lustre/ptlrpc/import.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/import.c
  37  *
  38  * Author: Mike Shaver <shaver@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_RPC
  42 #ifndef __KERNEL__
  43 # include <liblustre.h>
  44 #endif
  45
  46 #include <obd_support.h>
  47 #include <lustre_ha.h>
  48 #include <lustre_net.h>
  49 #include <lustre_import.h>
  50 #include <lustre_export.h>
  51 #include <obd.h>
  52 #include <obd_cksum.h>
  53 #include <obd_class.h>
  54
  55 #include "ptlrpc_internal.h"
  56
  57 struct ptlrpc_connect_async_args {
  58          __u64 pcaa_peer_committed;
  59         int pcaa_initial_connect;
  60 };
  61
  62 /* A CLOSED import should remain so. */
  63 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  64 do {                                                                           \
  65         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  66                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  67                       imp, obd2cli_tgt(imp->imp_obd),                          \
  68                       ptlrpc_import_state_name(imp->imp_state),                \
  69                       ptlrpc_import_state_name(state));                        \
  70                imp->imp_state = state;                                         \
  71         }                                                                      \
  72 } while(0)
  73
  74 #define IMPORT_SET_STATE(imp, state)            \
  75 do {                                            \
  76         spin_lock(&imp->imp_lock);              \
  77         IMPORT_SET_STATE_NOLOCK(imp, state);    \
  78         spin_unlock(&imp->imp_lock);            \
  79 } while(0)
  80
  81
  82 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
  83                                     void * data, int rc);
  84 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
  85
  86 /* Only this function is allowed to change the import state when it is
  87  * CLOSED. I would rather refcount the import and free it after
  88  * disconnection like we do with exports. To do that, the client_obd
  89  * will need to save the peer info somewhere other than in the import,
  90  * though. */
  91 int ptlrpc_init_import(struct obd_import *imp)
  92 {
  93         spin_lock(&imp->imp_lock);
  94
  95         imp->imp_generation++;
  96         imp->imp_state =  LUSTRE_IMP_NEW;
  97
  98         spin_unlock(&imp->imp_lock);
  99
 100         return 0;
 101 }
 102 EXPORT_SYMBOL(ptlrpc_init_import);
 103
 104 #define UUID_STR "_UUID"
 105 static void deuuidify(char *uuid, const char *prefix, char **uuid_start,
 106                       int *uuid_len)
 107 {
 108         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 109                 ? uuid : uuid + strlen(prefix);
 110
 111         *uuid_len = strlen(*uuid_start);
 112
 113         if (*uuid_len < strlen(UUID_STR))
 114                 return;
 115
 116         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 117                     UUID_STR, strlen(UUID_STR)))
 118                 *uuid_len -= strlen(UUID_STR);
 119 }
 120
 121 /* Returns true if import was FULL, false if import was already not
 122  * connected.
 123  * @imp - import to be disconnected
 124  * @conn_cnt - connection count (epoch) of the request that timed out
 125  *             and caused the disconnection.  In some cases, multiple
 126  *             inflight requests can fail to a single target (e.g. OST
 127  *             bulk requests) and if one has already caused a reconnection
 128  *             (increasing the import->conn_cnt) the older failure should
 129  *             not also cause a reconnection.  If zero it forces a reconnect.
 130  */
 131 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 132 {
 133         int rc = 0;
 134
 135         spin_lock(&imp->imp_lock);
 136
 137         if (imp->imp_state == LUSTRE_IMP_FULL &&
 138             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 139                 char *target_start;
 140                 int   target_len;
 141
 142                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 143                           &target_start, &target_len);
 144
 145                 if (imp->imp_replayable) {
 146                         LCONSOLE_WARN("%s: Connection to service %.*s via nid "
 147                                "%s was lost; in progress operations using this "
 148                                "service will wait for recovery to complete.\n",
 149                                imp->imp_obd->obd_name, target_len, target_start,
 150                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 151                 } else {
 152                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to service "
 153                                            "%.*s via nid %s was lost; in progress"
 154                                            "operations using this service will"
 155                                            "fail.\n",
 156                                            imp->imp_obd->obd_name,
 157                                            target_len, target_start,
 158                                  libcfs_nid2str(imp->imp_connection->c_peer.nid));
 159                 }
 160                 ptlrpc_deactivate_timeouts(imp);
 161                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 162                 spin_unlock(&imp->imp_lock);
 163
 164                 if (obd_dump_on_timeout)
 165                         libcfs_debug_dumplog();
 166
 167                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 168                 rc = 1;
 169         } else {
 170                 spin_unlock(&imp->imp_lock);
 171                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 172                        imp->imp_client->cli_name, imp,
 173                        (imp->imp_state == LUSTRE_IMP_FULL &&
 174                         imp->imp_conn_cnt > conn_cnt) ?
 175                        "reconnected" : "not connected", imp->imp_conn_cnt,
 176                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 177         }
 178
 179         return rc;
 180 }
 181
 182 /* Must be called with imp_lock held! */
 183 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 184 {
 185         ENTRY;
 186         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 187
 188         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 189         imp->imp_invalid = 1;
 190         imp->imp_generation++;
 191         spin_unlock(&imp->imp_lock);
 192
 193         ptlrpc_abort_inflight(imp);
 194         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 195
 196         EXIT;
 197 }
 198
 199 /*
 200  * This acts as a barrier; all existing requests are rejected, and
 201  * no new requests will be accepted until the import is valid again.
 202  */
 203 void ptlrpc_deactivate_import(struct obd_import *imp)
 204 {
 205         spin_lock(&imp->imp_lock);
 206         ptlrpc_deactivate_and_unlock_import(imp);
 207 }
 208
 209 /*
 210  * This function will invalidate the import, if necessary, then block
 211  * for all the RPC completions, and finally notify the obd to
 212  * invalidate its state (ie cancel locks, clear pending requests,
 213  * etc).
 214  */
 215 void ptlrpc_invalidate_import(struct obd_import *imp)
 216 {
 217         struct list_head *tmp, *n;
 218         struct ptlrpc_request *req;
 219         struct l_wait_info lwi;
 220         int rc;
 221
 222         atomic_inc(&imp->imp_inval_count);
 223
 224         /*
 225          * If this is an invalid MGC connection, then don't bother
 226          * waiting for imp_inflight to drop to 0.
 227          */
 228         if (imp->imp_invalid && imp->imp_recon_bk && !imp->imp_obd->obd_no_recov)
 229                 goto out;
 230
 231         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 232                 ptlrpc_deactivate_import(imp);
 233
 234         LASSERT(imp->imp_invalid);
 235
 236         /* wait for all requests to error out and call completion callbacks.
 237            Cap it at obd_timeout -- these should all have been locally
 238            cancelled by ptlrpc_abort_inflight. */
 239         lwi = LWI_TIMEOUT_INTERVAL(
 240                 cfs_timeout_cap(cfs_time_seconds(obd_timeout)),
 241                 cfs_time_seconds(1), NULL, NULL);
 242         rc = l_wait_event(imp->imp_recovery_waitq,
 243                           (atomic_read(&imp->imp_inflight) == 0), &lwi);
 244
 245         if (rc) {
 246                 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 247                          obd2cli_tgt(imp->imp_obd), rc,
 248                          atomic_read(&imp->imp_inflight));
 249                 spin_lock(&imp->imp_lock);
 250                 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 251                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
 252                         DEBUG_REQ(D_ERROR, req, "still on sending list");
 253                 }
 254                 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
 255                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
 256                         DEBUG_REQ(D_ERROR, req, "still on delayed list");
 257                 }
 258                 spin_unlock(&imp->imp_lock);
 259                 LASSERT(atomic_read(&imp->imp_inflight) == 0);
 260         }
 261
 262 out:
 263         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 264         sptlrpc_import_flush_all_ctx(imp);
 265
 266         atomic_dec(&imp->imp_inval_count);
 267         cfs_waitq_signal(&imp->imp_recovery_waitq);
 268 }
 269
 270 /* unset imp_invalid */
 271 void ptlrpc_activate_import(struct obd_import *imp)
 272 {
 273         struct obd_device *obd = imp->imp_obd;
 274
 275         spin_lock(&imp->imp_lock);
 276         imp->imp_invalid = 0;
 277         ptlrpc_activate_timeouts(imp);
 278         spin_unlock(&imp->imp_lock);
 279         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 280 }
 281
 282 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 283 {
 284         ENTRY;
 285
 286         LASSERT(!imp->imp_dlm_fake);
 287
 288         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 289                 if (!imp->imp_replayable) {
 290                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 291                                "auto-deactivating\n",
 292                                obd2cli_tgt(imp->imp_obd),
 293                                imp->imp_connection->c_remote_uuid.uuid,
 294                                imp->imp_obd->obd_name);
 295                         ptlrpc_deactivate_import(imp);
 296                 }
 297
 298                 CDEBUG(D_HA, "%s: waking up pinger\n",
 299                        obd2cli_tgt(imp->imp_obd));
 300
 301                 spin_lock(&imp->imp_lock);
 302                 imp->imp_force_verify = 1;
 303                 spin_unlock(&imp->imp_lock);
 304
 305                 ptlrpc_pinger_wake_up();
 306         }
 307         EXIT;
 308 }
 309
 310 int ptlrpc_reconnect_import(struct obd_import *imp)
 311 {
 312
 313         ptlrpc_set_import_discon(imp, 0);
 314         /* Force a new connect attempt */
 315         ptlrpc_invalidate_import(imp);
 316         /* Do a fresh connect next time by zeroing the handle */
 317         ptlrpc_disconnect_import(imp, 1);
 318         /* Wait for all invalidate calls to finish */
 319         if (atomic_read(&imp->imp_inval_count) > 0) {
 320                 int rc;
 321                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 322                 rc = l_wait_event(imp->imp_recovery_waitq,
 323                                   (atomic_read(&imp->imp_inval_count) == 0),
 324                                   &lwi);
 325                 if (rc)
 326                         CERROR("Interrupted, inval=%d\n",
 327                                atomic_read(&imp->imp_inval_count));
 328         }
 329
 330         /* Allow reconnect attempts */
 331         imp->imp_obd->obd_no_recov = 0;
 332         /* Remove 'invalid' flag */
 333         ptlrpc_activate_import(imp);
 334         /* Attempt a new connect */
 335         ptlrpc_recover_import(imp, NULL);
 336         return 0;
 337 }
 338
 339 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 340
 341 static int import_select_connection(struct obd_import *imp)
 342 {
 343         struct obd_import_conn *imp_conn = NULL, *conn;
 344         struct obd_export *dlmexp;
 345         int tried_all = 1;
 346         ENTRY;
 347
 348         spin_lock(&imp->imp_lock);
 349
 350         if (list_empty(&imp->imp_conn_list)) {
 351                 CERROR("%s: no connections available\n",
 352                         imp->imp_obd->obd_name);
 353                 spin_unlock(&imp->imp_lock);
 354                 RETURN(-EINVAL);
 355         }
 356
 357         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 358                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 359                        imp->imp_obd->obd_name,
 360                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 361                        conn->oic_last_attempt);
 362                 /* Don't thrash connections */
 363                 if (cfs_time_before_64(cfs_time_current_64(),
 364                                      conn->oic_last_attempt +
 365                                      cfs_time_seconds(CONNECTION_SWITCH_MIN))) {
 366                         continue;
 367                 }
 368
 369                 /* If we have not tried this connection since the
 370                    the last successful attempt, go with this one */
 371                 if ((conn->oic_last_attempt == 0) ||
 372                     cfs_time_beforeq_64(conn->oic_last_attempt,
 373                                        imp->imp_last_success_conn)) {
 374                         imp_conn = conn;
 375                         tried_all = 0;
 376                         break;
 377                 }
 378
 379                 /* If all of the connections have already been tried
 380                    since the last successful connection; just choose the
 381                    least recently used */
 382                 if (!imp_conn)
 383                         imp_conn = conn;
 384                 else if (cfs_time_before_64(conn->oic_last_attempt,
 385                                             imp_conn->oic_last_attempt))
 386                         imp_conn = conn;
 387         }
 388
 389         /* if not found, simply choose the current one */
 390         if (!imp_conn) {
 391                 LASSERT(imp->imp_conn_current);
 392                 imp_conn = imp->imp_conn_current;
 393                 tried_all = 0;
 394         }
 395         LASSERT(imp_conn->oic_conn);
 396
 397         /* If we've tried everything, and we're back to the beginning of the
 398            list, increase our timeout and try again. It will be reset when
 399            we do finally connect. (FIXME: really we should wait for all network
 400            state associated with the last connection attempt to drain before
 401            trying to reconnect on it.) */
 402         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item) &&
 403             !imp->imp_recon_bk /* not retrying */) {
 404                 if (at_get(&imp->imp_at.iat_net_latency) <
 405                     CONNECTION_SWITCH_MAX) {
 406                         at_add(&imp->imp_at.iat_net_latency,
 407                                at_get(&imp->imp_at.iat_net_latency) +
 408                                CONNECTION_SWITCH_INC);
 409                 }
 410                 LASSERT(imp_conn->oic_last_attempt);
 411                 CWARN("%s: tried all connections, increasing latency to %ds\n",
 412                       imp->imp_obd->obd_name,
 413                       at_get(&imp->imp_at.iat_net_latency));
 414         }
 415
 416         imp_conn->oic_last_attempt = cfs_time_current_64();
 417
 418         /* switch connection, don't mind if it's same as the current one */
 419         if (imp->imp_connection)
 420                 ptlrpc_put_connection(imp->imp_connection);
 421         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 422
 423         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 424         LASSERT(dlmexp != NULL);
 425         if (dlmexp->exp_connection)
 426                 ptlrpc_put_connection(dlmexp->exp_connection);
 427         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 428         class_export_put(dlmexp);
 429
 430         if (imp->imp_conn_current != imp_conn) {
 431                 if (imp->imp_conn_current)
 432                         LCONSOLE_INFO("Changing connection for %s to %s/%s\n",
 433                                       imp->imp_obd->obd_name,
 434                                       imp_conn->oic_uuid.uuid,
 435                                       libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 436                 imp->imp_conn_current = imp_conn;
 437         }
 438
 439         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 440                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 441                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 442
 443         spin_unlock(&imp->imp_lock);
 444
 445         RETURN(0);
 446 }
 447
 448 /*
 449  * must be called under imp_lock
 450  */
 451 int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 452 {
 453         struct ptlrpc_request *req;
 454         struct list_head *tmp;
 455
 456         if (list_empty(&imp->imp_replay_list))
 457                 return 0;
 458         tmp = imp->imp_replay_list.next;
 459         req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 460         *transno = req->rq_transno;
 461         if (req->rq_transno == 0) {
 462                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 463                 LBUG();
 464         }
 465
 466         return 1;
 467 }
 468
 469 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
 470 {
 471         struct obd_device *obd = imp->imp_obd;
 472         int initial_connect = 0;
 473         int set_transno = 0;
 474         __u64 committed_before_reconnect = 0;
 475         struct ptlrpc_request *request;
 476         char *bufs[] = { NULL,
 477                          obd2cli_tgt(imp->imp_obd),
 478                          obd->obd_uuid.uuid,
 479                          (char *)&imp->imp_dlm_handle,
 480                          (char *)&imp->imp_connect_data };
 481         struct ptlrpc_connect_async_args *aa;
 482         int rc;
 483         ENTRY;
 484
 485         spin_lock(&imp->imp_lock);
 486         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 487                 spin_unlock(&imp->imp_lock);
 488                 CERROR("can't connect to a closed import\n");
 489                 RETURN(-EINVAL);
 490         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 491                 spin_unlock(&imp->imp_lock);
 492                 CERROR("already connected\n");
 493                 RETURN(0);
 494         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 495                 spin_unlock(&imp->imp_lock);
 496                 CERROR("already connecting\n");
 497                 RETURN(-EALREADY);
 498         }
 499
 500         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 501
 502         imp->imp_conn_cnt++;
 503         imp->imp_resend_replay = 0;
 504
 505         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 506                 initial_connect = 1;
 507         else
 508                 committed_before_reconnect = imp->imp_peer_committed_transno;
 509
 510         set_transno = ptlrpc_first_transno(imp, &imp->imp_connect_data.ocd_transno);
 511         spin_unlock(&imp->imp_lock);
 512
 513         if (new_uuid) {
 514                 struct obd_uuid uuid;
 515
 516                 obd_str2uuid(&uuid, new_uuid);
 517                 rc = import_set_conn_priority(imp, &uuid);
 518                 if (rc)
 519                         GOTO(out, rc);
 520         }
 521
 522         rc = import_select_connection(imp);
 523         if (rc)
 524                 GOTO(out, rc);
 525
 526         /* last in connection list */
 527         if (imp->imp_conn_current->oic_item.next == &imp->imp_conn_list) {
 528                 if (imp->imp_initial_recov_bk && initial_connect) {
 529                         CDEBUG(D_HA, "Last connection attempt (%d) for %s\n",
 530                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 531                         /* Don't retry if connect fails */
 532                         rc = 0;
 533                         obd_set_info_async(obd->obd_self_export,
 534                                            sizeof(KEY_INIT_RECOV),
 535                                            KEY_INIT_RECOV,
 536                                            sizeof(rc), &rc, NULL);
 537                 }
 538                 if (imp->imp_recon_bk) {
 539                         CDEBUG(D_HA, "Last reconnection attempt (%d) for %s\n",
 540                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 541                         spin_lock(&imp->imp_lock);
 542                         imp->imp_last_recon = 1;
 543                         spin_unlock(&imp->imp_lock);
 544                 }
 545         }
 546
 547         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 548         if (rc)
 549                 GOTO(out, rc);
 550
 551         /* Reset connect flags to the originally requested flags, in case
 552          * the server is updated on-the-fly we will get the new features. */
 553         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 554         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 555
 556         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 557                            &obd->obd_uuid, &imp->imp_connect_data);
 558         if (rc)
 559                 GOTO(out, rc);
 560
 561         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 562         if (request == NULL)
 563                 GOTO(out, rc = -ENOMEM);
 564
 565         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 566                                       imp->imp_connect_op, bufs, NULL);
 567         if (rc) {
 568                 ptlrpc_request_free(request);
 569                 GOTO(out, rc);
 570         }
 571
 572 #ifndef __KERNEL__
 573         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 574 #endif
 575         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 576
 577         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 578         /* Allow a slightly larger reply for future growth compatibility */
 579         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 580                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 581         ptlrpc_request_set_replen(request);
 582         request->rq_interpret_reply = ptlrpc_connect_interpret;
 583
 584         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 585         aa = ptlrpc_req_async_args(request);
 586         memset(aa, 0, sizeof *aa);
 587
 588         aa->pcaa_peer_committed = committed_before_reconnect;
 589         aa->pcaa_initial_connect = initial_connect;
 590
 591         if (aa->pcaa_initial_connect) {
 592                 spin_lock(&imp->imp_lock);
 593                 imp->imp_replayable = 1;
 594                 spin_unlock(&imp->imp_lock);
 595                 lustre_msg_add_op_flags(request->rq_reqmsg,
 596                                         MSG_CONNECT_INITIAL);
 597                 if (AT_OFF)
 598                         /* AT will use INITIAL_CONNECT_TIMEOUT the first
 599                            time, adaptive after that. */
 600                         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 601         }
 602
 603         if (set_transno)
 604                 lustre_msg_add_op_flags(request->rq_reqmsg,
 605                                         MSG_CONNECT_TRANSNO);
 606
 607         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request");
 608         ptlrpcd_add_req(request);
 609         rc = 0;
 610 out:
 611         if (rc != 0) {
 612                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 613         }
 614
 615         RETURN(rc);
 616 }
 617 EXPORT_SYMBOL(ptlrpc_connect_import);
 618
 619 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 620 {
 621 #ifdef __KERNEL__
 622         struct obd_import_conn *imp_conn;
 623 #endif
 624         int wake_pinger = 0;
 625
 626         ENTRY;
 627
 628         spin_lock(&imp->imp_lock);
 629         if (list_empty(&imp->imp_conn_list))
 630                 GOTO(unlock, 0);
 631
 632 #ifdef __KERNEL__
 633         imp_conn = list_entry(imp->imp_conn_list.prev,
 634                               struct obd_import_conn,
 635                               oic_item);
 636
 637         /* XXX: When the failover node is the primary node, it is possible
 638          * to have two identical connections in imp_conn_list. We must
 639          * compare not conn's pointers but NIDs, otherwise we can defeat
 640          * connection throttling. (See bug 14774.) */
 641         if (imp->imp_conn_current->oic_conn->c_self !=
 642                                 imp_conn->oic_conn->c_self) {
 643                 ptlrpc_ping_import_soon(imp);
 644                 wake_pinger = 1;
 645         }
 646 #else
 647         /* liblustre has no pinger thead, so we wakup pinger anyway */
 648         wake_pinger = 1;
 649 #endif
 650
 651  unlock:
 652         spin_unlock(&imp->imp_lock);
 653
 654         if (wake_pinger)
 655                 ptlrpc_pinger_wake_up();
 656
 657         EXIT;
 658 }
 659
 660 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
 661                                     void * data, int rc)
 662 {
 663         struct ptlrpc_connect_async_args *aa = data;
 664         struct obd_import *imp = request->rq_import;
 665         struct client_obd *cli = &imp->imp_obd->u.cli;
 666         struct lustre_handle old_hdl;
 667         int msg_flags;
 668         ENTRY;
 669
 670         spin_lock(&imp->imp_lock);
 671         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 672                 spin_unlock(&imp->imp_lock);
 673                 RETURN(0);
 674         }
 675         spin_unlock(&imp->imp_lock);
 676
 677         if (rc)
 678                 GOTO(out, rc);
 679
 680         LASSERT(imp->imp_conn_current);
 681
 682         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 683
 684         /* All imports are pingable */
 685         spin_lock(&imp->imp_lock);
 686         imp->imp_pingable = 1;
 687
 688         if (aa->pcaa_initial_connect) {
 689                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 690                         imp->imp_replayable = 1;
 691                         spin_unlock(&imp->imp_lock);
 692                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 693                                obd2cli_tgt(imp->imp_obd));
 694                 } else {
 695                         imp->imp_replayable = 0;
 696                         spin_unlock(&imp->imp_lock);
 697                 }
 698
 699                 /* if applies, adjust the imp->imp_msg_magic here
 700                  * according to reply flags */
 701
 702                 imp->imp_remote_handle =
 703                                 *lustre_msg_get_handle(request->rq_repmsg);
 704
 705                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 706                 spin_lock(&imp->imp_lock);
 707                 if (imp->imp_invalid) {
 708                         spin_unlock(&imp->imp_lock);
 709                         ptlrpc_activate_import(imp);
 710                 } else {
 711                         spin_unlock(&imp->imp_lock);
 712                 }
 713
 714                 GOTO(finish, rc = 0);
 715         } else {
 716                 spin_unlock(&imp->imp_lock);
 717         }
 718
 719         /* Determine what recovery state to move the import to. */
 720         if (MSG_CONNECT_RECONNECT & msg_flags) {
 721                 memset(&old_hdl, 0, sizeof(old_hdl));
 722                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 723                             sizeof (old_hdl))) {
 724                         CERROR("%s@%s didn't like our handle "LPX64
 725                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 726                                imp->imp_connection->c_remote_uuid.uuid,
 727                                imp->imp_dlm_handle.cookie);
 728                         GOTO(out, rc = -ENOTCONN);
 729                 }
 730
 731                 if (memcmp(&imp->imp_remote_handle,
 732                            lustre_msg_get_handle(request->rq_repmsg),
 733                            sizeof(imp->imp_remote_handle))) {
 734                         int level = msg_flags & MSG_CONNECT_RECOVERING ? D_HA :
 735                                                                          D_WARNING;
 736
 737                         /* Bug 16611/14775: if server handle have changed,
 738                          * that means some sort of disconnection happened.
 739                          * If the server is not in recovery, that also means it
 740                          * already erased all of our state because of previous
 741                          * eviction. If it is in recovery - we are safe to
 742                          * participate since we can reestablish all of our state
 743                          * with server again */
 744                         CDEBUG(level,"%s@%s changed server handle from "
 745                                      LPX64" to "LPX64"%s \n" "but is still in recovery \n",
 746                                      obd2cli_tgt(imp->imp_obd),
 747                                      imp->imp_connection->c_remote_uuid.uuid,
 748                                      imp->imp_remote_handle.cookie,
 749                                      lustre_msg_get_handle(request->rq_repmsg)->
 750                                                                         cookie,
 751                                      (MSG_CONNECT_RECOVERING & msg_flags) ?
 752                                          "but is still in recovery" : "");
 753
 754                         imp->imp_remote_handle =
 755                                      *lustre_msg_get_handle(request->rq_repmsg);
 756
 757                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
 758                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 759                                 GOTO(finish, rc = 0);
 760                         }
 761
 762                 } else {
 763                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 764                                obd2cli_tgt(imp->imp_obd),
 765                                imp->imp_connection->c_remote_uuid.uuid);
 766                 }
 767
 768                 if (imp->imp_invalid) {
 769                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
 770                                "marking evicted\n", imp->imp_obd->obd_name);
 771                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 772                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 773                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 774                                imp->imp_obd->obd_name,
 775                                obd2cli_tgt(imp->imp_obd));
 776
 777                         spin_lock(&imp->imp_lock);
 778                         imp->imp_resend_replay = 1;
 779                         spin_unlock(&imp->imp_lock);
 780
 781                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 782                 } else {
 783                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 784                 }
 785         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 786                 LASSERT(imp->imp_replayable);
 787                 imp->imp_remote_handle =
 788                                 *lustre_msg_get_handle(request->rq_repmsg);
 789                 imp->imp_last_replay_transno = 0;
 790                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 791         } else {
 792                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
 793                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
 794                 imp->imp_remote_handle =
 795                                 *lustre_msg_get_handle(request->rq_repmsg);
 796                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 797         }
 798
 799         /* Sanity checks for a reconnected import. */
 800         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 801                 CERROR("imp_replayable flag does not match server "
 802                        "after reconnect. We should LBUG right here.\n");
 803         }
 804
 805         if (lustre_msg_get_last_committed(request->rq_repmsg) <
 806             aa->pcaa_peer_committed) {
 807                 CERROR("%s went back in time (transno "LPD64
 808                        " was previously committed, server now claims "LPD64
 809                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
 810                        "id=9646\n",
 811                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 812                        lustre_msg_get_last_committed(request->rq_repmsg));
 813         }
 814
 815 finish:
 816         rc = ptlrpc_import_recovery_state_machine(imp);
 817         if (rc != 0) {
 818                 if (rc == -ENOTCONN) {
 819                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 820                                "invalidating and reconnecting\n",
 821                                obd2cli_tgt(imp->imp_obd),
 822                                imp->imp_connection->c_remote_uuid.uuid);
 823                         ptlrpc_connect_import(imp, NULL);
 824                         RETURN(0);
 825                 }
 826         } else {
 827                 struct obd_connect_data *ocd;
 828                 struct obd_export *exp;
 829                 int ret;
 830                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 831                                            RCL_SERVER);
 832                 /* server replied obd_connect_data is always bigger */
 833                 ocd = req_capsule_server_sized_get(&request->rq_pill,
 834                                                    &RMF_CONNECT_DATA, ret);
 835
 836                 spin_lock(&imp->imp_lock);
 837                 list_del(&imp->imp_conn_current->oic_item);
 838                 list_add(&imp->imp_conn_current->oic_item, &imp->imp_conn_list);
 839                 imp->imp_last_success_conn =
 840                         imp->imp_conn_current->oic_last_attempt;
 841
 842                 if (ocd == NULL) {
 843                         spin_unlock(&imp->imp_lock);
 844                         CERROR("Wrong connect data from server\n");
 845                         rc = -EPROTO;
 846                         GOTO(out, rc);
 847                 }
 848
 849                 imp->imp_connect_data = *ocd;
 850
 851                 exp = class_conn2export(&imp->imp_dlm_handle);
 852                 spin_unlock(&imp->imp_lock);
 853
 854                 /* check that server granted subset of flags we asked for. */
 855                 LASSERTF((ocd->ocd_connect_flags &
 856                           imp->imp_connect_flags_orig) ==
 857                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 858                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 859
 860                 if (!exp) {
 861                         /* This could happen if export is cleaned during the
 862                            connect attempt */
 863                         CERROR("Missing export for %s\n",
 864                                imp->imp_obd->obd_name);
 865                         GOTO(out, rc = -ENODEV);
 866                 }
 867                 exp->exp_connect_flags = ocd->ocd_connect_flags;
 868                 imp->imp_obd->obd_self_export->exp_connect_flags =
 869                                                         ocd->ocd_connect_flags;
 870                 class_export_put(exp);
 871
 872                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 873
 874                 if (!ocd->ocd_ibits_known &&
 875                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 876                         CERROR("Inodebits aware server returned zero compatible"
 877                                " bits?\n");
 878
 879                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 880                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
 881                                         LUSTRE_VERSION_OFFSET_WARN ||
 882                      ocd->ocd_version < LUSTRE_VERSION_CODE -
 883                                         LUSTRE_VERSION_OFFSET_WARN)) {
 884                         /* Sigh, some compilers do not like #ifdef in the middle
 885                            of macro arguments */
 886 #ifdef __KERNEL__
 887                         const char *older =
 888                                 "older. Consider upgrading this client";
 889 #else
 890                         const char *older =
 891                                 "older. Consider recompiling this application";
 892 #endif
 893                         const char *newer = "newer than client version";
 894
 895                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 896                                       "is much %s (%s)\n",
 897                                       obd2cli_tgt(imp->imp_obd),
 898                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 899                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 900                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 901                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
 902                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
 903                                       newer : older, LUSTRE_VERSION_STRING);
 904                 }
 905
 906                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
 907                         /* We sent to the server ocd_cksum_types with bits set
 908                          * for algorithms we understand. The server masked off
 909                          * the checksum types it doesn't support */
 910                         if ((ocd->ocd_cksum_types & OBD_CKSUM_ALL) == 0) {
 911                                 LCONSOLE_WARN("The negotiation of the checksum "
 912                                               "alogrithm to use with server %s "
 913                                               "failed (%x/%x), disabling "
 914                                               "checksums\n",
 915                                               obd2cli_tgt(imp->imp_obd),
 916                                               ocd->ocd_cksum_types,
 917                                               OBD_CKSUM_ALL);
 918                                 cli->cl_checksum = 0;
 919                                 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 920                                 cli->cl_cksum_type = OBD_CKSUM_CRC32;
 921                         } else {
 922                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
 923
 924                                 if (ocd->ocd_cksum_types & OSC_DEFAULT_CKSUM)
 925                                         cli->cl_cksum_type = OSC_DEFAULT_CKSUM;
 926                                 else if (ocd->ocd_cksum_types & OBD_CKSUM_ADLER)
 927                                         cli->cl_cksum_type = OBD_CKSUM_ADLER;
 928                                 else
 929                                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
 930                         }
 931                 } else {
 932                         /* The server does not support OBD_CONNECT_CKSUM.
 933                          * Enforce CRC32 for backward compatibility*/
 934                         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 935                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
 936                 }
 937
 938                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
 939                         cli->cl_max_pages_per_rpc =
 940                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
 941                 }
 942
 943                 imp->imp_obd->obd_namespace->ns_connect_flags =
 944                                                         ocd->ocd_connect_flags;
 945                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
 946                                                         ocd->ocd_connect_flags;
 947
 948                 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
 949                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
 950                         /* We need a per-message support flag, because
 951                            a. we don't know if the incoming connect reply
 952                               supports AT or not (in reply_in_callback)
 953                               until we unpack it.
 954                            b. failovered server means export and flags are gone
 955                               (in ptlrpc_send_reply).
 956                            Can only be set when we know AT is supported at
 957                            both ends */
 958                         imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 959                 else
 960                         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 961
 962                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 963                         (cli->cl_max_pages_per_rpc > 0));
 964         }
 965
 966 out:
 967         if (rc != 0) {
 968                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 969                 spin_lock(&imp->imp_lock);
 970                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov &&
 971                     (request->rq_import_generation == imp->imp_generation))
 972                         ptlrpc_deactivate_and_unlock_import(imp);
 973                 else
 974                         spin_unlock(&imp->imp_lock);
 975
 976                 if ((imp->imp_recon_bk && imp->imp_last_recon) ||
 977                     (rc == -EACCES)) {
 978                         /*
 979                          * Give up trying to reconnect
 980                          * EACCES means client has no permission for connection
 981                          */
 982                         imp->imp_obd->obd_no_recov = 1;
 983                         ptlrpc_deactivate_import(imp);
 984                 }
 985
 986                 if (rc == -EPROTO) {
 987                         struct obd_connect_data *ocd;
 988
 989                         /* reply message might not be ready */
 990                         if (request->rq_repmsg == NULL)
 991                                 RETURN(-EPROTO);
 992
 993                         ocd = req_capsule_server_get(&request->rq_pill,
 994                                                      &RMF_CONNECT_DATA);
 995                         if (ocd &&
 996                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 997                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
 998                            /* Actually servers are only supposed to refuse
 999                               connection from liblustre clients, so we should
1000                               never see this from VFS context */
1001                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1002                                         "(%d.%d.%d.%d)"
1003                                         " refused connection from this client "
1004                                         "with an incompatible version (%s).  "
1005                                         "Client must be recompiled\n",
1006                                         obd2cli_tgt(imp->imp_obd),
1007                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1008                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1009                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1010                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1011                                         LUSTRE_VERSION_STRING);
1012                                 ptlrpc_deactivate_import(imp);
1013                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1014                         }
1015                         RETURN(-EPROTO);
1016                 }
1017
1018                 ptlrpc_maybe_ping_import_soon(imp);
1019
1020                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1021                        obd2cli_tgt(imp->imp_obd),
1022                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1023         }
1024
1025         spin_lock(&imp->imp_lock);
1026         imp->imp_last_recon = 0;
1027         spin_unlock(&imp->imp_lock);
1028
1029         cfs_waitq_signal(&imp->imp_recovery_waitq);
1030         RETURN(rc);
1031 }
1032
1033 static int completed_replay_interpret(struct ptlrpc_request *req,
1034                                     void * data, int rc)
1035 {
1036         ENTRY;
1037         atomic_dec(&req->rq_import->imp_replay_inflight);
1038         if (req->rq_status == 0) {
1039                 ptlrpc_import_recovery_state_machine(req->rq_import);
1040         } else {
1041                 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1042                        "reconnecting\n",
1043                        req->rq_import->imp_obd->obd_name, req->rq_status);
1044                 ptlrpc_connect_import(req->rq_import, NULL);
1045         }
1046
1047         RETURN(0);
1048 }
1049
1050 static int signal_completed_replay(struct obd_import *imp)
1051 {
1052         struct ptlrpc_request *req;
1053         ENTRY;
1054
1055         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1056         atomic_inc(&imp->imp_replay_inflight);
1057
1058         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1059                                         OBD_PING);
1060         if (req == NULL) {
1061                 atomic_dec(&imp->imp_replay_inflight);
1062                 RETURN(-ENOMEM);
1063         }
1064
1065         ptlrpc_request_set_replen(req);
1066         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1067         lustre_msg_add_flags(req->rq_reqmsg,
1068                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1069         req->rq_timeout *= 3;
1070         req->rq_interpret_reply = completed_replay_interpret;
1071
1072         ptlrpcd_add_req(req);
1073         RETURN(0);
1074 }
1075
1076 #ifdef __KERNEL__
1077 static int ptlrpc_invalidate_import_thread(void *data)
1078 {
1079         struct obd_import *imp = data;
1080
1081         ENTRY;
1082
1083         ptlrpc_daemonize("ll_imp_inval");
1084
1085         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1086                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1087                imp->imp_connection->c_remote_uuid.uuid);
1088
1089         ptlrpc_invalidate_import(imp);
1090
1091         if (obd_dump_on_eviction) {
1092                 CERROR("dump the log upon eviction\n");
1093                 libcfs_debug_dumplog();
1094         }
1095
1096         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1097         ptlrpc_import_recovery_state_machine(imp);
1098
1099         RETURN(0);
1100 }
1101 #endif
1102
1103 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1104 {
1105         int rc = 0;
1106         int inflight;
1107         char *target_start;
1108         int target_len;
1109
1110         ENTRY;
1111         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1112                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1113                           &target_start, &target_len);
1114                 /* Don't care about MGC eviction */
1115                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1116                            LUSTRE_MGC_NAME) != 0) {
1117                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1118                                            "%.*s; in progress operations using "
1119                                            "this service will fail.\n",
1120                                            target_len, target_start);
1121                 }
1122                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1123                        obd2cli_tgt(imp->imp_obd),
1124                        imp->imp_connection->c_remote_uuid.uuid);
1125
1126 #ifdef __KERNEL__
1127                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
1128                                        CLONE_VM | CLONE_FILES);
1129                 if (rc < 0)
1130                         CERROR("error starting invalidate thread: %d\n", rc);
1131                 else
1132                         rc = 0;
1133                 RETURN(rc);
1134 #else
1135                 ptlrpc_invalidate_import(imp);
1136
1137                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1138 #endif
1139         }
1140
1141         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1142                 CDEBUG(D_HA, "replay requested by %s\n",
1143                        obd2cli_tgt(imp->imp_obd));
1144                 rc = ptlrpc_replay_next(imp, &inflight);
1145                 if (inflight == 0 &&
1146                     atomic_read(&imp->imp_replay_inflight) == 0) {
1147                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1148                         rc = ldlm_replay_locks(imp);
1149                         if (rc)
1150                                 GOTO(out, rc);
1151                 }
1152                 rc = 0;
1153         }
1154
1155         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1156                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1157                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1158                         rc = signal_completed_replay(imp);
1159                         if (rc)
1160                                 GOTO(out, rc);
1161                 }
1162
1163         }
1164
1165         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1166                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1167                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1168                 }
1169         }
1170
1171         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1172                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1173                        obd2cli_tgt(imp->imp_obd),
1174                        imp->imp_connection->c_remote_uuid.uuid);
1175
1176                 rc = ptlrpc_resend(imp);
1177                 if (rc)
1178                         GOTO(out, rc);
1179                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1180                 ptlrpc_activate_import(imp);
1181
1182                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1183                           &target_start, &target_len);
1184                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1185                               "using nid %s.\n", imp->imp_obd->obd_name,
1186                               target_len, target_start,
1187                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1188         }
1189
1190         if (imp->imp_state == LUSTRE_IMP_FULL) {
1191                 cfs_waitq_signal(&imp->imp_recovery_waitq);
1192                 ptlrpc_wake_delayed(imp);
1193         }
1194
1195 out:
1196         RETURN(rc);
1197 }
1198
1199 static int back_to_sleep(void *unused)
1200 {
1201         return 0;
1202 }
1203
1204 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1205 {
1206         struct ptlrpc_request *req;
1207         int rq_opc, rc = 0;
1208         int nowait = imp->imp_obd->obd_force;
1209         ENTRY;
1210
1211         if (nowait)
1212                 GOTO(set_state, rc);
1213
1214         switch (imp->imp_connect_op) {
1215         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1216         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1217         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1218         default:
1219                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1220                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1221                 RETURN(-EINVAL);
1222         }
1223
1224         if (ptlrpc_import_in_recovery(imp)) {
1225                 struct l_wait_info lwi;
1226                 cfs_duration_t timeout;
1227
1228
1229                 if (AT_OFF) {
1230                         if (imp->imp_server_timeout)
1231                                 timeout = cfs_time_seconds(obd_timeout / 2);
1232                         else
1233                                 timeout = cfs_time_seconds(obd_timeout);
1234                 } else {
1235                         int idx = import_at_get_index(imp,
1236                                 imp->imp_client->cli_request_portal);
1237                         timeout = cfs_time_seconds(
1238                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1239                 }
1240
1241                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1242                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1243                 rc = l_wait_event(imp->imp_recovery_waitq,
1244                                   !ptlrpc_import_in_recovery(imp), &lwi);
1245
1246         }
1247
1248         spin_lock(&imp->imp_lock);
1249         if (imp->imp_state != LUSTRE_IMP_FULL)
1250                 GOTO(out, 0);
1251
1252         spin_unlock(&imp->imp_lock);
1253
1254         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1255                                         LUSTRE_OBD_VERSION, rq_opc);
1256         if (req) {
1257                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1258                  * it fails.  We can get through the above with a down server
1259                  * if the client doesn't know the server is gone yet. */
1260                 req->rq_no_resend = 1;
1261
1262 #ifndef CRAY_XT3
1263                 /* We want client umounts to happen quickly, no matter the
1264                    server state... */
1265                 req->rq_timeout = min_t(int, req->rq_timeout,
1266                                         INITIAL_CONNECT_TIMEOUT);
1267 #else
1268                 /* ... but we always want liblustre clients to nicely
1269                    disconnect, so only use the adaptive value. */
1270                 if (AT_OFF)
1271                         req->rq_timeout = obd_timeout / 3;
1272 #endif
1273
1274                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1275                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1276                 ptlrpc_request_set_replen(req);
1277                 rc = ptlrpc_queue_wait(req);
1278                 ptlrpc_req_finished(req);
1279         }
1280
1281 set_state:
1282         spin_lock(&imp->imp_lock);
1283 out:
1284         if (noclose)
1285                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1286         else
1287                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1288         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1289         imp->imp_conn_cnt = 0;
1290         /* Try all connections in the future - bz 12758 */
1291         imp->imp_last_recon = 0;
1292         spin_unlock(&imp->imp_lock);
1293
1294         RETURN(rc);
1295 }
1296
1297
1298 /* Adaptive Timeout utils */
1299 extern unsigned int at_min, at_max, at_history;
1300
1301 /* Bin into timeslices using AT_BINS bins.
1302    This gives us a max of the last binlimit*AT_BINS secs without the storage,
1303    but still smoothing out a return to normalcy from a slow response.
1304    (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1305 int at_add(struct adaptive_timeout *at, unsigned int val)
1306 {
1307         unsigned int old = at->at_current;
1308         time_t now = cfs_time_current_sec();
1309         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1310
1311         LASSERT(at);
1312 #if 0
1313         CDEBUG(D_INFO, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1314                val, at, now - at->at_binstart, at->at_current,
1315                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1316 #endif
1317         if (val == 0)
1318                 /* 0's don't count, because we never want our timeout to
1319                    drop to 0, and because 0 could mean an error */
1320                 return 0;
1321
1322         spin_lock(&at->at_lock);
1323
1324         if (unlikely(at->at_binstart == 0)) {
1325                 /* Special case to remove default from history */
1326                 at->at_current = val;
1327                 at->at_worst_ever = val;
1328                 at->at_worst_time = now;
1329                 at->at_hist[0] = val;
1330                 at->at_binstart = now;
1331         } else if (now - at->at_binstart < binlimit ) {
1332                 /* in bin 0 */
1333                 at->at_hist[0] = max(val, at->at_hist[0]);
1334                 at->at_current = max(val, at->at_current);
1335         } else {
1336                 int i, shift;
1337                 unsigned int maxv = val;
1338                 /* move bins over */
1339                 shift = (now - at->at_binstart) / binlimit;
1340                 LASSERT(shift > 0);
1341                 for(i = AT_BINS - 1; i >= 0; i--) {
1342                         if (i >= shift) {
1343                                 at->at_hist[i] = at->at_hist[i - shift];
1344                                 maxv = max(maxv, at->at_hist[i]);
1345                         } else {
1346                                 at->at_hist[i] = 0;
1347                         }
1348                 }
1349                 at->at_hist[0] = val;
1350                 at->at_current = maxv;
1351                 at->at_binstart += shift * binlimit;
1352         }
1353
1354         if (at->at_current > at->at_worst_ever) {
1355                 at->at_worst_ever = at->at_current;
1356                 at->at_worst_time = now;
1357         }
1358
1359         if (at->at_flags & AT_FLG_NOHIST)
1360                 /* Only keep last reported val; keeping the rest of the history
1361                    for proc only */
1362                 at->at_current = val;
1363
1364         if (at_max > 0)
1365                 at->at_current =  min(at->at_current, at_max);
1366         at->at_current =  max(at->at_current, at_min);
1367
1368 #if 0
1369         if (at->at_current != old)
1370                 CDEBUG(D_ADAPTTO, "AT %p change: old=%u new=%u delta=%d "
1371                        "(val=%u) hist %u %u %u %u\n", at,
1372                        old, at->at_current, at->at_current - old, val,
1373                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1374                        at->at_hist[3]);
1375 #endif
1376
1377         /* if we changed, report the old value */
1378         old = (at->at_current != old) ? old : 0;
1379
1380         spin_unlock(&at->at_lock);
1381         return old;
1382 }
1383
1384 /* Find the imp_at index for a given portal; assign if space available */
1385 int import_at_get_index(struct obd_import *imp, int portal)
1386 {
1387         struct imp_at *at = &imp->imp_at;
1388         int i;
1389
1390         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1391                 if (at->iat_portal[i] == portal)
1392                         return i;
1393                 if (at->iat_portal[i] == 0)
1394                         /* unused */
1395                         break;
1396         }
1397
1398         /* Not found in list, add it under a lock */
1399         spin_lock(&imp->imp_lock);
1400
1401         /* Check unused under lock */
1402         for (; i < IMP_AT_MAX_PORTALS; i++) {
1403                 if (at->iat_portal[i] == portal)
1404                         goto out;
1405                 if (at->iat_portal[i] == 0)
1406                         /* unused */
1407                         break;
1408         }
1409
1410         /* Not enough portals? */
1411         LASSERT(i < IMP_AT_MAX_PORTALS);
1412
1413         at->iat_portal[i] = portal;
1414 out:
1415         spin_unlock(&imp->imp_lock);
1416         return i;
1417 }