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_cksum.h>
  38 #include <obd_class.h>
  39
  40 #include "ptlrpc_internal.h"
  41
  42 struct ptlrpc_connect_async_args {
  43          __u64 pcaa_peer_committed;
  44         int pcaa_initial_connect;
  45 };
  46
  47 /* A CLOSED import should remain so. */
  48 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  49 do {                                                                           \
  50         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  51                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  52                       imp, obd2cli_tgt(imp->imp_obd),                          \
  53                       ptlrpc_import_state_name(imp->imp_state),                \
  54                       ptlrpc_import_state_name(state));                        \
  55                imp->imp_state = state;                                         \
  56         }                                                                      \
  57 } while(0)
  58
  59 #define IMPORT_SET_STATE(imp, state)            \
  60 do {                                            \
  61         spin_lock(&imp->imp_lock);              \
  62         IMPORT_SET_STATE_NOLOCK(imp, state);    \
  63         spin_unlock(&imp->imp_lock);            \
  64 } while(0)
  65
  66
  67 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
  68                                     void * data, int rc);
  69 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
  70
  71 /* Only this function is allowed to change the import state when it is
  72  * CLOSED. I would rather refcount the import and free it after
  73  * disconnection like we do with exports. To do that, the client_obd
  74  * will need to save the peer info somewhere other than in the import,
  75  * though. */
  76 int ptlrpc_init_import(struct obd_import *imp)
  77 {
  78         spin_lock(&imp->imp_lock);
  79
  80         imp->imp_generation++;
  81         imp->imp_state =  LUSTRE_IMP_NEW;
  82
  83         spin_unlock(&imp->imp_lock);
  84
  85         return 0;
  86 }
  87 EXPORT_SYMBOL(ptlrpc_init_import);
  88
  89 #define UUID_STR "_UUID"
  90 static void deuuidify(char *uuid, const char *prefix, char **uuid_start,
  91                       int *uuid_len)
  92 {
  93         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
  94                 ? uuid : uuid + strlen(prefix);
  95
  96         *uuid_len = strlen(*uuid_start);
  97
  98         if (*uuid_len < strlen(UUID_STR))
  99                 return;
 100
 101         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 102                     UUID_STR, strlen(UUID_STR)))
 103                 *uuid_len -= strlen(UUID_STR);
 104 }
 105
 106 /* Returns true if import was FULL, false if import was already not
 107  * connected.
 108  * @imp - import to be disconnected
 109  * @conn_cnt - connection count (epoch) of the request that timed out
 110  *             and caused the disconnection.  In some cases, multiple
 111  *             inflight requests can fail to a single target (e.g. OST
 112  *             bulk requests) and if one has already caused a reconnection
 113  *             (increasing the import->conn_cnt) the older failure should
 114  *             not also cause a reconnection.  If zero it forces a reconnect.
 115  */
 116 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 117 {
 118         int rc = 0;
 119
 120         spin_lock(&imp->imp_lock);
 121
 122         if (imp->imp_state == LUSTRE_IMP_FULL &&
 123             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 124                 char *target_start;
 125                 int   target_len;
 126
 127                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 128                           &target_start, &target_len);
 129
 130                 if (imp->imp_replayable) {
 131                         LCONSOLE_WARN("%s: Connection to service %.*s via nid "
 132                                "%s was lost; in progress operations using this "
 133                                "service will wait for recovery to complete.\n",
 134                                imp->imp_obd->obd_name, target_len, target_start,
 135                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 136                 } else {
 137                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to service "
 138                                            "%.*s via nid %s was lost; in progress"
 139                                            "operations using this service will"
 140                                            "fail.\n",
 141                                            imp->imp_obd->obd_name,
 142                                            target_len, target_start,
 143                                  libcfs_nid2str(imp->imp_connection->c_peer.nid));
 144                 }
 145                 ptlrpc_deactivate_timeouts(imp);
 146                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 147                 spin_unlock(&imp->imp_lock);
 148
 149                 if (obd_dump_on_timeout)
 150                         libcfs_debug_dumplog();
 151
 152                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 153                 rc = 1;
 154         } else {
 155                 spin_unlock(&imp->imp_lock);
 156                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 157                        imp->imp_client->cli_name, imp,
 158                        (imp->imp_state == LUSTRE_IMP_FULL &&
 159                         imp->imp_conn_cnt > conn_cnt) ?
 160                        "reconnected" : "not connected", imp->imp_conn_cnt,
 161                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 162         }
 163
 164         return rc;
 165 }
 166
 167 /* Must be called with imp_lock held! */
 168 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 169 {
 170         ENTRY;
 171         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 172
 173         if (imp->imp_invalid) {
 174                 spin_unlock(&imp->imp_lock);
 175                 EXIT;
 176                 return;
 177         }
 178
 179         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 180         imp->imp_invalid = 1;
 181         imp->imp_generation++;
 182         spin_unlock(&imp->imp_lock);
 183
 184         ptlrpc_abort_inflight(imp);
 185         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 186
 187         EXIT;
 188 }
 189
 190 /*
 191  * This acts as a barrier; all existing requests are rejected, and
 192  * no new requests will be accepted until the import is valid again.
 193  */
 194 void ptlrpc_deactivate_import(struct obd_import *imp)
 195 {
 196         spin_lock(&imp->imp_lock);
 197         ptlrpc_deactivate_and_unlock_import(imp);
 198 }
 199
 200 /*
 201  * This function will invalidate the import, if necessary, then block
 202  * for all the RPC completions, and finally notify the obd to
 203  * invalidate its state (ie cancel locks, clear pending requests,
 204  * etc).
 205  */
 206 void ptlrpc_invalidate_import(struct obd_import *imp)
 207 {
 208         struct l_wait_info lwi;
 209         int rc;
 210
 211         atomic_inc(&imp->imp_inval_count);
 212
 213         ptlrpc_deactivate_import(imp);
 214
 215         LASSERT(imp->imp_invalid);
 216
 217         /* wait for all requests to error out and call completion callbacks */
 218         lwi = LWI_TIMEOUT_INTERVAL(cfs_timeout_cap(cfs_time_seconds(obd_timeout)),
 219                                    HZ, NULL, NULL);
 220         rc = l_wait_event(imp->imp_recovery_waitq,
 221                           (atomic_read(&imp->imp_inflight) == 0), &lwi);
 222
 223         if (rc)
 224                 CDEBUG(D_HA, "%s: rc = %d waiting for callback (%d != 0)\n",
 225                        obd2cli_tgt(imp->imp_obd), rc,
 226                        atomic_read(&imp->imp_inflight));
 227
 228         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 229         sptlrpc_import_flush_all_ctx(imp);
 230
 231         atomic_dec(&imp->imp_inval_count);
 232         cfs_waitq_signal(&imp->imp_recovery_waitq);
 233 }
 234
 235 /* unset imp_invalid */
 236 void ptlrpc_activate_import(struct obd_import *imp)
 237 {
 238         struct obd_device *obd = imp->imp_obd;
 239
 240         spin_lock(&imp->imp_lock);
 241         imp->imp_invalid = 0;
 242         ptlrpc_activate_timeouts(imp);
 243         spin_unlock(&imp->imp_lock);
 244         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 245 }
 246
 247 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 248 {
 249         ENTRY;
 250
 251         LASSERT(!imp->imp_dlm_fake);
 252
 253         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 254                 if (!imp->imp_replayable) {
 255                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 256                                "auto-deactivating\n",
 257                                obd2cli_tgt(imp->imp_obd),
 258                                imp->imp_connection->c_remote_uuid.uuid,
 259                                imp->imp_obd->obd_name);
 260                         ptlrpc_deactivate_import(imp);
 261                 }
 262
 263                 CDEBUG(D_HA, "%s: waking up pinger\n",
 264                        obd2cli_tgt(imp->imp_obd));
 265
 266                 spin_lock(&imp->imp_lock);
 267                 imp->imp_force_verify = 1;
 268                 spin_unlock(&imp->imp_lock);
 269
 270                 ptlrpc_pinger_wake_up();
 271         }
 272         EXIT;
 273 }
 274
 275 int ptlrpc_reconnect_import(struct obd_import *imp)
 276 {
 277
 278         ptlrpc_set_import_discon(imp, 0);
 279         /* Force a new connect attempt */
 280         ptlrpc_invalidate_import(imp);
 281         /* Do a fresh connect next time by zeroing the handle */
 282         ptlrpc_disconnect_import(imp, 1);
 283         /* Wait for all invalidate calls to finish */
 284         if (atomic_read(&imp->imp_inval_count) > 0) {
 285                 int rc;
 286                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 287                 rc = l_wait_event(imp->imp_recovery_waitq,
 288                                   (atomic_read(&imp->imp_inval_count) == 0),
 289                                   &lwi);
 290                 if (rc)
 291                         CERROR("Interrupted, inval=%d\n",
 292                                atomic_read(&imp->imp_inval_count));
 293         }
 294
 295         /* Allow reconnect attempts */
 296         imp->imp_obd->obd_no_recov = 0;
 297         /* Remove 'invalid' flag */
 298         ptlrpc_activate_import(imp);
 299         /* Attempt a new connect */
 300         ptlrpc_recover_import(imp, NULL);
 301         return 0;
 302 }
 303
 304 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 305
 306 static int import_select_connection(struct obd_import *imp)
 307 {
 308         struct obd_import_conn *imp_conn = NULL, *conn;
 309         struct obd_export *dlmexp;
 310         ENTRY;
 311
 312         spin_lock(&imp->imp_lock);
 313
 314         if (list_empty(&imp->imp_conn_list)) {
 315                 CERROR("%s: no connections available\n",
 316                         imp->imp_obd->obd_name);
 317                 spin_unlock(&imp->imp_lock);
 318                 RETURN(-EINVAL);
 319         }
 320
 321         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 322                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 323                        imp->imp_obd->obd_name,
 324                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 325                        conn->oic_last_attempt);
 326                 /* Throttle the reconnect rate to once per RECONNECT_INTERVAL */
 327                 if (cfs_time_before_64(conn->oic_last_attempt +
 328                                        RECONNECT_INTERVAL * HZ,
 329                                        cfs_time_current_64())) {
 330                         /* If we have never tried this connection since the
 331                            the last successful attempt, go with this one */
 332                         if (cfs_time_beforeq_64(conn->oic_last_attempt,
 333                                                imp->imp_last_success_conn)) {
 334                                 imp_conn = conn;
 335                                 break;
 336                         }
 337
 338                         /* Both of these connections have already been tried
 339                            since the last successful connection; just choose the
 340                            least recently used */
 341                         if (!imp_conn)
 342                                 imp_conn = conn;
 343                         else if (cfs_time_before_64(conn->oic_last_attempt,
 344                                                     imp_conn->oic_last_attempt))
 345                                 imp_conn = conn;
 346                 }
 347         }
 348
 349         /* if not found, simply choose the current one */
 350         if (!imp_conn) {
 351                 LASSERT(imp->imp_conn_current);
 352                 imp_conn = imp->imp_conn_current;
 353         }
 354         LASSERT(imp_conn->oic_conn);
 355
 356         imp_conn->oic_last_attempt = cfs_time_current_64();
 357
 358         /* switch connection, don't mind if it's same as the current one */
 359         if (imp->imp_connection)
 360                 ptlrpc_put_connection(imp->imp_connection);
 361         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 362
 363         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 364         LASSERT(dlmexp != NULL);
 365         if (dlmexp->exp_connection)
 366                 ptlrpc_put_connection(dlmexp->exp_connection);
 367         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 368         class_export_put(dlmexp);
 369
 370         if (imp->imp_conn_current != imp_conn) {
 371                 if (imp->imp_conn_current)
 372                         LCONSOLE_INFO("Changing connection for %s to %s/%s\n",
 373                                       imp->imp_obd->obd_name,
 374                                       imp_conn->oic_uuid.uuid,
 375                                       libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 376                 imp->imp_conn_current = imp_conn;
 377         }
 378
 379         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 380                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 381                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 382
 383         spin_unlock(&imp->imp_lock);
 384
 385         RETURN(0);
 386 }
 387
 388 /*
 389  * must be called under imp_lock
 390  */
 391 int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 392 {
 393         struct ptlrpc_request *req;
 394         struct list_head *tmp;
 395
 396         if (list_empty(&imp->imp_replay_list))
 397                 return 0;
 398         tmp = imp->imp_replay_list.next;
 399         req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 400         *transno = req->rq_transno;
 401         if (req->rq_transno == 0) {
 402                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 403                 LBUG();
 404         }
 405
 406         return 1;
 407 }
 408
 409 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
 410 {
 411         struct obd_device *obd = imp->imp_obd;
 412         int initial_connect = 0;
 413         int set_transno = 0;
 414         __u64 committed_before_reconnect = 0;
 415         struct ptlrpc_request *request;
 416         char *bufs[] = { NULL,
 417                          obd2cli_tgt(imp->imp_obd),
 418                          obd->obd_uuid.uuid,
 419                          (char *)&imp->imp_dlm_handle,
 420                          (char *)&imp->imp_connect_data };
 421         struct ptlrpc_connect_async_args *aa;
 422         int rc;
 423         ENTRY;
 424
 425         spin_lock(&imp->imp_lock);
 426         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 427                 spin_unlock(&imp->imp_lock);
 428                 CERROR("can't connect to a closed import\n");
 429                 RETURN(-EINVAL);
 430         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 431                 spin_unlock(&imp->imp_lock);
 432                 CERROR("already connected\n");
 433                 RETURN(0);
 434         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 435                 spin_unlock(&imp->imp_lock);
 436                 CERROR("already connecting\n");
 437                 RETURN(-EALREADY);
 438         }
 439
 440         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 441
 442         imp->imp_conn_cnt++;
 443         imp->imp_resend_replay = 0;
 444
 445         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 446                 initial_connect = 1;
 447         else
 448                 committed_before_reconnect = imp->imp_peer_committed_transno;
 449
 450         set_transno = ptlrpc_first_transno(imp, &imp->imp_connect_data.ocd_transno);
 451         spin_unlock(&imp->imp_lock);
 452
 453         if (new_uuid) {
 454                 struct obd_uuid uuid;
 455
 456                 obd_str2uuid(&uuid, new_uuid);
 457                 rc = import_set_conn_priority(imp, &uuid);
 458                 if (rc)
 459                         GOTO(out, rc);
 460         }
 461
 462         rc = import_select_connection(imp);
 463         if (rc)
 464                 GOTO(out, rc);
 465
 466         /* last in connection list */
 467         if (imp->imp_conn_current->oic_item.next == &imp->imp_conn_list) {
 468                 if (imp->imp_initial_recov_bk && initial_connect) {
 469                         CDEBUG(D_HA, "Last connection attempt (%d) for %s\n",
 470                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 471                         /* Don't retry if connect fails */
 472                         rc = 0;
 473                         obd_set_info_async(obd->obd_self_export,
 474                                            strlen(KEY_INIT_RECOV),
 475                                            KEY_INIT_RECOV,
 476                                            sizeof(rc), &rc, NULL);
 477                 }
 478                 if (imp->imp_recon_bk) {
 479                         CDEBUG(D_HA, "Last reconnection attempt (%d) for %s\n",
 480                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 481                         spin_lock(&imp->imp_lock);
 482                         imp->imp_last_recon = 1;
 483                         spin_unlock(&imp->imp_lock);
 484                 }
 485         }
 486
 487         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 488         if (rc)
 489                 GOTO(out, rc);
 490
 491         /* Reset connect flags to the originally requested flags, in case
 492          * the server is updated on-the-fly we will get the new features. */
 493         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 494         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 495                            &obd->obd_uuid, &imp->imp_connect_data);
 496         if (rc)
 497                 GOTO(out, rc);
 498
 499         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 500         if (request == NULL)
 501                 GOTO(out, rc = -ENOMEM);
 502
 503         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 504                                       imp->imp_connect_op, bufs, NULL);
 505         if (rc) {
 506                 ptlrpc_request_free(request);
 507                 GOTO(out, rc);
 508         }
 509
 510 #ifndef __KERNEL__
 511         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 512 #endif
 513         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 514
 515         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 516         /* Allow a slightly larger reply for future growth compatibility */
 517         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 518                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 519         ptlrpc_request_set_replen(request);
 520         request->rq_interpret_reply = ptlrpc_connect_interpret;
 521
 522         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 523         aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
 524         memset(aa, 0, sizeof *aa);
 525
 526         aa->pcaa_peer_committed = committed_before_reconnect;
 527         aa->pcaa_initial_connect = initial_connect;
 528
 529         if (aa->pcaa_initial_connect) {
 530                 spin_lock(&imp->imp_lock);
 531                 imp->imp_replayable = 1;
 532                 spin_unlock(&imp->imp_lock);
 533                 /* On an initial connect, we don't know which one of a
 534                    failover server pair is up.  Don't wait long. */
 535 #ifdef CRAY_XT3
 536                 request->rq_timeout = max((int)(obd_timeout / 2), 5);
 537 #else
 538                 request->rq_timeout = max((int)(obd_timeout / 20), 5);
 539 #endif
 540                 lustre_msg_add_op_flags(request->rq_reqmsg,
 541                                         MSG_CONNECT_INITIAL);
 542         }
 543
 544         if (set_transno)
 545                 lustre_msg_add_op_flags(request->rq_reqmsg,
 546                                         MSG_CONNECT_TRANSNO);
 547
 548         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request");
 549         ptlrpcd_add_req(request);
 550         rc = 0;
 551 out:
 552         if (rc != 0) {
 553                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 554         }
 555
 556         RETURN(rc);
 557 }
 558 EXPORT_SYMBOL(ptlrpc_connect_import);
 559
 560 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 561 {
 562 #ifdef __KERNEL__
 563         struct obd_import_conn *imp_conn;
 564 #endif
 565         int wake_pinger = 0;
 566
 567         ENTRY;
 568
 569         spin_lock(&imp->imp_lock);
 570         if (list_empty(&imp->imp_conn_list))
 571                 GOTO(unlock, 0);
 572
 573 #ifdef __KERNEL__
 574         imp_conn = list_entry(imp->imp_conn_list.prev,
 575                               struct obd_import_conn,
 576                               oic_item);
 577
 578         /* XXX: When the failover node is the primary node, it is possible
 579          * to have two identical connections in imp_conn_list. We must
 580          * compare not conn's pointers but NIDs, otherwise we can defeat
 581          * connection throttling. (See bug 14774.) */
 582         if (imp->imp_conn_current->oic_conn->c_self !=
 583                                 imp_conn->oic_conn->c_self) {
 584                 ptlrpc_ping_import_soon(imp);
 585                 wake_pinger = 1;
 586         }
 587 #else
 588         /* liblustre has no pinger thead, so we wakup pinger anyway */
 589         wake_pinger = 1;
 590 #endif
 591
 592  unlock:
 593         spin_unlock(&imp->imp_lock);
 594
 595         if (wake_pinger)
 596                 ptlrpc_pinger_wake_up();
 597
 598         EXIT;
 599 }
 600
 601 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
 602                                     void * data, int rc)
 603 {
 604         struct ptlrpc_connect_async_args *aa = data;
 605         struct obd_import *imp = request->rq_import;
 606         struct client_obd *cli = &imp->imp_obd->u.cli;
 607         struct lustre_handle old_hdl;
 608         int msg_flags;
 609         ENTRY;
 610
 611         spin_lock(&imp->imp_lock);
 612         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 613                 spin_unlock(&imp->imp_lock);
 614                 RETURN(0);
 615         }
 616         spin_unlock(&imp->imp_lock);
 617
 618         if (rc)
 619                 GOTO(out, rc);
 620
 621         LASSERT(imp->imp_conn_current);
 622
 623         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 624
 625         /* All imports are pingable */
 626         spin_lock(&imp->imp_lock);
 627         imp->imp_pingable = 1;
 628
 629         if (aa->pcaa_initial_connect) {
 630                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 631                         imp->imp_replayable = 1;
 632                         spin_unlock(&imp->imp_lock);
 633                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 634                                obd2cli_tgt(imp->imp_obd));
 635                 } else {
 636                         imp->imp_replayable = 0;
 637                         spin_unlock(&imp->imp_lock);
 638                 }
 639
 640                 /* if applies, adjust the imp->imp_msg_magic here
 641                  * according to reply flags */
 642
 643                 imp->imp_remote_handle =
 644                                 *lustre_msg_get_handle(request->rq_repmsg);
 645
 646                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 647                 ptlrpc_activate_import(imp);
 648                 GOTO(finish, rc = 0);
 649         } else {
 650                 spin_unlock(&imp->imp_lock);
 651         }
 652
 653         /* Determine what recovery state to move the import to. */
 654         if (MSG_CONNECT_RECONNECT & msg_flags) {
 655                 memset(&old_hdl, 0, sizeof(old_hdl));
 656                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 657                             sizeof (old_hdl))) {
 658                         CERROR("%s@%s didn't like our handle "LPX64
 659                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 660                                imp->imp_connection->c_remote_uuid.uuid,
 661                                imp->imp_dlm_handle.cookie);
 662                         GOTO(out, rc = -ENOTCONN);
 663                 }
 664
 665                 if (memcmp(&imp->imp_remote_handle,
 666                            lustre_msg_get_handle(request->rq_repmsg),
 667                            sizeof(imp->imp_remote_handle))) {
 668
 669                         CWARN("%s@%s changed server handle from "
 670                                LPX64" to "LPX64" - evicting.\n",
 671                                obd2cli_tgt(imp->imp_obd),
 672                                imp->imp_connection->c_remote_uuid.uuid,
 673                                imp->imp_remote_handle.cookie,
 674                                lustre_msg_get_handle(request->rq_repmsg)->
 675                                          cookie);
 676                         imp->imp_remote_handle =
 677                                      *lustre_msg_get_handle(request->rq_repmsg);
 678
 679                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 680                         GOTO(finish, rc = 0);
 681                 } else {
 682                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 683                                obd2cli_tgt(imp->imp_obd),
 684                                imp->imp_connection->c_remote_uuid.uuid);
 685                 }
 686
 687                 if (imp->imp_invalid) {
 688                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 689                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 690                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 691                                imp->imp_obd->obd_name,
 692                                obd2cli_tgt(imp->imp_obd));
 693
 694                         spin_lock(&imp->imp_lock);
 695                         imp->imp_resend_replay = 1;
 696                         spin_unlock(&imp->imp_lock);
 697
 698                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 699                 } else {
 700                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 701                 }
 702         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 703                 LASSERT(imp->imp_replayable);
 704                 imp->imp_remote_handle =
 705                                 *lustre_msg_get_handle(request->rq_repmsg);
 706                 imp->imp_last_replay_transno = 0;
 707                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 708         } else {
 709                 imp->imp_remote_handle =
 710                                 *lustre_msg_get_handle(request->rq_repmsg);
 711                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 712         }
 713
 714         /* Sanity checks for a reconnected import. */
 715         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 716                 CERROR("imp_replayable flag does not match server "
 717                        "after reconnect. We should LBUG right here.\n");
 718         }
 719
 720         if (lustre_msg_get_last_committed(request->rq_repmsg) <
 721             aa->pcaa_peer_committed) {
 722                 CERROR("%s went back in time (transno "LPD64
 723                        " was previously committed, server now claims "LPD64
 724                        ")!  See https://bugzilla.clusterfs.com/"
 725                        "long_list.cgi?buglist=9646\n",
 726                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 727                        lustre_msg_get_last_committed(request->rq_repmsg));
 728         }
 729
 730 finish:
 731         rc = ptlrpc_import_recovery_state_machine(imp);
 732         if (rc != 0) {
 733                 if (rc == -ENOTCONN) {
 734                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 735                                "invalidating and reconnecting\n",
 736                                obd2cli_tgt(imp->imp_obd),
 737                                imp->imp_connection->c_remote_uuid.uuid);
 738                         ptlrpc_connect_import(imp, NULL);
 739                         RETURN(0);
 740                 }
 741         } else {
 742                 struct obd_connect_data *ocd;
 743                 struct obd_export *exp;
 744                 int ret;
 745                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 746                                            RCL_SERVER);
 747                 /* server replied obd_connect_data is always bigger */
 748                 ocd = req_capsule_server_sized_get(&request->rq_pill,
 749                                                    &RMF_CONNECT_DATA, ret);
 750
 751                 spin_lock(&imp->imp_lock);
 752                 list_del(&imp->imp_conn_current->oic_item);
 753                 list_add(&imp->imp_conn_current->oic_item, &imp->imp_conn_list);
 754                 imp->imp_last_success_conn =
 755                         imp->imp_conn_current->oic_last_attempt;
 756
 757                 if (ocd == NULL) {
 758                         spin_unlock(&imp->imp_lock);
 759                         CERROR("Wrong connect data from server\n");
 760                         rc = -EPROTO;
 761                         GOTO(out, rc);
 762                 }
 763
 764                 imp->imp_connect_data = *ocd;
 765
 766                 exp = class_conn2export(&imp->imp_dlm_handle);
 767                 spin_unlock(&imp->imp_lock);
 768
 769                 /* check that server granted subset of flags we asked for. */
 770                 LASSERTF((ocd->ocd_connect_flags &
 771                           imp->imp_connect_flags_orig) ==
 772                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 773                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 774
 775                 if (!exp) {
 776                         /* This could happen if export is cleaned during the
 777                            connect attempt */
 778                         CERROR("Missing export for %s\n",
 779                                imp->imp_obd->obd_name);
 780                         GOTO(out, rc = -ENODEV);
 781                 }
 782                 exp->exp_connect_flags = ocd->ocd_connect_flags;
 783                 imp->imp_obd->obd_self_export->exp_connect_flags =
 784                                                         ocd->ocd_connect_flags;
 785                 class_export_put(exp);
 786
 787                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 788
 789                 if (!ocd->ocd_ibits_known &&
 790                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 791                         CERROR("Inodebits aware server returned zero compatible"
 792                                " bits?\n");
 793
 794                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 795                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
 796                                         LUSTRE_VERSION_OFFSET_WARN ||
 797                      ocd->ocd_version < LUSTRE_VERSION_CODE -
 798                                         LUSTRE_VERSION_OFFSET_WARN)) {
 799                         /* Sigh, some compilers do not like #ifdef in the middle
 800                            of macro arguments */
 801 #ifdef __KERNEL__
 802                         const char *older =
 803                                 "older. Consider upgrading this client";
 804 #else
 805                         const char *older =
 806                                 "older. Consider recompiling this application";
 807 #endif
 808                         const char *newer = "newer than client version";
 809
 810                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 811                                       "is much %s (%s)\n",
 812                                       obd2cli_tgt(imp->imp_obd),
 813                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 814                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 815                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 816                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
 817                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
 818                                       newer : older, LUSTRE_VERSION_STRING);
 819                 }
 820
 821                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
 822                         /* We sent to the server ocd_cksum_types with bits set
 823                          * for algorithms we understand. The server masked off
 824                          * the checksum types it doesn't support */
 825                         if ((ocd->ocd_cksum_types & OBD_CKSUM_ALL) == 0) {
 826                                 LCONSOLE_WARN("The negotiation of the checksum "
 827                                               "alogrithm to use with server %s "
 828                                               "failed (%x/%x), disabling "
 829                                               "checksums\n",
 830                                               obd2cli_tgt(imp->imp_obd),
 831                                               ocd->ocd_cksum_types,
 832                                               OBD_CKSUM_ALL);
 833                                 cli->cl_checksum = 0;
 834                                 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 835                                 cli->cl_cksum_type = OBD_CKSUM_CRC32;
 836                         } else {
 837                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
 838
 839                                 if (ocd->ocd_cksum_types & OSC_DEFAULT_CKSUM)
 840                                         cli->cl_cksum_type = OSC_DEFAULT_CKSUM;
 841                                 else if (ocd->ocd_cksum_types & OBD_CKSUM_ADLER)
 842                                         cli->cl_cksum_type = OBD_CKSUM_ADLER;
 843                                 else
 844                                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
 845                         }
 846                 } else {
 847                         /* The server does not support OBD_CONNECT_CKSUM.
 848                          * Enforce CRC32 for backward compatibility*/
 849                         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 850                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
 851                 }
 852
 853                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
 854                         cli->cl_max_pages_per_rpc =
 855                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
 856                 }
 857
 858                 imp->imp_obd->obd_namespace->ns_connect_flags =
 859                                                         ocd->ocd_connect_flags;
 860                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
 861                                                         ocd->ocd_connect_flags;
 862
 863                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 864                         (cli->cl_max_pages_per_rpc > 0));
 865         }
 866
 867 out:
 868         if (rc != 0) {
 869                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 870                 spin_lock(&imp->imp_lock);
 871                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov &&
 872                     (request->rq_import_generation == imp->imp_generation))
 873                         ptlrpc_deactivate_and_unlock_import(imp);
 874                 else
 875                         spin_unlock(&imp->imp_lock);
 876
 877                 if ((imp->imp_recon_bk && imp->imp_last_recon) ||
 878                     (rc == -EACCES)) {
 879                         /*
 880                          * Give up trying to reconnect
 881                          * EACCES means client has no permission for connection
 882                          */
 883                         imp->imp_obd->obd_no_recov = 1;
 884                         ptlrpc_deactivate_import(imp);
 885                 }
 886
 887                 if (rc == -EPROTO) {
 888                         struct obd_connect_data *ocd;
 889
 890                         /* reply message might not be ready */
 891                         if (request->rq_repmsg == NULL)
 892                                 RETURN(-EPROTO);
 893
 894                         ocd = req_capsule_server_get(&request->rq_pill,
 895                                                      &RMF_CONNECT_DATA);
 896                         if (ocd &&
 897                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 898                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
 899                            /* Actually servers are only supposed to refuse
 900                               connection from liblustre clients, so we should
 901                               never see this from VFS context */
 902                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
 903                                         "(%d.%d.%d.%d)"
 904                                         " refused connection from this client "
 905                                         "with an incompatible version (%s).  "
 906                                         "Client must be recompiled\n",
 907                                         obd2cli_tgt(imp->imp_obd),
 908                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 909                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 910                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 911                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
 912                                         LUSTRE_VERSION_STRING);
 913                                 ptlrpc_deactivate_import(imp);
 914                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
 915                         }
 916                         RETURN(-EPROTO);
 917                 }
 918
 919                 ptlrpc_maybe_ping_import_soon(imp);
 920
 921                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
 922                        obd2cli_tgt(imp->imp_obd),
 923                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
 924         }
 925
 926         spin_lock(&imp->imp_lock);
 927         imp->imp_last_recon = 0;
 928         spin_unlock(&imp->imp_lock);
 929
 930         cfs_waitq_signal(&imp->imp_recovery_waitq);
 931         RETURN(rc);
 932 }
 933
 934 static int completed_replay_interpret(struct ptlrpc_request *req,
 935                                     void * data, int rc)
 936 {
 937         ENTRY;
 938         atomic_dec(&req->rq_import->imp_replay_inflight);
 939         if (req->rq_status == 0) {
 940                 ptlrpc_import_recovery_state_machine(req->rq_import);
 941         } else {
 942                 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
 943                        "reconnecting\n",
 944                        req->rq_import->imp_obd->obd_name, req->rq_status);
 945                 ptlrpc_connect_import(req->rq_import, NULL);
 946         }
 947
 948         RETURN(0);
 949 }
 950
 951 static int signal_completed_replay(struct obd_import *imp)
 952 {
 953         struct ptlrpc_request *req;
 954         ENTRY;
 955
 956         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 957         atomic_inc(&imp->imp_replay_inflight);
 958
 959         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
 960                                         OBD_PING);
 961         if (req == NULL) {
 962                 atomic_dec(&imp->imp_replay_inflight);
 963                 RETURN(-ENOMEM);
 964         }
 965
 966         ptlrpc_request_set_replen(req);
 967         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
 968         lustre_msg_add_flags(req->rq_reqmsg,
 969                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
 970         req->rq_timeout *= 3;
 971         req->rq_interpret_reply = completed_replay_interpret;
 972
 973         ptlrpcd_add_req(req);
 974         RETURN(0);
 975 }
 976
 977 #ifdef __KERNEL__
 978 static int ptlrpc_invalidate_import_thread(void *data)
 979 {
 980         struct obd_import *imp = data;
 981
 982         ENTRY;
 983
 984         ptlrpc_daemonize("ll_imp_inval");
 985
 986         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
 987                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 988                imp->imp_connection->c_remote_uuid.uuid);
 989
 990         ptlrpc_invalidate_import(imp);
 991
 992         if (obd_dump_on_eviction) {
 993                 CERROR("dump the log upon eviction\n");
 994                 libcfs_debug_dumplog();
 995         }
 996
 997         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 998         ptlrpc_import_recovery_state_machine(imp);
 999
1000         RETURN(0);
1001 }
1002 #endif
1003
1004 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1005 {
1006         int rc = 0;
1007         int inflight;
1008         char *target_start;
1009         int target_len;
1010
1011         ENTRY;
1012         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1013                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1014                           &target_start, &target_len);
1015                 /* Don't care about MGC eviction */
1016                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1017                            LUSTRE_MGC_NAME) != 0) {
1018                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1019                                            "%.*s; in progress operations using "
1020                                            "this service will fail.\n",
1021                                            target_len, target_start);
1022                 }
1023                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1024                        obd2cli_tgt(imp->imp_obd),
1025                        imp->imp_connection->c_remote_uuid.uuid);
1026
1027 #ifdef __KERNEL__
1028                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
1029                                        CLONE_VM | CLONE_FILES);
1030                 if (rc < 0)
1031                         CERROR("error starting invalidate thread: %d\n", rc);
1032                 else
1033                         rc = 0;
1034                 RETURN(rc);
1035 #else
1036                 ptlrpc_invalidate_import(imp);
1037
1038                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1039 #endif
1040         }
1041
1042         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1043                 CDEBUG(D_HA, "replay requested by %s\n",
1044                        obd2cli_tgt(imp->imp_obd));
1045                 rc = ptlrpc_replay_next(imp, &inflight);
1046                 if (inflight == 0 &&
1047                     atomic_read(&imp->imp_replay_inflight) == 0) {
1048                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1049                         rc = ldlm_replay_locks(imp);
1050                         if (rc)
1051                                 GOTO(out, rc);
1052                 }
1053                 rc = 0;
1054         }
1055
1056         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1057                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1058                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1059                         rc = signal_completed_replay(imp);
1060                         if (rc)
1061                                 GOTO(out, rc);
1062                 }
1063
1064         }
1065
1066         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1067                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1068                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1069                 }
1070         }
1071
1072         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1073                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1074                        obd2cli_tgt(imp->imp_obd),
1075                        imp->imp_connection->c_remote_uuid.uuid);
1076
1077                 rc = ptlrpc_resend(imp);
1078                 if (rc)
1079                         GOTO(out, rc);
1080                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1081                 ptlrpc_activate_import(imp);
1082
1083                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1084                           &target_start, &target_len);
1085                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1086                               "using nid %s.\n", imp->imp_obd->obd_name,
1087                               target_len, target_start,
1088                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1089         }
1090
1091         if (imp->imp_state == LUSTRE_IMP_FULL) {
1092                 cfs_waitq_signal(&imp->imp_recovery_waitq);
1093                 ptlrpc_wake_delayed(imp);
1094         }
1095
1096 out:
1097         RETURN(rc);
1098 }
1099
1100 static int back_to_sleep(void *unused)
1101 {
1102         return 0;
1103 }
1104
1105 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1106 {
1107         struct ptlrpc_request *req;
1108         int rq_opc, rc = 0;
1109         int nowait = imp->imp_obd->obd_force;
1110         ENTRY;
1111
1112         if (nowait)
1113                 GOTO(set_state, rc);
1114
1115         switch (imp->imp_connect_op) {
1116         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1117         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1118         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1119         default:
1120                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1121                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1122                 RETURN(-EINVAL);
1123         }
1124
1125         if (ptlrpc_import_in_recovery(imp)) {
1126                 struct l_wait_info lwi;
1127                 cfs_duration_t timeout;
1128                 if (imp->imp_server_timeout)
1129                         timeout = cfs_time_seconds(obd_timeout / 2);
1130                 else
1131                         timeout = cfs_time_seconds(obd_timeout);
1132
1133                 timeout = MAX(timeout * HZ, 1);
1134
1135                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1136                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1137                 rc = l_wait_event(imp->imp_recovery_waitq,
1138                                   !ptlrpc_import_in_recovery(imp), &lwi);
1139
1140         }
1141
1142         spin_lock(&imp->imp_lock);
1143         if (imp->imp_state != LUSTRE_IMP_FULL)
1144                 GOTO(out, 0);
1145
1146         spin_unlock(&imp->imp_lock);
1147
1148         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1149                                         LUSTRE_OBD_VERSION, rq_opc);
1150         if (req) {
1151                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1152                  * it fails.  We can get through the above with a down server
1153                  * if the client doesn't know the server is gone yet. */
1154                 req->rq_no_resend = 1;
1155 #ifdef CRAY_XT3
1156                 req->rq_timeout = obd_timeout / 3;
1157 #else
1158                 req->rq_timeout = 5;
1159 #endif
1160                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1161                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1162                 ptlrpc_request_set_replen(req);
1163                 rc = ptlrpc_queue_wait(req);
1164                 ptlrpc_req_finished(req);
1165         }
1166
1167 set_state:
1168         spin_lock(&imp->imp_lock);
1169 out:
1170         if (noclose)
1171                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1172         else
1173                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1174         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1175         imp->imp_conn_cnt = 0;
1176         imp->imp_last_recon = 0;
1177         spin_unlock(&imp->imp_lock);
1178
1179         RETURN(rc);
1180 }
1181