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, 2014, 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 <obd_support.h>
  44 #include <lustre_ha.h>
  45 #include <lustre_net.h>
  46 #include <lustre_import.h>
  47 #include <lustre_export.h>
  48 #include <obd.h>
  49 #include <obd_cksum.h>
  50 #include <obd_class.h>
  51
  52 #include "ptlrpc_internal.h"
  53
  54 struct ptlrpc_connect_async_args {
  55          __u64 pcaa_peer_committed;
  56         int pcaa_initial_connect;
  57 };
  58
  59 /**
  60  * Updates import \a imp current state to provided \a state value
  61  * Helper function. Must be called under imp_lock.
  62  */
  63 static void __import_set_state(struct obd_import *imp,
  64                                enum lustre_imp_state state)
  65 {
  66         switch (state) {
  67         case LUSTRE_IMP_CLOSED:
  68         case LUSTRE_IMP_NEW:
  69         case LUSTRE_IMP_DISCON:
  70         case LUSTRE_IMP_CONNECTING:
  71                 break;
  72         case LUSTRE_IMP_REPLAY_WAIT:
  73                 imp->imp_replay_state = LUSTRE_IMP_REPLAY_LOCKS;
  74                 break;
  75         default:
  76                 imp->imp_replay_state = LUSTRE_IMP_REPLAY;
  77         }
  78         imp->imp_state = state;
  79         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
  80         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
  81                 cfs_time_current_sec();
  82         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
  83                 IMP_STATE_HIST_LEN;
  84 }
  85
  86 /* A CLOSED import should remain so. */
  87 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  88 do {                                                                           \
  89         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  90                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  91                       imp, obd2cli_tgt(imp->imp_obd),                          \
  92                       ptlrpc_import_state_name(imp->imp_state),                \
  93                       ptlrpc_import_state_name(state));                        \
  94                __import_set_state(imp, state);                                 \
  95         }                                                                      \
  96 } while(0)
  97
  98 #define IMPORT_SET_STATE(imp, state)                                    \
  99 do {                                                                    \
 100         spin_lock(&imp->imp_lock);                                      \
 101         IMPORT_SET_STATE_NOLOCK(imp, state);                            \
 102         spin_unlock(&imp->imp_lock);                                    \
 103 } while(0)
 104
 105
 106 static int ptlrpc_connect_interpret(const struct lu_env *env,
 107                                     struct ptlrpc_request *request,
 108                                     void * data, int rc);
 109 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
 110
 111 /* Only this function is allowed to change the import state when it is
 112  * CLOSED. I would rather refcount the import and free it after
 113  * disconnection like we do with exports. To do that, the client_obd
 114  * will need to save the peer info somewhere other than in the import,
 115  * though. */
 116 int ptlrpc_init_import(struct obd_import *imp)
 117 {
 118         spin_lock(&imp->imp_lock);
 119
 120         imp->imp_generation++;
 121         imp->imp_state =  LUSTRE_IMP_NEW;
 122
 123         spin_unlock(&imp->imp_lock);
 124
 125         return 0;
 126 }
 127 EXPORT_SYMBOL(ptlrpc_init_import);
 128
 129 #define UUID_STR "_UUID"
 130 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
 131 {
 132         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 133                 ? uuid : uuid + strlen(prefix);
 134
 135         *uuid_len = strlen(*uuid_start);
 136
 137         if (*uuid_len < strlen(UUID_STR))
 138                 return;
 139
 140         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 141                     UUID_STR, strlen(UUID_STR)))
 142                 *uuid_len -= strlen(UUID_STR);
 143 }
 144
 145 /**
 146  * Returns true if import was FULL, false if import was already not
 147  * connected.
 148  * @imp - import to be disconnected
 149  * @conn_cnt - connection count (epoch) of the request that timed out
 150  *             and caused the disconnection.  In some cases, multiple
 151  *             inflight requests can fail to a single target (e.g. OST
 152  *             bulk requests) and if one has already caused a reconnection
 153  *             (increasing the import->conn_cnt) the older failure should
 154  *             not also cause a reconnection.  If zero it forces a reconnect.
 155  */
 156 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 157 {
 158         int rc = 0;
 159
 160         spin_lock(&imp->imp_lock);
 161
 162         if (imp->imp_state == LUSTRE_IMP_FULL &&
 163             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 164                 char *target_start;
 165                 int   target_len;
 166
 167                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 168                           &target_start, &target_len);
 169
 170                 if (imp->imp_replayable) {
 171                         LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
 172                                "lost; in progress operations using this "
 173                                "service will wait for recovery to complete\n",
 174                                imp->imp_obd->obd_name, target_len, target_start,
 175                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 176                 } else {
 177                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
 178                                "%.*s (at %s) was lost; in progress "
 179                                "operations using this service will fail\n",
 180                                imp->imp_obd->obd_name,
 181                                target_len, target_start,
 182                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 183                 }
 184                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 185                 spin_unlock(&imp->imp_lock);
 186
 187                 if (obd_dump_on_timeout)
 188                         libcfs_debug_dumplog();
 189
 190                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 191                 rc = 1;
 192         } else {
 193                 spin_unlock(&imp->imp_lock);
 194                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 195                        imp->imp_client->cli_name, imp,
 196                        (imp->imp_state == LUSTRE_IMP_FULL &&
 197                         imp->imp_conn_cnt > conn_cnt) ?
 198                        "reconnected" : "not connected", imp->imp_conn_cnt,
 199                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 200         }
 201
 202         return rc;
 203 }
 204
 205 /* Must be called with imp_lock held! */
 206 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 207 {
 208         ENTRY;
 209         assert_spin_locked(&imp->imp_lock);
 210
 211         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 212         imp->imp_invalid = 1;
 213         imp->imp_generation++;
 214         spin_unlock(&imp->imp_lock);
 215
 216         ptlrpc_abort_inflight(imp);
 217         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 218
 219         EXIT;
 220 }
 221
 222 /*
 223  * This acts as a barrier; all existing requests are rejected, and
 224  * no new requests will be accepted until the import is valid again.
 225  */
 226 void ptlrpc_deactivate_import(struct obd_import *imp)
 227 {
 228         spin_lock(&imp->imp_lock);
 229         ptlrpc_deactivate_and_unlock_import(imp);
 230 }
 231 EXPORT_SYMBOL(ptlrpc_deactivate_import);
 232
 233 static unsigned int
 234 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
 235 {
 236         long dl;
 237
 238         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
 239               (req->rq_phase == RQ_PHASE_BULK) ||
 240               (req->rq_phase == RQ_PHASE_NEW)))
 241                 return 0;
 242
 243         if (req->rq_timedout)
 244                 return 0;
 245
 246         if (req->rq_phase == RQ_PHASE_NEW)
 247                 dl = req->rq_sent;
 248         else
 249                 dl = req->rq_deadline;
 250
 251         if (dl <= now)
 252                 return 0;
 253
 254         return dl - now;
 255 }
 256
 257 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
 258 {
 259         time_t now = cfs_time_current_sec();
 260         struct list_head *tmp, *n;
 261         struct ptlrpc_request *req;
 262         unsigned int timeout = 0;
 263
 264         spin_lock(&imp->imp_lock);
 265         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 266                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 267                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 268         }
 269         spin_unlock(&imp->imp_lock);
 270         return timeout;
 271 }
 272
 273 /**
 274  * This function will invalidate the import, if necessary, then block
 275  * for all the RPC completions, and finally notify the obd to
 276  * invalidate its state (ie cancel locks, clear pending requests,
 277  * etc).
 278  */
 279 void ptlrpc_invalidate_import(struct obd_import *imp)
 280 {
 281         struct list_head *tmp, *n;
 282         struct ptlrpc_request *req;
 283         struct l_wait_info lwi;
 284         unsigned int timeout;
 285         int rc;
 286
 287         atomic_inc(&imp->imp_inval_count);
 288
 289         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 290                 ptlrpc_deactivate_import(imp);
 291
 292         CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
 293         LASSERT(imp->imp_invalid);
 294
 295         /* Wait forever until inflight == 0. We really can't do it another
 296          * way because in some cases we need to wait for very long reply
 297          * unlink. We can't do anything before that because there is really
 298          * no guarantee that some rdma transfer is not in progress right now. */
 299         do {
 300                 /* Calculate max timeout for waiting on rpcs to error
 301                  * out. Use obd_timeout if calculated value is smaller
 302                  * than it. */
 303                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 304                         timeout = ptlrpc_inflight_timeout(imp);
 305                         timeout += timeout / 3;
 306
 307                         if (timeout == 0)
 308                                 timeout = obd_timeout;
 309                 } else {
 310                         /* decrease the interval to increase race condition */
 311                         timeout = 1;
 312                 }
 313
 314                 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
 315                        timeout);
 316
 317                 /* Wait for all requests to error out and call completion
 318                  * callbacks. Cap it at obd_timeout -- these should all
 319                  * have been locally cancelled by ptlrpc_abort_inflight. */
 320                 lwi = LWI_TIMEOUT_INTERVAL(
 321                         cfs_timeout_cap(cfs_time_seconds(timeout)),
 322                         (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
 323                         NULL, NULL);
 324                 rc = l_wait_event(imp->imp_recovery_waitq,
 325                                   (atomic_read(&imp->imp_inflight) == 0),
 326                                   &lwi);
 327                 if (rc) {
 328                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 329
 330                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 331                                cli_tgt, rc, atomic_read(&imp->imp_inflight));
 332
 333                         spin_lock(&imp->imp_lock);
 334                         if (atomic_read(&imp->imp_inflight) == 0) {
 335                                 int count = atomic_read(&imp->imp_unregistering);
 336
 337                                 /* We know that "unregistering" rpcs only can
 338                                  * survive in sending or delaying lists (they
 339                                  * maybe waiting for long reply unlink in
 340                                  * sluggish nets). Let's check this. If there
 341                                  * is no inflight and unregistering != 0, this
 342                                  * is bug. */
 343                                 LASSERTF(count == 0, "Some RPCs are still "
 344                                          "unregistering: %d\n", count);
 345
 346                                 /* Let's save one loop as soon as inflight have
 347                                  * dropped to zero. No new inflights possible at
 348                                  * this point. */
 349                                 rc = 0;
 350                         } else {
 351                                 list_for_each_safe(tmp, n,
 352                                                    &imp->imp_sending_list) {
 353                                         req = list_entry(tmp,
 354                                                          struct ptlrpc_request,
 355                                                          rq_list);
 356                                         DEBUG_REQ(D_ERROR, req,
 357                                                   "still on sending list");
 358                                 }
 359                                 list_for_each_safe(tmp, n,
 360                                                    &imp->imp_delayed_list) {
 361                                         req = list_entry(tmp,
 362                                                          struct ptlrpc_request,
 363                                                          rq_list);
 364                                         DEBUG_REQ(D_ERROR, req,
 365                                                   "still on delayed list");
 366                                 }
 367
 368                                 CERROR("%s: RPCs in \"%s\" phase found (%d). "
 369                                        "Network is sluggish? Waiting them "
 370                                        "to error out.\n", cli_tgt,
 371                                        ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
 372                                        atomic_read(&imp->imp_unregistering));
 373                         }
 374                         spin_unlock(&imp->imp_lock);
 375                 }
 376         } while (rc != 0);
 377
 378         /*
 379          * Let's additionally check that no new rpcs added to import in
 380          * "invalidate" state.
 381          */
 382         LASSERT(atomic_read(&imp->imp_inflight) == 0);
 383         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 384         sptlrpc_import_flush_all_ctx(imp);
 385
 386         atomic_dec(&imp->imp_inval_count);
 387         wake_up_all(&imp->imp_recovery_waitq);
 388 }
 389 EXPORT_SYMBOL(ptlrpc_invalidate_import);
 390
 391 /* unset imp_invalid */
 392 void ptlrpc_activate_import(struct obd_import *imp)
 393 {
 394         struct obd_device *obd = imp->imp_obd;
 395
 396         spin_lock(&imp->imp_lock);
 397         if (imp->imp_deactive != 0) {
 398                 spin_unlock(&imp->imp_lock);
 399                 return;
 400         }
 401
 402         imp->imp_invalid = 0;
 403         spin_unlock(&imp->imp_lock);
 404         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 405 }
 406 EXPORT_SYMBOL(ptlrpc_activate_import);
 407
 408 static void ptlrpc_pinger_force(struct obd_import *imp)
 409 {
 410         CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
 411                ptlrpc_import_state_name(imp->imp_state));
 412
 413         spin_lock(&imp->imp_lock);
 414         imp->imp_force_verify = 1;
 415         spin_unlock(&imp->imp_lock);
 416
 417         if (imp->imp_state != LUSTRE_IMP_CONNECTING)
 418                 ptlrpc_pinger_wake_up();
 419 }
 420
 421 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 422 {
 423         ENTRY;
 424
 425         LASSERT(!imp->imp_dlm_fake);
 426
 427         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 428                 if (!imp->imp_replayable) {
 429                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 430                                "auto-deactivating\n",
 431                                obd2cli_tgt(imp->imp_obd),
 432                                imp->imp_connection->c_remote_uuid.uuid,
 433                                imp->imp_obd->obd_name);
 434                         ptlrpc_deactivate_import(imp);
 435                 }
 436
 437                 ptlrpc_pinger_force(imp);
 438         }
 439         EXIT;
 440 }
 441
 442 int ptlrpc_reconnect_import(struct obd_import *imp)
 443 {
 444 #ifdef ENABLE_PINGER
 445         struct l_wait_info lwi;
 446         int secs = cfs_time_seconds(obd_timeout);
 447         int rc;
 448
 449         ptlrpc_pinger_force(imp);
 450
 451         CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 452                obd2cli_tgt(imp->imp_obd), secs);
 453
 454         lwi = LWI_TIMEOUT(secs, NULL, NULL);
 455         rc = l_wait_event(imp->imp_recovery_waitq,
 456                           !ptlrpc_import_in_recovery(imp), &lwi);
 457         CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
 458                ptlrpc_import_state_name(imp->imp_state));
 459         return rc;
 460 #else
 461         ptlrpc_set_import_discon(imp, 0);
 462         /* Force a new connect attempt */
 463         ptlrpc_invalidate_import(imp);
 464         /* Do a fresh connect next time by zeroing the handle */
 465         ptlrpc_disconnect_import(imp, 1);
 466         /* Wait for all invalidate calls to finish */
 467         if (atomic_read(&imp->imp_inval_count) > 0) {
 468                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 469                 int rc;
 470
 471                 rc = l_wait_event(imp->imp_recovery_waitq,
 472                                   (atomic_read(&imp->imp_inval_count) == 0),
 473                                   &lwi);
 474                 if (rc)
 475                         CERROR("Interrupted, inval=%d\n",
 476                                atomic_read(&imp->imp_inval_count));
 477         }
 478
 479         /* Allow reconnect attempts */
 480         imp->imp_obd->obd_no_recov = 0;
 481         /* Remove 'invalid' flag */
 482         ptlrpc_activate_import(imp);
 483         /* Attempt a new connect */
 484         ptlrpc_recover_import(imp, NULL, 0);
 485         return 0;
 486 #endif
 487 }
 488 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 489
 490 /**
 491  * Connection on import \a imp is changed to another one (if more than one is
 492  * present). We typically chose connection that we have not tried to connect to
 493  * the longest
 494  */
 495 static int import_select_connection(struct obd_import *imp)
 496 {
 497         struct obd_import_conn *imp_conn = NULL, *conn;
 498         struct obd_export *dlmexp;
 499         char *target_start;
 500         int target_len, tried_all = 1;
 501         ENTRY;
 502
 503         spin_lock(&imp->imp_lock);
 504
 505         if (list_empty(&imp->imp_conn_list)) {
 506                 CERROR("%s: no connections available\n",
 507                        imp->imp_obd->obd_name);
 508                 spin_unlock(&imp->imp_lock);
 509                 RETURN(-EINVAL);
 510         }
 511
 512         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 513                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 514                        imp->imp_obd->obd_name,
 515                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 516                        conn->oic_last_attempt);
 517
 518                 /* If we have not tried this connection since
 519                    the last successful attempt, go with this one */
 520                 if ((conn->oic_last_attempt == 0) ||
 521                     cfs_time_beforeq_64(conn->oic_last_attempt,
 522                                        imp->imp_last_success_conn)) {
 523                         imp_conn = conn;
 524                         tried_all = 0;
 525                         break;
 526                 }
 527
 528                 /* If all of the connections have already been tried
 529                    since the last successful connection; just choose the
 530                    least recently used */
 531                 if (!imp_conn)
 532                         imp_conn = conn;
 533                 else if (cfs_time_before_64(conn->oic_last_attempt,
 534                                             imp_conn->oic_last_attempt))
 535                         imp_conn = conn;
 536         }
 537
 538         /* if not found, simply choose the current one */
 539         if (!imp_conn || imp->imp_force_reconnect) {
 540                 LASSERT(imp->imp_conn_current);
 541                 imp_conn = imp->imp_conn_current;
 542                 tried_all = 0;
 543         }
 544         LASSERT(imp_conn->oic_conn);
 545
 546         /* If we've tried everything, and we're back to the beginning of the
 547            list, increase our timeout and try again. It will be reset when
 548            we do finally connect. (FIXME: really we should wait for all network
 549            state associated with the last connection attempt to drain before
 550            trying to reconnect on it.) */
 551         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 552                 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
 553                 if (at_get(at) < CONNECTION_SWITCH_MAX) {
 554                         at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
 555                         if (at_get(at) > CONNECTION_SWITCH_MAX)
 556                                 at_reset(at, CONNECTION_SWITCH_MAX);
 557                 }
 558                 LASSERT(imp_conn->oic_last_attempt);
 559                 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
 560                         "to %ds\n", imp->imp_obd->obd_name, at_get(at));
 561         }
 562
 563         imp_conn->oic_last_attempt = cfs_time_current_64();
 564
 565         /* switch connection, don't mind if it's same as the current one */
 566         if (imp->imp_connection)
 567                 ptlrpc_connection_put(imp->imp_connection);
 568         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 569
 570         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 571         LASSERT(dlmexp != NULL);
 572         if (dlmexp->exp_connection)
 573                 ptlrpc_connection_put(dlmexp->exp_connection);
 574         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 575         class_export_put(dlmexp);
 576
 577         if (imp->imp_conn_current != imp_conn) {
 578                 if (imp->imp_conn_current) {
 579                         deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 580                                   &target_start, &target_len);
 581
 582                         CDEBUG(D_HA, "%s: Connection changing to"
 583                                " %.*s (at %s)\n",
 584                                imp->imp_obd->obd_name,
 585                                target_len, target_start,
 586                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 587                 }
 588
 589                 imp->imp_conn_current = imp_conn;
 590         }
 591
 592         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 593                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 594                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 595
 596         spin_unlock(&imp->imp_lock);
 597
 598         RETURN(0);
 599 }
 600
 601 /*
 602  * must be called under imp_lock
 603  */
 604 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 605 {
 606         struct ptlrpc_request   *req;
 607         struct list_head        *tmp;
 608
 609         /* The requests in committed_list always have smaller transnos than
 610          * the requests in replay_list */
 611         if (!list_empty(&imp->imp_committed_list)) {
 612                 tmp = imp->imp_committed_list.next;
 613                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 614                 *transno = req->rq_transno;
 615                 if (req->rq_transno == 0) {
 616                         DEBUG_REQ(D_ERROR, req, "zero transno in committed_list");
 617                         LBUG();
 618                 }
 619                 return 1;
 620         }
 621         if (!list_empty(&imp->imp_replay_list)) {
 622                 tmp = imp->imp_replay_list.next;
 623                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 624                 *transno = req->rq_transno;
 625                 if (req->rq_transno == 0) {
 626                         DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
 627                         LBUG();
 628                 }
 629                 return 1;
 630         }
 631         return 0;
 632 }
 633
 634 /**
 635  * Attempt to (re)connect import \a imp. This includes all preparations,
 636  * initializing CONNECT RPC request and passing it to ptlrpcd for
 637  * actual sending.
 638  * Returns 0 on success or error code.
 639  */
 640 int ptlrpc_connect_import(struct obd_import *imp)
 641 {
 642         struct obd_device *obd = imp->imp_obd;
 643         int initial_connect = 0;
 644         int set_transno = 0;
 645         __u64 committed_before_reconnect = 0;
 646         struct ptlrpc_request *request;
 647         char *bufs[] = { NULL,
 648                          obd2cli_tgt(imp->imp_obd),
 649                          obd->obd_uuid.uuid,
 650                          (char *)&imp->imp_dlm_handle,
 651                          (char *)&imp->imp_connect_data };
 652         struct ptlrpc_connect_async_args *aa;
 653         int rc;
 654         ENTRY;
 655
 656         spin_lock(&imp->imp_lock);
 657         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 658                 spin_unlock(&imp->imp_lock);
 659                 CERROR("can't connect to a closed import\n");
 660                 RETURN(-EINVAL);
 661         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 662                 spin_unlock(&imp->imp_lock);
 663                 CERROR("already connected\n");
 664                 RETURN(0);
 665         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 666                 spin_unlock(&imp->imp_lock);
 667                 CERROR("already connecting\n");
 668                 RETURN(-EALREADY);
 669         }
 670
 671         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 672
 673         imp->imp_conn_cnt++;
 674         imp->imp_resend_replay = 0;
 675
 676         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 677                 initial_connect = 1;
 678         else
 679                 committed_before_reconnect = imp->imp_peer_committed_transno;
 680
 681         set_transno = ptlrpc_first_transno(imp,
 682                                            &imp->imp_connect_data.ocd_transno);
 683         spin_unlock(&imp->imp_lock);
 684
 685         rc = import_select_connection(imp);
 686         if (rc)
 687                 GOTO(out, rc);
 688
 689         rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
 690         if (rc)
 691                 GOTO(out, rc);
 692
 693         /* Reset connect flags to the originally requested flags, in case
 694          * the server is updated on-the-fly we will get the new features. */
 695         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 696         /* Reset ocd_version each time so the server knows the exact versions */
 697         imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
 698         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 699         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 700
 701         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 702                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 703         if (rc)
 704                 GOTO(out, rc);
 705
 706         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 707         if (request == NULL)
 708                 GOTO(out, rc = -ENOMEM);
 709
 710         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 711                                       imp->imp_connect_op, bufs, NULL);
 712         if (rc) {
 713                 ptlrpc_request_free(request);
 714                 GOTO(out, rc);
 715         }
 716
 717         /* Report the rpc service time to the server so that it knows how long
 718          * to wait for clients to join recovery */
 719         lustre_msg_set_service_time(request->rq_reqmsg,
 720                                     at_timeout2est(request->rq_timeout));
 721
 722         /* The amount of time we give the server to process the connect req.
 723          * import_select_connection will increase the net latency on
 724          * repeated reconnect attempts to cover slow networks.
 725          * We override/ignore the server rpc completion estimate here,
 726          * which may be large if this is a reconnect attempt */
 727         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 728         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 729
 730         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 731
 732         request->rq_no_resend = request->rq_no_delay = 1;
 733         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 734         /* Allow a slightly larger reply for future growth compatibility */
 735         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 736                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 737         ptlrpc_request_set_replen(request);
 738         request->rq_interpret_reply = ptlrpc_connect_interpret;
 739
 740         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 741         aa = ptlrpc_req_async_args(request);
 742         memset(aa, 0, sizeof *aa);
 743
 744         aa->pcaa_peer_committed = committed_before_reconnect;
 745         aa->pcaa_initial_connect = initial_connect;
 746
 747         if (aa->pcaa_initial_connect) {
 748                 spin_lock(&imp->imp_lock);
 749                 imp->imp_replayable = 1;
 750                 spin_unlock(&imp->imp_lock);
 751                 lustre_msg_add_op_flags(request->rq_reqmsg,
 752                                         MSG_CONNECT_INITIAL);
 753         }
 754
 755         if (set_transno)
 756                 lustre_msg_add_op_flags(request->rq_reqmsg,
 757                                         MSG_CONNECT_TRANSNO);
 758
 759         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
 760                   request->rq_timeout);
 761         ptlrpcd_add_req(request, PDL_POLICY_ROUND, -1);
 762         rc = 0;
 763 out:
 764         if (rc != 0) {
 765                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 766         }
 767
 768         RETURN(rc);
 769 }
 770 EXPORT_SYMBOL(ptlrpc_connect_import);
 771
 772 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 773 {
 774         int force_verify;
 775
 776         spin_lock(&imp->imp_lock);
 777         force_verify = imp->imp_force_verify != 0;
 778         spin_unlock(&imp->imp_lock);
 779
 780         if (force_verify)
 781                 ptlrpc_pinger_wake_up();
 782 }
 783
 784 static int ptlrpc_busy_reconnect(int rc)
 785 {
 786         return (rc == -EBUSY) || (rc == -EAGAIN);
 787 }
 788
 789 static int ptlrpc_connect_set_flags(struct obd_import *imp,
 790                                      struct obd_connect_data *ocd,
 791                                      __u64 old_connect_flags,
 792                                      struct obd_export *exp, int init_connect)
 793 {
 794         static bool warned;
 795         struct client_obd *cli = &imp->imp_obd->u.cli;
 796
 797         if ((imp->imp_connect_flags_orig & OBD_CONNECT_IBITS) &&
 798             !(ocd->ocd_connect_flags & OBD_CONNECT_IBITS)) {
 799                 LCONSOLE_WARN("%s: MDS %s does not support ibits "
 800                               "lock, either very old or invalid: "
 801                               "requested "LPX64", replied "LPX64"\n",
 802                               imp->imp_obd->obd_name,
 803                               imp->imp_connection->c_remote_uuid.uuid,
 804                               imp->imp_connect_flags_orig,
 805                               ocd->ocd_connect_flags);
 806                 return -EPROTO;
 807         }
 808
 809         spin_lock(&imp->imp_lock);
 810         list_del(&imp->imp_conn_current->oic_item);
 811         list_add(&imp->imp_conn_current->oic_item,
 812                  &imp->imp_conn_list);
 813         imp->imp_last_success_conn =
 814                 imp->imp_conn_current->oic_last_attempt;
 815
 816         spin_unlock(&imp->imp_lock);
 817
 818
 819         if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 820             (ocd->ocd_version > LUSTRE_VERSION_CODE +
 821                                 LUSTRE_VERSION_OFFSET_WARN ||
 822              ocd->ocd_version < LUSTRE_VERSION_CODE -
 823                                 LUSTRE_VERSION_OFFSET_WARN)) {
 824                 /* Sigh, some compilers do not like #ifdef in the middle
 825                    of macro arguments */
 826                 const char *older = "older than client. "
 827                                     "Consider upgrading server";
 828                 const char *newer = "newer than client. "
 829                                     "Consider recompiling application";
 830
 831                 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 832                               "is much %s (%s)\n",
 833                               obd2cli_tgt(imp->imp_obd),
 834                               OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 835                               OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 836                               OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 837                               OBD_OCD_VERSION_FIX(ocd->ocd_version),
 838                               ocd->ocd_version > LUSTRE_VERSION_CODE ?
 839                               newer : older, LUSTRE_VERSION_STRING);
 840                 warned = true;
 841         }
 842
 843 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
 844         /* Check if server has LU-1252 fix applied to not always swab
 845          * the IR MNE entries. Do this only once per connection.  This
 846          * fixup is version-limited, because we don't want to carry the
 847          * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
 848          * need interop with unpatched 2.2 servers.  For newer servers,
 849          * the client will do MNE swabbing only as needed.  LU-1644 */
 850         if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 851                      !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
 852                      OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
 853                      OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
 854                      OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
 855                      strcmp(imp->imp_obd->obd_type->typ_name,
 856                             LUSTRE_MGC_NAME) == 0))
 857                 imp->imp_need_mne_swab = 1;
 858         else /* clear if server was upgraded since last connect */
 859                 imp->imp_need_mne_swab = 0;
 860 #endif
 861
 862         if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
 863                 /* We sent to the server ocd_cksum_types with bits set
 864                  * for algorithms we understand. The server masked off
 865                  * the checksum types it doesn't support */
 866                 if ((ocd->ocd_cksum_types &
 867                      cksum_types_supported_client()) == 0) {
 868                         LCONSOLE_WARN("The negotiation of the checksum "
 869                                       "alogrithm to use with server %s "
 870                                       "failed (%x/%x), disabling "
 871                                       "checksums\n",
 872                                       obd2cli_tgt(imp->imp_obd),
 873                                       ocd->ocd_cksum_types,
 874                                       cksum_types_supported_client());
 875                         cli->cl_checksum = 0;
 876                         cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
 877                 } else {
 878                         cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
 879                 }
 880         } else {
 881                 /* The server does not support OBD_CONNECT_CKSUM.
 882                  * Enforce ADLER for backward compatibility*/
 883                 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
 884         }
 885         cli->cl_cksum_type = cksum_type_select(cli->cl_supp_cksum_types);
 886
 887         if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
 888                 cli->cl_max_pages_per_rpc =
 889                         min(ocd->ocd_brw_size >> PAGE_CACHE_SHIFT,
 890                             cli->cl_max_pages_per_rpc);
 891         else if (imp->imp_connect_op == MDS_CONNECT ||
 892                  imp->imp_connect_op == MGS_CONNECT)
 893                 cli->cl_max_pages_per_rpc = 1;
 894
 895         LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 896                 (cli->cl_max_pages_per_rpc > 0));
 897
 898         client_adjust_max_dirty(cli);
 899
 900
 901         /* Reset ns_connect_flags only for initial connect. It might be
 902          * changed in while using FS and if we reset it in reconnect
 903          * this leads to losing user settings done before such as
 904          * disable lru_resize, etc. */
 905         if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
 906                 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
 907                              "flags: "LPX64"\n", imp->imp_obd->obd_name,
 908                              ocd->ocd_connect_flags);
 909                 imp->imp_obd->obd_namespace->ns_connect_flags =
 910                         ocd->ocd_connect_flags;
 911                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
 912                         ocd->ocd_connect_flags;
 913         }
 914
 915         if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
 916                 (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
 917                 /* We need a per-message support flag, because
 918                  * a. we don't know if the incoming connect reply
 919                  *    supports AT or not (in reply_in_callback)
 920                  *    until we unpack it.
 921                  * b. failovered server means export and flags are gone
 922                  *    (in ptlrpc_send_reply).
 923                  *    Can only be set when we know AT is supported at
 924                  *    both ends */
 925                 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 926         else
 927                 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 928
 929         if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) &&
 930             (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
 931                 imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 932         else
 933                 imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 934
 935         return 0;
 936 }
 937
 938 /**
 939  * interpret_reply callback for connect RPCs.
 940  * Looks into returned status of connect operation and decides
 941  * what to do with the import - i.e enter recovery, promote it to
 942  * full state for normal operations of disconnect it due to an error.
 943  */
 944 static int ptlrpc_connect_interpret(const struct lu_env *env,
 945                                     struct ptlrpc_request *request,
 946                                     void *data, int rc)
 947 {
 948         struct ptlrpc_connect_async_args *aa = data;
 949         struct obd_import *imp = request->rq_import;
 950         struct lustre_handle old_hdl;
 951         __u64 old_connect_flags;
 952         int msg_flags;
 953         struct obd_connect_data *ocd;
 954         struct obd_export *exp;
 955         int ret;
 956         ENTRY;
 957
 958         spin_lock(&imp->imp_lock);
 959         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 960                 imp->imp_connect_tried = 1;
 961                 spin_unlock(&imp->imp_lock);
 962                 RETURN(0);
 963         }
 964
 965         if (rc) {
 966                 /* if this reconnect to busy export - not need select new target
 967                  * for connecting*/
 968                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 969                 spin_unlock(&imp->imp_lock);
 970                 ptlrpc_maybe_ping_import_soon(imp);
 971                 GOTO(out, rc);
 972         }
 973         spin_unlock(&imp->imp_lock);
 974
 975         LASSERT(imp->imp_conn_current);
 976
 977         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 978
 979         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 980                                    RCL_SERVER);
 981         /* server replied obd_connect_data is always bigger */
 982         ocd = req_capsule_server_sized_get(&request->rq_pill,
 983                                            &RMF_CONNECT_DATA, ret);
 984
 985         if (ocd == NULL) {
 986                 CERROR("%s: no connect data from server\n",
 987                        imp->imp_obd->obd_name);
 988                 rc = -EPROTO;
 989                 GOTO(out, rc);
 990         }
 991
 992         spin_lock(&imp->imp_lock);
 993
 994         /* All imports are pingable */
 995         imp->imp_pingable = 1;
 996         imp->imp_force_reconnect = 0;
 997         imp->imp_force_verify = 0;
 998
 999         imp->imp_connect_data = *ocd;
1000
1001         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1002                imp->imp_obd->obd_name, ocd->ocd_instance);
1003         exp = class_conn2export(&imp->imp_dlm_handle);
1004
1005         spin_unlock(&imp->imp_lock);
1006
1007         /* check that server granted subset of flags we asked for. */
1008         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1009             ocd->ocd_connect_flags) {
1010                 CERROR("%s: Server didn't granted asked subset of flags: "
1011                        "asked="LPX64" grranted="LPX64"\n",
1012                        imp->imp_obd->obd_name,imp->imp_connect_flags_orig,
1013                        ocd->ocd_connect_flags);
1014                 GOTO(out, rc = -EPROTO);
1015         }
1016
1017         if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1018             (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1019             (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1020             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1021                 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1022                 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1023                 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1024
1025                 /* We do not support the MDT-MDT interoperations with
1026                  * different version MDT because of protocol changes. */
1027                 if (unlikely(major != LUSTRE_MAJOR ||
1028                              minor != LUSTRE_MINOR ||
1029                              abs(patch - LUSTRE_PATCH) > 3)) {
1030                         LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1031                                       "connection to different version MDT "
1032                                       "(%d.%d.%d.%d) %s\n",
1033                                       imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1034                                       LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1035                                       major, minor, patch,
1036                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1037                                       imp->imp_connection->c_remote_uuid.uuid);
1038
1039                         GOTO(out, rc = -EPROTO);
1040                 }
1041         }
1042
1043         if (!exp) {
1044                 /* This could happen if export is cleaned during the
1045                    connect attempt */
1046                 CERROR("%s: missing export after connect\n",
1047                        imp->imp_obd->obd_name);
1048                 GOTO(out, rc = -ENODEV);
1049         }
1050         old_connect_flags = exp_connect_flags(exp);
1051         exp->exp_connect_data = *ocd;
1052         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1053
1054         /* The net statistics after (re-)connect is not valid anymore,
1055          * because may reflect other routing, etc. */
1056         at_init(&imp->imp_at.iat_net_latency, 0, 0);
1057         ptlrpc_at_adj_net_latency(request,
1058                         lustre_msg_get_service_time(request->rq_repmsg));
1059
1060         /* Import flags should be updated before waking import at FULL state */
1061         rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1062                                  aa->pcaa_initial_connect);
1063         class_export_put(exp);
1064         if (rc != 0)
1065                 GOTO(out, rc);
1066
1067         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1068
1069         if (aa->pcaa_initial_connect) {
1070                 spin_lock(&imp->imp_lock);
1071                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1072                         imp->imp_replayable = 1;
1073                         spin_unlock(&imp->imp_lock);
1074                         CDEBUG(D_HA, "connected to replayable target: %s\n",
1075                                obd2cli_tgt(imp->imp_obd));
1076                 } else {
1077                         imp->imp_replayable = 0;
1078                         spin_unlock(&imp->imp_lock);
1079                 }
1080
1081                 /* if applies, adjust the imp->imp_msg_magic here
1082                  * according to reply flags */
1083
1084                 imp->imp_remote_handle =
1085                                 *lustre_msg_get_handle(request->rq_repmsg);
1086
1087                 /* Initial connects are allowed for clients with non-random
1088                  * uuids when servers are in recovery.  Simply signal the
1089                  * servers replay is complete and wait in REPLAY_WAIT. */
1090                 if (msg_flags & MSG_CONNECT_RECOVERING) {
1091                         CDEBUG(D_HA, "connect to %s during recovery\n",
1092                                obd2cli_tgt(imp->imp_obd));
1093                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1094                 } else {
1095                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1096                         ptlrpc_activate_import(imp);
1097                 }
1098
1099                 GOTO(finish, rc = 0);
1100         }
1101
1102         /* Determine what recovery state to move the import to. */
1103         if (MSG_CONNECT_RECONNECT & msg_flags) {
1104                 memset(&old_hdl, 0, sizeof(old_hdl));
1105                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1106                             sizeof (old_hdl))) {
1107                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1108                                       "bad handle "LPX64"\n",
1109                                       obd2cli_tgt(imp->imp_obd),
1110                                       imp->imp_connection->c_remote_uuid.uuid,
1111                                       imp->imp_dlm_handle.cookie);
1112                         GOTO(out, rc = -ENOTCONN);
1113                 }
1114
1115                 if (memcmp(&imp->imp_remote_handle,
1116                            lustre_msg_get_handle(request->rq_repmsg),
1117                            sizeof(imp->imp_remote_handle))) {
1118                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
1119                                 D_HA : D_WARNING;
1120
1121                         /* Bug 16611/14775: if server handle have changed,
1122                          * that means some sort of disconnection happened.
1123                          * If the server is not in recovery, that also means it
1124                          * already erased all of our state because of previous
1125                          * eviction. If it is in recovery - we are safe to
1126                          * participate since we can reestablish all of our state
1127                          * with server again */
1128                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1129                                 CDEBUG(level,"%s@%s changed server handle from "
1130                                        LPX64" to "LPX64
1131                                        " but is still in recovery\n",
1132                                        obd2cli_tgt(imp->imp_obd),
1133                                        imp->imp_connection->c_remote_uuid.uuid,
1134                                        imp->imp_remote_handle.cookie,
1135                                        lustre_msg_get_handle(
1136                                        request->rq_repmsg)->cookie);
1137                         } else {
1138                                 LCONSOLE_WARN("Evicted from %s (at %s) "
1139                                               "after server handle changed from "
1140                                               LPX64" to "LPX64"\n",
1141                                               obd2cli_tgt(imp->imp_obd),
1142                                               imp->imp_connection-> \
1143                                               c_remote_uuid.uuid,
1144                                               imp->imp_remote_handle.cookie,
1145                                               lustre_msg_get_handle(
1146                                               request->rq_repmsg)->cookie);
1147                         }
1148
1149
1150                         imp->imp_remote_handle =
1151                                      *lustre_msg_get_handle(request->rq_repmsg);
1152
1153                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1154                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1155                                 GOTO(finish, rc = 0);
1156                         }
1157
1158                 } else {
1159                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1160                                obd2cli_tgt(imp->imp_obd),
1161                                imp->imp_connection->c_remote_uuid.uuid);
1162                 }
1163
1164                 if (imp->imp_invalid) {
1165                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1166                                "marking evicted\n", imp->imp_obd->obd_name);
1167                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1168                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1169                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1170                                imp->imp_obd->obd_name,
1171                                obd2cli_tgt(imp->imp_obd));
1172
1173                         spin_lock(&imp->imp_lock);
1174                         imp->imp_resend_replay = 1;
1175                         spin_unlock(&imp->imp_lock);
1176
1177                         IMPORT_SET_STATE(imp, imp->imp_replay_state);
1178                 } else {
1179                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1180                 }
1181         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1182                 LASSERT(imp->imp_replayable);
1183                 imp->imp_remote_handle =
1184                                 *lustre_msg_get_handle(request->rq_repmsg);
1185                 imp->imp_last_replay_transno = 0;
1186                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
1187         } else {
1188                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
1189                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
1190                 imp->imp_remote_handle =
1191                                 *lustre_msg_get_handle(request->rq_repmsg);
1192                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1193         }
1194
1195         /* Sanity checks for a reconnected import. */
1196         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
1197                 CERROR("imp_replayable flag does not match server "
1198                        "after reconnect. We should LBUG right here.\n");
1199         }
1200
1201         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1202             lustre_msg_get_last_committed(request->rq_repmsg) <
1203             aa->pcaa_peer_committed) {
1204                 CERROR("%s went back in time (transno "LPD64
1205                        " was previously committed, server now claims "LPD64
1206                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
1207                        "id=9646\n",
1208                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1209                        lustre_msg_get_last_committed(request->rq_repmsg));
1210         }
1211
1212 finish:
1213         rc = ptlrpc_import_recovery_state_machine(imp);
1214         if (rc == -ENOTCONN) {
1215                 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1216                        "invalidating and reconnecting\n",
1217                        obd2cli_tgt(imp->imp_obd),
1218                        imp->imp_connection->c_remote_uuid.uuid);
1219                 ptlrpc_connect_import(imp);
1220                 imp->imp_connect_tried = 1;
1221                 RETURN(0);
1222         }
1223
1224 out:
1225         imp->imp_connect_tried = 1;
1226
1227         if (rc != 0) {
1228                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1229                 if (rc == -EACCES) {
1230                         /*
1231                          * Give up trying to reconnect
1232                          * EACCES means client has no permission for connection
1233                          */
1234                         imp->imp_obd->obd_no_recov = 1;
1235                         ptlrpc_deactivate_import(imp);
1236                 }
1237
1238                 if (rc == -EPROTO) {
1239                         struct obd_connect_data *ocd;
1240
1241                         /* reply message might not be ready */
1242                         if (request->rq_repmsg == NULL)
1243                                 RETURN(-EPROTO);
1244
1245                         ocd = req_capsule_server_get(&request->rq_pill,
1246                                                      &RMF_CONNECT_DATA);
1247                         if (ocd &&
1248                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1249                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1250                            /* Actually servers are only supposed to refuse
1251                               connection from liblustre clients, so we should
1252                               never see this from VFS context */
1253                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1254                                         "(%d.%d.%d.%d)"
1255                                         " refused connection from this client "
1256                                         "with an incompatible version (%s).  "
1257                                         "Client must be recompiled\n",
1258                                         obd2cli_tgt(imp->imp_obd),
1259                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1260                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1261                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1262                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1263                                         LUSTRE_VERSION_STRING);
1264                                 ptlrpc_deactivate_import(imp);
1265                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1266                         }
1267                         RETURN(-EPROTO);
1268                 }
1269
1270                 ptlrpc_maybe_ping_import_soon(imp);
1271
1272                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1273                        obd2cli_tgt(imp->imp_obd),
1274                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1275         }
1276
1277         wake_up_all(&imp->imp_recovery_waitq);
1278         RETURN(rc);
1279 }
1280
1281 /**
1282  * interpret callback for "completed replay" RPCs.
1283  * \see signal_completed_replay
1284  */
1285 static int completed_replay_interpret(const struct lu_env *env,
1286                                       struct ptlrpc_request *req,
1287                                       void * data, int rc)
1288 {
1289         ENTRY;
1290         atomic_dec(&req->rq_import->imp_replay_inflight);
1291         if (req->rq_status == 0 &&
1292             !req->rq_import->imp_vbr_failed) {
1293                 ptlrpc_import_recovery_state_machine(req->rq_import);
1294         } else {
1295                 if (req->rq_import->imp_vbr_failed) {
1296                         CDEBUG(D_WARNING,
1297                                "%s: version recovery fails, reconnecting\n",
1298                                req->rq_import->imp_obd->obd_name);
1299                 } else {
1300                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1301                                      "reconnecting\n",
1302                                req->rq_import->imp_obd->obd_name,
1303                                req->rq_status);
1304                 }
1305                 ptlrpc_connect_import(req->rq_import);
1306         }
1307
1308         RETURN(0);
1309 }
1310
1311 /**
1312  * Let server know that we have no requests to replay anymore.
1313  * Achieved by just sending a PING request
1314  */
1315 static int signal_completed_replay(struct obd_import *imp)
1316 {
1317         struct ptlrpc_request *req;
1318         ENTRY;
1319
1320         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1321                 RETURN(0);
1322
1323         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1324         atomic_inc(&imp->imp_replay_inflight);
1325
1326         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1327                                         OBD_PING);
1328         if (req == NULL) {
1329                 atomic_dec(&imp->imp_replay_inflight);
1330                 RETURN(-ENOMEM);
1331         }
1332
1333         ptlrpc_request_set_replen(req);
1334         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1335         lustre_msg_add_flags(req->rq_reqmsg,
1336                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1337         if (AT_OFF)
1338                 req->rq_timeout *= 3;
1339         req->rq_interpret_reply = completed_replay_interpret;
1340
1341         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
1342         RETURN(0);
1343 }
1344
1345 /**
1346  * In kernel code all import invalidation happens in its own
1347  * separate thread, so that whatever application happened to encounter
1348  * a problem could still be killed or otherwise continue
1349  */
1350 static int ptlrpc_invalidate_import_thread(void *data)
1351 {
1352         struct obd_import *imp = data;
1353
1354         ENTRY;
1355
1356         unshare_fs_struct();
1357
1358         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1359                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1360                imp->imp_connection->c_remote_uuid.uuid);
1361
1362         ptlrpc_invalidate_import(imp);
1363
1364         if (obd_dump_on_eviction) {
1365                 CERROR("dump the log upon eviction\n");
1366                 libcfs_debug_dumplog();
1367         }
1368
1369         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1370         ptlrpc_import_recovery_state_machine(imp);
1371
1372         class_import_put(imp);
1373         RETURN(0);
1374 }
1375
1376 /**
1377  * This is the state machine for client-side recovery on import.
1378  *
1379  * Typicaly we have two possibly paths. If we came to server and it is not
1380  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1381  * state and reconnect from scratch.
1382  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1383  * We go through our list of requests to replay and send them to server one by
1384  * one.
1385  * After sending all request from the list we change import state to
1386  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1387  * and also all the locks we don't yet have and wait for server to grant us.
1388  * After that we send a special "replay completed" request and change import
1389  * state to IMP_REPLAY_WAIT.
1390  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1391  * state and resend all requests from sending list.
1392  * After that we promote import to FULL state and send all delayed requests
1393  * and import is fully operational after that.
1394  *
1395  */
1396 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1397 {
1398         int rc = 0;
1399         int inflight;
1400         char *target_start;
1401         int target_len;
1402
1403         ENTRY;
1404         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1405                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1406                           &target_start, &target_len);
1407                 /* Don't care about MGC eviction */
1408                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1409                            LUSTRE_MGC_NAME) != 0) {
1410                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1411                                            "by %.*s; in progress operations "
1412                                            "using this service will fail.\n",
1413                                            imp->imp_obd->obd_name, target_len,
1414                                            target_start);
1415                 }
1416                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1417                        obd2cli_tgt(imp->imp_obd),
1418                        imp->imp_connection->c_remote_uuid.uuid);
1419                 /* reset vbr_failed flag upon eviction */
1420                 spin_lock(&imp->imp_lock);
1421                 imp->imp_vbr_failed = 0;
1422                 spin_unlock(&imp->imp_lock);
1423
1424                 {
1425                 struct task_struct *task;
1426                 /* bug 17802:  XXX client_disconnect_export vs connect request
1427                  * race. if client is evicted at this time then we start
1428                  * invalidate thread without reference to import and import can
1429                  * be freed at same time. */
1430                 class_import_get(imp);
1431                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1432                                      "ll_imp_inval");
1433                 if (IS_ERR(task)) {
1434                         class_import_put(imp);
1435                         CERROR("error starting invalidate thread: %d\n", rc);
1436                         rc = PTR_ERR(task);
1437                 } else {
1438                         rc = 0;
1439                 }
1440                 RETURN(rc);
1441                 }
1442         }
1443
1444         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1445                 CDEBUG(D_HA, "replay requested by %s\n",
1446                        obd2cli_tgt(imp->imp_obd));
1447                 rc = ptlrpc_replay_next(imp, &inflight);
1448                 if (inflight == 0 &&
1449                     atomic_read(&imp->imp_replay_inflight) == 0) {
1450                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1451                         rc = ldlm_replay_locks(imp);
1452                         if (rc)
1453                                 GOTO(out, rc);
1454                 }
1455                 rc = 0;
1456         }
1457
1458         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1459                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1460                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1461                         rc = signal_completed_replay(imp);
1462                         if (rc)
1463                                 GOTO(out, rc);
1464                 }
1465         }
1466
1467         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1468                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1469                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1470                 }
1471         }
1472
1473         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1474                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1475                        obd2cli_tgt(imp->imp_obd),
1476                        imp->imp_connection->c_remote_uuid.uuid);
1477
1478                 rc = ptlrpc_resend(imp);
1479                 if (rc)
1480                         GOTO(out, rc);
1481                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1482                 ptlrpc_activate_import(imp);
1483
1484                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1485                           &target_start, &target_len);
1486                 LCONSOLE_INFO("%s: Connection restored to %.*s (at %s)\n",
1487                               imp->imp_obd->obd_name,
1488                               target_len, target_start,
1489                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1490         }
1491
1492         if (imp->imp_state == LUSTRE_IMP_FULL) {
1493                 wake_up_all(&imp->imp_recovery_waitq);
1494                 ptlrpc_wake_delayed(imp);
1495         }
1496
1497 out:
1498         RETURN(rc);
1499 }
1500
1501 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1502 {
1503         struct ptlrpc_request *req;
1504         int rq_opc, rc = 0;
1505         ENTRY;
1506
1507         if (imp->imp_obd->obd_force)
1508                 GOTO(set_state, rc);
1509
1510         switch (imp->imp_connect_op) {
1511         case OST_CONNECT:
1512                 rq_opc = OST_DISCONNECT;
1513                 break;
1514         case MDS_CONNECT:
1515                 rq_opc = MDS_DISCONNECT;
1516                 break;
1517         case MGS_CONNECT:
1518                 rq_opc = MGS_DISCONNECT;
1519                 break;
1520         default:
1521                 rc = -EINVAL;
1522                 CERROR("%s: don't know how to disconnect from %s "
1523                        "(connect_op %d): rc = %d\n",
1524                        imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1525                        imp->imp_connect_op, rc);
1526                 RETURN(rc);
1527         }
1528
1529         if (ptlrpc_import_in_recovery(imp)) {
1530                 struct l_wait_info lwi;
1531                 cfs_duration_t timeout;
1532
1533                 if (AT_OFF) {
1534                         if (imp->imp_server_timeout)
1535                                 timeout = cfs_time_seconds(obd_timeout / 2);
1536                         else
1537                                 timeout = cfs_time_seconds(obd_timeout);
1538                 } else {
1539                         int idx = import_at_get_index(imp,
1540                                 imp->imp_client->cli_request_portal);
1541                         timeout = cfs_time_seconds(
1542                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1543                 }
1544
1545                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1546                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1547                 rc = l_wait_event(imp->imp_recovery_waitq,
1548                                   !ptlrpc_import_in_recovery(imp), &lwi);
1549
1550         }
1551
1552         spin_lock(&imp->imp_lock);
1553         if (imp->imp_state != LUSTRE_IMP_FULL)
1554                 GOTO(out, rc);
1555         spin_unlock(&imp->imp_lock);
1556
1557         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1558                                         LUSTRE_OBD_VERSION, rq_opc);
1559         if (req) {
1560                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1561                  * it fails.  We can get through the above with a down server
1562                  * if the client doesn't know the server is gone yet. */
1563                 req->rq_no_resend = 1;
1564
1565                 /* We want client umounts to happen quickly, no matter the
1566                    server state... */
1567                 req->rq_timeout = min_t(int, req->rq_timeout,
1568                                         INITIAL_CONNECT_TIMEOUT);
1569
1570                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1571                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1572                 ptlrpc_request_set_replen(req);
1573                 rc = ptlrpc_queue_wait(req);
1574                 ptlrpc_req_finished(req);
1575         }
1576
1577 set_state:
1578         spin_lock(&imp->imp_lock);
1579 out:
1580         if (noclose)
1581                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1582         else
1583                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1584         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1585         spin_unlock(&imp->imp_lock);
1586
1587         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1588                 rc = 0;
1589         RETURN(rc);
1590 }
1591 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1592
1593 void ptlrpc_cleanup_imp(struct obd_import *imp)
1594 {
1595         ENTRY;
1596
1597         spin_lock(&imp->imp_lock);
1598         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1599         imp->imp_generation++;
1600         spin_unlock(&imp->imp_lock);
1601         ptlrpc_abort_inflight(imp);
1602
1603         EXIT;
1604 }
1605
1606 /* Adaptive Timeout utils */
1607 extern unsigned int at_min, at_max, at_history;
1608
1609 /* Update at_current with the specified value (bounded by at_min and at_max),
1610  * as well as the AT history "bins".
1611  *  - Bin into timeslices using AT_BINS bins.
1612  *  - This gives us a max of the last at_history seconds without the storage,
1613  *    but still smoothing out a return to normalcy from a slow response.
1614  *  - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1615  */
1616 int at_measured(struct adaptive_timeout *at, unsigned int val)
1617 {
1618         unsigned int old = at->at_current;
1619         time_t now = cfs_time_current_sec();
1620         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1621
1622         LASSERT(at);
1623         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1624                val, at, now - at->at_binstart, at->at_current,
1625                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1626
1627         if (val == 0)
1628                 /* 0's don't count, because we never want our timeout to
1629                    drop to 0, and because 0 could mean an error */
1630                 return 0;
1631
1632         spin_lock(&at->at_lock);
1633
1634         if (unlikely(at->at_binstart == 0)) {
1635                 /* Special case to remove default from history */
1636                 at->at_current = val;
1637                 at->at_worst_ever = val;
1638                 at->at_worst_time = now;
1639                 at->at_hist[0] = val;
1640                 at->at_binstart = now;
1641         } else if (now - at->at_binstart < binlimit ) {
1642                 /* in bin 0 */
1643                 at->at_hist[0] = max(val, at->at_hist[0]);
1644                 at->at_current = max(val, at->at_current);
1645         } else {
1646                 int i, shift;
1647                 unsigned int maxv = val;
1648                 /* move bins over */
1649                 shift = (now - at->at_binstart) / binlimit;
1650                 LASSERT(shift > 0);
1651                 for(i = AT_BINS - 1; i >= 0; i--) {
1652                         if (i >= shift) {
1653                                 at->at_hist[i] = at->at_hist[i - shift];
1654                                 maxv = max(maxv, at->at_hist[i]);
1655                         } else {
1656                                 at->at_hist[i] = 0;
1657                         }
1658                 }
1659                 at->at_hist[0] = val;
1660                 at->at_current = maxv;
1661                 at->at_binstart += shift * binlimit;
1662         }
1663
1664         if (at->at_current > at->at_worst_ever) {
1665                 at->at_worst_ever = at->at_current;
1666                 at->at_worst_time = now;
1667         }
1668
1669         if (at->at_flags & AT_FLG_NOHIST)
1670                 /* Only keep last reported val; keeping the rest of the history
1671                    for proc only */
1672                 at->at_current = val;
1673
1674         if (at_max > 0)
1675                 at->at_current =  min(at->at_current, at_max);
1676         at->at_current =  max(at->at_current, at_min);
1677
1678         if (at->at_current != old)
1679                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1680                        "(val=%u) hist %u %u %u %u\n", at,
1681                        old, at->at_current, at->at_current - old, val,
1682                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1683                        at->at_hist[3]);
1684
1685         /* if we changed, report the old value */
1686         old = (at->at_current != old) ? old : 0;
1687
1688         spin_unlock(&at->at_lock);
1689         return old;
1690 }
1691
1692 /* Find the imp_at index for a given portal; assign if space available */
1693 int import_at_get_index(struct obd_import *imp, int portal)
1694 {
1695         struct imp_at *at = &imp->imp_at;
1696         int i;
1697
1698         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1699                 if (at->iat_portal[i] == portal)
1700                         return i;
1701                 if (at->iat_portal[i] == 0)
1702                         /* unused */
1703                         break;
1704         }
1705
1706         /* Not found in list, add it under a lock */
1707         spin_lock(&imp->imp_lock);
1708
1709         /* Check unused under lock */
1710         for (; i < IMP_AT_MAX_PORTALS; i++) {
1711                 if (at->iat_portal[i] == portal)
1712                         goto out;
1713                 if (at->iat_portal[i] == 0)
1714                         /* unused */
1715                         break;
1716         }
1717
1718         /* Not enough portals? */
1719         LASSERT(i < IMP_AT_MAX_PORTALS);
1720
1721         at->iat_portal[i] = portal;
1722 out:
1723         spin_unlock(&imp->imp_lock);
1724         return i;
1725 }