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: RPCs in \"%s\" phase found (%d). "
 376                                        "Network is sluggish? Waiting them "
 377                                        "to error out.\n", cli_tgt,
 378                                        ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
 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         struct l_wait_info lwi;
 454         int secs = cfs_time_seconds(obd_timeout);
 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), secs);
 461
 462         lwi = LWI_TIMEOUT(secs, 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 "LPU64"\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         /* 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 "LPX64", replied "LPX64"\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: "LPX64"\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 granted asked subset of flags: "
1056                        "asked="LPX64" grranted="LPX64"\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 (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1063             (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1064             (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1065             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1066                 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1067                 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1068                 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1069
1070                 /* We do not support the MDT-MDT interoperations with
1071                  * different version MDT because of protocol changes. */
1072                 if (unlikely(major != LUSTRE_MAJOR ||
1073                              minor != LUSTRE_MINOR ||
1074                              abs(patch - LUSTRE_PATCH) > 3)) {
1075                         LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1076                                       "connection to different version MDT "
1077                                       "(%d.%d.%d.%d) %s\n",
1078                                       imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1079                                       LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1080                                       major, minor, patch,
1081                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1082                                       imp->imp_connection->c_remote_uuid.uuid);
1083
1084                         GOTO(out, rc = -EPROTO);
1085                 }
1086         }
1087
1088         if (!exp) {
1089                 /* This could happen if export is cleaned during the
1090                    connect attempt */
1091                 CERROR("%s: missing export after connect\n",
1092                        imp->imp_obd->obd_name);
1093                 GOTO(out, rc = -ENODEV);
1094         }
1095         old_connect_flags = exp_connect_flags(exp);
1096         exp->exp_connect_data = *ocd;
1097         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1098
1099         /* The net statistics after (re-)connect is not valid anymore,
1100          * because may reflect other routing, etc. */
1101         at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1102         ptlrpc_at_adj_net_latency(request,
1103                         lustre_msg_get_service_time(request->rq_repmsg));
1104
1105         /* Import flags should be updated before waking import at FULL state */
1106         rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1107                                  aa->pcaa_initial_connect);
1108         class_export_put(exp);
1109         if (rc != 0)
1110                 GOTO(out, rc);
1111
1112         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1113
1114         if (aa->pcaa_initial_connect) {
1115                 spin_lock(&imp->imp_lock);
1116                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1117                         imp->imp_replayable = 1;
1118                         spin_unlock(&imp->imp_lock);
1119                         CDEBUG(D_HA, "connected to replayable target: %s\n",
1120                                obd2cli_tgt(imp->imp_obd));
1121                 } else {
1122                         imp->imp_replayable = 0;
1123                         spin_unlock(&imp->imp_lock);
1124                 }
1125
1126                 /* if applies, adjust the imp->imp_msg_magic here
1127                  * according to reply flags */
1128
1129                 imp->imp_remote_handle =
1130                                 *lustre_msg_get_handle(request->rq_repmsg);
1131
1132                 /* Initial connects are allowed for clients with non-random
1133                  * uuids when servers are in recovery.  Simply signal the
1134                  * servers replay is complete and wait in REPLAY_WAIT. */
1135                 if (msg_flags & MSG_CONNECT_RECOVERING) {
1136                         CDEBUG(D_HA, "connect to %s during recovery\n",
1137                                obd2cli_tgt(imp->imp_obd));
1138                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1139                 } else {
1140                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1141                         ptlrpc_activate_import(imp);
1142                 }
1143
1144                 GOTO(finish, rc = 0);
1145         }
1146
1147         /* Determine what recovery state to move the import to. */
1148         if (MSG_CONNECT_RECONNECT & msg_flags) {
1149                 memset(&old_hdl, 0, sizeof(old_hdl));
1150                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1151                             sizeof (old_hdl))) {
1152                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1153                                       "bad handle "LPX64"\n",
1154                                       obd2cli_tgt(imp->imp_obd),
1155                                       imp->imp_connection->c_remote_uuid.uuid,
1156                                       imp->imp_dlm_handle.cookie);
1157                         GOTO(out, rc = -ENOTCONN);
1158                 }
1159
1160                 if (memcmp(&imp->imp_remote_handle,
1161                            lustre_msg_get_handle(request->rq_repmsg),
1162                            sizeof(imp->imp_remote_handle))) {
1163                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
1164                                 D_HA : D_WARNING;
1165
1166                         /* Bug 16611/14775: if server handle have changed,
1167                          * that means some sort of disconnection happened.
1168                          * If the server is not in recovery, that also means it
1169                          * already erased all of our state because of previous
1170                          * eviction. If it is in recovery - we are safe to
1171                          * participate since we can reestablish all of our state
1172                          * with server again */
1173                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1174                                 CDEBUG(level,"%s@%s changed server handle from "
1175                                        LPX64" to "LPX64
1176                                        " but is still in recovery\n",
1177                                        obd2cli_tgt(imp->imp_obd),
1178                                        imp->imp_connection->c_remote_uuid.uuid,
1179                                        imp->imp_remote_handle.cookie,
1180                                        lustre_msg_get_handle(
1181                                        request->rq_repmsg)->cookie);
1182                         } else {
1183                                 LCONSOLE_WARN("Evicted from %s (at %s) "
1184                                               "after server handle changed from "
1185                                               LPX64" to "LPX64"\n",
1186                                               obd2cli_tgt(imp->imp_obd),
1187                                               imp->imp_connection-> \
1188                                               c_remote_uuid.uuid,
1189                                               imp->imp_remote_handle.cookie,
1190                                               lustre_msg_get_handle(
1191                                               request->rq_repmsg)->cookie);
1192                         }
1193
1194
1195                         imp->imp_remote_handle =
1196                                      *lustre_msg_get_handle(request->rq_repmsg);
1197
1198                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1199                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1200                                 GOTO(finish, rc = 0);
1201                         }
1202
1203                 } else {
1204                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1205                                obd2cli_tgt(imp->imp_obd),
1206                                imp->imp_connection->c_remote_uuid.uuid);
1207                 }
1208
1209                 if (imp->imp_invalid) {
1210                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1211                                "marking evicted\n", imp->imp_obd->obd_name);
1212                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1213                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1214                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1215                                imp->imp_obd->obd_name,
1216                                obd2cli_tgt(imp->imp_obd));
1217
1218                         spin_lock(&imp->imp_lock);
1219                         imp->imp_resend_replay = 1;
1220                         spin_unlock(&imp->imp_lock);
1221
1222                         IMPORT_SET_STATE(imp, imp->imp_replay_state);
1223                 } else {
1224                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1225                 }
1226         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1227                 LASSERT(imp->imp_replayable);
1228                 imp->imp_remote_handle =
1229                                 *lustre_msg_get_handle(request->rq_repmsg);
1230                 imp->imp_last_replay_transno = 0;
1231                 imp->imp_replay_cursor = &imp->imp_committed_list;
1232                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
1233         } else {
1234                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
1235                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
1236                 imp->imp_remote_handle =
1237                                 *lustre_msg_get_handle(request->rq_repmsg);
1238                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1239         }
1240
1241         /* Sanity checks for a reconnected import. */
1242         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
1243                 CERROR("imp_replayable flag does not match server "
1244                        "after reconnect. We should LBUG right here.\n");
1245         }
1246
1247         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1248             lustre_msg_get_last_committed(request->rq_repmsg) <
1249             aa->pcaa_peer_committed) {
1250                 CERROR("%s went back in time (transno "LPD64
1251                        " was previously committed, server now claims "LPD64
1252                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
1253                        "id=9646\n",
1254                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1255                        lustre_msg_get_last_committed(request->rq_repmsg));
1256         }
1257
1258 finish:
1259         ptlrpc_prepare_replay(imp);
1260         rc = ptlrpc_import_recovery_state_machine(imp);
1261         if (rc == -ENOTCONN) {
1262                 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1263                        "invalidating and reconnecting\n",
1264                        obd2cli_tgt(imp->imp_obd),
1265                        imp->imp_connection->c_remote_uuid.uuid);
1266                 ptlrpc_connect_import(imp);
1267                 spin_lock(&imp->imp_lock);
1268                 imp->imp_connected = 0;
1269                 imp->imp_connect_tried = 1;
1270                 spin_unlock(&imp->imp_lock);
1271                 RETURN(0);
1272         }
1273
1274 out:
1275         spin_lock(&imp->imp_lock);
1276         imp->imp_connected = 0;
1277         imp->imp_connect_tried = 1;
1278         spin_unlock(&imp->imp_lock);
1279
1280         if (rc != 0) {
1281                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1282                 if (rc == -EACCES) {
1283                         /*
1284                          * Give up trying to reconnect
1285                          * EACCES means client has no permission for connection
1286                          */
1287                         imp->imp_obd->obd_no_recov = 1;
1288                         ptlrpc_deactivate_import(imp);
1289                 }
1290
1291                 if (rc == -EPROTO) {
1292                         struct obd_connect_data *ocd;
1293
1294                         /* reply message might not be ready */
1295                         if (request->rq_repmsg == NULL)
1296                                 RETURN(-EPROTO);
1297
1298                         ocd = req_capsule_server_get(&request->rq_pill,
1299                                                      &RMF_CONNECT_DATA);
1300                         if (ocd &&
1301                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1302                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1303                            /* Actually servers are only supposed to refuse
1304                               connection from liblustre clients, so we should
1305                               never see this from VFS context */
1306                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1307                                         "(%d.%d.%d.%d)"
1308                                         " refused connection from this client "
1309                                         "with an incompatible version (%s).  "
1310                                         "Client must be recompiled\n",
1311                                         obd2cli_tgt(imp->imp_obd),
1312                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1313                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1314                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1315                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1316                                         LUSTRE_VERSION_STRING);
1317                                 ptlrpc_deactivate_import(imp);
1318                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1319                         }
1320                         RETURN(-EPROTO);
1321                 }
1322
1323                 ptlrpc_maybe_ping_import_soon(imp);
1324
1325                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1326                        obd2cli_tgt(imp->imp_obd),
1327                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1328         }
1329
1330         wake_up_all(&imp->imp_recovery_waitq);
1331         RETURN(rc);
1332 }
1333
1334 /**
1335  * interpret callback for "completed replay" RPCs.
1336  * \see signal_completed_replay
1337  */
1338 static int completed_replay_interpret(const struct lu_env *env,
1339                                       struct ptlrpc_request *req,
1340                                       void * data, int rc)
1341 {
1342         ENTRY;
1343         atomic_dec(&req->rq_import->imp_replay_inflight);
1344         if (req->rq_status == 0 &&
1345             !req->rq_import->imp_vbr_failed) {
1346                 ptlrpc_import_recovery_state_machine(req->rq_import);
1347         } else {
1348                 if (req->rq_import->imp_vbr_failed) {
1349                         CDEBUG(D_WARNING,
1350                                "%s: version recovery fails, reconnecting\n",
1351                                req->rq_import->imp_obd->obd_name);
1352                 } else {
1353                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1354                                      "reconnecting\n",
1355                                req->rq_import->imp_obd->obd_name,
1356                                req->rq_status);
1357                 }
1358                 ptlrpc_connect_import(req->rq_import);
1359         }
1360
1361         RETURN(0);
1362 }
1363
1364 /**
1365  * Let server know that we have no requests to replay anymore.
1366  * Achieved by just sending a PING request
1367  */
1368 static int signal_completed_replay(struct obd_import *imp)
1369 {
1370         struct ptlrpc_request *req;
1371         ENTRY;
1372
1373         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1374                 RETURN(0);
1375
1376         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1377         atomic_inc(&imp->imp_replay_inflight);
1378
1379         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1380                                         OBD_PING);
1381         if (req == NULL) {
1382                 atomic_dec(&imp->imp_replay_inflight);
1383                 RETURN(-ENOMEM);
1384         }
1385
1386         ptlrpc_request_set_replen(req);
1387         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1388         lustre_msg_add_flags(req->rq_reqmsg,
1389                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1390         if (AT_OFF)
1391                 req->rq_timeout *= 3;
1392         req->rq_interpret_reply = completed_replay_interpret;
1393
1394         ptlrpcd_add_req(req);
1395         RETURN(0);
1396 }
1397
1398 /**
1399  * In kernel code all import invalidation happens in its own
1400  * separate thread, so that whatever application happened to encounter
1401  * a problem could still be killed or otherwise continue
1402  */
1403 static int ptlrpc_invalidate_import_thread(void *data)
1404 {
1405         struct obd_import *imp = data;
1406
1407         ENTRY;
1408
1409         unshare_fs_struct();
1410
1411         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1412                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1413                imp->imp_connection->c_remote_uuid.uuid);
1414
1415         ptlrpc_invalidate_import(imp);
1416
1417         if (obd_dump_on_eviction) {
1418                 CERROR("dump the log upon eviction\n");
1419                 libcfs_debug_dumplog();
1420         }
1421
1422         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1423         ptlrpc_import_recovery_state_machine(imp);
1424
1425         class_import_put(imp);
1426         RETURN(0);
1427 }
1428
1429 /**
1430  * This is the state machine for client-side recovery on import.
1431  *
1432  * Typicaly we have two possibly paths. If we came to server and it is not
1433  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1434  * state and reconnect from scratch.
1435  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1436  * We go through our list of requests to replay and send them to server one by
1437  * one.
1438  * After sending all request from the list we change import state to
1439  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1440  * and also all the locks we don't yet have and wait for server to grant us.
1441  * After that we send a special "replay completed" request and change import
1442  * state to IMP_REPLAY_WAIT.
1443  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1444  * state and resend all requests from sending list.
1445  * After that we promote import to FULL state and send all delayed requests
1446  * and import is fully operational after that.
1447  *
1448  */
1449 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1450 {
1451         int rc = 0;
1452         int inflight;
1453         char *target_start;
1454         int target_len;
1455
1456         ENTRY;
1457         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1458                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1459                           &target_start, &target_len);
1460                 /* Don't care about MGC eviction */
1461                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1462                            LUSTRE_MGC_NAME) != 0) {
1463                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1464                                            "by %.*s; in progress operations "
1465                                            "using this service will fail.\n",
1466                                            imp->imp_obd->obd_name, target_len,
1467                                            target_start);
1468                 }
1469                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1470                        obd2cli_tgt(imp->imp_obd),
1471                        imp->imp_connection->c_remote_uuid.uuid);
1472                 /* reset vbr_failed flag upon eviction */
1473                 spin_lock(&imp->imp_lock);
1474                 imp->imp_vbr_failed = 0;
1475                 spin_unlock(&imp->imp_lock);
1476
1477                 {
1478                 struct task_struct *task;
1479                 /* bug 17802:  XXX client_disconnect_export vs connect request
1480                  * race. if client is evicted at this time then we start
1481                  * invalidate thread without reference to import and import can
1482                  * be freed at same time. */
1483                 class_import_get(imp);
1484                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1485                                      "ll_imp_inval");
1486                 if (IS_ERR(task)) {
1487                         class_import_put(imp);
1488                         CERROR("error starting invalidate thread: %d\n", rc);
1489                         rc = PTR_ERR(task);
1490                 } else {
1491                         rc = 0;
1492                 }
1493                 RETURN(rc);
1494                 }
1495         }
1496
1497         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1498                 CDEBUG(D_HA, "replay requested by %s\n",
1499                        obd2cli_tgt(imp->imp_obd));
1500                 rc = ptlrpc_replay_next(imp, &inflight);
1501                 if (inflight == 0 &&
1502                     atomic_read(&imp->imp_replay_inflight) == 0) {
1503                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1504                         rc = ldlm_replay_locks(imp);
1505                         if (rc)
1506                                 GOTO(out, rc);
1507                 }
1508                 rc = 0;
1509         }
1510
1511         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1512                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1513                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1514                         rc = signal_completed_replay(imp);
1515                         if (rc)
1516                                 GOTO(out, rc);
1517                 }
1518         }
1519
1520         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1521                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1522                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1523                 }
1524         }
1525
1526         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1527                 struct ptlrpc_connection *conn = imp->imp_connection;
1528
1529                 rc = ptlrpc_resend(imp);
1530                 if (rc)
1531                         GOTO(out, rc);
1532                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1533                 ptlrpc_activate_import(imp);
1534
1535                 LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n",
1536                               imp->imp_obd->obd_name,
1537                               obd_uuid2str(&conn->c_remote_uuid),
1538                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1539         }
1540
1541         if (imp->imp_state == LUSTRE_IMP_FULL) {
1542                 wake_up_all(&imp->imp_recovery_waitq);
1543                 ptlrpc_wake_delayed(imp);
1544         }
1545
1546 out:
1547         RETURN(rc);
1548 }
1549
1550 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1551 {
1552         struct ptlrpc_request *req;
1553         int rq_opc, rc = 0;
1554         ENTRY;
1555
1556         if (imp->imp_obd->obd_force)
1557                 GOTO(set_state, rc);
1558
1559         switch (imp->imp_connect_op) {
1560         case OST_CONNECT:
1561                 rq_opc = OST_DISCONNECT;
1562                 break;
1563         case MDS_CONNECT:
1564                 rq_opc = MDS_DISCONNECT;
1565                 break;
1566         case MGS_CONNECT:
1567                 rq_opc = MGS_DISCONNECT;
1568                 break;
1569         default:
1570                 rc = -EINVAL;
1571                 CERROR("%s: don't know how to disconnect from %s "
1572                        "(connect_op %d): rc = %d\n",
1573                        imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1574                        imp->imp_connect_op, rc);
1575                 RETURN(rc);
1576         }
1577
1578         if (ptlrpc_import_in_recovery(imp)) {
1579                 struct l_wait_info lwi;
1580                 cfs_duration_t timeout;
1581
1582                 if (AT_OFF) {
1583                         if (imp->imp_server_timeout)
1584                                 timeout = cfs_time_seconds(obd_timeout / 2);
1585                         else
1586                                 timeout = cfs_time_seconds(obd_timeout);
1587                 } else {
1588                         int idx = import_at_get_index(imp,
1589                                 imp->imp_client->cli_request_portal);
1590                         timeout = cfs_time_seconds(
1591                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1592                 }
1593
1594                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1595                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1596                 rc = l_wait_event(imp->imp_recovery_waitq,
1597                                   !ptlrpc_import_in_recovery(imp), &lwi);
1598
1599         }
1600
1601         spin_lock(&imp->imp_lock);
1602         if (imp->imp_state != LUSTRE_IMP_FULL)
1603                 GOTO(out, rc);
1604         spin_unlock(&imp->imp_lock);
1605
1606         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1607                                         LUSTRE_OBD_VERSION, rq_opc);
1608         if (req) {
1609                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1610                  * it fails.  We can get through the above with a down server
1611                  * if the client doesn't know the server is gone yet. */
1612                 req->rq_no_resend = 1;
1613
1614                 /* We want client umounts to happen quickly, no matter the
1615                    server state... */
1616                 req->rq_timeout = min_t(int, req->rq_timeout,
1617                                         INITIAL_CONNECT_TIMEOUT);
1618
1619                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1620                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1621                 ptlrpc_request_set_replen(req);
1622                 rc = ptlrpc_queue_wait(req);
1623                 ptlrpc_req_finished(req);
1624         }
1625
1626 set_state:
1627         spin_lock(&imp->imp_lock);
1628 out:
1629         if (noclose)
1630                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1631         else
1632                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1633         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1634         spin_unlock(&imp->imp_lock);
1635
1636         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1637                 rc = 0;
1638         RETURN(rc);
1639 }
1640 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1641
1642 void ptlrpc_cleanup_imp(struct obd_import *imp)
1643 {
1644         ENTRY;
1645
1646         spin_lock(&imp->imp_lock);
1647         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1648         imp->imp_generation++;
1649         spin_unlock(&imp->imp_lock);
1650         ptlrpc_abort_inflight(imp);
1651
1652         EXIT;
1653 }
1654
1655 /* Adaptive Timeout utils */
1656 extern unsigned int at_min, at_max, at_history;
1657
1658 /* Update at_current with the specified value (bounded by at_min and at_max),
1659  * as well as the AT history "bins".
1660  *  - Bin into timeslices using AT_BINS bins.
1661  *  - This gives us a max of the last at_history seconds without the storage,
1662  *    but still smoothing out a return to normalcy from a slow response.
1663  *  - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1664  */
1665 int at_measured(struct adaptive_timeout *at, unsigned int val)
1666 {
1667         unsigned int old = at->at_current;
1668         time_t now = cfs_time_current_sec();
1669         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1670
1671         LASSERT(at);
1672         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1673                val, at, now - at->at_binstart, at->at_current,
1674                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1675
1676         if (val == 0)
1677                 /* 0's don't count, because we never want our timeout to
1678                    drop to 0, and because 0 could mean an error */
1679                 return 0;
1680
1681         spin_lock(&at->at_lock);
1682
1683         if (unlikely(at->at_binstart == 0)) {
1684                 /* Special case to remove default from history */
1685                 at->at_current = val;
1686                 at->at_worst_ever = val;
1687                 at->at_worst_time = now;
1688                 at->at_hist[0] = val;
1689                 at->at_binstart = now;
1690         } else if (now - at->at_binstart < binlimit ) {
1691                 /* in bin 0 */
1692                 at->at_hist[0] = max(val, at->at_hist[0]);
1693                 at->at_current = max(val, at->at_current);
1694         } else {
1695                 int i, shift;
1696                 unsigned int maxv = val;
1697                 /* move bins over */
1698                 shift = (now - at->at_binstart) / binlimit;
1699                 LASSERT(shift > 0);
1700                 for(i = AT_BINS - 1; i >= 0; i--) {
1701                         if (i >= shift) {
1702                                 at->at_hist[i] = at->at_hist[i - shift];
1703                                 maxv = max(maxv, at->at_hist[i]);
1704                         } else {
1705                                 at->at_hist[i] = 0;
1706                         }
1707                 }
1708                 at->at_hist[0] = val;
1709                 at->at_current = maxv;
1710                 at->at_binstart += shift * binlimit;
1711         }
1712
1713         if (at->at_current > at->at_worst_ever) {
1714                 at->at_worst_ever = at->at_current;
1715                 at->at_worst_time = now;
1716         }
1717
1718         if (at->at_flags & AT_FLG_NOHIST)
1719                 /* Only keep last reported val; keeping the rest of the history
1720                    for proc only */
1721                 at->at_current = val;
1722
1723         if (at_max > 0)
1724                 at->at_current =  min(at->at_current, at_max);
1725         at->at_current =  max(at->at_current, at_min);
1726
1727         if (at->at_current != old)
1728                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1729                        "(val=%u) hist %u %u %u %u\n", at,
1730                        old, at->at_current, at->at_current - old, val,
1731                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1732                        at->at_hist[3]);
1733
1734         /* if we changed, report the old value */
1735         old = (at->at_current != old) ? old : 0;
1736
1737         spin_unlock(&at->at_lock);
1738         return old;
1739 }
1740
1741 /* Find the imp_at index for a given portal; assign if space available */
1742 int import_at_get_index(struct obd_import *imp, int portal)
1743 {
1744         struct imp_at *at = &imp->imp_at;
1745         int i;
1746
1747         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1748                 if (at->iat_portal[i] == portal)
1749                         return i;
1750                 if (at->iat_portal[i] == 0)
1751                         /* unused */
1752                         break;
1753         }
1754
1755         /* Not found in list, add it under a lock */
1756         spin_lock(&imp->imp_lock);
1757
1758         /* Check unused under lock */
1759         for (; i < IMP_AT_MAX_PORTALS; i++) {
1760                 if (at->iat_portal[i] == portal)
1761                         goto out;
1762                 if (at->iat_portal[i] == 0)
1763                         /* unused */
1764                         break;
1765         }
1766
1767         /* Not enough portals? */
1768         LASSERT(i < IMP_AT_MAX_PORTALS);
1769
1770         at->iat_portal[i] = portal;
1771 out:
1772         spin_unlock(&imp->imp_lock);
1773         return i;
1774 }