lustre/ptlrpc/import.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ptlrpc/import.c
  33  *
  34  * Author: Mike Shaver <shaver@clusterfs.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38
  39 #include <linux/kthread.h>
  40 #include <obd_support.h>
  41 #include <lustre_ha.h>
  42 #include <lustre_net.h>
  43 #include <lustre_import.h>
  44 #include <lustre_export.h>
  45 #include <obd.h>
  46 #include <obd_cksum.h>
  47 #include <obd_class.h>
  48
  49 #include "ptlrpc_internal.h"
  50
  51 struct ptlrpc_connect_async_args {
  52          __u64 pcaa_peer_committed;
  53         int pcaa_initial_connect;
  54 };
  55
  56 /**
  57  * Updates import \a imp current state to provided \a state value
  58  * Helper function. Must be called under imp_lock.
  59  */
  60 static void __import_set_state(struct obd_import *imp,
  61                                enum lustre_imp_state state)
  62 {
  63         switch (state) {
  64         case LUSTRE_IMP_CLOSED:
  65         case LUSTRE_IMP_NEW:
  66         case LUSTRE_IMP_DISCON:
  67         case LUSTRE_IMP_CONNECTING:
  68                 break;
  69         case LUSTRE_IMP_REPLAY_WAIT:
  70                 imp->imp_replay_state = LUSTRE_IMP_REPLAY_LOCKS;
  71                 break;
  72         default:
  73                 imp->imp_replay_state = LUSTRE_IMP_REPLAY;
  74         }
  75         imp->imp_state = state;
  76         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
  77         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
  78                 ktime_get_real_seconds();
  79         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
  80                 IMP_STATE_HIST_LEN;
  81 }
  82
  83 /* A CLOSED import should remain so. */
  84 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  85 do {                                                                           \
  86         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  87                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  88                       imp, obd2cli_tgt(imp->imp_obd),                          \
  89                       ptlrpc_import_state_name(imp->imp_state),                \
  90                       ptlrpc_import_state_name(state));                        \
  91                __import_set_state(imp, state);                                 \
  92         }                                                                      \
  93 } while(0)
  94
  95 #define IMPORT_SET_STATE(imp, state)                                    \
  96 do {                                                                    \
  97         spin_lock(&imp->imp_lock);                                      \
  98         IMPORT_SET_STATE_NOLOCK(imp, state);                            \
  99         spin_unlock(&imp->imp_lock);                                    \
 100 } while(0)
 101
 102 void ptlrpc_import_enter_resend(struct obd_import *imp)
 103 {
 104         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 105 }
 106 EXPORT_SYMBOL(ptlrpc_import_enter_resend);
 107
 108
 109 static int ptlrpc_connect_interpret(const struct lu_env *env,
 110                                     struct ptlrpc_request *request,
 111                                     void * data, int rc);
 112 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
 113
 114 /* Only this function is allowed to change the import state when it is
 115  * CLOSED. I would rather refcount the import and free it after
 116  * disconnection like we do with exports. To do that, the client_obd
 117  * will need to save the peer info somewhere other than in the import,
 118  * though. */
 119 int ptlrpc_init_import(struct obd_import *imp)
 120 {
 121         spin_lock(&imp->imp_lock);
 122
 123         imp->imp_generation++;
 124         imp->imp_state =  LUSTRE_IMP_NEW;
 125
 126         spin_unlock(&imp->imp_lock);
 127
 128         return 0;
 129 }
 130 EXPORT_SYMBOL(ptlrpc_init_import);
 131
 132 #define UUID_STR "_UUID"
 133 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
 134 {
 135         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 136                 ? uuid : uuid + strlen(prefix);
 137
 138         *uuid_len = strlen(*uuid_start);
 139
 140         if (*uuid_len < strlen(UUID_STR))
 141                 return;
 142
 143         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 144                     UUID_STR, strlen(UUID_STR)))
 145                 *uuid_len -= strlen(UUID_STR);
 146 }
 147
 148 /**
 149  * Returns true if import was FULL, false if import was already not
 150  * connected.
 151  * @imp - import to be disconnected
 152  * @conn_cnt - connection count (epoch) of the request that timed out
 153  *             and caused the disconnection.  In some cases, multiple
 154  *             inflight requests can fail to a single target (e.g. OST
 155  *             bulk requests) and if one has already caused a reconnection
 156  *             (increasing the import->conn_cnt) the older failure should
 157  *             not also cause a reconnection.  If zero it forces a reconnect.
 158  */
 159 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 160 {
 161         int rc = 0;
 162
 163         spin_lock(&imp->imp_lock);
 164
 165         if (imp->imp_state == LUSTRE_IMP_FULL &&
 166             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 167                 char *target_start;
 168                 int   target_len;
 169
 170                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 171                           &target_start, &target_len);
 172
 173                 if (imp->imp_replayable) {
 174                         LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
 175                                "lost; in progress operations using this "
 176                                "service will wait for recovery to complete\n",
 177                                imp->imp_obd->obd_name, target_len, target_start,
 178                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 179                 } else {
 180                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
 181                                "%.*s (at %s) was lost; in progress "
 182                                "operations using this service will fail\n",
 183                                imp->imp_obd->obd_name,
 184                                target_len, target_start,
 185                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 186                 }
 187                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 188                 spin_unlock(&imp->imp_lock);
 189
 190                 if (obd_dump_on_timeout)
 191                         libcfs_debug_dumplog();
 192
 193                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 194                 rc = 1;
 195         } else {
 196                 spin_unlock(&imp->imp_lock);
 197                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 198                        imp->imp_client->cli_name, imp,
 199                        (imp->imp_state == LUSTRE_IMP_FULL &&
 200                         imp->imp_conn_cnt > conn_cnt) ?
 201                        "reconnected" : "not connected", imp->imp_conn_cnt,
 202                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 203         }
 204
 205         return rc;
 206 }
 207
 208 /* Must be called with imp_lock held! */
 209 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 210 {
 211         ENTRY;
 212         assert_spin_locked(&imp->imp_lock);
 213
 214         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 215         imp->imp_invalid = 1;
 216         imp->imp_generation++;
 217         spin_unlock(&imp->imp_lock);
 218
 219         ptlrpc_abort_inflight(imp);
 220         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 221
 222         EXIT;
 223 }
 224
 225 /*
 226  * This acts as a barrier; all existing requests are rejected, and
 227  * no new requests will be accepted until the import is valid again.
 228  */
 229 void ptlrpc_deactivate_import(struct obd_import *imp)
 230 {
 231         spin_lock(&imp->imp_lock);
 232         ptlrpc_deactivate_and_unlock_import(imp);
 233 }
 234 EXPORT_SYMBOL(ptlrpc_deactivate_import);
 235
 236 static time64_t ptlrpc_inflight_deadline(struct ptlrpc_request *req,
 237                                          time64_t now)
 238 {
 239         time64_t dl;
 240
 241         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
 242               (req->rq_phase == RQ_PHASE_BULK) ||
 243               (req->rq_phase == RQ_PHASE_NEW)))
 244                 return 0;
 245
 246         if (req->rq_timedout)
 247                 return 0;
 248
 249         if (req->rq_phase == RQ_PHASE_NEW)
 250                 dl = req->rq_sent;
 251         else
 252                 dl = req->rq_deadline;
 253
 254         if (dl <= now)
 255                 return 0;
 256
 257         return dl - now;
 258 }
 259
 260 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
 261 {
 262         time64_t now = ktime_get_real_seconds();
 263         struct list_head *tmp, *n;
 264         struct ptlrpc_request *req;
 265         time64_t timeout = 0;
 266
 267         spin_lock(&imp->imp_lock);
 268         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 269                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 270                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 271         }
 272         spin_unlock(&imp->imp_lock);
 273         return timeout;
 274 }
 275
 276 /**
 277  * This function will invalidate the import, if necessary, then block
 278  * for all the RPC completions, and finally notify the obd to
 279  * invalidate its state (ie cancel locks, clear pending requests,
 280  * etc).
 281  */
 282 void ptlrpc_invalidate_import(struct obd_import *imp)
 283 {
 284         struct list_head *tmp, *n;
 285         struct ptlrpc_request *req;
 286         struct l_wait_info lwi;
 287         time64_t timeout;
 288         int rc;
 289
 290         atomic_inc(&imp->imp_inval_count);
 291
 292         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 293                 ptlrpc_deactivate_import(imp);
 294
 295         CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
 296         LASSERT(imp->imp_invalid);
 297
 298         /* Wait forever until inflight == 0. We really can't do it another
 299          * way because in some cases we need to wait for very long reply
 300          * unlink. We can't do anything before that because there is really
 301          * no guarantee that some rdma transfer is not in progress right now. */
 302         do {
 303                 long timeout_jiffies;
 304
 305                 /* Calculate max timeout for waiting on rpcs to error
 306                  * out. Use obd_timeout if calculated value is smaller
 307                  * than it.
 308                  */
 309                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 310                         timeout = ptlrpc_inflight_timeout(imp);
 311                         timeout += timeout / 3;
 312
 313                         if (timeout == 0)
 314                                 timeout = obd_timeout;
 315                 } else {
 316                         /* decrease the interval to increase race condition */
 317                         timeout = 1;
 318                 }
 319
 320                 CDEBUG(D_RPCTRACE, "Sleeping %llds for inflight to error out\n",
 321                        timeout);
 322
 323                 /* Wait for all requests to error out and call completion
 324                  * callbacks. Cap it at obd_timeout -- these should all
 325                  * have been locally cancelled by ptlrpc_abort_inflight.
 326                  */
 327                 timeout_jiffies = max_t(long, cfs_time_seconds(timeout), 1);
 328                 lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
 329                                            (timeout > 1) ? cfs_time_seconds(1) :
 330                                                            cfs_time_seconds(1) / 2,
 331                                                            NULL, NULL);
 332                 rc = l_wait_event(imp->imp_recovery_waitq,
 333                                   (atomic_read(&imp->imp_inflight) == 0),
 334                                   &lwi);
 335                 if (rc) {
 336                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 337
 338                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 339                                cli_tgt, rc, atomic_read(&imp->imp_inflight));
 340
 341                         spin_lock(&imp->imp_lock);
 342                         if (atomic_read(&imp->imp_inflight) == 0) {
 343                                 int count = atomic_read(&imp->imp_unregistering);
 344
 345                                 /* We know that "unregistering" rpcs only can
 346                                  * survive in sending or delaying lists (they
 347                                  * maybe waiting for long reply unlink in
 348                                  * sluggish nets). Let's check this. If there
 349                                  * is no inflight and unregistering != 0, this
 350                                  * is bug. */
 351                                 LASSERTF(count == 0, "Some RPCs are still "
 352                                          "unregistering: %d\n", count);
 353
 354                                 /* Let's save one loop as soon as inflight have
 355                                  * dropped to zero. No new inflights possible at
 356                                  * this point. */
 357                                 rc = 0;
 358                         } else {
 359                                 list_for_each_safe(tmp, n,
 360                                                    &imp->imp_sending_list) {
 361                                         req = list_entry(tmp,
 362                                                          struct ptlrpc_request,
 363                                                          rq_list);
 364                                         DEBUG_REQ(D_ERROR, req,
 365                                                   "still on sending list");
 366                                 }
 367                                 list_for_each_safe(tmp, n,
 368                                                    &imp->imp_delayed_list) {
 369                                         req = list_entry(tmp,
 370                                                          struct ptlrpc_request,
 371                                                          rq_list);
 372                                         DEBUG_REQ(D_ERROR, req,
 373                                                   "still on delayed list");
 374                                 }
 375
 376                                 CERROR("%s: Unregistering RPCs found (%d). "
 377                                        "Network is sluggish? Waiting them "
 378                                        "to error out.\n", cli_tgt,
 379                                        atomic_read(&imp->imp_unregistering));
 380                         }
 381                         spin_unlock(&imp->imp_lock);
 382                 }
 383         } while (rc != 0);
 384
 385         /*
 386          * Let's additionally check that no new rpcs added to import in
 387          * "invalidate" state.
 388          */
 389         LASSERT(atomic_read(&imp->imp_inflight) == 0);
 390         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 391         sptlrpc_import_flush_all_ctx(imp);
 392
 393         atomic_dec(&imp->imp_inval_count);
 394         wake_up_all(&imp->imp_recovery_waitq);
 395 }
 396 EXPORT_SYMBOL(ptlrpc_invalidate_import);
 397
 398 /* unset imp_invalid */
 399 void ptlrpc_activate_import(struct obd_import *imp)
 400 {
 401         struct obd_device *obd = imp->imp_obd;
 402
 403         spin_lock(&imp->imp_lock);
 404         if (imp->imp_deactive != 0) {
 405                 spin_unlock(&imp->imp_lock);
 406                 return;
 407         }
 408
 409         imp->imp_invalid = 0;
 410         spin_unlock(&imp->imp_lock);
 411         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 412 }
 413 EXPORT_SYMBOL(ptlrpc_activate_import);
 414
 415 void ptlrpc_pinger_force(struct obd_import *imp)
 416 {
 417         CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
 418                ptlrpc_import_state_name(imp->imp_state));
 419
 420         spin_lock(&imp->imp_lock);
 421         imp->imp_force_verify = 1;
 422         spin_unlock(&imp->imp_lock);
 423
 424         if (imp->imp_state != LUSTRE_IMP_CONNECTING)
 425                 ptlrpc_pinger_wake_up();
 426 }
 427 EXPORT_SYMBOL(ptlrpc_pinger_force);
 428
 429 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 430 {
 431         ENTRY;
 432
 433         LASSERT(!imp->imp_dlm_fake);
 434
 435         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 436                 if (!imp->imp_replayable) {
 437                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 438                                "auto-deactivating\n",
 439                                obd2cli_tgt(imp->imp_obd),
 440                                imp->imp_connection->c_remote_uuid.uuid,
 441                                imp->imp_obd->obd_name);
 442                         ptlrpc_deactivate_import(imp);
 443                 }
 444
 445                 ptlrpc_pinger_force(imp);
 446         }
 447         EXIT;
 448 }
 449
 450 int ptlrpc_reconnect_import(struct obd_import *imp)
 451 {
 452 #ifdef ENABLE_PINGER
 453         long timeout_jiffies = cfs_time_seconds(obd_timeout);
 454         struct l_wait_info lwi;
 455         int rc;
 456
 457         ptlrpc_pinger_force(imp);
 458
 459         CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 460                obd2cli_tgt(imp->imp_obd), obd_timeout);
 461
 462         lwi = LWI_TIMEOUT(timeout_jiffies, NULL, NULL);
 463         rc = l_wait_event(imp->imp_recovery_waitq,
 464                           !ptlrpc_import_in_recovery(imp), &lwi);
 465         CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
 466                ptlrpc_import_state_name(imp->imp_state));
 467         return rc;
 468 #else
 469         ptlrpc_set_import_discon(imp, 0);
 470         /* Force a new connect attempt */
 471         ptlrpc_invalidate_import(imp);
 472         /* Do a fresh connect next time by zeroing the handle */
 473         ptlrpc_disconnect_import(imp, 1);
 474         /* Wait for all invalidate calls to finish */
 475         if (atomic_read(&imp->imp_inval_count) > 0) {
 476                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 477                 int rc;
 478
 479                 rc = l_wait_event(imp->imp_recovery_waitq,
 480                                   (atomic_read(&imp->imp_inval_count) == 0),
 481                                   &lwi);
 482                 if (rc)
 483                         CERROR("Interrupted, inval=%d\n",
 484                                atomic_read(&imp->imp_inval_count));
 485         }
 486
 487         /* Allow reconnect attempts */
 488         imp->imp_obd->obd_no_recov = 0;
 489         /* Remove 'invalid' flag */
 490         ptlrpc_activate_import(imp);
 491         /* Attempt a new connect */
 492         ptlrpc_recover_import(imp, NULL, 0);
 493         return 0;
 494 #endif
 495 }
 496 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 497
 498 /**
 499  * Connection on import \a imp is changed to another one (if more than one is
 500  * present). We typically chose connection that we have not tried to connect to
 501  * the longest
 502  */
 503 static int import_select_connection(struct obd_import *imp)
 504 {
 505         struct obd_import_conn *imp_conn = NULL, *conn;
 506         struct obd_export *dlmexp;
 507         char *target_start;
 508         int target_len, tried_all = 1;
 509         ENTRY;
 510
 511         spin_lock(&imp->imp_lock);
 512
 513         if (list_empty(&imp->imp_conn_list)) {
 514                 CERROR("%s: no connections available\n",
 515                        imp->imp_obd->obd_name);
 516                 spin_unlock(&imp->imp_lock);
 517                 RETURN(-EINVAL);
 518         }
 519
 520         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 521                 CDEBUG(D_HA, "%s: connect to NID %s last attempt %lld\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                     conn->oic_last_attempt <= 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 (imp_conn->oic_last_attempt > conn->oic_last_attempt)
 541                         imp_conn = conn;
 542         }
 543
 544         /* if not found, simply choose the current one */
 545         if (!imp_conn || imp->imp_force_reconnect) {
 546                 LASSERT(imp->imp_conn_current);
 547                 imp_conn = imp->imp_conn_current;
 548                 tried_all = 0;
 549         }
 550         LASSERT(imp_conn->oic_conn);
 551
 552         /* If we've tried everything, and we're back to the beginning of the
 553            list, increase our timeout and try again. It will be reset when
 554            we do finally connect. (FIXME: really we should wait for all network
 555            state associated with the last connection attempt to drain before
 556            trying to reconnect on it.) */
 557         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 558                 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
 559                 if (at_get(at) < CONNECTION_SWITCH_MAX) {
 560                         at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
 561                         if (at_get(at) > CONNECTION_SWITCH_MAX)
 562                                 at_reset(at, CONNECTION_SWITCH_MAX);
 563                 }
 564                 LASSERT(imp_conn->oic_last_attempt);
 565                 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
 566                         "to %ds\n", imp->imp_obd->obd_name, at_get(at));
 567         }
 568
 569         imp_conn->oic_last_attempt = ktime_get_seconds();
 570
 571         /* switch connection, don't mind if it's same as the current one */
 572         if (imp->imp_connection)
 573                 ptlrpc_connection_put(imp->imp_connection);
 574         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 575
 576         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 577         LASSERT(dlmexp != NULL);
 578         if (dlmexp->exp_connection)
 579                 ptlrpc_connection_put(dlmexp->exp_connection);
 580         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 581         class_export_put(dlmexp);
 582
 583         if (imp->imp_conn_current != imp_conn) {
 584                 if (imp->imp_conn_current) {
 585                         deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 586                                   &target_start, &target_len);
 587
 588                         CDEBUG(D_HA, "%s: Connection changing to"
 589                                " %.*s (at %s)\n",
 590                                imp->imp_obd->obd_name,
 591                                target_len, target_start,
 592                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 593                 }
 594
 595                 imp->imp_conn_current = imp_conn;
 596         }
 597
 598         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 599                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 600                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 601
 602         spin_unlock(&imp->imp_lock);
 603
 604         RETURN(0);
 605 }
 606
 607 /*
 608  * must be called under imp_lock
 609  */
 610 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 611 {
 612         struct ptlrpc_request   *req;
 613         struct list_head        *tmp;
 614
 615         /* The requests in committed_list always have smaller transnos than
 616          * the requests in replay_list */
 617         if (!list_empty(&imp->imp_committed_list)) {
 618                 tmp = imp->imp_committed_list.next;
 619                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 620                 *transno = req->rq_transno;
 621                 if (req->rq_transno == 0) {
 622                         DEBUG_REQ(D_ERROR, req, "zero transno in committed_list");
 623                         LBUG();
 624                 }
 625                 return 1;
 626         }
 627         if (!list_empty(&imp->imp_replay_list)) {
 628                 tmp = imp->imp_replay_list.next;
 629                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 630                 *transno = req->rq_transno;
 631                 if (req->rq_transno == 0) {
 632                         DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
 633                         LBUG();
 634                 }
 635                 return 1;
 636         }
 637         return 0;
 638 }
 639
 640 /**
 641  * Attempt to (re)connect import \a imp. This includes all preparations,
 642  * initializing CONNECT RPC request and passing it to ptlrpcd for
 643  * actual sending.
 644  * Returns 0 on success or error code.
 645  */
 646 int ptlrpc_connect_import(struct obd_import *imp)
 647 {
 648         struct obd_device *obd = imp->imp_obd;
 649         int initial_connect = 0;
 650         int set_transno = 0;
 651         __u64 committed_before_reconnect = 0;
 652         struct ptlrpc_request *request;
 653         char *bufs[] = { NULL,
 654                          obd2cli_tgt(imp->imp_obd),
 655                          obd->obd_uuid.uuid,
 656                          (char *)&imp->imp_dlm_handle,
 657                          (char *)&imp->imp_connect_data };
 658         struct ptlrpc_connect_async_args *aa;
 659         int rc;
 660         ENTRY;
 661
 662         spin_lock(&imp->imp_lock);
 663         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 664                 spin_unlock(&imp->imp_lock);
 665                 CERROR("can't connect to a closed import\n");
 666                 RETURN(-EINVAL);
 667         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 668                 spin_unlock(&imp->imp_lock);
 669                 CERROR("already connected\n");
 670                 RETURN(0);
 671         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING ||
 672                    imp->imp_connected) {
 673                 spin_unlock(&imp->imp_lock);
 674                 CERROR("already connecting\n");
 675                 RETURN(-EALREADY);
 676         }
 677
 678         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 679
 680         imp->imp_conn_cnt++;
 681         imp->imp_resend_replay = 0;
 682
 683         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 684                 initial_connect = 1;
 685         else
 686                 committed_before_reconnect = imp->imp_peer_committed_transno;
 687
 688         set_transno = ptlrpc_first_transno(imp,
 689                                            &imp->imp_connect_data.ocd_transno);
 690         spin_unlock(&imp->imp_lock);
 691
 692         rc = import_select_connection(imp);
 693         if (rc)
 694                 GOTO(out, rc);
 695
 696         rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
 697         if (rc)
 698                 GOTO(out, rc);
 699
 700         /* Reset connect flags to the originally requested flags, in case
 701          * the server is updated on-the-fly we will get the new features. */
 702         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 703         imp->imp_connect_data.ocd_connect_flags2 = imp->imp_connect_flags2_orig;
 704         /* Reset ocd_version each time so the server knows the exact versions */
 705         imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
 706         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 707         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 708
 709         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 710                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 711         if (rc)
 712                 GOTO(out, rc);
 713
 714         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 715         if (request == NULL)
 716                 GOTO(out, rc = -ENOMEM);
 717
 718         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 719                                       imp->imp_connect_op, bufs, NULL);
 720         if (rc) {
 721                 ptlrpc_request_free(request);
 722                 GOTO(out, rc);
 723         }
 724
 725         /* Report the rpc service time to the server so that it knows how long
 726          * to wait for clients to join recovery */
 727         lustre_msg_set_service_time(request->rq_reqmsg,
 728                                     at_timeout2est(request->rq_timeout));
 729
 730         /* The amount of time we give the server to process the connect req.
 731          * import_select_connection will increase the net latency on
 732          * repeated reconnect attempts to cover slow networks.
 733          * We override/ignore the server rpc completion estimate here,
 734          * which may be large if this is a reconnect attempt */
 735         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 736         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 737
 738         request->rq_no_resend = request->rq_no_delay = 1;
 739         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 740         /* Allow a slightly larger reply for future growth compatibility */
 741         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 742                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 743         ptlrpc_request_set_replen(request);
 744         request->rq_interpret_reply = ptlrpc_connect_interpret;
 745
 746         CLASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
 747         aa = ptlrpc_req_async_args(request);
 748         memset(aa, 0, sizeof *aa);
 749
 750         aa->pcaa_peer_committed = committed_before_reconnect;
 751         aa->pcaa_initial_connect = initial_connect;
 752
 753         if (aa->pcaa_initial_connect) {
 754                 spin_lock(&imp->imp_lock);
 755                 imp->imp_replayable = 1;
 756                 spin_unlock(&imp->imp_lock);
 757                 lustre_msg_add_op_flags(request->rq_reqmsg,
 758                                         MSG_CONNECT_INITIAL);
 759         }
 760
 761         if (set_transno)
 762                 lustre_msg_add_op_flags(request->rq_reqmsg,
 763                                         MSG_CONNECT_TRANSNO);
 764
 765         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %ld)",
 766                   request->rq_timeout);
 767         ptlrpcd_add_req(request);
 768         rc = 0;
 769 out:
 770         if (rc != 0)
 771                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 772
 773         RETURN(rc);
 774 }
 775 EXPORT_SYMBOL(ptlrpc_connect_import);
 776
 777 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 778 {
 779         int force_verify;
 780
 781         spin_lock(&imp->imp_lock);
 782         force_verify = imp->imp_force_verify != 0;
 783         spin_unlock(&imp->imp_lock);
 784
 785         if (force_verify)
 786                 ptlrpc_pinger_wake_up();
 787 }
 788
 789 static int ptlrpc_busy_reconnect(int rc)
 790 {
 791         return (rc == -EBUSY) || (rc == -EAGAIN);
 792 }
 793
 794 static int ptlrpc_connect_set_flags(struct obd_import *imp,
 795                                      struct obd_connect_data *ocd,
 796                                      __u64 old_connect_flags,
 797                                      struct obd_export *exp, int init_connect)
 798 {
 799         static bool warned;
 800         struct client_obd *cli = &imp->imp_obd->u.cli;
 801
 802         spin_lock(&imp->imp_lock);
 803         list_del(&imp->imp_conn_current->oic_item);
 804         list_add(&imp->imp_conn_current->oic_item,
 805                  &imp->imp_conn_list);
 806         imp->imp_last_success_conn =
 807                 imp->imp_conn_current->oic_last_attempt;
 808
 809         spin_unlock(&imp->imp_lock);
 810
 811         if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 812             (ocd->ocd_version > LUSTRE_VERSION_CODE +
 813                                 LUSTRE_VERSION_OFFSET_WARN ||
 814              ocd->ocd_version < LUSTRE_VERSION_CODE -
 815                                 LUSTRE_VERSION_OFFSET_WARN)) {
 816                 /* Sigh, some compilers do not like #ifdef in the middle
 817                    of macro arguments */
 818                 const char *older = "older than client. "
 819                                     "Consider upgrading server";
 820                 const char *newer = "newer than client. "
 821                                     "Consider upgrading client";
 822
 823                 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 824                               "is much %s (%s)\n",
 825                               obd2cli_tgt(imp->imp_obd),
 826                               OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 827                               OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 828                               OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 829                               OBD_OCD_VERSION_FIX(ocd->ocd_version),
 830                               ocd->ocd_version > LUSTRE_VERSION_CODE ?
 831                               newer : older, LUSTRE_VERSION_STRING);
 832                 warned = true;
 833         }
 834
 835 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
 836         /* Check if server has LU-1252 fix applied to not always swab
 837          * the IR MNE entries. Do this only once per connection.  This
 838          * fixup is version-limited, because we don't want to carry the
 839          * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
 840          * need interop with unpatched 2.2 servers.  For newer servers,
 841          * the client will do MNE swabbing only as needed.  LU-1644 */
 842         if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 843                      !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
 844                      OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
 845                      OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
 846                      OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
 847                      strcmp(imp->imp_obd->obd_type->typ_name,
 848                             LUSTRE_MGC_NAME) == 0))
 849                 imp->imp_need_mne_swab = 1;
 850         else /* clear if server was upgraded since last connect */
 851                 imp->imp_need_mne_swab = 0;
 852 #endif
 853
 854         if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
 855                 /* We sent to the server ocd_cksum_types with bits set
 856                  * for algorithms we understand. The server masked off
 857                  * the checksum types it doesn't support */
 858                 if ((ocd->ocd_cksum_types &
 859                      cksum_types_supported_client()) == 0) {
 860                         LCONSOLE_ERROR("The negotiation of the checksum "
 861                                        "alogrithm to use with server %s "
 862                                        "failed (%x/%x)\n",
 863                                        obd2cli_tgt(imp->imp_obd),
 864                                        ocd->ocd_cksum_types,
 865                                        cksum_types_supported_client());
 866                         return -EPROTO;
 867                 } else {
 868                         cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
 869                 }
 870         } else {
 871                 /* The server does not support OBD_CONNECT_CKSUM.
 872                  * Enforce ADLER for backward compatibility*/
 873                 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
 874         }
 875         cli->cl_cksum_type = cksum_type_select(cli->cl_supp_cksum_types);
 876
 877         if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
 878                 cli->cl_max_pages_per_rpc =
 879                         min(ocd->ocd_brw_size >> PAGE_SHIFT,
 880                             cli->cl_max_pages_per_rpc);
 881         else if (imp->imp_connect_op == MDS_CONNECT ||
 882                  imp->imp_connect_op == MGS_CONNECT)
 883                 cli->cl_max_pages_per_rpc = 1;
 884
 885         LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 886                 (cli->cl_max_pages_per_rpc > 0));
 887
 888         client_adjust_max_dirty(cli);
 889
 890         /* Update client max modify RPCs in flight with value returned
 891          * by the server */
 892         if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
 893                 cli->cl_max_mod_rpcs_in_flight = min(
 894                                         cli->cl_max_mod_rpcs_in_flight,
 895                                         ocd->ocd_maxmodrpcs);
 896         else
 897                 cli->cl_max_mod_rpcs_in_flight = 1;
 898
 899         /* Reset ns_connect_flags only for initial connect. It might be
 900          * changed in while using FS and if we reset it in reconnect
 901          * this leads to losing user settings done before such as
 902          * disable lru_resize, etc. */
 903         if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
 904                 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
 905                              "flags: %#llx\n", imp->imp_obd->obd_name,
 906                              ocd->ocd_connect_flags);
 907                 imp->imp_obd->obd_namespace->ns_connect_flags =
 908                         ocd->ocd_connect_flags;
 909                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
 910                         ocd->ocd_connect_flags;
 911         }
 912
 913         if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
 914                 /* We need a per-message support flag, because
 915                  * a. we don't know if the incoming connect reply
 916                  *    supports AT or not (in reply_in_callback)
 917                  *    until we unpack it.
 918                  * b. failovered server means export and flags are gone
 919                  *    (in ptlrpc_send_reply).
 920                  *    Can only be set when we know AT is supported at
 921                  *    both ends */
 922                 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 923         else
 924                 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 925
 926         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 927
 928         return 0;
 929 }
 930
 931 /**
 932  * Add all replay requests back to unreplied list before start replay,
 933  * so that we can make sure the known replied XID is always increased
 934  * only even if when replaying requests.
 935  */
 936 static void ptlrpc_prepare_replay(struct obd_import *imp)
 937 {
 938         struct ptlrpc_request *req;
 939
 940         if (imp->imp_state != LUSTRE_IMP_REPLAY ||
 941             imp->imp_resend_replay)
 942                 return;
 943
 944         /* If the server was restart during repaly, the requests may
 945          * have been added to the unreplied list in former replay. */
 946         spin_lock(&imp->imp_lock);
 947
 948         list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
 949                 if (list_empty(&req->rq_unreplied_list))
 950                         ptlrpc_add_unreplied(req);
 951         }
 952
 953         list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
 954                 if (list_empty(&req->rq_unreplied_list))
 955                         ptlrpc_add_unreplied(req);
 956         }
 957
 958         imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
 959         spin_unlock(&imp->imp_lock);
 960 }
 961
 962 /**
 963  * interpret_reply callback for connect RPCs.
 964  * Looks into returned status of connect operation and decides
 965  * what to do with the import - i.e enter recovery, promote it to
 966  * full state for normal operations of disconnect it due to an error.
 967  */
 968 static int ptlrpc_connect_interpret(const struct lu_env *env,
 969                                     struct ptlrpc_request *request,
 970                                     void *data, int rc)
 971 {
 972         struct ptlrpc_connect_async_args *aa = data;
 973         struct obd_import *imp = request->rq_import;
 974         struct lustre_handle old_hdl;
 975         __u64 old_connect_flags;
 976         int msg_flags;
 977         struct obd_connect_data *ocd;
 978         struct obd_export *exp;
 979         int ret;
 980         ENTRY;
 981
 982         spin_lock(&imp->imp_lock);
 983         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 984                 imp->imp_connect_tried = 1;
 985                 spin_unlock(&imp->imp_lock);
 986                 RETURN(0);
 987         }
 988
 989         if (rc) {
 990                 /* if this reconnect to busy export - not need select new target
 991                  * for connecting*/
 992                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 993                 spin_unlock(&imp->imp_lock);
 994                 ptlrpc_maybe_ping_import_soon(imp);
 995                 GOTO(out, rc);
 996         }
 997
 998         /* LU-7558: indicate that we are interpretting connect reply,
 999          * pltrpc_connect_import() will not try to reconnect until
1000          * interpret will finish. */
1001         imp->imp_connected = 1;
1002         spin_unlock(&imp->imp_lock);
1003
1004         LASSERT(imp->imp_conn_current);
1005
1006         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1007
1008         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1009                                    RCL_SERVER);
1010         /* server replied obd_connect_data is always bigger */
1011         ocd = req_capsule_server_sized_get(&request->rq_pill,
1012                                            &RMF_CONNECT_DATA, ret);
1013
1014         if (ocd == NULL) {
1015                 CERROR("%s: no connect data from server\n",
1016                        imp->imp_obd->obd_name);
1017                 rc = -EPROTO;
1018                 GOTO(out, rc);
1019         }
1020
1021         spin_lock(&imp->imp_lock);
1022
1023         /* All imports are pingable */
1024         imp->imp_pingable = 1;
1025         imp->imp_force_reconnect = 0;
1026         imp->imp_force_verify = 0;
1027
1028         imp->imp_connect_data = *ocd;
1029
1030         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1031                imp->imp_obd->obd_name, ocd->ocd_instance);
1032         exp = class_conn2export(&imp->imp_dlm_handle);
1033
1034         spin_unlock(&imp->imp_lock);
1035
1036         if (!exp) {
1037                 /* This could happen if export is cleaned during the
1038                    connect attempt */
1039                 CERROR("%s: missing export after connect\n",
1040                        imp->imp_obd->obd_name);
1041                 GOTO(out, rc = -ENODEV);
1042         }
1043
1044         /* check that server granted subset of flags we asked for. */
1045         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1046             ocd->ocd_connect_flags) {
1047                 CERROR("%s: Server didn't grant requested subset of flags: "
1048                        "asked=%#llx granted=%#llx\n",
1049                        imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
1050                        ocd->ocd_connect_flags);
1051                 GOTO(out, rc = -EPROTO);
1052         }
1053
1054         if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) !=
1055             ocd->ocd_connect_flags2) {
1056                 CERROR("%s: Server didn't grant requested subset of flags2: "
1057                        "asked=%#llx granted=%#llx\n",
1058                        imp->imp_obd->obd_name, imp->imp_connect_flags2_orig,
1059                        ocd->ocd_connect_flags2);
1060                 GOTO(out, rc = -EPROTO);
1061         }
1062
1063         if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1064             (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1065             (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1066             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1067                 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1068                 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1069                 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1070
1071                 /* We do not support the MDT-MDT interoperations with
1072                  * different version MDT because of protocol changes. */
1073                 if (unlikely(major != LUSTRE_MAJOR ||
1074                              minor != LUSTRE_MINOR ||
1075                              abs(patch - LUSTRE_PATCH) > 3)) {
1076                         LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1077                                       "connection to different version MDT "
1078                                       "(%d.%d.%d.%d) %s\n",
1079                                       imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1080                                       LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1081                                       major, minor, patch,
1082                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1083                                       imp->imp_connection->c_remote_uuid.uuid);
1084
1085                         GOTO(out, rc = -EPROTO);
1086                 }
1087         }
1088
1089         old_connect_flags = exp_connect_flags(exp);
1090         exp->exp_connect_data = *ocd;
1091         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1092
1093         /* The net statistics after (re-)connect is not valid anymore,
1094          * because may reflect other routing, etc. */
1095         at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1096         ptlrpc_at_adj_net_latency(request,
1097                         lustre_msg_get_service_time(request->rq_repmsg));
1098
1099         /* Import flags should be updated before waking import at FULL state */
1100         rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1101                                       aa->pcaa_initial_connect);
1102         class_export_put(exp);
1103         if (rc != 0)
1104                 GOTO(out, rc);
1105
1106         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1107
1108         if (aa->pcaa_initial_connect) {
1109                 spin_lock(&imp->imp_lock);
1110                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1111                         imp->imp_replayable = 1;
1112                         spin_unlock(&imp->imp_lock);
1113                         CDEBUG(D_HA, "connected to replayable target: %s\n",
1114                                obd2cli_tgt(imp->imp_obd));
1115                 } else {
1116                         imp->imp_replayable = 0;
1117                         spin_unlock(&imp->imp_lock);
1118                 }
1119
1120                 /* if applies, adjust the imp->imp_msg_magic here
1121                  * according to reply flags */
1122
1123                 imp->imp_remote_handle =
1124                                 *lustre_msg_get_handle(request->rq_repmsg);
1125
1126                 /* Initial connects are allowed for clients with non-random
1127                  * uuids when servers are in recovery.  Simply signal the
1128                  * servers replay is complete and wait in REPLAY_WAIT. */
1129                 if (msg_flags & MSG_CONNECT_RECOVERING) {
1130                         CDEBUG(D_HA, "connect to %s during recovery\n",
1131                                obd2cli_tgt(imp->imp_obd));
1132                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1133                 } else {
1134                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1135                         ptlrpc_activate_import(imp);
1136                 }
1137
1138                 GOTO(finish, rc = 0);
1139         }
1140
1141         /* Determine what recovery state to move the import to. */
1142         if (MSG_CONNECT_RECONNECT & msg_flags) {
1143                 memset(&old_hdl, 0, sizeof(old_hdl));
1144                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1145                             sizeof (old_hdl))) {
1146                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1147                                       "bad handle %#llx\n",
1148                                       obd2cli_tgt(imp->imp_obd),
1149                                       imp->imp_connection->c_remote_uuid.uuid,
1150                                       imp->imp_dlm_handle.cookie);
1151                         GOTO(out, rc = -ENOTCONN);
1152                 }
1153
1154                 if (memcmp(&imp->imp_remote_handle,
1155                            lustre_msg_get_handle(request->rq_repmsg),
1156                            sizeof(imp->imp_remote_handle))) {
1157                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
1158                                 D_HA : D_WARNING;
1159
1160                         /* Bug 16611/14775: if server handle have changed,
1161                          * that means some sort of disconnection happened.
1162                          * If the server is not in recovery, that also means it
1163                          * already erased all of our state because of previous
1164                          * eviction. If it is in recovery - we are safe to
1165                          * participate since we can reestablish all of our state
1166                          * with server again */
1167                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1168                                 CDEBUG(level,"%s@%s changed server handle from "
1169                                        "%#llx to %#llx"
1170                                        " but is still in recovery\n",
1171                                        obd2cli_tgt(imp->imp_obd),
1172                                        imp->imp_connection->c_remote_uuid.uuid,
1173                                        imp->imp_remote_handle.cookie,
1174                                        lustre_msg_get_handle(
1175                                        request->rq_repmsg)->cookie);
1176                         } else {
1177                                 LCONSOLE_WARN("Evicted from %s (at %s) "
1178                                               "after server handle changed from "
1179                                               "%#llx to %#llx\n",
1180                                               obd2cli_tgt(imp->imp_obd),
1181                                               imp->imp_connection-> \
1182                                               c_remote_uuid.uuid,
1183                                               imp->imp_remote_handle.cookie,
1184                                               lustre_msg_get_handle(
1185                                               request->rq_repmsg)->cookie);
1186                         }
1187
1188
1189                         imp->imp_remote_handle =
1190                                      *lustre_msg_get_handle(request->rq_repmsg);
1191
1192                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1193                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1194                                 GOTO(finish, rc = 0);
1195                         }
1196
1197                 } else {
1198                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1199                                obd2cli_tgt(imp->imp_obd),
1200                                imp->imp_connection->c_remote_uuid.uuid);
1201                 }
1202
1203                 if (imp->imp_invalid) {
1204                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1205                                "marking evicted\n", imp->imp_obd->obd_name);
1206                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1207                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1208                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1209                                imp->imp_obd->obd_name,
1210                                obd2cli_tgt(imp->imp_obd));
1211
1212                         spin_lock(&imp->imp_lock);
1213                         imp->imp_resend_replay = 1;
1214                         spin_unlock(&imp->imp_lock);
1215
1216                         IMPORT_SET_STATE(imp, imp->imp_replay_state);
1217                 } else {
1218                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1219                 }
1220         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1221                 LASSERT(imp->imp_replayable);
1222                 imp->imp_remote_handle =
1223                                 *lustre_msg_get_handle(request->rq_repmsg);
1224                 imp->imp_last_replay_transno = 0;
1225                 imp->imp_replay_cursor = &imp->imp_committed_list;
1226                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
1227         } else {
1228                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
1229                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
1230                 imp->imp_remote_handle =
1231                                 *lustre_msg_get_handle(request->rq_repmsg);
1232                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1233         }
1234
1235         /* Sanity checks for a reconnected import. */
1236         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
1237                 CERROR("imp_replayable flag does not match server "
1238                        "after reconnect. We should LBUG right here.\n");
1239         }
1240
1241         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1242             lustre_msg_get_last_committed(request->rq_repmsg) <
1243             aa->pcaa_peer_committed) {
1244                 CERROR("%s went back in time (transno %lld"
1245                        " was previously committed, server now claims %lld"
1246                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
1247                        "id=9646\n",
1248                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1249                        lustre_msg_get_last_committed(request->rq_repmsg));
1250         }
1251
1252 finish:
1253         ptlrpc_prepare_replay(imp);
1254         rc = ptlrpc_import_recovery_state_machine(imp);
1255         if (rc == -ENOTCONN) {
1256                 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1257                        "invalidating and reconnecting\n",
1258                        obd2cli_tgt(imp->imp_obd),
1259                        imp->imp_connection->c_remote_uuid.uuid);
1260                 ptlrpc_connect_import(imp);
1261                 spin_lock(&imp->imp_lock);
1262                 imp->imp_connected = 0;
1263                 imp->imp_connect_tried = 1;
1264                 spin_unlock(&imp->imp_lock);
1265                 RETURN(0);
1266         }
1267
1268 out:
1269         spin_lock(&imp->imp_lock);
1270         imp->imp_connected = 0;
1271         imp->imp_connect_tried = 1;
1272         spin_unlock(&imp->imp_lock);
1273
1274         if (rc != 0) {
1275                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1276                 if (rc == -EACCES) {
1277                         /*
1278                          * Give up trying to reconnect
1279                          * EACCES means client has no permission for connection
1280                          */
1281                         imp->imp_obd->obd_no_recov = 1;
1282                         ptlrpc_deactivate_import(imp);
1283                 }
1284
1285                 if (rc == -EPROTO) {
1286                         struct obd_connect_data *ocd;
1287
1288                         /* reply message might not be ready */
1289                         if (request->rq_repmsg == NULL)
1290                                 RETURN(-EPROTO);
1291
1292                         ocd = req_capsule_server_get(&request->rq_pill,
1293                                                      &RMF_CONNECT_DATA);
1294                         /* Servers are not supposed to refuse connections from
1295                          * clients based on version, only connection feature
1296                          * flags.  We should never see this from llite, but it
1297                          * may be useful for debugging in the future. */
1298                         if (ocd &&
1299                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1300                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1301                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1302                                         "(%d.%d.%d.%d)"
1303                                         " refused connection from this client "
1304                                         "with an incompatible version (%s).  "
1305                                         "Client must be recompiled\n",
1306                                         obd2cli_tgt(imp->imp_obd),
1307                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1308                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1309                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1310                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1311                                         LUSTRE_VERSION_STRING);
1312                                 ptlrpc_deactivate_import(imp);
1313                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1314                         }
1315                         RETURN(-EPROTO);
1316                 }
1317
1318                 ptlrpc_maybe_ping_import_soon(imp);
1319
1320                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1321                        obd2cli_tgt(imp->imp_obd),
1322                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1323         }
1324
1325         wake_up_all(&imp->imp_recovery_waitq);
1326         RETURN(rc);
1327 }
1328
1329 /**
1330  * interpret callback for "completed replay" RPCs.
1331  * \see signal_completed_replay
1332  */
1333 static int completed_replay_interpret(const struct lu_env *env,
1334                                       struct ptlrpc_request *req,
1335                                       void * data, int rc)
1336 {
1337         ENTRY;
1338         atomic_dec(&req->rq_import->imp_replay_inflight);
1339         if (req->rq_status == 0 &&
1340             !req->rq_import->imp_vbr_failed) {
1341                 ptlrpc_import_recovery_state_machine(req->rq_import);
1342         } else {
1343                 if (req->rq_import->imp_vbr_failed) {
1344                         CDEBUG(D_WARNING,
1345                                "%s: version recovery fails, reconnecting\n",
1346                                req->rq_import->imp_obd->obd_name);
1347                 } else {
1348                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1349                                      "reconnecting\n",
1350                                req->rq_import->imp_obd->obd_name,
1351                                req->rq_status);
1352                 }
1353                 ptlrpc_connect_import(req->rq_import);
1354         }
1355
1356         RETURN(0);
1357 }
1358
1359 /**
1360  * Let server know that we have no requests to replay anymore.
1361  * Achieved by just sending a PING request
1362  */
1363 static int signal_completed_replay(struct obd_import *imp)
1364 {
1365         struct ptlrpc_request *req;
1366         ENTRY;
1367
1368         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1369                 RETURN(0);
1370
1371         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1372         atomic_inc(&imp->imp_replay_inflight);
1373
1374         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1375                                         OBD_PING);
1376         if (req == NULL) {
1377                 atomic_dec(&imp->imp_replay_inflight);
1378                 RETURN(-ENOMEM);
1379         }
1380
1381         ptlrpc_request_set_replen(req);
1382         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1383         lustre_msg_add_flags(req->rq_reqmsg,
1384                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1385         if (AT_OFF)
1386                 req->rq_timeout *= 3;
1387         req->rq_interpret_reply = completed_replay_interpret;
1388
1389         ptlrpcd_add_req(req);
1390         RETURN(0);
1391 }
1392
1393 /**
1394  * In kernel code all import invalidation happens in its own
1395  * separate thread, so that whatever application happened to encounter
1396  * a problem could still be killed or otherwise continue
1397  */
1398 static int ptlrpc_invalidate_import_thread(void *data)
1399 {
1400         struct obd_import *imp = data;
1401
1402         ENTRY;
1403
1404         unshare_fs_struct();
1405
1406         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1407                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1408                imp->imp_connection->c_remote_uuid.uuid);
1409
1410         ptlrpc_invalidate_import(imp);
1411
1412         if (obd_dump_on_eviction) {
1413                 CERROR("dump the log upon eviction\n");
1414                 libcfs_debug_dumplog();
1415         }
1416
1417         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1418         ptlrpc_import_recovery_state_machine(imp);
1419
1420         class_import_put(imp);
1421         RETURN(0);
1422 }
1423
1424 /**
1425  * This is the state machine for client-side recovery on import.
1426  *
1427  * Typicaly we have two possibly paths. If we came to server and it is not
1428  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1429  * state and reconnect from scratch.
1430  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1431  * We go through our list of requests to replay and send them to server one by
1432  * one.
1433  * After sending all request from the list we change import state to
1434  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1435  * and also all the locks we don't yet have and wait for server to grant us.
1436  * After that we send a special "replay completed" request and change import
1437  * state to IMP_REPLAY_WAIT.
1438  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1439  * state and resend all requests from sending list.
1440  * After that we promote import to FULL state and send all delayed requests
1441  * and import is fully operational after that.
1442  *
1443  */
1444 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1445 {
1446         int rc = 0;
1447         int inflight;
1448         char *target_start;
1449         int target_len;
1450
1451         ENTRY;
1452         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1453                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1454                           &target_start, &target_len);
1455                 /* Don't care about MGC eviction */
1456                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1457                            LUSTRE_MGC_NAME) != 0) {
1458                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1459                                            "by %.*s; in progress operations "
1460                                            "using this service will fail.\n",
1461                                            imp->imp_obd->obd_name, target_len,
1462                                            target_start);
1463                 }
1464                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1465                        obd2cli_tgt(imp->imp_obd),
1466                        imp->imp_connection->c_remote_uuid.uuid);
1467                 /* reset vbr_failed flag upon eviction */
1468                 spin_lock(&imp->imp_lock);
1469                 imp->imp_vbr_failed = 0;
1470                 spin_unlock(&imp->imp_lock);
1471
1472                 {
1473                 struct task_struct *task;
1474                 /* bug 17802:  XXX client_disconnect_export vs connect request
1475                  * race. if client is evicted at this time then we start
1476                  * invalidate thread without reference to import and import can
1477                  * be freed at same time. */
1478                 class_import_get(imp);
1479                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1480                                      "ll_imp_inval");
1481                 if (IS_ERR(task)) {
1482                         class_import_put(imp);
1483                         CERROR("error starting invalidate thread: %d\n", rc);
1484                         rc = PTR_ERR(task);
1485                 } else {
1486                         rc = 0;
1487                 }
1488                 RETURN(rc);
1489                 }
1490         }
1491
1492         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1493                 CDEBUG(D_HA, "replay requested by %s\n",
1494                        obd2cli_tgt(imp->imp_obd));
1495                 rc = ptlrpc_replay_next(imp, &inflight);
1496                 if (inflight == 0 &&
1497                     atomic_read(&imp->imp_replay_inflight) == 0) {
1498                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1499                         rc = ldlm_replay_locks(imp);
1500                         if (rc)
1501                                 GOTO(out, rc);
1502                 }
1503                 rc = 0;
1504         }
1505
1506         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1507                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1508                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1509                         rc = signal_completed_replay(imp);
1510                         if (rc)
1511                                 GOTO(out, rc);
1512                 }
1513         }
1514
1515         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1516                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1517                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1518                 }
1519         }
1520
1521         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1522                 struct ptlrpc_connection *conn = imp->imp_connection;
1523
1524                 rc = ptlrpc_resend(imp);
1525                 if (rc)
1526                         GOTO(out, rc);
1527                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1528                 ptlrpc_activate_import(imp);
1529
1530                 LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n",
1531                               imp->imp_obd->obd_name,
1532                               obd_uuid2str(&conn->c_remote_uuid),
1533                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1534         }
1535
1536         if (imp->imp_state == LUSTRE_IMP_FULL) {
1537                 wake_up_all(&imp->imp_recovery_waitq);
1538                 ptlrpc_wake_delayed(imp);
1539         }
1540
1541 out:
1542         RETURN(rc);
1543 }
1544
1545 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1546 {
1547         struct ptlrpc_request *req;
1548         int rq_opc, rc = 0;
1549         ENTRY;
1550
1551         if (imp->imp_obd->obd_force)
1552                 GOTO(set_state, rc);
1553
1554         switch (imp->imp_connect_op) {
1555         case OST_CONNECT:
1556                 rq_opc = OST_DISCONNECT;
1557                 break;
1558         case MDS_CONNECT:
1559                 rq_opc = MDS_DISCONNECT;
1560                 break;
1561         case MGS_CONNECT:
1562                 rq_opc = MGS_DISCONNECT;
1563                 break;
1564         default:
1565                 rc = -EINVAL;
1566                 CERROR("%s: don't know how to disconnect from %s "
1567                        "(connect_op %d): rc = %d\n",
1568                        imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1569                        imp->imp_connect_op, rc);
1570                 RETURN(rc);
1571         }
1572
1573         if (ptlrpc_import_in_recovery(imp)) {
1574                 struct l_wait_info lwi;
1575                 long timeout_jiffies;
1576                 time64_t timeout;
1577
1578                 if (AT_OFF) {
1579                         if (imp->imp_server_timeout)
1580                                 timeout = obd_timeout >> 1;
1581                         else
1582                                 timeout = obd_timeout;
1583                 } else {
1584                         u32 req_portal;
1585                         int idx;
1586
1587                         req_portal = imp->imp_client->cli_request_portal;
1588                         idx = import_at_get_index(imp, req_portal);
1589                         timeout = at_get(&imp->imp_at.iat_service_estimate[idx]);
1590                 }
1591
1592                 timeout_jiffies = cfs_time_seconds(timeout);
1593                 lwi = LWI_TIMEOUT_INTR(max_t(long, timeout_jiffies, 1),
1594                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1595                 rc = l_wait_event(imp->imp_recovery_waitq,
1596                                   !ptlrpc_import_in_recovery(imp), &lwi);
1597
1598         }
1599
1600         spin_lock(&imp->imp_lock);
1601         if (imp->imp_state != LUSTRE_IMP_FULL)
1602                 GOTO(out, rc);
1603         spin_unlock(&imp->imp_lock);
1604
1605         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1606                                         LUSTRE_OBD_VERSION, rq_opc);
1607         if (req) {
1608                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1609                  * it fails.  We can get through the above with a down server
1610                  * if the client doesn't know the server is gone yet. */
1611                 req->rq_no_resend = 1;
1612
1613                 /* We want client umounts to happen quickly, no matter the
1614                    server state... */
1615                 req->rq_timeout = min_t(int, req->rq_timeout,
1616                                         INITIAL_CONNECT_TIMEOUT);
1617
1618                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1619                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1620                 ptlrpc_request_set_replen(req);
1621                 rc = ptlrpc_queue_wait(req);
1622                 ptlrpc_req_finished(req);
1623         }
1624
1625 set_state:
1626         spin_lock(&imp->imp_lock);
1627 out:
1628         if (noclose)
1629                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1630         else
1631                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1632         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1633         spin_unlock(&imp->imp_lock);
1634
1635         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1636                 rc = 0;
1637         RETURN(rc);
1638 }
1639 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1640
1641 void ptlrpc_cleanup_imp(struct obd_import *imp)
1642 {
1643         ENTRY;
1644
1645         spin_lock(&imp->imp_lock);
1646         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1647         imp->imp_generation++;
1648         spin_unlock(&imp->imp_lock);
1649         ptlrpc_abort_inflight(imp);
1650
1651         EXIT;
1652 }
1653
1654 /* Adaptive Timeout utils */
1655 extern unsigned int at_min, at_max, at_history;
1656
1657 /* Update at_current with the specified value (bounded by at_min and at_max),
1658  * as well as the AT history "bins".
1659  *  - Bin into timeslices using AT_BINS bins.
1660  *  - This gives us a max of the last at_history seconds without the storage,
1661  *    but still smoothing out a return to normalcy from a slow response.
1662  *  - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1663  */
1664 int at_measured(struct adaptive_timeout *at, unsigned int val)
1665 {
1666         unsigned int old = at->at_current;
1667         time64_t now = ktime_get_real_seconds();
1668         long binlimit = max_t(long, at_history / AT_BINS, 1);
1669
1670         LASSERT(at);
1671         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1672                val, at, (long)(now - at->at_binstart), at->at_current,
1673                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1674
1675         if (val == 0)
1676                 /* 0's don't count, because we never want our timeout to
1677                    drop to 0, and because 0 could mean an error */
1678                 return 0;
1679
1680         spin_lock(&at->at_lock);
1681
1682         if (unlikely(at->at_binstart == 0)) {
1683                 /* Special case to remove default from history */
1684                 at->at_current = val;
1685                 at->at_worst_ever = val;
1686                 at->at_worst_time = now;
1687                 at->at_hist[0] = val;
1688                 at->at_binstart = now;
1689         } else if (now - at->at_binstart < binlimit ) {
1690                 /* in bin 0 */
1691                 at->at_hist[0] = max(val, at->at_hist[0]);
1692                 at->at_current = max(val, at->at_current);
1693         } else {
1694                 int i, shift;
1695                 unsigned int maxv = val;
1696
1697                 /* move bins over */
1698                 shift = (u32)(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 }