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_cksum.h>
  38 #include <obd_class.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                 if (imp->imp_replayable) {
 131                         LCONSOLE_WARN("%s: Connection to service %.*s via nid "
 132                                "%s was lost; in progress operations using this "
 133                                "service will wait for recovery to complete.\n",
 134                                imp->imp_obd->obd_name, target_len, target_start,
 135                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 136                 } else {
 137                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to service "
 138                                            "%.*s via nid %s was lost; in progress"
 139                                            "operations using this service will"
 140                                            "fail.\n",
 141                                            imp->imp_obd->obd_name,
 142                                            target_len, target_start,
 143                                  libcfs_nid2str(imp->imp_connection->c_peer.nid));
 144                 }
 145                 ptlrpc_deactivate_timeouts(imp);
 146                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 147                 spin_unlock(&imp->imp_lock);
 148
 149                 if (obd_dump_on_timeout)
 150                         libcfs_debug_dumplog();
 151
 152                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 153                 rc = 1;
 154         } else {
 155                 spin_unlock(&imp->imp_lock);
 156                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 157                        imp->imp_client->cli_name, imp,
 158                        (imp->imp_state == LUSTRE_IMP_FULL &&
 159                         imp->imp_conn_cnt > conn_cnt) ?
 160                        "reconnected" : "not connected", imp->imp_conn_cnt,
 161                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 162         }
 163
 164         return rc;
 165 }
 166
 167 /* Must be called with imp_lock held! */
 168 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 169 {
 170         ENTRY;
 171         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 172
 173         if (imp->imp_invalid) {
 174                 spin_unlock(&imp->imp_lock);
 175                 EXIT;
 176                 return;
 177         }
 178
 179         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 180         imp->imp_invalid = 1;
 181         imp->imp_generation++;
 182         spin_unlock(&imp->imp_lock);
 183
 184         ptlrpc_abort_inflight(imp);
 185         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 186
 187         EXIT;
 188 }
 189
 190 /*
 191  * This acts as a barrier; all existing requests are rejected, and
 192  * no new requests will be accepted until the import is valid again.
 193  */
 194 void ptlrpc_deactivate_import(struct obd_import *imp)
 195 {
 196         spin_lock(&imp->imp_lock);
 197         ptlrpc_deactivate_and_unlock_import(imp);
 198 }
 199
 200 /*
 201  * This function will invalidate the import, if necessary, then block
 202  * for all the RPC completions, and finally notify the obd to
 203  * invalidate its state (ie cancel locks, clear pending requests,
 204  * etc).
 205  */
 206 void ptlrpc_invalidate_import(struct obd_import *imp)
 207 {
 208         struct l_wait_info lwi;
 209         int rc;
 210
 211         atomic_inc(&imp->imp_inval_count);
 212
 213         /*
 214          * If this is an invalid MGC connection, then don't bother
 215          * waiting for imp_inflight to drop to 0.
 216          */
 217         if (imp->imp_invalid && imp->imp_recon_bk && !imp->imp_obd->obd_no_recov)
 218                 goto out;
 219
 220         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 221                 ptlrpc_deactivate_import(imp);
 222
 223         LASSERT(imp->imp_invalid);
 224
 225         /* wait for all requests to error out and call completion callbacks */
 226         lwi = LWI_TIMEOUT_INTERVAL(cfs_timeout_cap(cfs_time_seconds(obd_timeout)),
 227                                    HZ, NULL, NULL);
 228         rc = l_wait_event(imp->imp_recovery_waitq,
 229                           (atomic_read(&imp->imp_inflight) == 0), &lwi);
 230
 231         if (rc) {
 232                 struct list_head *tmp, *n;
 233                 struct ptlrpc_request *req;
 234
 235                 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 236                        obd2cli_tgt(imp->imp_obd), rc,
 237                        atomic_read(&imp->imp_inflight));
 238                 spin_lock(&imp->imp_lock);
 239                 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 240                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
 241                         DEBUG_REQ(D_ERROR, req, "still on sending list");
 242                 }
 243                 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
 244                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
 245                         DEBUG_REQ(D_ERROR, req, "still on delayed list");
 246                 }
 247                 spin_unlock(&imp->imp_lock);
 248         }
 249
 250 out:
 251         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 252         sptlrpc_import_flush_all_ctx(imp);
 253
 254         atomic_dec(&imp->imp_inval_count);
 255         cfs_waitq_signal(&imp->imp_recovery_waitq);
 256 }
 257
 258 /* unset imp_invalid */
 259 void ptlrpc_activate_import(struct obd_import *imp)
 260 {
 261         struct obd_device *obd = imp->imp_obd;
 262
 263         spin_lock(&imp->imp_lock);
 264         imp->imp_invalid = 0;
 265         ptlrpc_activate_timeouts(imp);
 266         spin_unlock(&imp->imp_lock);
 267         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 268 }
 269
 270 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 271 {
 272         ENTRY;
 273
 274         LASSERT(!imp->imp_dlm_fake);
 275
 276         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 277                 if (!imp->imp_replayable) {
 278                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 279                                "auto-deactivating\n",
 280                                obd2cli_tgt(imp->imp_obd),
 281                                imp->imp_connection->c_remote_uuid.uuid,
 282                                imp->imp_obd->obd_name);
 283                         ptlrpc_deactivate_import(imp);
 284                 }
 285
 286                 CDEBUG(D_HA, "%s: waking up pinger\n",
 287                        obd2cli_tgt(imp->imp_obd));
 288
 289                 spin_lock(&imp->imp_lock);
 290                 imp->imp_force_verify = 1;
 291                 spin_unlock(&imp->imp_lock);
 292
 293                 ptlrpc_pinger_wake_up();
 294         }
 295         EXIT;
 296 }
 297
 298 int ptlrpc_reconnect_import(struct obd_import *imp)
 299 {
 300
 301         ptlrpc_set_import_discon(imp, 0);
 302         /* Force a new connect attempt */
 303         ptlrpc_invalidate_import(imp);
 304         /* Do a fresh connect next time by zeroing the handle */
 305         ptlrpc_disconnect_import(imp, 1);
 306         /* Wait for all invalidate calls to finish */
 307         if (atomic_read(&imp->imp_inval_count) > 0) {
 308                 int rc;
 309                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 310                 rc = l_wait_event(imp->imp_recovery_waitq,
 311                                   (atomic_read(&imp->imp_inval_count) == 0),
 312                                   &lwi);
 313                 if (rc)
 314                         CERROR("Interrupted, inval=%d\n",
 315                                atomic_read(&imp->imp_inval_count));
 316         }
 317
 318         /* Allow reconnect attempts */
 319         imp->imp_obd->obd_no_recov = 0;
 320         /* Remove 'invalid' flag */
 321         ptlrpc_activate_import(imp);
 322         /* Attempt a new connect */
 323         ptlrpc_recover_import(imp, NULL);
 324         return 0;
 325 }
 326
 327 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 328
 329 static int import_select_connection(struct obd_import *imp)
 330 {
 331         struct obd_import_conn *imp_conn = NULL, *conn;
 332         struct obd_export *dlmexp;
 333         ENTRY;
 334
 335         spin_lock(&imp->imp_lock);
 336
 337         if (list_empty(&imp->imp_conn_list)) {
 338                 CERROR("%s: no connections available\n",
 339                         imp->imp_obd->obd_name);
 340                 spin_unlock(&imp->imp_lock);
 341                 RETURN(-EINVAL);
 342         }
 343
 344         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 345                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 346                        imp->imp_obd->obd_name,
 347                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 348                        conn->oic_last_attempt);
 349                 /* Throttle the reconnect rate to once per RECONNECT_INTERVAL */
 350                 if (cfs_time_before_64(conn->oic_last_attempt +
 351                                        RECONNECT_INTERVAL * HZ,
 352                                        cfs_time_current_64())) {
 353                         /* If we have never tried this connection since the
 354                            the last successful attempt, go with this one */
 355                         if (cfs_time_beforeq_64(conn->oic_last_attempt,
 356                                                imp->imp_last_success_conn)) {
 357                                 imp_conn = conn;
 358                                 break;
 359                         }
 360
 361                         /* Both of these connections have already been tried
 362                            since the last successful connection; just choose the
 363                            least recently used */
 364                         if (!imp_conn)
 365                                 imp_conn = conn;
 366                         else if (cfs_time_before_64(conn->oic_last_attempt,
 367                                                     imp_conn->oic_last_attempt))
 368                                 imp_conn = conn;
 369                 }
 370         }
 371
 372         /* if not found, simply choose the current one */
 373         if (!imp_conn) {
 374                 LASSERT(imp->imp_conn_current);
 375                 imp_conn = imp->imp_conn_current;
 376         }
 377         LASSERT(imp_conn->oic_conn);
 378
 379         imp_conn->oic_last_attempt = cfs_time_current_64();
 380
 381         /* switch connection, don't mind if it's same as the current one */
 382         if (imp->imp_connection)
 383                 ptlrpc_put_connection(imp->imp_connection);
 384         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 385
 386         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 387         LASSERT(dlmexp != NULL);
 388         if (dlmexp->exp_connection)
 389                 ptlrpc_put_connection(dlmexp->exp_connection);
 390         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 391         class_export_put(dlmexp);
 392
 393         if (imp->imp_conn_current != imp_conn) {
 394                 if (imp->imp_conn_current)
 395                         LCONSOLE_INFO("Changing connection for %s to %s/%s\n",
 396                                       imp->imp_obd->obd_name,
 397                                       imp_conn->oic_uuid.uuid,
 398                                       libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 399                 imp->imp_conn_current = imp_conn;
 400         }
 401
 402         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 403                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 404                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 405
 406         spin_unlock(&imp->imp_lock);
 407
 408         RETURN(0);
 409 }
 410
 411 /*
 412  * must be called under imp_lock
 413  */
 414 int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 415 {
 416         struct ptlrpc_request *req;
 417         struct list_head *tmp;
 418
 419         if (list_empty(&imp->imp_replay_list))
 420                 return 0;
 421         tmp = imp->imp_replay_list.next;
 422         req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 423         *transno = req->rq_transno;
 424         if (req->rq_transno == 0) {
 425                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 426                 LBUG();
 427         }
 428
 429         return 1;
 430 }
 431
 432 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
 433 {
 434         struct obd_device *obd = imp->imp_obd;
 435         int initial_connect = 0;
 436         int set_transno = 0;
 437         __u64 committed_before_reconnect = 0;
 438         struct ptlrpc_request *request;
 439         char *bufs[] = { NULL,
 440                          obd2cli_tgt(imp->imp_obd),
 441                          obd->obd_uuid.uuid,
 442                          (char *)&imp->imp_dlm_handle,
 443                          (char *)&imp->imp_connect_data };
 444         struct ptlrpc_connect_async_args *aa;
 445         int rc;
 446         ENTRY;
 447
 448         spin_lock(&imp->imp_lock);
 449         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 450                 spin_unlock(&imp->imp_lock);
 451                 CERROR("can't connect to a closed import\n");
 452                 RETURN(-EINVAL);
 453         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 454                 spin_unlock(&imp->imp_lock);
 455                 CERROR("already connected\n");
 456                 RETURN(0);
 457         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 458                 spin_unlock(&imp->imp_lock);
 459                 CERROR("already connecting\n");
 460                 RETURN(-EALREADY);
 461         }
 462
 463         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 464
 465         imp->imp_conn_cnt++;
 466         imp->imp_resend_replay = 0;
 467
 468         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 469                 initial_connect = 1;
 470         else
 471                 committed_before_reconnect = imp->imp_peer_committed_transno;
 472
 473         set_transno = ptlrpc_first_transno(imp, &imp->imp_connect_data.ocd_transno);
 474         spin_unlock(&imp->imp_lock);
 475
 476         if (new_uuid) {
 477                 struct obd_uuid uuid;
 478
 479                 obd_str2uuid(&uuid, new_uuid);
 480                 rc = import_set_conn_priority(imp, &uuid);
 481                 if (rc)
 482                         GOTO(out, rc);
 483         }
 484
 485         rc = import_select_connection(imp);
 486         if (rc)
 487                 GOTO(out, rc);
 488
 489         /* last in connection list */
 490         if (imp->imp_conn_current->oic_item.next == &imp->imp_conn_list) {
 491                 if (imp->imp_initial_recov_bk && initial_connect) {
 492                         CDEBUG(D_HA, "Last connection attempt (%d) for %s\n",
 493                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 494                         /* Don't retry if connect fails */
 495                         rc = 0;
 496                         obd_set_info_async(obd->obd_self_export,
 497                                            strlen(KEY_INIT_RECOV),
 498                                            KEY_INIT_RECOV,
 499                                            sizeof(rc), &rc, NULL);
 500                 }
 501                 if (imp->imp_recon_bk) {
 502                         CDEBUG(D_HA, "Last reconnection attempt (%d) for %s\n",
 503                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 504                         spin_lock(&imp->imp_lock);
 505                         imp->imp_last_recon = 1;
 506                         spin_unlock(&imp->imp_lock);
 507                 }
 508         }
 509
 510         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 511         if (rc)
 512                 GOTO(out, rc);
 513
 514         /* Reset connect flags to the originally requested flags, in case
 515          * the server is updated on-the-fly we will get the new features. */
 516         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 517         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 518                            &obd->obd_uuid, &imp->imp_connect_data);
 519         if (rc)
 520                 GOTO(out, rc);
 521
 522         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 523         if (request == NULL)
 524                 GOTO(out, rc = -ENOMEM);
 525
 526         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 527                                       imp->imp_connect_op, bufs, NULL);
 528         if (rc) {
 529                 ptlrpc_request_free(request);
 530                 GOTO(out, rc);
 531         }
 532
 533 #ifndef __KERNEL__
 534         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 535 #endif
 536         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 537
 538         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 539         /* Allow a slightly larger reply for future growth compatibility */
 540         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 541                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 542         ptlrpc_request_set_replen(request);
 543         request->rq_interpret_reply = ptlrpc_connect_interpret;
 544
 545         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 546         aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
 547         memset(aa, 0, sizeof *aa);
 548
 549         aa->pcaa_peer_committed = committed_before_reconnect;
 550         aa->pcaa_initial_connect = initial_connect;
 551
 552         if (aa->pcaa_initial_connect) {
 553                 spin_lock(&imp->imp_lock);
 554                 imp->imp_replayable = 1;
 555                 spin_unlock(&imp->imp_lock);
 556                 /* On an initial connect, we don't know which one of a
 557                    failover server pair is up.  Don't wait long. */
 558 #ifdef CRAY_XT3
 559                 request->rq_timeout = max((int)(obd_timeout / 2), 5);
 560 #else
 561                 request->rq_timeout = max((int)(obd_timeout / 20), 5);
 562 #endif
 563                 lustre_msg_add_op_flags(request->rq_reqmsg,
 564                                         MSG_CONNECT_INITIAL);
 565         }
 566
 567         if (set_transno)
 568                 lustre_msg_add_op_flags(request->rq_reqmsg,
 569                                         MSG_CONNECT_TRANSNO);
 570
 571         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request");
 572         ptlrpcd_add_req(request);
 573         rc = 0;
 574 out:
 575         if (rc != 0) {
 576                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 577         }
 578
 579         RETURN(rc);
 580 }
 581 EXPORT_SYMBOL(ptlrpc_connect_import);
 582
 583 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 584 {
 585 #ifdef __KERNEL__
 586         struct obd_import_conn *imp_conn;
 587 #endif
 588         int wake_pinger = 0;
 589
 590         ENTRY;
 591
 592         spin_lock(&imp->imp_lock);
 593         if (list_empty(&imp->imp_conn_list))
 594                 GOTO(unlock, 0);
 595
 596 #ifdef __KERNEL__
 597         imp_conn = list_entry(imp->imp_conn_list.prev,
 598                               struct obd_import_conn,
 599                               oic_item);
 600
 601         /* XXX: When the failover node is the primary node, it is possible
 602          * to have two identical connections in imp_conn_list. We must
 603          * compare not conn's pointers but NIDs, otherwise we can defeat
 604          * connection throttling. (See bug 14774.) */
 605         if (imp->imp_conn_current->oic_conn->c_self !=
 606                                 imp_conn->oic_conn->c_self) {
 607                 ptlrpc_ping_import_soon(imp);
 608                 wake_pinger = 1;
 609         }
 610 #else
 611         /* liblustre has no pinger thead, so we wakup pinger anyway */
 612         wake_pinger = 1;
 613 #endif
 614
 615  unlock:
 616         spin_unlock(&imp->imp_lock);
 617
 618         if (wake_pinger)
 619                 ptlrpc_pinger_wake_up();
 620
 621         EXIT;
 622 }
 623
 624 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
 625                                     void * data, int rc)
 626 {
 627         struct ptlrpc_connect_async_args *aa = data;
 628         struct obd_import *imp = request->rq_import;
 629         struct client_obd *cli = &imp->imp_obd->u.cli;
 630         struct lustre_handle old_hdl;
 631         int msg_flags;
 632         ENTRY;
 633
 634         spin_lock(&imp->imp_lock);
 635         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 636                 spin_unlock(&imp->imp_lock);
 637                 RETURN(0);
 638         }
 639         spin_unlock(&imp->imp_lock);
 640
 641         if (rc)
 642                 GOTO(out, rc);
 643
 644         LASSERT(imp->imp_conn_current);
 645
 646         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 647
 648         /* All imports are pingable */
 649         spin_lock(&imp->imp_lock);
 650         imp->imp_pingable = 1;
 651
 652         if (aa->pcaa_initial_connect) {
 653                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 654                         imp->imp_replayable = 1;
 655                         spin_unlock(&imp->imp_lock);
 656                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 657                                obd2cli_tgt(imp->imp_obd));
 658                 } else {
 659                         imp->imp_replayable = 0;
 660                         spin_unlock(&imp->imp_lock);
 661                 }
 662
 663                 /* if applies, adjust the imp->imp_msg_magic here
 664                  * according to reply flags */
 665
 666                 imp->imp_remote_handle =
 667                                 *lustre_msg_get_handle(request->rq_repmsg);
 668
 669                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 670                 spin_lock(&imp->imp_lock);
 671                 if (imp->imp_invalid) {
 672                         spin_unlock(&imp->imp_lock);
 673                         ptlrpc_activate_import(imp);
 674                 } else {
 675                         spin_unlock(&imp->imp_lock);
 676                 }
 677
 678                 GOTO(finish, rc = 0);
 679         } else {
 680                 spin_unlock(&imp->imp_lock);
 681         }
 682
 683         /* Determine what recovery state to move the import to. */
 684         if (MSG_CONNECT_RECONNECT & msg_flags) {
 685                 memset(&old_hdl, 0, sizeof(old_hdl));
 686                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 687                             sizeof (old_hdl))) {
 688                         CERROR("%s@%s didn't like our handle "LPX64
 689                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 690                                imp->imp_connection->c_remote_uuid.uuid,
 691                                imp->imp_dlm_handle.cookie);
 692                         GOTO(out, rc = -ENOTCONN);
 693                 }
 694
 695                 if (memcmp(&imp->imp_remote_handle,
 696                            lustre_msg_get_handle(request->rq_repmsg),
 697                            sizeof(imp->imp_remote_handle))) {
 698
 699                         CWARN("%s@%s changed server handle from "
 700                                LPX64" to "LPX64" - evicting.\n",
 701                                obd2cli_tgt(imp->imp_obd),
 702                                imp->imp_connection->c_remote_uuid.uuid,
 703                                imp->imp_remote_handle.cookie,
 704                                lustre_msg_get_handle(request->rq_repmsg)->
 705                                          cookie);
 706                         imp->imp_remote_handle =
 707                                      *lustre_msg_get_handle(request->rq_repmsg);
 708
 709                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 710                         GOTO(finish, rc = 0);
 711                 } else {
 712                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 713                                obd2cli_tgt(imp->imp_obd),
 714                                imp->imp_connection->c_remote_uuid.uuid);
 715                 }
 716
 717                 if (imp->imp_invalid) {
 718                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 719                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 720                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 721                                imp->imp_obd->obd_name,
 722                                obd2cli_tgt(imp->imp_obd));
 723
 724                         spin_lock(&imp->imp_lock);
 725                         imp->imp_resend_replay = 1;
 726                         spin_unlock(&imp->imp_lock);
 727
 728                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 729                 } else {
 730                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 731                 }
 732         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 733                 LASSERT(imp->imp_replayable);
 734                 imp->imp_remote_handle =
 735                                 *lustre_msg_get_handle(request->rq_repmsg);
 736                 imp->imp_last_replay_transno = 0;
 737                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 738         } else {
 739                 imp->imp_remote_handle =
 740                                 *lustre_msg_get_handle(request->rq_repmsg);
 741                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 742         }
 743
 744         /* Sanity checks for a reconnected import. */
 745         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 746                 CERROR("imp_replayable flag does not match server "
 747                        "after reconnect. We should LBUG right here.\n");
 748         }
 749
 750         if (lustre_msg_get_last_committed(request->rq_repmsg) <
 751             aa->pcaa_peer_committed) {
 752                 CERROR("%s went back in time (transno "LPD64
 753                        " was previously committed, server now claims "LPD64
 754                        ")!  See https://bugzilla.clusterfs.com/"
 755                        "long_list.cgi?buglist=9646\n",
 756                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 757                        lustre_msg_get_last_committed(request->rq_repmsg));
 758         }
 759
 760 finish:
 761         rc = ptlrpc_import_recovery_state_machine(imp);
 762         if (rc != 0) {
 763                 if (rc == -ENOTCONN) {
 764                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 765                                "invalidating and reconnecting\n",
 766                                obd2cli_tgt(imp->imp_obd),
 767                                imp->imp_connection->c_remote_uuid.uuid);
 768                         ptlrpc_connect_import(imp, NULL);
 769                         RETURN(0);
 770                 }
 771         } else {
 772                 struct obd_connect_data *ocd;
 773                 struct obd_export *exp;
 774                 int ret;
 775                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 776                                            RCL_SERVER);
 777                 /* server replied obd_connect_data is always bigger */
 778                 ocd = req_capsule_server_sized_get(&request->rq_pill,
 779                                                    &RMF_CONNECT_DATA, ret);
 780
 781                 spin_lock(&imp->imp_lock);
 782                 list_del(&imp->imp_conn_current->oic_item);
 783                 list_add(&imp->imp_conn_current->oic_item, &imp->imp_conn_list);
 784                 imp->imp_last_success_conn =
 785                         imp->imp_conn_current->oic_last_attempt;
 786
 787                 if (ocd == NULL) {
 788                         spin_unlock(&imp->imp_lock);
 789                         CERROR("Wrong connect data from server\n");
 790                         rc = -EPROTO;
 791                         GOTO(out, rc);
 792                 }
 793
 794                 imp->imp_connect_data = *ocd;
 795
 796                 exp = class_conn2export(&imp->imp_dlm_handle);
 797                 spin_unlock(&imp->imp_lock);
 798
 799                 /* check that server granted subset of flags we asked for. */
 800                 LASSERTF((ocd->ocd_connect_flags &
 801                           imp->imp_connect_flags_orig) ==
 802                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 803                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 804
 805                 if (!exp) {
 806                         /* This could happen if export is cleaned during the
 807                            connect attempt */
 808                         CERROR("Missing export for %s\n",
 809                                imp->imp_obd->obd_name);
 810                         GOTO(out, rc = -ENODEV);
 811                 }
 812                 exp->exp_connect_flags = ocd->ocd_connect_flags;
 813                 imp->imp_obd->obd_self_export->exp_connect_flags =
 814                                                         ocd->ocd_connect_flags;
 815                 class_export_put(exp);
 816
 817                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 818
 819                 if (!ocd->ocd_ibits_known &&
 820                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 821                         CERROR("Inodebits aware server returned zero compatible"
 822                                " bits?\n");
 823
 824                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 825                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
 826                                         LUSTRE_VERSION_OFFSET_WARN ||
 827                      ocd->ocd_version < LUSTRE_VERSION_CODE -
 828                                         LUSTRE_VERSION_OFFSET_WARN)) {
 829                         /* Sigh, some compilers do not like #ifdef in the middle
 830                            of macro arguments */
 831 #ifdef __KERNEL__
 832                         const char *older =
 833                                 "older. Consider upgrading this client";
 834 #else
 835                         const char *older =
 836                                 "older. Consider recompiling this application";
 837 #endif
 838                         const char *newer = "newer than client version";
 839
 840                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 841                                       "is much %s (%s)\n",
 842                                       obd2cli_tgt(imp->imp_obd),
 843                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 844                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 845                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 846                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
 847                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
 848                                       newer : older, LUSTRE_VERSION_STRING);
 849                 }
 850
 851                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
 852                         /* We sent to the server ocd_cksum_types with bits set
 853                          * for algorithms we understand. The server masked off
 854                          * the checksum types it doesn't support */
 855                         if ((ocd->ocd_cksum_types & OBD_CKSUM_ALL) == 0) {
 856                                 LCONSOLE_WARN("The negotiation of the checksum "
 857                                               "alogrithm to use with server %s "
 858                                               "failed (%x/%x), disabling "
 859                                               "checksums\n",
 860                                               obd2cli_tgt(imp->imp_obd),
 861                                               ocd->ocd_cksum_types,
 862                                               OBD_CKSUM_ALL);
 863                                 cli->cl_checksum = 0;
 864                                 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 865                                 cli->cl_cksum_type = OBD_CKSUM_CRC32;
 866                         } else {
 867                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
 868
 869                                 if (ocd->ocd_cksum_types & OSC_DEFAULT_CKSUM)
 870                                         cli->cl_cksum_type = OSC_DEFAULT_CKSUM;
 871                                 else if (ocd->ocd_cksum_types & OBD_CKSUM_ADLER)
 872                                         cli->cl_cksum_type = OBD_CKSUM_ADLER;
 873                                 else
 874                                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
 875                         }
 876                 } else {
 877                         /* The server does not support OBD_CONNECT_CKSUM.
 878                          * Enforce CRC32 for backward compatibility*/
 879                         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 880                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
 881                 }
 882
 883                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
 884                         cli->cl_max_pages_per_rpc =
 885                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
 886                 }
 887
 888                 imp->imp_obd->obd_namespace->ns_connect_flags =
 889                                                         ocd->ocd_connect_flags;
 890                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
 891                                                         ocd->ocd_connect_flags;
 892
 893                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 894                         (cli->cl_max_pages_per_rpc > 0));
 895         }
 896
 897 out:
 898         if (rc != 0) {
 899                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 900                 spin_lock(&imp->imp_lock);
 901                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov &&
 902                     (request->rq_import_generation == imp->imp_generation))
 903                         ptlrpc_deactivate_and_unlock_import(imp);
 904                 else
 905                         spin_unlock(&imp->imp_lock);
 906
 907                 if ((imp->imp_recon_bk && imp->imp_last_recon) ||
 908                     (rc == -EACCES)) {
 909                         /*
 910                          * Give up trying to reconnect
 911                          * EACCES means client has no permission for connection
 912                          */
 913                         imp->imp_obd->obd_no_recov = 1;
 914                         ptlrpc_deactivate_import(imp);
 915                 }
 916
 917                 if (rc == -EPROTO) {
 918                         struct obd_connect_data *ocd;
 919
 920                         /* reply message might not be ready */
 921                         if (request->rq_repmsg == NULL)
 922                                 RETURN(-EPROTO);
 923
 924                         ocd = req_capsule_server_get(&request->rq_pill,
 925                                                      &RMF_CONNECT_DATA);
 926                         if (ocd &&
 927                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 928                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
 929                            /* Actually servers are only supposed to refuse
 930                               connection from liblustre clients, so we should
 931                               never see this from VFS context */
 932                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
 933                                         "(%d.%d.%d.%d)"
 934                                         " refused connection from this client "
 935                                         "with an incompatible version (%s).  "
 936                                         "Client must be recompiled\n",
 937                                         obd2cli_tgt(imp->imp_obd),
 938                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 939                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 940                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 941                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
 942                                         LUSTRE_VERSION_STRING);
 943                                 ptlrpc_deactivate_import(imp);
 944                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
 945                         }
 946                         RETURN(-EPROTO);
 947                 }
 948
 949                 ptlrpc_maybe_ping_import_soon(imp);
 950
 951                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
 952                        obd2cli_tgt(imp->imp_obd),
 953                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
 954         }
 955
 956         spin_lock(&imp->imp_lock);
 957         imp->imp_last_recon = 0;
 958         spin_unlock(&imp->imp_lock);
 959
 960         cfs_waitq_signal(&imp->imp_recovery_waitq);
 961         RETURN(rc);
 962 }
 963
 964 static int completed_replay_interpret(struct ptlrpc_request *req,
 965                                     void * data, int rc)
 966 {
 967         ENTRY;
 968         atomic_dec(&req->rq_import->imp_replay_inflight);
 969         if (req->rq_status == 0) {
 970                 ptlrpc_import_recovery_state_machine(req->rq_import);
 971         } else {
 972                 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
 973                        "reconnecting\n",
 974                        req->rq_import->imp_obd->obd_name, req->rq_status);
 975                 ptlrpc_connect_import(req->rq_import, NULL);
 976         }
 977
 978         RETURN(0);
 979 }
 980
 981 static int signal_completed_replay(struct obd_import *imp)
 982 {
 983         struct ptlrpc_request *req;
 984         ENTRY;
 985
 986         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 987         atomic_inc(&imp->imp_replay_inflight);
 988
 989         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
 990                                         OBD_PING);
 991         if (req == NULL) {
 992                 atomic_dec(&imp->imp_replay_inflight);
 993                 RETURN(-ENOMEM);
 994         }
 995
 996         ptlrpc_request_set_replen(req);
 997         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
 998         lustre_msg_add_flags(req->rq_reqmsg,
 999                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1000         req->rq_timeout *= 3;
1001         req->rq_interpret_reply = completed_replay_interpret;
1002
1003         ptlrpcd_add_req(req);
1004         RETURN(0);
1005 }
1006
1007 #ifdef __KERNEL__
1008 static int ptlrpc_invalidate_import_thread(void *data)
1009 {
1010         struct obd_import *imp = data;
1011
1012         ENTRY;
1013
1014         ptlrpc_daemonize("ll_imp_inval");
1015
1016         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1017                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1018                imp->imp_connection->c_remote_uuid.uuid);
1019
1020         ptlrpc_invalidate_import(imp);
1021
1022         if (obd_dump_on_eviction) {
1023                 CERROR("dump the log upon eviction\n");
1024                 libcfs_debug_dumplog();
1025         }
1026
1027         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1028         ptlrpc_import_recovery_state_machine(imp);
1029
1030         RETURN(0);
1031 }
1032 #endif
1033
1034 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1035 {
1036         int rc = 0;
1037         int inflight;
1038         char *target_start;
1039         int target_len;
1040
1041         ENTRY;
1042         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1043                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1044                           &target_start, &target_len);
1045                 /* Don't care about MGC eviction */
1046                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1047                            LUSTRE_MGC_NAME) != 0) {
1048                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1049                                            "%.*s; in progress operations using "
1050                                            "this service will fail.\n",
1051                                            target_len, target_start);
1052                 }
1053                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1054                        obd2cli_tgt(imp->imp_obd),
1055                        imp->imp_connection->c_remote_uuid.uuid);
1056
1057 #ifdef __KERNEL__
1058                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
1059                                        CLONE_VM | CLONE_FILES);
1060                 if (rc < 0)
1061                         CERROR("error starting invalidate thread: %d\n", rc);
1062                 else
1063                         rc = 0;
1064                 RETURN(rc);
1065 #else
1066                 ptlrpc_invalidate_import(imp);
1067
1068                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1069 #endif
1070         }
1071
1072         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1073                 CDEBUG(D_HA, "replay requested by %s\n",
1074                        obd2cli_tgt(imp->imp_obd));
1075                 rc = ptlrpc_replay_next(imp, &inflight);
1076                 if (inflight == 0 &&
1077                     atomic_read(&imp->imp_replay_inflight) == 0) {
1078                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1079                         rc = ldlm_replay_locks(imp);
1080                         if (rc)
1081                                 GOTO(out, rc);
1082                 }
1083                 rc = 0;
1084         }
1085
1086         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1087                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1088                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1089                         rc = signal_completed_replay(imp);
1090                         if (rc)
1091                                 GOTO(out, rc);
1092                 }
1093
1094         }
1095
1096         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1097                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1098                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1099                 }
1100         }
1101
1102         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1103                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1104                        obd2cli_tgt(imp->imp_obd),
1105                        imp->imp_connection->c_remote_uuid.uuid);
1106
1107                 rc = ptlrpc_resend(imp);
1108                 if (rc)
1109                         GOTO(out, rc);
1110                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1111                 ptlrpc_activate_import(imp);
1112
1113                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1114                           &target_start, &target_len);
1115                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1116                               "using nid %s.\n", imp->imp_obd->obd_name,
1117                               target_len, target_start,
1118                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1119         }
1120
1121         if (imp->imp_state == LUSTRE_IMP_FULL) {
1122                 cfs_waitq_signal(&imp->imp_recovery_waitq);
1123                 ptlrpc_wake_delayed(imp);
1124         }
1125
1126 out:
1127         RETURN(rc);
1128 }
1129
1130 static int back_to_sleep(void *unused)
1131 {
1132         return 0;
1133 }
1134
1135 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1136 {
1137         struct ptlrpc_request *req;
1138         int rq_opc, rc = 0;
1139         int nowait = imp->imp_obd->obd_force;
1140         ENTRY;
1141
1142         if (nowait)
1143                 GOTO(set_state, rc);
1144
1145         switch (imp->imp_connect_op) {
1146         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1147         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1148         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1149         default:
1150                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1151                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1152                 RETURN(-EINVAL);
1153         }
1154
1155         if (ptlrpc_import_in_recovery(imp)) {
1156                 struct l_wait_info lwi;
1157                 cfs_duration_t timeout;
1158                 if (imp->imp_server_timeout)
1159                         timeout = cfs_time_seconds(obd_timeout / 2);
1160                 else
1161                         timeout = cfs_time_seconds(obd_timeout);
1162
1163                 timeout = MAX(timeout * HZ, 1);
1164
1165                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1166                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1167                 rc = l_wait_event(imp->imp_recovery_waitq,
1168                                   !ptlrpc_import_in_recovery(imp), &lwi);
1169
1170         }
1171
1172         spin_lock(&imp->imp_lock);
1173         if (imp->imp_state != LUSTRE_IMP_FULL)
1174                 GOTO(out, 0);
1175
1176         spin_unlock(&imp->imp_lock);
1177
1178         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1179                                         LUSTRE_OBD_VERSION, rq_opc);
1180         if (req) {
1181                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1182                  * it fails.  We can get through the above with a down server
1183                  * if the client doesn't know the server is gone yet. */
1184                 req->rq_no_resend = 1;
1185 #ifdef CRAY_XT3
1186                 req->rq_timeout = obd_timeout / 3;
1187 #else
1188                 req->rq_timeout = 5;
1189 #endif
1190                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1191                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1192                 ptlrpc_request_set_replen(req);
1193                 rc = ptlrpc_queue_wait(req);
1194                 ptlrpc_req_finished(req);
1195         }
1196
1197 set_state:
1198         spin_lock(&imp->imp_lock);
1199 out:
1200         if (noclose)
1201                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1202         else
1203                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1204         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1205         imp->imp_conn_cnt = 0;
1206         imp->imp_last_recon = 0;
1207         spin_unlock(&imp->imp_lock);
1208
1209         RETURN(rc);
1210 }
1211