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 <obd_support.h>
  41 #include <lustre_ha.h>
  42 #include <lustre_net.h>
  43 #include <lustre_import.h>
  44 #include <lustre_export.h>
  45 #include <obd.h>
  46 #include <obd_cksum.h>
  47 #include <obd_class.h>
  48
  49 #include "ptlrpc_internal.h"
  50
  51 struct ptlrpc_connect_async_args {
  52          __u64 pcaa_peer_committed;
  53         int pcaa_initial_connect;
  54 };
  55
  56 /**
  57  * Updates import \a imp current state to provided \a state value
  58  * Helper function. Must be called under imp_lock.
  59  */
  60 static void __import_set_state(struct obd_import *imp,
  61                                enum lustre_imp_state state)
  62 {
  63         switch (state) {
  64         case LUSTRE_IMP_CLOSED:
  65         case LUSTRE_IMP_NEW:
  66         case LUSTRE_IMP_DISCON:
  67         case LUSTRE_IMP_CONNECTING:
  68                 break;
  69         case LUSTRE_IMP_REPLAY_WAIT:
  70                 imp->imp_replay_state = LUSTRE_IMP_REPLAY_LOCKS;
  71                 break;
  72         default:
  73                 imp->imp_replay_state = LUSTRE_IMP_REPLAY;
  74         }
  75         imp->imp_state = state;
  76         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
  77         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
  78                 ktime_get_real_seconds();
  79         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
  80                 IMP_STATE_HIST_LEN;
  81 }
  82
  83 /* A CLOSED import should remain so. */
  84 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  85 do {                                                                           \
  86         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  87                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  88                       imp, obd2cli_tgt(imp->imp_obd),                          \
  89                       ptlrpc_import_state_name(imp->imp_state),                \
  90                       ptlrpc_import_state_name(state));                        \
  91                __import_set_state(imp, state);                                 \
  92         }                                                                      \
  93 } while(0)
  94
  95 #define IMPORT_SET_STATE(imp, state)                                    \
  96 do {                                                                    \
  97         spin_lock(&imp->imp_lock);                                      \
  98         IMPORT_SET_STATE_NOLOCK(imp, state);                            \
  99         spin_unlock(&imp->imp_lock);                                    \
 100 } while(0)
 101
 102 void ptlrpc_import_enter_resend(struct obd_import *imp)
 103 {
 104         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 105 }
 106 EXPORT_SYMBOL(ptlrpc_import_enter_resend);
 107
 108
 109 static int ptlrpc_connect_interpret(const struct lu_env *env,
 110                                     struct ptlrpc_request *request,
 111                                     void * data, int rc);
 112 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
 113
 114 /* Only this function is allowed to change the import state when it is
 115  * CLOSED. I would rather refcount the import and free it after
 116  * disconnection like we do with exports. To do that, the client_obd
 117  * will need to save the peer info somewhere other than in the import,
 118  * though. */
 119 int ptlrpc_init_import(struct obd_import *imp)
 120 {
 121         spin_lock(&imp->imp_lock);
 122
 123         imp->imp_generation++;
 124         imp->imp_state =  LUSTRE_IMP_NEW;
 125
 126         spin_unlock(&imp->imp_lock);
 127
 128         return 0;
 129 }
 130 EXPORT_SYMBOL(ptlrpc_init_import);
 131
 132 #define UUID_STR "_UUID"
 133 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
 134 {
 135         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 136                 ? uuid : uuid + strlen(prefix);
 137
 138         *uuid_len = strlen(*uuid_start);
 139
 140         if (*uuid_len < strlen(UUID_STR))
 141                 return;
 142
 143         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 144                     UUID_STR, strlen(UUID_STR)))
 145                 *uuid_len -= strlen(UUID_STR);
 146 }
 147
 148 /**
 149  * Returns true if import was FULL, false if import was already not
 150  * connected.
 151  * @imp - import to be disconnected
 152  * @conn_cnt - connection count (epoch) of the request that timed out
 153  *             and caused the disconnection.  In some cases, multiple
 154  *             inflight requests can fail to a single target (e.g. OST
 155  *             bulk requests) and if one has already caused a reconnection
 156  *             (increasing the import->conn_cnt) the older failure should
 157  *             not also cause a reconnection.  If zero it forces a reconnect.
 158  */
 159 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 160 {
 161         int rc = 0;
 162
 163         spin_lock(&imp->imp_lock);
 164
 165         if (imp->imp_state == LUSTRE_IMP_FULL &&
 166             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 167                 char *target_start;
 168                 int   target_len;
 169
 170                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 171                           &target_start, &target_len);
 172
 173                 if (imp->imp_replayable) {
 174                         LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
 175                                "lost; in progress operations using this "
 176                                "service will wait for recovery to complete\n",
 177                                imp->imp_obd->obd_name, target_len, target_start,
 178                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 179                 } else {
 180                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
 181                                "%.*s (at %s) was lost; in progress "
 182                                "operations using this service will fail\n",
 183                                imp->imp_obd->obd_name,
 184                                target_len, target_start,
 185                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 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 unsigned int 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         CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
 296         LASSERT(imp->imp_invalid);
 297
 298         /* Wait forever until inflight == 0. We really can't do it another
 299          * way because in some cases we need to wait for very long reply
 300          * unlink. We can't do anything before that because there is really
 301          * no guarantee that some rdma transfer is not in progress right now. */
 302         do {
 303                 long timeout_jiffies;
 304
 305                 /* Calculate max timeout for waiting on rpcs to error
 306                  * out. Use obd_timeout if calculated value is smaller
 307                  * than it.
 308                  */
 309                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 310                         timeout = ptlrpc_inflight_timeout(imp);
 311                         timeout += timeout / 3;
 312
 313                         if (timeout == 0)
 314                                 timeout = obd_timeout;
 315                 } else {
 316                         /* decrease the interval to increase race condition */
 317                         timeout = 1;
 318                 }
 319
 320                 CDEBUG(D_RPCTRACE, "Sleeping %llds for inflight to error out\n",
 321                        timeout);
 322
 323                 /* Wait for all requests to error out and call completion
 324                  * callbacks. Cap it at obd_timeout -- these should all
 325                  * have been locally cancelled by ptlrpc_abort_inflight.
 326                  */
 327                 timeout_jiffies = max_t(long, cfs_time_seconds(timeout), 1);
 328                 lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
 329                                            (timeout > 1) ? cfs_time_seconds(1) :
 330                                                            cfs_time_seconds(1) / 2,
 331                                                            NULL, NULL);
 332                 rc = l_wait_event(imp->imp_recovery_waitq,
 333                                   (atomic_read(&imp->imp_inflight) == 0),
 334                                   &lwi);
 335                 if (rc) {
 336                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 337
 338                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 339                                cli_tgt, rc, atomic_read(&imp->imp_inflight));
 340
 341                         spin_lock(&imp->imp_lock);
 342                         if (atomic_read(&imp->imp_inflight) == 0) {
 343                                 int count = atomic_read(&imp->imp_unregistering);
 344
 345                                 /* We know that "unregistering" rpcs only can
 346                                  * survive in sending or delaying lists (they
 347                                  * maybe waiting for long reply unlink in
 348                                  * sluggish nets). Let's check this. If there
 349                                  * is no inflight and unregistering != 0, this
 350                                  * is bug. */
 351                                 LASSERTF(count == 0, "Some RPCs are still "
 352                                          "unregistering: %d\n", count);
 353
 354                                 /* Let's save one loop as soon as inflight have
 355                                  * dropped to zero. No new inflights possible at
 356                                  * this point. */
 357                                 rc = 0;
 358                         } else {
 359                                 list_for_each_safe(tmp, n,
 360                                                    &imp->imp_sending_list) {
 361                                         req = list_entry(tmp,
 362                                                          struct ptlrpc_request,
 363                                                          rq_list);
 364                                         DEBUG_REQ(D_ERROR, req,
 365                                                   "still on sending list");
 366                                 }
 367                                 list_for_each_safe(tmp, n,
 368                                                    &imp->imp_delayed_list) {
 369                                         req = list_entry(tmp,
 370                                                          struct ptlrpc_request,
 371                                                          rq_list);
 372                                         DEBUG_REQ(D_ERROR, req,
 373                                                   "still on delayed list");
 374                                 }
 375
 376                                 CERROR("%s: Unregistering RPCs found (%d). "
 377                                        "Network is sluggish? Waiting them "
 378                                        "to error out.\n", cli_tgt,
 379                                        atomic_read(&imp->imp_unregistering));
 380                         }
 381                         spin_unlock(&imp->imp_lock);
 382                 }
 383         } while (rc != 0);
 384
 385         /*
 386          * Let's additionally check that no new rpcs added to import in
 387          * "invalidate" state.
 388          */
 389         LASSERT(atomic_read(&imp->imp_inflight) == 0);
 390         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 391         sptlrpc_import_flush_all_ctx(imp);
 392
 393         atomic_dec(&imp->imp_inval_count);
 394         wake_up_all(&imp->imp_recovery_waitq);
 395 }
 396 EXPORT_SYMBOL(ptlrpc_invalidate_import);
 397
 398 /* unset imp_invalid */
 399 void ptlrpc_activate_import(struct obd_import *imp)
 400 {
 401         struct obd_device *obd = imp->imp_obd;
 402
 403         spin_lock(&imp->imp_lock);
 404         if (imp->imp_deactive != 0) {
 405                 spin_unlock(&imp->imp_lock);
 406                 return;
 407         }
 408
 409         imp->imp_invalid = 0;
 410         spin_unlock(&imp->imp_lock);
 411         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 412 }
 413 EXPORT_SYMBOL(ptlrpc_activate_import);
 414
 415 void ptlrpc_pinger_force(struct obd_import *imp)
 416 {
 417         CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
 418                ptlrpc_import_state_name(imp->imp_state));
 419
 420         spin_lock(&imp->imp_lock);
 421         imp->imp_force_verify = 1;
 422         spin_unlock(&imp->imp_lock);
 423
 424         if (imp->imp_state != LUSTRE_IMP_CONNECTING)
 425                 ptlrpc_pinger_wake_up();
 426 }
 427 EXPORT_SYMBOL(ptlrpc_pinger_force);
 428
 429 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 430 {
 431         ENTRY;
 432
 433         LASSERT(!imp->imp_dlm_fake);
 434
 435         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 436                 if (!imp->imp_replayable) {
 437                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 438                                "auto-deactivating\n",
 439                                obd2cli_tgt(imp->imp_obd),
 440                                imp->imp_connection->c_remote_uuid.uuid,
 441                                imp->imp_obd->obd_name);
 442                         ptlrpc_deactivate_import(imp);
 443                 }
 444
 445                 ptlrpc_pinger_force(imp);
 446         }
 447         EXIT;
 448 }
 449
 450 int ptlrpc_reconnect_import(struct obd_import *imp)
 451 {
 452 #ifdef ENABLE_PINGER
 453         long timeout_jiffies = cfs_time_seconds(obd_timeout);
 454         struct l_wait_info lwi;
 455         int rc;
 456
 457         ptlrpc_pinger_force(imp);
 458
 459         CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 460                obd2cli_tgt(imp->imp_obd), obd_timeout);
 461
 462         lwi = LWI_TIMEOUT(timeout_jiffies, NULL, NULL);
 463         rc = l_wait_event(imp->imp_recovery_waitq,
 464                           !ptlrpc_import_in_recovery(imp), &lwi);
 465         CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
 466                ptlrpc_import_state_name(imp->imp_state));
 467         return rc;
 468 #else
 469         ptlrpc_set_import_discon(imp, 0);
 470         /* Force a new connect attempt */
 471         ptlrpc_invalidate_import(imp);
 472         /* Do a fresh connect next time by zeroing the handle */
 473         ptlrpc_disconnect_import(imp, 1);
 474         /* Wait for all invalidate calls to finish */
 475         if (atomic_read(&imp->imp_inval_count) > 0) {
 476                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 477                 int rc;
 478
 479                 rc = l_wait_event(imp->imp_recovery_waitq,
 480                                   (atomic_read(&imp->imp_inval_count) == 0),
 481                                   &lwi);
 482                 if (rc)
 483                         CERROR("Interrupted, inval=%d\n",
 484                                atomic_read(&imp->imp_inval_count));
 485         }
 486
 487         /* Allow reconnect attempts */
 488         imp->imp_obd->obd_no_recov = 0;
 489         /* Remove 'invalid' flag */
 490         ptlrpc_activate_import(imp);
 491         /* Attempt a new connect */
 492         ptlrpc_recover_import(imp, NULL, 0);
 493         return 0;
 494 #endif
 495 }
 496 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 497
 498 /**
 499  * Connection on import \a imp is changed to another one (if more than one is
 500  * present). We typically chose connection that we have not tried to connect to
 501  * the longest
 502  */
 503 static int import_select_connection(struct obd_import *imp)
 504 {
 505         struct obd_import_conn *imp_conn = NULL, *conn;
 506         struct obd_export *dlmexp;
 507         char *target_start;
 508         int target_len, tried_all = 1;
 509         ENTRY;
 510
 511         spin_lock(&imp->imp_lock);
 512
 513         if (list_empty(&imp->imp_conn_list)) {
 514                 CERROR("%s: no connections available\n",
 515                        imp->imp_obd->obd_name);
 516                 spin_unlock(&imp->imp_lock);
 517                 RETURN(-EINVAL);
 518         }
 519
 520         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 521                 CDEBUG(D_HA, "%s: connect to NID %s last attempt %llu\n",
 522                        imp->imp_obd->obd_name,
 523                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 524                        conn->oic_last_attempt);
 525
 526                 /* If we have not tried this connection since
 527                    the last successful attempt, go with this one */
 528                 if ((conn->oic_last_attempt == 0) ||
 529                     cfs_time_beforeq_64(conn->oic_last_attempt,
 530                                        imp->imp_last_success_conn)) {
 531                         imp_conn = conn;
 532                         tried_all = 0;
 533                         break;
 534                 }
 535
 536                 /* If all of the connections have already been tried
 537                    since the last successful connection; just choose the
 538                    least recently used */
 539                 if (!imp_conn)
 540                         imp_conn = conn;
 541                 else if (cfs_time_before_64(conn->oic_last_attempt,
 542                                             imp_conn->oic_last_attempt))
 543                         imp_conn = conn;
 544         }
 545
 546         /* if not found, simply choose the current one */
 547         if (!imp_conn || imp->imp_force_reconnect) {
 548                 LASSERT(imp->imp_conn_current);
 549                 imp_conn = imp->imp_conn_current;
 550                 tried_all = 0;
 551         }
 552         LASSERT(imp_conn->oic_conn);
 553
 554         /* If we've tried everything, and we're back to the beginning of the
 555            list, increase our timeout and try again. It will be reset when
 556            we do finally connect. (FIXME: really we should wait for all network
 557            state associated with the last connection attempt to drain before
 558            trying to reconnect on it.) */
 559         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 560                 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
 561                 if (at_get(at) < CONNECTION_SWITCH_MAX) {
 562                         at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
 563                         if (at_get(at) > CONNECTION_SWITCH_MAX)
 564                                 at_reset(at, CONNECTION_SWITCH_MAX);
 565                 }
 566                 LASSERT(imp_conn->oic_last_attempt);
 567                 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
 568                         "to %ds\n", imp->imp_obd->obd_name, at_get(at));
 569         }
 570
 571         imp_conn->oic_last_attempt = cfs_time_current_64();
 572
 573         /* switch connection, don't mind if it's same as the current one */
 574         if (imp->imp_connection)
 575                 ptlrpc_connection_put(imp->imp_connection);
 576         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 577
 578         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 579         LASSERT(dlmexp != NULL);
 580         if (dlmexp->exp_connection)
 581                 ptlrpc_connection_put(dlmexp->exp_connection);
 582         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 583         class_export_put(dlmexp);
 584
 585         if (imp->imp_conn_current != imp_conn) {
 586                 if (imp->imp_conn_current) {
 587                         deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 588                                   &target_start, &target_len);
 589
 590                         CDEBUG(D_HA, "%s: Connection changing to"
 591                                " %.*s (at %s)\n",
 592                                imp->imp_obd->obd_name,
 593                                target_len, target_start,
 594                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 595                 }
 596
 597                 imp->imp_conn_current = imp_conn;
 598         }
 599
 600         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 601                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 602                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 603
 604         spin_unlock(&imp->imp_lock);
 605
 606         RETURN(0);
 607 }
 608
 609 /*
 610  * must be called under imp_lock
 611  */
 612 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 613 {
 614         struct ptlrpc_request   *req;
 615         struct list_head        *tmp;
 616
 617         /* The requests in committed_list always have smaller transnos than
 618          * the requests in replay_list */
 619         if (!list_empty(&imp->imp_committed_list)) {
 620                 tmp = imp->imp_committed_list.next;
 621                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 622                 *transno = req->rq_transno;
 623                 if (req->rq_transno == 0) {
 624                         DEBUG_REQ(D_ERROR, req, "zero transno in committed_list");
 625                         LBUG();
 626                 }
 627                 return 1;
 628         }
 629         if (!list_empty(&imp->imp_replay_list)) {
 630                 tmp = imp->imp_replay_list.next;
 631                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 632                 *transno = req->rq_transno;
 633                 if (req->rq_transno == 0) {
 634                         DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
 635                         LBUG();
 636                 }
 637                 return 1;
 638         }
 639         return 0;
 640 }
 641
 642 /**
 643  * Attempt to (re)connect import \a imp. This includes all preparations,
 644  * initializing CONNECT RPC request and passing it to ptlrpcd for
 645  * actual sending.
 646  * Returns 0 on success or error code.
 647  */
 648 int ptlrpc_connect_import(struct obd_import *imp)
 649 {
 650         struct obd_device *obd = imp->imp_obd;
 651         int initial_connect = 0;
 652         int set_transno = 0;
 653         __u64 committed_before_reconnect = 0;
 654         struct ptlrpc_request *request;
 655         char *bufs[] = { NULL,
 656                          obd2cli_tgt(imp->imp_obd),
 657                          obd->obd_uuid.uuid,
 658                          (char *)&imp->imp_dlm_handle,
 659                          (char *)&imp->imp_connect_data };
 660         struct ptlrpc_connect_async_args *aa;
 661         int rc;
 662         ENTRY;
 663
 664         spin_lock(&imp->imp_lock);
 665         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 666                 spin_unlock(&imp->imp_lock);
 667                 CERROR("can't connect to a closed import\n");
 668                 RETURN(-EINVAL);
 669         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 670                 spin_unlock(&imp->imp_lock);
 671                 CERROR("already connected\n");
 672                 RETURN(0);
 673         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING ||
 674                    imp->imp_connected) {
 675                 spin_unlock(&imp->imp_lock);
 676                 CERROR("already connecting\n");
 677                 RETURN(-EALREADY);
 678         }
 679
 680         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 681
 682         imp->imp_conn_cnt++;
 683         imp->imp_resend_replay = 0;
 684
 685         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 686                 initial_connect = 1;
 687         else
 688                 committed_before_reconnect = imp->imp_peer_committed_transno;
 689
 690         set_transno = ptlrpc_first_transno(imp,
 691                                            &imp->imp_connect_data.ocd_transno);
 692         spin_unlock(&imp->imp_lock);
 693
 694         rc = import_select_connection(imp);
 695         if (rc)
 696                 GOTO(out, rc);
 697
 698         rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
 699         if (rc)
 700                 GOTO(out, rc);
 701
 702         /* Reset connect flags to the originally requested flags, in case
 703          * the server is updated on-the-fly we will get the new features. */
 704         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 705         imp->imp_connect_data.ocd_connect_flags2 = imp->imp_connect_flags2_orig;
 706         /* Reset ocd_version each time so the server knows the exact versions */
 707         imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
 708         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 709         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 710
 711         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 712                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 713         if (rc)
 714                 GOTO(out, rc);
 715
 716         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 717         if (request == NULL)
 718                 GOTO(out, rc = -ENOMEM);
 719
 720         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 721                                       imp->imp_connect_op, bufs, NULL);
 722         if (rc) {
 723                 ptlrpc_request_free(request);
 724                 GOTO(out, rc);
 725         }
 726
 727         /* Report the rpc service time to the server so that it knows how long
 728          * to wait for clients to join recovery */
 729         lustre_msg_set_service_time(request->rq_reqmsg,
 730                                     at_timeout2est(request->rq_timeout));
 731
 732         /* The amount of time we give the server to process the connect req.
 733          * import_select_connection will increase the net latency on
 734          * repeated reconnect attempts to cover slow networks.
 735          * We override/ignore the server rpc completion estimate here,
 736          * which may be large if this is a reconnect attempt */
 737         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 738         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 739
 740         request->rq_no_resend = request->rq_no_delay = 1;
 741         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 742         /* Allow a slightly larger reply for future growth compatibility */
 743         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 744                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 745         ptlrpc_request_set_replen(request);
 746         request->rq_interpret_reply = ptlrpc_connect_interpret;
 747
 748         CLASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
 749         aa = ptlrpc_req_async_args(request);
 750         memset(aa, 0, sizeof *aa);
 751
 752         aa->pcaa_peer_committed = committed_before_reconnect;
 753         aa->pcaa_initial_connect = initial_connect;
 754
 755         if (aa->pcaa_initial_connect) {
 756                 spin_lock(&imp->imp_lock);
 757                 imp->imp_replayable = 1;
 758                 spin_unlock(&imp->imp_lock);
 759                 lustre_msg_add_op_flags(request->rq_reqmsg,
 760                                         MSG_CONNECT_INITIAL);
 761         }
 762
 763         if (set_transno)
 764                 lustre_msg_add_op_flags(request->rq_reqmsg,
 765                                         MSG_CONNECT_TRANSNO);
 766
 767         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %ld)",
 768                   request->rq_timeout);
 769         ptlrpcd_add_req(request);
 770         rc = 0;
 771 out:
 772         if (rc != 0)
 773                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 774
 775         RETURN(rc);
 776 }
 777 EXPORT_SYMBOL(ptlrpc_connect_import);
 778
 779 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 780 {
 781         int force_verify;
 782
 783         spin_lock(&imp->imp_lock);
 784         force_verify = imp->imp_force_verify != 0;
 785         spin_unlock(&imp->imp_lock);
 786
 787         if (force_verify)
 788                 ptlrpc_pinger_wake_up();
 789 }
 790
 791 static int ptlrpc_busy_reconnect(int rc)
 792 {
 793         return (rc == -EBUSY) || (rc == -EAGAIN);
 794 }
 795
 796 static int ptlrpc_connect_set_flags(struct obd_import *imp,
 797                                      struct obd_connect_data *ocd,
 798                                      __u64 old_connect_flags,
 799                                      struct obd_export *exp, int init_connect)
 800 {
 801         static bool warned;
 802         struct client_obd *cli = &imp->imp_obd->u.cli;
 803
 804         spin_lock(&imp->imp_lock);
 805         list_del(&imp->imp_conn_current->oic_item);
 806         list_add(&imp->imp_conn_current->oic_item,
 807                  &imp->imp_conn_list);
 808         imp->imp_last_success_conn =
 809                 imp->imp_conn_current->oic_last_attempt;
 810
 811         spin_unlock(&imp->imp_lock);
 812
 813         if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 814             (ocd->ocd_version > LUSTRE_VERSION_CODE +
 815                                 LUSTRE_VERSION_OFFSET_WARN ||
 816              ocd->ocd_version < LUSTRE_VERSION_CODE -
 817                                 LUSTRE_VERSION_OFFSET_WARN)) {
 818                 /* Sigh, some compilers do not like #ifdef in the middle
 819                    of macro arguments */
 820                 const char *older = "older than client. "
 821                                     "Consider upgrading server";
 822                 const char *newer = "newer than client. "
 823                                     "Consider upgrading client";
 824
 825                 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 826                               "is much %s (%s)\n",
 827                               obd2cli_tgt(imp->imp_obd),
 828                               OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 829                               OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 830                               OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 831                               OBD_OCD_VERSION_FIX(ocd->ocd_version),
 832                               ocd->ocd_version > LUSTRE_VERSION_CODE ?
 833                               newer : older, LUSTRE_VERSION_STRING);
 834                 warned = true;
 835         }
 836
 837 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
 838         /* Check if server has LU-1252 fix applied to not always swab
 839          * the IR MNE entries. Do this only once per connection.  This
 840          * fixup is version-limited, because we don't want to carry the
 841          * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
 842          * need interop with unpatched 2.2 servers.  For newer servers,
 843          * the client will do MNE swabbing only as needed.  LU-1644 */
 844         if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 845                      !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
 846                      OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
 847                      OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
 848                      OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
 849                      strcmp(imp->imp_obd->obd_type->typ_name,
 850                             LUSTRE_MGC_NAME) == 0))
 851                 imp->imp_need_mne_swab = 1;
 852         else /* clear if server was upgraded since last connect */
 853                 imp->imp_need_mne_swab = 0;
 854 #endif
 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                      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                                        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 = cksum_type_select(cli->cl_supp_cksum_types);
 878
 879         if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
 880                 cli->cl_max_pages_per_rpc =
 881                         min(ocd->ocd_brw_size >> PAGE_SHIFT,
 882                             cli->cl_max_pages_per_rpc);
 883         else if (imp->imp_connect_op == MDS_CONNECT ||
 884                  imp->imp_connect_op == MGS_CONNECT)
 885                 cli->cl_max_pages_per_rpc = 1;
 886
 887         LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 888                 (cli->cl_max_pages_per_rpc > 0));
 889
 890         client_adjust_max_dirty(cli);
 891
 892         /* Update client max modify RPCs in flight with value returned
 893          * by the server */
 894         if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
 895                 cli->cl_max_mod_rpcs_in_flight = min(
 896                                         cli->cl_max_mod_rpcs_in_flight,
 897                                         ocd->ocd_maxmodrpcs);
 898         else
 899                 cli->cl_max_mod_rpcs_in_flight = 1;
 900
 901         /* Reset ns_connect_flags only for initial connect. It might be
 902          * changed in while using FS and if we reset it in reconnect
 903          * this leads to losing user settings done before such as
 904          * disable lru_resize, etc. */
 905         if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
 906                 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
 907                              "flags: %#llx\n", imp->imp_obd->obd_name,
 908                              ocd->ocd_connect_flags);
 909                 imp->imp_obd->obd_namespace->ns_connect_flags =
 910                         ocd->ocd_connect_flags;
 911                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
 912                         ocd->ocd_connect_flags;
 913         }
 914
 915         if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
 916                 /* We need a per-message support flag, because
 917                  * a. we don't know if the incoming connect reply
 918                  *    supports AT or not (in reply_in_callback)
 919                  *    until we unpack it.
 920                  * b. failovered server means export and flags are gone
 921                  *    (in ptlrpc_send_reply).
 922                  *    Can only be set when we know AT is supported at
 923                  *    both ends */
 924                 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 925         else
 926                 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 927
 928         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 929
 930         return 0;
 931 }
 932
 933 /**
 934  * Add all replay requests back to unreplied list before start replay,
 935  * so that we can make sure the known replied XID is always increased
 936  * only even if when replaying requests.
 937  */
 938 static void ptlrpc_prepare_replay(struct obd_import *imp)
 939 {
 940         struct ptlrpc_request *req;
 941
 942         if (imp->imp_state != LUSTRE_IMP_REPLAY ||
 943             imp->imp_resend_replay)
 944                 return;
 945
 946         /* If the server was restart during repaly, the requests may
 947          * have been added to the unreplied list in former replay. */
 948         spin_lock(&imp->imp_lock);
 949
 950         list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
 951                 if (list_empty(&req->rq_unreplied_list))
 952                         ptlrpc_add_unreplied(req);
 953         }
 954
 955         list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
 956                 if (list_empty(&req->rq_unreplied_list))
 957                         ptlrpc_add_unreplied(req);
 958         }
 959
 960         imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
 961         spin_unlock(&imp->imp_lock);
 962 }
 963
 964 /**
 965  * interpret_reply callback for connect RPCs.
 966  * Looks into returned status of connect operation and decides
 967  * what to do with the import - i.e enter recovery, promote it to
 968  * full state for normal operations of disconnect it due to an error.
 969  */
 970 static int ptlrpc_connect_interpret(const struct lu_env *env,
 971                                     struct ptlrpc_request *request,
 972                                     void *data, int rc)
 973 {
 974         struct ptlrpc_connect_async_args *aa = data;
 975         struct obd_import *imp = request->rq_import;
 976         struct lustre_handle old_hdl;
 977         __u64 old_connect_flags;
 978         int msg_flags;
 979         struct obd_connect_data *ocd;
 980         struct obd_export *exp;
 981         int ret;
 982         ENTRY;
 983
 984         spin_lock(&imp->imp_lock);
 985         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 986                 imp->imp_connect_tried = 1;
 987                 spin_unlock(&imp->imp_lock);
 988                 RETURN(0);
 989         }
 990
 991         if (rc) {
 992                 /* if this reconnect to busy export - not need select new target
 993                  * for connecting*/
 994                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 995                 spin_unlock(&imp->imp_lock);
 996                 ptlrpc_maybe_ping_import_soon(imp);
 997                 GOTO(out, rc);
 998         }
 999
1000         /* LU-7558: indicate that we are interpretting connect reply,
1001          * pltrpc_connect_import() will not try to reconnect until
1002          * interpret will finish. */
1003         imp->imp_connected = 1;
1004         spin_unlock(&imp->imp_lock);
1005
1006         LASSERT(imp->imp_conn_current);
1007
1008         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1009
1010         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1011                                    RCL_SERVER);
1012         /* server replied obd_connect_data is always bigger */
1013         ocd = req_capsule_server_sized_get(&request->rq_pill,
1014                                            &RMF_CONNECT_DATA, ret);
1015
1016         if (ocd == NULL) {
1017                 CERROR("%s: no connect data from server\n",
1018                        imp->imp_obd->obd_name);
1019                 rc = -EPROTO;
1020                 GOTO(out, rc);
1021         }
1022
1023         spin_lock(&imp->imp_lock);
1024
1025         /* All imports are pingable */
1026         imp->imp_pingable = 1;
1027         imp->imp_force_reconnect = 0;
1028         imp->imp_force_verify = 0;
1029
1030         imp->imp_connect_data = *ocd;
1031
1032         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1033                imp->imp_obd->obd_name, ocd->ocd_instance);
1034         exp = class_conn2export(&imp->imp_dlm_handle);
1035
1036         spin_unlock(&imp->imp_lock);
1037
1038         if (!exp) {
1039                 /* This could happen if export is cleaned during the
1040                    connect attempt */
1041                 CERROR("%s: missing export after connect\n",
1042                        imp->imp_obd->obd_name);
1043                 GOTO(out, rc = -ENODEV);
1044         }
1045
1046         /* check that server granted subset of flags we asked for. */
1047         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1048             ocd->ocd_connect_flags) {
1049                 CERROR("%s: Server didn't grant requested subset of flags: "
1050                        "asked=%#llx granted=%#llx\n",
1051                        imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
1052                        ocd->ocd_connect_flags);
1053                 GOTO(out, rc = -EPROTO);
1054         }
1055
1056         if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) !=
1057             ocd->ocd_connect_flags2) {
1058                 CERROR("%s: Server didn't grant requested subset of flags2: "
1059                        "asked=%#llx granted=%#llx\n",
1060                        imp->imp_obd->obd_name, imp->imp_connect_flags2_orig,
1061                        ocd->ocd_connect_flags2);
1062                 GOTO(out, rc = -EPROTO);
1063         }
1064
1065         if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1066             (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1067             (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1068             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1069                 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1070                 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1071                 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1072
1073                 /* We do not support the MDT-MDT interoperations with
1074                  * different version MDT because of protocol changes. */
1075                 if (unlikely(major != LUSTRE_MAJOR ||
1076                              minor != LUSTRE_MINOR ||
1077                              abs(patch - LUSTRE_PATCH) > 3)) {
1078                         LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1079                                       "connection to different version MDT "
1080                                       "(%d.%d.%d.%d) %s\n",
1081                                       imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1082                                       LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1083                                       major, minor, patch,
1084                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1085                                       imp->imp_connection->c_remote_uuid.uuid);
1086
1087                         GOTO(out, rc = -EPROTO);
1088                 }
1089         }
1090
1091         old_connect_flags = exp_connect_flags(exp);
1092         exp->exp_connect_data = *ocd;
1093         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1094
1095         /* The net statistics after (re-)connect is not valid anymore,
1096          * because may reflect other routing, etc. */
1097         at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1098         ptlrpc_at_adj_net_latency(request,
1099                         lustre_msg_get_service_time(request->rq_repmsg));
1100
1101         /* Import flags should be updated before waking import at FULL state */
1102         rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1103                                       aa->pcaa_initial_connect);
1104         class_export_put(exp);
1105         if (rc != 0)
1106                 GOTO(out, rc);
1107
1108         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1109
1110         if (aa->pcaa_initial_connect) {
1111                 spin_lock(&imp->imp_lock);
1112                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1113                         imp->imp_replayable = 1;
1114                         spin_unlock(&imp->imp_lock);
1115                         CDEBUG(D_HA, "connected to replayable target: %s\n",
1116                                obd2cli_tgt(imp->imp_obd));
1117                 } else {
1118                         imp->imp_replayable = 0;
1119                         spin_unlock(&imp->imp_lock);
1120                 }
1121
1122                 /* if applies, adjust the imp->imp_msg_magic here
1123                  * according to reply flags */
1124
1125                 imp->imp_remote_handle =
1126                                 *lustre_msg_get_handle(request->rq_repmsg);
1127
1128                 /* Initial connects are allowed for clients with non-random
1129                  * uuids when servers are in recovery.  Simply signal the
1130                  * servers replay is complete and wait in REPLAY_WAIT. */
1131                 if (msg_flags & MSG_CONNECT_RECOVERING) {
1132                         CDEBUG(D_HA, "connect to %s during recovery\n",
1133                                obd2cli_tgt(imp->imp_obd));
1134                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1135                 } else {
1136                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1137                         ptlrpc_activate_import(imp);
1138                 }
1139
1140                 GOTO(finish, rc = 0);
1141         }
1142
1143         /* Determine what recovery state to move the import to. */
1144         if (MSG_CONNECT_RECONNECT & msg_flags) {
1145                 memset(&old_hdl, 0, sizeof(old_hdl));
1146                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1147                             sizeof (old_hdl))) {
1148                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1149                                       "bad handle %#llx\n",
1150                                       obd2cli_tgt(imp->imp_obd),
1151                                       imp->imp_connection->c_remote_uuid.uuid,
1152                                       imp->imp_dlm_handle.cookie);
1153                         GOTO(out, rc = -ENOTCONN);
1154                 }
1155
1156                 if (memcmp(&imp->imp_remote_handle,
1157                            lustre_msg_get_handle(request->rq_repmsg),
1158                            sizeof(imp->imp_remote_handle))) {
1159                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
1160                                 D_HA : D_WARNING;
1161
1162                         /* Bug 16611/14775: if server handle have changed,
1163                          * that means some sort of disconnection happened.
1164                          * If the server is not in recovery, that also means it
1165                          * already erased all of our state because of previous
1166                          * eviction. If it is in recovery - we are safe to
1167                          * participate since we can reestablish all of our state
1168                          * with server again */
1169                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1170                                 CDEBUG(level,"%s@%s changed server handle from "
1171                                        "%#llx to %#llx"
1172                                        " but is still in recovery\n",
1173                                        obd2cli_tgt(imp->imp_obd),
1174                                        imp->imp_connection->c_remote_uuid.uuid,
1175                                        imp->imp_remote_handle.cookie,
1176                                        lustre_msg_get_handle(
1177                                        request->rq_repmsg)->cookie);
1178                         } else {
1179                                 LCONSOLE_WARN("Evicted from %s (at %s) "
1180                                               "after server handle changed from "
1181                                               "%#llx to %#llx\n",
1182                                               obd2cli_tgt(imp->imp_obd),
1183                                               imp->imp_connection-> \
1184                                               c_remote_uuid.uuid,
1185                                               imp->imp_remote_handle.cookie,
1186                                               lustre_msg_get_handle(
1187                                               request->rq_repmsg)->cookie);
1188                         }
1189
1190
1191                         imp->imp_remote_handle =
1192                                      *lustre_msg_get_handle(request->rq_repmsg);
1193
1194                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1195                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1196                                 GOTO(finish, rc = 0);
1197                         }
1198
1199                 } else {
1200                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1201                                obd2cli_tgt(imp->imp_obd),
1202                                imp->imp_connection->c_remote_uuid.uuid);
1203                 }
1204
1205                 if (imp->imp_invalid) {
1206                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1207                                "marking evicted\n", imp->imp_obd->obd_name);
1208                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1209                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1210                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1211                                imp->imp_obd->obd_name,
1212                                obd2cli_tgt(imp->imp_obd));
1213
1214                         spin_lock(&imp->imp_lock);
1215                         imp->imp_resend_replay = 1;
1216                         spin_unlock(&imp->imp_lock);
1217
1218                         IMPORT_SET_STATE(imp, imp->imp_replay_state);
1219                 } else {
1220                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1221                 }
1222         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1223                 LASSERT(imp->imp_replayable);
1224                 imp->imp_remote_handle =
1225                                 *lustre_msg_get_handle(request->rq_repmsg);
1226                 imp->imp_last_replay_transno = 0;
1227                 imp->imp_replay_cursor = &imp->imp_committed_list;
1228                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
1229         } else {
1230                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
1231                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
1232                 imp->imp_remote_handle =
1233                                 *lustre_msg_get_handle(request->rq_repmsg);
1234                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1235         }
1236
1237         /* Sanity checks for a reconnected import. */
1238         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
1239                 CERROR("imp_replayable flag does not match server "
1240                        "after reconnect. We should LBUG right here.\n");
1241         }
1242
1243         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1244             lustre_msg_get_last_committed(request->rq_repmsg) <
1245             aa->pcaa_peer_committed) {
1246                 CERROR("%s went back in time (transno %lld"
1247                        " was previously committed, server now claims %lld"
1248                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
1249                        "id=9646\n",
1250                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1251                        lustre_msg_get_last_committed(request->rq_repmsg));
1252         }
1253
1254 finish:
1255         ptlrpc_prepare_replay(imp);
1256         rc = ptlrpc_import_recovery_state_machine(imp);
1257         if (rc == -ENOTCONN) {
1258                 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1259                        "invalidating and reconnecting\n",
1260                        obd2cli_tgt(imp->imp_obd),
1261                        imp->imp_connection->c_remote_uuid.uuid);
1262                 ptlrpc_connect_import(imp);
1263                 spin_lock(&imp->imp_lock);
1264                 imp->imp_connected = 0;
1265                 imp->imp_connect_tried = 1;
1266                 spin_unlock(&imp->imp_lock);
1267                 RETURN(0);
1268         }
1269
1270 out:
1271         spin_lock(&imp->imp_lock);
1272         imp->imp_connected = 0;
1273         imp->imp_connect_tried = 1;
1274         spin_unlock(&imp->imp_lock);
1275
1276         if (rc != 0) {
1277                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1278                 if (rc == -EACCES) {
1279                         /*
1280                          * Give up trying to reconnect
1281                          * EACCES means client has no permission for connection
1282                          */
1283                         imp->imp_obd->obd_no_recov = 1;
1284                         ptlrpc_deactivate_import(imp);
1285                 }
1286
1287                 if (rc == -EPROTO) {
1288                         struct obd_connect_data *ocd;
1289
1290                         /* reply message might not be ready */
1291                         if (request->rq_repmsg == NULL)
1292                                 RETURN(-EPROTO);
1293
1294                         ocd = req_capsule_server_get(&request->rq_pill,
1295                                                      &RMF_CONNECT_DATA);
1296                         /* Servers are not supposed to refuse connections from
1297                          * clients based on version, only connection feature
1298                          * flags.  We should never see this from llite, but it
1299                          * may be useful for debugging in the future. */
1300                         if (ocd &&
1301                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1302                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1303                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1304                                         "(%d.%d.%d.%d)"
1305                                         " refused connection from this client "
1306                                         "with an incompatible version (%s).  "
1307                                         "Client must be recompiled\n",
1308                                         obd2cli_tgt(imp->imp_obd),
1309                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1310                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1311                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1312                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1313                                         LUSTRE_VERSION_STRING);
1314                                 ptlrpc_deactivate_import(imp);
1315                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1316                         }
1317                         RETURN(-EPROTO);
1318                 }
1319
1320                 ptlrpc_maybe_ping_import_soon(imp);
1321
1322                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1323                        obd2cli_tgt(imp->imp_obd),
1324                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1325         }
1326
1327         wake_up_all(&imp->imp_recovery_waitq);
1328         RETURN(rc);
1329 }
1330
1331 /**
1332  * interpret callback for "completed replay" RPCs.
1333  * \see signal_completed_replay
1334  */
1335 static int completed_replay_interpret(const struct lu_env *env,
1336                                       struct ptlrpc_request *req,
1337                                       void * data, int rc)
1338 {
1339         ENTRY;
1340         atomic_dec(&req->rq_import->imp_replay_inflight);
1341         if (req->rq_status == 0 &&
1342             !req->rq_import->imp_vbr_failed) {
1343                 ptlrpc_import_recovery_state_machine(req->rq_import);
1344         } else {
1345                 if (req->rq_import->imp_vbr_failed) {
1346                         CDEBUG(D_WARNING,
1347                                "%s: version recovery fails, reconnecting\n",
1348                                req->rq_import->imp_obd->obd_name);
1349                 } else {
1350                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1351                                      "reconnecting\n",
1352                                req->rq_import->imp_obd->obd_name,
1353                                req->rq_status);
1354                 }
1355                 ptlrpc_connect_import(req->rq_import);
1356         }
1357
1358         RETURN(0);
1359 }
1360
1361 /**
1362  * Let server know that we have no requests to replay anymore.
1363  * Achieved by just sending a PING request
1364  */
1365 static int signal_completed_replay(struct obd_import *imp)
1366 {
1367         struct ptlrpc_request *req;
1368         ENTRY;
1369
1370         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1371                 RETURN(0);
1372
1373         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1374         atomic_inc(&imp->imp_replay_inflight);
1375
1376         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1377                                         OBD_PING);
1378         if (req == NULL) {
1379                 atomic_dec(&imp->imp_replay_inflight);
1380                 RETURN(-ENOMEM);
1381         }
1382
1383         ptlrpc_request_set_replen(req);
1384         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1385         lustre_msg_add_flags(req->rq_reqmsg,
1386                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1387         if (AT_OFF)
1388                 req->rq_timeout *= 3;
1389         req->rq_interpret_reply = completed_replay_interpret;
1390
1391         ptlrpcd_add_req(req);
1392         RETURN(0);
1393 }
1394
1395 /**
1396  * In kernel code all import invalidation happens in its own
1397  * separate thread, so that whatever application happened to encounter
1398  * a problem could still be killed or otherwise continue
1399  */
1400 static int ptlrpc_invalidate_import_thread(void *data)
1401 {
1402         struct obd_import *imp = data;
1403
1404         ENTRY;
1405
1406         unshare_fs_struct();
1407
1408         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1409                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1410                imp->imp_connection->c_remote_uuid.uuid);
1411
1412         ptlrpc_invalidate_import(imp);
1413
1414         if (obd_dump_on_eviction) {
1415                 CERROR("dump the log upon eviction\n");
1416                 libcfs_debug_dumplog();
1417         }
1418
1419         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1420         ptlrpc_import_recovery_state_machine(imp);
1421
1422         class_import_put(imp);
1423         RETURN(0);
1424 }
1425
1426 /**
1427  * This is the state machine for client-side recovery on import.
1428  *
1429  * Typicaly we have two possibly paths. If we came to server and it is not
1430  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1431  * state and reconnect from scratch.
1432  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1433  * We go through our list of requests to replay and send them to server one by
1434  * one.
1435  * After sending all request from the list we change import state to
1436  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1437  * and also all the locks we don't yet have and wait for server to grant us.
1438  * After that we send a special "replay completed" request and change import
1439  * state to IMP_REPLAY_WAIT.
1440  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1441  * state and resend all requests from sending list.
1442  * After that we promote import to FULL state and send all delayed requests
1443  * and import is fully operational after that.
1444  *
1445  */
1446 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1447 {
1448         int rc = 0;
1449         int inflight;
1450         char *target_start;
1451         int target_len;
1452
1453         ENTRY;
1454         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1455                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1456                           &target_start, &target_len);
1457                 /* Don't care about MGC eviction */
1458                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1459                            LUSTRE_MGC_NAME) != 0) {
1460                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1461                                            "by %.*s; in progress operations "
1462                                            "using this service will fail.\n",
1463                                            imp->imp_obd->obd_name, target_len,
1464                                            target_start);
1465                 }
1466                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1467                        obd2cli_tgt(imp->imp_obd),
1468                        imp->imp_connection->c_remote_uuid.uuid);
1469                 /* reset vbr_failed flag upon eviction */
1470                 spin_lock(&imp->imp_lock);
1471                 imp->imp_vbr_failed = 0;
1472                 spin_unlock(&imp->imp_lock);
1473
1474                 {
1475                 struct task_struct *task;
1476                 /* bug 17802:  XXX client_disconnect_export vs connect request
1477                  * race. if client is evicted at this time then we start
1478                  * invalidate thread without reference to import and import can
1479                  * be freed at same time. */
1480                 class_import_get(imp);
1481                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1482                                      "ll_imp_inval");
1483                 if (IS_ERR(task)) {
1484                         class_import_put(imp);
1485                         CERROR("error starting invalidate thread: %d\n", rc);
1486                         rc = PTR_ERR(task);
1487                 } else {
1488                         rc = 0;
1489                 }
1490                 RETURN(rc);
1491                 }
1492         }
1493
1494         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1495                 CDEBUG(D_HA, "replay requested by %s\n",
1496                        obd2cli_tgt(imp->imp_obd));
1497                 rc = ptlrpc_replay_next(imp, &inflight);
1498                 if (inflight == 0 &&
1499                     atomic_read(&imp->imp_replay_inflight) == 0) {
1500                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1501                         rc = ldlm_replay_locks(imp);
1502                         if (rc)
1503                                 GOTO(out, rc);
1504                 }
1505                 rc = 0;
1506         }
1507
1508         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1509                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1510                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1511                         rc = signal_completed_replay(imp);
1512                         if (rc)
1513                                 GOTO(out, rc);
1514                 }
1515         }
1516
1517         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1518                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1519                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1520                 }
1521         }
1522
1523         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1524                 struct ptlrpc_connection *conn = imp->imp_connection;
1525
1526                 rc = ptlrpc_resend(imp);
1527                 if (rc)
1528                         GOTO(out, rc);
1529                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1530                 ptlrpc_activate_import(imp);
1531
1532                 LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n",
1533                               imp->imp_obd->obd_name,
1534                               obd_uuid2str(&conn->c_remote_uuid),
1535                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1536         }
1537
1538         if (imp->imp_state == LUSTRE_IMP_FULL) {
1539                 wake_up_all(&imp->imp_recovery_waitq);
1540                 ptlrpc_wake_delayed(imp);
1541         }
1542
1543 out:
1544         RETURN(rc);
1545 }
1546
1547 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1548 {
1549         struct ptlrpc_request *req;
1550         int rq_opc, rc = 0;
1551         ENTRY;
1552
1553         if (imp->imp_obd->obd_force)
1554                 GOTO(set_state, rc);
1555
1556         switch (imp->imp_connect_op) {
1557         case OST_CONNECT:
1558                 rq_opc = OST_DISCONNECT;
1559                 break;
1560         case MDS_CONNECT:
1561                 rq_opc = MDS_DISCONNECT;
1562                 break;
1563         case MGS_CONNECT:
1564                 rq_opc = MGS_DISCONNECT;
1565                 break;
1566         default:
1567                 rc = -EINVAL;
1568                 CERROR("%s: don't know how to disconnect from %s "
1569                        "(connect_op %d): rc = %d\n",
1570                        imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1571                        imp->imp_connect_op, rc);
1572                 RETURN(rc);
1573         }
1574
1575         if (ptlrpc_import_in_recovery(imp)) {
1576                 struct l_wait_info lwi;
1577                 long timeout_jiffies;
1578                 time64_t timeout;
1579
1580                 if (AT_OFF) {
1581                         if (imp->imp_server_timeout)
1582                                 timeout = obd_timeout >> 1;
1583                         else
1584                                 timeout = obd_timeout;
1585                 } else {
1586                         u32 req_portal;
1587                         int idx;
1588
1589                         req_portal = imp->imp_client->cli_request_portal;
1590                         idx = import_at_get_index(imp, req_portal);
1591                         timeout = at_get(&imp->imp_at.iat_service_estimate[idx]);
1592                 }
1593
1594                 timeout_jiffies = cfs_time_seconds(timeout);
1595                 lwi = LWI_TIMEOUT_INTR(max_t(long, timeout_jiffies, 1),
1596                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1597                 rc = l_wait_event(imp->imp_recovery_waitq,
1598                                   !ptlrpc_import_in_recovery(imp), &lwi);
1599
1600         }
1601
1602         spin_lock(&imp->imp_lock);
1603         if (imp->imp_state != LUSTRE_IMP_FULL)
1604                 GOTO(out, rc);
1605         spin_unlock(&imp->imp_lock);
1606
1607         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1608                                         LUSTRE_OBD_VERSION, rq_opc);
1609         if (req) {
1610                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1611                  * it fails.  We can get through the above with a down server
1612                  * if the client doesn't know the server is gone yet. */
1613                 req->rq_no_resend = 1;
1614
1615                 /* We want client umounts to happen quickly, no matter the
1616                    server state... */
1617                 req->rq_timeout = min_t(int, req->rq_timeout,
1618                                         INITIAL_CONNECT_TIMEOUT);
1619
1620                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1621                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1622                 ptlrpc_request_set_replen(req);
1623                 rc = ptlrpc_queue_wait(req);
1624                 ptlrpc_req_finished(req);
1625         }
1626
1627 set_state:
1628         spin_lock(&imp->imp_lock);
1629 out:
1630         if (noclose)
1631                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1632         else
1633                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1634         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1635         spin_unlock(&imp->imp_lock);
1636
1637         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1638                 rc = 0;
1639         RETURN(rc);
1640 }
1641 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1642
1643 void ptlrpc_cleanup_imp(struct obd_import *imp)
1644 {
1645         ENTRY;
1646
1647         spin_lock(&imp->imp_lock);
1648         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1649         imp->imp_generation++;
1650         spin_unlock(&imp->imp_lock);
1651         ptlrpc_abort_inflight(imp);
1652
1653         EXIT;
1654 }
1655
1656 /* Adaptive Timeout utils */
1657 extern unsigned int at_min, at_max, at_history;
1658
1659 /* Update at_current with the specified value (bounded by at_min and at_max),
1660  * as well as the AT history "bins".
1661  *  - Bin into timeslices using AT_BINS bins.
1662  *  - This gives us a max of the last at_history seconds without the storage,
1663  *    but still smoothing out a return to normalcy from a slow response.
1664  *  - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1665  */
1666 int at_measured(struct adaptive_timeout *at, unsigned int val)
1667 {
1668         unsigned int old = at->at_current;
1669         time64_t now = ktime_get_real_seconds();
1670         long binlimit = max_t(long, at_history / AT_BINS, 1);
1671
1672         LASSERT(at);
1673         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1674                val, at, (long)(now - at->at_binstart), at->at_current,
1675                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1676
1677         if (val == 0)
1678                 /* 0's don't count, because we never want our timeout to
1679                    drop to 0, and because 0 could mean an error */
1680                 return 0;
1681
1682         spin_lock(&at->at_lock);
1683
1684         if (unlikely(at->at_binstart == 0)) {
1685                 /* Special case to remove default from history */
1686                 at->at_current = val;
1687                 at->at_worst_ever = val;
1688                 at->at_worst_time = now;
1689                 at->at_hist[0] = val;
1690                 at->at_binstart = now;
1691         } else if (now - at->at_binstart < binlimit ) {
1692                 /* in bin 0 */
1693                 at->at_hist[0] = max(val, at->at_hist[0]);
1694                 at->at_current = max(val, at->at_current);
1695         } else {
1696                 int i, shift;
1697                 unsigned int maxv = val;
1698
1699                 /* move bins over */
1700                 shift = (u32)(now - at->at_binstart) / binlimit;
1701                 LASSERT(shift > 0);
1702                 for(i = AT_BINS - 1; i >= 0; i--) {
1703                         if (i >= shift) {
1704                                 at->at_hist[i] = at->at_hist[i - shift];
1705                                 maxv = max(maxv, at->at_hist[i]);
1706                         } else {
1707                                 at->at_hist[i] = 0;
1708                         }
1709                 }
1710                 at->at_hist[0] = val;
1711                 at->at_current = maxv;
1712                 at->at_binstart += shift * binlimit;
1713         }
1714
1715         if (at->at_current > at->at_worst_ever) {
1716                 at->at_worst_ever = at->at_current;
1717                 at->at_worst_time = now;
1718         }
1719
1720         if (at->at_flags & AT_FLG_NOHIST)
1721                 /* Only keep last reported val; keeping the rest of the history
1722                    for proc only */
1723                 at->at_current = val;
1724
1725         if (at_max > 0)
1726                 at->at_current =  min(at->at_current, at_max);
1727         at->at_current =  max(at->at_current, at_min);
1728
1729         if (at->at_current != old)
1730                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1731                        "(val=%u) hist %u %u %u %u\n", at,
1732                        old, at->at_current, at->at_current - old, val,
1733                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1734                        at->at_hist[3]);
1735
1736         /* if we changed, report the old value */
1737         old = (at->at_current != old) ? old : 0;
1738
1739         spin_unlock(&at->at_lock);
1740         return old;
1741 }
1742
1743 /* Find the imp_at index for a given portal; assign if space available */
1744 int import_at_get_index(struct obd_import *imp, int portal)
1745 {
1746         struct imp_at *at = &imp->imp_at;
1747         int i;
1748
1749         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1750                 if (at->iat_portal[i] == portal)
1751                         return i;
1752                 if (at->iat_portal[i] == 0)
1753                         /* unused */
1754                         break;
1755         }
1756
1757         /* Not found in list, add it under a lock */
1758         spin_lock(&imp->imp_lock);
1759
1760         /* Check unused under lock */
1761         for (; i < IMP_AT_MAX_PORTALS; i++) {
1762                 if (at->iat_portal[i] == portal)
1763                         goto out;
1764                 if (at->iat_portal[i] == 0)
1765                         /* unused */
1766                         break;
1767         }
1768
1769         /* Not enough portals? */
1770         LASSERT(i < IMP_AT_MAX_PORTALS);
1771
1772         at->iat_portal[i] = portal;
1773 out:
1774         spin_unlock(&imp->imp_lock);
1775         return i;
1776 }