lustre/ptlrpc/import.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Mike Shaver <shaver@clusterfs.com>
   6  *
   7  *   This file is part of the Lustre file system, http://www.lustre.org
   8  *   Lustre is a trademark of Cluster File Systems, Inc.
   9  *
  10  *   You may have signed or agreed to another license before downloading
  11  *   this software.  If so, you are bound by the terms and conditions
  12  *   of that agreement, and the following does not apply to you.  See the
  13  *   LICENSE file included with this distribution for more information.
  14  *
  15  *   If you did not agree to a different license, then this copy of Lustre
  16  *   is open source software; you can redistribute it and/or modify it
  17  *   under the terms of version 2 of the GNU General Public License as
  18  *   published by the Free Software Foundation.
  19  *
  20  *   In either case, Lustre is distributed in the hope that it will be
  21  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  22  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  23  *   license text for more details.
  24  */
  25
  26 #define DEBUG_SUBSYSTEM S_RPC
  27 #ifndef __KERNEL__
  28 # include <liblustre.h>
  29 #endif
  30
  31 #include <obd_support.h>
  32 #include <lustre_ha.h>
  33 #include <lustre_net.h>
  34 #include <lustre_import.h>
  35 #include <lustre_export.h>
  36 #include <obd.h>
  37 #include <obd_class.h>
  38
  39 #include "ptlrpc_internal.h"
  40
  41 struct ptlrpc_connect_async_args {
  42          __u64 pcaa_peer_committed;
  43         int pcaa_initial_connect;
  44 };
  45
  46 /* A CLOSED import should remain so. */
  47 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  48 do {                                                                           \
  49         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  50                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  51                       imp, obd2cli_tgt(imp->imp_obd),                          \
  52                       ptlrpc_import_state_name(imp->imp_state),                \
  53                       ptlrpc_import_state_name(state));                        \
  54                imp->imp_state = state;                                         \
  55         }                                                                      \
  56 } while(0)
  57
  58 #define IMPORT_SET_STATE(imp, state)            \
  59 do {                                            \
  60         spin_lock(&imp->imp_lock);              \
  61         IMPORT_SET_STATE_NOLOCK(imp, state);    \
  62         spin_unlock(&imp->imp_lock);            \
  63 } while(0)
  64
  65
  66 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
  67                                     void * data, int rc);
  68 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
  69
  70 /* Only this function is allowed to change the import state when it is
  71  * CLOSED. I would rather refcount the import and free it after
  72  * disconnection like we do with exports. To do that, the client_obd
  73  * will need to save the peer info somewhere other than in the import,
  74  * though. */
  75 int ptlrpc_init_import(struct obd_import *imp)
  76 {
  77         spin_lock(&imp->imp_lock);
  78
  79         imp->imp_generation++;
  80         imp->imp_state =  LUSTRE_IMP_NEW;
  81
  82         spin_unlock(&imp->imp_lock);
  83
  84         return 0;
  85 }
  86 EXPORT_SYMBOL(ptlrpc_init_import);
  87
  88 #define UUID_STR "_UUID"
  89 static void deuuidify(char *uuid, const char *prefix, char **uuid_start,
  90                       int *uuid_len)
  91 {
  92         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
  93                 ? uuid : uuid + strlen(prefix);
  94
  95         *uuid_len = strlen(*uuid_start);
  96
  97         if (*uuid_len < strlen(UUID_STR))
  98                 return;
  99
 100         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 101                     UUID_STR, strlen(UUID_STR)))
 102                 *uuid_len -= strlen(UUID_STR);
 103 }
 104
 105 /* Returns true if import was FULL, false if import was already not
 106  * connected.
 107  * @imp - import to be disconnected
 108  * @conn_cnt - connection count (epoch) of the request that timed out
 109  *             and caused the disconnection.  In some cases, multiple
 110  *             inflight requests can fail to a single target (e.g. OST
 111  *             bulk requests) and if one has already caused a reconnection
 112  *             (increasing the import->conn_cnt) the older failure should
 113  *             not also cause a reconnection.  If zero it forces a reconnect.
 114  */
 115 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 116 {
 117         int rc = 0;
 118
 119         spin_lock(&imp->imp_lock);
 120
 121         if (imp->imp_state == LUSTRE_IMP_FULL &&
 122             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 123                 char *target_start;
 124                 int   target_len;
 125
 126                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 127                           &target_start, &target_len);
 128
 129                 if (imp->imp_replayable) {
 130                         LCONSOLE_WARN("%s: Connection to service %.*s via nid "
 131                                "%s was lost; in progress operations using this "
 132                                "service will wait for recovery to complete.\n",
 133                                imp->imp_obd->obd_name, target_len, target_start,
 134                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 135                 } else {
 136                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to service "
 137                                            "%.*s via nid %s was lost; in progress"
 138                                            "operations using this service will"
 139                                            "fail.\n",
 140                                            imp->imp_obd->obd_name,
 141                                            target_len, target_start,
 142                                  libcfs_nid2str(imp->imp_connection->c_peer.nid));
 143                 }
 144                 ptlrpc_deactivate_timeouts(imp);
 145                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 146                 spin_unlock(&imp->imp_lock);
 147
 148                 if (obd_dump_on_timeout)
 149                         libcfs_debug_dumplog();
 150
 151                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 152                 rc = 1;
 153         } else {
 154                 spin_unlock(&imp->imp_lock);
 155                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 156                        imp->imp_client->cli_name, imp,
 157                        (imp->imp_state == LUSTRE_IMP_FULL &&
 158                         imp->imp_conn_cnt > conn_cnt) ?
 159                        "reconnected" : "not connected", imp->imp_conn_cnt,
 160                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 161         }
 162
 163         return rc;
 164 }
 165
 166 /*
 167  * This acts as a barrier; all existing requests are rejected, and
 168  * no new requests will be accepted until the import is valid again.
 169  */
 170 void ptlrpc_deactivate_import(struct obd_import *imp)
 171 {
 172         ENTRY;
 173
 174         spin_lock(&imp->imp_lock);
 175         if (imp->imp_invalid) {
 176                 spin_unlock(&imp->imp_lock);
 177                 EXIT;
 178                 return;
 179         }
 180
 181         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 182         imp->imp_invalid = 1;
 183         imp->imp_generation++;
 184         spin_unlock(&imp->imp_lock);
 185
 186         ptlrpc_abort_inflight(imp);
 187         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 188
 189         EXIT;
 190 }
 191
 192 /*
 193  * This function will invalidate the import, if necessary, then block
 194  * for all the RPC completions, and finally notify the obd to
 195  * invalidate its state (ie cancel locks, clear pending requests,
 196  * etc).
 197  */
 198 void ptlrpc_invalidate_import(struct obd_import *imp)
 199 {
 200         struct l_wait_info lwi;
 201         int rc;
 202
 203         atomic_inc(&imp->imp_inval_count);
 204
 205         ptlrpc_deactivate_import(imp);
 206
 207         LASSERT(imp->imp_invalid);
 208
 209         /* wait for all requests to error out and call completion callbacks */
 210         lwi = LWI_TIMEOUT_INTERVAL(cfs_timeout_cap(cfs_time_seconds(obd_timeout)),
 211                                    HZ, NULL, NULL);
 212         rc = l_wait_event(imp->imp_recovery_waitq,
 213                           (atomic_read(&imp->imp_inflight) == 0), &lwi);
 214
 215         if (rc)
 216                 CDEBUG(D_HA, "%s: rc = %d waiting for callback (%d != 0)\n",
 217                        obd2cli_tgt(imp->imp_obd), rc,
 218                        atomic_read(&imp->imp_inflight));
 219
 220         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 221         sptlrpc_import_flush_all_ctx(imp);
 222
 223         atomic_dec(&imp->imp_inval_count);
 224         cfs_waitq_signal(&imp->imp_recovery_waitq);
 225 }
 226
 227 /* unset imp_invalid */
 228 void ptlrpc_activate_import(struct obd_import *imp)
 229 {
 230         struct obd_device *obd = imp->imp_obd;
 231
 232         spin_lock(&imp->imp_lock);
 233         imp->imp_invalid = 0;
 234         ptlrpc_activate_timeouts(imp);
 235         spin_unlock(&imp->imp_lock);
 236         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 237 }
 238
 239 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 240 {
 241         ENTRY;
 242
 243         LASSERT(!imp->imp_dlm_fake);
 244
 245         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 246                 if (!imp->imp_replayable) {
 247                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 248                                "auto-deactivating\n",
 249                                obd2cli_tgt(imp->imp_obd),
 250                                imp->imp_connection->c_remote_uuid.uuid,
 251                                imp->imp_obd->obd_name);
 252                         ptlrpc_deactivate_import(imp);
 253                 }
 254
 255                 CDEBUG(D_HA, "%s: waking up pinger\n",
 256                        obd2cli_tgt(imp->imp_obd));
 257
 258                 spin_lock(&imp->imp_lock);
 259                 imp->imp_force_verify = 1;
 260                 spin_unlock(&imp->imp_lock);
 261
 262                 ptlrpc_pinger_wake_up();
 263         }
 264         EXIT;
 265 }
 266
 267 static int import_select_connection(struct obd_import *imp)
 268 {
 269         struct obd_import_conn *imp_conn = NULL, *conn;
 270         struct obd_export *dlmexp;
 271         ENTRY;
 272
 273         spin_lock(&imp->imp_lock);
 274
 275         if (list_empty(&imp->imp_conn_list)) {
 276                 CERROR("%s: no connections available\n",
 277                         imp->imp_obd->obd_name);
 278                 spin_unlock(&imp->imp_lock);
 279                 RETURN(-EINVAL);
 280         }
 281
 282         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 283                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 284                        imp->imp_obd->obd_name,
 285                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 286                        conn->oic_last_attempt);
 287                 /* Throttle the reconnect rate to once per RECONNECT_INTERVAL */
 288                 if (cfs_time_before_64(conn->oic_last_attempt +
 289                                        RECONNECT_INTERVAL * HZ,
 290                                        cfs_time_current_64())) {
 291                         /* If we have never tried this connection since the
 292                            the last successful attempt, go with this one */
 293                         if (cfs_time_beforeq_64(conn->oic_last_attempt,
 294                                                imp->imp_last_success_conn)) {
 295                                 imp_conn = conn;
 296                                 break;
 297                         }
 298
 299                         /* Both of these connections have already been tried
 300                            since the last successful connection; just choose the
 301                            least recently used */
 302                         if (!imp_conn)
 303                                 imp_conn = conn;
 304                         else if (cfs_time_before_64(conn->oic_last_attempt,
 305                                                     imp_conn->oic_last_attempt))
 306                                 imp_conn = conn;
 307                 }
 308         }
 309
 310         /* if not found, simply choose the current one */
 311         if (!imp_conn) {
 312                 LASSERT(imp->imp_conn_current);
 313                 imp_conn = imp->imp_conn_current;
 314         }
 315         LASSERT(imp_conn->oic_conn);
 316
 317         imp_conn->oic_last_attempt = cfs_time_current_64();
 318
 319         /* switch connection, don't mind if it's same as the current one */
 320         if (imp->imp_connection)
 321                 ptlrpc_put_connection(imp->imp_connection);
 322         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 323
 324         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 325         LASSERT(dlmexp != NULL);
 326         if (dlmexp->exp_connection)
 327                 ptlrpc_put_connection(dlmexp->exp_connection);
 328         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 329         class_export_put(dlmexp);
 330
 331         if (imp->imp_conn_current != imp_conn) {
 332                 if (imp->imp_conn_current)
 333                         LCONSOLE_INFO("Changing connection for %s to %s/%s\n",
 334                                       imp->imp_obd->obd_name,
 335                                       imp_conn->oic_uuid.uuid,
 336                                       libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 337                 imp->imp_conn_current = imp_conn;
 338         }
 339
 340         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 341                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 342                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 343
 344         spin_unlock(&imp->imp_lock);
 345
 346         RETURN(0);
 347 }
 348
 349 /*
 350  * must be called under imp_lock
 351  */
 352 int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 353 {
 354         struct ptlrpc_request *req;
 355         struct list_head *tmp;
 356
 357         if (list_empty(&imp->imp_replay_list))
 358                 return 0;
 359         tmp = imp->imp_replay_list.next;
 360         req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 361         *transno = req->rq_transno;
 362         return 1;
 363 }
 364
 365 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
 366 {
 367         struct obd_device *obd = imp->imp_obd;
 368         int initial_connect = 0;
 369         int set_transno = 0;
 370         int rc;
 371         __u64 committed_before_reconnect = 0;
 372         struct ptlrpc_request *request;
 373         int size[] = { sizeof(struct ptlrpc_body),
 374                        sizeof(imp->imp_obd->u.cli.cl_target_uuid),
 375                        sizeof(obd->obd_uuid),
 376                        sizeof(imp->imp_dlm_handle),
 377                        sizeof(imp->imp_connect_data) };
 378         char *tmp[] = { NULL,
 379                         obd2cli_tgt(imp->imp_obd),
 380                         obd->obd_uuid.uuid,
 381                         (char *)&imp->imp_dlm_handle,
 382                         (char *)&imp->imp_connect_data };
 383         struct ptlrpc_connect_async_args *aa;
 384         ENTRY;
 385
 386         spin_lock(&imp->imp_lock);
 387         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 388                 spin_unlock(&imp->imp_lock);
 389                 CERROR("can't connect to a closed import\n");
 390                 RETURN(-EINVAL);
 391         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 392                 spin_unlock(&imp->imp_lock);
 393                 CERROR("already connected\n");
 394                 RETURN(0);
 395         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 396                 spin_unlock(&imp->imp_lock);
 397                 CERROR("already connecting\n");
 398                 RETURN(-EALREADY);
 399         }
 400
 401         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 402
 403         imp->imp_conn_cnt++;
 404         imp->imp_resend_replay = 0;
 405
 406         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 407                 initial_connect = 1;
 408         else
 409                 committed_before_reconnect = imp->imp_peer_committed_transno;
 410
 411         set_transno = ptlrpc_first_transno(imp, &imp->imp_connect_data.ocd_transno);
 412         spin_unlock(&imp->imp_lock);
 413
 414         if (new_uuid) {
 415                 struct obd_uuid uuid;
 416
 417                 obd_str2uuid(&uuid, new_uuid);
 418                 rc = import_set_conn_priority(imp, &uuid);
 419                 if (rc)
 420                         GOTO(out, rc);
 421         }
 422
 423         rc = import_select_connection(imp);
 424         if (rc)
 425                 GOTO(out, rc);
 426
 427         /* last in connection list */
 428         if (imp->imp_conn_current->oic_item.next == &imp->imp_conn_list) {
 429                 if (imp->imp_initial_recov_bk && initial_connect) {
 430                         CDEBUG(D_HA, "Last connection attempt (%d) for %s\n",
 431                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 432                         /* Don't retry if connect fails */
 433                         rc = 0;
 434                         obd_set_info_async(obd->obd_self_export,
 435                                            strlen(KEY_INIT_RECOV),
 436                                            KEY_INIT_RECOV,
 437                                            sizeof(rc), &rc, NULL);
 438                 }
 439                 if (imp->imp_recon_bk) {
 440                         CDEBUG(D_HA, "Last reconnection attempt (%d) for %s\n",
 441                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 442                         spin_lock(&imp->imp_lock);
 443                         imp->imp_last_recon = 1;
 444                         spin_unlock(&imp->imp_lock);
 445                 }
 446         }
 447
 448         /* Reset connect flags to the originally requested flags, in case
 449          * the server is updated on-the-fly we will get the new features. */
 450         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 451         rc = obd_reconnect(imp->imp_obd->obd_self_export, obd,
 452                            &obd->obd_uuid, &imp->imp_connect_data);
 453         if (rc)
 454                 GOTO(out, rc);
 455
 456         request = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, imp->imp_connect_op,
 457                                   5, size, tmp);
 458         if (!request)
 459                 GOTO(out, rc = -ENOMEM);
 460
 461 #ifndef __KERNEL__
 462         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 463 #endif
 464         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 465
 466         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 467         /* Allow a slightly larger reply for future growth compatibility */
 468         size[REPLY_REC_OFF] = sizeof(struct obd_connect_data) +
 469                               16 * sizeof(__u64);
 470         ptlrpc_req_set_repsize(request, 2, size);
 471         request->rq_interpret_reply = ptlrpc_connect_interpret;
 472
 473         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 474         aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
 475         memset(aa, 0, sizeof *aa);
 476
 477         aa->pcaa_peer_committed = committed_before_reconnect;
 478         aa->pcaa_initial_connect = initial_connect;
 479
 480         if (aa->pcaa_initial_connect) {
 481                 spin_lock(&imp->imp_lock);
 482                 imp->imp_replayable = 1;
 483                 spin_unlock(&imp->imp_lock);
 484                 /* On an initial connect, we don't know which one of a
 485                    failover server pair is up.  Don't wait long. */
 486 #ifdef CRAY_XT3
 487                 request->rq_timeout = max((int)(obd_timeout / 2), 5);
 488 #else
 489                 request->rq_timeout = max((int)(obd_timeout / 20), 5);
 490 #endif
 491                 lustre_msg_add_op_flags(request->rq_reqmsg,
 492                                         MSG_CONNECT_INITIAL);
 493         }
 494
 495         if (set_transno)
 496                 lustre_msg_add_op_flags(request->rq_reqmsg,
 497                                         MSG_CONNECT_TRANSNO);
 498
 499         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request");
 500         ptlrpcd_add_req(request);
 501         rc = 0;
 502 out:
 503         if (rc != 0) {
 504                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 505         }
 506
 507         RETURN(rc);
 508 }
 509 EXPORT_SYMBOL(ptlrpc_connect_import);
 510
 511 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 512 {
 513 #ifdef __KERNEL__
 514         struct obd_import_conn *imp_conn;
 515 #endif
 516         int wake_pinger = 0;
 517
 518         ENTRY;
 519
 520         spin_lock(&imp->imp_lock);
 521         if (list_empty(&imp->imp_conn_list))
 522                 GOTO(unlock, 0);
 523
 524 #ifdef __KERNEL__
 525         imp_conn = list_entry(imp->imp_conn_list.prev,
 526                               struct obd_import_conn,
 527                               oic_item);
 528
 529         if (imp->imp_conn_current != imp_conn) {
 530                 ptlrpc_ping_import_soon(imp);
 531                 wake_pinger = 1;
 532         }
 533 #else
 534         /* liblustre has no pinger thead, so we wakup pinger anyway */
 535         wake_pinger = 1;
 536 #endif
 537
 538  unlock:
 539         spin_unlock(&imp->imp_lock);
 540
 541         if (wake_pinger)
 542                 ptlrpc_pinger_wake_up();
 543
 544         EXIT;
 545 }
 546
 547 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
 548                                     void * data, int rc)
 549 {
 550         struct ptlrpc_connect_async_args *aa = data;
 551         struct obd_import *imp = request->rq_import;
 552         struct client_obd *cli = &imp->imp_obd->u.cli;
 553         struct lustre_handle old_hdl;
 554         int msg_flags;
 555         ENTRY;
 556
 557         spin_lock(&imp->imp_lock);
 558         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 559                 spin_unlock(&imp->imp_lock);
 560                 RETURN(0);
 561         }
 562         spin_unlock(&imp->imp_lock);
 563
 564         if (rc)
 565                 GOTO(out, rc);
 566
 567         LASSERT(imp->imp_conn_current);
 568
 569         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 570
 571         /* All imports are pingable */
 572         spin_lock(&imp->imp_lock);
 573         imp->imp_pingable = 1;
 574
 575         if (aa->pcaa_initial_connect) {
 576                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 577                         imp->imp_replayable = 1;
 578                         spin_unlock(&imp->imp_lock);
 579                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 580                                obd2cli_tgt(imp->imp_obd));
 581                 } else {
 582                         imp->imp_replayable = 0;
 583                         spin_unlock(&imp->imp_lock);
 584                 }
 585
 586                 if (msg_flags & MSG_CONNECT_NEXT_VER) {
 587                         imp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
 588                         CDEBUG(D_RPCTRACE, "connect to %s with lustre_msg_v2\n",
 589                                obd2cli_tgt(imp->imp_obd));
 590                 } else {
 591                         CDEBUG(D_RPCTRACE, "connect to %s with lustre_msg_v1\n",
 592                                obd2cli_tgt(imp->imp_obd));
 593                 }
 594
 595                 imp->imp_remote_handle =
 596                                 *lustre_msg_get_handle(request->rq_repmsg);
 597
 598                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 599                 GOTO(finish, rc = 0);
 600         } else {
 601                 spin_unlock(&imp->imp_lock);
 602         }
 603
 604         /* Determine what recovery state to move the import to. */
 605         if (MSG_CONNECT_RECONNECT & msg_flags) {
 606                 memset(&old_hdl, 0, sizeof(old_hdl));
 607                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 608                             sizeof (old_hdl))) {
 609                         CERROR("%s@%s didn't like our handle "LPX64
 610                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 611                                imp->imp_connection->c_remote_uuid.uuid,
 612                                imp->imp_dlm_handle.cookie);
 613                         GOTO(out, rc = -ENOTCONN);
 614                 }
 615
 616                 if (memcmp(&imp->imp_remote_handle,
 617                            lustre_msg_get_handle(request->rq_repmsg),
 618                            sizeof(imp->imp_remote_handle))) {
 619                         int level = D_ERROR;
 620                         /* Old MGC can reconnect to a restarted MGS */
 621                         if (strcmp(imp->imp_obd->obd_type->typ_name,
 622                                    LUSTRE_MGC_NAME) == 0) {
 623                                 level = D_CONFIG;
 624                         }
 625                         CDEBUG(level,
 626                                "%s@%s changed handle from "LPX64" to "LPX64
 627                                "; copying, but this may foreshadow disaster\n",
 628                                obd2cli_tgt(imp->imp_obd),
 629                                imp->imp_connection->c_remote_uuid.uuid,
 630                                imp->imp_remote_handle.cookie,
 631                                lustre_msg_get_handle(request->rq_repmsg)->
 632                                         cookie);
 633                         imp->imp_remote_handle =
 634                                      *lustre_msg_get_handle(request->rq_repmsg);
 635                 } else {
 636                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 637                                obd2cli_tgt(imp->imp_obd),
 638                                imp->imp_connection->c_remote_uuid.uuid);
 639                 }
 640
 641                 if (imp->imp_invalid) {
 642                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 643                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 644                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 645                                imp->imp_obd->obd_name,
 646                                obd2cli_tgt(imp->imp_obd));
 647
 648                         spin_lock(&imp->imp_lock);
 649                         imp->imp_resend_replay = 1;
 650                         spin_unlock(&imp->imp_lock);
 651
 652                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 653                 } else {
 654                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 655                 }
 656         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 657                 LASSERT(imp->imp_replayable);
 658                 imp->imp_remote_handle =
 659                                 *lustre_msg_get_handle(request->rq_repmsg);
 660                 imp->imp_last_replay_transno = 0;
 661                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 662         } else {
 663                 imp->imp_remote_handle =
 664                                 *lustre_msg_get_handle(request->rq_repmsg);
 665                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 666         }
 667
 668         /* Sanity checks for a reconnected import. */
 669         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 670                 CERROR("imp_replayable flag does not match server "
 671                        "after reconnect. We should LBUG right here.\n");
 672         }
 673
 674         if (lustre_msg_get_last_committed(request->rq_repmsg) <
 675             aa->pcaa_peer_committed) {
 676                 CERROR("%s went back in time (transno "LPD64
 677                        " was previously committed, server now claims "LPD64
 678                        ")!  See https://bugzilla.clusterfs.com/"
 679                        "long_list.cgi?buglist=9646\n",
 680                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 681                        lustre_msg_get_last_committed(request->rq_repmsg));
 682         }
 683
 684 finish:
 685         rc = ptlrpc_import_recovery_state_machine(imp);
 686         if (rc != 0) {
 687                 if (rc == -ENOTCONN) {
 688                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 689                                "invalidating and reconnecting\n",
 690                                obd2cli_tgt(imp->imp_obd),
 691                                imp->imp_connection->c_remote_uuid.uuid);
 692                         ptlrpc_connect_import(imp, NULL);
 693                         RETURN(0);
 694                 }
 695         } else {
 696                 struct obd_connect_data *ocd;
 697                 struct obd_export *exp;
 698
 699                 ocd = lustre_swab_repbuf(request, REPLY_REC_OFF, sizeof(*ocd),
 700                                          lustre_swab_connect);
 701
 702                 spin_lock(&imp->imp_lock);
 703                 list_del(&imp->imp_conn_current->oic_item);
 704                 list_add(&imp->imp_conn_current->oic_item, &imp->imp_conn_list);
 705                 imp->imp_last_success_conn =
 706                         imp->imp_conn_current->oic_last_attempt;
 707
 708                 if (ocd == NULL) {
 709                         spin_unlock(&imp->imp_lock);
 710                         CERROR("Wrong connect data from server\n");
 711                         rc = -EPROTO;
 712                         GOTO(out, rc);
 713                 }
 714
 715                 imp->imp_connect_data = *ocd;
 716
 717                 exp = class_conn2export(&imp->imp_dlm_handle);
 718                 spin_unlock(&imp->imp_lock);
 719
 720                 /* check that server granted subset of flags we asked for. */
 721                 LASSERTF((ocd->ocd_connect_flags &
 722                           imp->imp_connect_flags_orig) ==
 723                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 724                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 725
 726                 if (!exp) {
 727                         /* This could happen if export is cleaned during the
 728                            connect attempt */
 729                         CERROR("Missing export for %s\n",
 730                                imp->imp_obd->obd_name);
 731                         GOTO(out, rc = -ENODEV);
 732                 }
 733                 exp->exp_connect_flags = ocd->ocd_connect_flags;
 734                 imp->imp_obd->obd_self_export->exp_connect_flags =
 735                                                         ocd->ocd_connect_flags;
 736                 class_export_put(exp);
 737
 738                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 739
 740                 if (!ocd->ocd_ibits_known &&
 741                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 742                         CERROR("Inodebits aware server returned zero compatible"
 743                                " bits?\n");
 744
 745                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 746                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
 747                                         LUSTRE_VERSION_OFFSET_WARN ||
 748                      ocd->ocd_version < LUSTRE_VERSION_CODE -
 749                                         LUSTRE_VERSION_OFFSET_WARN)) {
 750                         /* Sigh, some compilers do not like #ifdef in the middle
 751                            of macro arguments */
 752 #ifdef __KERNEL__
 753                         const char *older =
 754                                 "older. Consider upgrading this client";
 755 #else
 756                         const char *older =
 757                                 "older. Consider recompiling this application";
 758 #endif
 759                         const char *newer = "newer than client version";
 760
 761                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 762                                       "is much %s (%s)\n",
 763                                       obd2cli_tgt(imp->imp_obd),
 764                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 765                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 766                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 767                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
 768                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
 769                                       newer : older, LUSTRE_VERSION_STRING);
 770                 }
 771
 772                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
 773                         cli->cl_max_pages_per_rpc =
 774                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
 775                 }
 776
 777                 imp->imp_obd->obd_namespace->ns_connect_flags =
 778                                                         ocd->ocd_connect_flags;
 779                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
 780                                                         ocd->ocd_connect_flags;
 781
 782                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 783                         (cli->cl_max_pages_per_rpc > 0));
 784         }
 785
 786 out:
 787         if (rc != 0) {
 788                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 789                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov)
 790                         ptlrpc_deactivate_import(imp);
 791
 792                 if ((imp->imp_recon_bk && imp->imp_last_recon) ||
 793                     (rc == -EACCES)) {
 794                         /*
 795                          * Give up trying to reconnect
 796                          * EACCES means client has no permission for connection
 797                          */
 798                         imp->imp_obd->obd_no_recov = 1;
 799                         ptlrpc_deactivate_import(imp);
 800                 }
 801
 802                 if (rc == -EPROTO) {
 803                         struct obd_connect_data *ocd;
 804
 805                         /* reply message might not be ready */
 806                         if (request->rq_repmsg != NULL)
 807                                 RETURN(-EPROTO);
 808
 809                         ocd = lustre_swab_repbuf(request, REPLY_REC_OFF,
 810                                                  sizeof *ocd,
 811                                                  lustre_swab_connect);
 812                         if (ocd &&
 813                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 814                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
 815                            /* Actually servers are only supposed to refuse
 816                               connection from liblustre clients, so we should
 817                               never see this from VFS context */
 818                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
 819                                         "(%d.%d.%d.%d)"
 820                                         " refused connection from this client "
 821                                         "with an incompatible version (%s).  "
 822                                         "Client must be recompiled\n",
 823                                         obd2cli_tgt(imp->imp_obd),
 824                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 825                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 826                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 827                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
 828                                         LUSTRE_VERSION_STRING);
 829                                 ptlrpc_deactivate_import(imp);
 830                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
 831                         }
 832                         RETURN(-EPROTO);
 833                 }
 834
 835                 ptlrpc_maybe_ping_import_soon(imp);
 836
 837                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
 838                        obd2cli_tgt(imp->imp_obd),
 839                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
 840         }
 841
 842         spin_lock(&imp->imp_lock);
 843         imp->imp_last_recon = 0;
 844         spin_unlock(&imp->imp_lock);
 845
 846         cfs_waitq_signal(&imp->imp_recovery_waitq);
 847         RETURN(rc);
 848 }
 849
 850 static int completed_replay_interpret(struct ptlrpc_request *req,
 851                                     void * data, int rc)
 852 {
 853         ENTRY;
 854         atomic_dec(&req->rq_import->imp_replay_inflight);
 855         if (req->rq_status == 0) {
 856                 ptlrpc_import_recovery_state_machine(req->rq_import);
 857         } else {
 858                 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
 859                        "reconnecting\n",
 860                        req->rq_import->imp_obd->obd_name, req->rq_status);
 861                 ptlrpc_connect_import(req->rq_import, NULL);
 862         }
 863
 864         RETURN(0);
 865 }
 866
 867 static int signal_completed_replay(struct obd_import *imp)
 868 {
 869         struct ptlrpc_request *req;
 870         ENTRY;
 871
 872         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 873         atomic_inc(&imp->imp_replay_inflight);
 874
 875         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING, 1, NULL, NULL);
 876         if (!req) {
 877                 atomic_dec(&imp->imp_replay_inflight);
 878                 RETURN(-ENOMEM);
 879         }
 880
 881         ptlrpc_req_set_repsize(req, 1, NULL);
 882         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
 883         lustre_msg_add_flags(req->rq_reqmsg,
 884                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
 885         req->rq_timeout *= 3;
 886         req->rq_interpret_reply = completed_replay_interpret;
 887
 888         ptlrpcd_add_req(req);
 889         RETURN(0);
 890 }
 891
 892 #ifdef __KERNEL__
 893 static int ptlrpc_invalidate_import_thread(void *data)
 894 {
 895         struct obd_import *imp = data;
 896
 897         ENTRY;
 898
 899         ptlrpc_daemonize("ll_imp_inval");
 900
 901         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
 902                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 903                imp->imp_connection->c_remote_uuid.uuid);
 904
 905         ptlrpc_invalidate_import(imp);
 906
 907         if (obd_dump_on_eviction) {
 908                 CERROR("dump the log upon eviction\n");
 909                 libcfs_debug_dumplog();
 910         }
 911
 912         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 913         ptlrpc_import_recovery_state_machine(imp);
 914
 915         RETURN(0);
 916 }
 917 #endif
 918
 919 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
 920 {
 921         int rc = 0;
 922         int inflight;
 923         char *target_start;
 924         int target_len;
 925
 926         ENTRY;
 927         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
 928                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 929                           &target_start, &target_len);
 930                 /* Don't care about MGC eviction */
 931                 if (strcmp(imp->imp_obd->obd_type->typ_name,
 932                            LUSTRE_MGC_NAME) != 0) {
 933                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
 934                                            "%.*s; in progress operations using "
 935                                            "this service will fail.\n",
 936                                            target_len, target_start);
 937                 }
 938                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
 939                        obd2cli_tgt(imp->imp_obd),
 940                        imp->imp_connection->c_remote_uuid.uuid);
 941
 942 #ifdef __KERNEL__
 943                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
 944                                        CLONE_VM | CLONE_FILES);
 945                 if (rc < 0)
 946                         CERROR("error starting invalidate thread: %d\n", rc);
 947                 else
 948                         rc = 0;
 949                 RETURN(rc);
 950 #else
 951                 ptlrpc_invalidate_import(imp);
 952
 953                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 954 #endif
 955         }
 956
 957         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
 958                 CDEBUG(D_HA, "replay requested by %s\n",
 959                        obd2cli_tgt(imp->imp_obd));
 960                 rc = ptlrpc_replay_next(imp, &inflight);
 961                 if (inflight == 0 &&
 962                     atomic_read(&imp->imp_replay_inflight) == 0) {
 963                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 964                         rc = ldlm_replay_locks(imp);
 965                         if (rc)
 966                                 GOTO(out, rc);
 967                 }
 968                 rc = 0;
 969         }
 970
 971         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
 972                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
 973                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
 974                         rc = signal_completed_replay(imp);
 975                         if (rc)
 976                                 GOTO(out, rc);
 977                 }
 978
 979         }
 980
 981         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
 982                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
 983                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 984                 }
 985         }
 986
 987         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
 988                 CDEBUG(D_HA, "reconnected to %s@%s\n",
 989                        obd2cli_tgt(imp->imp_obd),
 990                        imp->imp_connection->c_remote_uuid.uuid);
 991
 992                 rc = ptlrpc_resend(imp);
 993                 if (rc)
 994                         GOTO(out, rc);
 995                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 996                 ptlrpc_activate_import(imp);
 997
 998                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 999                           &target_start, &target_len);
1000                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1001                               "using nid %s.\n", imp->imp_obd->obd_name,
1002                               target_len, target_start,
1003                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1004         }
1005
1006         if (imp->imp_state == LUSTRE_IMP_FULL) {
1007                 cfs_waitq_signal(&imp->imp_recovery_waitq);
1008                 ptlrpc_wake_delayed(imp);
1009         }
1010
1011 out:
1012         RETURN(rc);
1013 }
1014
1015 static int back_to_sleep(void *unused)
1016 {
1017         return 0;
1018 }
1019
1020 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1021 {
1022         struct ptlrpc_request *req;
1023         int rq_opc, rc = 0;
1024         ENTRY;
1025
1026         switch (imp->imp_connect_op) {
1027         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1028         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1029         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1030         default:
1031                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1032                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1033                 RETURN(-EINVAL);
1034         }
1035
1036         if (ptlrpc_import_in_recovery(imp)) {
1037                 struct l_wait_info lwi;
1038                 cfs_duration_t timeout;
1039                 if (imp->imp_server_timeout)
1040                         timeout = cfs_time_seconds(obd_timeout / 2);
1041                 else
1042                         timeout = cfs_time_seconds(obd_timeout);
1043
1044                 timeout = MAX(timeout * HZ, 1);
1045
1046                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1047                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1048                 rc = l_wait_event(imp->imp_recovery_waitq,
1049                                   !ptlrpc_import_in_recovery(imp), &lwi);
1050
1051         }
1052
1053         spin_lock(&imp->imp_lock);
1054         if (imp->imp_state != LUSTRE_IMP_FULL)
1055                 GOTO(out, 0);
1056
1057         spin_unlock(&imp->imp_lock);
1058
1059         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, rq_opc, 1, NULL, NULL);
1060         if (req) {
1061                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1062                  * it fails.  We can get through the above with a down server
1063                  * if the client doesn't know the server is gone yet. */
1064                 req->rq_no_resend = 1;
1065 #ifdef CRAY_XT3
1066                 req->rq_timeout = obd_timeout / 3;
1067 #else
1068                 req->rq_timeout = 5;
1069 #endif
1070                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1071                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1072                 ptlrpc_req_set_repsize(req, 1, NULL);
1073                 rc = ptlrpc_queue_wait(req);
1074                 ptlrpc_req_finished(req);
1075         }
1076
1077         spin_lock(&imp->imp_lock);
1078 out:
1079         if (noclose)
1080                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1081         else
1082                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1083         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1084         imp->imp_conn_cnt = 0;
1085         imp->imp_last_recon = 0;
1086         spin_unlock(&imp->imp_lock);
1087
1088         RETURN(rc);
1089 }
1090