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         if (req->rq_transno == 0) {
 363                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 364                 LBUG();
 365         }
 366
 367         return 1;
 368 }
 369
 370 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
 371 {
 372         struct obd_device *obd = imp->imp_obd;
 373         int initial_connect = 0;
 374         int set_transno = 0;
 375         __u64 committed_before_reconnect = 0;
 376         struct ptlrpc_request *request;
 377         char *bufs[] = { NULL,
 378                          obd2cli_tgt(imp->imp_obd),
 379                          obd->obd_uuid.uuid,
 380                          (char *)&imp->imp_dlm_handle,
 381                          (char *)&imp->imp_connect_data };
 382         struct ptlrpc_connect_async_args *aa;
 383         int rc;
 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         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 449         if (rc)
 450                 GOTO(out, rc);
 451
 452         /* Reset connect flags to the originally requested flags, in case
 453          * the server is updated on-the-fly we will get the new features. */
 454         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 455         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 456                            &obd->obd_uuid, &imp->imp_connect_data);
 457         if (rc)
 458                 GOTO(out, rc);
 459
 460         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 461         if (request == NULL)
 462                 GOTO(out, rc = -ENOMEM);
 463
 464         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 465                                       imp->imp_connect_op, bufs, NULL);
 466         if (rc) {
 467                 ptlrpc_request_free(request);
 468                 GOTO(out, rc);
 469         }
 470
 471 #ifndef __KERNEL__
 472         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 473 #endif
 474         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 475
 476         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 477         /* Allow a slightly larger reply for future growth compatibility */
 478         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 479                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 480         ptlrpc_request_set_replen(request);
 481         request->rq_interpret_reply = ptlrpc_connect_interpret;
 482
 483         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 484         aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
 485         memset(aa, 0, sizeof *aa);
 486
 487         aa->pcaa_peer_committed = committed_before_reconnect;
 488         aa->pcaa_initial_connect = initial_connect;
 489
 490         if (aa->pcaa_initial_connect) {
 491                 spin_lock(&imp->imp_lock);
 492                 imp->imp_replayable = 1;
 493                 spin_unlock(&imp->imp_lock);
 494                 /* On an initial connect, we don't know which one of a
 495                    failover server pair is up.  Don't wait long. */
 496 #ifdef CRAY_XT3
 497                 request->rq_timeout = max((int)(obd_timeout / 2), 5);
 498 #else
 499                 request->rq_timeout = max((int)(obd_timeout / 20), 5);
 500 #endif
 501                 lustre_msg_add_op_flags(request->rq_reqmsg,
 502                                         MSG_CONNECT_INITIAL);
 503         }
 504
 505         if (set_transno)
 506                 lustre_msg_add_op_flags(request->rq_reqmsg,
 507                                         MSG_CONNECT_TRANSNO);
 508
 509         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request");
 510         ptlrpcd_add_req(request);
 511         rc = 0;
 512 out:
 513         if (rc != 0) {
 514                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 515         }
 516
 517         RETURN(rc);
 518 }
 519 EXPORT_SYMBOL(ptlrpc_connect_import);
 520
 521 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 522 {
 523 #ifdef __KERNEL__
 524         struct obd_import_conn *imp_conn;
 525 #endif
 526         int wake_pinger = 0;
 527
 528         ENTRY;
 529
 530         spin_lock(&imp->imp_lock);
 531         if (list_empty(&imp->imp_conn_list))
 532                 GOTO(unlock, 0);
 533
 534 #ifdef __KERNEL__
 535         imp_conn = list_entry(imp->imp_conn_list.prev,
 536                               struct obd_import_conn,
 537                               oic_item);
 538
 539         /* XXX: When the failover node is the primary node, it is possible
 540          * to have two identical connections in imp_conn_list. We must
 541          * compare not conn's pointers but NIDs, otherwise we can defeat
 542          * connection throttling. (See bug 14774.) */
 543         if (imp->imp_conn_current->oic_conn->c_self !=
 544                                 imp_conn->oic_conn->c_self) {
 545                 ptlrpc_ping_import_soon(imp);
 546                 wake_pinger = 1;
 547         }
 548 #else
 549         /* liblustre has no pinger thead, so we wakup pinger anyway */
 550         wake_pinger = 1;
 551 #endif
 552
 553  unlock:
 554         spin_unlock(&imp->imp_lock);
 555
 556         if (wake_pinger)
 557                 ptlrpc_pinger_wake_up();
 558
 559         EXIT;
 560 }
 561
 562 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
 563                                     void * data, int rc)
 564 {
 565         struct ptlrpc_connect_async_args *aa = data;
 566         struct obd_import *imp = request->rq_import;
 567         struct client_obd *cli = &imp->imp_obd->u.cli;
 568         struct lustre_handle old_hdl;
 569         int msg_flags;
 570         ENTRY;
 571
 572         spin_lock(&imp->imp_lock);
 573         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 574                 spin_unlock(&imp->imp_lock);
 575                 RETURN(0);
 576         }
 577         spin_unlock(&imp->imp_lock);
 578
 579         if (rc)
 580                 GOTO(out, rc);
 581
 582         LASSERT(imp->imp_conn_current);
 583
 584         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 585
 586         /* All imports are pingable */
 587         spin_lock(&imp->imp_lock);
 588         imp->imp_pingable = 1;
 589
 590         if (aa->pcaa_initial_connect) {
 591                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 592                         imp->imp_replayable = 1;
 593                         spin_unlock(&imp->imp_lock);
 594                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 595                                obd2cli_tgt(imp->imp_obd));
 596                 } else {
 597                         imp->imp_replayable = 0;
 598                         spin_unlock(&imp->imp_lock);
 599                 }
 600
 601                 if (msg_flags & MSG_CONNECT_NEXT_VER) {
 602                         imp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
 603                         CDEBUG(D_RPCTRACE, "connect to %s with lustre_msg_v2\n",
 604                                obd2cli_tgt(imp->imp_obd));
 605                 } else {
 606                         CDEBUG(D_RPCTRACE, "connect to %s with lustre_msg_v1\n",
 607                                obd2cli_tgt(imp->imp_obd));
 608                 }
 609
 610                 imp->imp_remote_handle =
 611                                 *lustre_msg_get_handle(request->rq_repmsg);
 612
 613                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 614                 ptlrpc_activate_import(imp);
 615                 GOTO(finish, rc = 0);
 616         } else {
 617                 spin_unlock(&imp->imp_lock);
 618         }
 619
 620         /* Determine what recovery state to move the import to. */
 621         if (MSG_CONNECT_RECONNECT & msg_flags) {
 622                 memset(&old_hdl, 0, sizeof(old_hdl));
 623                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 624                             sizeof (old_hdl))) {
 625                         CERROR("%s@%s didn't like our handle "LPX64
 626                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 627                                imp->imp_connection->c_remote_uuid.uuid,
 628                                imp->imp_dlm_handle.cookie);
 629                         GOTO(out, rc = -ENOTCONN);
 630                 }
 631
 632                 if (memcmp(&imp->imp_remote_handle,
 633                            lustre_msg_get_handle(request->rq_repmsg),
 634                            sizeof(imp->imp_remote_handle))) {
 635                         int level = D_ERROR;
 636                         /* Old MGC can reconnect to a restarted MGS */
 637                         if (strcmp(imp->imp_obd->obd_type->typ_name,
 638                                    LUSTRE_MGC_NAME) == 0) {
 639                                 level = D_CONFIG;
 640                         }
 641                         CDEBUG(level,
 642                                "%s@%s changed handle from "LPX64" to "LPX64
 643                                "; copying, but this may foreshadow disaster\n",
 644                                obd2cli_tgt(imp->imp_obd),
 645                                imp->imp_connection->c_remote_uuid.uuid,
 646                                imp->imp_remote_handle.cookie,
 647                                lustre_msg_get_handle(request->rq_repmsg)->
 648                                         cookie);
 649                         imp->imp_remote_handle =
 650                                      *lustre_msg_get_handle(request->rq_repmsg);
 651                 } else {
 652                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 653                                obd2cli_tgt(imp->imp_obd),
 654                                imp->imp_connection->c_remote_uuid.uuid);
 655                 }
 656
 657                 if (imp->imp_invalid) {
 658                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 659                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 660                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 661                                imp->imp_obd->obd_name,
 662                                obd2cli_tgt(imp->imp_obd));
 663
 664                         spin_lock(&imp->imp_lock);
 665                         imp->imp_resend_replay = 1;
 666                         spin_unlock(&imp->imp_lock);
 667
 668                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 669                 } else {
 670                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 671                 }
 672         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 673                 LASSERT(imp->imp_replayable);
 674                 imp->imp_remote_handle =
 675                                 *lustre_msg_get_handle(request->rq_repmsg);
 676                 imp->imp_last_replay_transno = 0;
 677                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 678         } else {
 679                 imp->imp_remote_handle =
 680                                 *lustre_msg_get_handle(request->rq_repmsg);
 681                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 682         }
 683
 684         /* Sanity checks for a reconnected import. */
 685         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 686                 CERROR("imp_replayable flag does not match server "
 687                        "after reconnect. We should LBUG right here.\n");
 688         }
 689
 690         if (lustre_msg_get_last_committed(request->rq_repmsg) <
 691             aa->pcaa_peer_committed) {
 692                 CERROR("%s went back in time (transno "LPD64
 693                        " was previously committed, server now claims "LPD64
 694                        ")!  See https://bugzilla.clusterfs.com/"
 695                        "long_list.cgi?buglist=9646\n",
 696                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 697                        lustre_msg_get_last_committed(request->rq_repmsg));
 698         }
 699
 700 finish:
 701         rc = ptlrpc_import_recovery_state_machine(imp);
 702         if (rc != 0) {
 703                 if (rc == -ENOTCONN) {
 704                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 705                                "invalidating and reconnecting\n",
 706                                obd2cli_tgt(imp->imp_obd),
 707                                imp->imp_connection->c_remote_uuid.uuid);
 708                         ptlrpc_connect_import(imp, NULL);
 709                         RETURN(0);
 710                 }
 711         } else {
 712                 struct obd_connect_data *ocd;
 713                 struct obd_export *exp;
 714                 int ret;
 715                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 716                                            RCL_SERVER);
 717                 /* server replied obd_connect_data is always bigger */
 718                 ocd = req_capsule_server_sized_get(&request->rq_pill,
 719                                                    &RMF_CONNECT_DATA, ret);
 720
 721                 spin_lock(&imp->imp_lock);
 722                 list_del(&imp->imp_conn_current->oic_item);
 723                 list_add(&imp->imp_conn_current->oic_item, &imp->imp_conn_list);
 724                 imp->imp_last_success_conn =
 725                         imp->imp_conn_current->oic_last_attempt;
 726
 727                 if (ocd == NULL) {
 728                         spin_unlock(&imp->imp_lock);
 729                         CERROR("Wrong connect data from server\n");
 730                         rc = -EPROTO;
 731                         GOTO(out, rc);
 732                 }
 733
 734                 imp->imp_connect_data = *ocd;
 735
 736                 exp = class_conn2export(&imp->imp_dlm_handle);
 737                 spin_unlock(&imp->imp_lock);
 738
 739                 /* check that server granted subset of flags we asked for. */
 740                 LASSERTF((ocd->ocd_connect_flags &
 741                           imp->imp_connect_flags_orig) ==
 742                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 743                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 744
 745                 if (!exp) {
 746                         /* This could happen if export is cleaned during the
 747                            connect attempt */
 748                         CERROR("Missing export for %s\n",
 749                                imp->imp_obd->obd_name);
 750                         GOTO(out, rc = -ENODEV);
 751                 }
 752                 exp->exp_connect_flags = ocd->ocd_connect_flags;
 753                 imp->imp_obd->obd_self_export->exp_connect_flags =
 754                                                         ocd->ocd_connect_flags;
 755                 class_export_put(exp);
 756
 757                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 758
 759                 if (!ocd->ocd_ibits_known &&
 760                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 761                         CERROR("Inodebits aware server returned zero compatible"
 762                                " bits?\n");
 763
 764                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 765                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
 766                                         LUSTRE_VERSION_OFFSET_WARN ||
 767                      ocd->ocd_version < LUSTRE_VERSION_CODE -
 768                                         LUSTRE_VERSION_OFFSET_WARN)) {
 769                         /* Sigh, some compilers do not like #ifdef in the middle
 770                            of macro arguments */
 771 #ifdef __KERNEL__
 772                         const char *older =
 773                                 "older. Consider upgrading this client";
 774 #else
 775                         const char *older =
 776                                 "older. Consider recompiling this application";
 777 #endif
 778                         const char *newer = "newer than client version";
 779
 780                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 781                                       "is much %s (%s)\n",
 782                                       obd2cli_tgt(imp->imp_obd),
 783                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 784                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 785                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 786                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
 787                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
 788                                       newer : older, LUSTRE_VERSION_STRING);
 789                 }
 790
 791                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
 792                         cli->cl_max_pages_per_rpc =
 793                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
 794                 }
 795
 796                 imp->imp_obd->obd_namespace->ns_connect_flags =
 797                                                         ocd->ocd_connect_flags;
 798                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
 799                                                         ocd->ocd_connect_flags;
 800
 801                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 802                         (cli->cl_max_pages_per_rpc > 0));
 803         }
 804
 805 out:
 806         if (rc != 0) {
 807                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 808                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov &&
 809                     (request->rq_import_generation == imp->imp_generation))
 810                         ptlrpc_deactivate_import(imp);
 811
 812                 if ((imp->imp_recon_bk && imp->imp_last_recon) ||
 813                     (rc == -EACCES)) {
 814                         /*
 815                          * Give up trying to reconnect
 816                          * EACCES means client has no permission for connection
 817                          */
 818                         imp->imp_obd->obd_no_recov = 1;
 819                         ptlrpc_deactivate_import(imp);
 820                 }
 821
 822                 if (rc == -EPROTO) {
 823                         struct obd_connect_data *ocd;
 824
 825                         /* reply message might not be ready */
 826                         if (request->rq_repmsg == NULL)
 827                                 RETURN(-EPROTO);
 828
 829                         ocd = req_capsule_server_get(&request->rq_pill,
 830                                                      &RMF_CONNECT_DATA);
 831                         if (ocd &&
 832                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 833                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
 834                            /* Actually servers are only supposed to refuse
 835                               connection from liblustre clients, so we should
 836                               never see this from VFS context */
 837                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
 838                                         "(%d.%d.%d.%d)"
 839                                         " refused connection from this client "
 840                                         "with an incompatible version (%s).  "
 841                                         "Client must be recompiled\n",
 842                                         obd2cli_tgt(imp->imp_obd),
 843                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 844                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 845                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 846                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
 847                                         LUSTRE_VERSION_STRING);
 848                                 ptlrpc_deactivate_import(imp);
 849                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
 850                         }
 851                         RETURN(-EPROTO);
 852                 }
 853
 854                 ptlrpc_maybe_ping_import_soon(imp);
 855
 856                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
 857                        obd2cli_tgt(imp->imp_obd),
 858                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
 859         }
 860
 861         spin_lock(&imp->imp_lock);
 862         imp->imp_last_recon = 0;
 863         spin_unlock(&imp->imp_lock);
 864
 865         cfs_waitq_signal(&imp->imp_recovery_waitq);
 866         RETURN(rc);
 867 }
 868
 869 static int completed_replay_interpret(struct ptlrpc_request *req,
 870                                     void * data, int rc)
 871 {
 872         ENTRY;
 873         atomic_dec(&req->rq_import->imp_replay_inflight);
 874         if (req->rq_status == 0) {
 875                 ptlrpc_import_recovery_state_machine(req->rq_import);
 876         } else {
 877                 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
 878                        "reconnecting\n",
 879                        req->rq_import->imp_obd->obd_name, req->rq_status);
 880                 ptlrpc_connect_import(req->rq_import, NULL);
 881         }
 882
 883         RETURN(0);
 884 }
 885
 886 static int signal_completed_replay(struct obd_import *imp)
 887 {
 888         struct ptlrpc_request *req;
 889         ENTRY;
 890
 891         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 892         atomic_inc(&imp->imp_replay_inflight);
 893
 894         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
 895                                         OBD_PING);
 896         if (req == NULL) {
 897                 atomic_dec(&imp->imp_replay_inflight);
 898                 RETURN(-ENOMEM);
 899         }
 900
 901         ptlrpc_request_set_replen(req);
 902         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
 903         lustre_msg_add_flags(req->rq_reqmsg,
 904                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
 905         req->rq_timeout *= 3;
 906         req->rq_interpret_reply = completed_replay_interpret;
 907
 908         ptlrpcd_add_req(req);
 909         RETURN(0);
 910 }
 911
 912 #ifdef __KERNEL__
 913 static int ptlrpc_invalidate_import_thread(void *data)
 914 {
 915         struct obd_import *imp = data;
 916
 917         ENTRY;
 918
 919         ptlrpc_daemonize("ll_imp_inval");
 920
 921         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
 922                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 923                imp->imp_connection->c_remote_uuid.uuid);
 924
 925         ptlrpc_invalidate_import(imp);
 926
 927         if (obd_dump_on_eviction) {
 928                 CERROR("dump the log upon eviction\n");
 929                 libcfs_debug_dumplog();
 930         }
 931
 932         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 933         ptlrpc_import_recovery_state_machine(imp);
 934
 935         RETURN(0);
 936 }
 937 #endif
 938
 939 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
 940 {
 941         int rc = 0;
 942         int inflight;
 943         char *target_start;
 944         int target_len;
 945
 946         ENTRY;
 947         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
 948                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 949                           &target_start, &target_len);
 950                 /* Don't care about MGC eviction */
 951                 if (strcmp(imp->imp_obd->obd_type->typ_name,
 952                            LUSTRE_MGC_NAME) != 0) {
 953                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
 954                                            "%.*s; in progress operations using "
 955                                            "this service will fail.\n",
 956                                            target_len, target_start);
 957                 }
 958                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
 959                        obd2cli_tgt(imp->imp_obd),
 960                        imp->imp_connection->c_remote_uuid.uuid);
 961
 962 #ifdef __KERNEL__
 963                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
 964                                        CLONE_VM | CLONE_FILES);
 965                 if (rc < 0)
 966                         CERROR("error starting invalidate thread: %d\n", rc);
 967                 else
 968                         rc = 0;
 969                 RETURN(rc);
 970 #else
 971                 ptlrpc_invalidate_import(imp);
 972
 973                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 974 #endif
 975         }
 976
 977         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
 978                 CDEBUG(D_HA, "replay requested by %s\n",
 979                        obd2cli_tgt(imp->imp_obd));
 980                 rc = ptlrpc_replay_next(imp, &inflight);
 981                 if (inflight == 0 &&
 982                     atomic_read(&imp->imp_replay_inflight) == 0) {
 983                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 984                         rc = ldlm_replay_locks(imp);
 985                         if (rc)
 986                                 GOTO(out, rc);
 987                 }
 988                 rc = 0;
 989         }
 990
 991         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
 992                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
 993                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
 994                         rc = signal_completed_replay(imp);
 995                         if (rc)
 996                                 GOTO(out, rc);
 997                 }
 998
 999         }
1000
1001         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1002                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1003                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1004                 }
1005         }
1006
1007         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1008                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1009                        obd2cli_tgt(imp->imp_obd),
1010                        imp->imp_connection->c_remote_uuid.uuid);
1011
1012                 rc = ptlrpc_resend(imp);
1013                 if (rc)
1014                         GOTO(out, rc);
1015                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1016                 ptlrpc_activate_import(imp);
1017
1018                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1019                           &target_start, &target_len);
1020                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1021                               "using nid %s.\n", imp->imp_obd->obd_name,
1022                               target_len, target_start,
1023                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1024         }
1025
1026         if (imp->imp_state == LUSTRE_IMP_FULL) {
1027                 cfs_waitq_signal(&imp->imp_recovery_waitq);
1028                 ptlrpc_wake_delayed(imp);
1029         }
1030
1031 out:
1032         RETURN(rc);
1033 }
1034
1035 static int back_to_sleep(void *unused)
1036 {
1037         return 0;
1038 }
1039
1040 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1041 {
1042         struct ptlrpc_request *req;
1043         int rq_opc, rc = 0;
1044         int nowait = imp->imp_obd->obd_force;
1045         ENTRY;
1046
1047         if (nowait)
1048                 GOTO(set_state, rc);
1049
1050         switch (imp->imp_connect_op) {
1051         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1052         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1053         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1054         default:
1055                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1056                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1057                 RETURN(-EINVAL);
1058         }
1059
1060         if (ptlrpc_import_in_recovery(imp)) {
1061                 struct l_wait_info lwi;
1062                 cfs_duration_t timeout;
1063                 if (imp->imp_server_timeout)
1064                         timeout = cfs_time_seconds(obd_timeout / 2);
1065                 else
1066                         timeout = cfs_time_seconds(obd_timeout);
1067
1068                 timeout = MAX(timeout * HZ, 1);
1069
1070                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1071                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1072                 rc = l_wait_event(imp->imp_recovery_waitq,
1073                                   !ptlrpc_import_in_recovery(imp), &lwi);
1074
1075         }
1076
1077         spin_lock(&imp->imp_lock);
1078         if (imp->imp_state != LUSTRE_IMP_FULL)
1079                 GOTO(out, 0);
1080
1081         spin_unlock(&imp->imp_lock);
1082
1083         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1084                                         LUSTRE_OBD_VERSION, rq_opc);
1085         if (req) {
1086                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1087                  * it fails.  We can get through the above with a down server
1088                  * if the client doesn't know the server is gone yet. */
1089                 req->rq_no_resend = 1;
1090 #ifdef CRAY_XT3
1091                 req->rq_timeout = obd_timeout / 3;
1092 #else
1093                 req->rq_timeout = 5;
1094 #endif
1095                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1096                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1097                 ptlrpc_request_set_replen(req);
1098                 rc = ptlrpc_queue_wait(req);
1099                 ptlrpc_req_finished(req);
1100         }
1101
1102 set_state:
1103         spin_lock(&imp->imp_lock);
1104 out:
1105         if (noclose)
1106                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1107         else
1108                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1109         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1110         imp->imp_conn_cnt = 0;
1111         imp->imp_last_recon = 0;
1112         spin_unlock(&imp->imp_lock);
1113
1114         RETURN(rc);
1115 }
1116