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 /* Must be called with imp_lock held! */
 167 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 168 {
 169         ENTRY;
 170         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 171
 172         if (imp->imp_invalid) {
 173                 spin_unlock(&imp->imp_lock);
 174                 EXIT;
 175                 return;
 176         }
 177
 178         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 179         imp->imp_invalid = 1;
 180         imp->imp_generation++;
 181         spin_unlock(&imp->imp_lock);
 182
 183         ptlrpc_abort_inflight(imp);
 184         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 185
 186         EXIT;
 187 }
 188
 189 /*
 190  * This acts as a barrier; all existing requests are rejected, and
 191  * no new requests will be accepted until the import is valid again.
 192  */
 193 void ptlrpc_deactivate_import(struct obd_import *imp)
 194 {
 195         spin_lock(&imp->imp_lock);
 196         ptlrpc_deactivate_and_unlock_import(imp);
 197 }
 198
 199 /*
 200  * This function will invalidate the import, if necessary, then block
 201  * for all the RPC completions, and finally notify the obd to
 202  * invalidate its state (ie cancel locks, clear pending requests,
 203  * etc).
 204  */
 205 void ptlrpc_invalidate_import(struct obd_import *imp)
 206 {
 207         struct l_wait_info lwi;
 208         int rc;
 209
 210         atomic_inc(&imp->imp_inval_count);
 211
 212         ptlrpc_deactivate_import(imp);
 213
 214         LASSERT(imp->imp_invalid);
 215
 216         /* wait for all requests to error out and call completion callbacks */
 217         lwi = LWI_TIMEOUT_INTERVAL(cfs_timeout_cap(cfs_time_seconds(obd_timeout)),
 218                                    HZ, NULL, NULL);
 219         rc = l_wait_event(imp->imp_recovery_waitq,
 220                           (atomic_read(&imp->imp_inflight) == 0), &lwi);
 221
 222         if (rc)
 223                 CDEBUG(D_HA, "%s: rc = %d waiting for callback (%d != 0)\n",
 224                        obd2cli_tgt(imp->imp_obd), rc,
 225                        atomic_read(&imp->imp_inflight));
 226
 227         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 228         sptlrpc_import_flush_all_ctx(imp);
 229
 230         atomic_dec(&imp->imp_inval_count);
 231         cfs_waitq_signal(&imp->imp_recovery_waitq);
 232 }
 233
 234 /* unset imp_invalid */
 235 void ptlrpc_activate_import(struct obd_import *imp)
 236 {
 237         struct obd_device *obd = imp->imp_obd;
 238
 239         spin_lock(&imp->imp_lock);
 240         imp->imp_invalid = 0;
 241         ptlrpc_activate_timeouts(imp);
 242         spin_unlock(&imp->imp_lock);
 243         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 244 }
 245
 246 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 247 {
 248         ENTRY;
 249
 250         LASSERT(!imp->imp_dlm_fake);
 251
 252         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 253                 if (!imp->imp_replayable) {
 254                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 255                                "auto-deactivating\n",
 256                                obd2cli_tgt(imp->imp_obd),
 257                                imp->imp_connection->c_remote_uuid.uuid,
 258                                imp->imp_obd->obd_name);
 259                         ptlrpc_deactivate_import(imp);
 260                 }
 261
 262                 CDEBUG(D_HA, "%s: waking up pinger\n",
 263                        obd2cli_tgt(imp->imp_obd));
 264
 265                 spin_lock(&imp->imp_lock);
 266                 imp->imp_force_verify = 1;
 267                 spin_unlock(&imp->imp_lock);
 268
 269                 ptlrpc_pinger_wake_up();
 270         }
 271         EXIT;
 272 }
 273
 274 static int import_select_connection(struct obd_import *imp)
 275 {
 276         struct obd_import_conn *imp_conn = NULL, *conn;
 277         struct obd_export *dlmexp;
 278         ENTRY;
 279
 280         spin_lock(&imp->imp_lock);
 281
 282         if (list_empty(&imp->imp_conn_list)) {
 283                 CERROR("%s: no connections available\n",
 284                         imp->imp_obd->obd_name);
 285                 spin_unlock(&imp->imp_lock);
 286                 RETURN(-EINVAL);
 287         }
 288
 289         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 290                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 291                        imp->imp_obd->obd_name,
 292                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 293                        conn->oic_last_attempt);
 294                 /* Throttle the reconnect rate to once per RECONNECT_INTERVAL */
 295                 if (cfs_time_before_64(conn->oic_last_attempt +
 296                                        RECONNECT_INTERVAL * HZ,
 297                                        cfs_time_current_64())) {
 298                         /* If we have never tried this connection since the
 299                            the last successful attempt, go with this one */
 300                         if (cfs_time_beforeq_64(conn->oic_last_attempt,
 301                                                imp->imp_last_success_conn)) {
 302                                 imp_conn = conn;
 303                                 break;
 304                         }
 305
 306                         /* Both of these connections have already been tried
 307                            since the last successful connection; just choose the
 308                            least recently used */
 309                         if (!imp_conn)
 310                                 imp_conn = conn;
 311                         else if (cfs_time_before_64(conn->oic_last_attempt,
 312                                                     imp_conn->oic_last_attempt))
 313                                 imp_conn = conn;
 314                 }
 315         }
 316
 317         /* if not found, simply choose the current one */
 318         if (!imp_conn) {
 319                 LASSERT(imp->imp_conn_current);
 320                 imp_conn = imp->imp_conn_current;
 321         }
 322         LASSERT(imp_conn->oic_conn);
 323
 324         imp_conn->oic_last_attempt = cfs_time_current_64();
 325
 326         /* switch connection, don't mind if it's same as the current one */
 327         if (imp->imp_connection)
 328                 ptlrpc_put_connection(imp->imp_connection);
 329         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 330
 331         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 332         LASSERT(dlmexp != NULL);
 333         if (dlmexp->exp_connection)
 334                 ptlrpc_put_connection(dlmexp->exp_connection);
 335         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 336         class_export_put(dlmexp);
 337
 338         if (imp->imp_conn_current != imp_conn) {
 339                 if (imp->imp_conn_current)
 340                         LCONSOLE_INFO("Changing connection for %s to %s/%s\n",
 341                                       imp->imp_obd->obd_name,
 342                                       imp_conn->oic_uuid.uuid,
 343                                       libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 344                 imp->imp_conn_current = imp_conn;
 345         }
 346
 347         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 348                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 349                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 350
 351         spin_unlock(&imp->imp_lock);
 352
 353         RETURN(0);
 354 }
 355
 356 /*
 357  * must be called under imp_lock
 358  */
 359 int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 360 {
 361         struct ptlrpc_request *req;
 362         struct list_head *tmp;
 363
 364         if (list_empty(&imp->imp_replay_list))
 365                 return 0;
 366         tmp = imp->imp_replay_list.next;
 367         req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 368         *transno = req->rq_transno;
 369         if (req->rq_transno == 0) {
 370                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 371                 LBUG();
 372         }
 373
 374         return 1;
 375 }
 376
 377 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
 378 {
 379         struct obd_device *obd = imp->imp_obd;
 380         int initial_connect = 0;
 381         int set_transno = 0;
 382         __u64 committed_before_reconnect = 0;
 383         struct ptlrpc_request *request;
 384         char *bufs[] = { NULL,
 385                          obd2cli_tgt(imp->imp_obd),
 386                          obd->obd_uuid.uuid,
 387                          (char *)&imp->imp_dlm_handle,
 388                          (char *)&imp->imp_connect_data };
 389         struct ptlrpc_connect_async_args *aa;
 390         int rc;
 391         ENTRY;
 392
 393         spin_lock(&imp->imp_lock);
 394         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 395                 spin_unlock(&imp->imp_lock);
 396                 CERROR("can't connect to a closed import\n");
 397                 RETURN(-EINVAL);
 398         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 399                 spin_unlock(&imp->imp_lock);
 400                 CERROR("already connected\n");
 401                 RETURN(0);
 402         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 403                 spin_unlock(&imp->imp_lock);
 404                 CERROR("already connecting\n");
 405                 RETURN(-EALREADY);
 406         }
 407
 408         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 409
 410         imp->imp_conn_cnt++;
 411         imp->imp_resend_replay = 0;
 412
 413         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 414                 initial_connect = 1;
 415         else
 416                 committed_before_reconnect = imp->imp_peer_committed_transno;
 417
 418         set_transno = ptlrpc_first_transno(imp, &imp->imp_connect_data.ocd_transno);
 419         spin_unlock(&imp->imp_lock);
 420
 421         if (new_uuid) {
 422                 struct obd_uuid uuid;
 423
 424                 obd_str2uuid(&uuid, new_uuid);
 425                 rc = import_set_conn_priority(imp, &uuid);
 426                 if (rc)
 427                         GOTO(out, rc);
 428         }
 429
 430         rc = import_select_connection(imp);
 431         if (rc)
 432                 GOTO(out, rc);
 433
 434         /* last in connection list */
 435         if (imp->imp_conn_current->oic_item.next == &imp->imp_conn_list) {
 436                 if (imp->imp_initial_recov_bk && initial_connect) {
 437                         CDEBUG(D_HA, "Last connection attempt (%d) for %s\n",
 438                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 439                         /* Don't retry if connect fails */
 440                         rc = 0;
 441                         obd_set_info_async(obd->obd_self_export,
 442                                            strlen(KEY_INIT_RECOV),
 443                                            KEY_INIT_RECOV,
 444                                            sizeof(rc), &rc, NULL);
 445                 }
 446                 if (imp->imp_recon_bk) {
 447                         CDEBUG(D_HA, "Last reconnection attempt (%d) for %s\n",
 448                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 449                         spin_lock(&imp->imp_lock);
 450                         imp->imp_last_recon = 1;
 451                         spin_unlock(&imp->imp_lock);
 452                 }
 453         }
 454
 455         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 456         if (rc)
 457                 GOTO(out, rc);
 458
 459         /* Reset connect flags to the originally requested flags, in case
 460          * the server is updated on-the-fly we will get the new features. */
 461         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 462         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 463                            &obd->obd_uuid, &imp->imp_connect_data);
 464         if (rc)
 465                 GOTO(out, rc);
 466
 467         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 468         if (request == NULL)
 469                 GOTO(out, rc = -ENOMEM);
 470
 471         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 472                                       imp->imp_connect_op, bufs, NULL);
 473         if (rc) {
 474                 ptlrpc_request_free(request);
 475                 GOTO(out, rc);
 476         }
 477
 478 #ifndef __KERNEL__
 479         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 480 #endif
 481         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 482
 483         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 484         /* Allow a slightly larger reply for future growth compatibility */
 485         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 486                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 487         ptlrpc_request_set_replen(request);
 488         request->rq_interpret_reply = ptlrpc_connect_interpret;
 489
 490         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 491         aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
 492         memset(aa, 0, sizeof *aa);
 493
 494         aa->pcaa_peer_committed = committed_before_reconnect;
 495         aa->pcaa_initial_connect = initial_connect;
 496
 497         if (aa->pcaa_initial_connect) {
 498                 spin_lock(&imp->imp_lock);
 499                 imp->imp_replayable = 1;
 500                 spin_unlock(&imp->imp_lock);
 501                 /* On an initial connect, we don't know which one of a
 502                    failover server pair is up.  Don't wait long. */
 503 #ifdef CRAY_XT3
 504                 request->rq_timeout = max((int)(obd_timeout / 2), 5);
 505 #else
 506                 request->rq_timeout = max((int)(obd_timeout / 20), 5);
 507 #endif
 508                 lustre_msg_add_op_flags(request->rq_reqmsg,
 509                                         MSG_CONNECT_INITIAL);
 510         }
 511
 512         if (set_transno)
 513                 lustre_msg_add_op_flags(request->rq_reqmsg,
 514                                         MSG_CONNECT_TRANSNO);
 515
 516         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request");
 517         ptlrpcd_add_req(request);
 518         rc = 0;
 519 out:
 520         if (rc != 0) {
 521                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 522         }
 523
 524         RETURN(rc);
 525 }
 526 EXPORT_SYMBOL(ptlrpc_connect_import);
 527
 528 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 529 {
 530 #ifdef __KERNEL__
 531         struct obd_import_conn *imp_conn;
 532 #endif
 533         int wake_pinger = 0;
 534
 535         ENTRY;
 536
 537         spin_lock(&imp->imp_lock);
 538         if (list_empty(&imp->imp_conn_list))
 539                 GOTO(unlock, 0);
 540
 541 #ifdef __KERNEL__
 542         imp_conn = list_entry(imp->imp_conn_list.prev,
 543                               struct obd_import_conn,
 544                               oic_item);
 545
 546         /* XXX: When the failover node is the primary node, it is possible
 547          * to have two identical connections in imp_conn_list. We must
 548          * compare not conn's pointers but NIDs, otherwise we can defeat
 549          * connection throttling. (See bug 14774.) */
 550         if (imp->imp_conn_current->oic_conn->c_self !=
 551                                 imp_conn->oic_conn->c_self) {
 552                 ptlrpc_ping_import_soon(imp);
 553                 wake_pinger = 1;
 554         }
 555 #else
 556         /* liblustre has no pinger thead, so we wakup pinger anyway */
 557         wake_pinger = 1;
 558 #endif
 559
 560  unlock:
 561         spin_unlock(&imp->imp_lock);
 562
 563         if (wake_pinger)
 564                 ptlrpc_pinger_wake_up();
 565
 566         EXIT;
 567 }
 568
 569 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
 570                                     void * data, int rc)
 571 {
 572         struct ptlrpc_connect_async_args *aa = data;
 573         struct obd_import *imp = request->rq_import;
 574         struct client_obd *cli = &imp->imp_obd->u.cli;
 575         struct lustre_handle old_hdl;
 576         int msg_flags;
 577         ENTRY;
 578
 579         spin_lock(&imp->imp_lock);
 580         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 581                 spin_unlock(&imp->imp_lock);
 582                 RETURN(0);
 583         }
 584         spin_unlock(&imp->imp_lock);
 585
 586         if (rc)
 587                 GOTO(out, rc);
 588
 589         LASSERT(imp->imp_conn_current);
 590
 591         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 592
 593         /* All imports are pingable */
 594         spin_lock(&imp->imp_lock);
 595         imp->imp_pingable = 1;
 596
 597         if (aa->pcaa_initial_connect) {
 598                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 599                         imp->imp_replayable = 1;
 600                         spin_unlock(&imp->imp_lock);
 601                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 602                                obd2cli_tgt(imp->imp_obd));
 603                 } else {
 604                         imp->imp_replayable = 0;
 605                         spin_unlock(&imp->imp_lock);
 606                 }
 607
 608                 if (msg_flags & MSG_CONNECT_NEXT_VER) {
 609                         imp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
 610                         CDEBUG(D_RPCTRACE, "connect to %s with lustre_msg_v2\n",
 611                                obd2cli_tgt(imp->imp_obd));
 612                 } else {
 613                         CDEBUG(D_RPCTRACE, "connect to %s with lustre_msg_v1\n",
 614                                obd2cli_tgt(imp->imp_obd));
 615                 }
 616
 617                 imp->imp_remote_handle =
 618                                 *lustre_msg_get_handle(request->rq_repmsg);
 619
 620                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 621                 ptlrpc_activate_import(imp);
 622                 GOTO(finish, rc = 0);
 623         } else {
 624                 spin_unlock(&imp->imp_lock);
 625         }
 626
 627         /* Determine what recovery state to move the import to. */
 628         if (MSG_CONNECT_RECONNECT & msg_flags) {
 629                 memset(&old_hdl, 0, sizeof(old_hdl));
 630                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 631                             sizeof (old_hdl))) {
 632                         CERROR("%s@%s didn't like our handle "LPX64
 633                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 634                                imp->imp_connection->c_remote_uuid.uuid,
 635                                imp->imp_dlm_handle.cookie);
 636                         GOTO(out, rc = -ENOTCONN);
 637                 }
 638
 639                 if (memcmp(&imp->imp_remote_handle,
 640                            lustre_msg_get_handle(request->rq_repmsg),
 641                            sizeof(imp->imp_remote_handle))) {
 642
 643                         CWARN("%s@%s changed server handle from "
 644                                LPX64" to "LPX64" - evicting.\n",
 645                                obd2cli_tgt(imp->imp_obd),
 646                                imp->imp_connection->c_remote_uuid.uuid,
 647                                imp->imp_remote_handle.cookie,
 648                                lustre_msg_get_handle(request->rq_repmsg)->
 649                                          cookie);
 650                         imp->imp_remote_handle =
 651                                      *lustre_msg_get_handle(request->rq_repmsg);
 652
 653                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 654                         GOTO(finish, rc = 0);
 655                 } else {
 656                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 657                                obd2cli_tgt(imp->imp_obd),
 658                                imp->imp_connection->c_remote_uuid.uuid);
 659                 }
 660
 661                 if (imp->imp_invalid) {
 662                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 663                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 664                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 665                                imp->imp_obd->obd_name,
 666                                obd2cli_tgt(imp->imp_obd));
 667
 668                         spin_lock(&imp->imp_lock);
 669                         imp->imp_resend_replay = 1;
 670                         spin_unlock(&imp->imp_lock);
 671
 672                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 673                 } else {
 674                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 675                 }
 676         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 677                 LASSERT(imp->imp_replayable);
 678                 imp->imp_remote_handle =
 679                                 *lustre_msg_get_handle(request->rq_repmsg);
 680                 imp->imp_last_replay_transno = 0;
 681                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 682         } else {
 683                 imp->imp_remote_handle =
 684                                 *lustre_msg_get_handle(request->rq_repmsg);
 685                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 686         }
 687
 688         /* Sanity checks for a reconnected import. */
 689         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 690                 CERROR("imp_replayable flag does not match server "
 691                        "after reconnect. We should LBUG right here.\n");
 692         }
 693
 694         if (lustre_msg_get_last_committed(request->rq_repmsg) <
 695             aa->pcaa_peer_committed) {
 696                 CERROR("%s went back in time (transno "LPD64
 697                        " was previously committed, server now claims "LPD64
 698                        ")!  See https://bugzilla.clusterfs.com/"
 699                        "long_list.cgi?buglist=9646\n",
 700                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 701                        lustre_msg_get_last_committed(request->rq_repmsg));
 702         }
 703
 704 finish:
 705         rc = ptlrpc_import_recovery_state_machine(imp);
 706         if (rc != 0) {
 707                 if (rc == -ENOTCONN) {
 708                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 709                                "invalidating and reconnecting\n",
 710                                obd2cli_tgt(imp->imp_obd),
 711                                imp->imp_connection->c_remote_uuid.uuid);
 712                         ptlrpc_connect_import(imp, NULL);
 713                         RETURN(0);
 714                 }
 715         } else {
 716                 struct obd_connect_data *ocd;
 717                 struct obd_export *exp;
 718                 int ret;
 719                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 720                                            RCL_SERVER);
 721                 /* server replied obd_connect_data is always bigger */
 722                 ocd = req_capsule_server_sized_get(&request->rq_pill,
 723                                                    &RMF_CONNECT_DATA, ret);
 724
 725                 spin_lock(&imp->imp_lock);
 726                 list_del(&imp->imp_conn_current->oic_item);
 727                 list_add(&imp->imp_conn_current->oic_item, &imp->imp_conn_list);
 728                 imp->imp_last_success_conn =
 729                         imp->imp_conn_current->oic_last_attempt;
 730
 731                 if (ocd == NULL) {
 732                         spin_unlock(&imp->imp_lock);
 733                         CERROR("Wrong connect data from server\n");
 734                         rc = -EPROTO;
 735                         GOTO(out, rc);
 736                 }
 737
 738                 imp->imp_connect_data = *ocd;
 739
 740                 exp = class_conn2export(&imp->imp_dlm_handle);
 741                 spin_unlock(&imp->imp_lock);
 742
 743                 /* check that server granted subset of flags we asked for. */
 744                 LASSERTF((ocd->ocd_connect_flags &
 745                           imp->imp_connect_flags_orig) ==
 746                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 747                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 748
 749                 if (!exp) {
 750                         /* This could happen if export is cleaned during the
 751                            connect attempt */
 752                         CERROR("Missing export for %s\n",
 753                                imp->imp_obd->obd_name);
 754                         GOTO(out, rc = -ENODEV);
 755                 }
 756                 exp->exp_connect_flags = ocd->ocd_connect_flags;
 757                 imp->imp_obd->obd_self_export->exp_connect_flags =
 758                                                         ocd->ocd_connect_flags;
 759                 class_export_put(exp);
 760
 761                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 762
 763                 if (!ocd->ocd_ibits_known &&
 764                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 765                         CERROR("Inodebits aware server returned zero compatible"
 766                                " bits?\n");
 767
 768                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 769                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
 770                                         LUSTRE_VERSION_OFFSET_WARN ||
 771                      ocd->ocd_version < LUSTRE_VERSION_CODE -
 772                                         LUSTRE_VERSION_OFFSET_WARN)) {
 773                         /* Sigh, some compilers do not like #ifdef in the middle
 774                            of macro arguments */
 775 #ifdef __KERNEL__
 776                         const char *older =
 777                                 "older. Consider upgrading this client";
 778 #else
 779                         const char *older =
 780                                 "older. Consider recompiling this application";
 781 #endif
 782                         const char *newer = "newer than client version";
 783
 784                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 785                                       "is much %s (%s)\n",
 786                                       obd2cli_tgt(imp->imp_obd),
 787                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 788                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 789                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 790                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
 791                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
 792                                       newer : older, LUSTRE_VERSION_STRING);
 793                 }
 794
 795                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
 796                         cli->cl_max_pages_per_rpc =
 797                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
 798                 }
 799
 800                 imp->imp_obd->obd_namespace->ns_connect_flags =
 801                                                         ocd->ocd_connect_flags;
 802                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
 803                                                         ocd->ocd_connect_flags;
 804
 805                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 806                         (cli->cl_max_pages_per_rpc > 0));
 807         }
 808
 809 out:
 810         if (rc != 0) {
 811                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 812                 spin_lock(&imp->imp_lock);
 813                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov &&
 814                     (request->rq_import_generation == imp->imp_generation))
 815                         ptlrpc_deactivate_and_unlock_import(imp);
 816                 else
 817                         spin_unlock(&imp->imp_lock);
 818
 819                 if ((imp->imp_recon_bk && imp->imp_last_recon) ||
 820                     (rc == -EACCES)) {
 821                         /*
 822                          * Give up trying to reconnect
 823                          * EACCES means client has no permission for connection
 824                          */
 825                         imp->imp_obd->obd_no_recov = 1;
 826                         ptlrpc_deactivate_import(imp);
 827                 }
 828
 829                 if (rc == -EPROTO) {
 830                         struct obd_connect_data *ocd;
 831
 832                         /* reply message might not be ready */
 833                         if (request->rq_repmsg == NULL)
 834                                 RETURN(-EPROTO);
 835
 836                         ocd = req_capsule_server_get(&request->rq_pill,
 837                                                      &RMF_CONNECT_DATA);
 838                         if (ocd &&
 839                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 840                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
 841                            /* Actually servers are only supposed to refuse
 842                               connection from liblustre clients, so we should
 843                               never see this from VFS context */
 844                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
 845                                         "(%d.%d.%d.%d)"
 846                                         " refused connection from this client "
 847                                         "with an incompatible version (%s).  "
 848                                         "Client must be recompiled\n",
 849                                         obd2cli_tgt(imp->imp_obd),
 850                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 851                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 852                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 853                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
 854                                         LUSTRE_VERSION_STRING);
 855                                 ptlrpc_deactivate_import(imp);
 856                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
 857                         }
 858                         RETURN(-EPROTO);
 859                 }
 860
 861                 ptlrpc_maybe_ping_import_soon(imp);
 862
 863                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
 864                        obd2cli_tgt(imp->imp_obd),
 865                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
 866         }
 867
 868         spin_lock(&imp->imp_lock);
 869         imp->imp_last_recon = 0;
 870         spin_unlock(&imp->imp_lock);
 871
 872         cfs_waitq_signal(&imp->imp_recovery_waitq);
 873         RETURN(rc);
 874 }
 875
 876 static int completed_replay_interpret(struct ptlrpc_request *req,
 877                                     void * data, int rc)
 878 {
 879         ENTRY;
 880         atomic_dec(&req->rq_import->imp_replay_inflight);
 881         if (req->rq_status == 0) {
 882                 ptlrpc_import_recovery_state_machine(req->rq_import);
 883         } else {
 884                 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
 885                        "reconnecting\n",
 886                        req->rq_import->imp_obd->obd_name, req->rq_status);
 887                 ptlrpc_connect_import(req->rq_import, NULL);
 888         }
 889
 890         RETURN(0);
 891 }
 892
 893 static int signal_completed_replay(struct obd_import *imp)
 894 {
 895         struct ptlrpc_request *req;
 896         ENTRY;
 897
 898         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 899         atomic_inc(&imp->imp_replay_inflight);
 900
 901         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
 902                                         OBD_PING);
 903         if (req == NULL) {
 904                 atomic_dec(&imp->imp_replay_inflight);
 905                 RETURN(-ENOMEM);
 906         }
 907
 908         ptlrpc_request_set_replen(req);
 909         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
 910         lustre_msg_add_flags(req->rq_reqmsg,
 911                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
 912         req->rq_timeout *= 3;
 913         req->rq_interpret_reply = completed_replay_interpret;
 914
 915         ptlrpcd_add_req(req);
 916         RETURN(0);
 917 }
 918
 919 #ifdef __KERNEL__
 920 static int ptlrpc_invalidate_import_thread(void *data)
 921 {
 922         struct obd_import *imp = data;
 923
 924         ENTRY;
 925
 926         ptlrpc_daemonize("ll_imp_inval");
 927
 928         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
 929                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 930                imp->imp_connection->c_remote_uuid.uuid);
 931
 932         ptlrpc_invalidate_import(imp);
 933
 934         if (obd_dump_on_eviction) {
 935                 CERROR("dump the log upon eviction\n");
 936                 libcfs_debug_dumplog();
 937         }
 938
 939         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 940         ptlrpc_import_recovery_state_machine(imp);
 941
 942         RETURN(0);
 943 }
 944 #endif
 945
 946 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
 947 {
 948         int rc = 0;
 949         int inflight;
 950         char *target_start;
 951         int target_len;
 952
 953         ENTRY;
 954         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
 955                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 956                           &target_start, &target_len);
 957                 /* Don't care about MGC eviction */
 958                 if (strcmp(imp->imp_obd->obd_type->typ_name,
 959                            LUSTRE_MGC_NAME) != 0) {
 960                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
 961                                            "%.*s; in progress operations using "
 962                                            "this service will fail.\n",
 963                                            target_len, target_start);
 964                 }
 965                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
 966                        obd2cli_tgt(imp->imp_obd),
 967                        imp->imp_connection->c_remote_uuid.uuid);
 968
 969 #ifdef __KERNEL__
 970                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
 971                                        CLONE_VM | CLONE_FILES);
 972                 if (rc < 0)
 973                         CERROR("error starting invalidate thread: %d\n", rc);
 974                 else
 975                         rc = 0;
 976                 RETURN(rc);
 977 #else
 978                 ptlrpc_invalidate_import(imp);
 979
 980                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 981 #endif
 982         }
 983
 984         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
 985                 CDEBUG(D_HA, "replay requested by %s\n",
 986                        obd2cli_tgt(imp->imp_obd));
 987                 rc = ptlrpc_replay_next(imp, &inflight);
 988                 if (inflight == 0 &&
 989                     atomic_read(&imp->imp_replay_inflight) == 0) {
 990                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 991                         rc = ldlm_replay_locks(imp);
 992                         if (rc)
 993                                 GOTO(out, rc);
 994                 }
 995                 rc = 0;
 996         }
 997
 998         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
 999                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1000                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1001                         rc = signal_completed_replay(imp);
1002                         if (rc)
1003                                 GOTO(out, rc);
1004                 }
1005
1006         }
1007
1008         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1009                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1010                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1011                 }
1012         }
1013
1014         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1015                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1016                        obd2cli_tgt(imp->imp_obd),
1017                        imp->imp_connection->c_remote_uuid.uuid);
1018
1019                 rc = ptlrpc_resend(imp);
1020                 if (rc)
1021                         GOTO(out, rc);
1022                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1023                 ptlrpc_activate_import(imp);
1024
1025                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1026                           &target_start, &target_len);
1027                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1028                               "using nid %s.\n", imp->imp_obd->obd_name,
1029                               target_len, target_start,
1030                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1031         }
1032
1033         if (imp->imp_state == LUSTRE_IMP_FULL) {
1034                 cfs_waitq_signal(&imp->imp_recovery_waitq);
1035                 ptlrpc_wake_delayed(imp);
1036         }
1037
1038 out:
1039         RETURN(rc);
1040 }
1041
1042 static int back_to_sleep(void *unused)
1043 {
1044         return 0;
1045 }
1046
1047 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1048 {
1049         struct ptlrpc_request *req;
1050         int rq_opc, rc = 0;
1051         int nowait = imp->imp_obd->obd_force;
1052         ENTRY;
1053
1054         if (nowait)
1055                 GOTO(set_state, rc);
1056
1057         switch (imp->imp_connect_op) {
1058         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1059         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1060         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1061         default:
1062                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1063                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1064                 RETURN(-EINVAL);
1065         }
1066
1067         if (ptlrpc_import_in_recovery(imp)) {
1068                 struct l_wait_info lwi;
1069                 cfs_duration_t timeout;
1070                 if (imp->imp_server_timeout)
1071                         timeout = cfs_time_seconds(obd_timeout / 2);
1072                 else
1073                         timeout = cfs_time_seconds(obd_timeout);
1074
1075                 timeout = MAX(timeout * HZ, 1);
1076
1077                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1078                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1079                 rc = l_wait_event(imp->imp_recovery_waitq,
1080                                   !ptlrpc_import_in_recovery(imp), &lwi);
1081
1082         }
1083
1084         spin_lock(&imp->imp_lock);
1085         if (imp->imp_state != LUSTRE_IMP_FULL)
1086                 GOTO(out, 0);
1087
1088         spin_unlock(&imp->imp_lock);
1089
1090         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1091                                         LUSTRE_OBD_VERSION, rq_opc);
1092         if (req) {
1093                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1094                  * it fails.  We can get through the above with a down server
1095                  * if the client doesn't know the server is gone yet. */
1096                 req->rq_no_resend = 1;
1097 #ifdef CRAY_XT3
1098                 req->rq_timeout = obd_timeout / 3;
1099 #else
1100                 req->rq_timeout = 5;
1101 #endif
1102                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1103                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1104                 ptlrpc_request_set_replen(req);
1105                 rc = ptlrpc_queue_wait(req);
1106                 ptlrpc_req_finished(req);
1107         }
1108
1109 set_state:
1110         spin_lock(&imp->imp_lock);
1111 out:
1112         if (noclose)
1113                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1114         else
1115                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1116         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1117         imp->imp_conn_cnt = 0;
1118         imp->imp_last_recon = 0;
1119         spin_unlock(&imp->imp_lock);
1120
1121         RETURN(rc);
1122 }
1123