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  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ptlrpc/import.c
  33  *
  34  * Author: Mike Shaver <shaver@clusterfs.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38
  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 (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 int ptlrpc_init_import(struct obd_import *imp)
 123 {
 124         spin_lock(&imp->imp_lock);
 125
 126         imp->imp_generation++;
 127         imp->imp_state =  LUSTRE_IMP_NEW;
 128
 129         spin_unlock(&imp->imp_lock);
 130
 131         return 0;
 132 }
 133 EXPORT_SYMBOL(ptlrpc_init_import);
 134
 135 #define UUID_STR "_UUID"
 136 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
 137 {
 138         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 139                 ? uuid : uuid + strlen(prefix);
 140
 141         *uuid_len = strlen(*uuid_start);
 142
 143         if (*uuid_len < strlen(UUID_STR))
 144                 return;
 145
 146         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 147                     UUID_STR, strlen(UUID_STR)))
 148                 *uuid_len -= strlen(UUID_STR);
 149 }
 150
 151 /* Must be called with imp_lock held! */
 152 static void ptlrpc_deactivate_import_nolock(struct obd_import *imp)
 153 {
 154         ENTRY;
 155
 156         assert_spin_locked(&imp->imp_lock);
 157         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 158         imp->imp_invalid = 1;
 159         imp->imp_generation++;
 160
 161         ptlrpc_abort_inflight(imp);
 162
 163         EXIT;
 164 }
 165
 166 /**
 167  * Returns true if import was FULL, false if import was already not
 168  * connected.
 169  * @imp - import to be disconnected
 170  * @conn_cnt - connection count (epoch) of the request that timed out
 171  *             and caused the disconnection.  In some cases, multiple
 172  *             inflight requests can fail to a single target (e.g. OST
 173  *             bulk requests) and if one has already caused a reconnection
 174  *             (increasing the import->conn_cnt) the older failure should
 175  *             not also cause a reconnection.  If zero it forces a reconnect.
 176  * @invalid - set import invalid flag
 177  */
 178 int ptlrpc_set_import_discon(struct obd_import *imp,
 179                              __u32 conn_cnt, bool invalid)
 180 {
 181         int rc = 0;
 182
 183         spin_lock(&imp->imp_lock);
 184
 185         if (imp->imp_state == LUSTRE_IMP_FULL &&
 186             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 187                 char *target_start;
 188                 int   target_len;
 189                 bool  inact = false;
 190
 191                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 192                           &target_start, &target_len);
 193
 194                 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
 195                 if (imp->imp_replayable) {
 196                         LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
 197                                "lost; in progress operations using this "
 198                                "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 "
 203                                "%.*s (at %s) was lost; in progress "
 204                                "operations using this service will fail\n",
 205                                imp->imp_obd->obd_name, target_len, target_start,
 206                                obd_import_nid2str(imp));
 207                         if (invalid) {
 208                                 CDEBUG(D_HA, "import %s@%s for %s not "
 209                                        "replayable, auto-deactivating\n",
 210                                        obd2cli_tgt(imp->imp_obd),
 211                                        imp->imp_connection->c_remote_uuid.uuid,
 212                                        imp->imp_obd->obd_name);
 213                                 ptlrpc_deactivate_import_nolock(imp);
 214                                 inact = true;
 215                         }
 216                 }
 217                 spin_unlock(&imp->imp_lock);
 218
 219                 if (obd_dump_on_timeout)
 220                         libcfs_debug_dumplog();
 221
 222                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 223
 224                 if (inact)
 225                         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 226                 rc = 1;
 227         } else {
 228                 spin_unlock(&imp->imp_lock);
 229                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 230                        imp->imp_client->cli_name, imp,
 231                        (imp->imp_state == LUSTRE_IMP_FULL &&
 232                         imp->imp_conn_cnt > conn_cnt) ?
 233                        "reconnected" : "not connected", imp->imp_conn_cnt,
 234                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 235         }
 236
 237         return rc;
 238 }
 239
 240 /*
 241  * This acts as a barrier; all existing requests are rejected, and
 242  * no new requests will be accepted until the import is valid again.
 243  */
 244 void ptlrpc_deactivate_import(struct obd_import *imp)
 245 {
 246         spin_lock(&imp->imp_lock);
 247         ptlrpc_deactivate_import_nolock(imp);
 248         spin_unlock(&imp->imp_lock);
 249
 250         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 251 }
 252 EXPORT_SYMBOL(ptlrpc_deactivate_import);
 253
 254 static time64_t ptlrpc_inflight_deadline(struct ptlrpc_request *req,
 255                                          time64_t now)
 256 {
 257         time64_t dl;
 258
 259         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
 260               (req->rq_phase == RQ_PHASE_BULK) ||
 261               (req->rq_phase == RQ_PHASE_NEW)))
 262                 return 0;
 263
 264         if (req->rq_timedout)
 265                 return 0;
 266
 267         if (req->rq_phase == RQ_PHASE_NEW)
 268                 dl = req->rq_sent;
 269         else
 270                 dl = req->rq_deadline;
 271
 272         if (dl <= now)
 273                 return 0;
 274
 275         return dl - now;
 276 }
 277
 278 static time64_t ptlrpc_inflight_timeout(struct obd_import *imp)
 279 {
 280         time64_t now = ktime_get_real_seconds();
 281         struct list_head *tmp, *n;
 282         struct ptlrpc_request *req;
 283         time64_t timeout = 0;
 284
 285         spin_lock(&imp->imp_lock);
 286         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 287                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 288                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 289         }
 290         spin_unlock(&imp->imp_lock);
 291         return timeout;
 292 }
 293
 294 /**
 295  * This function will invalidate the import, if necessary, then block
 296  * for all the RPC completions, and finally notify the obd to
 297  * invalidate its state (ie cancel locks, clear pending requests,
 298  * etc).
 299  */
 300 void ptlrpc_invalidate_import(struct obd_import *imp)
 301 {
 302         struct list_head *tmp, *n;
 303         struct ptlrpc_request *req;
 304         struct l_wait_info lwi;
 305         time64_t timeout;
 306         int rc;
 307
 308         atomic_inc(&imp->imp_inval_count);
 309
 310         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 311                 ptlrpc_deactivate_import(imp);
 312
 313         if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CONNECT_RACE)) {
 314                 OBD_RACE(OBD_FAIL_PTLRPC_CONNECT_RACE);
 315                 msleep(10 * MSEC_PER_SEC);
 316         }
 317         CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
 318         LASSERT(imp->imp_invalid);
 319
 320         /* Wait forever until inflight == 0. We really can't do it another
 321          * way because in some cases we need to wait for very long reply
 322          * unlink. We can't do anything before that because there is really
 323          * no guarantee that some rdma transfer is not in progress right now. */
 324         do {
 325                 long timeout_jiffies;
 326
 327                 /* Calculate max timeout for waiting on rpcs to error
 328                  * out. Use obd_timeout if calculated value is smaller
 329                  * than it.
 330                  */
 331                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 332                         timeout = ptlrpc_inflight_timeout(imp);
 333                         timeout += div_u64(timeout, 3);
 334
 335                         if (timeout == 0)
 336                                 timeout = obd_timeout;
 337                 } else {
 338                         /* decrease the interval to increase race condition */
 339                         timeout = 1;
 340                 }
 341
 342                 CDEBUG(D_RPCTRACE, "Sleeping %llds for inflight to error out\n",
 343                        timeout);
 344
 345                 /* Wait for all requests to error out and call completion
 346                  * callbacks. Cap it at obd_timeout -- these should all
 347                  * have been locally cancelled by ptlrpc_abort_inflight.
 348                  */
 349                 timeout_jiffies = max_t(long, cfs_time_seconds(timeout), 1);
 350                 lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
 351                                            (timeout > 1) ? cfs_time_seconds(1) :
 352                                                            cfs_time_seconds(1) / 2,
 353                                                            NULL, NULL);
 354                 rc = l_wait_event(imp->imp_recovery_waitq,
 355                                   (atomic_read(&imp->imp_inflight) == 0),
 356                                   &lwi);
 357                 if (rc) {
 358                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 359
 360                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 361                                cli_tgt, rc, atomic_read(&imp->imp_inflight));
 362
 363                         spin_lock(&imp->imp_lock);
 364                         if (atomic_read(&imp->imp_inflight) == 0) {
 365                                 int count = atomic_read(&imp->imp_unregistering);
 366
 367                                 /* We know that "unregistering" rpcs only can
 368                                  * survive in sending or delaying lists (they
 369                                  * maybe waiting for long reply unlink in
 370                                  * sluggish nets). Let's check this. If there
 371                                  * is no inflight and unregistering != 0, this
 372                                  * is bug. */
 373                                 LASSERTF(count == 0, "Some RPCs are still "
 374                                          "unregistering: %d\n", count);
 375
 376                                 /* Let's save one loop as soon as inflight have
 377                                  * dropped to zero. No new inflights possible at
 378                                  * this point. */
 379                                 rc = 0;
 380                         } else {
 381                                 list_for_each_safe(tmp, n,
 382                                                    &imp->imp_sending_list) {
 383                                         req = list_entry(tmp,
 384                                                          struct ptlrpc_request,
 385                                                          rq_list);
 386                                         DEBUG_REQ(D_ERROR, req,
 387                                                   "still on sending list");
 388                                 }
 389                                 list_for_each_safe(tmp, n,
 390                                                    &imp->imp_delayed_list) {
 391                                         req = list_entry(tmp,
 392                                                          struct ptlrpc_request,
 393                                                          rq_list);
 394                                         DEBUG_REQ(D_ERROR, req,
 395                                                   "still on delayed list");
 396                                 }
 397
 398                                 CERROR("%s: Unregistering RPCs found (%d). "
 399                                        "Network is sluggish? Waiting them "
 400                                        "to error out.\n", cli_tgt,
 401                                        atomic_read(&imp->imp_unregistering));
 402                         }
 403                         spin_unlock(&imp->imp_lock);
 404                 }
 405         } while (rc != 0);
 406
 407         /*
 408          * Let's additionally check that no new rpcs added to import in
 409          * "invalidate" state.
 410          */
 411         LASSERT(atomic_read(&imp->imp_inflight) == 0);
 412         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 413         sptlrpc_import_flush_all_ctx(imp);
 414
 415         atomic_dec(&imp->imp_inval_count);
 416         wake_up_all(&imp->imp_recovery_waitq);
 417 }
 418 EXPORT_SYMBOL(ptlrpc_invalidate_import);
 419
 420 /* unset imp_invalid */
 421 void ptlrpc_activate_import(struct obd_import *imp, bool set_state_full)
 422 {
 423         struct obd_device *obd = imp->imp_obd;
 424
 425         spin_lock(&imp->imp_lock);
 426         if (imp->imp_deactive != 0) {
 427                 LASSERT(imp->imp_state != LUSTRE_IMP_FULL);
 428                 if (imp->imp_state != LUSTRE_IMP_DISCON)
 429                         import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
 430                 spin_unlock(&imp->imp_lock);
 431                 return;
 432         }
 433         if (set_state_full)
 434                 import_set_state_nolock(imp, LUSTRE_IMP_FULL);
 435
 436         imp->imp_invalid = 0;
 437
 438         spin_unlock(&imp->imp_lock);
 439         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 440 }
 441 EXPORT_SYMBOL(ptlrpc_activate_import);
 442
 443 void ptlrpc_pinger_force(struct obd_import *imp)
 444 {
 445         CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
 446                ptlrpc_import_state_name(imp->imp_state));
 447
 448         spin_lock(&imp->imp_lock);
 449         imp->imp_force_verify = 1;
 450         spin_unlock(&imp->imp_lock);
 451
 452         if (imp->imp_state != LUSTRE_IMP_CONNECTING)
 453                 ptlrpc_pinger_wake_up();
 454 }
 455 EXPORT_SYMBOL(ptlrpc_pinger_force);
 456
 457 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 458 {
 459         ENTRY;
 460
 461         LASSERT(!imp->imp_dlm_fake);
 462
 463         if (ptlrpc_set_import_discon(imp, conn_cnt, true))
 464                 ptlrpc_pinger_force(imp);
 465
 466         EXIT;
 467 }
 468
 469 int ptlrpc_reconnect_import(struct obd_import *imp)
 470 {
 471 #ifdef ENABLE_PINGER
 472         long timeout_jiffies = cfs_time_seconds(obd_timeout);
 473         struct l_wait_info lwi;
 474         int rc;
 475
 476         ptlrpc_pinger_force(imp);
 477
 478         CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 479                obd2cli_tgt(imp->imp_obd), obd_timeout);
 480
 481         lwi = LWI_TIMEOUT(timeout_jiffies, NULL, NULL);
 482         rc = l_wait_event(imp->imp_recovery_waitq,
 483                           !ptlrpc_import_in_recovery(imp), &lwi);
 484         CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
 485                ptlrpc_import_state_name(imp->imp_state));
 486         return rc;
 487 #else
 488         ptlrpc_set_import_discon(imp, 0, false);
 489         /* Force a new connect attempt */
 490         ptlrpc_invalidate_import(imp);
 491         /* Do a fresh connect next time by zeroing the handle */
 492         ptlrpc_disconnect_import(imp, 1);
 493         /* Wait for all invalidate calls to finish */
 494         if (atomic_read(&imp->imp_inval_count) > 0) {
 495                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 496                 int rc;
 497
 498                 rc = l_wait_event(imp->imp_recovery_waitq,
 499                                   (atomic_read(&imp->imp_inval_count) == 0),
 500                                   &lwi);
 501                 if (rc)
 502                         CERROR("Interrupted, inval=%d\n",
 503                                atomic_read(&imp->imp_inval_count));
 504         }
 505
 506         /* Allow reconnect attempts */
 507         imp->imp_obd->obd_no_recov = 0;
 508         /* Remove 'invalid' flag */
 509         ptlrpc_activate_import(imp, false);
 510         /* Attempt a new connect */
 511         ptlrpc_recover_import(imp, NULL, 0);
 512         return 0;
 513 #endif
 514 }
 515 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 516
 517 /**
 518  * Connection on import \a imp is changed to another one (if more than one is
 519  * present). We typically chose connection that we have not tried to connect to
 520  * the longest
 521  */
 522 static int import_select_connection(struct obd_import *imp)
 523 {
 524         struct obd_import_conn *imp_conn = NULL, *conn;
 525         struct obd_export *dlmexp;
 526         char *target_start;
 527         int target_len, tried_all = 1;
 528         ENTRY;
 529
 530         spin_lock(&imp->imp_lock);
 531
 532         if (list_empty(&imp->imp_conn_list)) {
 533                 CERROR("%s: no connections available\n",
 534                        imp->imp_obd->obd_name);
 535                 spin_unlock(&imp->imp_lock);
 536                 RETURN(-EINVAL);
 537         }
 538
 539         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 540                 CDEBUG(D_HA, "%s: connect to NID %s last attempt %lld\n",
 541                        imp->imp_obd->obd_name,
 542                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 543                        conn->oic_last_attempt);
 544
 545                 /* If we have not tried this connection since
 546                    the last successful attempt, go with this one */
 547                 if ((conn->oic_last_attempt == 0) ||
 548                     conn->oic_last_attempt <= imp->imp_last_success_conn) {
 549                         imp_conn = conn;
 550                         tried_all = 0;
 551                         break;
 552                 }
 553
 554                 /* If all of the connections have already been tried
 555                    since the last successful connection; just choose the
 556                    least recently used */
 557                 if (!imp_conn)
 558                         imp_conn = conn;
 559                 else if (imp_conn->oic_last_attempt > conn->oic_last_attempt)
 560                         imp_conn = conn;
 561         }
 562
 563         /* if not found, simply choose the current one */
 564         if (!imp_conn || imp->imp_force_reconnect) {
 565                 LASSERT(imp->imp_conn_current);
 566                 imp_conn = imp->imp_conn_current;
 567                 tried_all = 0;
 568         }
 569         LASSERT(imp_conn->oic_conn);
 570
 571         /* If we've tried everything, and we're back to the beginning of the
 572            list, increase our timeout and try again. It will be reset when
 573            we do finally connect. (FIXME: really we should wait for all network
 574            state associated with the last connection attempt to drain before
 575            trying to reconnect on it.) */
 576         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 577                 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
 578                 if (at_get(at) < CONNECTION_SWITCH_MAX) {
 579                         at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
 580                         if (at_get(at) > CONNECTION_SWITCH_MAX)
 581                                 at_reset(at, CONNECTION_SWITCH_MAX);
 582                 }
 583                 LASSERT(imp_conn->oic_last_attempt);
 584                 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
 585                         "to %ds\n", imp->imp_obd->obd_name, at_get(at));
 586         }
 587
 588         imp_conn->oic_last_attempt = ktime_get_seconds();
 589
 590         /* switch connection, don't mind if it's same as the current one */
 591         if (imp->imp_connection)
 592                 ptlrpc_connection_put(imp->imp_connection);
 593         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 594
 595         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 596         LASSERT(dlmexp != NULL);
 597         if (dlmexp->exp_connection)
 598                 ptlrpc_connection_put(dlmexp->exp_connection);
 599         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 600         class_export_put(dlmexp);
 601
 602         if (imp->imp_conn_current != imp_conn) {
 603                 if (imp->imp_conn_current) {
 604                         deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 605                                   &target_start, &target_len);
 606
 607                         CDEBUG(D_HA, "%s: Connection changing to"
 608                                " %.*s (at %s)\n",
 609                                imp->imp_obd->obd_name,
 610                                target_len, target_start,
 611                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 612                 }
 613
 614                 imp->imp_conn_current = imp_conn;
 615         }
 616
 617         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 618                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 619                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 620
 621         spin_unlock(&imp->imp_lock);
 622
 623         RETURN(0);
 624 }
 625
 626 /*
 627  * must be called under imp_lock
 628  */
 629 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 630 {
 631         struct ptlrpc_request   *req;
 632         struct list_head        *tmp;
 633
 634         /* The requests in committed_list always have smaller transnos than
 635          * the requests in replay_list */
 636         if (!list_empty(&imp->imp_committed_list)) {
 637                 tmp = imp->imp_committed_list.next;
 638                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 639                 *transno = req->rq_transno;
 640                 if (req->rq_transno == 0) {
 641                         DEBUG_REQ(D_ERROR, req, "zero transno in committed_list");
 642                         LBUG();
 643                 }
 644                 return 1;
 645         }
 646         if (!list_empty(&imp->imp_replay_list)) {
 647                 tmp = imp->imp_replay_list.next;
 648                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 649                 *transno = req->rq_transno;
 650                 if (req->rq_transno == 0) {
 651                         DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
 652                         LBUG();
 653                 }
 654                 return 1;
 655         }
 656         return 0;
 657 }
 658
 659 /**
 660  * Attempt to (re)connect import \a imp. This includes all preparations,
 661  * initializing CONNECT RPC request and passing it to ptlrpcd for
 662  * actual sending.
 663  * Returns 0 on success or error code.
 664  */
 665 int ptlrpc_connect_import(struct obd_import *imp)
 666 {
 667         struct obd_device *obd = imp->imp_obd;
 668         int initial_connect = 0;
 669         int set_transno = 0;
 670         __u64 committed_before_reconnect = 0;
 671         struct ptlrpc_request *request;
 672         struct obd_connect_data ocd;
 673         char *bufs[] = { NULL,
 674                          obd2cli_tgt(imp->imp_obd),
 675                          obd->obd_uuid.uuid,
 676                          (char *)&imp->imp_dlm_handle,
 677                          (char *)&ocd,
 678                          NULL };
 679         struct ptlrpc_connect_async_args *aa;
 680         int rc;
 681         ENTRY;
 682
 683         spin_lock(&imp->imp_lock);
 684         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 685                 spin_unlock(&imp->imp_lock);
 686                 CERROR("can't connect to a closed import\n");
 687                 RETURN(-EINVAL);
 688         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 689                 spin_unlock(&imp->imp_lock);
 690                 CERROR("already connected\n");
 691                 RETURN(0);
 692         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING ||
 693                    imp->imp_state == LUSTRE_IMP_EVICTED ||
 694                    imp->imp_connected) {
 695                 spin_unlock(&imp->imp_lock);
 696                 CERROR("already connecting\n");
 697                 RETURN(-EALREADY);
 698         }
 699
 700         import_set_state_nolock(imp, LUSTRE_IMP_CONNECTING);
 701
 702         imp->imp_conn_cnt++;
 703         imp->imp_resend_replay = 0;
 704
 705         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 706                 initial_connect = 1;
 707         else
 708                 committed_before_reconnect = imp->imp_peer_committed_transno;
 709
 710         set_transno = ptlrpc_first_transno(imp,
 711                                            &imp->imp_connect_data.ocd_transno);
 712         spin_unlock(&imp->imp_lock);
 713
 714         rc = import_select_connection(imp);
 715         if (rc)
 716                 GOTO(out, rc);
 717
 718         rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
 719         if (rc)
 720                 GOTO(out, rc);
 721
 722         /* Reset connect flags to the originally requested flags, in case
 723          * the server is updated on-the-fly we will get the new features. */
 724         ocd = imp->imp_connect_data;
 725         ocd.ocd_connect_flags = imp->imp_connect_flags_orig;
 726         ocd.ocd_connect_flags2 = imp->imp_connect_flags2_orig;
 727         /* Reset ocd_version each time so the server knows the exact versions */
 728         ocd.ocd_version = LUSTRE_VERSION_CODE;
 729         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 730         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 731
 732         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 733                            &obd->obd_uuid, &ocd, NULL);
 734         if (rc)
 735                 GOTO(out, rc);
 736
 737         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 738         if (request == NULL)
 739                 GOTO(out, rc = -ENOMEM);
 740
 741         /* get SELinux policy info if any */
 742         rc = sptlrpc_get_sepol(request);
 743         if (rc < 0) {
 744                 ptlrpc_request_free(request);
 745                 GOTO(out, rc);
 746         }
 747
 748         bufs[5] = request->rq_sepol;
 749
 750         req_capsule_set_size(&request->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
 751                              strlen(request->rq_sepol) ?
 752                              strlen(request->rq_sepol) + 1 : 0);
 753
 754         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 755                                       imp->imp_connect_op, bufs, NULL);
 756         if (rc) {
 757                 ptlrpc_request_free(request);
 758                 GOTO(out, rc);
 759         }
 760
 761         /* Report the rpc service time to the server so that it knows how long
 762          * to wait for clients to join recovery */
 763         lustre_msg_set_service_time(request->rq_reqmsg,
 764                                     at_timeout2est(request->rq_timeout));
 765
 766         /* The amount of time we give the server to process the connect req.
 767          * import_select_connection will increase the net latency on
 768          * repeated reconnect attempts to cover slow networks.
 769          * We override/ignore the server rpc completion estimate here,
 770          * which may be large if this is a reconnect attempt */
 771         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 772         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 773
 774         request->rq_no_resend = request->rq_no_delay = 1;
 775         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 776         /* Allow a slightly larger reply for future growth compatibility */
 777         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 778                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 779         ptlrpc_request_set_replen(request);
 780         request->rq_interpret_reply = ptlrpc_connect_interpret;
 781
 782         aa = ptlrpc_req_async_args(aa, request);
 783         memset(aa, 0, sizeof *aa);
 784
 785         aa->pcaa_peer_committed = committed_before_reconnect;
 786         aa->pcaa_initial_connect = initial_connect;
 787
 788         if (aa->pcaa_initial_connect) {
 789                 spin_lock(&imp->imp_lock);
 790                 imp->imp_replayable = 1;
 791                 spin_unlock(&imp->imp_lock);
 792                 lustre_msg_add_op_flags(request->rq_reqmsg,
 793                                         MSG_CONNECT_INITIAL);
 794         }
 795
 796         if (set_transno)
 797                 lustre_msg_add_op_flags(request->rq_reqmsg,
 798                                         MSG_CONNECT_TRANSNO);
 799
 800         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %ld)",
 801                   request->rq_timeout);
 802         ptlrpcd_add_req(request);
 803         rc = 0;
 804 out:
 805         if (rc != 0)
 806                 import_set_state(imp, LUSTRE_IMP_DISCON);
 807
 808         RETURN(rc);
 809 }
 810 EXPORT_SYMBOL(ptlrpc_connect_import);
 811
 812 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 813 {
 814         int force_verify;
 815
 816         spin_lock(&imp->imp_lock);
 817         force_verify = imp->imp_force_verify != 0;
 818         spin_unlock(&imp->imp_lock);
 819
 820         if (force_verify)
 821                 ptlrpc_pinger_wake_up();
 822 }
 823
 824 static int ptlrpc_busy_reconnect(int rc)
 825 {
 826         return (rc == -EBUSY) || (rc == -EAGAIN);
 827 }
 828
 829 static int ptlrpc_connect_set_flags(struct obd_import *imp,
 830                                     struct obd_connect_data *ocd,
 831                                     __u64 old_connect_flags,
 832                                     struct obd_export *exp, int init_connect)
 833 {
 834         static bool warned;
 835         struct client_obd *cli = &imp->imp_obd->u.cli;
 836
 837         spin_lock(&imp->imp_lock);
 838         list_move(&imp->imp_conn_current->oic_item,
 839                   &imp->imp_conn_list);
 840         imp->imp_last_success_conn =
 841                 imp->imp_conn_current->oic_last_attempt;
 842
 843         spin_unlock(&imp->imp_lock);
 844
 845         if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 846             (ocd->ocd_version > LUSTRE_VERSION_CODE +
 847                                 LUSTRE_VERSION_OFFSET_WARN ||
 848              ocd->ocd_version < LUSTRE_VERSION_CODE -
 849                                 LUSTRE_VERSION_OFFSET_WARN)) {
 850                 /* Sigh, some compilers do not like #ifdef in the middle
 851                    of macro arguments */
 852                 const char *older = "older than client. "
 853                                     "Consider upgrading server";
 854                 const char *newer = "newer than client. "
 855                                     "Consider upgrading client";
 856
 857                 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 858                               "is much %s (%s)\n",
 859                               obd2cli_tgt(imp->imp_obd),
 860                               OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 861                               OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 862                               OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 863                               OBD_OCD_VERSION_FIX(ocd->ocd_version),
 864                               ocd->ocd_version > LUSTRE_VERSION_CODE ?
 865                               newer : older, LUSTRE_VERSION_STRING);
 866                 warned = true;
 867         }
 868
 869         if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
 870                 /* We sent to the server ocd_cksum_types with bits set
 871                  * for algorithms we understand. The server masked off
 872                  * the checksum types it doesn't support */
 873                 if ((ocd->ocd_cksum_types &
 874                      obd_cksum_types_supported_client()) == 0) {
 875                         LCONSOLE_ERROR("The negotiation of the checksum "
 876                                        "alogrithm to use with server %s "
 877                                        "failed (%x/%x)\n",
 878                                        obd2cli_tgt(imp->imp_obd),
 879                                        ocd->ocd_cksum_types,
 880                                        obd_cksum_types_supported_client());
 881                         return -EPROTO;
 882                 } else {
 883                         cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
 884                 }
 885         } else {
 886                 /* The server does not support OBD_CONNECT_CKSUM.
 887                  * Enforce ADLER for backward compatibility*/
 888                 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
 889         }
 890         cli->cl_cksum_type = obd_cksum_type_select(imp->imp_obd->obd_name,
 891                                                   cli->cl_supp_cksum_types,
 892                                                   cli->cl_preferred_cksum_type);
 893
 894         if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
 895                 cli->cl_max_pages_per_rpc =
 896                         min(ocd->ocd_brw_size >> PAGE_SHIFT,
 897                             cli->cl_max_pages_per_rpc);
 898         else if (imp->imp_connect_op == MDS_CONNECT ||
 899                  imp->imp_connect_op == MGS_CONNECT)
 900                 cli->cl_max_pages_per_rpc = 1;
 901
 902         LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 903                 (cli->cl_max_pages_per_rpc > 0));
 904
 905         client_adjust_max_dirty(cli);
 906
 907         /* Update client max modify RPCs in flight with value returned
 908          * by the server */
 909         if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
 910                 cli->cl_max_mod_rpcs_in_flight = min(
 911                                         cli->cl_max_mod_rpcs_in_flight,
 912                                         ocd->ocd_maxmodrpcs);
 913         else
 914                 cli->cl_max_mod_rpcs_in_flight = 1;
 915
 916         /* Reset ns_connect_flags only for initial connect. It might be
 917          * changed in while using FS and if we reset it in reconnect
 918          * this leads to losing user settings done before such as
 919          * disable lru_resize, etc. */
 920         if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
 921                 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
 922                 __u64 changed_flags;
 923
 924                 changed_flags =
 925                         ns->ns_connect_flags ^ ns->ns_orig_connect_flags;
 926                 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
 927                              "flags: %#llx\n", imp->imp_obd->obd_name,
 928                              ocd->ocd_connect_flags);
 929                 ns->ns_connect_flags = (ns->ns_connect_flags & changed_flags) |
 930                                       (ocd->ocd_connect_flags & ~changed_flags);
 931                 ns->ns_orig_connect_flags = ocd->ocd_connect_flags;
 932         }
 933
 934         if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
 935                 /* We need a per-message support flag, because
 936                  * a. we don't know if the incoming connect reply
 937                  *    supports AT or not (in reply_in_callback)
 938                  *    until we unpack it.
 939                  * b. failovered server means export and flags are gone
 940                  *    (in ptlrpc_send_reply).
 941                  *    Can only be set when we know AT is supported at
 942                  *    both ends */
 943                 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 944         else
 945                 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 946
 947         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 948
 949         return 0;
 950 }
 951
 952 /**
 953  * Add all replay requests back to unreplied list before start replay,
 954  * so that we can make sure the known replied XID is always increased
 955  * only even if when replaying requests.
 956  */
 957 static void ptlrpc_prepare_replay(struct obd_import *imp)
 958 {
 959         struct ptlrpc_request *req;
 960
 961         if (imp->imp_state != LUSTRE_IMP_REPLAY ||
 962             imp->imp_resend_replay)
 963                 return;
 964
 965         /* If the server was restart during repaly, the requests may
 966          * have been added to the unreplied list in former replay. */
 967         spin_lock(&imp->imp_lock);
 968
 969         list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
 970                 if (list_empty(&req->rq_unreplied_list))
 971                         ptlrpc_add_unreplied(req);
 972         }
 973
 974         list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
 975                 if (list_empty(&req->rq_unreplied_list))
 976                         ptlrpc_add_unreplied(req);
 977         }
 978
 979         imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
 980         spin_unlock(&imp->imp_lock);
 981 }
 982
 983 /**
 984  * interpret_reply callback for connect RPCs.
 985  * Looks into returned status of connect operation and decides
 986  * what to do with the import - i.e enter recovery, promote it to
 987  * full state for normal operations of disconnect it due to an error.
 988  */
 989 static int ptlrpc_connect_interpret(const struct lu_env *env,
 990                                     struct ptlrpc_request *request,
 991                                     void *data, int rc)
 992 {
 993         struct ptlrpc_connect_async_args *aa = data;
 994         struct obd_import *imp = request->rq_import;
 995         struct lustre_handle old_hdl;
 996         __u64 old_connect_flags;
 997         int msg_flags;
 998         struct obd_connect_data *ocd;
 999         struct obd_export *exp = NULL;
1000         int ret;
1001         ENTRY;
1002
1003         spin_lock(&imp->imp_lock);
1004         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1005                 imp->imp_connect_tried = 1;
1006                 spin_unlock(&imp->imp_lock);
1007                 RETURN(0);
1008         }
1009
1010         imp->imp_connect_error = rc;
1011         if (rc) {
1012                 struct ptlrpc_request *free_req;
1013                 struct ptlrpc_request *tmp;
1014
1015                 /* abort all delayed requests initiated connection */
1016                 list_for_each_entry_safe(free_req, tmp, &imp->imp_delayed_list,
1017                                          rq_list) {
1018                         spin_lock(&free_req->rq_lock);
1019                         if (free_req->rq_no_resend) {
1020                                 free_req->rq_err = 1;
1021                                 free_req->rq_status = -EIO;
1022                                 ptlrpc_client_wake_req(free_req);
1023                         }
1024                         spin_unlock(&free_req->rq_lock);
1025                 }
1026
1027                 /* if this reconnect to busy export - not need select new target
1028                  * for connecting*/
1029                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
1030                 spin_unlock(&imp->imp_lock);
1031                 ptlrpc_maybe_ping_import_soon(imp);
1032                 GOTO(out, rc);
1033         }
1034
1035         /* LU-7558: indicate that we are interpretting connect reply,
1036          * pltrpc_connect_import() will not try to reconnect until
1037          * interpret will finish. */
1038         imp->imp_connected = 1;
1039         spin_unlock(&imp->imp_lock);
1040
1041         LASSERT(imp->imp_conn_current);
1042
1043         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1044
1045         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1046                                    RCL_SERVER);
1047         /* server replied obd_connect_data is always bigger */
1048         ocd = req_capsule_server_sized_get(&request->rq_pill,
1049                                            &RMF_CONNECT_DATA, ret);
1050
1051         if (ocd == NULL) {
1052                 CERROR("%s: no connect data from server\n",
1053                        imp->imp_obd->obd_name);
1054                 rc = -EPROTO;
1055                 GOTO(out, rc);
1056         }
1057
1058         spin_lock(&imp->imp_lock);
1059
1060         /* All imports are pingable */
1061         imp->imp_pingable = 1;
1062         imp->imp_force_reconnect = 0;
1063         imp->imp_force_verify = 0;
1064
1065         imp->imp_connect_data = *ocd;
1066
1067         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1068                imp->imp_obd->obd_name, ocd->ocd_instance);
1069         exp = class_conn2export(&imp->imp_dlm_handle);
1070
1071         spin_unlock(&imp->imp_lock);
1072
1073         if (!exp) {
1074                 /* This could happen if export is cleaned during the
1075                    connect attempt */
1076                 CERROR("%s: missing export after connect\n",
1077                        imp->imp_obd->obd_name);
1078                 GOTO(out, rc = -ENODEV);
1079         }
1080
1081         /* check that server granted subset of flags we asked for. */
1082         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1083             ocd->ocd_connect_flags) {
1084                 CERROR("%s: Server didn't grant requested subset of flags: "
1085                        "asked=%#llx granted=%#llx\n",
1086                        imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
1087                        ocd->ocd_connect_flags);
1088                 GOTO(out, rc = -EPROTO);
1089         }
1090
1091         if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) !=
1092             ocd->ocd_connect_flags2) {
1093                 CERROR("%s: Server didn't grant requested subset of flags2: "
1094                        "asked=%#llx granted=%#llx\n",
1095                        imp->imp_obd->obd_name, imp->imp_connect_flags2_orig,
1096                        ocd->ocd_connect_flags2);
1097                 GOTO(out, rc = -EPROTO);
1098         }
1099
1100         if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1101             (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1102             (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1103             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1104                 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1105                 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1106                 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1107
1108                 /* We do not support the MDT-MDT interoperations with
1109                  * different version MDT because of protocol changes. */
1110                 if (unlikely(major != LUSTRE_MAJOR ||
1111                              minor != LUSTRE_MINOR ||
1112                              abs(patch - LUSTRE_PATCH) > 3)) {
1113                         LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1114                                       "connection to different version MDT "
1115                                       "(%d.%d.%d.%d) %s\n",
1116                                       imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1117                                       LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1118                                       major, minor, patch,
1119                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1120                                       imp->imp_connection->c_remote_uuid.uuid);
1121
1122                         GOTO(out, rc = -EPROTO);
1123                 }
1124         }
1125
1126         old_connect_flags = exp_connect_flags(exp);
1127         exp->exp_connect_data = *ocd;
1128         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1129
1130         /* The net statistics after (re-)connect is not valid anymore,
1131          * because may reflect other routing, etc. */
1132         at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1133         ptlrpc_at_adj_net_latency(request,
1134                         lustre_msg_get_service_time(request->rq_repmsg));
1135
1136         /* Import flags should be updated before waking import at FULL state */
1137         rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1138                                       aa->pcaa_initial_connect);
1139         class_export_put(exp);
1140         exp = NULL;
1141
1142         if (rc != 0)
1143                 GOTO(out, rc);
1144
1145         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1146
1147         if (aa->pcaa_initial_connect) {
1148                 spin_lock(&imp->imp_lock);
1149                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1150                         imp->imp_replayable = 1;
1151                         CDEBUG(D_HA, "connected to replayable target: %s\n",
1152                                obd2cli_tgt(imp->imp_obd));
1153                 } else {
1154                         imp->imp_replayable = 0;
1155                 }
1156
1157                 /* if applies, adjust the imp->imp_msg_magic here
1158                  * according to reply flags */
1159
1160                 imp->imp_remote_handle =
1161                                 *lustre_msg_get_handle(request->rq_repmsg);
1162
1163                 /* Initial connects are allowed for clients with non-random
1164                  * uuids when servers are in recovery.  Simply signal the
1165                  * servers replay is complete and wait in REPLAY_WAIT. */
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 "
1185                                       "bad handle %#llx\n",
1186                                       obd2cli_tgt(imp->imp_obd),
1187                                       imp->imp_connection->c_remote_uuid.uuid,
1188                                       imp->imp_dlm_handle.cookie);
1189                         GOTO(out, rc = -ENOTCONN);
1190                 }
1191
1192                 if (memcmp(&imp->imp_remote_handle,
1193                            lustre_msg_get_handle(request->rq_repmsg),
1194                            sizeof(imp->imp_remote_handle))) {
1195                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
1196                                 D_HA : D_WARNING;
1197
1198                         /* Bug 16611/14775: if server handle have changed,
1199                          * that means some sort of disconnection happened.
1200                          * If the server is not in recovery, that also means it
1201                          * already erased all of our state because of previous
1202                          * eviction. If it is in recovery - we are safe to
1203                          * participate since we can reestablish all of our state
1204                          * with server again */
1205                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1206                                 CDEBUG(level,"%s@%s changed server handle from "
1207                                        "%#llx to %#llx"
1208                                        " but is still in recovery\n",
1209                                        obd2cli_tgt(imp->imp_obd),
1210                                        imp->imp_connection->c_remote_uuid.uuid,
1211                                        imp->imp_remote_handle.cookie,
1212                                        lustre_msg_get_handle(
1213                                        request->rq_repmsg)->cookie);
1214                         } else {
1215                                 LCONSOLE_WARN("Evicted from %s (at %s) "
1216                                               "after server handle changed from "
1217                                               "%#llx to %#llx\n",
1218                                               obd2cli_tgt(imp->imp_obd),
1219                                               imp->imp_connection-> \
1220                                               c_remote_uuid.uuid,
1221                                               imp->imp_remote_handle.cookie,
1222                                               lustre_msg_get_handle(
1223                                               request->rq_repmsg)->cookie);
1224                         }
1225
1226
1227                         imp->imp_remote_handle =
1228                                      *lustre_msg_get_handle(request->rq_repmsg);
1229
1230                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1231                                 import_set_state(imp, LUSTRE_IMP_EVICTED);
1232                                 GOTO(finish, rc = 0);
1233                         }
1234
1235                 } else {
1236                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1237                                obd2cli_tgt(imp->imp_obd),
1238                                imp->imp_connection->c_remote_uuid.uuid);
1239                 }
1240
1241                 if (imp->imp_invalid) {
1242                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1243                                "marking evicted\n", imp->imp_obd->obd_name);
1244                         import_set_state(imp, LUSTRE_IMP_EVICTED);
1245                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1246                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1247                                imp->imp_obd->obd_name,
1248                                obd2cli_tgt(imp->imp_obd));
1249
1250                         spin_lock(&imp->imp_lock);
1251                         imp->imp_resend_replay = 1;
1252                         spin_unlock(&imp->imp_lock);
1253
1254                         import_set_state(imp, imp->imp_replay_state);
1255                 } else {
1256                         import_set_state(imp, LUSTRE_IMP_RECOVER);
1257                 }
1258         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1259                 LASSERT(imp->imp_replayable);
1260                 imp->imp_remote_handle =
1261                                 *lustre_msg_get_handle(request->rq_repmsg);
1262                 imp->imp_last_replay_transno = 0;
1263                 imp->imp_replay_cursor = &imp->imp_committed_list;
1264                 import_set_state(imp, LUSTRE_IMP_REPLAY);
1265         } else if ((ocd->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0 &&
1266                    !imp->imp_invalid) {
1267
1268                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
1269                 DEBUG_REQ(D_HA, request, "%s: lwp recover",
1270                           imp->imp_obd->obd_name);
1271                 imp->imp_remote_handle =
1272                         *lustre_msg_get_handle(request->rq_repmsg);
1273                 import_set_state(imp, LUSTRE_IMP_RECOVER);
1274         } else {
1275                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
1276                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
1277                 imp->imp_remote_handle =
1278                                 *lustre_msg_get_handle(request->rq_repmsg);
1279                 import_set_state(imp, LUSTRE_IMP_EVICTED);
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 "
1285                        "after reconnect. We should LBUG right here.\n");
1286         }
1287
1288         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1289             lustre_msg_get_last_committed(request->rq_repmsg) <
1290             aa->pcaa_peer_committed) {
1291                 CERROR("%s went back in time (transno %lld"
1292                        " was previously committed, server now claims %lld"
1293                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
1294                        "id=9646\n",
1295                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1296                        lustre_msg_get_last_committed(request->rq_repmsg));
1297         }
1298
1299 finish:
1300         ptlrpc_prepare_replay(imp);
1301         rc = ptlrpc_import_recovery_state_machine(imp);
1302         if (rc == -ENOTCONN) {
1303                 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1304                        "invalidating and reconnecting\n",
1305                        obd2cli_tgt(imp->imp_obd),
1306                        imp->imp_connection->c_remote_uuid.uuid);
1307                 ptlrpc_connect_import(imp);
1308                 spin_lock(&imp->imp_lock);
1309                 imp->imp_connected = 0;
1310                 imp->imp_connect_tried = 1;
1311                 spin_unlock(&imp->imp_lock);
1312                 RETURN(0);
1313         }
1314
1315 out:
1316         if (exp != NULL)
1317                 class_export_put(exp);
1318
1319         spin_lock(&imp->imp_lock);
1320         imp->imp_connected = 0;
1321         imp->imp_connect_tried = 1;
1322
1323         if (rc != 0) {
1324                 bool inact = false;
1325
1326                 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
1327                 if (rc == -EACCES) {
1328                         /*
1329                          * Give up trying to reconnect
1330                          * EACCES means client has no permission for connection
1331                          */
1332                         imp->imp_obd->obd_no_recov = 1;
1333                         ptlrpc_deactivate_import_nolock(imp);
1334                         inact = true;
1335                 } else if (rc == -EPROTO) {
1336                         struct obd_connect_data *ocd;
1337
1338                         /* reply message might not be ready */
1339                         if (request->rq_repmsg == NULL) {
1340                                 spin_unlock(&imp->imp_lock);
1341                                 RETURN(-EPROTO);
1342                         }
1343
1344                         ocd = req_capsule_server_get(&request->rq_pill,
1345                                                      &RMF_CONNECT_DATA);
1346                         /* Servers are not supposed to refuse connections from
1347                          * clients based on version, only connection feature
1348                          * flags.  We should never see this from llite, but it
1349                          * may be useful for debugging in the future. */
1350                         if (ocd &&
1351                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1352                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1353                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1354                                         "(%d.%d.%d.%d)"
1355                                         " refused connection from this client "
1356                                         "with an incompatible version (%s).  "
1357                                         "Client must be recompiled\n",
1358                                         obd2cli_tgt(imp->imp_obd),
1359                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1360                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1361                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1362                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1363                                         LUSTRE_VERSION_STRING);
1364                                 ptlrpc_deactivate_import_nolock(imp);
1365                                 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1366                                 inact = true;
1367                         }
1368                 }
1369                 spin_unlock(&imp->imp_lock);
1370
1371                 if (inact)
1372                         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
1373
1374                 if (rc == -EPROTO)
1375                         RETURN(rc);
1376
1377                 ptlrpc_maybe_ping_import_soon(imp);
1378
1379                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1380                        obd2cli_tgt(imp->imp_obd),
1381                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1382         } else {
1383                 spin_unlock(&imp->imp_lock);
1384         }
1385
1386         wake_up_all(&imp->imp_recovery_waitq);
1387         RETURN(rc);
1388 }
1389
1390 /**
1391  * interpret callback for "completed replay" RPCs.
1392  * \see signal_completed_replay
1393  */
1394 static int completed_replay_interpret(const struct lu_env *env,
1395                                       struct ptlrpc_request *req,
1396                                       void *args, int rc)
1397 {
1398         ENTRY;
1399         atomic_dec(&req->rq_import->imp_replay_inflight);
1400         if (req->rq_status == 0 && !req->rq_import->imp_vbr_failed) {
1401                 ptlrpc_import_recovery_state_machine(req->rq_import);
1402         } else {
1403                 if (req->rq_import->imp_vbr_failed) {
1404                         CDEBUG(D_WARNING,
1405                                "%s: version recovery fails, reconnecting\n",
1406                                req->rq_import->imp_obd->obd_name);
1407                 } else {
1408                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1409                                      "reconnecting\n",
1410                                req->rq_import->imp_obd->obd_name,
1411                                req->rq_status);
1412                 }
1413                 ptlrpc_connect_import(req->rq_import);
1414         }
1415
1416         RETURN(0);
1417 }
1418
1419 /**
1420  * Let server know that we have no requests to replay anymore.
1421  * Achieved by just sending a PING request
1422  */
1423 static int signal_completed_replay(struct obd_import *imp)
1424 {
1425         struct ptlrpc_request *req;
1426         ENTRY;
1427
1428         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1429                 RETURN(0);
1430
1431         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1432         atomic_inc(&imp->imp_replay_inflight);
1433
1434         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1435                                         OBD_PING);
1436         if (req == NULL) {
1437                 atomic_dec(&imp->imp_replay_inflight);
1438                 RETURN(-ENOMEM);
1439         }
1440
1441         ptlrpc_request_set_replen(req);
1442         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1443         lustre_msg_add_flags(req->rq_reqmsg,
1444                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1445         if (AT_OFF)
1446                 req->rq_timeout *= 3;
1447         req->rq_interpret_reply = completed_replay_interpret;
1448
1449         ptlrpcd_add_req(req);
1450         RETURN(0);
1451 }
1452
1453 /**
1454  * In kernel code all import invalidation happens in its own
1455  * separate thread, so that whatever application happened to encounter
1456  * a problem could still be killed or otherwise continue
1457  */
1458 static int ptlrpc_invalidate_import_thread(void *data)
1459 {
1460         struct obd_import *imp = data;
1461
1462         ENTRY;
1463
1464         unshare_fs_struct();
1465
1466         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1467                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1468                imp->imp_connection->c_remote_uuid.uuid);
1469
1470         ptlrpc_invalidate_import(imp);
1471
1472         if (obd_dump_on_eviction) {
1473                 CERROR("dump the log upon eviction\n");
1474                 libcfs_debug_dumplog();
1475         }
1476
1477         import_set_state(imp, LUSTRE_IMP_RECOVER);
1478         ptlrpc_import_recovery_state_machine(imp);
1479
1480         class_import_put(imp);
1481         RETURN(0);
1482 }
1483
1484 /**
1485  * This is the state machine for client-side recovery on import.
1486  *
1487  * Typicaly we have two possibly paths. If we came to server and it is not
1488  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1489  * state and reconnect from scratch.
1490  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1491  * We go through our list of requests to replay and send them to server one by
1492  * one.
1493  * After sending all request from the list we change import state to
1494  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1495  * and also all the locks we don't yet have and wait for server to grant us.
1496  * After that we send a special "replay completed" request and change import
1497  * state to IMP_REPLAY_WAIT.
1498  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1499  * state and resend all requests from sending list.
1500  * After that we promote import to FULL state and send all delayed requests
1501  * and import is fully operational after that.
1502  *
1503  */
1504 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1505 {
1506         int rc = 0;
1507         int inflight;
1508         char *target_start;
1509         int target_len;
1510
1511         ENTRY;
1512         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1513                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1514                           &target_start, &target_len);
1515                 /* Don't care about MGC eviction */
1516                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1517                            LUSTRE_MGC_NAME) != 0) {
1518                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1519                                            "by %.*s; in progress operations "
1520                                            "using this service will fail.\n",
1521                                            imp->imp_obd->obd_name, target_len,
1522                                            target_start);
1523                         LASSERTF(!obd_lbug_on_eviction, "LBUG upon eviction");
1524                 }
1525                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1526                        obd2cli_tgt(imp->imp_obd),
1527                        imp->imp_connection->c_remote_uuid.uuid);
1528                 /* reset vbr_failed flag upon eviction */
1529                 spin_lock(&imp->imp_lock);
1530                 imp->imp_vbr_failed = 0;
1531                 spin_unlock(&imp->imp_lock);
1532
1533                 {
1534                 struct task_struct *task;
1535                 /* bug 17802:  XXX client_disconnect_export vs connect request
1536                  * race. if client is evicted at this time then we start
1537                  * invalidate thread without reference to import and import can
1538                  * be freed at same time. */
1539                 class_import_get(imp);
1540                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1541                                      "ll_imp_inval");
1542                 if (IS_ERR(task)) {
1543                         class_import_put(imp);
1544                         CERROR("error starting invalidate thread: %d\n", rc);
1545                         rc = PTR_ERR(task);
1546                 } else {
1547                         rc = 0;
1548                 }
1549                 RETURN(rc);
1550                 }
1551         }
1552
1553         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1554                 CDEBUG(D_HA, "replay requested by %s\n",
1555                        obd2cli_tgt(imp->imp_obd));
1556                 rc = ptlrpc_replay_next(imp, &inflight);
1557                 if (inflight == 0 &&
1558                     atomic_read(&imp->imp_replay_inflight) == 0) {
1559                         import_set_state(imp, LUSTRE_IMP_REPLAY_LOCKS);
1560                         rc = ldlm_replay_locks(imp);
1561                         if (rc)
1562                                 GOTO(out, rc);
1563                 }
1564                 rc = 0;
1565         }
1566
1567         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1568                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1569                         import_set_state(imp, LUSTRE_IMP_REPLAY_WAIT);
1570                         rc = signal_completed_replay(imp);
1571                         if (rc)
1572                                 GOTO(out, rc);
1573                 }
1574         }
1575
1576         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1577                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1578                         import_set_state(imp, LUSTRE_IMP_RECOVER);
1579                 }
1580         }
1581
1582         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1583                 struct ptlrpc_connection *conn = imp->imp_connection;
1584
1585                 rc = ptlrpc_resend(imp);
1586                 if (rc)
1587                         GOTO(out, rc);
1588                 ptlrpc_activate_import(imp, true);
1589
1590                 LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n",
1591                               imp->imp_obd->obd_name,
1592                               obd_uuid2str(&conn->c_remote_uuid),
1593                               obd_import_nid2str(imp));
1594         }
1595
1596         if (imp->imp_state == LUSTRE_IMP_FULL) {
1597                 wake_up_all(&imp->imp_recovery_waitq);
1598                 ptlrpc_wake_delayed(imp);
1599         }
1600
1601 out:
1602         RETURN(rc);
1603 }
1604
1605 static struct ptlrpc_request *ptlrpc_disconnect_prep_req(struct obd_import *imp)
1606 {
1607         struct ptlrpc_request *req;
1608         int rq_opc, rc = 0;
1609         ENTRY;
1610
1611         switch (imp->imp_connect_op) {
1612         case OST_CONNECT:
1613                 rq_opc = OST_DISCONNECT;
1614                 break;
1615         case MDS_CONNECT:
1616                 rq_opc = MDS_DISCONNECT;
1617                 break;
1618         case MGS_CONNECT:
1619                 rq_opc = MGS_DISCONNECT;
1620                 break;
1621         default:
1622                 rc = -EINVAL;
1623                 CERROR("%s: don't know how to disconnect from %s "
1624                        "(connect_op %d): rc = %d\n",
1625                        imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1626                        imp->imp_connect_op, rc);
1627                 RETURN(ERR_PTR(rc));
1628         }
1629
1630         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1631                                         LUSTRE_OBD_VERSION, rq_opc);
1632         if (req == NULL)
1633                 RETURN(NULL);
1634
1635         /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1636          * it fails.  We can get through the above with a down server
1637          * if the client doesn't know the server is gone yet. */
1638         req->rq_no_resend = 1;
1639
1640         /* We want client umounts to happen quickly, no matter the
1641            server state... */
1642         req->rq_timeout = min_t(int, req->rq_timeout,
1643                                 INITIAL_CONNECT_TIMEOUT);
1644
1645         import_set_state(imp, LUSTRE_IMP_CONNECTING);
1646         req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1647         ptlrpc_request_set_replen(req);
1648
1649         RETURN(req);
1650 }
1651
1652 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1653 {
1654         struct ptlrpc_request *req;
1655         int rc = 0;
1656         ENTRY;
1657
1658         if (imp->imp_obd->obd_force)
1659                 GOTO(set_state, rc);
1660
1661         /* probably the import has been disconnected already being idle */
1662         spin_lock(&imp->imp_lock);
1663         if (imp->imp_state == LUSTRE_IMP_IDLE)
1664                 GOTO(out, rc);
1665         spin_unlock(&imp->imp_lock);
1666
1667         if (ptlrpc_import_in_recovery(imp)) {
1668                 struct l_wait_info lwi;
1669                 long timeout_jiffies;
1670                 time64_t timeout;
1671
1672                 if (AT_OFF) {
1673                         if (imp->imp_server_timeout)
1674                                 timeout = obd_timeout >> 1;
1675                         else
1676                                 timeout = obd_timeout;
1677                 } else {
1678                         u32 req_portal;
1679                         int idx;
1680
1681                         req_portal = imp->imp_client->cli_request_portal;
1682                         idx = import_at_get_index(imp, req_portal);
1683                         timeout = at_get(&imp->imp_at.iat_service_estimate[idx]);
1684                 }
1685
1686                 timeout_jiffies = cfs_time_seconds(timeout);
1687                 lwi = LWI_TIMEOUT_INTR(max_t(long, timeout_jiffies, 1),
1688                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1689                 rc = l_wait_event(imp->imp_recovery_waitq,
1690                                   !ptlrpc_import_in_recovery(imp), &lwi);
1691
1692         }
1693
1694         spin_lock(&imp->imp_lock);
1695         if (imp->imp_state != LUSTRE_IMP_FULL)
1696                 GOTO(out, rc);
1697         spin_unlock(&imp->imp_lock);
1698
1699         req = ptlrpc_disconnect_prep_req(imp);
1700         if (IS_ERR(req))
1701                 GOTO(set_state, rc = PTR_ERR(req));
1702         rc = ptlrpc_queue_wait(req);
1703         ptlrpc_req_finished(req);
1704
1705 set_state:
1706         spin_lock(&imp->imp_lock);
1707 out:
1708         if (noclose)
1709                 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
1710         else
1711                 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1712         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1713         spin_unlock(&imp->imp_lock);
1714
1715         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1716                 rc = 0;
1717         RETURN(rc);
1718 }
1719 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1720
1721 static void ptlrpc_reset_reqs_generation(struct obd_import *imp)
1722 {
1723         struct ptlrpc_request *old, *tmp;
1724
1725         /* tag all resendable requests generated before disconnection
1726          * notice this code is part of disconnect-at-idle path only */
1727         list_for_each_entry_safe(old, tmp, &imp->imp_delayed_list,
1728                         rq_list) {
1729                 spin_lock(&old->rq_lock);
1730                 if (old->rq_import_generation == imp->imp_generation - 1 &&
1731                     !old->rq_no_resend)
1732                         old->rq_import_generation = imp->imp_generation;
1733                 spin_unlock(&old->rq_lock);
1734         }
1735 }
1736
1737 static int ptlrpc_disconnect_idle_interpret(const struct lu_env *env,
1738                                             struct ptlrpc_request *req,
1739                                             void *args, int rc)
1740 {
1741         struct obd_import *imp = req->rq_import;
1742         int connect = 0;
1743
1744         DEBUG_REQ(D_HA, req, "inflight=%d, refcount=%d: rc = %d ",
1745                   atomic_read(&imp->imp_inflight),
1746                   atomic_read(&imp->imp_refcount), rc);
1747
1748         spin_lock(&imp->imp_lock);
1749         /* DISCONNECT reply can be late and another connection can just
1750          * be initiated. so we have to abort disconnection. */
1751         if (req->rq_import_generation == imp->imp_generation &&
1752             imp->imp_state != LUSTRE_IMP_CLOSED) {
1753                 LASSERTF(imp->imp_state == LUSTRE_IMP_CONNECTING,
1754                          "%s\n", ptlrpc_import_state_name(imp->imp_state));
1755                 imp->imp_state = LUSTRE_IMP_IDLE;
1756                 memset(&imp->imp_remote_handle, 0,
1757                        sizeof(imp->imp_remote_handle));
1758                 /* take our DISCONNECT into account */
1759                 if (atomic_read(&imp->imp_inflight) > 1) {
1760                         imp->imp_generation++;
1761                         imp->imp_initiated_at = imp->imp_generation;
1762                         import_set_state_nolock(imp, LUSTRE_IMP_NEW);
1763                         ptlrpc_reset_reqs_generation(imp);
1764                         connect = 1;
1765                 }
1766         }
1767         spin_unlock(&imp->imp_lock);
1768
1769         if (connect) {
1770                 rc = ptlrpc_connect_import(imp);
1771                 if (rc >= 0)
1772                         ptlrpc_pinger_add_import(imp);
1773         }
1774
1775         return 0;
1776 }
1777
1778 int ptlrpc_disconnect_and_idle_import(struct obd_import *imp)
1779 {
1780         struct ptlrpc_request *req;
1781         ENTRY;
1782
1783         if (imp->imp_obd->obd_force)
1784                 RETURN(0);
1785
1786         if (ptlrpc_import_in_recovery(imp))
1787                 RETURN(0);
1788
1789         spin_lock(&imp->imp_lock);
1790         if (imp->imp_state != LUSTRE_IMP_FULL) {
1791                 spin_unlock(&imp->imp_lock);
1792                 RETURN(0);
1793         }
1794         spin_unlock(&imp->imp_lock);
1795
1796         req = ptlrpc_disconnect_prep_req(imp);
1797         if (IS_ERR(req))
1798                 RETURN(PTR_ERR(req));
1799
1800         CDEBUG_LIMIT(imp->imp_idle_debug, "%s: disconnect after %llus idle\n",
1801                      imp->imp_obd->obd_name,
1802                      ktime_get_real_seconds() - imp->imp_last_reply_time);
1803         req->rq_interpret_reply = ptlrpc_disconnect_idle_interpret;
1804         ptlrpcd_add_req(req);
1805
1806         RETURN(0);
1807 }
1808 EXPORT_SYMBOL(ptlrpc_disconnect_and_idle_import);
1809
1810 void ptlrpc_cleanup_imp(struct obd_import *imp)
1811 {
1812         ENTRY;
1813
1814         spin_lock(&imp->imp_lock);
1815
1816         import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1817         imp->imp_generation++;
1818         ptlrpc_abort_inflight(imp);
1819
1820         spin_unlock(&imp->imp_lock);
1821
1822         EXIT;
1823 }
1824
1825 /* Adaptive Timeout utils */
1826
1827 /* Update at_current with the specified value (bounded by at_min and at_max),
1828  * as well as the AT history "bins".
1829  *  - Bin into timeslices using AT_BINS bins.
1830  *  - This gives us a max of the last at_history seconds without the storage,
1831  *    but still smoothing out a return to normalcy from a slow response.
1832  *  - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1833  */
1834 int at_measured(struct adaptive_timeout *at, unsigned int val)
1835 {
1836         unsigned int old = at->at_current;
1837         time64_t now = ktime_get_real_seconds();
1838         long binlimit = max_t(long, at_history / AT_BINS, 1);
1839
1840         LASSERT(at);
1841         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1842                val, at, (long)(now - at->at_binstart), at->at_current,
1843                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1844
1845         if (val == 0)
1846                 /* 0's don't count, because we never want our timeout to
1847                    drop to 0, and because 0 could mean an error */
1848                 return 0;
1849
1850         spin_lock(&at->at_lock);
1851
1852         if (unlikely(at->at_binstart == 0)) {
1853                 /* Special case to remove default from history */
1854                 at->at_current = val;
1855                 at->at_worst_ever = val;
1856                 at->at_worst_time = now;
1857                 at->at_hist[0] = val;
1858                 at->at_binstart = now;
1859         } else if (now - at->at_binstart < binlimit ) {
1860                 /* in bin 0 */
1861                 at->at_hist[0] = max(val, at->at_hist[0]);
1862                 at->at_current = max(val, at->at_current);
1863         } else {
1864                 int i, shift;
1865                 unsigned int maxv = val;
1866
1867                 /* move bins over */
1868                 shift = (u32)(now - at->at_binstart) / binlimit;
1869                 LASSERT(shift > 0);
1870                 for(i = AT_BINS - 1; i >= 0; i--) {
1871                         if (i >= shift) {
1872                                 at->at_hist[i] = at->at_hist[i - shift];
1873                                 maxv = max(maxv, at->at_hist[i]);
1874                         } else {
1875                                 at->at_hist[i] = 0;
1876                         }
1877                 }
1878                 at->at_hist[0] = val;
1879                 at->at_current = maxv;
1880                 at->at_binstart += shift * binlimit;
1881         }
1882
1883         if (at->at_current > at->at_worst_ever) {
1884                 at->at_worst_ever = at->at_current;
1885                 at->at_worst_time = now;
1886         }
1887
1888         if (at->at_flags & AT_FLG_NOHIST)
1889                 /* Only keep last reported val; keeping the rest of the history
1890                    for proc only */
1891                 at->at_current = val;
1892
1893         if (at_max > 0)
1894                 at->at_current =  min(at->at_current, at_max);
1895         at->at_current =  max(at->at_current, at_min);
1896
1897         if (at->at_current != old)
1898                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1899                        "(val=%u) hist %u %u %u %u\n", at,
1900                        old, at->at_current, at->at_current - old, val,
1901                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1902                        at->at_hist[3]);
1903
1904         /* if we changed, report the old value */
1905         old = (at->at_current != old) ? old : 0;
1906
1907         spin_unlock(&at->at_lock);
1908         return old;
1909 }
1910
1911 /* Find the imp_at index for a given portal; assign if space available */
1912 int import_at_get_index(struct obd_import *imp, int portal)
1913 {
1914         struct imp_at *at = &imp->imp_at;
1915         int i;
1916
1917         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1918                 if (at->iat_portal[i] == portal)
1919                         return i;
1920                 if (at->iat_portal[i] == 0)
1921                         /* unused */
1922                         break;
1923         }
1924
1925         /* Not found in list, add it under a lock */
1926         spin_lock(&imp->imp_lock);
1927
1928         /* Check unused under lock */
1929         for (; i < IMP_AT_MAX_PORTALS; i++) {
1930                 if (at->iat_portal[i] == portal)
1931                         goto out;
1932                 if (at->iat_portal[i] == 0)
1933                         /* unused */
1934                         break;
1935         }
1936
1937         /* Not enough portals? */
1938         LASSERT(i < IMP_AT_MAX_PORTALS);
1939
1940         at->iat_portal[i] = portal;
1941 out:
1942         spin_unlock(&imp->imp_lock);
1943         return i;
1944 }