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. Must be called under imp_lock.
  60  */
  61 static void __import_set_state(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         }
  76         imp->imp_state = state;
  77         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
  78         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
  79                 ktime_get_real_seconds();
  80         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
  81                 IMP_STATE_HIST_LEN;
  82 }
  83
  84 /* A CLOSED import should remain so. */
  85 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  86 do {                                                                           \
  87         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  88                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  89                       imp, obd2cli_tgt(imp->imp_obd),                          \
  90                       ptlrpc_import_state_name(imp->imp_state),                \
  91                       ptlrpc_import_state_name(state));                        \
  92                __import_set_state(imp, state);                                 \
  93         }                                                                      \
  94 } while(0)
  95
  96 #define IMPORT_SET_STATE(imp, state)                                    \
  97 do {                                                                    \
  98         spin_lock(&imp->imp_lock);                                      \
  99         IMPORT_SET_STATE_NOLOCK(imp, state);                            \
 100         spin_unlock(&imp->imp_lock);                                    \
 101 } while(0)
 102
 103 void ptlrpc_import_enter_resend(struct obd_import *imp)
 104 {
 105         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 106 }
 107 EXPORT_SYMBOL(ptlrpc_import_enter_resend);
 108
 109
 110 static int ptlrpc_connect_interpret(const struct lu_env *env,
 111                                     struct ptlrpc_request *request,
 112                                     void *args, int rc);
 113 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
 114
 115 /* Only this function is allowed to change the import state when it is
 116  * CLOSED. I would rather refcount the import and free it after
 117  * disconnection like we do with exports. To do that, the client_obd
 118  * will need to save the peer info somewhere other than in the import,
 119  * though. */
 120 int ptlrpc_init_import(struct obd_import *imp)
 121 {
 122         spin_lock(&imp->imp_lock);
 123
 124         imp->imp_generation++;
 125         imp->imp_state =  LUSTRE_IMP_NEW;
 126
 127         spin_unlock(&imp->imp_lock);
 128
 129         return 0;
 130 }
 131 EXPORT_SYMBOL(ptlrpc_init_import);
 132
 133 #define UUID_STR "_UUID"
 134 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
 135 {
 136         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 137                 ? uuid : uuid + strlen(prefix);
 138
 139         *uuid_len = strlen(*uuid_start);
 140
 141         if (*uuid_len < strlen(UUID_STR))
 142                 return;
 143
 144         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 145                     UUID_STR, strlen(UUID_STR)))
 146                 *uuid_len -= strlen(UUID_STR);
 147 }
 148
 149 /**
 150  * Returns true if import was FULL, false if import was already not
 151  * connected.
 152  * @imp - import to be disconnected
 153  * @conn_cnt - connection count (epoch) of the request that timed out
 154  *             and caused the disconnection.  In some cases, multiple
 155  *             inflight requests can fail to a single target (e.g. OST
 156  *             bulk requests) and if one has already caused a reconnection
 157  *             (increasing the import->conn_cnt) the older failure should
 158  *             not also cause a reconnection.  If zero it forces a reconnect.
 159  */
 160 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 161 {
 162         int rc = 0;
 163
 164         spin_lock(&imp->imp_lock);
 165
 166         if (imp->imp_state == LUSTRE_IMP_FULL &&
 167             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 168                 char *target_start;
 169                 int   target_len;
 170
 171                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 172                           &target_start, &target_len);
 173
 174                 if (imp->imp_replayable) {
 175                         LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
 176                                "lost; in progress operations using this "
 177                                "service will wait for recovery to complete\n",
 178                                imp->imp_obd->obd_name, target_len, target_start,
 179                                obd_import_nid2str(imp));
 180                 } else {
 181                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
 182                                "%.*s (at %s) was lost; in progress "
 183                                "operations using this service will fail\n",
 184                                imp->imp_obd->obd_name, target_len, target_start,
 185                                obd_import_nid2str(imp));
 186                 }
 187                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 188                 spin_unlock(&imp->imp_lock);
 189
 190                 if (obd_dump_on_timeout)
 191                         libcfs_debug_dumplog();
 192
 193                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 194                 rc = 1;
 195         } else {
 196                 spin_unlock(&imp->imp_lock);
 197                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 198                        imp->imp_client->cli_name, imp,
 199                        (imp->imp_state == LUSTRE_IMP_FULL &&
 200                         imp->imp_conn_cnt > conn_cnt) ?
 201                        "reconnected" : "not connected", imp->imp_conn_cnt,
 202                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 203         }
 204
 205         return rc;
 206 }
 207
 208 /* Must be called with imp_lock held! */
 209 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 210 {
 211         ENTRY;
 212         assert_spin_locked(&imp->imp_lock);
 213
 214         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 215         imp->imp_invalid = 1;
 216         imp->imp_generation++;
 217         spin_unlock(&imp->imp_lock);
 218
 219         ptlrpc_abort_inflight(imp);
 220         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 221
 222         EXIT;
 223 }
 224
 225 /*
 226  * This acts as a barrier; all existing requests are rejected, and
 227  * no new requests will be accepted until the import is valid again.
 228  */
 229 void ptlrpc_deactivate_import(struct obd_import *imp)
 230 {
 231         spin_lock(&imp->imp_lock);
 232         ptlrpc_deactivate_and_unlock_import(imp);
 233 }
 234 EXPORT_SYMBOL(ptlrpc_deactivate_import);
 235
 236 static time64_t ptlrpc_inflight_deadline(struct ptlrpc_request *req,
 237                                          time64_t now)
 238 {
 239         time64_t dl;
 240
 241         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
 242               (req->rq_phase == RQ_PHASE_BULK) ||
 243               (req->rq_phase == RQ_PHASE_NEW)))
 244                 return 0;
 245
 246         if (req->rq_timedout)
 247                 return 0;
 248
 249         if (req->rq_phase == RQ_PHASE_NEW)
 250                 dl = req->rq_sent;
 251         else
 252                 dl = req->rq_deadline;
 253
 254         if (dl <= now)
 255                 return 0;
 256
 257         return dl - now;
 258 }
 259
 260 static time64_t ptlrpc_inflight_timeout(struct obd_import *imp)
 261 {
 262         time64_t now = ktime_get_real_seconds();
 263         struct list_head *tmp, *n;
 264         struct ptlrpc_request *req;
 265         time64_t timeout = 0;
 266
 267         spin_lock(&imp->imp_lock);
 268         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 269                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 270                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 271         }
 272         spin_unlock(&imp->imp_lock);
 273         return timeout;
 274 }
 275
 276 /**
 277  * This function will invalidate the import, if necessary, then block
 278  * for all the RPC completions, and finally notify the obd to
 279  * invalidate its state (ie cancel locks, clear pending requests,
 280  * etc).
 281  */
 282 void ptlrpc_invalidate_import(struct obd_import *imp)
 283 {
 284         struct list_head *tmp, *n;
 285         struct ptlrpc_request *req;
 286         struct l_wait_info lwi;
 287         time64_t timeout;
 288         int rc;
 289
 290         atomic_inc(&imp->imp_inval_count);
 291
 292         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 293                 ptlrpc_deactivate_import(imp);
 294
 295         if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CONNECT_RACE)) {
 296                 OBD_RACE(OBD_FAIL_PTLRPC_CONNECT_RACE);
 297                 msleep(10 * MSEC_PER_SEC);
 298         }
 299         CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
 300         LASSERT(imp->imp_invalid);
 301
 302         /* Wait forever until inflight == 0. We really can't do it another
 303          * way because in some cases we need to wait for very long reply
 304          * unlink. We can't do anything before that because there is really
 305          * no guarantee that some rdma transfer is not in progress right now. */
 306         do {
 307                 long timeout_jiffies;
 308
 309                 /* Calculate max timeout for waiting on rpcs to error
 310                  * out. Use obd_timeout if calculated value is smaller
 311                  * than it.
 312                  */
 313                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 314                         timeout = ptlrpc_inflight_timeout(imp);
 315                         timeout += div_u64(timeout, 3);
 316
 317                         if (timeout == 0)
 318                                 timeout = obd_timeout;
 319                 } else {
 320                         /* decrease the interval to increase race condition */
 321                         timeout = 1;
 322                 }
 323
 324                 CDEBUG(D_RPCTRACE, "Sleeping %llds for inflight to error out\n",
 325                        timeout);
 326
 327                 /* Wait for all requests to error out and call completion
 328                  * callbacks. Cap it at obd_timeout -- these should all
 329                  * have been locally cancelled by ptlrpc_abort_inflight.
 330                  */
 331                 timeout_jiffies = max_t(long, cfs_time_seconds(timeout), 1);
 332                 lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
 333                                            (timeout > 1) ? cfs_time_seconds(1) :
 334                                                            cfs_time_seconds(1) / 2,
 335                                                            NULL, NULL);
 336                 rc = l_wait_event(imp->imp_recovery_waitq,
 337                                   (atomic_read(&imp->imp_inflight) == 0),
 338                                   &lwi);
 339                 if (rc) {
 340                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 341
 342                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 343                                cli_tgt, rc, atomic_read(&imp->imp_inflight));
 344
 345                         spin_lock(&imp->imp_lock);
 346                         if (atomic_read(&imp->imp_inflight) == 0) {
 347                                 int count = atomic_read(&imp->imp_unregistering);
 348
 349                                 /* We know that "unregistering" rpcs only can
 350                                  * survive in sending or delaying lists (they
 351                                  * maybe waiting for long reply unlink in
 352                                  * sluggish nets). Let's check this. If there
 353                                  * is no inflight and unregistering != 0, this
 354                                  * is bug. */
 355                                 LASSERTF(count == 0, "Some RPCs are still "
 356                                          "unregistering: %d\n", count);
 357
 358                                 /* Let's save one loop as soon as inflight have
 359                                  * dropped to zero. No new inflights possible at
 360                                  * this point. */
 361                                 rc = 0;
 362                         } else {
 363                                 list_for_each_safe(tmp, n,
 364                                                    &imp->imp_sending_list) {
 365                                         req = list_entry(tmp,
 366                                                          struct ptlrpc_request,
 367                                                          rq_list);
 368                                         DEBUG_REQ(D_ERROR, req,
 369                                                   "still on sending list");
 370                                 }
 371                                 list_for_each_safe(tmp, n,
 372                                                    &imp->imp_delayed_list) {
 373                                         req = list_entry(tmp,
 374                                                          struct ptlrpc_request,
 375                                                          rq_list);
 376                                         DEBUG_REQ(D_ERROR, req,
 377                                                   "still on delayed list");
 378                                 }
 379
 380                                 CERROR("%s: Unregistering RPCs found (%d). "
 381                                        "Network is sluggish? Waiting them "
 382                                        "to error out.\n", cli_tgt,
 383                                        atomic_read(&imp->imp_unregistering));
 384                         }
 385                         spin_unlock(&imp->imp_lock);
 386                 }
 387         } while (rc != 0);
 388
 389         /*
 390          * Let's additionally check that no new rpcs added to import in
 391          * "invalidate" state.
 392          */
 393         LASSERT(atomic_read(&imp->imp_inflight) == 0);
 394         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 395         sptlrpc_import_flush_all_ctx(imp);
 396
 397         atomic_dec(&imp->imp_inval_count);
 398         wake_up_all(&imp->imp_recovery_waitq);
 399 }
 400 EXPORT_SYMBOL(ptlrpc_invalidate_import);
 401
 402 /* unset imp_invalid */
 403 void ptlrpc_activate_import(struct obd_import *imp)
 404 {
 405         struct obd_device *obd = imp->imp_obd;
 406
 407         spin_lock(&imp->imp_lock);
 408         if (imp->imp_deactive != 0) {
 409                 spin_unlock(&imp->imp_lock);
 410                 return;
 411         }
 412
 413         imp->imp_invalid = 0;
 414         spin_unlock(&imp->imp_lock);
 415         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 416 }
 417 EXPORT_SYMBOL(ptlrpc_activate_import);
 418
 419 void ptlrpc_pinger_force(struct obd_import *imp)
 420 {
 421         CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
 422                ptlrpc_import_state_name(imp->imp_state));
 423
 424         spin_lock(&imp->imp_lock);
 425         imp->imp_force_verify = 1;
 426         spin_unlock(&imp->imp_lock);
 427
 428         if (imp->imp_state != LUSTRE_IMP_CONNECTING)
 429                 ptlrpc_pinger_wake_up();
 430 }
 431 EXPORT_SYMBOL(ptlrpc_pinger_force);
 432
 433 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 434 {
 435         ENTRY;
 436
 437         LASSERT(!imp->imp_dlm_fake);
 438
 439         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 440                 if (!imp->imp_replayable) {
 441                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 442                                "auto-deactivating\n",
 443                                obd2cli_tgt(imp->imp_obd),
 444                                imp->imp_connection->c_remote_uuid.uuid,
 445                                imp->imp_obd->obd_name);
 446                         ptlrpc_deactivate_import(imp);
 447                 }
 448
 449                 ptlrpc_pinger_force(imp);
 450         }
 451         EXIT;
 452 }
 453
 454 int ptlrpc_reconnect_import(struct obd_import *imp)
 455 {
 456 #ifdef ENABLE_PINGER
 457         long timeout_jiffies = cfs_time_seconds(obd_timeout);
 458         struct l_wait_info lwi;
 459         int rc;
 460
 461         ptlrpc_pinger_force(imp);
 462
 463         CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 464                obd2cli_tgt(imp->imp_obd), obd_timeout);
 465
 466         lwi = LWI_TIMEOUT(timeout_jiffies, NULL, NULL);
 467         rc = l_wait_event(imp->imp_recovery_waitq,
 468                           !ptlrpc_import_in_recovery(imp), &lwi);
 469         CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
 470                ptlrpc_import_state_name(imp->imp_state));
 471         return rc;
 472 #else
 473         ptlrpc_set_import_discon(imp, 0);
 474         /* Force a new connect attempt */
 475         ptlrpc_invalidate_import(imp);
 476         /* Do a fresh connect next time by zeroing the handle */
 477         ptlrpc_disconnect_import(imp, 1);
 478         /* Wait for all invalidate calls to finish */
 479         if (atomic_read(&imp->imp_inval_count) > 0) {
 480                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 481                 int rc;
 482
 483                 rc = l_wait_event(imp->imp_recovery_waitq,
 484                                   (atomic_read(&imp->imp_inval_count) == 0),
 485                                   &lwi);
 486                 if (rc)
 487                         CERROR("Interrupted, inval=%d\n",
 488                                atomic_read(&imp->imp_inval_count));
 489         }
 490
 491         /* Allow reconnect attempts */
 492         imp->imp_obd->obd_no_recov = 0;
 493         /* Remove 'invalid' flag */
 494         ptlrpc_activate_import(imp);
 495         /* Attempt a new connect */
 496         ptlrpc_recover_import(imp, NULL, 0);
 497         return 0;
 498 #endif
 499 }
 500 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 501
 502 /**
 503  * Connection on import \a imp is changed to another one (if more than one is
 504  * present). We typically chose connection that we have not tried to connect to
 505  * the longest
 506  */
 507 static int import_select_connection(struct obd_import *imp)
 508 {
 509         struct obd_import_conn *imp_conn = NULL, *conn;
 510         struct obd_export *dlmexp;
 511         char *target_start;
 512         int target_len, tried_all = 1;
 513         ENTRY;
 514
 515         spin_lock(&imp->imp_lock);
 516
 517         if (list_empty(&imp->imp_conn_list)) {
 518                 CERROR("%s: no connections available\n",
 519                        imp->imp_obd->obd_name);
 520                 spin_unlock(&imp->imp_lock);
 521                 RETURN(-EINVAL);
 522         }
 523
 524         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 525                 CDEBUG(D_HA, "%s: connect to NID %s last attempt %lld\n",
 526                        imp->imp_obd->obd_name,
 527                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 528                        conn->oic_last_attempt);
 529
 530                 /* If we have not tried this connection since
 531                    the last successful attempt, go with this one */
 532                 if ((conn->oic_last_attempt == 0) ||
 533                     conn->oic_last_attempt <= imp->imp_last_success_conn) {
 534                         imp_conn = conn;
 535                         tried_all = 0;
 536                         break;
 537                 }
 538
 539                 /* If all of the connections have already been tried
 540                    since the last successful connection; just choose the
 541                    least recently used */
 542                 if (!imp_conn)
 543                         imp_conn = conn;
 544                 else if (imp_conn->oic_last_attempt > conn->oic_last_attempt)
 545                         imp_conn = conn;
 546         }
 547
 548         /* if not found, simply choose the current one */
 549         if (!imp_conn || imp->imp_force_reconnect) {
 550                 LASSERT(imp->imp_conn_current);
 551                 imp_conn = imp->imp_conn_current;
 552                 tried_all = 0;
 553         }
 554         LASSERT(imp_conn->oic_conn);
 555
 556         /* If we've tried everything, and we're back to the beginning of the
 557            list, increase our timeout and try again. It will be reset when
 558            we do finally connect. (FIXME: really we should wait for all network
 559            state associated with the last connection attempt to drain before
 560            trying to reconnect on it.) */
 561         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 562                 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
 563                 if (at_get(at) < CONNECTION_SWITCH_MAX) {
 564                         at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
 565                         if (at_get(at) > CONNECTION_SWITCH_MAX)
 566                                 at_reset(at, CONNECTION_SWITCH_MAX);
 567                 }
 568                 LASSERT(imp_conn->oic_last_attempt);
 569                 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
 570                         "to %ds\n", imp->imp_obd->obd_name, at_get(at));
 571         }
 572
 573         imp_conn->oic_last_attempt = ktime_get_seconds();
 574
 575         /* switch connection, don't mind if it's same as the current one */
 576         if (imp->imp_connection)
 577                 ptlrpc_connection_put(imp->imp_connection);
 578         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 579
 580         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 581         LASSERT(dlmexp != NULL);
 582         if (dlmexp->exp_connection)
 583                 ptlrpc_connection_put(dlmexp->exp_connection);
 584         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 585         class_export_put(dlmexp);
 586
 587         if (imp->imp_conn_current != imp_conn) {
 588                 if (imp->imp_conn_current) {
 589                         deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 590                                   &target_start, &target_len);
 591
 592                         CDEBUG(D_HA, "%s: Connection changing to"
 593                                " %.*s (at %s)\n",
 594                                imp->imp_obd->obd_name,
 595                                target_len, target_start,
 596                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 597                 }
 598
 599                 imp->imp_conn_current = imp_conn;
 600         }
 601
 602         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 603                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 604                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 605
 606         spin_unlock(&imp->imp_lock);
 607
 608         RETURN(0);
 609 }
 610
 611 /*
 612  * must be called under imp_lock
 613  */
 614 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 615 {
 616         struct ptlrpc_request   *req;
 617         struct list_head        *tmp;
 618
 619         /* The requests in committed_list always have smaller transnos than
 620          * the requests in replay_list */
 621         if (!list_empty(&imp->imp_committed_list)) {
 622                 tmp = imp->imp_committed_list.next;
 623                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 624                 *transno = req->rq_transno;
 625                 if (req->rq_transno == 0) {
 626                         DEBUG_REQ(D_ERROR, req, "zero transno in committed_list");
 627                         LBUG();
 628                 }
 629                 return 1;
 630         }
 631         if (!list_empty(&imp->imp_replay_list)) {
 632                 tmp = imp->imp_replay_list.next;
 633                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 634                 *transno = req->rq_transno;
 635                 if (req->rq_transno == 0) {
 636                         DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
 637                         LBUG();
 638                 }
 639                 return 1;
 640         }
 641         return 0;
 642 }
 643
 644 /**
 645  * Attempt to (re)connect import \a imp. This includes all preparations,
 646  * initializing CONNECT RPC request and passing it to ptlrpcd for
 647  * actual sending.
 648  * Returns 0 on success or error code.
 649  */
 650 int ptlrpc_connect_import(struct obd_import *imp)
 651 {
 652         struct obd_device *obd = imp->imp_obd;
 653         int initial_connect = 0;
 654         int set_transno = 0;
 655         __u64 committed_before_reconnect = 0;
 656         struct ptlrpc_request *request;
 657         struct obd_connect_data ocd;
 658         char *bufs[] = { NULL,
 659                          obd2cli_tgt(imp->imp_obd),
 660                          obd->obd_uuid.uuid,
 661                          (char *)&imp->imp_dlm_handle,
 662                          (char *)&ocd,
 663                          NULL };
 664         struct ptlrpc_connect_async_args *aa;
 665         int rc;
 666         ENTRY;
 667
 668         spin_lock(&imp->imp_lock);
 669         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 670                 spin_unlock(&imp->imp_lock);
 671                 CERROR("can't connect to a closed import\n");
 672                 RETURN(-EINVAL);
 673         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 674                 spin_unlock(&imp->imp_lock);
 675                 CERROR("already connected\n");
 676                 RETURN(0);
 677         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING ||
 678                    imp->imp_state == LUSTRE_IMP_EVICTED ||
 679                    imp->imp_connected) {
 680                 spin_unlock(&imp->imp_lock);
 681                 CERROR("already connecting\n");
 682                 RETURN(-EALREADY);
 683         }
 684
 685         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 686
 687         imp->imp_conn_cnt++;
 688         imp->imp_resend_replay = 0;
 689
 690         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 691                 initial_connect = 1;
 692         else
 693                 committed_before_reconnect = imp->imp_peer_committed_transno;
 694
 695         set_transno = ptlrpc_first_transno(imp,
 696                                            &imp->imp_connect_data.ocd_transno);
 697         spin_unlock(&imp->imp_lock);
 698
 699         rc = import_select_connection(imp);
 700         if (rc)
 701                 GOTO(out, rc);
 702
 703         rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
 704         if (rc)
 705                 GOTO(out, rc);
 706
 707         /* Reset connect flags to the originally requested flags, in case
 708          * the server is updated on-the-fly we will get the new features. */
 709         ocd = imp->imp_connect_data;
 710         ocd.ocd_connect_flags = imp->imp_connect_flags_orig;
 711         ocd.ocd_connect_flags2 = imp->imp_connect_flags2_orig;
 712         /* Reset ocd_version each time so the server knows the exact versions */
 713         ocd.ocd_version = LUSTRE_VERSION_CODE;
 714         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 715         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 716
 717         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 718                            &obd->obd_uuid, &ocd, NULL);
 719         if (rc)
 720                 GOTO(out, rc);
 721
 722         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 723         if (request == NULL)
 724                 GOTO(out, rc = -ENOMEM);
 725
 726         /* get SELinux policy info if any */
 727         rc = sptlrpc_get_sepol(request);
 728         if (rc < 0) {
 729                 ptlrpc_request_free(request);
 730                 GOTO(out, rc);
 731         }
 732
 733         bufs[5] = request->rq_sepol;
 734
 735         req_capsule_set_size(&request->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
 736                              strlen(request->rq_sepol) ?
 737                              strlen(request->rq_sepol) + 1 : 0);
 738
 739         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 740                                       imp->imp_connect_op, bufs, NULL);
 741         if (rc) {
 742                 ptlrpc_request_free(request);
 743                 GOTO(out, rc);
 744         }
 745
 746         /* Report the rpc service time to the server so that it knows how long
 747          * to wait for clients to join recovery */
 748         lustre_msg_set_service_time(request->rq_reqmsg,
 749                                     at_timeout2est(request->rq_timeout));
 750
 751         /* The amount of time we give the server to process the connect req.
 752          * import_select_connection will increase the net latency on
 753          * repeated reconnect attempts to cover slow networks.
 754          * We override/ignore the server rpc completion estimate here,
 755          * which may be large if this is a reconnect attempt */
 756         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 757         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 758
 759         request->rq_no_resend = request->rq_no_delay = 1;
 760         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 761         /* Allow a slightly larger reply for future growth compatibility */
 762         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 763                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 764         ptlrpc_request_set_replen(request);
 765         request->rq_interpret_reply = ptlrpc_connect_interpret;
 766
 767         CLASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
 768         aa = ptlrpc_req_async_args(request);
 769         memset(aa, 0, sizeof *aa);
 770
 771         aa->pcaa_peer_committed = committed_before_reconnect;
 772         aa->pcaa_initial_connect = initial_connect;
 773
 774         if (aa->pcaa_initial_connect) {
 775                 spin_lock(&imp->imp_lock);
 776                 imp->imp_replayable = 1;
 777                 spin_unlock(&imp->imp_lock);
 778                 lustre_msg_add_op_flags(request->rq_reqmsg,
 779                                         MSG_CONNECT_INITIAL);
 780         }
 781
 782         if (set_transno)
 783                 lustre_msg_add_op_flags(request->rq_reqmsg,
 784                                         MSG_CONNECT_TRANSNO);
 785
 786         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %ld)",
 787                   request->rq_timeout);
 788         ptlrpcd_add_req(request);
 789         rc = 0;
 790 out:
 791         if (rc != 0)
 792                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 793
 794         RETURN(rc);
 795 }
 796 EXPORT_SYMBOL(ptlrpc_connect_import);
 797
 798 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 799 {
 800         int force_verify;
 801
 802         spin_lock(&imp->imp_lock);
 803         force_verify = imp->imp_force_verify != 0;
 804         spin_unlock(&imp->imp_lock);
 805
 806         if (force_verify)
 807                 ptlrpc_pinger_wake_up();
 808 }
 809
 810 static int ptlrpc_busy_reconnect(int rc)
 811 {
 812         return (rc == -EBUSY) || (rc == -EAGAIN);
 813 }
 814
 815 static int ptlrpc_connect_set_flags(struct obd_import *imp,
 816                                     struct obd_connect_data *ocd,
 817                                     __u64 old_connect_flags,
 818                                     struct obd_export *exp, int init_connect)
 819 {
 820         static bool warned;
 821         struct client_obd *cli = &imp->imp_obd->u.cli;
 822
 823         spin_lock(&imp->imp_lock);
 824         list_del(&imp->imp_conn_current->oic_item);
 825         list_add(&imp->imp_conn_current->oic_item,
 826                  &imp->imp_conn_list);
 827         imp->imp_last_success_conn =
 828                 imp->imp_conn_current->oic_last_attempt;
 829
 830         spin_unlock(&imp->imp_lock);
 831
 832         if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 833             (ocd->ocd_version > LUSTRE_VERSION_CODE +
 834                                 LUSTRE_VERSION_OFFSET_WARN ||
 835              ocd->ocd_version < LUSTRE_VERSION_CODE -
 836                                 LUSTRE_VERSION_OFFSET_WARN)) {
 837                 /* Sigh, some compilers do not like #ifdef in the middle
 838                    of macro arguments */
 839                 const char *older = "older than client. "
 840                                     "Consider upgrading server";
 841                 const char *newer = "newer than client. "
 842                                     "Consider upgrading client";
 843
 844                 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 845                               "is much %s (%s)\n",
 846                               obd2cli_tgt(imp->imp_obd),
 847                               OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 848                               OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 849                               OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 850                               OBD_OCD_VERSION_FIX(ocd->ocd_version),
 851                               ocd->ocd_version > LUSTRE_VERSION_CODE ?
 852                               newer : older, LUSTRE_VERSION_STRING);
 853                 warned = true;
 854         }
 855
 856         if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
 857                 /* We sent to the server ocd_cksum_types with bits set
 858                  * for algorithms we understand. The server masked off
 859                  * the checksum types it doesn't support */
 860                 if ((ocd->ocd_cksum_types &
 861                      obd_cksum_types_supported_client()) == 0) {
 862                         LCONSOLE_ERROR("The negotiation of the checksum "
 863                                        "alogrithm to use with server %s "
 864                                        "failed (%x/%x)\n",
 865                                        obd2cli_tgt(imp->imp_obd),
 866                                        ocd->ocd_cksum_types,
 867                                        obd_cksum_types_supported_client());
 868                         return -EPROTO;
 869                 } else {
 870                         cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
 871                 }
 872         } else {
 873                 /* The server does not support OBD_CONNECT_CKSUM.
 874                  * Enforce ADLER for backward compatibility*/
 875                 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
 876         }
 877         cli->cl_cksum_type = obd_cksum_type_select(imp->imp_obd->obd_name,
 878                                                    cli->cl_supp_cksum_types);
 879
 880         if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
 881                 cli->cl_max_pages_per_rpc =
 882                         min(ocd->ocd_brw_size >> PAGE_SHIFT,
 883                             cli->cl_max_pages_per_rpc);
 884         else if (imp->imp_connect_op == MDS_CONNECT ||
 885                  imp->imp_connect_op == MGS_CONNECT)
 886                 cli->cl_max_pages_per_rpc = 1;
 887
 888         LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 889                 (cli->cl_max_pages_per_rpc > 0));
 890
 891         client_adjust_max_dirty(cli);
 892
 893         /* Update client max modify RPCs in flight with value returned
 894          * by the server */
 895         if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
 896                 cli->cl_max_mod_rpcs_in_flight = min(
 897                                         cli->cl_max_mod_rpcs_in_flight,
 898                                         ocd->ocd_maxmodrpcs);
 899         else
 900                 cli->cl_max_mod_rpcs_in_flight = 1;
 901
 902         /* Reset ns_connect_flags only for initial connect. It might be
 903          * changed in while using FS and if we reset it in reconnect
 904          * this leads to losing user settings done before such as
 905          * disable lru_resize, etc. */
 906         if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
 907                 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
 908                 __u64 changed_flags;
 909
 910                 changed_flags =
 911                         ns->ns_connect_flags ^ ns->ns_orig_connect_flags;
 912                 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
 913                              "flags: %#llx\n", imp->imp_obd->obd_name,
 914                              ocd->ocd_connect_flags);
 915                 ns->ns_connect_flags = (ns->ns_connect_flags & changed_flags) |
 916                                       (ocd->ocd_connect_flags & ~changed_flags);
 917                 ns->ns_orig_connect_flags = ocd->ocd_connect_flags;
 918         }
 919
 920         if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
 921                 /* We need a per-message support flag, because
 922                  * a. we don't know if the incoming connect reply
 923                  *    supports AT or not (in reply_in_callback)
 924                  *    until we unpack it.
 925                  * b. failovered server means export and flags are gone
 926                  *    (in ptlrpc_send_reply).
 927                  *    Can only be set when we know AT is supported at
 928                  *    both ends */
 929                 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 930         else
 931                 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 932
 933         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 934
 935         return 0;
 936 }
 937
 938 /**
 939  * Add all replay requests back to unreplied list before start replay,
 940  * so that we can make sure the known replied XID is always increased
 941  * only even if when replaying requests.
 942  */
 943 static void ptlrpc_prepare_replay(struct obd_import *imp)
 944 {
 945         struct ptlrpc_request *req;
 946
 947         if (imp->imp_state != LUSTRE_IMP_REPLAY ||
 948             imp->imp_resend_replay)
 949                 return;
 950
 951         /* If the server was restart during repaly, the requests may
 952          * have been added to the unreplied list in former replay. */
 953         spin_lock(&imp->imp_lock);
 954
 955         list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
 956                 if (list_empty(&req->rq_unreplied_list))
 957                         ptlrpc_add_unreplied(req);
 958         }
 959
 960         list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
 961                 if (list_empty(&req->rq_unreplied_list))
 962                         ptlrpc_add_unreplied(req);
 963         }
 964
 965         imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
 966         spin_unlock(&imp->imp_lock);
 967 }
 968
 969 /**
 970  * interpret_reply callback for connect RPCs.
 971  * Looks into returned status of connect operation and decides
 972  * what to do with the import - i.e enter recovery, promote it to
 973  * full state for normal operations of disconnect it due to an error.
 974  */
 975 static int ptlrpc_connect_interpret(const struct lu_env *env,
 976                                     struct ptlrpc_request *request,
 977                                     void *data, int rc)
 978 {
 979         struct ptlrpc_connect_async_args *aa = data;
 980         struct obd_import *imp = request->rq_import;
 981         struct lustre_handle old_hdl;
 982         __u64 old_connect_flags;
 983         int msg_flags;
 984         struct obd_connect_data *ocd;
 985         struct obd_export *exp = NULL;
 986         int ret;
 987         ENTRY;
 988
 989         spin_lock(&imp->imp_lock);
 990         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 991                 imp->imp_connect_tried = 1;
 992                 spin_unlock(&imp->imp_lock);
 993                 RETURN(0);
 994         }
 995
 996         imp->imp_connect_error = rc;
 997         if (rc) {
 998                 struct ptlrpc_request *free_req;
 999                 struct ptlrpc_request *tmp;
1000
1001                 /* abort all delayed requests initiated connection */
1002                 list_for_each_entry_safe(free_req, tmp, &imp->imp_delayed_list,
1003                                          rq_list) {
1004                         spin_lock(&free_req->rq_lock);
1005                         if (free_req->rq_no_resend) {
1006                                 free_req->rq_err = 1;
1007                                 free_req->rq_status = -EIO;
1008                                 ptlrpc_client_wake_req(free_req);
1009                         }
1010                         spin_unlock(&free_req->rq_lock);
1011                 }
1012
1013                 /* if this reconnect to busy export - not need select new target
1014                  * for connecting*/
1015                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
1016                 spin_unlock(&imp->imp_lock);
1017                 ptlrpc_maybe_ping_import_soon(imp);
1018                 GOTO(out, rc);
1019         }
1020
1021         /* LU-7558: indicate that we are interpretting connect reply,
1022          * pltrpc_connect_import() will not try to reconnect until
1023          * interpret will finish. */
1024         imp->imp_connected = 1;
1025         spin_unlock(&imp->imp_lock);
1026
1027         LASSERT(imp->imp_conn_current);
1028
1029         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1030
1031         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1032                                    RCL_SERVER);
1033         /* server replied obd_connect_data is always bigger */
1034         ocd = req_capsule_server_sized_get(&request->rq_pill,
1035                                            &RMF_CONNECT_DATA, ret);
1036
1037         if (ocd == NULL) {
1038                 CERROR("%s: no connect data from server\n",
1039                        imp->imp_obd->obd_name);
1040                 rc = -EPROTO;
1041                 GOTO(out, rc);
1042         }
1043
1044         spin_lock(&imp->imp_lock);
1045
1046         /* All imports are pingable */
1047         imp->imp_pingable = 1;
1048         imp->imp_force_reconnect = 0;
1049         imp->imp_force_verify = 0;
1050
1051         imp->imp_connect_data = *ocd;
1052
1053         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1054                imp->imp_obd->obd_name, ocd->ocd_instance);
1055         exp = class_conn2export(&imp->imp_dlm_handle);
1056
1057         spin_unlock(&imp->imp_lock);
1058
1059         if (!exp) {
1060                 /* This could happen if export is cleaned during the
1061                    connect attempt */
1062                 CERROR("%s: missing export after connect\n",
1063                        imp->imp_obd->obd_name);
1064                 GOTO(out, rc = -ENODEV);
1065         }
1066
1067         /* check that server granted subset of flags we asked for. */
1068         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1069             ocd->ocd_connect_flags) {
1070                 CERROR("%s: Server didn't grant requested subset of flags: "
1071                        "asked=%#llx granted=%#llx\n",
1072                        imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
1073                        ocd->ocd_connect_flags);
1074                 GOTO(out, rc = -EPROTO);
1075         }
1076
1077         if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) !=
1078             ocd->ocd_connect_flags2) {
1079                 CERROR("%s: Server didn't grant requested subset of flags2: "
1080                        "asked=%#llx granted=%#llx\n",
1081                        imp->imp_obd->obd_name, imp->imp_connect_flags2_orig,
1082                        ocd->ocd_connect_flags2);
1083                 GOTO(out, rc = -EPROTO);
1084         }
1085
1086         if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1087             (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1088             (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1089             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1090                 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1091                 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1092                 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1093
1094                 /* We do not support the MDT-MDT interoperations with
1095                  * different version MDT because of protocol changes. */
1096                 if (unlikely(major != LUSTRE_MAJOR ||
1097                              minor != LUSTRE_MINOR ||
1098                              abs(patch - LUSTRE_PATCH) > 3)) {
1099                         LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1100                                       "connection to different version MDT "
1101                                       "(%d.%d.%d.%d) %s\n",
1102                                       imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1103                                       LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1104                                       major, minor, patch,
1105                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1106                                       imp->imp_connection->c_remote_uuid.uuid);
1107
1108                         GOTO(out, rc = -EPROTO);
1109                 }
1110         }
1111
1112         old_connect_flags = exp_connect_flags(exp);
1113         exp->exp_connect_data = *ocd;
1114         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1115
1116         /* The net statistics after (re-)connect is not valid anymore,
1117          * because may reflect other routing, etc. */
1118         at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1119         ptlrpc_at_adj_net_latency(request,
1120                         lustre_msg_get_service_time(request->rq_repmsg));
1121
1122         /* Import flags should be updated before waking import at FULL state */
1123         rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1124                                       aa->pcaa_initial_connect);
1125         class_export_put(exp);
1126         exp = NULL;
1127
1128         if (rc != 0)
1129                 GOTO(out, rc);
1130
1131         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1132
1133         if (aa->pcaa_initial_connect) {
1134                 spin_lock(&imp->imp_lock);
1135                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1136                         imp->imp_replayable = 1;
1137                         spin_unlock(&imp->imp_lock);
1138                         CDEBUG(D_HA, "connected to replayable target: %s\n",
1139                                obd2cli_tgt(imp->imp_obd));
1140                 } else {
1141                         imp->imp_replayable = 0;
1142                         spin_unlock(&imp->imp_lock);
1143                 }
1144
1145                 /* if applies, adjust the imp->imp_msg_magic here
1146                  * according to reply flags */
1147
1148                 imp->imp_remote_handle =
1149                                 *lustre_msg_get_handle(request->rq_repmsg);
1150
1151                 /* Initial connects are allowed for clients with non-random
1152                  * uuids when servers are in recovery.  Simply signal the
1153                  * servers replay is complete and wait in REPLAY_WAIT. */
1154                 if (msg_flags & MSG_CONNECT_RECOVERING) {
1155                         CDEBUG(D_HA, "connect to %s during recovery\n",
1156                                obd2cli_tgt(imp->imp_obd));
1157                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1158                 } else {
1159                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1160                         ptlrpc_activate_import(imp);
1161                 }
1162
1163                 GOTO(finish, rc = 0);
1164         }
1165
1166         /* Determine what recovery state to move the import to. */
1167         if (MSG_CONNECT_RECONNECT & msg_flags) {
1168                 memset(&old_hdl, 0, sizeof(old_hdl));
1169                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1170                             sizeof (old_hdl))) {
1171                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1172                                       "bad handle %#llx\n",
1173                                       obd2cli_tgt(imp->imp_obd),
1174                                       imp->imp_connection->c_remote_uuid.uuid,
1175                                       imp->imp_dlm_handle.cookie);
1176                         GOTO(out, rc = -ENOTCONN);
1177                 }
1178
1179                 if (memcmp(&imp->imp_remote_handle,
1180                            lustre_msg_get_handle(request->rq_repmsg),
1181                            sizeof(imp->imp_remote_handle))) {
1182                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
1183                                 D_HA : D_WARNING;
1184
1185                         /* Bug 16611/14775: if server handle have changed,
1186                          * that means some sort of disconnection happened.
1187                          * If the server is not in recovery, that also means it
1188                          * already erased all of our state because of previous
1189                          * eviction. If it is in recovery - we are safe to
1190                          * participate since we can reestablish all of our state
1191                          * with server again */
1192                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1193                                 CDEBUG(level,"%s@%s changed server handle from "
1194                                        "%#llx to %#llx"
1195                                        " but is still in recovery\n",
1196                                        obd2cli_tgt(imp->imp_obd),
1197                                        imp->imp_connection->c_remote_uuid.uuid,
1198                                        imp->imp_remote_handle.cookie,
1199                                        lustre_msg_get_handle(
1200                                        request->rq_repmsg)->cookie);
1201                         } else {
1202                                 LCONSOLE_WARN("Evicted from %s (at %s) "
1203                                               "after server handle changed from "
1204                                               "%#llx to %#llx\n",
1205                                               obd2cli_tgt(imp->imp_obd),
1206                                               imp->imp_connection-> \
1207                                               c_remote_uuid.uuid,
1208                                               imp->imp_remote_handle.cookie,
1209                                               lustre_msg_get_handle(
1210                                               request->rq_repmsg)->cookie);
1211                         }
1212
1213
1214                         imp->imp_remote_handle =
1215                                      *lustre_msg_get_handle(request->rq_repmsg);
1216
1217                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1218                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1219                                 GOTO(finish, rc = 0);
1220                         }
1221
1222                 } else {
1223                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1224                                obd2cli_tgt(imp->imp_obd),
1225                                imp->imp_connection->c_remote_uuid.uuid);
1226                 }
1227
1228                 if (imp->imp_invalid) {
1229                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1230                                "marking evicted\n", imp->imp_obd->obd_name);
1231                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1232                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1233                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1234                                imp->imp_obd->obd_name,
1235                                obd2cli_tgt(imp->imp_obd));
1236
1237                         spin_lock(&imp->imp_lock);
1238                         imp->imp_resend_replay = 1;
1239                         spin_unlock(&imp->imp_lock);
1240
1241                         IMPORT_SET_STATE(imp, imp->imp_replay_state);
1242                 } else {
1243                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1244                 }
1245         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1246                 LASSERT(imp->imp_replayable);
1247                 imp->imp_remote_handle =
1248                                 *lustre_msg_get_handle(request->rq_repmsg);
1249                 imp->imp_last_replay_transno = 0;
1250                 imp->imp_replay_cursor = &imp->imp_committed_list;
1251                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
1252         } else if ((ocd->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0 &&
1253                    !imp->imp_invalid) {
1254
1255                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
1256                 DEBUG_REQ(D_HA, request, "%s: lwp recover",
1257                           imp->imp_obd->obd_name);
1258                 imp->imp_remote_handle =
1259                         *lustre_msg_get_handle(request->rq_repmsg);
1260                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1261         } else {
1262                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
1263                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
1264                 imp->imp_remote_handle =
1265                                 *lustre_msg_get_handle(request->rq_repmsg);
1266                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1267         }
1268
1269         /* Sanity checks for a reconnected import. */
1270         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
1271                 CERROR("imp_replayable flag does not match server "
1272                        "after reconnect. We should LBUG right here.\n");
1273         }
1274
1275         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1276             lustre_msg_get_last_committed(request->rq_repmsg) <
1277             aa->pcaa_peer_committed) {
1278                 CERROR("%s went back in time (transno %lld"
1279                        " was previously committed, server now claims %lld"
1280                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
1281                        "id=9646\n",
1282                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1283                        lustre_msg_get_last_committed(request->rq_repmsg));
1284         }
1285
1286 finish:
1287         ptlrpc_prepare_replay(imp);
1288         rc = ptlrpc_import_recovery_state_machine(imp);
1289         if (rc == -ENOTCONN) {
1290                 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1291                        "invalidating and reconnecting\n",
1292                        obd2cli_tgt(imp->imp_obd),
1293                        imp->imp_connection->c_remote_uuid.uuid);
1294                 ptlrpc_connect_import(imp);
1295                 spin_lock(&imp->imp_lock);
1296                 imp->imp_connected = 0;
1297                 imp->imp_connect_tried = 1;
1298                 spin_unlock(&imp->imp_lock);
1299                 RETURN(0);
1300         }
1301
1302 out:
1303         spin_lock(&imp->imp_lock);
1304         imp->imp_connected = 0;
1305         imp->imp_connect_tried = 1;
1306         spin_unlock(&imp->imp_lock);
1307
1308         if (exp != NULL)
1309                 class_export_put(exp);
1310
1311         if (rc != 0) {
1312                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1313                 if (rc == -EACCES) {
1314                         /*
1315                          * Give up trying to reconnect
1316                          * EACCES means client has no permission for connection
1317                          */
1318                         imp->imp_obd->obd_no_recov = 1;
1319                         ptlrpc_deactivate_import(imp);
1320                 }
1321
1322                 if (rc == -EPROTO) {
1323                         struct obd_connect_data *ocd;
1324
1325                         /* reply message might not be ready */
1326                         if (request->rq_repmsg == NULL)
1327                                 RETURN(-EPROTO);
1328
1329                         ocd = req_capsule_server_get(&request->rq_pill,
1330                                                      &RMF_CONNECT_DATA);
1331                         /* Servers are not supposed to refuse connections from
1332                          * clients based on version, only connection feature
1333                          * flags.  We should never see this from llite, but it
1334                          * may be useful for debugging in the future. */
1335                         if (ocd &&
1336                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1337                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1338                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1339                                         "(%d.%d.%d.%d)"
1340                                         " refused connection from this client "
1341                                         "with an incompatible version (%s).  "
1342                                         "Client must be recompiled\n",
1343                                         obd2cli_tgt(imp->imp_obd),
1344                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1345                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1346                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1347                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1348                                         LUSTRE_VERSION_STRING);
1349                                 ptlrpc_deactivate_import(imp);
1350                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1351                         }
1352                         RETURN(-EPROTO);
1353                 }
1354
1355                 ptlrpc_maybe_ping_import_soon(imp);
1356
1357                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1358                        obd2cli_tgt(imp->imp_obd),
1359                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1360         }
1361
1362         wake_up_all(&imp->imp_recovery_waitq);
1363         RETURN(rc);
1364 }
1365
1366 /**
1367  * interpret callback for "completed replay" RPCs.
1368  * \see signal_completed_replay
1369  */
1370 static int completed_replay_interpret(const struct lu_env *env,
1371                                       struct ptlrpc_request *req,
1372                                       void *args, int rc)
1373 {
1374         ENTRY;
1375         atomic_dec(&req->rq_import->imp_replay_inflight);
1376         if (req->rq_status == 0 && !req->rq_import->imp_vbr_failed) {
1377                 ptlrpc_import_recovery_state_machine(req->rq_import);
1378         } else {
1379                 if (req->rq_import->imp_vbr_failed) {
1380                         CDEBUG(D_WARNING,
1381                                "%s: version recovery fails, reconnecting\n",
1382                                req->rq_import->imp_obd->obd_name);
1383                 } else {
1384                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1385                                      "reconnecting\n",
1386                                req->rq_import->imp_obd->obd_name,
1387                                req->rq_status);
1388                 }
1389                 ptlrpc_connect_import(req->rq_import);
1390         }
1391
1392         RETURN(0);
1393 }
1394
1395 /**
1396  * Let server know that we have no requests to replay anymore.
1397  * Achieved by just sending a PING request
1398  */
1399 static int signal_completed_replay(struct obd_import *imp)
1400 {
1401         struct ptlrpc_request *req;
1402         ENTRY;
1403
1404         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1405                 RETURN(0);
1406
1407         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1408         atomic_inc(&imp->imp_replay_inflight);
1409
1410         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1411                                         OBD_PING);
1412         if (req == NULL) {
1413                 atomic_dec(&imp->imp_replay_inflight);
1414                 RETURN(-ENOMEM);
1415         }
1416
1417         ptlrpc_request_set_replen(req);
1418         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1419         lustre_msg_add_flags(req->rq_reqmsg,
1420                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1421         if (AT_OFF)
1422                 req->rq_timeout *= 3;
1423         req->rq_interpret_reply = completed_replay_interpret;
1424
1425         ptlrpcd_add_req(req);
1426         RETURN(0);
1427 }
1428
1429 /**
1430  * In kernel code all import invalidation happens in its own
1431  * separate thread, so that whatever application happened to encounter
1432  * a problem could still be killed or otherwise continue
1433  */
1434 static int ptlrpc_invalidate_import_thread(void *data)
1435 {
1436         struct obd_import *imp = data;
1437
1438         ENTRY;
1439
1440         unshare_fs_struct();
1441
1442         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1443                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1444                imp->imp_connection->c_remote_uuid.uuid);
1445
1446         ptlrpc_invalidate_import(imp);
1447
1448         if (obd_dump_on_eviction) {
1449                 CERROR("dump the log upon eviction\n");
1450                 libcfs_debug_dumplog();
1451         }
1452
1453         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1454         ptlrpc_import_recovery_state_machine(imp);
1455
1456         class_import_put(imp);
1457         RETURN(0);
1458 }
1459
1460 /**
1461  * This is the state machine for client-side recovery on import.
1462  *
1463  * Typicaly we have two possibly paths. If we came to server and it is not
1464  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1465  * state and reconnect from scratch.
1466  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1467  * We go through our list of requests to replay and send them to server one by
1468  * one.
1469  * After sending all request from the list we change import state to
1470  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1471  * and also all the locks we don't yet have and wait for server to grant us.
1472  * After that we send a special "replay completed" request and change import
1473  * state to IMP_REPLAY_WAIT.
1474  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1475  * state and resend all requests from sending list.
1476  * After that we promote import to FULL state and send all delayed requests
1477  * and import is fully operational after that.
1478  *
1479  */
1480 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1481 {
1482         int rc = 0;
1483         int inflight;
1484         char *target_start;
1485         int target_len;
1486
1487         ENTRY;
1488         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1489                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1490                           &target_start, &target_len);
1491                 /* Don't care about MGC eviction */
1492                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1493                            LUSTRE_MGC_NAME) != 0) {
1494                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1495                                            "by %.*s; in progress operations "
1496                                            "using this service will fail.\n",
1497                                            imp->imp_obd->obd_name, target_len,
1498                                            target_start);
1499                         LASSERTF(!obd_lbug_on_eviction, "LBUG upon eviction");
1500                 }
1501                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1502                        obd2cli_tgt(imp->imp_obd),
1503                        imp->imp_connection->c_remote_uuid.uuid);
1504                 /* reset vbr_failed flag upon eviction */
1505                 spin_lock(&imp->imp_lock);
1506                 imp->imp_vbr_failed = 0;
1507                 spin_unlock(&imp->imp_lock);
1508
1509                 {
1510                 struct task_struct *task;
1511                 /* bug 17802:  XXX client_disconnect_export vs connect request
1512                  * race. if client is evicted at this time then we start
1513                  * invalidate thread without reference to import and import can
1514                  * be freed at same time. */
1515                 class_import_get(imp);
1516                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1517                                      "ll_imp_inval");
1518                 if (IS_ERR(task)) {
1519                         class_import_put(imp);
1520                         CERROR("error starting invalidate thread: %d\n", rc);
1521                         rc = PTR_ERR(task);
1522                 } else {
1523                         rc = 0;
1524                 }
1525                 RETURN(rc);
1526                 }
1527         }
1528
1529         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1530                 CDEBUG(D_HA, "replay requested by %s\n",
1531                        obd2cli_tgt(imp->imp_obd));
1532                 rc = ptlrpc_replay_next(imp, &inflight);
1533                 if (inflight == 0 &&
1534                     atomic_read(&imp->imp_replay_inflight) == 0) {
1535                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1536                         rc = ldlm_replay_locks(imp);
1537                         if (rc)
1538                                 GOTO(out, rc);
1539                 }
1540                 rc = 0;
1541         }
1542
1543         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1544                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1545                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1546                         rc = signal_completed_replay(imp);
1547                         if (rc)
1548                                 GOTO(out, rc);
1549                 }
1550         }
1551
1552         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1553                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1554                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1555                 }
1556         }
1557
1558         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1559                 struct ptlrpc_connection *conn = imp->imp_connection;
1560
1561                 rc = ptlrpc_resend(imp);
1562                 if (rc)
1563                         GOTO(out, rc);
1564                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1565                 ptlrpc_activate_import(imp);
1566
1567                 LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n",
1568                               imp->imp_obd->obd_name,
1569                               obd_uuid2str(&conn->c_remote_uuid),
1570                               obd_import_nid2str(imp));
1571         }
1572
1573         if (imp->imp_state == LUSTRE_IMP_FULL) {
1574                 wake_up_all(&imp->imp_recovery_waitq);
1575                 ptlrpc_wake_delayed(imp);
1576         }
1577
1578 out:
1579         RETURN(rc);
1580 }
1581
1582 static struct ptlrpc_request *ptlrpc_disconnect_prep_req(struct obd_import *imp)
1583 {
1584         struct ptlrpc_request *req;
1585         int rq_opc, rc = 0;
1586         ENTRY;
1587
1588         switch (imp->imp_connect_op) {
1589         case OST_CONNECT:
1590                 rq_opc = OST_DISCONNECT;
1591                 break;
1592         case MDS_CONNECT:
1593                 rq_opc = MDS_DISCONNECT;
1594                 break;
1595         case MGS_CONNECT:
1596                 rq_opc = MGS_DISCONNECT;
1597                 break;
1598         default:
1599                 rc = -EINVAL;
1600                 CERROR("%s: don't know how to disconnect from %s "
1601                        "(connect_op %d): rc = %d\n",
1602                        imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1603                        imp->imp_connect_op, rc);
1604                 RETURN(ERR_PTR(rc));
1605         }
1606
1607         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1608                                         LUSTRE_OBD_VERSION, rq_opc);
1609         if (req == NULL)
1610                 RETURN(NULL);
1611
1612         /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1613          * it fails.  We can get through the above with a down server
1614          * if the client doesn't know the server is gone yet. */
1615         req->rq_no_resend = 1;
1616
1617         /* We want client umounts to happen quickly, no matter the
1618            server state... */
1619         req->rq_timeout = min_t(int, req->rq_timeout,
1620                                 INITIAL_CONNECT_TIMEOUT);
1621
1622         IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1623         req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1624         ptlrpc_request_set_replen(req);
1625
1626         RETURN(req);
1627 }
1628
1629 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1630 {
1631         struct ptlrpc_request *req;
1632         int rc = 0;
1633         ENTRY;
1634
1635         if (imp->imp_obd->obd_force)
1636                 GOTO(set_state, rc);
1637
1638         /* probably the import has been disconnected already being idle */
1639         spin_lock(&imp->imp_lock);
1640         if (imp->imp_state == LUSTRE_IMP_IDLE)
1641                 GOTO(out, rc);
1642         spin_unlock(&imp->imp_lock);
1643
1644         if (ptlrpc_import_in_recovery(imp)) {
1645                 struct l_wait_info lwi;
1646                 long timeout_jiffies;
1647                 time64_t timeout;
1648
1649                 if (AT_OFF) {
1650                         if (imp->imp_server_timeout)
1651                                 timeout = obd_timeout >> 1;
1652                         else
1653                                 timeout = obd_timeout;
1654                 } else {
1655                         u32 req_portal;
1656                         int idx;
1657
1658                         req_portal = imp->imp_client->cli_request_portal;
1659                         idx = import_at_get_index(imp, req_portal);
1660                         timeout = at_get(&imp->imp_at.iat_service_estimate[idx]);
1661                 }
1662
1663                 timeout_jiffies = cfs_time_seconds(timeout);
1664                 lwi = LWI_TIMEOUT_INTR(max_t(long, timeout_jiffies, 1),
1665                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1666                 rc = l_wait_event(imp->imp_recovery_waitq,
1667                                   !ptlrpc_import_in_recovery(imp), &lwi);
1668
1669         }
1670
1671         spin_lock(&imp->imp_lock);
1672         if (imp->imp_state != LUSTRE_IMP_FULL)
1673                 GOTO(out, rc);
1674         spin_unlock(&imp->imp_lock);
1675
1676         req = ptlrpc_disconnect_prep_req(imp);
1677         if (IS_ERR(req))
1678                 GOTO(set_state, rc = PTR_ERR(req));
1679         rc = ptlrpc_queue_wait(req);
1680         ptlrpc_req_finished(req);
1681
1682 set_state:
1683         spin_lock(&imp->imp_lock);
1684 out:
1685         if (noclose)
1686                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1687         else
1688                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1689         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1690         spin_unlock(&imp->imp_lock);
1691
1692         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1693                 rc = 0;
1694         RETURN(rc);
1695 }
1696 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1697
1698 static void ptlrpc_reset_reqs_generation(struct obd_import *imp)
1699 {
1700         struct ptlrpc_request *old, *tmp;
1701
1702         /* tag all resendable requests generated before disconnection
1703          * notice this code is part of disconnect-at-idle path only */
1704         list_for_each_entry_safe(old, tmp, &imp->imp_delayed_list,
1705                         rq_list) {
1706                 spin_lock(&old->rq_lock);
1707                 if (old->rq_import_generation == imp->imp_generation - 1 &&
1708                     !old->rq_no_resend)
1709                         old->rq_import_generation = imp->imp_generation;
1710                 spin_unlock(&old->rq_lock);
1711         }
1712 }
1713
1714 static int ptlrpc_disconnect_idle_interpret(const struct lu_env *env,
1715                                             struct ptlrpc_request *req,
1716                                             void *args, int rc)
1717 {
1718         struct obd_import *imp = req->rq_import;
1719         int connect = 0;
1720
1721         DEBUG_REQ(D_HA, req, "inflight=%d, refcount=%d: rc = %d ",
1722                   atomic_read(&imp->imp_inflight),
1723                   atomic_read(&imp->imp_refcount), rc);
1724
1725         spin_lock(&imp->imp_lock);
1726         /* DISCONNECT reply can be late and another connection can just
1727          * be initiated. so we have to abort disconnection. */
1728         if (req->rq_import_generation == imp->imp_generation &&
1729             imp->imp_state != LUSTRE_IMP_CLOSED) {
1730                 LASSERTF(imp->imp_state == LUSTRE_IMP_CONNECTING,
1731                          "%s\n", ptlrpc_import_state_name(imp->imp_state));
1732                 imp->imp_state = LUSTRE_IMP_IDLE;
1733                 memset(&imp->imp_remote_handle, 0,
1734                        sizeof(imp->imp_remote_handle));
1735                 /* take our DISCONNECT into account */
1736                 if (atomic_read(&imp->imp_inflight) > 1) {
1737                         imp->imp_generation++;
1738                         imp->imp_initiated_at = imp->imp_generation;
1739                         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_NEW);
1740                         ptlrpc_reset_reqs_generation(imp);
1741                         connect = 1;
1742                 }
1743         }
1744         spin_unlock(&imp->imp_lock);
1745
1746         if (connect) {
1747                 rc = ptlrpc_connect_import(imp);
1748                 if (rc >= 0)
1749                         ptlrpc_pinger_add_import(imp);
1750         }
1751
1752         return 0;
1753 }
1754
1755 int ptlrpc_disconnect_and_idle_import(struct obd_import *imp)
1756 {
1757         struct ptlrpc_request *req;
1758         ENTRY;
1759
1760         if (imp->imp_obd->obd_force)
1761                 RETURN(0);
1762
1763         if (ptlrpc_import_in_recovery(imp))
1764                 RETURN(0);
1765
1766         spin_lock(&imp->imp_lock);
1767         if (imp->imp_state != LUSTRE_IMP_FULL) {
1768                 spin_unlock(&imp->imp_lock);
1769                 RETURN(0);
1770         }
1771         spin_unlock(&imp->imp_lock);
1772
1773         req = ptlrpc_disconnect_prep_req(imp);
1774         if (IS_ERR(req))
1775                 RETURN(PTR_ERR(req));
1776
1777         CDEBUG_LIMIT(imp->imp_idle_debug, "%s: disconnect after %llus idle\n",
1778                      imp->imp_obd->obd_name,
1779                      ktime_get_real_seconds() - imp->imp_last_reply_time);
1780         req->rq_interpret_reply = ptlrpc_disconnect_idle_interpret;
1781         ptlrpcd_add_req(req);
1782
1783         RETURN(0);
1784 }
1785 EXPORT_SYMBOL(ptlrpc_disconnect_and_idle_import);
1786
1787 void ptlrpc_cleanup_imp(struct obd_import *imp)
1788 {
1789         ENTRY;
1790
1791         spin_lock(&imp->imp_lock);
1792         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1793         imp->imp_generation++;
1794         spin_unlock(&imp->imp_lock);
1795         ptlrpc_abort_inflight(imp);
1796
1797         EXIT;
1798 }
1799
1800 /* Adaptive Timeout utils */
1801
1802 /* Update at_current with the specified value (bounded by at_min and at_max),
1803  * as well as the AT history "bins".
1804  *  - Bin into timeslices using AT_BINS bins.
1805  *  - This gives us a max of the last at_history seconds without the storage,
1806  *    but still smoothing out a return to normalcy from a slow response.
1807  *  - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1808  */
1809 int at_measured(struct adaptive_timeout *at, unsigned int val)
1810 {
1811         unsigned int old = at->at_current;
1812         time64_t now = ktime_get_real_seconds();
1813         long binlimit = max_t(long, at_history / AT_BINS, 1);
1814
1815         LASSERT(at);
1816         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1817                val, at, (long)(now - at->at_binstart), at->at_current,
1818                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1819
1820         if (val == 0)
1821                 /* 0's don't count, because we never want our timeout to
1822                    drop to 0, and because 0 could mean an error */
1823                 return 0;
1824
1825         spin_lock(&at->at_lock);
1826
1827         if (unlikely(at->at_binstart == 0)) {
1828                 /* Special case to remove default from history */
1829                 at->at_current = val;
1830                 at->at_worst_ever = val;
1831                 at->at_worst_time = now;
1832                 at->at_hist[0] = val;
1833                 at->at_binstart = now;
1834         } else if (now - at->at_binstart < binlimit ) {
1835                 /* in bin 0 */
1836                 at->at_hist[0] = max(val, at->at_hist[0]);
1837                 at->at_current = max(val, at->at_current);
1838         } else {
1839                 int i, shift;
1840                 unsigned int maxv = val;
1841
1842                 /* move bins over */
1843                 shift = (u32)(now - at->at_binstart) / binlimit;
1844                 LASSERT(shift > 0);
1845                 for(i = AT_BINS - 1; i >= 0; i--) {
1846                         if (i >= shift) {
1847                                 at->at_hist[i] = at->at_hist[i - shift];
1848                                 maxv = max(maxv, at->at_hist[i]);
1849                         } else {
1850                                 at->at_hist[i] = 0;
1851                         }
1852                 }
1853                 at->at_hist[0] = val;
1854                 at->at_current = maxv;
1855                 at->at_binstart += shift * binlimit;
1856         }
1857
1858         if (at->at_current > at->at_worst_ever) {
1859                 at->at_worst_ever = at->at_current;
1860                 at->at_worst_time = now;
1861         }
1862
1863         if (at->at_flags & AT_FLG_NOHIST)
1864                 /* Only keep last reported val; keeping the rest of the history
1865                    for proc only */
1866                 at->at_current = val;
1867
1868         if (at_max > 0)
1869                 at->at_current =  min(at->at_current, at_max);
1870         at->at_current =  max(at->at_current, at_min);
1871
1872         if (at->at_current != old)
1873                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1874                        "(val=%u) hist %u %u %u %u\n", at,
1875                        old, at->at_current, at->at_current - old, val,
1876                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1877                        at->at_hist[3]);
1878
1879         /* if we changed, report the old value */
1880         old = (at->at_current != old) ? old : 0;
1881
1882         spin_unlock(&at->at_lock);
1883         return old;
1884 }
1885
1886 /* Find the imp_at index for a given portal; assign if space available */
1887 int import_at_get_index(struct obd_import *imp, int portal)
1888 {
1889         struct imp_at *at = &imp->imp_at;
1890         int i;
1891
1892         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1893                 if (at->iat_portal[i] == portal)
1894                         return i;
1895                 if (at->iat_portal[i] == 0)
1896                         /* unused */
1897                         break;
1898         }
1899
1900         /* Not found in list, add it under a lock */
1901         spin_lock(&imp->imp_lock);
1902
1903         /* Check unused under lock */
1904         for (; i < IMP_AT_MAX_PORTALS; i++) {
1905                 if (at->iat_portal[i] == portal)
1906                         goto out;
1907                 if (at->iat_portal[i] == 0)
1908                         /* unused */
1909                         break;
1910         }
1911
1912         /* Not enough portals? */
1913         LASSERT(i < IMP_AT_MAX_PORTALS);
1914
1915         at->iat_portal[i] = portal;
1916 out:
1917         spin_unlock(&imp->imp_lock);
1918         return i;
1919 }