lustre/ptlrpc/import.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Mike Shaver <shaver@clusterfs.com>
   6  *
   7  *   This file is part of the Lustre file system, http://www.lustre.org
   8  *   Lustre is a trademark of Cluster File Systems, Inc.
   9  *
  10  *   You may have signed or agreed to another license before downloading
  11  *   this software.  If so, you are bound by the terms and conditions
  12  *   of that agreement, and the following does not apply to you.  See the
  13  *   LICENSE file included with this distribution for more information.
  14  *
  15  *   If you did not agree to a different license, then this copy of Lustre
  16  *   is open source software; you can redistribute it and/or modify it
  17  *   under the terms of version 2 of the GNU General Public License as
  18  *   published by the Free Software Foundation.
  19  *
  20  *   In either case, Lustre is distributed in the hope that it will be
  21  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  22  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  23  *   license text for more details.
  24  */
  25
  26 #define DEBUG_SUBSYSTEM S_RPC
  27 #ifndef __KERNEL__
  28 # include <liblustre.h>
  29 #endif
  30
  31 #include <obd_support.h>
  32 #include <lustre_ha.h>
  33 #include <lustre_net.h>
  34 #include <lustre_import.h>
  35 #include <lustre_export.h>
  36 #include <obd.h>
  37 #include <obd_class.h>
  38
  39 #include "ptlrpc_internal.h"
  40
  41 struct ptlrpc_connect_async_args {
  42          __u64 pcaa_peer_committed;
  43         int pcaa_initial_connect;
  44 };
  45
  46 /* A CLOSED import should remain so. */
  47 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  48 do {                                                                           \
  49         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  50                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  51                       imp, obd2cli_tgt(imp->imp_obd),                          \
  52                       ptlrpc_import_state_name(imp->imp_state),                \
  53                       ptlrpc_import_state_name(state));                        \
  54                imp->imp_state = state;                                         \
  55         }                                                                      \
  56 } while(0)
  57
  58 #define IMPORT_SET_STATE(imp, state)            \
  59 do {                                            \
  60         spin_lock(&imp->imp_lock);              \
  61         IMPORT_SET_STATE_NOLOCK(imp, state);    \
  62         spin_unlock(&imp->imp_lock);            \
  63 } while(0)
  64
  65
  66 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
  67                                     void * data, int rc);
  68 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
  69
  70 /* Only this function is allowed to change the import state when it is
  71  * CLOSED. I would rather refcount the import and free it after
  72  * disconnection like we do with exports. To do that, the client_obd
  73  * will need to save the peer info somewhere other than in the import,
  74  * though. */
  75 int ptlrpc_init_import(struct obd_import *imp)
  76 {
  77         spin_lock(&imp->imp_lock);
  78
  79         imp->imp_generation++;
  80         imp->imp_state =  LUSTRE_IMP_NEW;
  81
  82         spin_unlock(&imp->imp_lock);
  83
  84         return 0;
  85 }
  86 EXPORT_SYMBOL(ptlrpc_init_import);
  87
  88 #define UUID_STR "_UUID"
  89 static void deuuidify(char *uuid, const char *prefix, char **uuid_start,
  90                       int *uuid_len)
  91 {
  92         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
  93                 ? uuid : uuid + strlen(prefix);
  94
  95         *uuid_len = strlen(*uuid_start);
  96
  97         if (*uuid_len < strlen(UUID_STR))
  98                 return;
  99
 100         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 101                     UUID_STR, strlen(UUID_STR)))
 102                 *uuid_len -= strlen(UUID_STR);
 103 }
 104
 105 /* Returns true if import was FULL, false if import was already not
 106  * connected.
 107  * @imp - import to be disconnected
 108  * @conn_cnt - connection count (epoch) of the request that timed out
 109  *             and caused the disconnection.  In some cases, multiple
 110  *             inflight requests can fail to a single target (e.g. OST
 111  *             bulk requests) and if one has already caused a reconnection
 112  *             (increasing the import->conn_cnt) the older failure should
 113  *             not also cause a reconnection.  If zero it forces a reconnect.
 114  */
 115 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 116 {
 117         int rc = 0;
 118
 119         spin_lock(&imp->imp_lock);
 120
 121         if (imp->imp_state == LUSTRE_IMP_FULL &&
 122             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 123                 char *target_start;
 124                 int   target_len;
 125
 126                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 127                           &target_start, &target_len);
 128                 if (imp->imp_replayable) {
 129                         LCONSOLE_WARN("%s: Connection to service %.*s via nid "
 130                                "%s was lost; in progress operations using this "
 131                                "service will wait for recovery to complete.\n",
 132                                imp->imp_obd->obd_name, target_len, target_start,
 133                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 134                 } else {
 135                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to service "
 136                                "%.*s via nid %s was lost; in progress "
 137                                "operations using this service will fail.\n",
 138                                imp->imp_obd->obd_name, target_len, target_start,
 139                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 140                 }
 141                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 142                 spin_unlock(&imp->imp_lock);
 143
 144                 if (obd_dump_on_timeout)
 145                         libcfs_debug_dumplog();
 146
 147                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 148                 rc = 1;
 149         } else {
 150                 spin_unlock(&imp->imp_lock);
 151                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 152                        imp->imp_client->cli_name, imp,
 153                        (imp->imp_state == LUSTRE_IMP_FULL &&
 154                         imp->imp_conn_cnt > conn_cnt) ?
 155                        "reconnected" : "not connected", imp->imp_conn_cnt,
 156                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 157         }
 158
 159         return rc;
 160 }
 161
 162 /*
 163  * This acts as a barrier; all existing requests are rejected, and
 164  * no new requests will be accepted until the import is valid again.
 165  */
 166 void ptlrpc_deactivate_import(struct obd_import *imp)
 167 {
 168         ENTRY;
 169
 170         spin_lock(&imp->imp_lock);
 171         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 172         imp->imp_invalid = 1;
 173         imp->imp_generation++;
 174         spin_unlock(&imp->imp_lock);
 175
 176         ptlrpc_abort_inflight(imp);
 177         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 178 }
 179
 180 /*
 181  * This function will invalidate the import, if necessary, then block
 182  * for all the RPC completions, and finally notify the obd to
 183  * invalidate its state (ie cancel locks, clear pending requests,
 184  * etc).
 185  */
 186 void ptlrpc_invalidate_import(struct obd_import *imp)
 187 {
 188         struct l_wait_info lwi;
 189         int rc;
 190
 191         if (!imp->imp_invalid)
 192                 ptlrpc_deactivate_import(imp);
 193
 194         LASSERT(imp->imp_invalid);
 195
 196         /* wait for all requests to error out and call completion callbacks */
 197         lwi = LWI_TIMEOUT_INTERVAL(cfs_timeout_cap(cfs_time_seconds(obd_timeout)),
 198                                    HZ, NULL, NULL);
 199         rc = l_wait_event(imp->imp_recovery_waitq,
 200                           (atomic_read(&imp->imp_inflight) == 0),
 201                           &lwi);
 202
 203         if (rc)
 204                 CDEBUG(D_HA, "%s: rc = %d waiting for callback (%d != 0)\n",
 205                        obd2cli_tgt(imp->imp_obd), rc,
 206                        atomic_read(&imp->imp_inflight));
 207
 208         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 209 }
 210
 211 /* unset imp_invalid */
 212 void ptlrpc_activate_import(struct obd_import *imp)
 213 {
 214         struct obd_device *obd = imp->imp_obd;
 215
 216         spin_lock(&imp->imp_lock);
 217         imp->imp_invalid = 0;
 218         spin_unlock(&imp->imp_lock);
 219
 220         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 221 }
 222
 223 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 224 {
 225         ENTRY;
 226
 227         LASSERT(!imp->imp_dlm_fake);
 228
 229         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 230                 if (!imp->imp_replayable) {
 231                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 232                                "auto-deactivating\n",
 233                                obd2cli_tgt(imp->imp_obd),
 234                                imp->imp_connection->c_remote_uuid.uuid,
 235                                imp->imp_obd->obd_name);
 236                         ptlrpc_deactivate_import(imp);
 237                 }
 238
 239                 CDEBUG(D_HA, "%s: waking up pinger\n",
 240                        obd2cli_tgt(imp->imp_obd));
 241
 242                 spin_lock(&imp->imp_lock);
 243                 imp->imp_force_verify = 1;
 244                 spin_unlock(&imp->imp_lock);
 245
 246                 ptlrpc_pinger_wake_up();
 247         }
 248         EXIT;
 249 }
 250
 251 static int import_select_connection(struct obd_import *imp)
 252 {
 253         struct obd_import_conn *imp_conn = NULL, *conn;
 254         struct obd_export *dlmexp;
 255         ENTRY;
 256
 257         spin_lock(&imp->imp_lock);
 258
 259         if (list_empty(&imp->imp_conn_list)) {
 260                 CERROR("%s: no connections available\n",
 261                         imp->imp_obd->obd_name);
 262                 spin_unlock(&imp->imp_lock);
 263                 RETURN(-EINVAL);
 264         }
 265
 266         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 267                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 268                        imp->imp_obd->obd_name,
 269                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 270                        conn->oic_last_attempt);
 271                 /* Throttle the reconnect rate to once per RECONNECT_INTERVAL */
 272                 if (cfs_time_before_64(conn->oic_last_attempt +
 273                                        RECONNECT_INTERVAL * HZ,
 274                                        cfs_time_current_64())) {
 275                         /* If we have never tried this connection since the
 276                            the last successful attempt, go with this one */
 277                         if (cfs_time_before_64(conn->oic_last_attempt,
 278                                                imp->imp_last_success_conn)) {
 279                                 imp_conn = conn;
 280                                 break;
 281                         }
 282
 283                         /* Both of these connections have already been tried
 284                            since the last successful connection; just choose the
 285                            least recently used */
 286                         if (!imp_conn)
 287                                 imp_conn = conn;
 288                         else if (cfs_time_before_64(conn->oic_last_attempt,
 289                                                     imp_conn->oic_last_attempt))
 290                                 imp_conn = conn;
 291                 }
 292         }
 293
 294         /* if not found, simply choose the current one */
 295         if (!imp_conn) {
 296                 LASSERT(imp->imp_conn_current);
 297                 imp_conn = imp->imp_conn_current;
 298         }
 299         LASSERT(imp_conn->oic_conn);
 300
 301         imp_conn->oic_last_attempt = cfs_time_current_64();
 302
 303         /* switch connection, don't mind if it's same as the current one */
 304         if (imp->imp_connection)
 305                 ptlrpc_put_connection(imp->imp_connection);
 306         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 307
 308         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 309         LASSERT(dlmexp != NULL);
 310         if (dlmexp->exp_connection)
 311                 ptlrpc_put_connection(dlmexp->exp_connection);
 312         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 313         class_export_put(dlmexp);
 314
 315         if (imp->imp_conn_current != imp_conn) {
 316                 if (imp->imp_conn_current)
 317                         LCONSOLE_INFO("Changing connection for %s to %s/%s\n",
 318                                       imp->imp_obd->obd_name,
 319                                       imp_conn->oic_uuid.uuid,
 320                                       libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 321                 imp->imp_conn_current = imp_conn;
 322         }
 323
 324         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 325                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 326                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 327
 328         spin_unlock(&imp->imp_lock);
 329
 330         RETURN(0);
 331 }
 332
 333 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
 334 {
 335         struct obd_device *obd = imp->imp_obd;
 336         int initial_connect = 0;
 337         int rc;
 338         __u64 committed_before_reconnect = 0;
 339         struct ptlrpc_request *request;
 340         int size[] = { sizeof(struct ptlrpc_body),
 341                        sizeof(imp->imp_obd->u.cli.cl_target_uuid),
 342                        sizeof(obd->obd_uuid),
 343                        sizeof(imp->imp_dlm_handle),
 344                        sizeof(imp->imp_connect_data) };
 345         char *tmp[] = { NULL,
 346                         obd2cli_tgt(imp->imp_obd),
 347                         obd->obd_uuid.uuid,
 348                         (char *)&imp->imp_dlm_handle,
 349                         (char *)&imp->imp_connect_data };
 350         struct ptlrpc_connect_async_args *aa;
 351
 352         ENTRY;
 353         spin_lock(&imp->imp_lock);
 354         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 355                 spin_unlock(&imp->imp_lock);
 356                 CERROR("can't connect to a closed import\n");
 357                 RETURN(-EINVAL);
 358         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 359                 spin_unlock(&imp->imp_lock);
 360                 CERROR("already connected\n");
 361                 RETURN(0);
 362         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 363                 spin_unlock(&imp->imp_lock);
 364                 CERROR("already connecting\n");
 365                 RETURN(-EALREADY);
 366         }
 367
 368         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 369
 370         imp->imp_conn_cnt++;
 371         imp->imp_resend_replay = 0;
 372
 373         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 374                 initial_connect = 1;
 375         else
 376                 committed_before_reconnect = imp->imp_peer_committed_transno;
 377
 378         spin_unlock(&imp->imp_lock);
 379
 380         if (new_uuid) {
 381                 struct obd_uuid uuid;
 382
 383                 obd_str2uuid(&uuid, new_uuid);
 384                 rc = import_set_conn_priority(imp, &uuid);
 385                 if (rc)
 386                         GOTO(out, rc);
 387         }
 388
 389         rc = import_select_connection(imp);
 390         if (rc)
 391                 GOTO(out, rc);
 392
 393         /* last in connection list */
 394         if (imp->imp_conn_current->oic_item.next == &imp->imp_conn_list) {
 395                 if (imp->imp_initial_recov_bk && initial_connect) {
 396                         CDEBUG(D_HA, "Last connection attempt (%d) for %s\n",
 397                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 398                         /* Don't retry if connect fails */
 399                         rc = 0;
 400                         obd_set_info_async(obd->obd_self_export,
 401                                            strlen(KEY_INIT_RECOV),
 402                                            KEY_INIT_RECOV,
 403                                            sizeof(rc), &rc, NULL);
 404                 }
 405                 if (imp->imp_recon_bk) {
 406                         CDEBUG(D_HA, "Last reconnection attempt (%d) for %s\n",
 407                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 408                         spin_lock(&imp->imp_lock);
 409                         imp->imp_last_recon = 1;
 410                         spin_unlock(&imp->imp_lock);
 411                 }
 412         }
 413
 414         /* Reset connect flags to the originally requested flags, in case
 415          * the server is updated on-the-fly we will get the new features. */
 416         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 417         rc = obd_reconnect(imp->imp_obd->obd_self_export, obd,
 418                            &obd->obd_uuid, &imp->imp_connect_data);
 419         if (rc)
 420                 GOTO(out, rc);
 421
 422         request = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, imp->imp_connect_op,
 423                                   5, size, tmp);
 424         if (!request)
 425                 GOTO(out, rc = -ENOMEM);
 426
 427 #ifndef __KERNEL__
 428         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 429 #endif
 430         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 431
 432         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 433         /* Allow a slightly larger reply for future growth compatibility */
 434         size[REPLY_REC_OFF] = sizeof(struct obd_connect_data) +
 435                               16 * sizeof(__u64);
 436         ptlrpc_req_set_repsize(request, 2, size);
 437         request->rq_interpret_reply = ptlrpc_connect_interpret;
 438
 439         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 440         aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
 441         memset(aa, 0, sizeof *aa);
 442
 443         aa->pcaa_peer_committed = committed_before_reconnect;
 444         aa->pcaa_initial_connect = initial_connect;
 445
 446         if (aa->pcaa_initial_connect) {
 447                 spin_lock(&imp->imp_lock);
 448                 imp->imp_replayable = 1;
 449                 spin_unlock(&imp->imp_lock);
 450                 /* On an initial connect, we don't know which one of a
 451                    failover server pair is up.  Don't wait long. */
 452 #ifdef CRAY_XT3
 453                 request->rq_timeout = max((int)(obd_timeout / 2), 5);
 454 #else
 455                 request->rq_timeout = max((int)(obd_timeout / 20), 5);
 456 #endif
 457         }
 458
 459         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request");
 460         ptlrpcd_add_req(request);
 461         rc = 0;
 462 out:
 463         if (rc != 0) {
 464                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 465         }
 466
 467         RETURN(rc);
 468 }
 469 EXPORT_SYMBOL(ptlrpc_connect_import);
 470
 471 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 472 {
 473 #ifdef __KERNEL__
 474         struct obd_import_conn *imp_conn;
 475 #endif
 476         int wake_pinger = 0;
 477
 478         ENTRY;
 479
 480         spin_lock(&imp->imp_lock);
 481         if (list_empty(&imp->imp_conn_list))
 482                 GOTO(unlock, 0);
 483
 484 #ifdef __KERNEL__
 485         imp_conn = list_entry(imp->imp_conn_list.prev,
 486                               struct obd_import_conn,
 487                               oic_item);
 488
 489         if (imp->imp_conn_current != imp_conn) {
 490                 ptlrpc_ping_import_soon(imp);
 491                 wake_pinger = 1;
 492         }
 493
 494 #else
 495         /* liblustre has no pinger thead, so we wakup pinger anyway */
 496         wake_pinger = 1;
 497 #endif
 498  unlock:
 499         spin_unlock(&imp->imp_lock);
 500
 501         if (wake_pinger)
 502                 ptlrpc_pinger_wake_up();
 503
 504         EXIT;
 505 }
 506
 507 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
 508                                     void * data, int rc)
 509 {
 510         struct ptlrpc_connect_async_args *aa = data;
 511         struct obd_import *imp = request->rq_import;
 512         struct client_obd *cli = &imp->imp_obd->u.cli;
 513         struct lustre_handle old_hdl;
 514         int msg_flags;
 515         ENTRY;
 516
 517         spin_lock(&imp->imp_lock);
 518         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 519                 spin_unlock(&imp->imp_lock);
 520                 RETURN(0);
 521         }
 522         spin_unlock(&imp->imp_lock);
 523
 524         if (rc)
 525                 GOTO(out, rc);
 526
 527         LASSERT(imp->imp_conn_current);
 528
 529         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 530
 531         /* All imports are pingable */
 532         spin_lock(&imp->imp_lock);
 533         imp->imp_pingable = 1;
 534
 535         if (aa->pcaa_initial_connect) {
 536                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 537                         imp->imp_replayable = 1;
 538                         spin_unlock(&imp->imp_lock);
 539                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 540                                obd2cli_tgt(imp->imp_obd));
 541                 } else {
 542                         imp->imp_replayable = 0;
 543                         spin_unlock(&imp->imp_lock);
 544                 }
 545
 546                 if (msg_flags & MSG_CONNECT_NEXT_VER) {
 547                         imp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
 548                         CDEBUG(D_RPCTRACE, "connect to %s with lustre_msg_v2\n",
 549                                obd2cli_tgt(imp->imp_obd));
 550                 } else {
 551                         CDEBUG(D_RPCTRACE, "connect to %s with lustre_msg_v1\n",
 552                                obd2cli_tgt(imp->imp_obd));
 553                 }
 554
 555                 imp->imp_remote_handle =
 556                                 *lustre_msg_get_handle(request->rq_repmsg);
 557
 558                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 559                 GOTO(finish, rc = 0);
 560         } else {
 561                 spin_unlock(&imp->imp_lock);
 562         }
 563
 564         /* Determine what recovery state to move the import to. */
 565         if (MSG_CONNECT_RECONNECT & msg_flags) {
 566                 memset(&old_hdl, 0, sizeof(old_hdl));
 567                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 568                             sizeof (old_hdl))) {
 569                         CERROR("%s@%s didn't like our handle "LPX64
 570                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 571                                imp->imp_connection->c_remote_uuid.uuid,
 572                                imp->imp_dlm_handle.cookie);
 573                         GOTO(out, rc = -ENOTCONN);
 574                 }
 575
 576                 if (memcmp(&imp->imp_remote_handle,
 577                            lustre_msg_get_handle(request->rq_repmsg),
 578                            sizeof(imp->imp_remote_handle))) {
 579                         int level = D_ERROR;
 580                         /* Old MGC can reconnect to a restarted MGS */
 581                         if (strcmp(imp->imp_obd->obd_type->typ_name,
 582                                    LUSTRE_MGC_NAME) == 0) {
 583                                 level = D_CONFIG;
 584                         }
 585                         CDEBUG(level,
 586                                "%s@%s changed handle from "LPX64" to "LPX64
 587                                "; copying, but this may foreshadow disaster\n",
 588                                obd2cli_tgt(imp->imp_obd),
 589                                imp->imp_connection->c_remote_uuid.uuid,
 590                                imp->imp_remote_handle.cookie,
 591                                lustre_msg_get_handle(request->rq_repmsg)->
 592                                         cookie);
 593                         imp->imp_remote_handle =
 594                                      *lustre_msg_get_handle(request->rq_repmsg);
 595                 } else {
 596                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 597                                obd2cli_tgt(imp->imp_obd),
 598                                imp->imp_connection->c_remote_uuid.uuid);
 599                 }
 600
 601                 if (imp->imp_invalid) {
 602                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 603                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 604                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 605                                imp->imp_obd->obd_name,
 606                                obd2cli_tgt(imp->imp_obd));
 607
 608                         spin_lock(&imp->imp_lock);
 609                         imp->imp_resend_replay = 1;
 610                         spin_unlock(&imp->imp_lock);
 611
 612                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 613                 } else {
 614                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 615                 }
 616         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 617                 LASSERT(imp->imp_replayable);
 618                 imp->imp_remote_handle =
 619                                 *lustre_msg_get_handle(request->rq_repmsg);
 620                 imp->imp_last_replay_transno = 0;
 621                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 622         } else {
 623                 imp->imp_remote_handle =
 624                                 *lustre_msg_get_handle(request->rq_repmsg);
 625                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 626         }
 627
 628         /* Sanity checks for a reconnected import. */
 629         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 630                 CERROR("imp_replayable flag does not match server "
 631                        "after reconnect. We should LBUG right here.\n");
 632         }
 633
 634         if (lustre_msg_get_last_committed(request->rq_repmsg) <
 635             aa->pcaa_peer_committed) {
 636                 CERROR("%s went back in time (transno "LPD64
 637                        " was previously committed, server now claims "LPD64
 638                        ")!  See https://bugzilla.clusterfs.com/"
 639                        "long_list.cgi?buglist=9646\n",
 640                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 641                        lustre_msg_get_last_committed(request->rq_repmsg));
 642         }
 643
 644 finish:
 645         rc = ptlrpc_import_recovery_state_machine(imp);
 646         if (rc != 0) {
 647                 if (rc == -ENOTCONN) {
 648                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 649                                "invalidating and reconnecting\n",
 650                                obd2cli_tgt(imp->imp_obd),
 651                                imp->imp_connection->c_remote_uuid.uuid);
 652                         ptlrpc_connect_import(imp, NULL);
 653                         RETURN(0);
 654                 }
 655         } else {
 656                 struct obd_connect_data *ocd;
 657                 struct obd_export *exp;
 658
 659                 ocd = lustre_swab_repbuf(request, REPLY_REC_OFF, sizeof(*ocd),
 660                                          lustre_swab_connect);
 661
 662                 spin_lock(&imp->imp_lock);
 663                 list_del(&imp->imp_conn_current->oic_item);
 664                 list_add(&imp->imp_conn_current->oic_item, &imp->imp_conn_list);
 665                 imp->imp_last_success_conn =
 666                         imp->imp_conn_current->oic_last_attempt;
 667
 668                 if (ocd == NULL) {
 669                         spin_unlock(&imp->imp_lock);
 670                         CERROR("Wrong connect data from server\n");
 671                         rc = -EPROTO;
 672                         GOTO(out, rc);
 673                 }
 674
 675                 imp->imp_connect_data = *ocd;
 676
 677                 exp = class_conn2export(&imp->imp_dlm_handle);
 678                 spin_unlock(&imp->imp_lock);
 679
 680                 /* check that server granted subset of flags we asked for. */
 681                 LASSERTF((ocd->ocd_connect_flags &
 682                           imp->imp_connect_flags_orig) ==
 683                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 684                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 685
 686                 if (!exp) {
 687                         /* This could happen if export is cleaned during the
 688                            connect attempt */
 689                         CERROR("Missing export for %s\n",
 690                                imp->imp_obd->obd_name);
 691                         GOTO(out, rc = -ENODEV);
 692                 }
 693                 exp->exp_connect_flags = ocd->ocd_connect_flags;
 694                 class_export_put(exp);
 695
 696                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 697
 698                 if (!ocd->ocd_ibits_known &&
 699                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 700                         CERROR("Inodebits aware server returned zero compatible"
 701                                " bits?\n");
 702
 703                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 704                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
 705                                         LUSTRE_VERSION_OFFSET_WARN ||
 706                      ocd->ocd_version < LUSTRE_VERSION_CODE -
 707                                         LUSTRE_VERSION_OFFSET_WARN)) {
 708                         /* Sigh, some compilers do not like #ifdef in the middle
 709                            of macro arguments */
 710 #ifdef __KERNEL__
 711                         const char *older =
 712                                 "older.  Consider upgrading this client";
 713 #else
 714                         const char *older =
 715                                 "older.  Consider recompiling this application";
 716 #endif
 717                         const char *newer = "newer than client version";
 718
 719                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 720                                       "is much %s (%s)\n",
 721                                       obd2cli_tgt(imp->imp_obd),
 722                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 723                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 724                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 725                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
 726                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
 727                                       newer : older, LUSTRE_VERSION_STRING);
 728                 }
 729
 730                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
 731                         cli->cl_max_pages_per_rpc =
 732                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
 733                 }
 734
 735                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 736                         (cli->cl_max_pages_per_rpc > 0));
 737         }
 738
 739  out:
 740         if (rc != 0) {
 741                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 742                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov)
 743                         ptlrpc_deactivate_import(imp);
 744
 745                 if (imp->imp_recon_bk && imp->imp_last_recon) {
 746                         /* Give up trying to reconnect */
 747                         imp->imp_obd->obd_no_recov = 1;
 748                         ptlrpc_deactivate_import(imp);
 749                 }
 750
 751                 if (rc == -EPROTO) {
 752                         struct obd_connect_data *ocd;
 753                         ocd = lustre_swab_repbuf(request, REPLY_REC_OFF,
 754                                                  sizeof *ocd,
 755                                                  lustre_swab_connect);
 756                         if (ocd &&
 757                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 758                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
 759                            /* Actually servers are only supposed to refuse
 760                               connection from liblustre clients, so we should
 761                               never see this from VFS context */
 762                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
 763                                         "(%d.%d.%d.%d)"
 764                                         " refused connection from this client "
 765                                         "with an incompatible version (%s).  "
 766                                         "Client must be recompiled\n",
 767                                         obd2cli_tgt(imp->imp_obd),
 768                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 769                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 770                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 771                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
 772                                         LUSTRE_VERSION_STRING);
 773                                 ptlrpc_deactivate_import(imp);
 774                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
 775                         }
 776                         RETURN(-EPROTO);
 777                 }
 778
 779                 ptlrpc_maybe_ping_import_soon(imp);
 780
 781                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
 782                        obd2cli_tgt(imp->imp_obd),
 783                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
 784         }
 785
 786         spin_lock(&imp->imp_lock);
 787         imp->imp_last_recon = 0;
 788         spin_unlock(&imp->imp_lock);
 789
 790         cfs_waitq_signal(&imp->imp_recovery_waitq);
 791         RETURN(rc);
 792 }
 793
 794 static int completed_replay_interpret(struct ptlrpc_request *req,
 795                                     void * data, int rc)
 796 {
 797         ENTRY;
 798         atomic_dec(&req->rq_import->imp_replay_inflight);
 799         if (req->rq_status == 0) {
 800                 ptlrpc_import_recovery_state_machine(req->rq_import);
 801         } else {
 802                 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
 803                        "reconnecting\n",
 804                        req->rq_import->imp_obd->obd_name, req->rq_status);
 805                 ptlrpc_connect_import(req->rq_import, NULL);
 806         }
 807
 808         RETURN(0);
 809 }
 810
 811 static int signal_completed_replay(struct obd_import *imp)
 812 {
 813         struct ptlrpc_request *req;
 814         ENTRY;
 815
 816         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 817         atomic_inc(&imp->imp_replay_inflight);
 818
 819         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING, 1, NULL, NULL);
 820         if (!req) {
 821                 atomic_dec(&imp->imp_replay_inflight);
 822                 RETURN(-ENOMEM);
 823         }
 824
 825         ptlrpc_req_set_repsize(req, 1, NULL);
 826         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
 827         lustre_msg_add_flags(req->rq_reqmsg, MSG_LAST_REPLAY);
 828         req->rq_timeout *= 3;
 829         req->rq_interpret_reply = completed_replay_interpret;
 830
 831         ptlrpcd_add_req(req);
 832         RETURN(0);
 833 }
 834
 835 #ifdef __KERNEL__
 836 static int ptlrpc_invalidate_import_thread(void *data)
 837 {
 838         struct obd_import *imp = data;
 839
 840         ENTRY;
 841
 842         ptlrpc_daemonize("ll_imp_inval");
 843
 844         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
 845                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 846                imp->imp_connection->c_remote_uuid.uuid);
 847
 848         ptlrpc_invalidate_import(imp);
 849
 850         if (obd_dump_on_eviction) {
 851                 CERROR("dump the log upon eviction\n");
 852                 libcfs_debug_dumplog();
 853         }
 854
 855         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 856         ptlrpc_import_recovery_state_machine(imp);
 857
 858         RETURN(0);
 859 }
 860 #endif
 861
 862 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
 863 {
 864         int rc = 0;
 865         int inflight;
 866         char *target_start;
 867         int target_len;
 868
 869         ENTRY;
 870         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
 871                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 872                           &target_start, &target_len);
 873                 /* Don't care about MGC eviction */
 874                 if (strcmp(imp->imp_obd->obd_type->typ_name,
 875                            LUSTRE_MGC_NAME) != 0) {
 876                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
 877                                            "%.*s; in progress operations using "
 878                                            "this service will fail.\n",
 879                                            target_len, target_start);
 880                 }
 881                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
 882                        obd2cli_tgt(imp->imp_obd),
 883                        imp->imp_connection->c_remote_uuid.uuid);
 884
 885 #ifdef __KERNEL__
 886                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
 887                                    CLONE_VM | CLONE_FILES);
 888                 if (rc < 0)
 889                         CERROR("error starting invalidate thread: %d\n", rc);
 890                 else
 891                         rc = 0;
 892                 RETURN(rc);
 893 #else
 894                 ptlrpc_invalidate_import(imp);
 895
 896                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 897 #endif
 898         }
 899
 900         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
 901                 CDEBUG(D_HA, "replay requested by %s\n",
 902                        obd2cli_tgt(imp->imp_obd));
 903                 rc = ptlrpc_replay_next(imp, &inflight);
 904                 if (inflight == 0 &&
 905                     atomic_read(&imp->imp_replay_inflight) == 0) {
 906                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 907                         rc = ldlm_replay_locks(imp);
 908                         if (rc)
 909                                 GOTO(out, rc);
 910                 }
 911                 rc = 0;
 912         }
 913
 914         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
 915                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
 916                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
 917                         rc = signal_completed_replay(imp);
 918                         if (rc)
 919                                 GOTO(out, rc);
 920                 }
 921
 922         }
 923
 924         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
 925                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
 926                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 927                 }
 928         }
 929
 930         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
 931                 CDEBUG(D_HA, "reconnected to %s@%s\n",
 932                        obd2cli_tgt(imp->imp_obd),
 933                        imp->imp_connection->c_remote_uuid.uuid);
 934
 935                 rc = ptlrpc_resend(imp);
 936                 if (rc)
 937                         GOTO(out, rc);
 938                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 939                 ptlrpc_activate_import(imp);
 940
 941                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 942                           &target_start, &target_len);
 943                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
 944                               "using nid %s.\n", imp->imp_obd->obd_name,
 945                               target_len, target_start,
 946                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
 947         }
 948
 949         if (imp->imp_state == LUSTRE_IMP_FULL) {
 950                 cfs_waitq_signal(&imp->imp_recovery_waitq);
 951                 ptlrpc_wake_delayed(imp);
 952         }
 953
 954  out:
 955         RETURN(rc);
 956 }
 957
 958 static int back_to_sleep(void *unused)
 959 {
 960         return 0;
 961 }
 962
 963 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
 964 {
 965         struct ptlrpc_request *req;
 966         int rq_opc, rc = 0;
 967         ENTRY;
 968
 969         switch (imp->imp_connect_op) {
 970         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
 971         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
 972         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
 973         default:
 974                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
 975                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
 976                 RETURN(-EINVAL);
 977         }
 978
 979         if (ptlrpc_import_in_recovery(imp)) {
 980                 struct l_wait_info lwi;
 981                 cfs_duration_t timeout = cfs_time_seconds(obd_timeout);
 982
 983                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
 984                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
 985                 rc = l_wait_event(imp->imp_recovery_waitq,
 986                                   !ptlrpc_import_in_recovery(imp), &lwi);
 987
 988         }
 989
 990         spin_lock(&imp->imp_lock);
 991         if (imp->imp_state != LUSTRE_IMP_FULL)
 992                 GOTO(out, 0);
 993
 994         spin_unlock(&imp->imp_lock);
 995
 996         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, rq_opc, 1, NULL, NULL);
 997         if (req) {
 998                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
 999                  * it fails.  We can get through the above with a down server
1000                  * if the client doesn't know the server is gone yet. */
1001                 req->rq_no_resend = 1;
1002 #ifdef CRAY_XT3
1003                 req->rq_timeout = obd_timeout / 3;
1004 #else
1005                 req->rq_timeout = 5;
1006 #endif
1007                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1008                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1009                 ptlrpc_req_set_repsize(req, 1, NULL);
1010                 rc = ptlrpc_queue_wait(req);
1011                 ptlrpc_req_finished(req);
1012         }
1013
1014         spin_lock(&imp->imp_lock);
1015 out:
1016         if (noclose)
1017                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1018         else
1019                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1020         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1021         spin_unlock(&imp->imp_lock);
1022
1023         RETURN(rc);
1024 }
1025