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         /*
 214          * If this is an invalid MGC connection, then don't bother
 215          * waiting for imp_inflight to drop to 0.
 216          */
 217         if (imp->imp_invalid && imp->imp_recon_bk && !imp->imp_obd->obd_no_recov)
 218                 goto out;
 219
 220         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 221                 ptlrpc_deactivate_import(imp);
 222
 223         LASSERT(imp->imp_invalid);
 224
 225         /* wait for all requests to error out and call completion callbacks */
 226         lwi = LWI_TIMEOUT_INTERVAL(cfs_timeout_cap(cfs_time_seconds(obd_timeout)),
 227                                    HZ, NULL, NULL);
 228         rc = l_wait_event(imp->imp_recovery_waitq,
 229                           (atomic_read(&imp->imp_inflight) == 0), &lwi);
 230
 231         if (rc) {
 232                 struct list_head *tmp, *n;
 233                 struct ptlrpc_request *req;
 234
 235                 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 236                        obd2cli_tgt(imp->imp_obd), rc,
 237                        atomic_read(&imp->imp_inflight));
 238                 spin_lock(&imp->imp_lock);
 239                 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 240                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
 241                         DEBUG_REQ(D_ERROR, req, "still on sending list");
 242                 }
 243                 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
 244                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
 245                         DEBUG_REQ(D_ERROR, req, "still on delayed list");
 246                 }
 247                 spin_unlock(&imp->imp_lock);
 248 //                LASSERT(atomic_read(&imp->imp_inflight) == 0);
 249         }
 250
 251 out:
 252         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 253         sptlrpc_import_flush_all_ctx(imp);
 254
 255         atomic_dec(&imp->imp_inval_count);
 256         cfs_waitq_signal(&imp->imp_recovery_waitq);
 257 }
 258
 259 /* unset imp_invalid */
 260 void ptlrpc_activate_import(struct obd_import *imp)
 261 {
 262         struct obd_device *obd = imp->imp_obd;
 263
 264         spin_lock(&imp->imp_lock);
 265         imp->imp_invalid = 0;
 266         ptlrpc_activate_timeouts(imp);
 267         spin_unlock(&imp->imp_lock);
 268         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 269 }
 270
 271 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 272 {
 273         ENTRY;
 274
 275         LASSERT(!imp->imp_dlm_fake);
 276
 277         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 278                 if (!imp->imp_replayable) {
 279                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 280                                "auto-deactivating\n",
 281                                obd2cli_tgt(imp->imp_obd),
 282                                imp->imp_connection->c_remote_uuid.uuid,
 283                                imp->imp_obd->obd_name);
 284                         ptlrpc_deactivate_import(imp);
 285                 }
 286
 287                 CDEBUG(D_HA, "%s: waking up pinger\n",
 288                        obd2cli_tgt(imp->imp_obd));
 289
 290                 spin_lock(&imp->imp_lock);
 291                 imp->imp_force_verify = 1;
 292                 spin_unlock(&imp->imp_lock);
 293
 294                 ptlrpc_pinger_wake_up();
 295         }
 296         EXIT;
 297 }
 298
 299 int ptlrpc_reconnect_import(struct obd_import *imp)
 300 {
 301
 302         ptlrpc_set_import_discon(imp, 0);
 303         /* Force a new connect attempt */
 304         ptlrpc_invalidate_import(imp);
 305         /* Do a fresh connect next time by zeroing the handle */
 306         ptlrpc_disconnect_import(imp, 1);
 307         /* Wait for all invalidate calls to finish */
 308         if (atomic_read(&imp->imp_inval_count) > 0) {
 309                 int rc;
 310                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 311                 rc = l_wait_event(imp->imp_recovery_waitq,
 312                                   (atomic_read(&imp->imp_inval_count) == 0),
 313                                   &lwi);
 314                 if (rc)
 315                         CERROR("Interrupted, inval=%d\n",
 316                                atomic_read(&imp->imp_inval_count));
 317         }
 318
 319         /* Allow reconnect attempts */
 320         imp->imp_obd->obd_no_recov = 0;
 321         /* Remove 'invalid' flag */
 322         ptlrpc_activate_import(imp);
 323         /* Attempt a new connect */
 324         ptlrpc_recover_import(imp, NULL);
 325         return 0;
 326 }
 327
 328 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 329
 330 static int import_select_connection(struct obd_import *imp)
 331 {
 332         struct obd_import_conn *imp_conn = NULL, *conn;
 333         struct obd_export *dlmexp;
 334         ENTRY;
 335
 336         spin_lock(&imp->imp_lock);
 337
 338         if (list_empty(&imp->imp_conn_list)) {
 339                 CERROR("%s: no connections available\n",
 340                         imp->imp_obd->obd_name);
 341                 spin_unlock(&imp->imp_lock);
 342                 RETURN(-EINVAL);
 343         }
 344
 345         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 346                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 347                        imp->imp_obd->obd_name,
 348                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 349                        conn->oic_last_attempt);
 350                 /* Throttle the reconnect rate to once per RECONNECT_INTERVAL */
 351                 if (cfs_time_before_64(conn->oic_last_attempt +
 352                                        RECONNECT_INTERVAL * HZ,
 353                                        cfs_time_current_64())) {
 354                         /* If we have never tried this connection since the
 355                            the last successful attempt, go with this one */
 356                         if (cfs_time_beforeq_64(conn->oic_last_attempt,
 357                                                imp->imp_last_success_conn)) {
 358                                 imp_conn = conn;
 359                                 break;
 360                         }
 361
 362                         /* Both of these connections have already been tried
 363                            since the last successful connection; just choose the
 364                            least recently used */
 365                         if (!imp_conn)
 366                                 imp_conn = conn;
 367                         else if (cfs_time_before_64(conn->oic_last_attempt,
 368                                                     imp_conn->oic_last_attempt))
 369                                 imp_conn = conn;
 370                 }
 371         }
 372
 373         /* if not found, simply choose the current one */
 374         if (!imp_conn) {
 375                 LASSERT(imp->imp_conn_current);
 376                 imp_conn = imp->imp_conn_current;
 377         }
 378         LASSERT(imp_conn->oic_conn);
 379
 380         imp_conn->oic_last_attempt = cfs_time_current_64();
 381
 382         /* switch connection, don't mind if it's same as the current one */
 383         if (imp->imp_connection)
 384                 ptlrpc_put_connection(imp->imp_connection);
 385         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 386
 387         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 388         LASSERT(dlmexp != NULL);
 389         if (dlmexp->exp_connection)
 390                 ptlrpc_put_connection(dlmexp->exp_connection);
 391         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 392         class_export_put(dlmexp);
 393
 394         if (imp->imp_conn_current != imp_conn) {
 395                 if (imp->imp_conn_current)
 396                         LCONSOLE_INFO("Changing connection for %s to %s/%s\n",
 397                                       imp->imp_obd->obd_name,
 398                                       imp_conn->oic_uuid.uuid,
 399                                       libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 400                 imp->imp_conn_current = imp_conn;
 401         }
 402
 403         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 404                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 405                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 406
 407         spin_unlock(&imp->imp_lock);
 408
 409         RETURN(0);
 410 }
 411
 412 /*
 413  * must be called under imp_lock
 414  */
 415 int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 416 {
 417         struct ptlrpc_request *req;
 418         struct list_head *tmp;
 419
 420         if (list_empty(&imp->imp_replay_list))
 421                 return 0;
 422         tmp = imp->imp_replay_list.next;
 423         req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 424         *transno = req->rq_transno;
 425         if (req->rq_transno == 0) {
 426                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 427                 LBUG();
 428         }
 429
 430         return 1;
 431 }
 432
 433 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
 434 {
 435         struct obd_device *obd = imp->imp_obd;
 436         int initial_connect = 0;
 437         int set_transno = 0;
 438         __u64 committed_before_reconnect = 0;
 439         struct ptlrpc_request *request;
 440         char *bufs[] = { NULL,
 441                          obd2cli_tgt(imp->imp_obd),
 442                          obd->obd_uuid.uuid,
 443                          (char *)&imp->imp_dlm_handle,
 444                          (char *)&imp->imp_connect_data };
 445         struct ptlrpc_connect_async_args *aa;
 446         int rc;
 447         ENTRY;
 448
 449         spin_lock(&imp->imp_lock);
 450         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 451                 spin_unlock(&imp->imp_lock);
 452                 CERROR("can't connect to a closed import\n");
 453                 RETURN(-EINVAL);
 454         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 455                 spin_unlock(&imp->imp_lock);
 456                 CERROR("already connected\n");
 457                 RETURN(0);
 458         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 459                 spin_unlock(&imp->imp_lock);
 460                 CERROR("already connecting\n");
 461                 RETURN(-EALREADY);
 462         }
 463
 464         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 465
 466         imp->imp_conn_cnt++;
 467         imp->imp_resend_replay = 0;
 468
 469         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 470                 initial_connect = 1;
 471         else
 472                 committed_before_reconnect = imp->imp_peer_committed_transno;
 473
 474         set_transno = ptlrpc_first_transno(imp, &imp->imp_connect_data.ocd_transno);
 475         spin_unlock(&imp->imp_lock);
 476
 477         if (new_uuid) {
 478                 struct obd_uuid uuid;
 479
 480                 obd_str2uuid(&uuid, new_uuid);
 481                 rc = import_set_conn_priority(imp, &uuid);
 482                 if (rc)
 483                         GOTO(out, rc);
 484         }
 485
 486         rc = import_select_connection(imp);
 487         if (rc)
 488                 GOTO(out, rc);
 489
 490         /* last in connection list */
 491         if (imp->imp_conn_current->oic_item.next == &imp->imp_conn_list) {
 492                 if (imp->imp_initial_recov_bk && initial_connect) {
 493                         CDEBUG(D_HA, "Last connection attempt (%d) for %s\n",
 494                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 495                         /* Don't retry if connect fails */
 496                         rc = 0;
 497                         obd_set_info_async(obd->obd_self_export,
 498                                            strlen(KEY_INIT_RECOV),
 499                                            KEY_INIT_RECOV,
 500                                            sizeof(rc), &rc, NULL);
 501                 }
 502                 if (imp->imp_recon_bk) {
 503                         CDEBUG(D_HA, "Last reconnection attempt (%d) for %s\n",
 504                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 505                         spin_lock(&imp->imp_lock);
 506                         imp->imp_last_recon = 1;
 507                         spin_unlock(&imp->imp_lock);
 508                 }
 509         }
 510
 511         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 512         if (rc)
 513                 GOTO(out, rc);
 514
 515         /* Reset connect flags to the originally requested flags, in case
 516          * the server is updated on-the-fly we will get the new features. */
 517         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 518         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 519                            &obd->obd_uuid, &imp->imp_connect_data);
 520         if (rc)
 521                 GOTO(out, rc);
 522
 523         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 524         if (request == NULL)
 525                 GOTO(out, rc = -ENOMEM);
 526
 527         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 528                                       imp->imp_connect_op, bufs, NULL);
 529         if (rc) {
 530                 ptlrpc_request_free(request);
 531                 GOTO(out, rc);
 532         }
 533
 534 #ifndef __KERNEL__
 535         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 536 #endif
 537         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 538
 539         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 540         /* Allow a slightly larger reply for future growth compatibility */
 541         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 542                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 543         ptlrpc_request_set_replen(request);
 544         request->rq_interpret_reply = ptlrpc_connect_interpret;
 545
 546         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 547         aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
 548         memset(aa, 0, sizeof *aa);
 549
 550         aa->pcaa_peer_committed = committed_before_reconnect;
 551         aa->pcaa_initial_connect = initial_connect;
 552
 553         if (aa->pcaa_initial_connect) {
 554                 spin_lock(&imp->imp_lock);
 555                 imp->imp_replayable = 1;
 556                 spin_unlock(&imp->imp_lock);
 557                 /* On an initial connect, we don't know which one of a
 558                    failover server pair is up.  Don't wait long. */
 559 #ifdef CRAY_XT3
 560                 request->rq_timeout = max((int)(obd_timeout / 2), 5);
 561 #else
 562                 request->rq_timeout = max((int)(obd_timeout / 20), 5);
 563 #endif
 564                 lustre_msg_add_op_flags(request->rq_reqmsg,
 565                                         MSG_CONNECT_INITIAL);
 566         }
 567
 568         if (set_transno)
 569                 lustre_msg_add_op_flags(request->rq_reqmsg,
 570                                         MSG_CONNECT_TRANSNO);
 571
 572         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request");
 573         ptlrpcd_add_req(request);
 574         rc = 0;
 575 out:
 576         if (rc != 0) {
 577                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 578         }
 579
 580         RETURN(rc);
 581 }
 582 EXPORT_SYMBOL(ptlrpc_connect_import);
 583
 584 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 585 {
 586 #ifdef __KERNEL__
 587         struct obd_import_conn *imp_conn;
 588 #endif
 589         int wake_pinger = 0;
 590
 591         ENTRY;
 592
 593         spin_lock(&imp->imp_lock);
 594         if (list_empty(&imp->imp_conn_list))
 595                 GOTO(unlock, 0);
 596
 597 #ifdef __KERNEL__
 598         imp_conn = list_entry(imp->imp_conn_list.prev,
 599                               struct obd_import_conn,
 600                               oic_item);
 601
 602         /* XXX: When the failover node is the primary node, it is possible
 603          * to have two identical connections in imp_conn_list. We must
 604          * compare not conn's pointers but NIDs, otherwise we can defeat
 605          * connection throttling. (See bug 14774.) */
 606         if (imp->imp_conn_current->oic_conn->c_self !=
 607                                 imp_conn->oic_conn->c_self) {
 608                 ptlrpc_ping_import_soon(imp);
 609                 wake_pinger = 1;
 610         }
 611 #else
 612         /* liblustre has no pinger thead, so we wakup pinger anyway */
 613         wake_pinger = 1;
 614 #endif
 615
 616  unlock:
 617         spin_unlock(&imp->imp_lock);
 618
 619         if (wake_pinger)
 620                 ptlrpc_pinger_wake_up();
 621
 622         EXIT;
 623 }
 624
 625 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
 626                                     void * data, int rc)
 627 {
 628         struct ptlrpc_connect_async_args *aa = data;
 629         struct obd_import *imp = request->rq_import;
 630         struct client_obd *cli = &imp->imp_obd->u.cli;
 631         struct lustre_handle old_hdl;
 632         int msg_flags;
 633         ENTRY;
 634
 635         spin_lock(&imp->imp_lock);
 636         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 637                 spin_unlock(&imp->imp_lock);
 638                 RETURN(0);
 639         }
 640         spin_unlock(&imp->imp_lock);
 641
 642         if (rc)
 643                 GOTO(out, rc);
 644
 645         LASSERT(imp->imp_conn_current);
 646
 647         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 648
 649         /* All imports are pingable */
 650         spin_lock(&imp->imp_lock);
 651         imp->imp_pingable = 1;
 652
 653         if (aa->pcaa_initial_connect) {
 654                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 655                         imp->imp_replayable = 1;
 656                         spin_unlock(&imp->imp_lock);
 657                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 658                                obd2cli_tgt(imp->imp_obd));
 659                 } else {
 660                         imp->imp_replayable = 0;
 661                         spin_unlock(&imp->imp_lock);
 662                 }
 663
 664                 /* if applies, adjust the imp->imp_msg_magic here
 665                  * according to reply flags */
 666
 667                 imp->imp_remote_handle =
 668                                 *lustre_msg_get_handle(request->rq_repmsg);
 669
 670                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 671                 spin_lock(&imp->imp_lock);
 672                 if (imp->imp_invalid) {
 673                         spin_unlock(&imp->imp_lock);
 674                         ptlrpc_activate_import(imp);
 675                 } else {
 676                         spin_unlock(&imp->imp_lock);
 677                 }
 678
 679                 GOTO(finish, rc = 0);
 680         } else {
 681                 spin_unlock(&imp->imp_lock);
 682         }
 683
 684         /* Determine what recovery state to move the import to. */
 685         if (MSG_CONNECT_RECONNECT & msg_flags) {
 686                 memset(&old_hdl, 0, sizeof(old_hdl));
 687                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 688                             sizeof (old_hdl))) {
 689                         CERROR("%s@%s didn't like our handle "LPX64
 690                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 691                                imp->imp_connection->c_remote_uuid.uuid,
 692                                imp->imp_dlm_handle.cookie);
 693                         GOTO(out, rc = -ENOTCONN);
 694                 }
 695
 696                 if (memcmp(&imp->imp_remote_handle,
 697                            lustre_msg_get_handle(request->rq_repmsg),
 698                            sizeof(imp->imp_remote_handle))) {
 699
 700                         CWARN("%s@%s changed server handle from "
 701                                LPX64" to "LPX64" - evicting.\n",
 702                                obd2cli_tgt(imp->imp_obd),
 703                                imp->imp_connection->c_remote_uuid.uuid,
 704                                imp->imp_remote_handle.cookie,
 705                                lustre_msg_get_handle(request->rq_repmsg)->
 706                                          cookie);
 707                         imp->imp_remote_handle =
 708                                      *lustre_msg_get_handle(request->rq_repmsg);
 709
 710                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 711                         GOTO(finish, rc = 0);
 712                 } else {
 713                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 714                                obd2cli_tgt(imp->imp_obd),
 715                                imp->imp_connection->c_remote_uuid.uuid);
 716                 }
 717
 718                 if (imp->imp_invalid) {
 719                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 720                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 721                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 722                                imp->imp_obd->obd_name,
 723                                obd2cli_tgt(imp->imp_obd));
 724
 725                         spin_lock(&imp->imp_lock);
 726                         imp->imp_resend_replay = 1;
 727                         spin_unlock(&imp->imp_lock);
 728
 729                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 730                 } else {
 731                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 732                 }
 733         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 734                 LASSERT(imp->imp_replayable);
 735                 imp->imp_remote_handle =
 736                                 *lustre_msg_get_handle(request->rq_repmsg);
 737                 imp->imp_last_replay_transno = 0;
 738                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 739         } else {
 740                 imp->imp_remote_handle =
 741                                 *lustre_msg_get_handle(request->rq_repmsg);
 742                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 743         }
 744
 745         /* Sanity checks for a reconnected import. */
 746         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 747                 CERROR("imp_replayable flag does not match server "
 748                        "after reconnect. We should LBUG right here.\n");
 749         }
 750
 751         if (lustre_msg_get_last_committed(request->rq_repmsg) <
 752             aa->pcaa_peer_committed) {
 753                 CERROR("%s went back in time (transno "LPD64
 754                        " was previously committed, server now claims "LPD64
 755                        ")!  See https://bugzilla.clusterfs.com/"
 756                        "long_list.cgi?buglist=9646\n",
 757                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 758                        lustre_msg_get_last_committed(request->rq_repmsg));
 759         }
 760
 761 finish:
 762         rc = ptlrpc_import_recovery_state_machine(imp);
 763         if (rc != 0) {
 764                 if (rc == -ENOTCONN) {
 765                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 766                                "invalidating and reconnecting\n",
 767                                obd2cli_tgt(imp->imp_obd),
 768                                imp->imp_connection->c_remote_uuid.uuid);
 769                         ptlrpc_connect_import(imp, NULL);
 770                         RETURN(0);
 771                 }
 772         } else {
 773                 struct obd_connect_data *ocd;
 774                 struct obd_export *exp;
 775                 int ret;
 776                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 777                                            RCL_SERVER);
 778                 /* server replied obd_connect_data is always bigger */
 779                 ocd = req_capsule_server_sized_get(&request->rq_pill,
 780                                                    &RMF_CONNECT_DATA, ret);
 781
 782                 spin_lock(&imp->imp_lock);
 783                 list_del(&imp->imp_conn_current->oic_item);
 784                 list_add(&imp->imp_conn_current->oic_item, &imp->imp_conn_list);
 785                 imp->imp_last_success_conn =
 786                         imp->imp_conn_current->oic_last_attempt;
 787
 788                 if (ocd == NULL) {
 789                         spin_unlock(&imp->imp_lock);
 790                         CERROR("Wrong connect data from server\n");
 791                         rc = -EPROTO;
 792                         GOTO(out, rc);
 793                 }
 794
 795                 imp->imp_connect_data = *ocd;
 796
 797                 exp = class_conn2export(&imp->imp_dlm_handle);
 798                 spin_unlock(&imp->imp_lock);
 799
 800                 /* check that server granted subset of flags we asked for. */
 801                 LASSERTF((ocd->ocd_connect_flags &
 802                           imp->imp_connect_flags_orig) ==
 803                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 804                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 805
 806                 if (!exp) {
 807                         /* This could happen if export is cleaned during the
 808                            connect attempt */
 809                         CERROR("Missing export for %s\n",
 810                                imp->imp_obd->obd_name);
 811                         GOTO(out, rc = -ENODEV);
 812                 }
 813                 exp->exp_connect_flags = ocd->ocd_connect_flags;
 814                 imp->imp_obd->obd_self_export->exp_connect_flags =
 815                                                         ocd->ocd_connect_flags;
 816                 class_export_put(exp);
 817
 818                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 819
 820                 if (!ocd->ocd_ibits_known &&
 821                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 822                         CERROR("Inodebits aware server returned zero compatible"
 823                                " bits?\n");
 824
 825                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 826                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
 827                                         LUSTRE_VERSION_OFFSET_WARN ||
 828                      ocd->ocd_version < LUSTRE_VERSION_CODE -
 829                                         LUSTRE_VERSION_OFFSET_WARN)) {
 830                         /* Sigh, some compilers do not like #ifdef in the middle
 831                            of macro arguments */
 832 #ifdef __KERNEL__
 833                         const char *older =
 834                                 "older. Consider upgrading this client";
 835 #else
 836                         const char *older =
 837                                 "older. Consider recompiling this application";
 838 #endif
 839                         const char *newer = "newer than client version";
 840
 841                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 842                                       "is much %s (%s)\n",
 843                                       obd2cli_tgt(imp->imp_obd),
 844                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 845                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 846                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 847                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
 848                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
 849                                       newer : older, LUSTRE_VERSION_STRING);
 850                 }
 851
 852                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
 853                         /* We sent to the server ocd_cksum_types with bits set
 854                          * for algorithms we understand. The server masked off
 855                          * the checksum types it doesn't support */
 856                         if ((ocd->ocd_cksum_types & OBD_CKSUM_ALL) == 0) {
 857                                 LCONSOLE_WARN("The negotiation of the checksum "
 858                                               "alogrithm to use with server %s "
 859                                               "failed (%x/%x), disabling "
 860                                               "checksums\n",
 861                                               obd2cli_tgt(imp->imp_obd),
 862                                               ocd->ocd_cksum_types,
 863                                               OBD_CKSUM_ALL);
 864                                 cli->cl_checksum = 0;
 865                                 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 866                                 cli->cl_cksum_type = OBD_CKSUM_CRC32;
 867                         } else {
 868                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
 869
 870                                 if (ocd->ocd_cksum_types & OSC_DEFAULT_CKSUM)
 871                                         cli->cl_cksum_type = OSC_DEFAULT_CKSUM;
 872                                 else if (ocd->ocd_cksum_types & OBD_CKSUM_ADLER)
 873                                         cli->cl_cksum_type = OBD_CKSUM_ADLER;
 874                                 else
 875                                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
 876                         }
 877                 } else {
 878                         /* The server does not support OBD_CONNECT_CKSUM.
 879                          * Enforce CRC32 for backward compatibility*/
 880                         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 881                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
 882                 }
 883
 884                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
 885                         cli->cl_max_pages_per_rpc =
 886                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
 887                 }
 888
 889                 imp->imp_obd->obd_namespace->ns_connect_flags =
 890                                                         ocd->ocd_connect_flags;
 891                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
 892                                                         ocd->ocd_connect_flags;
 893
 894                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 895                         (cli->cl_max_pages_per_rpc > 0));
 896         }
 897
 898 out:
 899         if (rc != 0) {
 900                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 901                 spin_lock(&imp->imp_lock);
 902                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov &&
 903                     (request->rq_import_generation == imp->imp_generation))
 904                         ptlrpc_deactivate_and_unlock_import(imp);
 905                 else
 906                         spin_unlock(&imp->imp_lock);
 907
 908                 if ((imp->imp_recon_bk && imp->imp_last_recon) ||
 909                     (rc == -EACCES)) {
 910                         /*
 911                          * Give up trying to reconnect
 912                          * EACCES means client has no permission for connection
 913                          */
 914                         imp->imp_obd->obd_no_recov = 1;
 915                         ptlrpc_deactivate_import(imp);
 916                 }
 917
 918                 if (rc == -EPROTO) {
 919                         struct obd_connect_data *ocd;
 920
 921                         /* reply message might not be ready */
 922                         if (request->rq_repmsg == NULL)
 923                                 RETURN(-EPROTO);
 924
 925                         ocd = req_capsule_server_get(&request->rq_pill,
 926                                                      &RMF_CONNECT_DATA);
 927                         if (ocd &&
 928                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 929                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
 930                            /* Actually servers are only supposed to refuse
 931                               connection from liblustre clients, so we should
 932                               never see this from VFS context */
 933                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
 934                                         "(%d.%d.%d.%d)"
 935                                         " refused connection from this client "
 936                                         "with an incompatible version (%s).  "
 937                                         "Client must be recompiled\n",
 938                                         obd2cli_tgt(imp->imp_obd),
 939                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 940                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 941                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 942                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
 943                                         LUSTRE_VERSION_STRING);
 944                                 ptlrpc_deactivate_import(imp);
 945                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
 946                         }
 947                         RETURN(-EPROTO);
 948                 }
 949
 950                 ptlrpc_maybe_ping_import_soon(imp);
 951
 952                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
 953                        obd2cli_tgt(imp->imp_obd),
 954                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
 955         }
 956
 957         spin_lock(&imp->imp_lock);
 958         imp->imp_last_recon = 0;
 959         spin_unlock(&imp->imp_lock);
 960
 961         cfs_waitq_signal(&imp->imp_recovery_waitq);
 962         RETURN(rc);
 963 }
 964
 965 static int completed_replay_interpret(struct ptlrpc_request *req,
 966                                     void * data, int rc)
 967 {
 968         ENTRY;
 969         atomic_dec(&req->rq_import->imp_replay_inflight);
 970         if (req->rq_status == 0) {
 971                 ptlrpc_import_recovery_state_machine(req->rq_import);
 972         } else {
 973                 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
 974                        "reconnecting\n",
 975                        req->rq_import->imp_obd->obd_name, req->rq_status);
 976                 ptlrpc_connect_import(req->rq_import, NULL);
 977         }
 978
 979         RETURN(0);
 980 }
 981
 982 static int signal_completed_replay(struct obd_import *imp)
 983 {
 984         struct ptlrpc_request *req;
 985         ENTRY;
 986
 987         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 988         atomic_inc(&imp->imp_replay_inflight);
 989
 990         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
 991                                         OBD_PING);
 992         if (req == NULL) {
 993                 atomic_dec(&imp->imp_replay_inflight);
 994                 RETURN(-ENOMEM);
 995         }
 996
 997         ptlrpc_request_set_replen(req);
 998         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
 999         lustre_msg_add_flags(req->rq_reqmsg,
1000                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1001         req->rq_timeout *= 3;
1002         req->rq_interpret_reply = completed_replay_interpret;
1003
1004         ptlrpcd_add_req(req);
1005         RETURN(0);
1006 }
1007
1008 #ifdef __KERNEL__
1009 static int ptlrpc_invalidate_import_thread(void *data)
1010 {
1011         struct obd_import *imp = data;
1012
1013         ENTRY;
1014
1015         ptlrpc_daemonize("ll_imp_inval");
1016
1017         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1018                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1019                imp->imp_connection->c_remote_uuid.uuid);
1020
1021         ptlrpc_invalidate_import(imp);
1022
1023         if (obd_dump_on_eviction) {
1024                 CERROR("dump the log upon eviction\n");
1025                 libcfs_debug_dumplog();
1026         }
1027
1028         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1029         ptlrpc_import_recovery_state_machine(imp);
1030
1031         RETURN(0);
1032 }
1033 #endif
1034
1035 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1036 {
1037         int rc = 0;
1038         int inflight;
1039         char *target_start;
1040         int target_len;
1041
1042         ENTRY;
1043         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1044                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1045                           &target_start, &target_len);
1046                 /* Don't care about MGC eviction */
1047                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1048                            LUSTRE_MGC_NAME) != 0) {
1049                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1050                                            "%.*s; in progress operations using "
1051                                            "this service will fail.\n",
1052                                            target_len, target_start);
1053                 }
1054                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1055                        obd2cli_tgt(imp->imp_obd),
1056                        imp->imp_connection->c_remote_uuid.uuid);
1057
1058 #ifdef __KERNEL__
1059                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
1060                                        CLONE_VM | CLONE_FILES);
1061                 if (rc < 0)
1062                         CERROR("error starting invalidate thread: %d\n", rc);
1063                 else
1064                         rc = 0;
1065                 RETURN(rc);
1066 #else
1067                 ptlrpc_invalidate_import(imp);
1068
1069                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1070 #endif
1071         }
1072
1073         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1074                 CDEBUG(D_HA, "replay requested by %s\n",
1075                        obd2cli_tgt(imp->imp_obd));
1076                 rc = ptlrpc_replay_next(imp, &inflight);
1077                 if (inflight == 0 &&
1078                     atomic_read(&imp->imp_replay_inflight) == 0) {
1079                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1080                         rc = ldlm_replay_locks(imp);
1081                         if (rc)
1082                                 GOTO(out, rc);
1083                 }
1084                 rc = 0;
1085         }
1086
1087         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1088                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1089                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1090                         rc = signal_completed_replay(imp);
1091                         if (rc)
1092                                 GOTO(out, rc);
1093                 }
1094
1095         }
1096
1097         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1098                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1099                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1100                 }
1101         }
1102
1103         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1104                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1105                        obd2cli_tgt(imp->imp_obd),
1106                        imp->imp_connection->c_remote_uuid.uuid);
1107
1108                 rc = ptlrpc_resend(imp);
1109                 if (rc)
1110                         GOTO(out, rc);
1111                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1112                 ptlrpc_activate_import(imp);
1113
1114                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1115                           &target_start, &target_len);
1116                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1117                               "using nid %s.\n", imp->imp_obd->obd_name,
1118                               target_len, target_start,
1119                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1120         }
1121
1122         if (imp->imp_state == LUSTRE_IMP_FULL) {
1123                 cfs_waitq_signal(&imp->imp_recovery_waitq);
1124                 ptlrpc_wake_delayed(imp);
1125         }
1126
1127 out:
1128         RETURN(rc);
1129 }
1130
1131 static int back_to_sleep(void *unused)
1132 {
1133         return 0;
1134 }
1135
1136 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1137 {
1138         struct ptlrpc_request *req;
1139         int rq_opc, rc = 0;
1140         int nowait = imp->imp_obd->obd_force;
1141         ENTRY;
1142
1143         if (nowait)
1144                 GOTO(set_state, rc);
1145
1146         switch (imp->imp_connect_op) {
1147         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1148         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1149         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1150         default:
1151                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1152                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1153                 RETURN(-EINVAL);
1154         }
1155
1156         if (ptlrpc_import_in_recovery(imp)) {
1157                 struct l_wait_info lwi;
1158                 cfs_duration_t timeout;
1159                 if (imp->imp_server_timeout)
1160                         timeout = cfs_time_seconds(obd_timeout / 2);
1161                 else
1162                         timeout = cfs_time_seconds(obd_timeout);
1163
1164                 timeout = MAX(timeout * HZ, 1);
1165
1166                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1167                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1168                 rc = l_wait_event(imp->imp_recovery_waitq,
1169                                   !ptlrpc_import_in_recovery(imp), &lwi);
1170
1171         }
1172
1173         spin_lock(&imp->imp_lock);
1174         if (imp->imp_state != LUSTRE_IMP_FULL)
1175                 GOTO(out, 0);
1176
1177         spin_unlock(&imp->imp_lock);
1178
1179         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1180                                         LUSTRE_OBD_VERSION, rq_opc);
1181         if (req) {
1182                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1183                  * it fails.  We can get through the above with a down server
1184                  * if the client doesn't know the server is gone yet. */
1185                 req->rq_no_resend = 1;
1186 #ifdef CRAY_XT3
1187                 req->rq_timeout = obd_timeout / 3;
1188 #else
1189                 req->rq_timeout = 5;
1190 #endif
1191                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1192                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1193                 ptlrpc_request_set_replen(req);
1194                 rc = ptlrpc_queue_wait(req);
1195                 ptlrpc_req_finished(req);
1196         }
1197
1198 set_state:
1199         spin_lock(&imp->imp_lock);
1200 out:
1201         if (noclose)
1202                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1203         else
1204                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1205         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1206         imp->imp_conn_cnt = 0;
1207         imp->imp_last_recon = 0;
1208         spin_unlock(&imp->imp_lock);
1209
1210         RETURN(rc);
1211 }
1212