
 268
 269                 ptlrpc_pinger_wake_up();
 270         }
 271         EXIT;
 272 }
 273
 274 int ptlrpc_reconnect_import(struct obd_import *imp)
 275 {
 276
 277         ptlrpc_set_import_discon(imp, 0);
 278         /* Force a new connect attempt */
 279         ptlrpc_invalidate_import(imp);
 280         /* Do a fresh connect next time by zeroing the handle */
 281         ptlrpc_disconnect_import(imp, 1);
 282         /* Wait for all invalidate calls to finish */
 283         if (atomic_read(&imp->imp_inval_count) > 0) {
 284                 int rc;
 285                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 286                 rc = l_wait_event(imp->imp_recovery_waitq,
 287                                   (atomic_read(&imp->imp_inval_count) == 0),
 288                                   &lwi);
 289                 if (rc)
 290                         CERROR("Interrupted, inval=%d\n",
 291                                atomic_read(&imp->imp_inval_count));
 292         }
 293
 294         /* Allow reconnect attempts */
 295         imp->imp_obd->obd_no_recov = 0;
 296         /* Remove 'invalid' flag */
 297         ptlrpc_activate_import(imp);
 298         /* Attempt a new connect */
 299         ptlrpc_recover_import(imp, NULL);
 300         return 0;
 301 }
 302
 303 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 304
 305 static int import_select_connection(struct obd_import *imp)
 306 {
 307         struct obd_import_conn *imp_conn = NULL, *conn;
 308         struct obd_export *dlmexp;
 309         ENTRY;
 310
 311         spin_lock(&imp->imp_lock);
 312
 313         if (list_empty(&imp->imp_conn_list)) {
 314                 CERROR("%s: no connections available\n",
 315                         imp->imp_obd->obd_name);
 316                 spin_unlock(&imp->imp_lock);
 317                 RETURN(-EINVAL);
 318         }
 319
 320         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 321                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 322                        imp->imp_obd->obd_name,
 323                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 324                        conn->oic_last_attempt);
 325                 /* Throttle the reconnect rate to once per RECONNECT_INTERVAL */
 326                 if (cfs_time_before_64(conn->oic_last_attempt +
 327                                        RECONNECT_INTERVAL * HZ,
 328                                        cfs_time_current_64())) {
 329                         /* If we have never tried this connection since the
 330                            the last successful attempt, go with this one */
 331                         if (cfs_time_beforeq_64(conn->oic_last_attempt,
 332                                                imp->imp_last_success_conn)) {
 333                                 imp_conn = conn;
 334                                 break;
 335                         }
 336
 337                         /* Both of these connections have already been tried
 338                            since the last successful connection; just choose the
 339                            least recently used */
 340                         if (!imp_conn)
 341                                 imp_conn = conn;
 342                         else if (cfs_time_before_64(conn->oic_last_attempt,
 343                                                     imp_conn->oic_last_attempt))
 344                                 imp_conn = conn;
 345                 }
 346         }
 347
 348         /* if not found, simply choose the current one */
 349         if (!imp_conn) {
 350                 LASSERT(imp->imp_conn_current);
 351                 imp_conn = imp->imp_conn_current;
 352         }
 353         LASSERT(imp_conn->oic_conn);
 354
 355         imp_conn->oic_last_attempt = cfs_time_current_64();
 356
 357         /* switch connection, don't mind if it's same as the current one */
 358         if (imp->imp_connection)
 359                 ptlrpc_put_connection(imp->imp_connection);
 360         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 361
 362         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 363         LASSERT(dlmexp != NULL);
 364         if (dlmexp->exp_connection)
 365                 ptlrpc_put_connection(dlmexp->exp_connection);
 366         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 367         class_export_put(dlmexp);
 368
 369         if (imp->imp_conn_current != imp_conn) {
 370                 if (imp->imp_conn_current)
 371                         LCONSOLE_INFO("Changing connection for %s to %s/%s\n",
 372                                       imp->imp_obd->obd_name,
 373                                       imp_conn->oic_uuid.uuid,
 374                                       libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 375                 imp->imp_conn_current = imp_conn;
 376         }
 377
 378         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 379                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 380                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 381
 382         spin_unlock(&imp->imp_lock);
 383
 384         RETURN(0);
 385 }
 386
 387 /*
 388  * must be called under imp_lock
 389  */
 390 int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 391 {
 392         struct ptlrpc_request *req;
 393         struct list_head *tmp;
 394
 395         if (list_empty(&imp->imp_replay_list))
 396                 return 0;
 397         tmp = imp->imp_replay_list.next;
 398         req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 399         *transno = req->rq_transno;
 400         if (req->rq_transno == 0) {
 401                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 402                 LBUG();
 403         }
 404
 405         return 1;
 406 }
 407
 408 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
 409 {
 410         struct obd_device *obd = imp->imp_obd;
 411         int initial_connect = 0;
 412         int set_transno = 0;
 413         __u64 committed_before_reconnect = 0;
 414         struct ptlrpc_request *request;
 415         char *bufs[] = { NULL,
 416                          obd2cli_tgt(imp->imp_obd),
 417                          obd->obd_uuid.uuid,
 418                          (char *)&imp->imp_dlm_handle,
 419                          (char *)&imp->imp_connect_data };
 420         struct ptlrpc_connect_async_args *aa;
 421         int rc;
 422         ENTRY;
 423
 424         spin_lock(&imp->imp_lock);
 425         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 426                 spin_unlock(&imp->imp_lock);
 427                 CERROR("can't connect to a closed import\n");
 428                 RETURN(-EINVAL);
 429         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 430                 spin_unlock(&imp->imp_lock);
 431                 CERROR("already connected\n");
 432                 RETURN(0);
 433         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 434                 spin_unlock(&imp->imp_lock);
 435                 CERROR("already connecting\n");
 436                 RETURN(-EALREADY);
 437         }
 438
 439         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 440
 441         imp->imp_conn_cnt++;
 442         imp->imp_resend_replay = 0;
 443
 444         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 445                 initial_connect = 1;
 446         else
 447                 committed_before_reconnect = imp->imp_peer_committed_transno;
 448
 449         set_transno = ptlrpc_first_transno(imp, &imp->imp_connect_data.ocd_transno);
 450         spin_unlock(&imp->imp_lock);
 451
 452         if (new_uuid) {
 453                 struct obd_uuid uuid;
 454
 455                 obd_str2uuid(&uuid, new_uuid);
 456                 rc = import_set_conn_priority(imp, &uuid);
 457                 if (rc)
 458                         GOTO(out, rc);
 459         }
 460
 461         rc = import_select_connection(imp);
 462         if (rc)
 463                 GOTO(out, rc);
 464
 465         /* last in connection list */
 466         if (imp->imp_conn_current->oic_item.next == &imp->imp_conn_list) {
 467                 if (imp->imp_initial_recov_bk && initial_connect) {
 468                         CDEBUG(D_HA, "Last connection attempt (%d) for %s\n",
 469                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 470                         /* Don't retry if connect fails */
 471                         rc = 0;
 472                         obd_set_info_async(obd->obd_self_export,
 473                                            strlen(KEY_INIT_RECOV),
 474                                            KEY_INIT_RECOV,
 475                                            sizeof(rc), &rc, NULL);
 476                 }
 477                 if (imp->imp_recon_bk) {
 478                         CDEBUG(D_HA, "Last reconnection attempt (%d) for %s\n",
 479                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 480                         spin_lock(&imp->imp_lock);
 481                         imp->imp_last_recon = 1;
 482                         spin_unlock(&imp->imp_lock);
 483                 }
 484         }
 485
 486         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 487         if (rc)
 488                 GOTO(out, rc);
 489
 490         /* Reset connect flags to the originally requested flags, in case
 491          * the server is updated on-the-fly we will get the new features. */
 492         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 493         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 494                            &obd->obd_uuid, &imp->imp_connect_data);
 495         if (rc)
 496                 GOTO(out, rc);
 497
 498         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 499         if (request == NULL)
 500                 GOTO(out, rc = -ENOMEM);
 501
 502         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 503                                       imp->imp_connect_op, bufs, NULL);
 504         if (rc) {
 505                 ptlrpc_request_free(request);
 506                 GOTO(out, rc);
 507         }
 508
 509 #ifndef __KERNEL__
 510         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 511 #endif
 512         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 513
 514         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 515         /* Allow a slightly larger reply for future growth compatibility */
 516         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 517                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 518         ptlrpc_request_set_replen(request);
 519         request->rq_interpret_reply = ptlrpc_connect_interpret;
 520
 521         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 522         aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
 523         memset(aa, 0, sizeof *aa);
 524
 525         aa->pcaa_peer_committed = committed_before_reconnect;
 526         aa->pcaa_initial_connect = initial_connect;
 527
 528         if (aa->pcaa_initial_connect) {
 529                 spin_lock(&imp->imp_lock);
 530                 imp->imp_replayable = 1;
 531                 spin_unlock(&imp->imp_lock);
 532                 /* On an initial connect, we don't know which one of a
 533                    failover server pair is up.  Don't wait long. */
 534 #ifdef CRAY_XT3
 535                 request->rq_timeout = max((int)(obd_timeout / 2), 5);
 536 #else
 537                 request->rq_timeout = max((int)(obd_timeout / 20), 5);
 538 #endif
 539                 lustre_msg_add_op_flags(request->rq_reqmsg,
 540                                         MSG_CONNECT_INITIAL);
 541         }
 542
 543         if (set_transno)
 544                 lustre_msg_add_op_flags(request->rq_reqmsg,
 545                                         MSG_CONNECT_TRANSNO);
 546
 547         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request");
 548         ptlrpcd_add_req(request);
 549         rc = 0;
 550 out:
 551         if (rc != 0) {
 552                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 553         }
 554
 555         RETURN(rc);
 556 }
 557 EXPORT_SYMBOL(ptlrpc_connect_import);
 558
 559 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 560 {
 561 #ifdef __KERNEL__
 562         struct obd_import_conn *imp_conn;
 563 #endif
 564         int wake_pinger = 0;
 565
 566         ENTRY;
 567
 568         spin_lock(&imp->imp_lock);
 569         if (list_empty(&imp->imp_conn_list))
 570                 GOTO(unlock, 0);
 571
 572 #ifdef __KERNEL__
 573         imp_conn = list_entry(imp->imp_conn_list.prev,
 574                               struct obd_import_conn,
 575                               oic_item);
 576
 577         /* XXX: When the failover node is the primary node, it is possible
 578          * to have two identical connections in imp_conn_list. We must
 579          * compare not conn's pointers but NIDs, otherwise we can defeat
 580          * connection throttling. (See bug 14774.) */
 581         if (imp->imp_conn_current->oic_conn->c_self !=
 582                                 imp_conn->oic_conn->c_self) {
 583                 ptlrpc_ping_import_soon(imp);
 584                 wake_pinger = 1;
 585         }
 586 #else
 587         /* liblustre has no pinger thead, so we wakup pinger anyway */
 588         wake_pinger = 1;
 589 #endif
 590
 591  unlock:
 592         spin_unlock(&imp->imp_lock);
 593
 594         if (wake_pinger)
 595                 ptlrpc_pinger_wake_up();
 596
 597         EXIT;
 598 }
 599
 600 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
 601                                     void * data, int rc)
 602 {
 603         struct ptlrpc_connect_async_args *aa = data;
 604         struct obd_import *imp = request->rq_import;
 605         struct client_obd *cli = &imp->imp_obd->u.cli;
 606         struct lustre_handle old_hdl;
 607         int msg_flags;
 608         ENTRY;
 609
 610         spin_lock(&imp->imp_lock);
 611         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 612                 spin_unlock(&imp->imp_lock);
 613                 RETURN(0);
 614         }
 615         spin_unlock(&imp->imp_lock);
 616
 617         if (rc)
 618                 GOTO(out, rc);
 619
 620         LASSERT(imp->imp_conn_current);
 621
 622         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 623
 624         /* All imports are pingable */
 625         spin_lock(&imp->imp_lock);
 626         imp->imp_pingable = 1;
 627
 628         if (aa->pcaa_initial_connect) {
 629                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 630                         imp->imp_replayable = 1;
 631                         spin_unlock(&imp->imp_lock);
 632                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 633                                obd2cli_tgt(imp->imp_obd));
 634                 } else {
 635                         imp->imp_replayable = 0;
 636                         spin_unlock(&imp->imp_lock);
 637                 }
 638
 639                 /* if applies, adjust the imp->imp_msg_magic here
 640                  * according to reply flags */
 641
 642                 imp->imp_remote_handle =
 643                                 *lustre_msg_get_handle(request->rq_repmsg);
 644
 645                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 646                 ptlrpc_activate_import(imp);
 647                 GOTO(finish, rc = 0);
 648         } else {
 649                 spin_unlock(&imp->imp_lock);
 650         }
 651
 652         /* Determine what recovery state to move the import to. */
 653         if (MSG_CONNECT_RECONNECT & msg_flags) {
 654                 memset(&old_hdl, 0, sizeof(old_hdl));
 655                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 656                             sizeof (old_hdl))) {
 657                         CERROR("%s@%s didn't like our handle "LPX64
 658                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 659                                imp->imp_connection->c_remote_uuid.uuid,
 660                                imp->imp_dlm_handle.cookie);
 661                         GOTO(out, rc = -ENOTCONN);
 662                 }
 663
 664                 if (memcmp(&imp->imp_remote_handle,
 665                            lustre_msg_get_handle(request->rq_repmsg),
 666                            sizeof(imp->imp_remote_handle))) {
 667
 668                         CWARN("%s@%s changed server handle from "
 669                                LPX64" to "LPX64" - evicting.\n",
 670                                obd2cli_tgt(imp->imp_obd),
 671                                imp->imp_connection->c_remote_uuid.uuid,
 672                                imp->imp_remote_handle.cookie,
 673                                lustre_msg_get_handle(request->rq_repmsg)->
 674                                          cookie);
 675                         imp->imp_remote_handle =
 676                                      *lustre_msg_get_handle(request->rq_repmsg);
 677
 678                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 679                         GOTO(finish, rc = 0);
 680                 } else {
 681                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 682                                obd2cli_tgt(imp->imp_obd),
 683                                imp->imp_connection->c_remote_uuid.uuid);
 684                 }
 685
 686                 if (imp->imp_invalid) {
 687                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 688                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 689                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 690                                imp->imp_obd->obd_name,
 691                                obd2cli_tgt(imp->imp_obd));
 692
 693                         spin_lock(&imp->imp_lock);
 694                         imp->imp_resend_replay = 1;
 695                         spin_unlock(&imp->imp_lock);
 696
 697                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 698                 } else {
 699                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 700                 }
 701         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 702                 LASSERT(imp->imp_replayable);
 703                 imp->imp_remote_handle =
 704                                 *lustre_msg_get_handle(request->rq_repmsg);
 705                 imp->imp_last_replay_transno = 0;
 706                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 707         } else {
 708                 imp->imp_remote_handle =
 709                                 *lustre_msg_get_handle(request->rq_repmsg);
 710                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 711         }
 712
 713         /* Sanity checks for a reconnected import. */
 714         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 715                 CERROR("imp_replayable flag does not match server "
 716                        "after reconnect. We should LBUG right here.\n");
 717         }
 718
 719         if (lustre_msg_get_last_committed(request->rq_repmsg) <
 720             aa->pcaa_peer_committed) {
 721                 CERROR("%s went back in time (transno "LPD64
 722                        " was previously committed, server now claims "LPD64
 723                        ")!  See https://bugzilla.clusterfs.com/"
 724                        "long_list.cgi?buglist=9646\n",
 725                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 726                        lustre_msg_get_last_committed(request->rq_repmsg));
 727         }
 728
 729 finish:
 730         rc = ptlrpc_import_recovery_state_machine(imp);
 731         if (rc != 0) {
 732                 if (rc == -ENOTCONN) {
 733                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 734                                "invalidating and reconnecting\n",
 735                                obd2cli_tgt(imp->imp_obd),
 736                                imp->imp_connection->c_remote_uuid.uuid);
 737                         ptlrpc_connect_import(imp, NULL);
 738                         RETURN(0);
 739                 }
 740         } else {
 741                 struct obd_connect_data *ocd;
 742                 struct obd_export *exp;
 743                 int ret;
 744                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 745                                            RCL_SERVER);
 746                 /* server replied obd_connect_data is always bigger */
 747                 ocd = req_capsule_server_sized_get(&request->rq_pill,
 748                                                    &RMF_CONNECT_DATA, ret);
 749
 750                 spin_lock(&imp->imp_lock);
 751                 list_del(&imp->imp_conn_current->oic_item);
 752                 list_add(&imp->imp_conn_current->oic_item, &imp->imp_conn_list);
 753                 imp->imp_last_success_conn =
 754                         imp->imp_conn_current->oic_last_attempt;
 755
 756                 if (ocd == NULL) {
 757                         spin_unlock(&imp->imp_lock);
 758                         CERROR("Wrong connect data from server\n");
 759                         rc = -EPROTO;
 760                         GOTO(out, rc);
 761                 }
 762
 763                 imp->imp_connect_data = *ocd;
 764
 765                 exp = class_conn2export(&imp->imp_dlm_handle);
 766                 spin_unlock(&imp->imp_lock);
 767
 768                 /* check that server granted subset of flags we asked for. */
 769                 LASSERTF((ocd->ocd_connect_flags &
 770                           imp->imp_connect_flags_orig) ==
 771                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 772                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 773
 774                 if (!exp) {
 775                         /* This could happen if export is cleaned during the
 776                            connect attempt */
 777                         CERROR("Missing export for %s\n",
 778                                imp->imp_obd->obd_name);
 779                         GOTO(out, rc = -ENODEV);
 780                 }
 781                 exp->exp_connect_flags = ocd->ocd_connect_flags;
 782                 imp->imp_obd->obd_self_export->exp_connect_flags =
 783                                                         ocd->ocd_connect_flags;
 784                 class_export_put(exp);
 785
 786                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 787
 788                 if (!ocd->ocd_ibits_known &&
 789                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 790                         CERROR("Inodebits aware server returned zero compatible"
 791                                " bits?\n");
 792
 793                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 794                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
 795                                         LUSTRE_VERSION_OFFSET_WARN ||
 796                      ocd->ocd_version < LUSTRE_VERSION_CODE -
 797                                         LUSTRE_VERSION_OFFSET_WARN)) {
 798                         /* Sigh, some compilers do not like #ifdef in the middle
 799                            of macro arguments */
 800 #ifdef __KERNEL__
 801                         const char *older =
 802                                 "older. Consider upgrading this client";
 803 #else
 804                         const char *older =
 805                                 "older. Consider recompiling this application";
 806 #endif
 807                         const char *newer = "newer than client version";
 808
 809                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 810                                       "is much %s (%s)\n",
 811                                       obd2cli_tgt(imp->imp_obd),
 812                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 813                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 814                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 815                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
 816                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
 817                                       newer : older, LUSTRE_VERSION_STRING);
 818                 }
 819
 820                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
 821                         /* We sent to the server ocd_cksum_types with bits set
 822                          * for algorithms we understand. The server masked off
 823                          * the checksum types it doesn't support */
 824                         if ((ocd->ocd_cksum_types & OBD_CKSUM_ALL) == 0) {
 825                                 LCONSOLE_WARN("The negotiation of the checksum "
 826                                               "alogrithm to use with server %s "
 827                                               "failed (%x/%x), disabling "
 828                                               "checksums\n",
 829                                               obd2cli_tgt(imp->imp_obd),
 830                                               ocd->ocd_cksum_types,
 831                                               OBD_CKSUM_ALL);
 832                                 cli->cl_checksum = 0;
 833                                 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 834                                 cli->cl_cksum_type = OBD_CKSUM_CRC32;
 835                         } else {
 836                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
 837
 838                                 if (ocd->ocd_cksum_types & OSC_DEFAULT_CKSUM)
 839                                         cli->cl_cksum_type = OSC_DEFAULT_CKSUM;
 840                                 else if (ocd->ocd_cksum_types & OBD_CKSUM_ADLER)
 841                                         cli->cl_cksum_type = OBD_CKSUM_ADLER;
 842                                 else
 843                                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
 844                         }
 845                 } else {
 846                         /* The server does not support OBD_CONNECT_CKSUM.
 847                          * Enforce CRC32 for backward compatibility*/
 848                         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 849                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
 850                 }
 851
 852                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
 853                         cli->cl_max_pages_per_rpc =
 854                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
 855                 }
 856
 857                 imp->imp_obd->obd_namespace->ns_connect_flags =
 858                                                         ocd->ocd_connect_flags;
 859                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
 860                                                         ocd->ocd_connect_flags;
 861
 862                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 863                         (cli->cl_max_pages_per_rpc > 0));
 864         }
 865
 866 out:
 867         if (rc != 0) {
 868                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 869                 spin_lock(&imp->imp_lock);
 870                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov &&
 871                     (request->rq_import_generation == imp->imp_generation))
 872                         ptlrpc_deactivate_and_unlock_import(imp);
 873                 else
 874                         spin_unlock(&imp->imp_lock);
 875
 876                 if ((imp->imp_recon_bk && imp->imp_last_recon) ||
 877                     (rc == -EACCES)) {
 878                         /*
 879                          * Give up trying to reconnect
 880                          * EACCES means client has no permission for connection
 881                          */
 882                         imp->imp_obd->obd_no_recov = 1;
 883                         ptlrpc_deactivate_import(imp);
 884                 }
 885
 886                 if (rc == -EPROTO) {
 887                         struct obd_connect_data *ocd;
 888
 889                         /* reply message might not be ready */
 890                         if (request->rq_repmsg == NULL)
 891                                 RETURN(-EPROTO);
 892
 893                         ocd = req_capsule_server_get(&request->rq_pill,
 894                                                      &RMF_CONNECT_DATA);
 895                         if (ocd &&
 896                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 897                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
 898                            /* Actually servers are only supposed to refuse
 899                               connection from liblustre clients, so we should
 900                               never see this from VFS context */
 901                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
 902                                         "(%d.%d.%d.%d)"
 903                                         " refused connection from this client "
 904                                         "with an incompatible version (%s).  "
 905                                         "Client must be recompiled\n",
 906                                         obd2cli_tgt(imp->imp_obd),
 907                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 908                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 909                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 910                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
 911                                         LUSTRE_VERSION_STRING);
 912                                 ptlrpc_deactivate_import(imp);
 913                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
 914                         }
 915                         RETURN(-EPROTO);
 916                 }
 917
 918                 ptlrpc_maybe_ping_import_soon(imp);
 919
 920                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
 921                        obd2cli_tgt(imp->imp_obd),
 922                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
 923         }
 924
 925         spin_lock(&imp->imp_lock);
 926         imp->imp_last_recon = 0;
 927         spin_unlock(&imp->imp_lock);
 928
 929         cfs_waitq_signal(&imp->imp_recovery_waitq);
 930         RETURN(rc);
 931 }
 932
 933 static int completed_replay_interpret(struct ptlrpc_request *req,
 934                                     void * data, int rc)
 935 {
 936         ENTRY;
 937         atomic_dec(&req->rq_import->imp_replay_inflight);
 938         if (req->rq_status == 0) {
 939                 ptlrpc_import_recovery_state_machine(req->rq_import);
 940         } else {
 941                 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
 942                        "reconnecting\n",
 943                        req->rq_import->imp_obd->obd_name, req->rq_status);
 944                 ptlrpc_connect_import(req->rq_import, NULL);
 945         }
 946
 947         RETURN(0);
 948 }
 949
 950 static int signal_completed_replay(struct obd_import *imp)
 951 {
 952         struct ptlrpc_request *req;
 953         ENTRY;
 954
 955         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 956         atomic_inc(&imp->imp_replay_inflight);
 957
 958         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
 959                                         OBD_PING);
 960         if (req == NULL) {
 961                 atomic_dec(&imp->imp_replay_inflight);
 962                 RETURN(-ENOMEM);
 963         }
 964
 965         ptlrpc_request_set_replen(req);
 966         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
 967         lustre_msg_add_flags(req->rq_reqmsg,
 968                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
 969         req->rq_timeout *= 3;
 970         req->rq_interpret_reply = completed_replay_interpret;
 971
 972         ptlrpcd_add_req(req);
 973         RETURN(0);
 974 }
 975
 976 #ifdef __KERNEL__
 977 static int ptlrpc_invalidate_import_thread(void *data)
 978 {
 979         struct obd_import *imp = data;
 980
 981         ENTRY;
 982
 983         ptlrpc_daemonize("ll_imp_inval");
 984
 985         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
 986                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 987                imp->imp_connection->c_remote_uuid.uuid);
 988
 989         ptlrpc_invalidate_import(imp);
 990
 991         if (obd_dump_on_eviction) {
 992                 CERROR("dump the log upon eviction\n");
 993                 libcfs_debug_dumplog();
 994         }
 995
 996         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 997         ptlrpc_import_recovery_state_machine(imp);
 998
 999         RETURN(0);
1000 }
1001 #endif
1002
1003 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1004 {
1005         int rc = 0;
1006         int inflight;
1007         char *target_start;
1008         int target_len;
1009
1010         ENTRY;
1011         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1012                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1013                           &target_start, &target_len);
1014                 /* Don't care about MGC eviction */
1015                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1016                            LUSTRE_MGC_NAME) != 0) {
1017                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1018                                            "%.*s; in progress operations using "
1019                                            "this service will fail.\n",
1020                                            target_len, target_start);
1021                 }
1022                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1023                        obd2cli_tgt(imp->imp_obd),
1024                        imp->imp_connection->c_remote_uuid.uuid);
1025
1026 #ifdef __KERNEL__
1027                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
1028                                        CLONE_VM | CLONE_FILES);
1029                 if (rc < 0)
1030                         CERROR("error starting invalidate thread: %d\n", rc);
1031                 else
1032                         rc = 0;
1033                 RETURN(rc);
1034 #else
1035                 ptlrpc_invalidate_import(imp);
1036
1037                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1038 #endif
1039         }
1040
1041         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1042                 CDEBUG(D_HA, "replay requested by %s\n",
1043                        obd2cli_tgt(imp->imp_obd));
1044                 rc = ptlrpc_replay_next(imp, &inflight);
1045                 if (inflight == 0 &&
1046                     atomic_read(&imp->imp_replay_inflight) == 0) {
1047                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1048                         rc = ldlm_replay_locks(imp);
1049                         if (rc)
1050                                 GOTO(out, rc);
1051                 }
1052                 rc = 0;
1053         }
1054
1055         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1056                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1057                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1058                         rc = signal_completed_replay(imp);
1059                         if (rc)
1060                                 GOTO(out, rc);
1061                 }
1062
1063         }
1064
1065         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1066                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1067                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1068                 }
1069         }
1070
1071         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1072                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1073                        obd2cli_tgt(imp->imp_obd),
1074                        imp->imp_connection->c_remote_uuid.uuid);
1075
1076                 rc = ptlrpc_resend(imp);
1077                 if (rc)
1078                         GOTO(out, rc);
1079                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1080                 ptlrpc_activate_import(imp);
1081
1082                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1083                           &target_start, &target_len);
1084                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1085                               "using nid %s.\n", imp->imp_obd->obd_name,
1086                               target_len, target_start,
1087                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1088         }
1089
1090         if (imp->imp_state == LUSTRE_IMP_FULL) {
1091                 cfs_waitq_signal(&imp->imp_recovery_waitq);
1092                 ptlrpc_wake_delayed(imp);
1093         }
1094
1095 out:
1096         RETURN(rc);
1097 }
1098
1099 static int back_to_sleep(void *unused)
1100 {
1101         return 0;
1102 }
1103
1104 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1105 {
1106         struct ptlrpc_request *req;
1107         int rq_opc, rc = 0;
1108         int nowait = imp->imp_obd->obd_force;
1109         ENTRY;
1110
1111         if (nowait)
1112                 GOTO(set_state, rc);
1113
1114         switch (imp->imp_connect_op) {
1115         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1116         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1117         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1118         default:
1119                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1120                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1121                 RETURN(-EINVAL);
1122         }
1123
1124         if (ptlrpc_import_in_recovery(imp)) {
1125                 struct l_wait_info lwi;
1126                 cfs_duration_t timeout;
1127                 if (imp->imp_server_timeout)
1128                         timeout = cfs_time_seconds(obd_timeout / 2);
1129                 else
1130                         timeout = cfs_time_seconds(obd_timeout);
1131
1132                 timeout = MAX(timeout * HZ, 1);
1133
1134                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1135                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1136                 rc = l_wait_event(imp->imp_recovery_waitq,
1137                                   !ptlrpc_import_in_recovery(imp), &lwi);
1138
1139         }
1140
1141         spin_lock(&imp->imp_lock);
1142         if (imp->imp_state != LUSTRE_IMP_FULL)
1143                 GOTO(out, 0);
1144
1145         spin_unlock(&imp->imp_lock);
1146
1147         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1148                                         LUSTRE_OBD_VERSION, rq_opc);
1149         if (req) {
1150                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1151                  * it fails.  We can get through the above with a down server
1152                  * if the client doesn't know the server is gone yet. */
1153                 req->rq_no_resend = 1;
1154 #ifdef CRAY_XT3
1155                 req->rq_timeout = obd_timeout / 3;
1156 #else
1157                 req->rq_timeout = 5;
1158 #endif
1159                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1160                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1161                 ptlrpc_request_set_replen(req);
1162                 rc = ptlrpc_queue_wait(req);
1163                 ptlrpc_req_finished(req);
1164         }
1165
1166 set_state:
1167         spin_lock(&imp->imp_lock);
1168 out:
1169         if (noclose)
1170                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1171         else
1172                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1173         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1174         imp->imp_conn_cnt = 0;
1175         imp->imp_last_recon = 0;
1176         spin_unlock(&imp->imp_lock);
1177
1178         RETURN(rc);
1179 }
1180