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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/import.c
  37  *
  38  * Author: Mike Shaver <shaver@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_RPC
  42
  43 #include <linux/kthread.h>
  44 #include <obd_support.h>
  45 #include <lustre_ha.h>
  46 #include <lustre_net.h>
  47 #include <lustre_import.h>
  48 #include <lustre_export.h>
  49 #include <obd.h>
  50 #include <obd_cksum.h>
  51 #include <obd_class.h>
  52
  53 #include "ptlrpc_internal.h"
  54
  55 struct ptlrpc_connect_async_args {
  56          __u64 pcaa_peer_committed;
  57         int pcaa_initial_connect;
  58 };
  59
  60 /**
  61  * Updates import \a imp current state to provided \a state value
  62  * Helper function. Must be called under imp_lock.
  63  */
  64 static void __import_set_state(struct obd_import *imp,
  65                                enum lustre_imp_state state)
  66 {
  67         switch (state) {
  68         case LUSTRE_IMP_CLOSED:
  69         case LUSTRE_IMP_NEW:
  70         case LUSTRE_IMP_DISCON:
  71         case LUSTRE_IMP_CONNECTING:
  72                 break;
  73         case LUSTRE_IMP_REPLAY_WAIT:
  74                 imp->imp_replay_state = LUSTRE_IMP_REPLAY_LOCKS;
  75                 break;
  76         default:
  77                 imp->imp_replay_state = LUSTRE_IMP_REPLAY;
  78         }
  79         imp->imp_state = state;
  80         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
  81         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
  82                 cfs_time_current_sec();
  83         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
  84                 IMP_STATE_HIST_LEN;
  85 }
  86
  87 /* A CLOSED import should remain so. */
  88 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  89 do {                                                                           \
  90         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  91                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  92                       imp, obd2cli_tgt(imp->imp_obd),                          \
  93                       ptlrpc_import_state_name(imp->imp_state),                \
  94                       ptlrpc_import_state_name(state));                        \
  95                __import_set_state(imp, state);                                 \
  96         }                                                                      \
  97 } while(0)
  98
  99 #define IMPORT_SET_STATE(imp, state)                                    \
 100 do {                                                                    \
 101         spin_lock(&imp->imp_lock);                                      \
 102         IMPORT_SET_STATE_NOLOCK(imp, state);                            \
 103         spin_unlock(&imp->imp_lock);                                    \
 104 } while(0)
 105
 106 void ptlrpc_import_enter_resend(struct obd_import *imp)
 107 {
 108         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 109 }
 110 EXPORT_SYMBOL(ptlrpc_import_enter_resend);
 111
 112
 113 static int ptlrpc_connect_interpret(const struct lu_env *env,
 114                                     struct ptlrpc_request *request,
 115                                     void * data, int rc);
 116 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
 117
 118 /* Only this function is allowed to change the import state when it is
 119  * CLOSED. I would rather refcount the import and free it after
 120  * disconnection like we do with exports. To do that, the client_obd
 121  * will need to save the peer info somewhere other than in the import,
 122  * though. */
 123 int ptlrpc_init_import(struct obd_import *imp)
 124 {
 125         spin_lock(&imp->imp_lock);
 126
 127         imp->imp_generation++;
 128         imp->imp_state =  LUSTRE_IMP_NEW;
 129
 130         spin_unlock(&imp->imp_lock);
 131
 132         return 0;
 133 }
 134 EXPORT_SYMBOL(ptlrpc_init_import);
 135
 136 #define UUID_STR "_UUID"
 137 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
 138 {
 139         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 140                 ? uuid : uuid + strlen(prefix);
 141
 142         *uuid_len = strlen(*uuid_start);
 143
 144         if (*uuid_len < strlen(UUID_STR))
 145                 return;
 146
 147         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 148                     UUID_STR, strlen(UUID_STR)))
 149                 *uuid_len -= strlen(UUID_STR);
 150 }
 151
 152 /**
 153  * Returns true if import was FULL, false if import was already not
 154  * connected.
 155  * @imp - import to be disconnected
 156  * @conn_cnt - connection count (epoch) of the request that timed out
 157  *             and caused the disconnection.  In some cases, multiple
 158  *             inflight requests can fail to a single target (e.g. OST
 159  *             bulk requests) and if one has already caused a reconnection
 160  *             (increasing the import->conn_cnt) the older failure should
 161  *             not also cause a reconnection.  If zero it forces a reconnect.
 162  */
 163 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 164 {
 165         int rc = 0;
 166
 167         spin_lock(&imp->imp_lock);
 168
 169         if (imp->imp_state == LUSTRE_IMP_FULL &&
 170             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 171                 char *target_start;
 172                 int   target_len;
 173
 174                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 175                           &target_start, &target_len);
 176
 177                 if (imp->imp_replayable) {
 178                         LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
 179                                "lost; in progress operations using this "
 180                                "service will wait for recovery to complete\n",
 181                                imp->imp_obd->obd_name, target_len, target_start,
 182                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 183                 } else {
 184                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
 185                                "%.*s (at %s) was lost; in progress "
 186                                "operations using this service will fail\n",
 187                                imp->imp_obd->obd_name,
 188                                target_len, target_start,
 189                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 190                 }
 191                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 192                 spin_unlock(&imp->imp_lock);
 193
 194                 if (obd_dump_on_timeout)
 195                         libcfs_debug_dumplog();
 196
 197                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 198                 rc = 1;
 199         } else {
 200                 spin_unlock(&imp->imp_lock);
 201                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 202                        imp->imp_client->cli_name, imp,
 203                        (imp->imp_state == LUSTRE_IMP_FULL &&
 204                         imp->imp_conn_cnt > conn_cnt) ?
 205                        "reconnected" : "not connected", imp->imp_conn_cnt,
 206                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 207         }
 208
 209         return rc;
 210 }
 211
 212 /* Must be called with imp_lock held! */
 213 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 214 {
 215         ENTRY;
 216         assert_spin_locked(&imp->imp_lock);
 217
 218         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 219         imp->imp_invalid = 1;
 220         imp->imp_generation++;
 221         spin_unlock(&imp->imp_lock);
 222
 223         ptlrpc_abort_inflight(imp);
 224         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 225
 226         EXIT;
 227 }
 228
 229 /*
 230  * This acts as a barrier; all existing requests are rejected, and
 231  * no new requests will be accepted until the import is valid again.
 232  */
 233 void ptlrpc_deactivate_import(struct obd_import *imp)
 234 {
 235         spin_lock(&imp->imp_lock);
 236         ptlrpc_deactivate_and_unlock_import(imp);
 237 }
 238 EXPORT_SYMBOL(ptlrpc_deactivate_import);
 239
 240 static unsigned int
 241 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
 242 {
 243         long dl;
 244
 245         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
 246               (req->rq_phase == RQ_PHASE_BULK) ||
 247               (req->rq_phase == RQ_PHASE_NEW)))
 248                 return 0;
 249
 250         if (req->rq_timedout)
 251                 return 0;
 252
 253         if (req->rq_phase == RQ_PHASE_NEW)
 254                 dl = req->rq_sent;
 255         else
 256                 dl = req->rq_deadline;
 257
 258         if (dl <= now)
 259                 return 0;
 260
 261         return dl - now;
 262 }
 263
 264 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
 265 {
 266         time_t now = cfs_time_current_sec();
 267         struct list_head *tmp, *n;
 268         struct ptlrpc_request *req;
 269         unsigned int timeout = 0;
 270
 271         spin_lock(&imp->imp_lock);
 272         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 273                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 274                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 275         }
 276         spin_unlock(&imp->imp_lock);
 277         return timeout;
 278 }
 279
 280 /**
 281  * This function will invalidate the import, if necessary, then block
 282  * for all the RPC completions, and finally notify the obd to
 283  * invalidate its state (ie cancel locks, clear pending requests,
 284  * etc).
 285  */
 286 void ptlrpc_invalidate_import(struct obd_import *imp)
 287 {
 288         struct list_head *tmp, *n;
 289         struct ptlrpc_request *req;
 290         struct l_wait_info lwi;
 291         unsigned int timeout;
 292         int rc;
 293
 294         atomic_inc(&imp->imp_inval_count);
 295
 296         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 297                 ptlrpc_deactivate_import(imp);
 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                 /* Calculate max timeout for waiting on rpcs to error
 308                  * out. Use obd_timeout if calculated value is smaller
 309                  * than it. */
 310                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 311                         timeout = ptlrpc_inflight_timeout(imp);
 312                         timeout += timeout / 3;
 313
 314                         if (timeout == 0)
 315                                 timeout = obd_timeout;
 316                 } else {
 317                         /* decrease the interval to increase race condition */
 318                         timeout = 1;
 319                 }
 320
 321                 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
 322                        timeout);
 323
 324                 /* Wait for all requests to error out and call completion
 325                  * callbacks. Cap it at obd_timeout -- these should all
 326                  * have been locally cancelled by ptlrpc_abort_inflight. */
 327                 lwi = LWI_TIMEOUT_INTERVAL(
 328                         cfs_timeout_cap(cfs_time_seconds(timeout)),
 329                         (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
 330                         NULL, NULL);
 331                 rc = l_wait_event(imp->imp_recovery_waitq,
 332                                   (atomic_read(&imp->imp_inflight) == 0),
 333                                   &lwi);
 334                 if (rc) {
 335                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 336
 337                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 338                                cli_tgt, rc, atomic_read(&imp->imp_inflight));
 339
 340                         spin_lock(&imp->imp_lock);
 341                         if (atomic_read(&imp->imp_inflight) == 0) {
 342                                 int count = atomic_read(&imp->imp_unregistering);
 343
 344                                 /* We know that "unregistering" rpcs only can
 345                                  * survive in sending or delaying lists (they
 346                                  * maybe waiting for long reply unlink in
 347                                  * sluggish nets). Let's check this. If there
 348                                  * is no inflight and unregistering != 0, this
 349                                  * is bug. */
 350                                 LASSERTF(count == 0, "Some RPCs are still "
 351                                          "unregistering: %d\n", count);
 352
 353                                 /* Let's save one loop as soon as inflight have
 354                                  * dropped to zero. No new inflights possible at
 355                                  * this point. */
 356                                 rc = 0;
 357                         } else {
 358                                 list_for_each_safe(tmp, n,
 359                                                    &imp->imp_sending_list) {
 360                                         req = list_entry(tmp,
 361                                                          struct ptlrpc_request,
 362                                                          rq_list);
 363                                         DEBUG_REQ(D_ERROR, req,
 364                                                   "still on sending list");
 365                                 }
 366                                 list_for_each_safe(tmp, n,
 367                                                    &imp->imp_delayed_list) {
 368                                         req = list_entry(tmp,
 369                                                          struct ptlrpc_request,
 370                                                          rq_list);
 371                                         DEBUG_REQ(D_ERROR, req,
 372                                                   "still on delayed list");
 373                                 }
 374
 375                                 CERROR("%s: Unregistering RPCs found (%d). "
 376                                        "Network is sluggish? Waiting them "
 377                                        "to error out.\n", cli_tgt,
 378                                        atomic_read(&imp->imp_unregistering));
 379                         }
 380                         spin_unlock(&imp->imp_lock);
 381                 }
 382         } while (rc != 0);
 383
 384         /*
 385          * Let's additionally check that no new rpcs added to import in
 386          * "invalidate" state.
 387          */
 388         LASSERT(atomic_read(&imp->imp_inflight) == 0);
 389         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 390         sptlrpc_import_flush_all_ctx(imp);
 391
 392         atomic_dec(&imp->imp_inval_count);
 393         wake_up_all(&imp->imp_recovery_waitq);
 394 }
 395 EXPORT_SYMBOL(ptlrpc_invalidate_import);
 396
 397 /* unset imp_invalid */
 398 void ptlrpc_activate_import(struct obd_import *imp)
 399 {
 400         struct obd_device *obd = imp->imp_obd;
 401
 402         spin_lock(&imp->imp_lock);
 403         if (imp->imp_deactive != 0) {
 404                 spin_unlock(&imp->imp_lock);
 405                 return;
 406         }
 407
 408         imp->imp_invalid = 0;
 409         spin_unlock(&imp->imp_lock);
 410         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 411 }
 412 EXPORT_SYMBOL(ptlrpc_activate_import);
 413
 414 void ptlrpc_pinger_force(struct obd_import *imp)
 415 {
 416         CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
 417                ptlrpc_import_state_name(imp->imp_state));
 418
 419         spin_lock(&imp->imp_lock);
 420         imp->imp_force_verify = 1;
 421         spin_unlock(&imp->imp_lock);
 422
 423         if (imp->imp_state != LUSTRE_IMP_CONNECTING)
 424                 ptlrpc_pinger_wake_up();
 425 }
 426 EXPORT_SYMBOL(ptlrpc_pinger_force);
 427
 428 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 429 {
 430         ENTRY;
 431
 432         LASSERT(!imp->imp_dlm_fake);
 433
 434         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 435                 if (!imp->imp_replayable) {
 436                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 437                                "auto-deactivating\n",
 438                                obd2cli_tgt(imp->imp_obd),
 439                                imp->imp_connection->c_remote_uuid.uuid,
 440                                imp->imp_obd->obd_name);
 441                         ptlrpc_deactivate_import(imp);
 442                 }
 443
 444                 ptlrpc_pinger_force(imp);
 445         }
 446         EXIT;
 447 }
 448
 449 int ptlrpc_reconnect_import(struct obd_import *imp)
 450 {
 451 #ifdef ENABLE_PINGER
 452         struct l_wait_info lwi;
 453         int secs = cfs_time_seconds(obd_timeout);
 454         int rc;
 455
 456         ptlrpc_pinger_force(imp);
 457
 458         CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 459                obd2cli_tgt(imp->imp_obd), secs);
 460
 461         lwi = LWI_TIMEOUT(secs, NULL, NULL);
 462         rc = l_wait_event(imp->imp_recovery_waitq,
 463                           !ptlrpc_import_in_recovery(imp), &lwi);
 464         CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
 465                ptlrpc_import_state_name(imp->imp_state));
 466         return rc;
 467 #else
 468         ptlrpc_set_import_discon(imp, 0);
 469         /* Force a new connect attempt */
 470         ptlrpc_invalidate_import(imp);
 471         /* Do a fresh connect next time by zeroing the handle */
 472         ptlrpc_disconnect_import(imp, 1);
 473         /* Wait for all invalidate calls to finish */
 474         if (atomic_read(&imp->imp_inval_count) > 0) {
 475                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 476                 int rc;
 477
 478                 rc = l_wait_event(imp->imp_recovery_waitq,
 479                                   (atomic_read(&imp->imp_inval_count) == 0),
 480                                   &lwi);
 481                 if (rc)
 482                         CERROR("Interrupted, inval=%d\n",
 483                                atomic_read(&imp->imp_inval_count));
 484         }
 485
 486         /* Allow reconnect attempts */
 487         imp->imp_obd->obd_no_recov = 0;
 488         /* Remove 'invalid' flag */
 489         ptlrpc_activate_import(imp);
 490         /* Attempt a new connect */
 491         ptlrpc_recover_import(imp, NULL, 0);
 492         return 0;
 493 #endif
 494 }
 495 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 496
 497 /**
 498  * Connection on import \a imp is changed to another one (if more than one is
 499  * present). We typically chose connection that we have not tried to connect to
 500  * the longest
 501  */
 502 static int import_select_connection(struct obd_import *imp)
 503 {
 504         struct obd_import_conn *imp_conn = NULL, *conn;
 505         struct obd_export *dlmexp;
 506         char *target_start;
 507         int target_len, tried_all = 1;
 508         ENTRY;
 509
 510         spin_lock(&imp->imp_lock);
 511
 512         if (list_empty(&imp->imp_conn_list)) {
 513                 CERROR("%s: no connections available\n",
 514                        imp->imp_obd->obd_name);
 515                 spin_unlock(&imp->imp_lock);
 516                 RETURN(-EINVAL);
 517         }
 518
 519         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 520                 CDEBUG(D_HA, "%s: connect to NID %s last attempt %llu\n",
 521                        imp->imp_obd->obd_name,
 522                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 523                        conn->oic_last_attempt);
 524
 525                 /* If we have not tried this connection since
 526                    the last successful attempt, go with this one */
 527                 if ((conn->oic_last_attempt == 0) ||
 528                     cfs_time_beforeq_64(conn->oic_last_attempt,
 529                                        imp->imp_last_success_conn)) {
 530                         imp_conn = conn;
 531                         tried_all = 0;
 532                         break;
 533                 }
 534
 535                 /* If all of the connections have already been tried
 536                    since the last successful connection; just choose the
 537                    least recently used */
 538                 if (!imp_conn)
 539                         imp_conn = conn;
 540                 else if (cfs_time_before_64(conn->oic_last_attempt,
 541                                             imp_conn->oic_last_attempt))
 542                         imp_conn = conn;
 543         }
 544
 545         /* if not found, simply choose the current one */
 546         if (!imp_conn || imp->imp_force_reconnect) {
 547                 LASSERT(imp->imp_conn_current);
 548                 imp_conn = imp->imp_conn_current;
 549                 tried_all = 0;
 550         }
 551         LASSERT(imp_conn->oic_conn);
 552
 553         /* If we've tried everything, and we're back to the beginning of the
 554            list, increase our timeout and try again. It will be reset when
 555            we do finally connect. (FIXME: really we should wait for all network
 556            state associated with the last connection attempt to drain before
 557            trying to reconnect on it.) */
 558         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 559                 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
 560                 if (at_get(at) < CONNECTION_SWITCH_MAX) {
 561                         at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
 562                         if (at_get(at) > CONNECTION_SWITCH_MAX)
 563                                 at_reset(at, CONNECTION_SWITCH_MAX);
 564                 }
 565                 LASSERT(imp_conn->oic_last_attempt);
 566                 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
 567                         "to %ds\n", imp->imp_obd->obd_name, at_get(at));
 568         }
 569
 570         imp_conn->oic_last_attempt = cfs_time_current_64();
 571
 572         /* switch connection, don't mind if it's same as the current one */
 573         if (imp->imp_connection)
 574                 ptlrpc_connection_put(imp->imp_connection);
 575         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 576
 577         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 578         LASSERT(dlmexp != NULL);
 579         if (dlmexp->exp_connection)
 580                 ptlrpc_connection_put(dlmexp->exp_connection);
 581         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 582         class_export_put(dlmexp);
 583
 584         if (imp->imp_conn_current != imp_conn) {
 585                 if (imp->imp_conn_current) {
 586                         deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 587                                   &target_start, &target_len);
 588
 589                         CDEBUG(D_HA, "%s: Connection changing to"
 590                                " %.*s (at %s)\n",
 591                                imp->imp_obd->obd_name,
 592                                target_len, target_start,
 593                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 594                 }
 595
 596                 imp->imp_conn_current = imp_conn;
 597         }
 598
 599         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 600                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 601                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 602
 603         spin_unlock(&imp->imp_lock);
 604
 605         RETURN(0);
 606 }
 607
 608 /*
 609  * must be called under imp_lock
 610  */
 611 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 612 {
 613         struct ptlrpc_request   *req;
 614         struct list_head        *tmp;
 615
 616         /* The requests in committed_list always have smaller transnos than
 617          * the requests in replay_list */
 618         if (!list_empty(&imp->imp_committed_list)) {
 619                 tmp = imp->imp_committed_list.next;
 620                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 621                 *transno = req->rq_transno;
 622                 if (req->rq_transno == 0) {
 623                         DEBUG_REQ(D_ERROR, req, "zero transno in committed_list");
 624                         LBUG();
 625                 }
 626                 return 1;
 627         }
 628         if (!list_empty(&imp->imp_replay_list)) {
 629                 tmp = imp->imp_replay_list.next;
 630                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 631                 *transno = req->rq_transno;
 632                 if (req->rq_transno == 0) {
 633                         DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
 634                         LBUG();
 635                 }
 636                 return 1;
 637         }
 638         return 0;
 639 }
 640
 641 /**
 642  * Attempt to (re)connect import \a imp. This includes all preparations,
 643  * initializing CONNECT RPC request and passing it to ptlrpcd for
 644  * actual sending.
 645  * Returns 0 on success or error code.
 646  */
 647 int ptlrpc_connect_import(struct obd_import *imp)
 648 {
 649         struct obd_device *obd = imp->imp_obd;
 650         int initial_connect = 0;
 651         int set_transno = 0;
 652         __u64 committed_before_reconnect = 0;
 653         struct ptlrpc_request *request;
 654         char *bufs[] = { NULL,
 655                          obd2cli_tgt(imp->imp_obd),
 656                          obd->obd_uuid.uuid,
 657                          (char *)&imp->imp_dlm_handle,
 658                          (char *)&imp->imp_connect_data };
 659         struct ptlrpc_connect_async_args *aa;
 660         int rc;
 661         ENTRY;
 662
 663         spin_lock(&imp->imp_lock);
 664         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 665                 spin_unlock(&imp->imp_lock);
 666                 CERROR("can't connect to a closed import\n");
 667                 RETURN(-EINVAL);
 668         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 669                 spin_unlock(&imp->imp_lock);
 670                 CERROR("already connected\n");
 671                 RETURN(0);
 672         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING ||
 673                    imp->imp_connected) {
 674                 spin_unlock(&imp->imp_lock);
 675                 CERROR("already connecting\n");
 676                 RETURN(-EALREADY);
 677         }
 678
 679         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 680
 681         imp->imp_conn_cnt++;
 682         imp->imp_resend_replay = 0;
 683
 684         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 685                 initial_connect = 1;
 686         else
 687                 committed_before_reconnect = imp->imp_peer_committed_transno;
 688
 689         set_transno = ptlrpc_first_transno(imp,
 690                                            &imp->imp_connect_data.ocd_transno);
 691         spin_unlock(&imp->imp_lock);
 692
 693         rc = import_select_connection(imp);
 694         if (rc)
 695                 GOTO(out, rc);
 696
 697         rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
 698         if (rc)
 699                 GOTO(out, rc);
 700
 701         /* Reset connect flags to the originally requested flags, in case
 702          * the server is updated on-the-fly we will get the new features. */
 703         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 704         imp->imp_connect_data.ocd_connect_flags2 = imp->imp_connect_flags2_orig;
 705         /* Reset ocd_version each time so the server knows the exact versions */
 706         imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
 707         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 708         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 709
 710         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 711                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 712         if (rc)
 713                 GOTO(out, rc);
 714
 715         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 716         if (request == NULL)
 717                 GOTO(out, rc = -ENOMEM);
 718
 719         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 720                                       imp->imp_connect_op, bufs, NULL);
 721         if (rc) {
 722                 ptlrpc_request_free(request);
 723                 GOTO(out, rc);
 724         }
 725
 726         /* Report the rpc service time to the server so that it knows how long
 727          * to wait for clients to join recovery */
 728         lustre_msg_set_service_time(request->rq_reqmsg,
 729                                     at_timeout2est(request->rq_timeout));
 730
 731         /* The amount of time we give the server to process the connect req.
 732          * import_select_connection will increase the net latency on
 733          * repeated reconnect attempts to cover slow networks.
 734          * We override/ignore the server rpc completion estimate here,
 735          * which may be large if this is a reconnect attempt */
 736         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 737         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 738
 739         request->rq_no_resend = request->rq_no_delay = 1;
 740         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 741         /* Allow a slightly larger reply for future growth compatibility */
 742         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 743                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 744         ptlrpc_request_set_replen(request);
 745         request->rq_interpret_reply = ptlrpc_connect_interpret;
 746
 747         CLASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
 748         aa = ptlrpc_req_async_args(request);
 749         memset(aa, 0, sizeof *aa);
 750
 751         aa->pcaa_peer_committed = committed_before_reconnect;
 752         aa->pcaa_initial_connect = initial_connect;
 753
 754         if (aa->pcaa_initial_connect) {
 755                 spin_lock(&imp->imp_lock);
 756                 imp->imp_replayable = 1;
 757                 spin_unlock(&imp->imp_lock);
 758                 lustre_msg_add_op_flags(request->rq_reqmsg,
 759                                         MSG_CONNECT_INITIAL);
 760         }
 761
 762         if (set_transno)
 763                 lustre_msg_add_op_flags(request->rq_reqmsg,
 764                                         MSG_CONNECT_TRANSNO);
 765
 766         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
 767                   request->rq_timeout);
 768         ptlrpcd_add_req(request);
 769         rc = 0;
 770 out:
 771         if (rc != 0)
 772                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 773
 774         RETURN(rc);
 775 }
 776 EXPORT_SYMBOL(ptlrpc_connect_import);
 777
 778 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 779 {
 780         int force_verify;
 781
 782         spin_lock(&imp->imp_lock);
 783         force_verify = imp->imp_force_verify != 0;
 784         spin_unlock(&imp->imp_lock);
 785
 786         if (force_verify)
 787                 ptlrpc_pinger_wake_up();
 788 }
 789
 790 static int ptlrpc_busy_reconnect(int rc)
 791 {
 792         return (rc == -EBUSY) || (rc == -EAGAIN);
 793 }
 794
 795 static int ptlrpc_connect_set_flags(struct obd_import *imp,
 796                                      struct obd_connect_data *ocd,
 797                                      __u64 old_connect_flags,
 798                                      struct obd_export *exp, int init_connect)
 799 {
 800         static bool warned;
 801         struct client_obd *cli = &imp->imp_obd->u.cli;
 802
 803         if ((imp->imp_connect_flags_orig & OBD_CONNECT_IBITS) &&
 804             !(ocd->ocd_connect_flags & OBD_CONNECT_IBITS)) {
 805                 LCONSOLE_WARN("%s: MDS %s does not support ibits "
 806                               "lock, either very old or invalid: "
 807                               "requested %#llx, replied %#llx\n",
 808                               imp->imp_obd->obd_name,
 809                               imp->imp_connection->c_remote_uuid.uuid,
 810                               imp->imp_connect_flags_orig,
 811                               ocd->ocd_connect_flags);
 812                 return -EPROTO;
 813         }
 814
 815         spin_lock(&imp->imp_lock);
 816         list_del(&imp->imp_conn_current->oic_item);
 817         list_add(&imp->imp_conn_current->oic_item,
 818                  &imp->imp_conn_list);
 819         imp->imp_last_success_conn =
 820                 imp->imp_conn_current->oic_last_attempt;
 821
 822         spin_unlock(&imp->imp_lock);
 823
 824
 825         if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 826             (ocd->ocd_version > LUSTRE_VERSION_CODE +
 827                                 LUSTRE_VERSION_OFFSET_WARN ||
 828              ocd->ocd_version < LUSTRE_VERSION_CODE -
 829                                 LUSTRE_VERSION_OFFSET_WARN)) {
 830                 /* Sigh, some compilers do not like #ifdef in the middle
 831                    of macro arguments */
 832                 const char *older = "older than client. "
 833                                     "Consider upgrading server";
 834                 const char *newer = "newer than client. "
 835                                     "Consider recompiling application";
 836
 837                 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 838                               "is much %s (%s)\n",
 839                               obd2cli_tgt(imp->imp_obd),
 840                               OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 841                               OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 842                               OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 843                               OBD_OCD_VERSION_FIX(ocd->ocd_version),
 844                               ocd->ocd_version > LUSTRE_VERSION_CODE ?
 845                               newer : older, LUSTRE_VERSION_STRING);
 846                 warned = true;
 847         }
 848
 849 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
 850         /* Check if server has LU-1252 fix applied to not always swab
 851          * the IR MNE entries. Do this only once per connection.  This
 852          * fixup is version-limited, because we don't want to carry the
 853          * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
 854          * need interop with unpatched 2.2 servers.  For newer servers,
 855          * the client will do MNE swabbing only as needed.  LU-1644 */
 856         if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 857                      !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
 858                      OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
 859                      OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
 860                      OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
 861                      strcmp(imp->imp_obd->obd_type->typ_name,
 862                             LUSTRE_MGC_NAME) == 0))
 863                 imp->imp_need_mne_swab = 1;
 864         else /* clear if server was upgraded since last connect */
 865                 imp->imp_need_mne_swab = 0;
 866 #endif
 867
 868         if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
 869                 /* We sent to the server ocd_cksum_types with bits set
 870                  * for algorithms we understand. The server masked off
 871                  * the checksum types it doesn't support */
 872                 if ((ocd->ocd_cksum_types &
 873                      cksum_types_supported_client()) == 0) {
 874                         LCONSOLE_WARN("The negotiation of the checksum "
 875                                       "alogrithm to use with server %s "
 876                                       "failed (%x/%x), disabling "
 877                                       "checksums\n",
 878                                       obd2cli_tgt(imp->imp_obd),
 879                                       ocd->ocd_cksum_types,
 880                                       cksum_types_supported_client());
 881                         cli->cl_checksum = 0;
 882                         cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
 883                 } else {
 884                         cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
 885                 }
 886         } else {
 887                 /* The server does not support OBD_CONNECT_CKSUM.
 888                  * Enforce ADLER for backward compatibility*/
 889                 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
 890         }
 891         cli->cl_cksum_type = cksum_type_select(cli->cl_supp_cksum_types);
 892
 893         if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
 894                 cli->cl_max_pages_per_rpc =
 895                         min(ocd->ocd_brw_size >> PAGE_CACHE_SHIFT,
 896                             cli->cl_max_pages_per_rpc);
 897         else if (imp->imp_connect_op == MDS_CONNECT ||
 898                  imp->imp_connect_op == MGS_CONNECT)
 899                 cli->cl_max_pages_per_rpc = 1;
 900
 901         LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 902                 (cli->cl_max_pages_per_rpc > 0));
 903
 904         client_adjust_max_dirty(cli);
 905
 906         /* Update client max modify RPCs in flight with value returned
 907          * by the server */
 908         if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
 909                 cli->cl_max_mod_rpcs_in_flight = min(
 910                                         cli->cl_max_mod_rpcs_in_flight,
 911                                         ocd->ocd_maxmodrpcs);
 912         else
 913                 cli->cl_max_mod_rpcs_in_flight = 1;
 914
 915         /* Reset ns_connect_flags only for initial connect. It might be
 916          * changed in while using FS and if we reset it in reconnect
 917          * this leads to losing user settings done before such as
 918          * disable lru_resize, etc. */
 919         if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
 920                 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
 921                              "flags: %#llx\n", imp->imp_obd->obd_name,
 922                              ocd->ocd_connect_flags);
 923                 imp->imp_obd->obd_namespace->ns_connect_flags =
 924                         ocd->ocd_connect_flags;
 925                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
 926                         ocd->ocd_connect_flags;
 927         }
 928
 929         if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
 930                 /* We need a per-message support flag, because
 931                  * a. we don't know if the incoming connect reply
 932                  *    supports AT or not (in reply_in_callback)
 933                  *    until we unpack it.
 934                  * b. failovered server means export and flags are gone
 935                  *    (in ptlrpc_send_reply).
 936                  *    Can only be set when we know AT is supported at
 937                  *    both ends */
 938                 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 939         else
 940                 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 941
 942         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 943
 944         return 0;
 945 }
 946
 947 /**
 948  * Add all replay requests back to unreplied list before start replay,
 949  * so that we can make sure the known replied XID is always increased
 950  * only even if when replaying requests.
 951  */
 952 static void ptlrpc_prepare_replay(struct obd_import *imp)
 953 {
 954         struct ptlrpc_request *req;
 955
 956         if (imp->imp_state != LUSTRE_IMP_REPLAY ||
 957             imp->imp_resend_replay)
 958                 return;
 959
 960         /* If the server was restart during repaly, the requests may
 961          * have been added to the unreplied list in former replay. */
 962         spin_lock(&imp->imp_lock);
 963
 964         list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
 965                 if (list_empty(&req->rq_unreplied_list))
 966                         ptlrpc_add_unreplied(req);
 967         }
 968
 969         list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
 970                 if (list_empty(&req->rq_unreplied_list))
 971                         ptlrpc_add_unreplied(req);
 972         }
 973
 974         imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
 975         spin_unlock(&imp->imp_lock);
 976 }
 977
 978 /**
 979  * interpret_reply callback for connect RPCs.
 980  * Looks into returned status of connect operation and decides
 981  * what to do with the import - i.e enter recovery, promote it to
 982  * full state for normal operations of disconnect it due to an error.
 983  */
 984 static int ptlrpc_connect_interpret(const struct lu_env *env,
 985                                     struct ptlrpc_request *request,
 986                                     void *data, int rc)
 987 {
 988         struct ptlrpc_connect_async_args *aa = data;
 989         struct obd_import *imp = request->rq_import;
 990         struct lustre_handle old_hdl;
 991         __u64 old_connect_flags;
 992         int msg_flags;
 993         struct obd_connect_data *ocd;
 994         struct obd_export *exp;
 995         int ret;
 996         ENTRY;
 997
 998         spin_lock(&imp->imp_lock);
 999         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1000                 imp->imp_connect_tried = 1;
1001                 spin_unlock(&imp->imp_lock);
1002                 RETURN(0);
1003         }
1004
1005         if (rc) {
1006                 /* if this reconnect to busy export - not need select new target
1007                  * for connecting*/
1008                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
1009                 spin_unlock(&imp->imp_lock);
1010                 ptlrpc_maybe_ping_import_soon(imp);
1011                 GOTO(out, rc);
1012         }
1013
1014         /* LU-7558: indicate that we are interpretting connect reply,
1015          * pltrpc_connect_import() will not try to reconnect until
1016          * interpret will finish. */
1017         imp->imp_connected = 1;
1018         spin_unlock(&imp->imp_lock);
1019
1020         LASSERT(imp->imp_conn_current);
1021
1022         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1023
1024         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1025                                    RCL_SERVER);
1026         /* server replied obd_connect_data is always bigger */
1027         ocd = req_capsule_server_sized_get(&request->rq_pill,
1028                                            &RMF_CONNECT_DATA, ret);
1029
1030         if (ocd == NULL) {
1031                 CERROR("%s: no connect data from server\n",
1032                        imp->imp_obd->obd_name);
1033                 rc = -EPROTO;
1034                 GOTO(out, rc);
1035         }
1036
1037         spin_lock(&imp->imp_lock);
1038
1039         /* All imports are pingable */
1040         imp->imp_pingable = 1;
1041         imp->imp_force_reconnect = 0;
1042         imp->imp_force_verify = 0;
1043
1044         imp->imp_connect_data = *ocd;
1045
1046         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1047                imp->imp_obd->obd_name, ocd->ocd_instance);
1048         exp = class_conn2export(&imp->imp_dlm_handle);
1049
1050         spin_unlock(&imp->imp_lock);
1051
1052         /* check that server granted subset of flags we asked for. */
1053         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1054             ocd->ocd_connect_flags) {
1055                 CERROR("%s: Server didn't grant requested subset of flags: "
1056                        "asked=%#llx granted=%#llx\n",
1057                        imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
1058                        ocd->ocd_connect_flags);
1059                 GOTO(out, rc = -EPROTO);
1060         }
1061
1062         if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) !=
1063             ocd->ocd_connect_flags2) {
1064                 CERROR("%s: Server didn't grant requested subset of flags2: "
1065                        "asked=%#llx granted=%#llx\n",
1066                        imp->imp_obd->obd_name, imp->imp_connect_flags2_orig,
1067                        ocd->ocd_connect_flags2);
1068                 GOTO(out, rc = -EPROTO);
1069         }
1070
1071         if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1072             (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1073             (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1074             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1075                 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1076                 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1077                 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1078
1079                 /* We do not support the MDT-MDT interoperations with
1080                  * different version MDT because of protocol changes. */
1081                 if (unlikely(major != LUSTRE_MAJOR ||
1082                              minor != LUSTRE_MINOR ||
1083                              abs(patch - LUSTRE_PATCH) > 3)) {
1084                         LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1085                                       "connection to different version MDT "
1086                                       "(%d.%d.%d.%d) %s\n",
1087                                       imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1088                                       LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1089                                       major, minor, patch,
1090                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1091                                       imp->imp_connection->c_remote_uuid.uuid);
1092
1093                         GOTO(out, rc = -EPROTO);
1094                 }
1095         }
1096
1097         if (!exp) {
1098                 /* This could happen if export is cleaned during the
1099                    connect attempt */
1100                 CERROR("%s: missing export after connect\n",
1101                        imp->imp_obd->obd_name);
1102                 GOTO(out, rc = -ENODEV);
1103         }
1104         old_connect_flags = exp_connect_flags(exp);
1105         exp->exp_connect_data = *ocd;
1106         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1107
1108         /* The net statistics after (re-)connect is not valid anymore,
1109          * because may reflect other routing, etc. */
1110         at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1111         ptlrpc_at_adj_net_latency(request,
1112                         lustre_msg_get_service_time(request->rq_repmsg));
1113
1114         /* Import flags should be updated before waking import at FULL state */
1115         rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1116                                  aa->pcaa_initial_connect);
1117         class_export_put(exp);
1118         if (rc != 0)
1119                 GOTO(out, rc);
1120
1121         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1122
1123         if (aa->pcaa_initial_connect) {
1124                 spin_lock(&imp->imp_lock);
1125                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1126                         imp->imp_replayable = 1;
1127                         spin_unlock(&imp->imp_lock);
1128                         CDEBUG(D_HA, "connected to replayable target: %s\n",
1129                                obd2cli_tgt(imp->imp_obd));
1130                 } else {
1131                         imp->imp_replayable = 0;
1132                         spin_unlock(&imp->imp_lock);
1133                 }
1134
1135                 /* if applies, adjust the imp->imp_msg_magic here
1136                  * according to reply flags */
1137
1138                 imp->imp_remote_handle =
1139                                 *lustre_msg_get_handle(request->rq_repmsg);
1140
1141                 /* Initial connects are allowed for clients with non-random
1142                  * uuids when servers are in recovery.  Simply signal the
1143                  * servers replay is complete and wait in REPLAY_WAIT. */
1144                 if (msg_flags & MSG_CONNECT_RECOVERING) {
1145                         CDEBUG(D_HA, "connect to %s during recovery\n",
1146                                obd2cli_tgt(imp->imp_obd));
1147                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1148                 } else {
1149                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1150                         ptlrpc_activate_import(imp);
1151                 }
1152
1153                 GOTO(finish, rc = 0);
1154         }
1155
1156         /* Determine what recovery state to move the import to. */
1157         if (MSG_CONNECT_RECONNECT & msg_flags) {
1158                 memset(&old_hdl, 0, sizeof(old_hdl));
1159                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1160                             sizeof (old_hdl))) {
1161                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1162                                       "bad handle %#llx\n",
1163                                       obd2cli_tgt(imp->imp_obd),
1164                                       imp->imp_connection->c_remote_uuid.uuid,
1165                                       imp->imp_dlm_handle.cookie);
1166                         GOTO(out, rc = -ENOTCONN);
1167                 }
1168
1169                 if (memcmp(&imp->imp_remote_handle,
1170                            lustre_msg_get_handle(request->rq_repmsg),
1171                            sizeof(imp->imp_remote_handle))) {
1172                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
1173                                 D_HA : D_WARNING;
1174
1175                         /* Bug 16611/14775: if server handle have changed,
1176                          * that means some sort of disconnection happened.
1177                          * If the server is not in recovery, that also means it
1178                          * already erased all of our state because of previous
1179                          * eviction. If it is in recovery - we are safe to
1180                          * participate since we can reestablish all of our state
1181                          * with server again */
1182                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1183                                 CDEBUG(level,"%s@%s changed server handle from "
1184                                        "%#llx to %#llx"
1185                                        " but is still in recovery\n",
1186                                        obd2cli_tgt(imp->imp_obd),
1187                                        imp->imp_connection->c_remote_uuid.uuid,
1188                                        imp->imp_remote_handle.cookie,
1189                                        lustre_msg_get_handle(
1190                                        request->rq_repmsg)->cookie);
1191                         } else {
1192                                 LCONSOLE_WARN("Evicted from %s (at %s) "
1193                                               "after server handle changed from "
1194                                               "%#llx to %#llx\n",
1195                                               obd2cli_tgt(imp->imp_obd),
1196                                               imp->imp_connection-> \
1197                                               c_remote_uuid.uuid,
1198                                               imp->imp_remote_handle.cookie,
1199                                               lustre_msg_get_handle(
1200                                               request->rq_repmsg)->cookie);
1201                         }
1202
1203
1204                         imp->imp_remote_handle =
1205                                      *lustre_msg_get_handle(request->rq_repmsg);
1206
1207                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1208                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1209                                 GOTO(finish, rc = 0);
1210                         }
1211
1212                 } else {
1213                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1214                                obd2cli_tgt(imp->imp_obd),
1215                                imp->imp_connection->c_remote_uuid.uuid);
1216                 }
1217
1218                 if (imp->imp_invalid) {
1219                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1220                                "marking evicted\n", imp->imp_obd->obd_name);
1221                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1222                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1223                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1224                                imp->imp_obd->obd_name,
1225                                obd2cli_tgt(imp->imp_obd));
1226
1227                         spin_lock(&imp->imp_lock);
1228                         imp->imp_resend_replay = 1;
1229                         spin_unlock(&imp->imp_lock);
1230
1231                         IMPORT_SET_STATE(imp, imp->imp_replay_state);
1232                 } else {
1233                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1234                 }
1235         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1236                 LASSERT(imp->imp_replayable);
1237                 imp->imp_remote_handle =
1238                                 *lustre_msg_get_handle(request->rq_repmsg);
1239                 imp->imp_last_replay_transno = 0;
1240                 imp->imp_replay_cursor = &imp->imp_committed_list;
1241                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
1242         } else {
1243                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
1244                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
1245                 imp->imp_remote_handle =
1246                                 *lustre_msg_get_handle(request->rq_repmsg);
1247                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1248         }
1249
1250         /* Sanity checks for a reconnected import. */
1251         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
1252                 CERROR("imp_replayable flag does not match server "
1253                        "after reconnect. We should LBUG right here.\n");
1254         }
1255
1256         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1257             lustre_msg_get_last_committed(request->rq_repmsg) <
1258             aa->pcaa_peer_committed) {
1259                 CERROR("%s went back in time (transno %lld"
1260                        " was previously committed, server now claims %lld"
1261                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
1262                        "id=9646\n",
1263                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1264                        lustre_msg_get_last_committed(request->rq_repmsg));
1265         }
1266
1267 finish:
1268         ptlrpc_prepare_replay(imp);
1269         rc = ptlrpc_import_recovery_state_machine(imp);
1270         if (rc == -ENOTCONN) {
1271                 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1272                        "invalidating and reconnecting\n",
1273                        obd2cli_tgt(imp->imp_obd),
1274                        imp->imp_connection->c_remote_uuid.uuid);
1275                 ptlrpc_connect_import(imp);
1276                 spin_lock(&imp->imp_lock);
1277                 imp->imp_connected = 0;
1278                 imp->imp_connect_tried = 1;
1279                 spin_unlock(&imp->imp_lock);
1280                 RETURN(0);
1281         }
1282
1283 out:
1284         spin_lock(&imp->imp_lock);
1285         imp->imp_connected = 0;
1286         imp->imp_connect_tried = 1;
1287         spin_unlock(&imp->imp_lock);
1288
1289         if (rc != 0) {
1290                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1291                 if (rc == -EACCES) {
1292                         /*
1293                          * Give up trying to reconnect
1294                          * EACCES means client has no permission for connection
1295                          */
1296                         imp->imp_obd->obd_no_recov = 1;
1297                         ptlrpc_deactivate_import(imp);
1298                 }
1299
1300                 if (rc == -EPROTO) {
1301                         struct obd_connect_data *ocd;
1302
1303                         /* reply message might not be ready */
1304                         if (request->rq_repmsg == NULL)
1305                                 RETURN(-EPROTO);
1306
1307                         ocd = req_capsule_server_get(&request->rq_pill,
1308                                                      &RMF_CONNECT_DATA);
1309                         if (ocd &&
1310                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1311                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1312                            /* Actually servers are only supposed to refuse
1313                               connection from liblustre clients, so we should
1314                               never see this from VFS context */
1315                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1316                                         "(%d.%d.%d.%d)"
1317                                         " refused connection from this client "
1318                                         "with an incompatible version (%s).  "
1319                                         "Client must be recompiled\n",
1320                                         obd2cli_tgt(imp->imp_obd),
1321                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1322                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1323                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1324                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1325                                         LUSTRE_VERSION_STRING);
1326                                 ptlrpc_deactivate_import(imp);
1327                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1328                         }
1329                         RETURN(-EPROTO);
1330                 }
1331
1332                 ptlrpc_maybe_ping_import_soon(imp);
1333
1334                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1335                        obd2cli_tgt(imp->imp_obd),
1336                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1337         }
1338
1339         wake_up_all(&imp->imp_recovery_waitq);
1340         RETURN(rc);
1341 }
1342
1343 /**
1344  * interpret callback for "completed replay" RPCs.
1345  * \see signal_completed_replay
1346  */
1347 static int completed_replay_interpret(const struct lu_env *env,
1348                                       struct ptlrpc_request *req,
1349                                       void * data, int rc)
1350 {
1351         ENTRY;
1352         atomic_dec(&req->rq_import->imp_replay_inflight);
1353         if (req->rq_status == 0 &&
1354             !req->rq_import->imp_vbr_failed) {
1355                 ptlrpc_import_recovery_state_machine(req->rq_import);
1356         } else {
1357                 if (req->rq_import->imp_vbr_failed) {
1358                         CDEBUG(D_WARNING,
1359                                "%s: version recovery fails, reconnecting\n",
1360                                req->rq_import->imp_obd->obd_name);
1361                 } else {
1362                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1363                                      "reconnecting\n",
1364                                req->rq_import->imp_obd->obd_name,
1365                                req->rq_status);
1366                 }
1367                 ptlrpc_connect_import(req->rq_import);
1368         }
1369
1370         RETURN(0);
1371 }
1372
1373 /**
1374  * Let server know that we have no requests to replay anymore.
1375  * Achieved by just sending a PING request
1376  */
1377 static int signal_completed_replay(struct obd_import *imp)
1378 {
1379         struct ptlrpc_request *req;
1380         ENTRY;
1381
1382         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1383                 RETURN(0);
1384
1385         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1386         atomic_inc(&imp->imp_replay_inflight);
1387
1388         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1389                                         OBD_PING);
1390         if (req == NULL) {
1391                 atomic_dec(&imp->imp_replay_inflight);
1392                 RETURN(-ENOMEM);
1393         }
1394
1395         ptlrpc_request_set_replen(req);
1396         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1397         lustre_msg_add_flags(req->rq_reqmsg,
1398                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1399         if (AT_OFF)
1400                 req->rq_timeout *= 3;
1401         req->rq_interpret_reply = completed_replay_interpret;
1402
1403         ptlrpcd_add_req(req);
1404         RETURN(0);
1405 }
1406
1407 /**
1408  * In kernel code all import invalidation happens in its own
1409  * separate thread, so that whatever application happened to encounter
1410  * a problem could still be killed or otherwise continue
1411  */
1412 static int ptlrpc_invalidate_import_thread(void *data)
1413 {
1414         struct obd_import *imp = data;
1415
1416         ENTRY;
1417
1418         unshare_fs_struct();
1419
1420         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1421                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1422                imp->imp_connection->c_remote_uuid.uuid);
1423
1424         ptlrpc_invalidate_import(imp);
1425
1426         if (obd_dump_on_eviction) {
1427                 CERROR("dump the log upon eviction\n");
1428                 libcfs_debug_dumplog();
1429         }
1430
1431         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1432         ptlrpc_import_recovery_state_machine(imp);
1433
1434         class_import_put(imp);
1435         RETURN(0);
1436 }
1437
1438 /**
1439  * This is the state machine for client-side recovery on import.
1440  *
1441  * Typicaly we have two possibly paths. If we came to server and it is not
1442  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1443  * state and reconnect from scratch.
1444  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1445  * We go through our list of requests to replay and send them to server one by
1446  * one.
1447  * After sending all request from the list we change import state to
1448  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1449  * and also all the locks we don't yet have and wait for server to grant us.
1450  * After that we send a special "replay completed" request and change import
1451  * state to IMP_REPLAY_WAIT.
1452  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1453  * state and resend all requests from sending list.
1454  * After that we promote import to FULL state and send all delayed requests
1455  * and import is fully operational after that.
1456  *
1457  */
1458 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1459 {
1460         int rc = 0;
1461         int inflight;
1462         char *target_start;
1463         int target_len;
1464
1465         ENTRY;
1466         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1467                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1468                           &target_start, &target_len);
1469                 /* Don't care about MGC eviction */
1470                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1471                            LUSTRE_MGC_NAME) != 0) {
1472                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1473                                            "by %.*s; in progress operations "
1474                                            "using this service will fail.\n",
1475                                            imp->imp_obd->obd_name, target_len,
1476                                            target_start);
1477                 }
1478                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1479                        obd2cli_tgt(imp->imp_obd),
1480                        imp->imp_connection->c_remote_uuid.uuid);
1481                 /* reset vbr_failed flag upon eviction */
1482                 spin_lock(&imp->imp_lock);
1483                 imp->imp_vbr_failed = 0;
1484                 spin_unlock(&imp->imp_lock);
1485
1486                 {
1487                 struct task_struct *task;
1488                 /* bug 17802:  XXX client_disconnect_export vs connect request
1489                  * race. if client is evicted at this time then we start
1490                  * invalidate thread without reference to import and import can
1491                  * be freed at same time. */
1492                 class_import_get(imp);
1493                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1494                                      "ll_imp_inval");
1495                 if (IS_ERR(task)) {
1496                         class_import_put(imp);
1497                         CERROR("error starting invalidate thread: %d\n", rc);
1498                         rc = PTR_ERR(task);
1499                 } else {
1500                         rc = 0;
1501                 }
1502                 RETURN(rc);
1503                 }
1504         }
1505
1506         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1507                 CDEBUG(D_HA, "replay requested by %s\n",
1508                        obd2cli_tgt(imp->imp_obd));
1509                 rc = ptlrpc_replay_next(imp, &inflight);
1510                 if (inflight == 0 &&
1511                     atomic_read(&imp->imp_replay_inflight) == 0) {
1512                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1513                         rc = ldlm_replay_locks(imp);
1514                         if (rc)
1515                                 GOTO(out, rc);
1516                 }
1517                 rc = 0;
1518         }
1519
1520         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1521                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1522                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1523                         rc = signal_completed_replay(imp);
1524                         if (rc)
1525                                 GOTO(out, rc);
1526                 }
1527         }
1528
1529         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1530                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1531                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1532                 }
1533         }
1534
1535         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1536                 struct ptlrpc_connection *conn = imp->imp_connection;
1537
1538                 rc = ptlrpc_resend(imp);
1539                 if (rc)
1540                         GOTO(out, rc);
1541                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1542                 ptlrpc_activate_import(imp);
1543
1544                 LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n",
1545                               imp->imp_obd->obd_name,
1546                               obd_uuid2str(&conn->c_remote_uuid),
1547                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1548         }
1549
1550         if (imp->imp_state == LUSTRE_IMP_FULL) {
1551                 wake_up_all(&imp->imp_recovery_waitq);
1552                 ptlrpc_wake_delayed(imp);
1553         }
1554
1555 out:
1556         RETURN(rc);
1557 }
1558
1559 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1560 {
1561         struct ptlrpc_request *req;
1562         int rq_opc, rc = 0;
1563         ENTRY;
1564
1565         if (imp->imp_obd->obd_force)
1566                 GOTO(set_state, rc);
1567
1568         switch (imp->imp_connect_op) {
1569         case OST_CONNECT:
1570                 rq_opc = OST_DISCONNECT;
1571                 break;
1572         case MDS_CONNECT:
1573                 rq_opc = MDS_DISCONNECT;
1574                 break;
1575         case MGS_CONNECT:
1576                 rq_opc = MGS_DISCONNECT;
1577                 break;
1578         default:
1579                 rc = -EINVAL;
1580                 CERROR("%s: don't know how to disconnect from %s "
1581                        "(connect_op %d): rc = %d\n",
1582                        imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1583                        imp->imp_connect_op, rc);
1584                 RETURN(rc);
1585         }
1586
1587         if (ptlrpc_import_in_recovery(imp)) {
1588                 struct l_wait_info lwi;
1589                 cfs_duration_t timeout;
1590
1591                 if (AT_OFF) {
1592                         if (imp->imp_server_timeout)
1593                                 timeout = cfs_time_seconds(obd_timeout / 2);
1594                         else
1595                                 timeout = cfs_time_seconds(obd_timeout);
1596                 } else {
1597                         int idx = import_at_get_index(imp,
1598                                 imp->imp_client->cli_request_portal);
1599                         timeout = cfs_time_seconds(
1600                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1601                 }
1602
1603                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1604                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1605                 rc = l_wait_event(imp->imp_recovery_waitq,
1606                                   !ptlrpc_import_in_recovery(imp), &lwi);
1607
1608         }
1609
1610         spin_lock(&imp->imp_lock);
1611         if (imp->imp_state != LUSTRE_IMP_FULL)
1612                 GOTO(out, rc);
1613         spin_unlock(&imp->imp_lock);
1614
1615         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1616                                         LUSTRE_OBD_VERSION, rq_opc);
1617         if (req) {
1618                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1619                  * it fails.  We can get through the above with a down server
1620                  * if the client doesn't know the server is gone yet. */
1621                 req->rq_no_resend = 1;
1622
1623                 /* We want client umounts to happen quickly, no matter the
1624                    server state... */
1625                 req->rq_timeout = min_t(int, req->rq_timeout,
1626                                         INITIAL_CONNECT_TIMEOUT);
1627
1628                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1629                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1630                 ptlrpc_request_set_replen(req);
1631                 rc = ptlrpc_queue_wait(req);
1632                 ptlrpc_req_finished(req);
1633         }
1634
1635 set_state:
1636         spin_lock(&imp->imp_lock);
1637 out:
1638         if (noclose)
1639                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1640         else
1641                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1642         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1643         spin_unlock(&imp->imp_lock);
1644
1645         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1646                 rc = 0;
1647         RETURN(rc);
1648 }
1649 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1650
1651 void ptlrpc_cleanup_imp(struct obd_import *imp)
1652 {
1653         ENTRY;
1654
1655         spin_lock(&imp->imp_lock);
1656         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1657         imp->imp_generation++;
1658         spin_unlock(&imp->imp_lock);
1659         ptlrpc_abort_inflight(imp);
1660
1661         EXIT;
1662 }
1663
1664 /* Adaptive Timeout utils */
1665 extern unsigned int at_min, at_max, at_history;
1666
1667 /* Update at_current with the specified value (bounded by at_min and at_max),
1668  * as well as the AT history "bins".
1669  *  - Bin into timeslices using AT_BINS bins.
1670  *  - This gives us a max of the last at_history seconds without the storage,
1671  *    but still smoothing out a return to normalcy from a slow response.
1672  *  - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1673  */
1674 int at_measured(struct adaptive_timeout *at, unsigned int val)
1675 {
1676         unsigned int old = at->at_current;
1677         time_t now = cfs_time_current_sec();
1678         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1679
1680         LASSERT(at);
1681         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1682                val, at, now - at->at_binstart, at->at_current,
1683                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1684
1685         if (val == 0)
1686                 /* 0's don't count, because we never want our timeout to
1687                    drop to 0, and because 0 could mean an error */
1688                 return 0;
1689
1690         spin_lock(&at->at_lock);
1691
1692         if (unlikely(at->at_binstart == 0)) {
1693                 /* Special case to remove default from history */
1694                 at->at_current = val;
1695                 at->at_worst_ever = val;
1696                 at->at_worst_time = now;
1697                 at->at_hist[0] = val;
1698                 at->at_binstart = now;
1699         } else if (now - at->at_binstart < binlimit ) {
1700                 /* in bin 0 */
1701                 at->at_hist[0] = max(val, at->at_hist[0]);
1702                 at->at_current = max(val, at->at_current);
1703         } else {
1704                 int i, shift;
1705                 unsigned int maxv = val;
1706                 /* move bins over */
1707                 shift = (now - at->at_binstart) / binlimit;
1708                 LASSERT(shift > 0);
1709                 for(i = AT_BINS - 1; i >= 0; i--) {
1710                         if (i >= shift) {
1711                                 at->at_hist[i] = at->at_hist[i - shift];
1712                                 maxv = max(maxv, at->at_hist[i]);
1713                         } else {
1714                                 at->at_hist[i] = 0;
1715                         }
1716                 }
1717                 at->at_hist[0] = val;
1718                 at->at_current = maxv;
1719                 at->at_binstart += shift * binlimit;
1720         }
1721
1722         if (at->at_current > at->at_worst_ever) {
1723                 at->at_worst_ever = at->at_current;
1724                 at->at_worst_time = now;
1725         }
1726
1727         if (at->at_flags & AT_FLG_NOHIST)
1728                 /* Only keep last reported val; keeping the rest of the history
1729                    for proc only */
1730                 at->at_current = val;
1731
1732         if (at_max > 0)
1733                 at->at_current =  min(at->at_current, at_max);
1734         at->at_current =  max(at->at_current, at_min);
1735
1736         if (at->at_current != old)
1737                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1738                        "(val=%u) hist %u %u %u %u\n", at,
1739                        old, at->at_current, at->at_current - old, val,
1740                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1741                        at->at_hist[3]);
1742
1743         /* if we changed, report the old value */
1744         old = (at->at_current != old) ? old : 0;
1745
1746         spin_unlock(&at->at_lock);
1747         return old;
1748 }
1749
1750 /* Find the imp_at index for a given portal; assign if space available */
1751 int import_at_get_index(struct obd_import *imp, int portal)
1752 {
1753         struct imp_at *at = &imp->imp_at;
1754         int i;
1755
1756         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1757                 if (at->iat_portal[i] == portal)
1758                         return i;
1759                 if (at->iat_portal[i] == 0)
1760                         /* unused */
1761                         break;
1762         }
1763
1764         /* Not found in list, add it under a lock */
1765         spin_lock(&imp->imp_lock);
1766
1767         /* Check unused under lock */
1768         for (; i < IMP_AT_MAX_PORTALS; i++) {
1769                 if (at->iat_portal[i] == portal)
1770                         goto out;
1771                 if (at->iat_portal[i] == 0)
1772                         /* unused */
1773                         break;
1774         }
1775
1776         /* Not enough portals? */
1777         LASSERT(i < IMP_AT_MAX_PORTALS);
1778
1779         at->iat_portal[i] = portal;
1780 out:
1781         spin_unlock(&imp->imp_lock);
1782         return i;
1783 }