lustre/ptlrpc/import.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lustre/ptlrpc/import.c
  32  *
  33  * Author: Mike Shaver <shaver@clusterfs.com>
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_RPC
  37
  38 #include <linux/fs_struct.h>
  39 #include <linux/kthread.h>
  40 #include <linux/delay.h>
  41 #include <obd_support.h>
  42 #include <lustre_ha.h>
  43 #include <lustre_net.h>
  44 #include <lustre_import.h>
  45 #include <lustre_export.h>
  46 #include <obd.h>
  47 #include <obd_cksum.h>
  48 #include <obd_class.h>
  49
  50 #include "ptlrpc_internal.h"
  51
  52 struct ptlrpc_connect_async_args {
  53          __u64 pcaa_peer_committed;
  54         int pcaa_initial_connect;
  55 };
  56
  57 /**
  58  * Updates import \a imp current state to provided \a state value
  59  * Helper function.
  60  */
  61 static void import_set_state_nolock(struct obd_import *imp,
  62                                     enum lustre_imp_state state)
  63 {
  64         switch (state) {
  65         case LUSTRE_IMP_CLOSED:
  66         case LUSTRE_IMP_NEW:
  67         case LUSTRE_IMP_DISCON:
  68         case LUSTRE_IMP_CONNECTING:
  69                 break;
  70         case LUSTRE_IMP_REPLAY_WAIT:
  71                 imp->imp_replay_state = LUSTRE_IMP_REPLAY_LOCKS;
  72                 break;
  73         default:
  74                 imp->imp_replay_state = LUSTRE_IMP_REPLAY;
  75                 break;
  76         }
  77
  78         /* A CLOSED import should remain so. */
  79         if (imp->imp_state == LUSTRE_IMP_CLOSED)
  80                 return;
  81
  82         if (imp->imp_state != LUSTRE_IMP_NEW) {
  83                 CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",
  84                        imp, obd2cli_tgt(imp->imp_obd),
  85                        ptlrpc_import_state_name(imp->imp_state),
  86                        ptlrpc_import_state_name(state));
  87         }
  88
  89         imp->imp_state = state;
  90         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
  91         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
  92                 ktime_get_real_seconds();
  93         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
  94                 IMP_STATE_HIST_LEN;
  95 }
  96
  97 static void import_set_state(struct obd_import *imp,
  98                              enum lustre_imp_state new_state)
  99 {
 100         spin_lock(&imp->imp_lock);
 101         import_set_state_nolock(imp, new_state);
 102         spin_unlock(&imp->imp_lock);
 103 }
 104
 105 void ptlrpc_import_enter_resend(struct obd_import *imp)
 106 {
 107         import_set_state(imp, LUSTRE_IMP_RECOVER);
 108 }
 109 EXPORT_SYMBOL(ptlrpc_import_enter_resend);
 110
 111
 112 static int ptlrpc_connect_interpret(const struct lu_env *env,
 113                                     struct ptlrpc_request *request,
 114                                     void *args, int rc);
 115 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
 116
 117 /* Only this function is allowed to change the import state when it is
 118  * CLOSED. I would rather refcount the import and free it after
 119  * disconnection like we do with exports. To do that, the client_obd
 120  * will need to save the peer info somewhere other than in the import,
 121  * though.
 122  */
 123 int ptlrpc_init_import(struct obd_import *imp)
 124 {
 125         spin_lock(&imp->imp_lock);
 126
 127         imp->imp_generation++;
 128         imp->imp_state =  LUSTRE_IMP_NEW;
 129
 130         spin_unlock(&imp->imp_lock);
 131
 132         return 0;
 133 }
 134 EXPORT_SYMBOL(ptlrpc_init_import);
 135
 136 #define UUID_STR "_UUID"
 137 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
 138 {
 139         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 140                 ? uuid : uuid + strlen(prefix);
 141
 142         *uuid_len = strlen(*uuid_start);
 143
 144         if (*uuid_len < strlen(UUID_STR))
 145                 return;
 146
 147         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 148                     UUID_STR, strlen(UUID_STR)))
 149                 *uuid_len -= strlen(UUID_STR);
 150 }
 151 EXPORT_SYMBOL(deuuidify);
 152
 153 /* Must be called with imp_lock held! */
 154 static void ptlrpc_deactivate_import_nolock(struct obd_import *imp)
 155 {
 156         ENTRY;
 157
 158         assert_spin_locked(&imp->imp_lock);
 159         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 160         imp->imp_invalid = 1;
 161         imp->imp_generation++;
 162
 163         ptlrpc_abort_inflight(imp);
 164
 165         EXIT;
 166 }
 167
 168 /**
 169  * Returns true if import was FULL, false if import was already not
 170  * connected.
 171  * @imp - import to be disconnected
 172  * @conn_cnt - connection count (epoch) of the request that timed out
 173  *             and caused the disconnection.  In some cases, multiple
 174  *             inflight requests can fail to a single target (e.g. OST
 175  *             bulk requests) and if one has already caused a reconnection
 176  *             (increasing the import->conn_cnt) the older failure should
 177  *             not also cause a reconnection.  If zero it forces a reconnect.
 178  * @invalid - set import invalid flag
 179  */
 180 int ptlrpc_set_import_discon(struct obd_import *imp,
 181                              __u32 conn_cnt, bool invalid)
 182 {
 183         int rc = 0;
 184
 185         spin_lock(&imp->imp_lock);
 186
 187         if (imp->imp_state == LUSTRE_IMP_FULL &&
 188             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 189                 char *target_start;
 190                 int   target_len;
 191                 bool  inact = false;
 192
 193                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 194                           &target_start, &target_len);
 195
 196                 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
 197                 if (imp->imp_replayable) {
 198                         LCONSOLE_WARN("%s: Connection to %.*s (at %s) was lost; in progress operations using this service will wait for recovery to complete\n",
 199                                imp->imp_obd->obd_name, target_len, target_start,
 200                                obd_import_nid2str(imp));
 201                 } else {
 202                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to %.*s (at %s) was lost; in progress operations using this service will fail\n",
 203                                imp->imp_obd->obd_name, target_len, target_start,
 204                                obd_import_nid2str(imp));
 205                         if (invalid) {
 206                                 CDEBUG(D_HA,
 207                                        "import %s@%s for %s not replayable, auto-deactivating\n",
 208                                        obd2cli_tgt(imp->imp_obd),
 209                                        imp->imp_connection->c_remote_uuid.uuid,
 210                                        imp->imp_obd->obd_name);
 211                                 ptlrpc_deactivate_import_nolock(imp);
 212                                 inact = true;
 213                         }
 214                 }
 215                 spin_unlock(&imp->imp_lock);
 216
 217                 if (obd_dump_on_timeout)
 218                         libcfs_debug_dumplog();
 219
 220                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 221
 222                 if (inact)
 223                         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 224                 rc = 1;
 225         } else {
 226                 spin_unlock(&imp->imp_lock);
 227                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 228                        imp->imp_client->cli_name, imp,
 229                        (imp->imp_state == LUSTRE_IMP_FULL &&
 230                         imp->imp_conn_cnt > conn_cnt) ?
 231                        "reconnected" : "not connected", imp->imp_conn_cnt,
 232                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 233         }
 234
 235         return rc;
 236 }
 237
 238 /*
 239  * This acts as a barrier; all existing requests are rejected, and
 240  * no new requests will be accepted until the import is valid again.
 241  */
 242 void ptlrpc_deactivate_import(struct obd_import *imp)
 243 {
 244         spin_lock(&imp->imp_lock);
 245         ptlrpc_deactivate_import_nolock(imp);
 246         spin_unlock(&imp->imp_lock);
 247
 248         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 249 }
 250 EXPORT_SYMBOL(ptlrpc_deactivate_import);
 251
 252 static time64_t ptlrpc_inflight_deadline(struct ptlrpc_request *req,
 253                                          time64_t now)
 254 {
 255         time64_t dl;
 256
 257         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
 258               (req->rq_phase == RQ_PHASE_BULK) ||
 259               (req->rq_phase == RQ_PHASE_NEW)))
 260                 return 0;
 261
 262         if (req->rq_timedout)
 263                 return 0;
 264
 265         if (req->rq_phase == RQ_PHASE_NEW)
 266                 dl = req->rq_sent;
 267         else
 268                 dl = req->rq_deadline;
 269
 270         if (dl <= now)
 271                 return 0;
 272
 273         return dl - now;
 274 }
 275
 276 static time64_t ptlrpc_inflight_timeout(struct obd_import *imp)
 277 {
 278         time64_t now = ktime_get_real_seconds();
 279         struct ptlrpc_request *req;
 280         time64_t timeout = 0;
 281
 282         spin_lock(&imp->imp_lock);
 283         list_for_each_entry(req, &imp->imp_sending_list, rq_list)
 284                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 285         spin_unlock(&imp->imp_lock);
 286         return timeout;
 287 }
 288
 289 /**
 290  * This function will invalidate the import, if necessary, then block
 291  * for all the RPC completions, and finally notify the obd to
 292  * invalidate its state (ie cancel locks, clear pending requests,
 293  * etc).
 294  */
 295 void ptlrpc_invalidate_import(struct obd_import *imp)
 296 {
 297         struct ptlrpc_request *req;
 298         time64_t timeout;
 299         int rc;
 300
 301         atomic_inc(&imp->imp_inval_count);
 302
 303         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 304                 ptlrpc_deactivate_import(imp);
 305
 306         if (CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CONNECT_RACE)) {
 307                 CFS_RACE(OBD_FAIL_PTLRPC_CONNECT_RACE);
 308                 msleep(10 * MSEC_PER_SEC);
 309         }
 310         CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
 311         LASSERT(imp->imp_invalid);
 312
 313         /* Wait forever until inflight == 0. We really can't do it another
 314          * way because in some cases we need to wait for very long reply
 315          * unlink. We can't do anything before that because there is really
 316          * no guarantee that some rdma transfer is not in progress right now.
 317          */
 318         do {
 319                 long timeout_jiffies;
 320
 321                 /* Calculate max timeout for waiting on rpcs to error
 322                  * out. Use obd_timeout if calculated value is smaller
 323                  * than it.
 324                  */
 325                 if (!CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 326                         timeout = ptlrpc_inflight_timeout(imp);
 327                         timeout += div_u64(timeout, 3);
 328
 329                         if (timeout == 0)
 330                                 timeout = obd_timeout;
 331                 } else {
 332                         /* decrease the interval to increase race condition */
 333                         timeout = 1;
 334                 }
 335
 336                 CDEBUG(D_RPCTRACE, "Sleeping %llds for inflight to error out\n",
 337                        timeout);
 338
 339                 /* Wait for all requests to error out and call completion
 340                  * callbacks. Cap it at obd_timeout -- these should all
 341                  * have been locally cancelled by ptlrpc_abort_inflight.
 342                  */
 343                 timeout_jiffies = max_t(long, cfs_time_seconds(timeout), 1);
 344                 rc = wait_event_idle_timeout(
 345                                     imp->imp_recovery_waitq,
 346                                     (atomic_read(&imp->imp_inflight) == 0),
 347                                     timeout_jiffies);
 348
 349                 if (rc == 0) {
 350                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 351
 352                         CERROR("%s: timeout waiting for callback (%d != 0)\n",
 353                                cli_tgt, atomic_read(&imp->imp_inflight));
 354
 355                         spin_lock(&imp->imp_lock);
 356                         if (atomic_read(&imp->imp_inflight) == 0) {
 357                                 int count = atomic_read(&imp->imp_unregistering);
 358
 359                                 /* We know that "unregistering" rpcs only can
 360                                  * survive in sending or delaying lists (they
 361                                  * maybe waiting for long reply unlink in
 362                                  * sluggish nets). Check this. If there is no
 363                                  * inflight and unregistering != 0, this is bug
 364                                  */
 365                                 LASSERTF(count == 0, "Some RPCs are still unregistering: %d\n",
 366                                          count);
 367
 368                                 /* Let's save one loop as soon as inflight have
 369                                  * dropped to zero. No new inflights possible at
 370                                  * this point.
 371                                  */
 372                                 rc = 1;
 373                         } else {
 374                                 list_for_each_entry(req, &imp->imp_sending_list,
 375                                                     rq_list) {
 376                                         DEBUG_REQ(D_ERROR, req,
 377                                                   "still on sending list");
 378                                 }
 379                                 list_for_each_entry(req, &imp->imp_delayed_list,
 380                                                     rq_list) {
 381                                         DEBUG_REQ(D_ERROR, req,
 382                                                   "still on delayed list");
 383                                 }
 384
 385                                 CERROR("%s: Unregistering RPCs found (%d). Network is sluggish? Waiting for them to error out.\n",
 386                                        cli_tgt,
 387                                        atomic_read(&imp->imp_unregistering));
 388                         }
 389                         spin_unlock(&imp->imp_lock);
 390                 }
 391         } while (rc == 0);
 392
 393         /*
 394          * Let's additionally check that no new rpcs added to import in
 395          * "invalidate" state.
 396          */
 397         LASSERT(atomic_read(&imp->imp_inflight) == 0);
 398         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 399         sptlrpc_import_flush_all_ctx(imp);
 400
 401         atomic_dec(&imp->imp_inval_count);
 402         wake_up(&imp->imp_recovery_waitq);
 403 }
 404 EXPORT_SYMBOL(ptlrpc_invalidate_import);
 405
 406 /* unset imp_invalid */
 407 void ptlrpc_activate_import(struct obd_import *imp, bool set_state_full)
 408 {
 409         struct obd_device *obd = imp->imp_obd;
 410
 411         spin_lock(&imp->imp_lock);
 412         if (imp->imp_deactive != 0) {
 413                 LASSERT(imp->imp_state != LUSTRE_IMP_FULL);
 414                 if (imp->imp_state != LUSTRE_IMP_DISCON)
 415                         import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
 416                 spin_unlock(&imp->imp_lock);
 417                 return;
 418         }
 419         if (set_state_full)
 420                 import_set_state_nolock(imp, LUSTRE_IMP_FULL);
 421
 422         imp->imp_invalid = 0;
 423
 424         spin_unlock(&imp->imp_lock);
 425         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 426 }
 427 EXPORT_SYMBOL(ptlrpc_activate_import);
 428
 429 void ptlrpc_pinger_force(struct obd_import *imp)
 430 {
 431         CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
 432                ptlrpc_import_state_name(imp->imp_state));
 433
 434         spin_lock(&imp->imp_lock);
 435         imp->imp_force_verify = 1;
 436         spin_unlock(&imp->imp_lock);
 437
 438         if (imp->imp_state != LUSTRE_IMP_CONNECTING)
 439                 ptlrpc_pinger_wake_up();
 440 }
 441 EXPORT_SYMBOL(ptlrpc_pinger_force);
 442
 443 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 444 {
 445         ENTRY;
 446
 447         LASSERT(!imp->imp_dlm_fake);
 448
 449         if (ptlrpc_set_import_discon(imp, conn_cnt, true))
 450                 ptlrpc_pinger_force(imp);
 451
 452         EXIT;
 453 }
 454
 455 int ptlrpc_reconnect_import(struct obd_import *imp)
 456 {
 457         int rc = 0;
 458
 459         ENTRY;
 460
 461         ptlrpc_set_import_discon(imp, 0, true);
 462         /* Force a new connect attempt */
 463         ptlrpc_invalidate_import(imp);
 464         /* Wait for all invalidate calls to finish */
 465         if (atomic_read(&imp->imp_inval_count) > 0) {
 466                 int rc;
 467
 468                 rc = l_wait_event_abortable(
 469                         imp->imp_recovery_waitq,
 470                         (atomic_read(&imp->imp_inval_count) == 0));
 471                 if (rc)
 472                         CERROR("Interrupted, inval=%d\n",
 473                                atomic_read(&imp->imp_inval_count));
 474         }
 475
 476         /* Allow reconnect attempts */
 477         imp->imp_obd->obd_no_recov = 0;
 478         imp->imp_remote_handle.cookie = 0;
 479         /* Attempt a new connect */
 480         rc = ptlrpc_recover_import(imp, NULL, 0);
 481
 482         RETURN(rc);
 483 }
 484 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 485
 486 /**
 487  * Connection on import \a imp is changed to another one (if more than one is
 488  * present). We typically chose connection that we have not tried to connect to
 489  * the longest
 490  */
 491 static int import_select_connection(struct obd_import *imp)
 492 {
 493         struct obd_import_conn *imp_conn = NULL, *conn;
 494         struct obd_export *dlmexp;
 495         char *target_start;
 496         int target_len, tried_all = 1;
 497         int rc = 0;
 498
 499         ENTRY;
 500
 501         spin_lock(&imp->imp_lock);
 502
 503         if (list_empty(&imp->imp_conn_list)) {
 504                 rc = -EINVAL;
 505                 CERROR("%s: no connections available: rc = %d\n",
 506                        imp->imp_obd->obd_name, rc);
 507                 GOTO(out_unlock, rc);
 508         }
 509
 510         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 511                 CDEBUG(D_HA, "%s: connect to NID %s last attempt %lld\n",
 512                        imp->imp_obd->obd_name,
 513                        libcfs_nidstr(&conn->oic_conn->c_peer.nid),
 514                        conn->oic_last_attempt);
 515
 516                 /* If we have not tried this connection since
 517                  * the last successful attempt, go with this one
 518                  */
 519                 if ((conn->oic_last_attempt == 0) ||
 520                     conn->oic_last_attempt <= imp->imp_last_success_conn) {
 521                         imp_conn = conn;
 522                         tried_all = 0;
 523                         break;
 524                 }
 525
 526                 /* If all of the connections have already been tried
 527                  * since the last successful connection; just choose the
 528                  * least recently used
 529                  */
 530                 if (!imp_conn)
 531                         imp_conn = conn;
 532                 else if (imp_conn->oic_last_attempt > conn->oic_last_attempt)
 533                         imp_conn = conn;
 534         }
 535
 536         /* if not found, simply choose the current one */
 537         if (!imp_conn || imp->imp_force_reconnect) {
 538                 LASSERT(imp->imp_conn_current);
 539                 imp_conn = imp->imp_conn_current;
 540                 tried_all = 0;
 541         }
 542         LASSERT(imp_conn->oic_conn);
 543
 544         /* If we've tried everything, and we're back to the beginning of the
 545          * list, increase our timeout and try again. It will be reset when
 546          * we do finally connect. (FIXME: really we should wait for all network
 547          * state associated with the last connection attempt to drain before
 548          * trying to reconnect on it.)
 549          */
 550         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 551                 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
 552                 timeout_t timeout = obd_at_get(imp->imp_obd, at);
 553
 554                 if (timeout < CONNECTION_SWITCH_MAX) {
 555                         obd_at_measure(imp->imp_obd, at,
 556                                        timeout + CONNECTION_SWITCH_INC);
 557                         if (timeout > CONNECTION_SWITCH_MAX)
 558                                 at_reset(at, CONNECTION_SWITCH_MAX);
 559                 }
 560                 LASSERT(imp_conn->oic_last_attempt);
 561                 CDEBUG(D_HA,
 562                        "%s: tried all connections, increasing latency to %ds\n",
 563                        imp->imp_obd->obd_name, timeout);
 564         }
 565
 566         imp_conn->oic_last_attempt = ktime_get_seconds();
 567
 568         /* switch connection, don't mind if it's same as the current one */
 569         ptlrpc_connection_put(imp->imp_connection);
 570         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 571
 572         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 573         if (!dlmexp)
 574                 GOTO(out_unlock, rc = -EINVAL);
 575         ptlrpc_connection_put(dlmexp->exp_connection);
 576         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 577         class_export_put(dlmexp);
 578
 579         if (imp->imp_conn_current != imp_conn) {
 580                 if (imp->imp_conn_current) {
 581                         deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 582                                   &target_start, &target_len);
 583
 584                         CDEBUG(D_HA, "%s: Connection changing to %.*s (at %s)\n",
 585                                imp->imp_obd->obd_name,
 586                                target_len, target_start,
 587                                libcfs_nidstr(&imp_conn->oic_conn->c_peer.nid));
 588                 }
 589
 590                 imp->imp_conn_current = imp_conn;
 591         }
 592
 593         /* The below message is checked in conf-sanity.sh test_35[ab] */
 594         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 595                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 596                libcfs_nidstr(&imp_conn->oic_conn->c_peer.nid));
 597
 598 out_unlock:
 599         spin_unlock(&imp->imp_lock);
 600         RETURN(rc);
 601 }
 602
 603 /*
 604  * must be called under imp_lock
 605  */
 606 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 607 {
 608         struct ptlrpc_request *req;
 609
 610         /* The requests in committed_list always have smaller transnos than
 611          * the requests in replay_list
 612          */
 613         if (!list_empty(&imp->imp_committed_list)) {
 614                 req = list_first_entry(&imp->imp_committed_list,
 615                                        struct ptlrpc_request, rq_replay_list);
 616                 *transno = req->rq_transno;
 617                 if (req->rq_transno == 0) {
 618                         DEBUG_REQ(D_ERROR, req,
 619                                   "zero transno in committed_list");
 620                         LBUG();
 621                 }
 622                 return 1;
 623         }
 624         if (!list_empty(&imp->imp_replay_list)) {
 625                 req = list_first_entry(&imp->imp_replay_list,
 626                                        struct ptlrpc_request, rq_replay_list);
 627                 *transno = req->rq_transno;
 628                 if (req->rq_transno == 0) {
 629                         DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
 630                         LBUG();
 631                 }
 632                 return 1;
 633         }
 634         return 0;
 635 }
 636
 637 int ptlrpc_connect_import(struct obd_import *imp)
 638 {
 639         spin_lock(&imp->imp_lock);
 640         return ptlrpc_connect_import_locked(imp);
 641 }
 642
 643 /**
 644  * Attempt to (re)connect import \a imp. This includes all preparations,
 645  * initializing CONNECT RPC request and passing it to ptlrpcd for
 646  * actual sending.
 647  *
 648  * Assumes imp->imp_lock is held, and releases it.
 649  *
 650  * Returns 0 on success or error code.
 651  */
 652 int ptlrpc_connect_import_locked(struct obd_import *imp)
 653 {
 654         struct obd_device *obd = imp->imp_obd;
 655         int initial_connect = 0;
 656         int set_transno = 0;
 657         __u64 committed_before_reconnect = 0;
 658         struct ptlrpc_request *request;
 659         struct sptlrpc_sepol *sepol;
 660         struct obd_connect_data ocd;
 661         char *bufs[] = { NULL,
 662                          obd2cli_tgt(imp->imp_obd),
 663                          obd->obd_uuid.uuid,
 664                          (char *)&imp->imp_dlm_handle,
 665                          (char *)&ocd,
 666                          NULL };
 667         struct ptlrpc_connect_async_args *aa;
 668         int rc;
 669
 670         ENTRY;
 671
 672         assert_spin_locked(&imp->imp_lock);
 673
 674         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 675                 spin_unlock(&imp->imp_lock);
 676                 CERROR("can't connect to a closed import\n");
 677                 RETURN(-EINVAL);
 678         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 679                 spin_unlock(&imp->imp_lock);
 680                 CERROR("already connected\n");
 681                 RETURN(0);
 682         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING ||
 683                    imp->imp_state == LUSTRE_IMP_EVICTED ||
 684                    imp->imp_connected) {
 685                 spin_unlock(&imp->imp_lock);
 686                 CERROR("already connecting\n");
 687                 RETURN(-EALREADY);
 688         }
 689
 690         import_set_state_nolock(imp, LUSTRE_IMP_CONNECTING);
 691
 692         imp->imp_conn_cnt++;
 693         imp->imp_resend_replay = 0;
 694
 695         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 696                 initial_connect = 1;
 697         else
 698                 committed_before_reconnect = imp->imp_peer_committed_transno;
 699
 700         set_transno = ptlrpc_first_transno(imp,
 701                                            &imp->imp_connect_data.ocd_transno);
 702         spin_unlock(&imp->imp_lock);
 703
 704         rc = import_select_connection(imp);
 705         if (rc)
 706                 GOTO(out, rc);
 707
 708         rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
 709         if (rc)
 710                 GOTO(out, rc);
 711
 712         /* Reset connect flags to the originally requested flags, in case
 713          * the server is updated on-the-fly we will get the new features.
 714          */
 715         ocd = imp->imp_connect_data;
 716         ocd.ocd_connect_flags = imp->imp_connect_flags_orig;
 717         ocd.ocd_connect_flags2 = imp->imp_connect_flags2_orig;
 718         /* Reset ocd_version each time so the server knows the exact versions */
 719         ocd.ocd_version = LUSTRE_VERSION_CODE;
 720         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 721         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 722
 723         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 724                            &obd->obd_uuid, &ocd, NULL);
 725         if (rc)
 726                 GOTO(out, rc);
 727
 728         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 729         if (request == NULL)
 730                 GOTO(out, rc = -ENOMEM);
 731
 732         /* get SELinux policy info if any */
 733         sepol = sptlrpc_sepol_get(request);
 734         if (IS_ERR(sepol)) {
 735                 ptlrpc_request_free(request);
 736                 GOTO(out, rc = PTR_ERR(sepol));
 737         }
 738
 739         bufs[5] = sepol->ssp_sepol;
 740
 741         req_capsule_set_size(&request->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
 742                              sptlrpc_sepol_size(sepol));
 743
 744         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 745                                       imp->imp_connect_op, bufs, NULL);
 746
 747         sptlrpc_sepol_put(sepol);
 748         if (rc) {
 749                 ptlrpc_request_free(request);
 750                 GOTO(out, rc);
 751         }
 752
 753         /* Report the rpc service time to the server so that it knows how long
 754          * to wait for clients to join recovery
 755          */
 756         lustre_msg_set_service_timeout(request->rq_reqmsg,
 757                                        at_timeout2est(request->rq_timeout));
 758
 759         /* The amount of time we give the server to process the connect req.
 760          * import_select_connection will increase the net latency on
 761          * repeated reconnect attempts to cover slow networks.
 762          * We override/ignore the server rpc completion estimate here,
 763          * which may be large if this is a reconnect attempt
 764          */
 765         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 766         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 767
 768         request->rq_no_resend = request->rq_no_delay = 1;
 769         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 770         /* Allow a slightly larger reply for future growth compatibility */
 771         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 772                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 773         ptlrpc_request_set_replen(request);
 774         request->rq_interpret_reply = ptlrpc_connect_interpret;
 775
 776         aa = ptlrpc_req_async_args(aa, request);
 777         memset(aa, 0, sizeof(*aa));
 778
 779         aa->pcaa_peer_committed = committed_before_reconnect;
 780         aa->pcaa_initial_connect = initial_connect;
 781
 782         if (aa->pcaa_initial_connect) {
 783                 spin_lock(&imp->imp_lock);
 784                 imp->imp_replayable = 1;
 785                 spin_unlock(&imp->imp_lock);
 786                 lustre_msg_add_op_flags(request->rq_reqmsg,
 787                                         MSG_CONNECT_INITIAL);
 788         }
 789
 790         if (set_transno)
 791                 lustre_msg_add_op_flags(request->rq_reqmsg,
 792                                         MSG_CONNECT_TRANSNO);
 793
 794         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
 795                   request->rq_timeout);
 796         ptlrpcd_add_req(request);
 797         rc = 0;
 798 out:
 799         if (rc != 0)
 800                 import_set_state(imp, LUSTRE_IMP_DISCON);
 801
 802         RETURN(rc);
 803 }
 804 EXPORT_SYMBOL(ptlrpc_connect_import);
 805
 806 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 807 {
 808         int force_verify;
 809
 810         spin_lock(&imp->imp_lock);
 811         force_verify = imp->imp_force_verify != 0;
 812         spin_unlock(&imp->imp_lock);
 813
 814         if (force_verify)
 815                 ptlrpc_pinger_wake_up();
 816 }
 817
 818 static int ptlrpc_busy_reconnect(int rc)
 819 {
 820         return (rc == -EBUSY) || (rc == -EAGAIN);
 821 }
 822
 823 static int ptlrpc_connect_set_flags(struct obd_import *imp,
 824                                     struct obd_connect_data *ocd,
 825                                     __u64 old_connect_flags,
 826                                     struct obd_export *exp, int init_connect)
 827 {
 828         static bool warned;
 829         struct client_obd *cli = &imp->imp_obd->u.cli;
 830
 831         spin_lock(&imp->imp_lock);
 832         list_move(&imp->imp_conn_current->oic_item,
 833                   &imp->imp_conn_list);
 834         imp->imp_last_success_conn =
 835                 imp->imp_conn_current->oic_last_attempt;
 836
 837         spin_unlock(&imp->imp_lock);
 838
 839         if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 840             (ocd->ocd_version > LUSTRE_VERSION_CODE +
 841                                 LUSTRE_VERSION_OFFSET_WARN ||
 842              ocd->ocd_version < LUSTRE_VERSION_CODE -
 843                                 LUSTRE_VERSION_OFFSET_WARN)) {
 844                 /* few compiler do not like #ifdef in middle of macro argument*/
 845                 const char *older = "older than client. Consider upgrading server"
 846                                     ;
 847                 const char *newer = "newer than client. Consider upgrading client"
 848                                     ;
 849
 850                 LCONSOLE_WARN("Client version (%s). Server %s version (%d.%d.%d.%d) is much %s\n",
 851                               LUSTRE_VERSION_STRING,
 852                               obd2cli_tgt(imp->imp_obd),
 853                               OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 854                               OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 855                               OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 856                               OBD_OCD_VERSION_FIX(ocd->ocd_version),
 857                               ocd->ocd_version > LUSTRE_VERSION_CODE ?
 858                               newer : older);
 859                 warned = true;
 860         }
 861
 862         if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
 863                 /* We sent to the server ocd_cksum_types with bits set
 864                  * for algorithms we understand. The server masked off
 865                  * the checksum types it doesn't support
 866                  */
 867                 if ((ocd->ocd_cksum_types &
 868                      obd_cksum_types_supported_client()) == 0) {
 869                         LCONSOLE_ERROR("The negotiation of the checksum alogrithm to use with server %s failed (%x/%x)\n",
 870                                        obd2cli_tgt(imp->imp_obd),
 871                                        ocd->ocd_cksum_types,
 872                                        obd_cksum_types_supported_client());
 873                         return -EPROTO;
 874                 }
 875                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
 876         } else {
 877                 /* The server does not support OBD_CONNECT_CKSUM.
 878                  * Enforce ADLER for backward compatibility
 879                  */
 880                 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
 881         }
 882         cli->cl_cksum_type = obd_cksum_type_select(imp->imp_obd->obd_name,
 883                                                   cli->cl_supp_cksum_types,
 884                                                   cli->cl_preferred_cksum_type);
 885
 886         if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
 887                 cli->cl_max_pages_per_rpc =
 888                         min(ocd->ocd_brw_size >> PAGE_SHIFT,
 889                             cli->cl_max_pages_per_rpc);
 890         else if (imp->imp_connect_op == MDS_CONNECT ||
 891                  imp->imp_connect_op == MGS_CONNECT)
 892                 cli->cl_max_pages_per_rpc = 1;
 893
 894         LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 895                 (cli->cl_max_pages_per_rpc > 0));
 896
 897         client_adjust_max_dirty(cli);
 898
 899         /* Update client max modify RPCs in flight with value returned
 900          * by the server
 901          */
 902         if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
 903                 cli->cl_max_mod_rpcs_in_flight = min(
 904                                         cli->cl_max_mod_rpcs_in_flight,
 905                                         ocd->ocd_maxmodrpcs);
 906         else
 907                 cli->cl_max_mod_rpcs_in_flight = 1;
 908
 909         /* Reset ns_connect_flags only for initial connect. Might be changed
 910          * while using FS and if we reset it in reconnect this leads to losing
 911          * user settings done before such as disable lru_resize, etc.
 912          */
 913         if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
 914                 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
 915                 __u64 changed_flags;
 916
 917                 changed_flags =
 918                         ns->ns_connect_flags ^ ns->ns_orig_connect_flags;
 919                 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server flags: %#llx\n",
 920                              imp->imp_obd->obd_name,
 921                              ocd->ocd_connect_flags);
 922                 ns->ns_connect_flags = (ns->ns_connect_flags & changed_flags) |
 923                                       (ocd->ocd_connect_flags & ~changed_flags);
 924                 ns->ns_orig_connect_flags = ocd->ocd_connect_flags;
 925         }
 926
 927         if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
 928                 /* We need a per-message support flag, because
 929                  * a. we don't know if the incoming connect reply
 930                  *    supports AT or not (in reply_in_callback)
 931                  *    until we unpack it.
 932                  * b. failovered server means export and flags are gone
 933                  *    (in ptlrpc_send_reply).
 934                  *    Can only be set when we know AT is supported at
 935                  *    both ends
 936                  */
 937                 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 938         else
 939                 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 940
 941         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 942
 943         return 0;
 944 }
 945
 946 /**
 947  * Add all replay requests back to unreplied list before start replay,
 948  * so that we can make sure the known replied XID is always increased
 949  * only even if when replaying requests.
 950  */
 951 static void ptlrpc_prepare_replay(struct obd_import *imp)
 952 {
 953         struct ptlrpc_request *req;
 954
 955         if (imp->imp_state != LUSTRE_IMP_REPLAY ||
 956             imp->imp_resend_replay)
 957                 return;
 958
 959         /* If the server was restart during repaly, the requests may
 960          * have been added to the unreplied list in former replay.
 961          */
 962         spin_lock(&imp->imp_lock);
 963
 964         list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
 965                 if (list_empty(&req->rq_unreplied_list))
 966                         ptlrpc_add_unreplied(req);
 967         }
 968
 969         list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
 970                 if (list_empty(&req->rq_unreplied_list))
 971                         ptlrpc_add_unreplied(req);
 972         }
 973
 974         imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
 975         spin_unlock(&imp->imp_lock);
 976 }
 977
 978 /**
 979  * interpret_reply callback for connect RPCs.
 980  * Looks into returned status of connect operation and decides
 981  * what to do with the import - i.e enter recovery, promote it to
 982  * full state for normal operations of disconnect it due to an error.
 983  */
 984 static int ptlrpc_connect_interpret(const struct lu_env *env,
 985                                     struct ptlrpc_request *request,
 986                                     void *data, int rc)
 987 {
 988         struct ptlrpc_connect_async_args *aa = data;
 989         struct obd_import *imp = request->rq_import;
 990         struct lustre_handle old_hdl;
 991         __u64 old_connect_flags;
 992         timeout_t service_timeout;
 993         int msg_flags;
 994         struct obd_connect_data *ocd;
 995         struct obd_export *exp = NULL;
 996         int ret;
 997
 998         ENTRY;
 999
1000         spin_lock(&imp->imp_lock);
1001         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1002                 imp->imp_connect_tried = 1;
1003                 spin_unlock(&imp->imp_lock);
1004                 RETURN(0);
1005         }
1006
1007         imp->imp_connect_error = rc;
1008         if (rc) {
1009                 struct ptlrpc_request *free_req;
1010                 struct ptlrpc_request *tmp;
1011
1012                 /* abort all delayed requests initiated connection */
1013                 list_for_each_entry_safe(free_req, tmp, &imp->imp_delayed_list,
1014                                          rq_list) {
1015                         spin_lock(&free_req->rq_lock);
1016                         if (free_req->rq_no_resend) {
1017                                 free_req->rq_err = 1;
1018                                 free_req->rq_status = -EIO;
1019                                 ptlrpc_client_wake_req(free_req);
1020                         }
1021                         spin_unlock(&free_req->rq_lock);
1022                 }
1023
1024                 /* if this reconnect to busy export - not need select new target
1025                  * for connecting
1026                  */
1027                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
1028                 spin_unlock(&imp->imp_lock);
1029                 GOTO(out, rc);
1030         }
1031
1032         /* LU-7558: indicate that we are interpretting connect reply,
1033          * pltrpc_connect_import() will not try to reconnect until
1034          * interpret will finish.
1035          */
1036         imp->imp_connected = 1;
1037         spin_unlock(&imp->imp_lock);
1038
1039         LASSERT(imp->imp_conn_current);
1040
1041         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1042
1043         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1044                                    RCL_SERVER);
1045         /* server replied obd_connect_data is always bigger */
1046         ocd = req_capsule_server_sized_get(&request->rq_pill,
1047                                            &RMF_CONNECT_DATA, ret);
1048
1049         if (ocd == NULL) {
1050                 CERROR("%s: no connect data from server\n",
1051                        imp->imp_obd->obd_name);
1052                 rc = -EPROTO;
1053                 GOTO(out, rc);
1054         }
1055
1056         spin_lock(&imp->imp_lock);
1057
1058         /* All imports are pingable */
1059         imp->imp_pingable = 1;
1060         imp->imp_force_reconnect = 0;
1061         imp->imp_force_verify = 0;
1062         imp->imp_setup_time = ktime_get_seconds();
1063
1064         imp->imp_connect_data = *ocd;
1065
1066         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1067                imp->imp_obd->obd_name, ocd->ocd_instance);
1068         exp = class_conn2export(&imp->imp_dlm_handle);
1069
1070         spin_unlock(&imp->imp_lock);
1071
1072         if (!exp) {
1073                 /* Could happen if export is cleaned during connect attempt */
1074                 CERROR("%s: missing export after connect\n",
1075                        imp->imp_obd->obd_name);
1076                 GOTO(out, rc = -ENODEV);
1077         }
1078
1079         /* check that server granted subset of flags we asked for. */
1080         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1081             ocd->ocd_connect_flags) {
1082                 CERROR("%s: Server didn't grant requested subset of flags: asked=%#llx granted=%#llx\n",
1083                        imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
1084                        ocd->ocd_connect_flags);
1085                 GOTO(out, rc = -EPROTO);
1086         }
1087
1088         if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) !=
1089             ocd->ocd_connect_flags2) {
1090                 CERROR("%s: Server didn't grant requested subset of flags2: asked=%#llx granted=%#llx\n",
1091                        imp->imp_obd->obd_name, imp->imp_connect_flags2_orig,
1092                        ocd->ocd_connect_flags2);
1093                 GOTO(out, rc = -EPROTO);
1094         }
1095
1096         if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1097             (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1098             (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1099             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1100                 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1101                 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1102                 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1103
1104                 /* We do not support the MDT-MDT interoperations with
1105                  * different version MDT because of protocol changes.
1106                  */
1107                 if (unlikely(major != LUSTRE_MAJOR ||
1108                              minor != LUSTRE_MINOR ||
1109                              abs(patch - LUSTRE_PATCH) > 3)) {
1110                         LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the connection to different version MDT (%d.%d.%d.%d) %s\n",
1111                                       imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1112                                       LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1113                                       major, minor, patch,
1114                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1115                                       imp->imp_connection->c_remote_uuid.uuid);
1116
1117                         GOTO(out, rc = -EPROTO);
1118                 }
1119         }
1120
1121         old_connect_flags = exp_connect_flags(exp);
1122         exp->exp_connect_data = *ocd;
1123         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1124
1125         /* The net statistics after (re-)connect is not valid anymore,
1126          * because may reflect other routing, etc.
1127          */
1128         service_timeout = lustre_msg_get_service_timeout(request->rq_repmsg);
1129         at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1130         ptlrpc_at_adj_net_latency(request, service_timeout);
1131
1132         /* Import flags should be updated before waking import at FULL state */
1133         rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1134                                       aa->pcaa_initial_connect);
1135         class_export_put(exp);
1136         exp = NULL;
1137
1138         if (rc != 0)
1139                 GOTO(out, rc);
1140
1141         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1142
1143         if (aa->pcaa_initial_connect) {
1144                 spin_lock(&imp->imp_lock);
1145                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1146                         imp->imp_replayable = 1;
1147                         CDEBUG(D_HA, "connected to replayable target: %s\n",
1148                                obd2cli_tgt(imp->imp_obd));
1149                 } else {
1150                         imp->imp_replayable = 0;
1151                 }
1152
1153                 /* if applies, adjust the imp->imp_msg_magic here
1154                  * according to reply flags
1155                  */
1156
1157                 imp->imp_remote_handle =
1158                         *lustre_msg_get_handle(request->rq_repmsg);
1159
1160                 imp->imp_no_cached_data = 1;
1161
1162                 /* Initial connects are allowed for clients with non-random
1163                  * uuids when servers are in recovery.  Simply signal the
1164                  * servers replay is complete and wait in REPLAY_WAIT.
1165                  */
1166                 if (msg_flags & MSG_CONNECT_RECOVERING) {
1167                         CDEBUG(D_HA, "connect to %s during recovery\n",
1168                                obd2cli_tgt(imp->imp_obd));
1169                         import_set_state_nolock(imp, LUSTRE_IMP_REPLAY_LOCKS);
1170                         spin_unlock(&imp->imp_lock);
1171                 } else {
1172                         spin_unlock(&imp->imp_lock);
1173                         ptlrpc_activate_import(imp, true);
1174                 }
1175
1176                 GOTO(finish, rc = 0);
1177         }
1178
1179         /* Determine what recovery state to move the import to. */
1180         if (MSG_CONNECT_RECONNECT & msg_flags) {
1181                 memset(&old_hdl, 0, sizeof(old_hdl));
1182                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1183                             sizeof(old_hdl))) {
1184                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due bad handle %#llx\n",
1185                                       obd2cli_tgt(imp->imp_obd),
1186                                       imp->imp_connection->c_remote_uuid.uuid,
1187                                       imp->imp_dlm_handle.cookie);
1188                         GOTO(out, rc = -ENOTCONN);
1189                 }
1190
1191                 if (memcmp(&imp->imp_remote_handle,
1192                            lustre_msg_get_handle(request->rq_repmsg),
1193                            sizeof(imp->imp_remote_handle))) {
1194                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
1195                                 D_HA : D_WARNING;
1196
1197                         /* Bug 16611/14775: if server handle have changed,
1198                          * that means some sort of disconnection happened.
1199                          * If the server is not in recovery, that also means it
1200                          * already erased all of our state because of previous
1201                          * eviction. If it is in recovery - we are safe to
1202                          * participate since we can reestablish all of our state
1203                          * with server again
1204                          */
1205                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1206                                 CDEBUG_LIMIT(level,
1207                                        "%s@%s changed server handle from %#llx to %#llx but is still in recovery\n",
1208                                        obd2cli_tgt(imp->imp_obd),
1209                                        imp->imp_connection->c_remote_uuid.uuid,
1210                                        imp->imp_remote_handle.cookie,
1211                                        lustre_msg_get_handle(
1212                                                request->rq_repmsg)->cookie);
1213                         } else {
1214                                 LCONSOLE_WARN("Evicted from %s (at %s) after server handle changed from %#llx to %#llx\n",
1215                                               obd2cli_tgt(imp->imp_obd),
1216                                               imp->imp_connection->c_remote_uuid.uuid,
1217                                               imp->imp_remote_handle.cookie,
1218                                               lustre_msg_get_handle(
1219                                                    request->rq_repmsg)->cookie);
1220                         }
1221
1222                         imp->imp_remote_handle =
1223                                 *lustre_msg_get_handle(request->rq_repmsg);
1224
1225                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1226                                 import_set_state(imp, LUSTRE_IMP_EVICTED);
1227                                 GOTO(finish, rc = 0);
1228                         }
1229                 } else {
1230                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1231                                obd2cli_tgt(imp->imp_obd),
1232                                imp->imp_connection->c_remote_uuid.uuid);
1233                 }
1234
1235                 if (imp->imp_invalid) {
1236                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1237                                "marking evicted\n", imp->imp_obd->obd_name);
1238                         import_set_state(imp, LUSTRE_IMP_EVICTED);
1239                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1240                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1241                                imp->imp_obd->obd_name,
1242                                obd2cli_tgt(imp->imp_obd));
1243
1244                         spin_lock(&imp->imp_lock);
1245                         imp->imp_resend_replay = 1;
1246                         spin_unlock(&imp->imp_lock);
1247
1248                         import_set_state(imp, imp->imp_replay_state);
1249                 } else {
1250                         import_set_state(imp, LUSTRE_IMP_RECOVER);
1251                 }
1252         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1253                 LASSERT(imp->imp_replayable);
1254                 imp->imp_remote_handle =
1255                         *lustre_msg_get_handle(request->rq_repmsg);
1256                 imp->imp_last_replay_transno = 0;
1257                 imp->imp_replay_cursor = &imp->imp_committed_list;
1258                 import_set_state(imp, LUSTRE_IMP_REPLAY);
1259         } else if ((ocd->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0 &&
1260                    !imp->imp_invalid) {
1261
1262                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
1263                 /* The below message is checked in recovery-small.sh test_106 */
1264                 DEBUG_REQ(D_HA, request, "%s: lwp recover",
1265                           imp->imp_obd->obd_name);
1266                 imp->imp_remote_handle =
1267                         *lustre_msg_get_handle(request->rq_repmsg);
1268                 import_set_state(imp, LUSTRE_IMP_RECOVER);
1269         } else {
1270                 imp->imp_remote_handle =
1271                         *lustre_msg_get_handle(request->rq_repmsg);
1272                 if (!imp->imp_no_cached_data) {
1273                         DEBUG_REQ(D_HA, request,
1274                                   "%s: evicting (reconnect/recover flags not set: %x)",
1275                                   imp->imp_obd->obd_name, msg_flags);
1276                         import_set_state(imp, LUSTRE_IMP_EVICTED);
1277                 } else {
1278                         ptlrpc_activate_import(imp, true);
1279                 }
1280         }
1281
1282         /* Sanity checks for a reconnected import. */
1283         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE))
1284                 CERROR("imp_replayable flag does not match server after reconnect. We should LBUG right here.\n");
1285
1286         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1287             lustre_msg_get_last_committed(request->rq_repmsg) <
1288             aa->pcaa_peer_committed) {
1289                 static bool printed;
1290
1291                 /* The below message is checked in recovery-small.sh test_54 */
1292                 CERROR("%s: went back in time (transno %lld was previously committed, server now claims %lld)!\n",
1293                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1294                        lustre_msg_get_last_committed(request->rq_repmsg));
1295                 if (!printed) {
1296                         CERROR("For further information, see http://doc.lustre.org/lustre_manual.xhtml#went_back_in_time\n");
1297                         printed = true;
1298                 }
1299         }
1300
1301 finish:
1302         ptlrpc_prepare_replay(imp);
1303         rc = ptlrpc_import_recovery_state_machine(imp);
1304         if (rc == -ENOTCONN) {
1305                 CDEBUG(D_HA,
1306                        "evicted/aborted by %s@%s during recovery; invalidating and reconnecting\n",
1307                        obd2cli_tgt(imp->imp_obd),
1308                        imp->imp_connection->c_remote_uuid.uuid);
1309                 ptlrpc_connect_import(imp);
1310                 spin_lock(&imp->imp_lock);
1311                 imp->imp_connected = 0;
1312                 imp->imp_connect_tried = 1;
1313                 spin_unlock(&imp->imp_lock);
1314                 RETURN(0);
1315         }
1316
1317 out:
1318         if (exp != NULL)
1319                 class_export_put(exp);
1320
1321         spin_lock(&imp->imp_lock);
1322         imp->imp_connected = 0;
1323         imp->imp_connect_tried = 1;
1324
1325         if (rc != 0) {
1326                 bool inact = false;
1327                 time64_t now = ktime_get_seconds();
1328                 time64_t next_connect;
1329
1330                 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
1331                 if (rc == -EACCES || rc == -EROFS) {
1332                         /*
1333                          * Give up trying to reconnect
1334                          * EACCES means client has no permission for connection
1335                          * EROFS means client must mount read-only
1336                          */
1337                         imp->imp_obd->obd_no_recov = 1;
1338                         ptlrpc_deactivate_import_nolock(imp);
1339                         inact = true;
1340                 } else if (rc == -EPROTO) {
1341                         struct obd_connect_data *ocd;
1342
1343                         /* reply message might not be ready */
1344                         if (request->rq_repmsg == NULL) {
1345                                 spin_unlock(&imp->imp_lock);
1346                                 RETURN(-EPROTO);
1347                         }
1348
1349                         ocd = req_capsule_server_get(&request->rq_pill,
1350                                                      &RMF_CONNECT_DATA);
1351                         /* Servers are not supposed to refuse connections from
1352                          * clients based on version, only connection feature
1353                          * flags.  We should never see this from llite, but it
1354                          * may be useful for debugging in the future.
1355                          */
1356                         if (ocd &&
1357                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1358                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1359                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version (%d.%d.%d.%d) refused connection from this client with an incompatible version (%s). Client must be recompiled\n",
1360                                                    obd2cli_tgt(imp->imp_obd),
1361                                                    OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1362                                                    OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1363                                                    OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1364                                                    OBD_OCD_VERSION_FIX(ocd->ocd_version),
1365                                                    LUSTRE_VERSION_STRING);
1366                                 ptlrpc_deactivate_import_nolock(imp);
1367                                 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1368                                 inact = true;
1369                         }
1370                 } else if (rc == -ENODEV || rc == -ETIMEDOUT) {
1371                         /* ENODEV means there is no service, force reconnection
1372                          * to a pair if attempt happen ptlrpc_next_reconnect
1373                          * before now. ETIMEDOUT could be set during network
1374                          * error and do not guarantee request deadline happened.
1375                          */
1376                         struct obd_import_conn *conn;
1377                         time64_t reconnect_time;
1378
1379                         /* Same as ptlrpc_next_reconnect, but in past */
1380                         reconnect_time = now - INITIAL_CONNECT_TIMEOUT;
1381                         list_for_each_entry(conn, &imp->imp_conn_list,
1382                                             oic_item) {
1383                                 if (conn->oic_last_attempt <= reconnect_time) {
1384                                         imp->imp_force_verify = 1;
1385                                         break;
1386                                 }
1387                         }
1388                 }
1389
1390                 next_connect = imp->imp_conn_current->oic_last_attempt +
1391                                (request->rq_deadline - request->rq_sent);
1392                 spin_unlock(&imp->imp_lock);
1393
1394                 if (inact)
1395                         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
1396
1397                 if (rc == -EPROTO)
1398                         RETURN(rc);
1399
1400                 /* adjust imp_next_ping to request deadline + 1 and reschedule
1401                  * a pinger if import lost processing during CONNECTING or far
1402                  * away from request deadline. It could happen when connection
1403                  * was initiated outside of pinger, like
1404                  * ptlrpc_set_import_discon().
1405                  */
1406                 if (!imp->imp_force_verify && (imp->imp_next_ping <= now ||
1407                     imp->imp_next_ping > next_connect)) {
1408                         imp->imp_next_ping = max(now, next_connect) + 1;
1409                         ptlrpc_pinger_wake_up();
1410                 }
1411
1412                 ptlrpc_maybe_ping_import_soon(imp);
1413
1414                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1415                        obd2cli_tgt(imp->imp_obd),
1416                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1417         } else {
1418                 spin_unlock(&imp->imp_lock);
1419         }
1420
1421         wake_up(&imp->imp_recovery_waitq);
1422         RETURN(rc);
1423 }
1424
1425 /**
1426  * interpret callback for "completed replay" RPCs.
1427  * \see signal_completed_replay
1428  */
1429 static int completed_replay_interpret(const struct lu_env *env,
1430                                       struct ptlrpc_request *req,
1431                                       void *args, int rc)
1432 {
1433         ENTRY;
1434         atomic_dec(&req->rq_import->imp_replay_inflight);
1435         if (req->rq_status == 0 && !req->rq_import->imp_vbr_failed) {
1436                 ptlrpc_import_recovery_state_machine(req->rq_import);
1437         } else {
1438                 if (req->rq_import->imp_vbr_failed) {
1439                         CDEBUG(D_WARNING,
1440                                "%s: version recovery fails, reconnecting\n",
1441                                req->rq_import->imp_obd->obd_name);
1442                 } else {
1443                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, reconnecting\n",
1444                                req->rq_import->imp_obd->obd_name,
1445                                req->rq_status);
1446                 }
1447                 ptlrpc_connect_import(req->rq_import);
1448         }
1449
1450         RETURN(0);
1451 }
1452
1453 /**
1454  * Let server know that we have no requests to replay anymore.
1455  * Achieved by just sending a PING request
1456  */
1457 static int signal_completed_replay(struct obd_import *imp)
1458 {
1459         struct ptlrpc_request *req;
1460
1461         ENTRY;
1462
1463         if (unlikely(CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1464                 RETURN(0);
1465
1466         if (!atomic_add_unless(&imp->imp_replay_inflight, 1, 1))
1467                 RETURN(0);
1468
1469         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1470                                         OBD_PING);
1471         if (req == NULL) {
1472                 atomic_dec(&imp->imp_replay_inflight);
1473                 RETURN(-ENOMEM);
1474         }
1475
1476         ptlrpc_request_set_replen(req);
1477         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1478         lustre_msg_add_flags(req->rq_reqmsg,
1479                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1480         if (obd_at_off(imp->imp_obd))
1481                 req->rq_timeout *= 3;
1482         req->rq_interpret_reply = completed_replay_interpret;
1483
1484         ptlrpcd_add_req(req);
1485         RETURN(0);
1486 }
1487
1488 /**
1489  * In kernel code all import invalidation happens in its own
1490  * separate thread, so that whatever application happened to encounter
1491  * a problem could still be killed or otherwise continue
1492  */
1493 static int ptlrpc_invalidate_import_thread(void *data)
1494 {
1495         struct obd_import *imp = data;
1496
1497         ENTRY;
1498         unshare_fs_struct();
1499         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1500                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1501                imp->imp_connection->c_remote_uuid.uuid);
1502
1503         if (do_dump_on_eviction(imp->imp_obd)) {
1504                 CERROR("dump the log upon eviction\n");
1505                 libcfs_debug_dumplog();
1506         }
1507
1508         ptlrpc_invalidate_import(imp);
1509         import_set_state(imp, LUSTRE_IMP_RECOVER);
1510         ptlrpc_import_recovery_state_machine(imp);
1511
1512         class_import_put(imp);
1513         RETURN(0);
1514 }
1515
1516 /**
1517  * This is the state machine for client-side recovery on import.
1518  *
1519  * Typicaly we have two possibly paths. If we came to server and it is not
1520  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1521  * state and reconnect from scratch.
1522  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1523  * We go through our list of requests to replay and send them to server one by
1524  * one.
1525  * After sending all request from the list we change import state to
1526  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1527  * and also all the locks we don't yet have and wait for server to grant us.
1528  * After that we send a special "replay completed" request and change import
1529  * state to IMP_REPLAY_WAIT.
1530  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1531  * state and resend all requests from sending list.
1532  * After that we promote import to FULL state and send all delayed requests
1533  * and import is fully operational after that.
1534  *
1535  */
1536 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1537 {
1538         int rc = 0;
1539         int inflight;
1540         char *target_start;
1541         int target_len;
1542
1543         ENTRY;
1544         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1545                 struct task_struct *task;
1546                 u64 connect_flags;
1547
1548                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1549                           &target_start, &target_len);
1550                 connect_flags = imp->imp_connect_data.ocd_connect_flags;
1551                 /* Don't care about MGC eviction */
1552                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1553                            LUSTRE_MGC_NAME) != 0 &&
1554                     (connect_flags & OBD_CONNECT_LIGHTWEIGHT) == 0) {
1555                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted by %.*s; in progress operations using this service will fail.\n",
1556                                            imp->imp_obd->obd_name, target_len,
1557                                            target_start);
1558                         LASSERTF(!obd_lbug_on_eviction, "LBUG upon eviction\n");
1559                 }
1560                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1561                        obd2cli_tgt(imp->imp_obd),
1562                        imp->imp_connection->c_remote_uuid.uuid);
1563                 /* reset vbr_failed flag upon eviction */
1564                 spin_lock(&imp->imp_lock);
1565                 imp->imp_vbr_failed = 0;
1566                 spin_unlock(&imp->imp_lock);
1567
1568                 /* bug 17802:  XXX client_disconnect_export vs connect request
1569                  * race. if client is evicted at this time then we start
1570                  * invalidate thread without reference to import and import can
1571                  * be freed at same time.
1572                  */
1573                 class_import_get(imp);
1574                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1575                                    "ll_imp_inval");
1576                 if (IS_ERR(task)) {
1577                         class_import_put(imp);
1578                         rc = PTR_ERR(task);
1579                         CERROR("%s: can't start invalidate thread: rc = %d\n",
1580                                imp->imp_obd->obd_name, rc);
1581                 } else {
1582                         rc = 0;
1583                 }
1584                 RETURN(rc);
1585         }
1586
1587         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1588                 CDEBUG(D_HA, "replay requested by %s\n",
1589                        obd2cli_tgt(imp->imp_obd));
1590                 rc = ptlrpc_replay_next(imp, &inflight);
1591                 if (inflight == 0 &&
1592                     atomic_read(&imp->imp_replay_inflight) == 0) {
1593                         import_set_state(imp, LUSTRE_IMP_REPLAY_LOCKS);
1594                         rc = ldlm_replay_locks(imp);
1595                         if (rc)
1596                                 GOTO(out, rc);
1597                 }
1598                 rc = 0;
1599         }
1600
1601         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1602                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1603                         import_set_state(imp, LUSTRE_IMP_REPLAY_WAIT);
1604                         rc = signal_completed_replay(imp);
1605                         if (rc)
1606                                 GOTO(out, rc);
1607                 }
1608         }
1609
1610         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1611                 if (atomic_read(&imp->imp_replay_inflight) == 0)
1612                         import_set_state(imp, LUSTRE_IMP_RECOVER);
1613         }
1614
1615         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1616                 struct ptlrpc_connection *conn = imp->imp_connection;
1617
1618                 rc = ptlrpc_resend(imp);
1619                 if (rc)
1620                         GOTO(out, rc);
1621                 ptlrpc_activate_import(imp, true);
1622
1623                 /* Reverse import are flagged with dlm_fake == 1.
1624                  * They do not do recovery and connection are not "restored".
1625                  */
1626                 if (!imp->imp_dlm_fake)
1627                         CDEBUG_LIMIT(imp->imp_was_idle ?
1628                                         imp->imp_idle_debug : D_CONSOLE,
1629                                      "%s: Connection restored to %s (at %s)\n",
1630                                      imp->imp_obd->obd_name,
1631                                      obd_uuid2str(&conn->c_remote_uuid),
1632                                      obd_import_nid2str(imp));
1633                 spin_lock(&imp->imp_lock);
1634                 imp->imp_was_idle = 0;
1635                 spin_unlock(&imp->imp_lock);
1636         }
1637
1638         if (imp->imp_state == LUSTRE_IMP_FULL) {
1639                 wake_up(&imp->imp_recovery_waitq);
1640                 ptlrpc_wake_delayed(imp);
1641         }
1642
1643 out:
1644         RETURN(rc);
1645 }
1646
1647 static struct ptlrpc_request *ptlrpc_disconnect_prep_req(struct obd_import *imp)
1648 {
1649         struct ptlrpc_request *req;
1650         int rq_opc, rc = 0;
1651
1652         ENTRY;
1653
1654         switch (imp->imp_connect_op) {
1655         case OST_CONNECT:
1656                 rq_opc = OST_DISCONNECT;
1657                 break;
1658         case MDS_CONNECT:
1659                 rq_opc = MDS_DISCONNECT;
1660                 break;
1661         case MGS_CONNECT:
1662                 rq_opc = MGS_DISCONNECT;
1663                 break;
1664         default:
1665                 rc = -EINVAL;
1666                 CERROR("%s: don't know how to disconnect from %s (connect_op %d): rc = %d\n",
1667                        imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1668                        imp->imp_connect_op, rc);
1669                 RETURN(ERR_PTR(rc));
1670         }
1671
1672         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1673                                         LUSTRE_OBD_VERSION, rq_opc);
1674         if (req == NULL)
1675                 RETURN(ERR_PTR(-ENOMEM));
1676
1677         /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1678          * it fails.  We can get through the above with a down server
1679          * if the client doesn't know the server is gone yet.
1680          */
1681         req->rq_no_resend = 1;
1682
1683         /* We want client umounts to happen quickly, no matter server state */
1684         req->rq_timeout = min_t(timeout_t, req->rq_timeout,
1685                                 INITIAL_CONNECT_TIMEOUT);
1686
1687         req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1688         ptlrpc_request_set_replen(req);
1689
1690         RETURN(req);
1691 }
1692
1693 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1694 {
1695         struct ptlrpc_request *req;
1696         int rc = 0;
1697
1698         ENTRY;
1699
1700         if (imp->imp_obd->obd_force)
1701                 GOTO(set_state, rc);
1702
1703         /* probably the import has been disconnected already being idle */
1704         spin_lock(&imp->imp_lock);
1705         if (imp->imp_state == LUSTRE_IMP_IDLE)
1706                 GOTO(out, rc);
1707         spin_unlock(&imp->imp_lock);
1708
1709         if (ptlrpc_import_in_recovery(imp)) {
1710                 long timeout_jiffies;
1711                 time64_t timeout;
1712
1713                 if (obd_at_off(imp->imp_obd)) {
1714                         if (imp->imp_server_timeout)
1715                                 timeout = obd_timeout >> 1;
1716                         else
1717                                 timeout = obd_timeout;
1718                 } else {
1719                         u32 req_portal;
1720                         int idx;
1721
1722                         req_portal = imp->imp_client->cli_request_portal;
1723                         idx = import_at_get_index(imp, req_portal);
1724                         timeout = obd_at_get(imp->imp_obd,
1725                                         &imp->imp_at.iat_service_estimate[idx]);
1726                 }
1727
1728                 timeout_jiffies = cfs_time_seconds(timeout);
1729                 if (wait_event_idle_timeout(imp->imp_recovery_waitq,
1730                                             !ptlrpc_import_in_recovery(imp),
1731                                             timeout_jiffies) == 0 &&
1732                     l_wait_event_abortable(imp->imp_recovery_waitq,
1733                                            !ptlrpc_import_in_recovery(imp)) < 0)
1734                         rc = -EINTR;
1735         }
1736
1737         req = ptlrpc_disconnect_prep_req(imp);
1738         if (IS_ERR(req))
1739                 GOTO(set_state, rc = PTR_ERR(req));
1740
1741         spin_lock(&imp->imp_lock);
1742         if (imp->imp_state != LUSTRE_IMP_FULL) {
1743                 ptlrpc_req_put_with_imp_lock(req);
1744                 GOTO(out, rc);
1745         }
1746         import_set_state_nolock(imp, LUSTRE_IMP_CONNECTING);
1747         spin_unlock(&imp->imp_lock);
1748
1749         rc = ptlrpc_queue_wait(req);
1750         ptlrpc_req_finished(req);
1751
1752 set_state:
1753         spin_lock(&imp->imp_lock);
1754 out:
1755         if (noclose)
1756                 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
1757         else
1758                 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1759         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1760         spin_unlock(&imp->imp_lock);
1761
1762         obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
1763         if (!noclose)
1764                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
1765
1766         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1767                 rc = 0;
1768         RETURN(rc);
1769 }
1770 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1771
1772 static void ptlrpc_reset_reqs_generation(struct obd_import *imp)
1773 {
1774         struct ptlrpc_request *old, *tmp;
1775
1776         /* tag all resendable requests generated before disconnection
1777          * notice this code is part of disconnect-at-idle path only
1778          */
1779         list_for_each_entry_safe(old, tmp, &imp->imp_delayed_list,
1780                         rq_list) {
1781                 spin_lock(&old->rq_lock);
1782                 if (old->rq_import_generation == imp->imp_generation - 1 &&
1783                     ((imp->imp_initiated_at == imp->imp_generation) ||
1784                      !old->rq_no_resend))
1785                         old->rq_import_generation = imp->imp_generation;
1786                 spin_unlock(&old->rq_lock);
1787         }
1788 }
1789
1790 static int ptlrpc_disconnect_idle_interpret(const struct lu_env *env,
1791                                             struct ptlrpc_request *req,
1792                                             void *args, int rc)
1793 {
1794         struct obd_import *imp = req->rq_import;
1795         int connect = 0;
1796
1797         DEBUG_REQ(D_HA, req, "inflight=%d, refcount=%d: rc = %d",
1798                   atomic_read(&imp->imp_inflight),
1799                   refcount_read(&imp->imp_refcount), rc);
1800
1801         spin_lock(&imp->imp_lock);
1802         /* DISCONNECT reply can be late and another connection can just
1803          * be initiated. so we have to abort disconnection.
1804          */
1805         if (req->rq_import_generation == imp->imp_generation &&
1806             imp->imp_state != LUSTRE_IMP_CLOSED) {
1807                 LASSERTF(imp->imp_state == LUSTRE_IMP_CONNECTING,
1808                          "%s\n", ptlrpc_import_state_name(imp->imp_state));
1809                 memset(&imp->imp_remote_handle, 0,
1810                        sizeof(imp->imp_remote_handle));
1811                 /* take our DISCONNECT into account */
1812                 if (atomic_read(&imp->imp_reqs) > 1) {
1813                         imp->imp_generation++;
1814                         imp->imp_initiated_at = imp->imp_generation;
1815                         import_set_state_nolock(imp, LUSTRE_IMP_NEW);
1816                         ptlrpc_reset_reqs_generation(imp);
1817                         connect = 1;
1818                 } else {
1819                         /* do not expose transient IDLE state */
1820                         import_set_state_nolock(imp, LUSTRE_IMP_IDLE);
1821                 }
1822         }
1823
1824         if (connect) {
1825                 rc = ptlrpc_connect_import_locked(imp);
1826                 if (rc >= 0)
1827                         ptlrpc_pinger_add_import(imp);
1828         } else {
1829                 spin_unlock(&imp->imp_lock);
1830         }
1831
1832         return 0;
1833 }
1834
1835 static bool ptlrpc_can_idle(struct obd_import *imp)
1836 {
1837         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1838
1839         /* one request for disconnect rpc */
1840         if (atomic_read(&imp->imp_reqs) > 1)
1841                 return false;
1842
1843         /* any lock increases ns_bref being a resource holder */
1844         if (ns && atomic_read(&ns->ns_bref) > 0)
1845                 return false;
1846
1847         return true;
1848 }
1849
1850 int ptlrpc_disconnect_and_idle_import(struct obd_import *imp)
1851 {
1852         struct ptlrpc_request *req;
1853
1854         ENTRY;
1855
1856         if (imp->imp_obd->obd_force)
1857                 RETURN(0);
1858
1859         if (ptlrpc_import_in_recovery(imp))
1860                 RETURN(0);
1861
1862         req = ptlrpc_disconnect_prep_req(imp);
1863         if (IS_ERR(req))
1864                 RETURN(PTR_ERR(req));
1865
1866         req->rq_interpret_reply = ptlrpc_disconnect_idle_interpret;
1867
1868         if (CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_IDLE_RACE)) {
1869                 __u32 idx;
1870
1871                 server_name2index(imp->imp_obd->obd_name, &idx, NULL);
1872                 if (idx == 0)
1873                         CFS_RACE(OBD_FAIL_PTLRPC_IDLE_RACE);
1874         }
1875
1876         spin_lock(&imp->imp_lock);
1877         if (imp->imp_state != LUSTRE_IMP_FULL || !ptlrpc_can_idle(imp)) {
1878                 ptlrpc_req_put_with_imp_lock(req);
1879                 spin_unlock(&imp->imp_lock);
1880                 RETURN(0);
1881         }
1882         import_set_state_nolock(imp, LUSTRE_IMP_CONNECTING);
1883         /* don't make noise at reconnection */
1884         imp->imp_was_idle = 1;
1885         spin_unlock(&imp->imp_lock);
1886
1887         CDEBUG_LIMIT(imp->imp_idle_debug, "%s: disconnect after %llus idle\n",
1888                      imp->imp_obd->obd_name,
1889                      ktime_get_real_seconds() - imp->imp_last_reply_time);
1890
1891         ptlrpcd_add_req(req);
1892
1893         RETURN(1);
1894 }
1895 EXPORT_SYMBOL(ptlrpc_disconnect_and_idle_import);
1896
1897 void ptlrpc_cleanup_imp(struct obd_import *imp)
1898 {
1899         ENTRY;
1900
1901         spin_lock(&imp->imp_lock);
1902
1903         import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1904         imp->imp_generation++;
1905         ptlrpc_abort_inflight(imp);
1906
1907         spin_unlock(&imp->imp_lock);
1908
1909         EXIT;
1910 }
1911
1912 /* Adaptive Timeout utils */
1913
1914 /* Update at_current_timeout with the specified value (bounded by at_min and
1915  * at_max), as well as the AT history "bins".
1916  *  - Bin into timeslices using AT_BINS bins.
1917  *  - This gives us a max of the last at_history seconds without the storage,
1918  *    but still smoothing out a return to normalcy from a slow response.
1919  *  - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1920  */
1921 timeout_t obd_at_measure(struct obd_device *obd, struct adaptive_timeout *at,
1922                             timeout_t timeout)
1923 {
1924         unsigned int l_at_min = obd_get_at_min(obd);
1925         unsigned int l_at_max = obd_get_at_max(obd);
1926         timeout_t old_timeout = at->at_current_timeout;
1927         time64_t now = ktime_get_real_seconds();
1928         long binlimit = max_t(long, obd_get_at_history(obd) / AT_BINS, 1);
1929
1930         LASSERT(at);
1931         CDEBUG(D_OTHER, "add %u to %p time=%lld v=%u (%u %u %u %u)\n",
1932                timeout, at, now - at->at_binstart, at->at_current_timeout,
1933                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1934
1935         if (timeout <= 0)
1936                 /* Negative timeouts and 0's don't count, because we never
1937                  * want our timeout to drop to 0 or below, and because 0 could
1938                  * mean an error
1939                  */
1940                 return 0;
1941
1942         spin_lock(&at->at_lock);
1943
1944         if (unlikely(at->at_binstart == 0)) {
1945                 /* Special case to remove default from history */
1946                 at->at_current_timeout = timeout;
1947                 at->at_worst_timeout_ever = timeout;
1948                 at->at_worst_timestamp = now;
1949                 at->at_hist[0] = timeout;
1950                 at->at_binstart = now;
1951         } else if (now - at->at_binstart < binlimit) {
1952                 /* in bin 0 */
1953                 at->at_hist[0] = max_t(timeout_t, timeout, at->at_hist[0]);
1954                 at->at_current_timeout = max_t(timeout_t, timeout,
1955                                                at->at_current_timeout);
1956         } else {
1957                 int i, shift;
1958                 timeout_t maxv = timeout;
1959
1960                 /* move bins over */
1961                 shift = (u32)(now - at->at_binstart) / binlimit;
1962                 LASSERT(shift > 0);
1963                 for (i = AT_BINS - 1; i >= 0; i--) {
1964                         if (i >= shift) {
1965                                 at->at_hist[i] = at->at_hist[i - shift];
1966                                 maxv = max_t(timeout_t, maxv, at->at_hist[i]);
1967                         } else {
1968                                 at->at_hist[i] = 0;
1969                         }
1970                 }
1971                 at->at_hist[0] = timeout;
1972                 at->at_current_timeout = maxv;
1973                 at->at_binstart += shift * binlimit;
1974         }
1975
1976         if (at->at_current_timeout > at->at_worst_timeout_ever) {
1977                 at->at_worst_timeout_ever = at->at_current_timeout;
1978                 at->at_worst_timestamp = now;
1979         }
1980
1981         if (at->at_flags & AT_FLG_NOHIST)
1982                 /* Only keep last reported val; keeping the rest of the history
1983                  * for debugfs only
1984                  */
1985                 at->at_current_timeout = timeout;
1986
1987         if (l_at_max > 0)
1988                 at->at_current_timeout = min_t(timeout_t,
1989                                                at->at_current_timeout,
1990                                                l_at_max);
1991         at->at_current_timeout = max_t(timeout_t, at->at_current_timeout,
1992                                        l_at_min);
1993         if (at->at_current_timeout != old_timeout)
1994                 CDEBUG(D_OTHER,
1995                        "AT %p change: old=%u new=%u delta=%d (val=%d) hist %u %u %u %u\n",
1996                        at, old_timeout, at->at_current_timeout,
1997                        at->at_current_timeout - old_timeout, timeout,
1998                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1999                        at->at_hist[3]);
2000
2001         /* if we changed, report the old timeout value */
2002         old_timeout = (at->at_current_timeout != old_timeout) ? old_timeout : 0;
2003
2004         spin_unlock(&at->at_lock);
2005         return old_timeout;
2006 }
2007
2008 /* Find the imp_at index for a given portal; assign if space available */
2009 int import_at_get_index(struct obd_import *imp, int portal)
2010 {
2011         struct imp_at *at = &imp->imp_at;
2012         int i;
2013
2014         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
2015                 if (at->iat_portal[i] == portal)
2016                         return i;
2017                 if (at->iat_portal[i] == 0)
2018                         /* unused */
2019                         break;
2020         }
2021
2022         /* Not found in list, add it under a lock */
2023         spin_lock(&imp->imp_lock);
2024
2025         /* Check unused under lock */
2026         for (; i < IMP_AT_MAX_PORTALS; i++) {
2027                 if (at->iat_portal[i] == portal)
2028                         goto out;
2029                 if (at->iat_portal[i] == 0)
2030                         /* unused */
2031                         break;
2032         }
2033
2034         /* Not enough portals? */
2035         LASSERT(i < IMP_AT_MAX_PORTALS);
2036
2037         at->iat_portal[i] = portal;
2038 out:
2039         spin_unlock(&imp->imp_lock);
2040         return i;
2041 }