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