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                                obd_import_nid2str(imp));
 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, target_len, target_start,
 184                                obd_import_nid2str(imp));
 185                 }
 186                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 187                 spin_unlock(&imp->imp_lock);
 188
 189                 if (obd_dump_on_timeout)
 190                         libcfs_debug_dumplog();
 191
 192                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 193                 rc = 1;
 194         } else {
 195                 spin_unlock(&imp->imp_lock);
 196                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 197                        imp->imp_client->cli_name, imp,
 198                        (imp->imp_state == LUSTRE_IMP_FULL &&
 199                         imp->imp_conn_cnt > conn_cnt) ?
 200                        "reconnected" : "not connected", imp->imp_conn_cnt,
 201                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 202         }
 203
 204         return rc;
 205 }
 206
 207 /* Must be called with imp_lock held! */
 208 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 209 {
 210         ENTRY;
 211         assert_spin_locked(&imp->imp_lock);
 212
 213         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 214         imp->imp_invalid = 1;
 215         imp->imp_generation++;
 216         spin_unlock(&imp->imp_lock);
 217
 218         ptlrpc_abort_inflight(imp);
 219         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 220
 221         EXIT;
 222 }
 223
 224 /*
 225  * This acts as a barrier; all existing requests are rejected, and
 226  * no new requests will be accepted until the import is valid again.
 227  */
 228 void ptlrpc_deactivate_import(struct obd_import *imp)
 229 {
 230         spin_lock(&imp->imp_lock);
 231         ptlrpc_deactivate_and_unlock_import(imp);
 232 }
 233 EXPORT_SYMBOL(ptlrpc_deactivate_import);
 234
 235 static time64_t ptlrpc_inflight_deadline(struct ptlrpc_request *req,
 236                                          time64_t now)
 237 {
 238         time64_t dl;
 239
 240         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
 241               (req->rq_phase == RQ_PHASE_BULK) ||
 242               (req->rq_phase == RQ_PHASE_NEW)))
 243                 return 0;
 244
 245         if (req->rq_timedout)
 246                 return 0;
 247
 248         if (req->rq_phase == RQ_PHASE_NEW)
 249                 dl = req->rq_sent;
 250         else
 251                 dl = req->rq_deadline;
 252
 253         if (dl <= now)
 254                 return 0;
 255
 256         return dl - now;
 257 }
 258
 259 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
 260 {
 261         time64_t now = ktime_get_real_seconds();
 262         struct list_head *tmp, *n;
 263         struct ptlrpc_request *req;
 264         time64_t timeout = 0;
 265
 266         spin_lock(&imp->imp_lock);
 267         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 268                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 269                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 270         }
 271         spin_unlock(&imp->imp_lock);
 272         return timeout;
 273 }
 274
 275 /**
 276  * This function will invalidate the import, if necessary, then block
 277  * for all the RPC completions, and finally notify the obd to
 278  * invalidate its state (ie cancel locks, clear pending requests,
 279  * etc).
 280  */
 281 void ptlrpc_invalidate_import(struct obd_import *imp)
 282 {
 283         struct list_head *tmp, *n;
 284         struct ptlrpc_request *req;
 285         struct l_wait_info lwi;
 286         time64_t timeout;
 287         int rc;
 288
 289         atomic_inc(&imp->imp_inval_count);
 290
 291         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 292                 ptlrpc_deactivate_import(imp);
 293
 294         CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
 295         LASSERT(imp->imp_invalid);
 296
 297         /* Wait forever until inflight == 0. We really can't do it another
 298          * way because in some cases we need to wait for very long reply
 299          * unlink. We can't do anything before that because there is really
 300          * no guarantee that some rdma transfer is not in progress right now. */
 301         do {
 302                 long timeout_jiffies;
 303
 304                 /* Calculate max timeout for waiting on rpcs to error
 305                  * out. Use obd_timeout if calculated value is smaller
 306                  * than it.
 307                  */
 308                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 309                         timeout = ptlrpc_inflight_timeout(imp);
 310                         timeout += timeout / 3;
 311
 312                         if (timeout == 0)
 313                                 timeout = obd_timeout;
 314                 } else {
 315                         /* decrease the interval to increase race condition */
 316                         timeout = 1;
 317                 }
 318
 319                 CDEBUG(D_RPCTRACE, "Sleeping %llds for inflight to error out\n",
 320                        timeout);
 321
 322                 /* Wait for all requests to error out and call completion
 323                  * callbacks. Cap it at obd_timeout -- these should all
 324                  * have been locally cancelled by ptlrpc_abort_inflight.
 325                  */
 326                 timeout_jiffies = max_t(long, cfs_time_seconds(timeout), 1);
 327                 lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
 328                                            (timeout > 1) ? cfs_time_seconds(1) :
 329                                                            cfs_time_seconds(1) / 2,
 330                                                            NULL, NULL);
 331                 rc = l_wait_event(imp->imp_recovery_waitq,
 332                                   (atomic_read(&imp->imp_inflight) == 0),
 333                                   &lwi);
 334                 if (rc) {
 335                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 336
 337                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 338                                cli_tgt, rc, atomic_read(&imp->imp_inflight));
 339
 340                         spin_lock(&imp->imp_lock);
 341                         if (atomic_read(&imp->imp_inflight) == 0) {
 342                                 int count = atomic_read(&imp->imp_unregistering);
 343
 344                                 /* We know that "unregistering" rpcs only can
 345                                  * survive in sending or delaying lists (they
 346                                  * maybe waiting for long reply unlink in
 347                                  * sluggish nets). Let's check this. If there
 348                                  * is no inflight and unregistering != 0, this
 349                                  * is bug. */
 350                                 LASSERTF(count == 0, "Some RPCs are still "
 351                                          "unregistering: %d\n", count);
 352
 353                                 /* Let's save one loop as soon as inflight have
 354                                  * dropped to zero. No new inflights possible at
 355                                  * this point. */
 356                                 rc = 0;
 357                         } else {
 358                                 list_for_each_safe(tmp, n,
 359                                                    &imp->imp_sending_list) {
 360                                         req = list_entry(tmp,
 361                                                          struct ptlrpc_request,
 362                                                          rq_list);
 363                                         DEBUG_REQ(D_ERROR, req,
 364                                                   "still on sending list");
 365                                 }
 366                                 list_for_each_safe(tmp, n,
 367                                                    &imp->imp_delayed_list) {
 368                                         req = list_entry(tmp,
 369                                                          struct ptlrpc_request,
 370                                                          rq_list);
 371                                         DEBUG_REQ(D_ERROR, req,
 372                                                   "still on delayed list");
 373                                 }
 374
 375                                 CERROR("%s: Unregistering RPCs found (%d). "
 376                                        "Network is sluggish? Waiting them "
 377                                        "to error out.\n", cli_tgt,
 378                                        atomic_read(&imp->imp_unregistering));
 379                         }
 380                         spin_unlock(&imp->imp_lock);
 381                 }
 382         } while (rc != 0);
 383
 384         /*
 385          * Let's additionally check that no new rpcs added to import in
 386          * "invalidate" state.
 387          */
 388         LASSERT(atomic_read(&imp->imp_inflight) == 0);
 389         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 390         sptlrpc_import_flush_all_ctx(imp);
 391
 392         atomic_dec(&imp->imp_inval_count);
 393         wake_up_all(&imp->imp_recovery_waitq);
 394 }
 395 EXPORT_SYMBOL(ptlrpc_invalidate_import);
 396
 397 /* unset imp_invalid */
 398 void ptlrpc_activate_import(struct obd_import *imp)
 399 {
 400         struct obd_device *obd = imp->imp_obd;
 401
 402         spin_lock(&imp->imp_lock);
 403         if (imp->imp_deactive != 0) {
 404                 spin_unlock(&imp->imp_lock);
 405                 return;
 406         }
 407
 408         imp->imp_invalid = 0;
 409         spin_unlock(&imp->imp_lock);
 410         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 411 }
 412 EXPORT_SYMBOL(ptlrpc_activate_import);
 413
 414 void ptlrpc_pinger_force(struct obd_import *imp)
 415 {
 416         CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
 417                ptlrpc_import_state_name(imp->imp_state));
 418
 419         spin_lock(&imp->imp_lock);
 420         imp->imp_force_verify = 1;
 421         spin_unlock(&imp->imp_lock);
 422
 423         if (imp->imp_state != LUSTRE_IMP_CONNECTING)
 424                 ptlrpc_pinger_wake_up();
 425 }
 426 EXPORT_SYMBOL(ptlrpc_pinger_force);
 427
 428 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 429 {
 430         ENTRY;
 431
 432         LASSERT(!imp->imp_dlm_fake);
 433
 434         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 435                 if (!imp->imp_replayable) {
 436                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 437                                "auto-deactivating\n",
 438                                obd2cli_tgt(imp->imp_obd),
 439                                imp->imp_connection->c_remote_uuid.uuid,
 440                                imp->imp_obd->obd_name);
 441                         ptlrpc_deactivate_import(imp);
 442                 }
 443
 444                 ptlrpc_pinger_force(imp);
 445         }
 446         EXIT;
 447 }
 448
 449 int ptlrpc_reconnect_import(struct obd_import *imp)
 450 {
 451 #ifdef ENABLE_PINGER
 452         long timeout_jiffies = cfs_time_seconds(obd_timeout);
 453         struct l_wait_info lwi;
 454         int rc;
 455
 456         ptlrpc_pinger_force(imp);
 457
 458         CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 459                obd2cli_tgt(imp->imp_obd), obd_timeout);
 460
 461         lwi = LWI_TIMEOUT(timeout_jiffies, NULL, NULL);
 462         rc = l_wait_event(imp->imp_recovery_waitq,
 463                           !ptlrpc_import_in_recovery(imp), &lwi);
 464         CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
 465                ptlrpc_import_state_name(imp->imp_state));
 466         return rc;
 467 #else
 468         ptlrpc_set_import_discon(imp, 0);
 469         /* Force a new connect attempt */
 470         ptlrpc_invalidate_import(imp);
 471         /* Do a fresh connect next time by zeroing the handle */
 472         ptlrpc_disconnect_import(imp, 1);
 473         /* Wait for all invalidate calls to finish */
 474         if (atomic_read(&imp->imp_inval_count) > 0) {
 475                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 476                 int rc;
 477
 478                 rc = l_wait_event(imp->imp_recovery_waitq,
 479                                   (atomic_read(&imp->imp_inval_count) == 0),
 480                                   &lwi);
 481                 if (rc)
 482                         CERROR("Interrupted, inval=%d\n",
 483                                atomic_read(&imp->imp_inval_count));
 484         }
 485
 486         /* Allow reconnect attempts */
 487         imp->imp_obd->obd_no_recov = 0;
 488         /* Remove 'invalid' flag */
 489         ptlrpc_activate_import(imp);
 490         /* Attempt a new connect */
 491         ptlrpc_recover_import(imp, NULL, 0);
 492         return 0;
 493 #endif
 494 }
 495 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 496
 497 /**
 498  * Connection on import \a imp is changed to another one (if more than one is
 499  * present). We typically chose connection that we have not tried to connect to
 500  * the longest
 501  */
 502 static int import_select_connection(struct obd_import *imp)
 503 {
 504         struct obd_import_conn *imp_conn = NULL, *conn;
 505         struct obd_export *dlmexp;
 506         char *target_start;
 507         int target_len, tried_all = 1;
 508         ENTRY;
 509
 510         spin_lock(&imp->imp_lock);
 511
 512         if (list_empty(&imp->imp_conn_list)) {
 513                 CERROR("%s: no connections available\n",
 514                        imp->imp_obd->obd_name);
 515                 spin_unlock(&imp->imp_lock);
 516                 RETURN(-EINVAL);
 517         }
 518
 519         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 520                 CDEBUG(D_HA, "%s: connect to NID %s last attempt %lld\n",
 521                        imp->imp_obd->obd_name,
 522                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 523                        conn->oic_last_attempt);
 524
 525                 /* If we have not tried this connection since
 526                    the last successful attempt, go with this one */
 527                 if ((conn->oic_last_attempt == 0) ||
 528                     conn->oic_last_attempt <= imp->imp_last_success_conn) {
 529                         imp_conn = conn;
 530                         tried_all = 0;
 531                         break;
 532                 }
 533
 534                 /* If all of the connections have already been tried
 535                    since the last successful connection; just choose the
 536                    least recently used */
 537                 if (!imp_conn)
 538                         imp_conn = conn;
 539                 else if (imp_conn->oic_last_attempt > conn->oic_last_attempt)
 540                         imp_conn = conn;
 541         }
 542
 543         /* if not found, simply choose the current one */
 544         if (!imp_conn || imp->imp_force_reconnect) {
 545                 LASSERT(imp->imp_conn_current);
 546                 imp_conn = imp->imp_conn_current;
 547                 tried_all = 0;
 548         }
 549         LASSERT(imp_conn->oic_conn);
 550
 551         /* If we've tried everything, and we're back to the beginning of the
 552            list, increase our timeout and try again. It will be reset when
 553            we do finally connect. (FIXME: really we should wait for all network
 554            state associated with the last connection attempt to drain before
 555            trying to reconnect on it.) */
 556         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 557                 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
 558                 if (at_get(at) < CONNECTION_SWITCH_MAX) {
 559                         at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
 560                         if (at_get(at) > CONNECTION_SWITCH_MAX)
 561                                 at_reset(at, CONNECTION_SWITCH_MAX);
 562                 }
 563                 LASSERT(imp_conn->oic_last_attempt);
 564                 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
 565                         "to %ds\n", imp->imp_obd->obd_name, at_get(at));
 566         }
 567
 568         imp_conn->oic_last_attempt = ktime_get_seconds();
 569
 570         /* switch connection, don't mind if it's same as the current one */
 571         if (imp->imp_connection)
 572                 ptlrpc_connection_put(imp->imp_connection);
 573         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 574
 575         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 576         LASSERT(dlmexp != NULL);
 577         if (dlmexp->exp_connection)
 578                 ptlrpc_connection_put(dlmexp->exp_connection);
 579         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 580         class_export_put(dlmexp);
 581
 582         if (imp->imp_conn_current != imp_conn) {
 583                 if (imp->imp_conn_current) {
 584                         deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 585                                   &target_start, &target_len);
 586
 587                         CDEBUG(D_HA, "%s: Connection changing to"
 588                                " %.*s (at %s)\n",
 589                                imp->imp_obd->obd_name,
 590                                target_len, target_start,
 591                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 592                 }
 593
 594                 imp->imp_conn_current = imp_conn;
 595         }
 596
 597         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 598                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 599                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 600
 601         spin_unlock(&imp->imp_lock);
 602
 603         RETURN(0);
 604 }
 605
 606 /*
 607  * must be called under imp_lock
 608  */
 609 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 610 {
 611         struct ptlrpc_request   *req;
 612         struct list_head        *tmp;
 613
 614         /* The requests in committed_list always have smaller transnos than
 615          * the requests in replay_list */
 616         if (!list_empty(&imp->imp_committed_list)) {
 617                 tmp = imp->imp_committed_list.next;
 618                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 619                 *transno = req->rq_transno;
 620                 if (req->rq_transno == 0) {
 621                         DEBUG_REQ(D_ERROR, req, "zero transno in committed_list");
 622                         LBUG();
 623                 }
 624                 return 1;
 625         }
 626         if (!list_empty(&imp->imp_replay_list)) {
 627                 tmp = imp->imp_replay_list.next;
 628                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 629                 *transno = req->rq_transno;
 630                 if (req->rq_transno == 0) {
 631                         DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
 632                         LBUG();
 633                 }
 634                 return 1;
 635         }
 636         return 0;
 637 }
 638
 639 /**
 640  * Attempt to (re)connect import \a imp. This includes all preparations,
 641  * initializing CONNECT RPC request and passing it to ptlrpcd for
 642  * actual sending.
 643  * Returns 0 on success or error code.
 644  */
 645 int ptlrpc_connect_import(struct obd_import *imp)
 646 {
 647         struct obd_device *obd = imp->imp_obd;
 648         int initial_connect = 0;
 649         int set_transno = 0;
 650         __u64 committed_before_reconnect = 0;
 651         struct ptlrpc_request *request;
 652         char *bufs[] = { NULL,
 653                          obd2cli_tgt(imp->imp_obd),
 654                          obd->obd_uuid.uuid,
 655                          (char *)&imp->imp_dlm_handle,
 656                          (char *)&imp->imp_connect_data };
 657         struct ptlrpc_connect_async_args *aa;
 658         int rc;
 659         ENTRY;
 660
 661         spin_lock(&imp->imp_lock);
 662         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 663                 spin_unlock(&imp->imp_lock);
 664                 CERROR("can't connect to a closed import\n");
 665                 RETURN(-EINVAL);
 666         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 667                 spin_unlock(&imp->imp_lock);
 668                 CERROR("already connected\n");
 669                 RETURN(0);
 670         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING ||
 671                    imp->imp_connected) {
 672                 spin_unlock(&imp->imp_lock);
 673                 CERROR("already connecting\n");
 674                 RETURN(-EALREADY);
 675         }
 676
 677         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 678
 679         imp->imp_conn_cnt++;
 680         imp->imp_resend_replay = 0;
 681
 682         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 683                 initial_connect = 1;
 684         else
 685                 committed_before_reconnect = imp->imp_peer_committed_transno;
 686
 687         set_transno = ptlrpc_first_transno(imp,
 688                                            &imp->imp_connect_data.ocd_transno);
 689         spin_unlock(&imp->imp_lock);
 690
 691         rc = import_select_connection(imp);
 692         if (rc)
 693                 GOTO(out, rc);
 694
 695         rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
 696         if (rc)
 697                 GOTO(out, rc);
 698
 699         /* Reset connect flags to the originally requested flags, in case
 700          * the server is updated on-the-fly we will get the new features. */
 701         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 702         imp->imp_connect_data.ocd_connect_flags2 = imp->imp_connect_flags2_orig;
 703         /* Reset ocd_version each time so the server knows the exact versions */
 704         imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
 705         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 706         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 707
 708         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 709                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 710         if (rc)
 711                 GOTO(out, rc);
 712
 713         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 714         if (request == NULL)
 715                 GOTO(out, rc = -ENOMEM);
 716
 717         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 718                                       imp->imp_connect_op, bufs, NULL);
 719         if (rc) {
 720                 ptlrpc_request_free(request);
 721                 GOTO(out, rc);
 722         }
 723
 724         /* Report the rpc service time to the server so that it knows how long
 725          * to wait for clients to join recovery */
 726         lustre_msg_set_service_time(request->rq_reqmsg,
 727                                     at_timeout2est(request->rq_timeout));
 728
 729         /* The amount of time we give the server to process the connect req.
 730          * import_select_connection will increase the net latency on
 731          * repeated reconnect attempts to cover slow networks.
 732          * We override/ignore the server rpc completion estimate here,
 733          * which may be large if this is a reconnect attempt */
 734         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 735         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 736
 737         request->rq_no_resend = request->rq_no_delay = 1;
 738         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 739         /* Allow a slightly larger reply for future growth compatibility */
 740         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 741                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 742         ptlrpc_request_set_replen(request);
 743         request->rq_interpret_reply = ptlrpc_connect_interpret;
 744
 745         CLASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
 746         aa = ptlrpc_req_async_args(request);
 747         memset(aa, 0, sizeof *aa);
 748
 749         aa->pcaa_peer_committed = committed_before_reconnect;
 750         aa->pcaa_initial_connect = initial_connect;
 751
 752         if (aa->pcaa_initial_connect) {
 753                 spin_lock(&imp->imp_lock);
 754                 imp->imp_replayable = 1;
 755                 spin_unlock(&imp->imp_lock);
 756                 lustre_msg_add_op_flags(request->rq_reqmsg,
 757                                         MSG_CONNECT_INITIAL);
 758         }
 759
 760         if (set_transno)
 761                 lustre_msg_add_op_flags(request->rq_reqmsg,
 762                                         MSG_CONNECT_TRANSNO);
 763
 764         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %ld)",
 765                   request->rq_timeout);
 766         ptlrpcd_add_req(request);
 767         rc = 0;
 768 out:
 769         if (rc != 0)
 770                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 771
 772         RETURN(rc);
 773 }
 774 EXPORT_SYMBOL(ptlrpc_connect_import);
 775
 776 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 777 {
 778         int force_verify;
 779
 780         spin_lock(&imp->imp_lock);
 781         force_verify = imp->imp_force_verify != 0;
 782         spin_unlock(&imp->imp_lock);
 783
 784         if (force_verify)
 785                 ptlrpc_pinger_wake_up();
 786 }
 787
 788 static int ptlrpc_busy_reconnect(int rc)
 789 {
 790         return (rc == -EBUSY) || (rc == -EAGAIN);
 791 }
 792
 793 static int ptlrpc_connect_set_flags(struct obd_import *imp,
 794                                      struct obd_connect_data *ocd,
 795                                      __u64 old_connect_flags,
 796                                      struct obd_export *exp, int init_connect)
 797 {
 798         static bool warned;
 799         struct client_obd *cli = &imp->imp_obd->u.cli;
 800
 801         spin_lock(&imp->imp_lock);
 802         list_del(&imp->imp_conn_current->oic_item);
 803         list_add(&imp->imp_conn_current->oic_item,
 804                  &imp->imp_conn_list);
 805         imp->imp_last_success_conn =
 806                 imp->imp_conn_current->oic_last_attempt;
 807
 808         spin_unlock(&imp->imp_lock);
 809
 810         if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 811             (ocd->ocd_version > LUSTRE_VERSION_CODE +
 812                                 LUSTRE_VERSION_OFFSET_WARN ||
 813              ocd->ocd_version < LUSTRE_VERSION_CODE -
 814                                 LUSTRE_VERSION_OFFSET_WARN)) {
 815                 /* Sigh, some compilers do not like #ifdef in the middle
 816                    of macro arguments */
 817                 const char *older = "older than client. "
 818                                     "Consider upgrading server";
 819                 const char *newer = "newer than client. "
 820                                     "Consider upgrading client";
 821
 822                 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 823                               "is much %s (%s)\n",
 824                               obd2cli_tgt(imp->imp_obd),
 825                               OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 826                               OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 827                               OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 828                               OBD_OCD_VERSION_FIX(ocd->ocd_version),
 829                               ocd->ocd_version > LUSTRE_VERSION_CODE ?
 830                               newer : older, LUSTRE_VERSION_STRING);
 831                 warned = true;
 832         }
 833
 834 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
 835         /* Check if server has LU-1252 fix applied to not always swab
 836          * the IR MNE entries. Do this only once per connection.  This
 837          * fixup is version-limited, because we don't want to carry the
 838          * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
 839          * need interop with unpatched 2.2 servers.  For newer servers,
 840          * the client will do MNE swabbing only as needed.  LU-1644 */
 841         if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 842                      !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
 843                      OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
 844                      OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
 845                      OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
 846                      strcmp(imp->imp_obd->obd_type->typ_name,
 847                             LUSTRE_MGC_NAME) == 0))
 848                 imp->imp_need_mne_swab = 1;
 849         else /* clear if server was upgraded since last connect */
 850                 imp->imp_need_mne_swab = 0;
 851 #endif
 852
 853         if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
 854                 /* We sent to the server ocd_cksum_types with bits set
 855                  * for algorithms we understand. The server masked off
 856                  * the checksum types it doesn't support */
 857                 if ((ocd->ocd_cksum_types &
 858                      cksum_types_supported_client()) == 0) {
 859                         LCONSOLE_ERROR("The negotiation of the checksum "
 860                                        "alogrithm to use with server %s "
 861                                        "failed (%x/%x)\n",
 862                                        obd2cli_tgt(imp->imp_obd),
 863                                        ocd->ocd_cksum_types,
 864                                        cksum_types_supported_client());
 865                         return -EPROTO;
 866                 } else {
 867                         cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
 868                 }
 869         } else {
 870                 /* The server does not support OBD_CONNECT_CKSUM.
 871                  * Enforce ADLER for backward compatibility*/
 872                 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
 873         }
 874         cli->cl_cksum_type = cksum_type_select(cli->cl_supp_cksum_types);
 875
 876         if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
 877                 cli->cl_max_pages_per_rpc =
 878                         min(ocd->ocd_brw_size >> PAGE_SHIFT,
 879                             cli->cl_max_pages_per_rpc);
 880         else if (imp->imp_connect_op == MDS_CONNECT ||
 881                  imp->imp_connect_op == MGS_CONNECT)
 882                 cli->cl_max_pages_per_rpc = 1;
 883
 884         LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 885                 (cli->cl_max_pages_per_rpc > 0));
 886
 887         client_adjust_max_dirty(cli);
 888
 889         /* Update client max modify RPCs in flight with value returned
 890          * by the server */
 891         if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
 892                 cli->cl_max_mod_rpcs_in_flight = min(
 893                                         cli->cl_max_mod_rpcs_in_flight,
 894                                         ocd->ocd_maxmodrpcs);
 895         else
 896                 cli->cl_max_mod_rpcs_in_flight = 1;
 897
 898         /* Reset ns_connect_flags only for initial connect. It might be
 899          * changed in while using FS and if we reset it in reconnect
 900          * this leads to losing user settings done before such as
 901          * disable lru_resize, etc. */
 902         if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
 903                 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
 904                              "flags: %#llx\n", imp->imp_obd->obd_name,
 905                              ocd->ocd_connect_flags);
 906                 imp->imp_obd->obd_namespace->ns_connect_flags =
 907                         ocd->ocd_connect_flags;
 908                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
 909                         ocd->ocd_connect_flags;
 910         }
 911
 912         if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
 913                 /* We need a per-message support flag, because
 914                  * a. we don't know if the incoming connect reply
 915                  *    supports AT or not (in reply_in_callback)
 916                  *    until we unpack it.
 917                  * b. failovered server means export and flags are gone
 918                  *    (in ptlrpc_send_reply).
 919                  *    Can only be set when we know AT is supported at
 920                  *    both ends */
 921                 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 922         else
 923                 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 924
 925         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 926
 927         return 0;
 928 }
 929
 930 /**
 931  * Add all replay requests back to unreplied list before start replay,
 932  * so that we can make sure the known replied XID is always increased
 933  * only even if when replaying requests.
 934  */
 935 static void ptlrpc_prepare_replay(struct obd_import *imp)
 936 {
 937         struct ptlrpc_request *req;
 938
 939         if (imp->imp_state != LUSTRE_IMP_REPLAY ||
 940             imp->imp_resend_replay)
 941                 return;
 942
 943         /* If the server was restart during repaly, the requests may
 944          * have been added to the unreplied list in former replay. */
 945         spin_lock(&imp->imp_lock);
 946
 947         list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
 948                 if (list_empty(&req->rq_unreplied_list))
 949                         ptlrpc_add_unreplied(req);
 950         }
 951
 952         list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
 953                 if (list_empty(&req->rq_unreplied_list))
 954                         ptlrpc_add_unreplied(req);
 955         }
 956
 957         imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
 958         spin_unlock(&imp->imp_lock);
 959 }
 960
 961 /**
 962  * interpret_reply callback for connect RPCs.
 963  * Looks into returned status of connect operation and decides
 964  * what to do with the import - i.e enter recovery, promote it to
 965  * full state for normal operations of disconnect it due to an error.
 966  */
 967 static int ptlrpc_connect_interpret(const struct lu_env *env,
 968                                     struct ptlrpc_request *request,
 969                                     void *data, int rc)
 970 {
 971         struct ptlrpc_connect_async_args *aa = data;
 972         struct obd_import *imp = request->rq_import;
 973         struct lustre_handle old_hdl;
 974         __u64 old_connect_flags;
 975         int msg_flags;
 976         struct obd_connect_data *ocd;
 977         struct obd_export *exp = NULL;
 978         int ret;
 979         ENTRY;
 980
 981         spin_lock(&imp->imp_lock);
 982         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 983                 imp->imp_connect_tried = 1;
 984                 spin_unlock(&imp->imp_lock);
 985                 RETURN(0);
 986         }
 987
 988         if (rc) {
 989                 /* if this reconnect to busy export - not need select new target
 990                  * for connecting*/
 991                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 992                 spin_unlock(&imp->imp_lock);
 993                 ptlrpc_maybe_ping_import_soon(imp);
 994                 GOTO(out, rc);
 995         }
 996
 997         /* LU-7558: indicate that we are interpretting connect reply,
 998          * pltrpc_connect_import() will not try to reconnect until
 999          * interpret will finish. */
1000         imp->imp_connected = 1;
1001         spin_unlock(&imp->imp_lock);
1002
1003         LASSERT(imp->imp_conn_current);
1004
1005         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1006
1007         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1008                                    RCL_SERVER);
1009         /* server replied obd_connect_data is always bigger */
1010         ocd = req_capsule_server_sized_get(&request->rq_pill,
1011                                            &RMF_CONNECT_DATA, ret);
1012
1013         if (ocd == NULL) {
1014                 CERROR("%s: no connect data from server\n",
1015                        imp->imp_obd->obd_name);
1016                 rc = -EPROTO;
1017                 GOTO(out, rc);
1018         }
1019
1020         spin_lock(&imp->imp_lock);
1021
1022         /* All imports are pingable */
1023         imp->imp_pingable = 1;
1024         imp->imp_force_reconnect = 0;
1025         imp->imp_force_verify = 0;
1026
1027         imp->imp_connect_data = *ocd;
1028
1029         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1030                imp->imp_obd->obd_name, ocd->ocd_instance);
1031         exp = class_conn2export(&imp->imp_dlm_handle);
1032
1033         spin_unlock(&imp->imp_lock);
1034
1035         if (!exp) {
1036                 /* This could happen if export is cleaned during the
1037                    connect attempt */
1038                 CERROR("%s: missing export after connect\n",
1039                        imp->imp_obd->obd_name);
1040                 GOTO(out, rc = -ENODEV);
1041         }
1042
1043         /* check that server granted subset of flags we asked for. */
1044         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1045             ocd->ocd_connect_flags) {
1046                 CERROR("%s: Server didn't grant requested subset of flags: "
1047                        "asked=%#llx granted=%#llx\n",
1048                        imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
1049                        ocd->ocd_connect_flags);
1050                 GOTO(out, rc = -EPROTO);
1051         }
1052
1053         if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) !=
1054             ocd->ocd_connect_flags2) {
1055                 CERROR("%s: Server didn't grant requested subset of flags2: "
1056                        "asked=%#llx granted=%#llx\n",
1057                        imp->imp_obd->obd_name, imp->imp_connect_flags2_orig,
1058                        ocd->ocd_connect_flags2);
1059                 GOTO(out, rc = -EPROTO);
1060         }
1061
1062         if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1063             (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1064             (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1065             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1066                 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1067                 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1068                 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1069
1070                 /* We do not support the MDT-MDT interoperations with
1071                  * different version MDT because of protocol changes. */
1072                 if (unlikely(major != LUSTRE_MAJOR ||
1073                              minor != LUSTRE_MINOR ||
1074                              abs(patch - LUSTRE_PATCH) > 3)) {
1075                         LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1076                                       "connection to different version MDT "
1077                                       "(%d.%d.%d.%d) %s\n",
1078                                       imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1079                                       LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1080                                       major, minor, patch,
1081                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1082                                       imp->imp_connection->c_remote_uuid.uuid);
1083
1084                         GOTO(out, rc = -EPROTO);
1085                 }
1086         }
1087
1088         old_connect_flags = exp_connect_flags(exp);
1089         exp->exp_connect_data = *ocd;
1090         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1091
1092         /* The net statistics after (re-)connect is not valid anymore,
1093          * because may reflect other routing, etc. */
1094         at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1095         ptlrpc_at_adj_net_latency(request,
1096                         lustre_msg_get_service_time(request->rq_repmsg));
1097
1098         /* Import flags should be updated before waking import at FULL state */
1099         rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1100                                       aa->pcaa_initial_connect);
1101         class_export_put(exp);
1102         exp = NULL;
1103
1104         if (rc != 0)
1105                 GOTO(out, rc);
1106
1107         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1108
1109         if (aa->pcaa_initial_connect) {
1110                 spin_lock(&imp->imp_lock);
1111                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1112                         imp->imp_replayable = 1;
1113                         spin_unlock(&imp->imp_lock);
1114                         CDEBUG(D_HA, "connected to replayable target: %s\n",
1115                                obd2cli_tgt(imp->imp_obd));
1116                 } else {
1117                         imp->imp_replayable = 0;
1118                         spin_unlock(&imp->imp_lock);
1119                 }
1120
1121                 /* if applies, adjust the imp->imp_msg_magic here
1122                  * according to reply flags */
1123
1124                 imp->imp_remote_handle =
1125                                 *lustre_msg_get_handle(request->rq_repmsg);
1126
1127                 /* Initial connects are allowed for clients with non-random
1128                  * uuids when servers are in recovery.  Simply signal the
1129                  * servers replay is complete and wait in REPLAY_WAIT. */
1130                 if (msg_flags & MSG_CONNECT_RECOVERING) {
1131                         CDEBUG(D_HA, "connect to %s during recovery\n",
1132                                obd2cli_tgt(imp->imp_obd));
1133                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1134                 } else {
1135                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1136                         ptlrpc_activate_import(imp);
1137                 }
1138
1139                 GOTO(finish, rc = 0);
1140         }
1141
1142         /* Determine what recovery state to move the import to. */
1143         if (MSG_CONNECT_RECONNECT & msg_flags) {
1144                 memset(&old_hdl, 0, sizeof(old_hdl));
1145                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1146                             sizeof (old_hdl))) {
1147                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1148                                       "bad handle %#llx\n",
1149                                       obd2cli_tgt(imp->imp_obd),
1150                                       imp->imp_connection->c_remote_uuid.uuid,
1151                                       imp->imp_dlm_handle.cookie);
1152                         GOTO(out, rc = -ENOTCONN);
1153                 }
1154
1155                 if (memcmp(&imp->imp_remote_handle,
1156                            lustre_msg_get_handle(request->rq_repmsg),
1157                            sizeof(imp->imp_remote_handle))) {
1158                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
1159                                 D_HA : D_WARNING;
1160
1161                         /* Bug 16611/14775: if server handle have changed,
1162                          * that means some sort of disconnection happened.
1163                          * If the server is not in recovery, that also means it
1164                          * already erased all of our state because of previous
1165                          * eviction. If it is in recovery - we are safe to
1166                          * participate since we can reestablish all of our state
1167                          * with server again */
1168                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1169                                 CDEBUG(level,"%s@%s changed server handle from "
1170                                        "%#llx to %#llx"
1171                                        " but is still in recovery\n",
1172                                        obd2cli_tgt(imp->imp_obd),
1173                                        imp->imp_connection->c_remote_uuid.uuid,
1174                                        imp->imp_remote_handle.cookie,
1175                                        lustre_msg_get_handle(
1176                                        request->rq_repmsg)->cookie);
1177                         } else {
1178                                 LCONSOLE_WARN("Evicted from %s (at %s) "
1179                                               "after server handle changed from "
1180                                               "%#llx to %#llx\n",
1181                                               obd2cli_tgt(imp->imp_obd),
1182                                               imp->imp_connection-> \
1183                                               c_remote_uuid.uuid,
1184                                               imp->imp_remote_handle.cookie,
1185                                               lustre_msg_get_handle(
1186                                               request->rq_repmsg)->cookie);
1187                         }
1188
1189
1190                         imp->imp_remote_handle =
1191                                      *lustre_msg_get_handle(request->rq_repmsg);
1192
1193                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1194                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1195                                 GOTO(finish, rc = 0);
1196                         }
1197
1198                 } else {
1199                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1200                                obd2cli_tgt(imp->imp_obd),
1201                                imp->imp_connection->c_remote_uuid.uuid);
1202                 }
1203
1204                 if (imp->imp_invalid) {
1205                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1206                                "marking evicted\n", imp->imp_obd->obd_name);
1207                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1208                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1209                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1210                                imp->imp_obd->obd_name,
1211                                obd2cli_tgt(imp->imp_obd));
1212
1213                         spin_lock(&imp->imp_lock);
1214                         imp->imp_resend_replay = 1;
1215                         spin_unlock(&imp->imp_lock);
1216
1217                         IMPORT_SET_STATE(imp, imp->imp_replay_state);
1218                 } else {
1219                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1220                 }
1221         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1222                 LASSERT(imp->imp_replayable);
1223                 imp->imp_remote_handle =
1224                                 *lustre_msg_get_handle(request->rq_repmsg);
1225                 imp->imp_last_replay_transno = 0;
1226                 imp->imp_replay_cursor = &imp->imp_committed_list;
1227                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
1228         } else {
1229                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
1230                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
1231                 imp->imp_remote_handle =
1232                                 *lustre_msg_get_handle(request->rq_repmsg);
1233                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1234         }
1235
1236         /* Sanity checks for a reconnected import. */
1237         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
1238                 CERROR("imp_replayable flag does not match server "
1239                        "after reconnect. We should LBUG right here.\n");
1240         }
1241
1242         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1243             lustre_msg_get_last_committed(request->rq_repmsg) <
1244             aa->pcaa_peer_committed) {
1245                 CERROR("%s went back in time (transno %lld"
1246                        " was previously committed, server now claims %lld"
1247                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
1248                        "id=9646\n",
1249                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1250                        lustre_msg_get_last_committed(request->rq_repmsg));
1251         }
1252
1253 finish:
1254         ptlrpc_prepare_replay(imp);
1255         rc = ptlrpc_import_recovery_state_machine(imp);
1256         if (rc == -ENOTCONN) {
1257                 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1258                        "invalidating and reconnecting\n",
1259                        obd2cli_tgt(imp->imp_obd),
1260                        imp->imp_connection->c_remote_uuid.uuid);
1261                 ptlrpc_connect_import(imp);
1262                 spin_lock(&imp->imp_lock);
1263                 imp->imp_connected = 0;
1264                 imp->imp_connect_tried = 1;
1265                 spin_unlock(&imp->imp_lock);
1266                 RETURN(0);
1267         }
1268
1269 out:
1270         spin_lock(&imp->imp_lock);
1271         imp->imp_connected = 0;
1272         imp->imp_connect_tried = 1;
1273         spin_unlock(&imp->imp_lock);
1274
1275         if (exp != NULL)
1276                 class_export_put(exp);
1277
1278         if (rc != 0) {
1279                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1280                 if (rc == -EACCES) {
1281                         /*
1282                          * Give up trying to reconnect
1283                          * EACCES means client has no permission for connection
1284                          */
1285                         imp->imp_obd->obd_no_recov = 1;
1286                         ptlrpc_deactivate_import(imp);
1287                 }
1288
1289                 if (rc == -EPROTO) {
1290                         struct obd_connect_data *ocd;
1291
1292                         /* reply message might not be ready */
1293                         if (request->rq_repmsg == NULL)
1294                                 RETURN(-EPROTO);
1295
1296                         ocd = req_capsule_server_get(&request->rq_pill,
1297                                                      &RMF_CONNECT_DATA);
1298                         /* Servers are not supposed to refuse connections from
1299                          * clients based on version, only connection feature
1300                          * flags.  We should never see this from llite, but it
1301                          * may be useful for debugging in the future. */
1302                         if (ocd &&
1303                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1304                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1305                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1306                                         "(%d.%d.%d.%d)"
1307                                         " refused connection from this client "
1308                                         "with an incompatible version (%s).  "
1309                                         "Client must be recompiled\n",
1310                                         obd2cli_tgt(imp->imp_obd),
1311                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1312                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1313                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1314                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1315                                         LUSTRE_VERSION_STRING);
1316                                 ptlrpc_deactivate_import(imp);
1317                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1318                         }
1319                         RETURN(-EPROTO);
1320                 }
1321
1322                 ptlrpc_maybe_ping_import_soon(imp);
1323
1324                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1325                        obd2cli_tgt(imp->imp_obd),
1326                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1327         }
1328
1329         wake_up_all(&imp->imp_recovery_waitq);
1330         RETURN(rc);
1331 }
1332
1333 /**
1334  * interpret callback for "completed replay" RPCs.
1335  * \see signal_completed_replay
1336  */
1337 static int completed_replay_interpret(const struct lu_env *env,
1338                                       struct ptlrpc_request *req,
1339                                       void * data, int rc)
1340 {
1341         ENTRY;
1342         atomic_dec(&req->rq_import->imp_replay_inflight);
1343         if (req->rq_status == 0 &&
1344             !req->rq_import->imp_vbr_failed) {
1345                 ptlrpc_import_recovery_state_machine(req->rq_import);
1346         } else {
1347                 if (req->rq_import->imp_vbr_failed) {
1348                         CDEBUG(D_WARNING,
1349                                "%s: version recovery fails, reconnecting\n",
1350                                req->rq_import->imp_obd->obd_name);
1351                 } else {
1352                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1353                                      "reconnecting\n",
1354                                req->rq_import->imp_obd->obd_name,
1355                                req->rq_status);
1356                 }
1357                 ptlrpc_connect_import(req->rq_import);
1358         }
1359
1360         RETURN(0);
1361 }
1362
1363 /**
1364  * Let server know that we have no requests to replay anymore.
1365  * Achieved by just sending a PING request
1366  */
1367 static int signal_completed_replay(struct obd_import *imp)
1368 {
1369         struct ptlrpc_request *req;
1370         ENTRY;
1371
1372         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1373                 RETURN(0);
1374
1375         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1376         atomic_inc(&imp->imp_replay_inflight);
1377
1378         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1379                                         OBD_PING);
1380         if (req == NULL) {
1381                 atomic_dec(&imp->imp_replay_inflight);
1382                 RETURN(-ENOMEM);
1383         }
1384
1385         ptlrpc_request_set_replen(req);
1386         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1387         lustre_msg_add_flags(req->rq_reqmsg,
1388                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1389         if (AT_OFF)
1390                 req->rq_timeout *= 3;
1391         req->rq_interpret_reply = completed_replay_interpret;
1392
1393         ptlrpcd_add_req(req);
1394         RETURN(0);
1395 }
1396
1397 /**
1398  * In kernel code all import invalidation happens in its own
1399  * separate thread, so that whatever application happened to encounter
1400  * a problem could still be killed or otherwise continue
1401  */
1402 static int ptlrpc_invalidate_import_thread(void *data)
1403 {
1404         struct obd_import *imp = data;
1405
1406         ENTRY;
1407
1408         unshare_fs_struct();
1409
1410         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1411                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1412                imp->imp_connection->c_remote_uuid.uuid);
1413
1414         ptlrpc_invalidate_import(imp);
1415
1416         if (obd_dump_on_eviction) {
1417                 CERROR("dump the log upon eviction\n");
1418                 libcfs_debug_dumplog();
1419         }
1420
1421         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1422         ptlrpc_import_recovery_state_machine(imp);
1423
1424         class_import_put(imp);
1425         RETURN(0);
1426 }
1427
1428 /**
1429  * This is the state machine for client-side recovery on import.
1430  *
1431  * Typicaly we have two possibly paths. If we came to server and it is not
1432  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1433  * state and reconnect from scratch.
1434  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1435  * We go through our list of requests to replay and send them to server one by
1436  * one.
1437  * After sending all request from the list we change import state to
1438  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1439  * and also all the locks we don't yet have and wait for server to grant us.
1440  * After that we send a special "replay completed" request and change import
1441  * state to IMP_REPLAY_WAIT.
1442  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1443  * state and resend all requests from sending list.
1444  * After that we promote import to FULL state and send all delayed requests
1445  * and import is fully operational after that.
1446  *
1447  */
1448 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1449 {
1450         int rc = 0;
1451         int inflight;
1452         char *target_start;
1453         int target_len;
1454
1455         ENTRY;
1456         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1457                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1458                           &target_start, &target_len);
1459                 /* Don't care about MGC eviction */
1460                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1461                            LUSTRE_MGC_NAME) != 0) {
1462                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1463                                            "by %.*s; in progress operations "
1464                                            "using this service will fail.\n",
1465                                            imp->imp_obd->obd_name, target_len,
1466                                            target_start);
1467                 }
1468                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1469                        obd2cli_tgt(imp->imp_obd),
1470                        imp->imp_connection->c_remote_uuid.uuid);
1471                 /* reset vbr_failed flag upon eviction */
1472                 spin_lock(&imp->imp_lock);
1473                 imp->imp_vbr_failed = 0;
1474                 spin_unlock(&imp->imp_lock);
1475
1476                 {
1477                 struct task_struct *task;
1478                 /* bug 17802:  XXX client_disconnect_export vs connect request
1479                  * race. if client is evicted at this time then we start
1480                  * invalidate thread without reference to import and import can
1481                  * be freed at same time. */
1482                 class_import_get(imp);
1483                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1484                                      "ll_imp_inval");
1485                 if (IS_ERR(task)) {
1486                         class_import_put(imp);
1487                         CERROR("error starting invalidate thread: %d\n", rc);
1488                         rc = PTR_ERR(task);
1489                 } else {
1490                         rc = 0;
1491                 }
1492                 RETURN(rc);
1493                 }
1494         }
1495
1496         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1497                 CDEBUG(D_HA, "replay requested by %s\n",
1498                        obd2cli_tgt(imp->imp_obd));
1499                 rc = ptlrpc_replay_next(imp, &inflight);
1500                 if (inflight == 0 &&
1501                     atomic_read(&imp->imp_replay_inflight) == 0) {
1502                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1503                         rc = ldlm_replay_locks(imp);
1504                         if (rc)
1505                                 GOTO(out, rc);
1506                 }
1507                 rc = 0;
1508         }
1509
1510         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1511                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1512                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1513                         rc = signal_completed_replay(imp);
1514                         if (rc)
1515                                 GOTO(out, rc);
1516                 }
1517         }
1518
1519         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1520                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1521                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1522                 }
1523         }
1524
1525         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1526                 struct ptlrpc_connection *conn = imp->imp_connection;
1527
1528                 rc = ptlrpc_resend(imp);
1529                 if (rc)
1530                         GOTO(out, rc);
1531                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1532                 ptlrpc_activate_import(imp);
1533
1534                 LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n",
1535                               imp->imp_obd->obd_name,
1536                               obd_uuid2str(&conn->c_remote_uuid),
1537                               obd_import_nid2str(imp));
1538         }
1539
1540         if (imp->imp_state == LUSTRE_IMP_FULL) {
1541                 wake_up_all(&imp->imp_recovery_waitq);
1542                 ptlrpc_wake_delayed(imp);
1543         }
1544
1545 out:
1546         RETURN(rc);
1547 }
1548
1549 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1550 {
1551         struct ptlrpc_request *req;
1552         int rq_opc, rc = 0;
1553         ENTRY;
1554
1555         if (imp->imp_obd->obd_force)
1556                 GOTO(set_state, rc);
1557
1558         switch (imp->imp_connect_op) {
1559         case OST_CONNECT:
1560                 rq_opc = OST_DISCONNECT;
1561                 break;
1562         case MDS_CONNECT:
1563                 rq_opc = MDS_DISCONNECT;
1564                 break;
1565         case MGS_CONNECT:
1566                 rq_opc = MGS_DISCONNECT;
1567                 break;
1568         default:
1569                 rc = -EINVAL;
1570                 CERROR("%s: don't know how to disconnect from %s "
1571                        "(connect_op %d): rc = %d\n",
1572                        imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1573                        imp->imp_connect_op, rc);
1574                 RETURN(rc);
1575         }
1576
1577         if (ptlrpc_import_in_recovery(imp)) {
1578                 struct l_wait_info lwi;
1579                 long timeout_jiffies;
1580                 time64_t timeout;
1581
1582                 if (AT_OFF) {
1583                         if (imp->imp_server_timeout)
1584                                 timeout = obd_timeout >> 1;
1585                         else
1586                                 timeout = obd_timeout;
1587                 } else {
1588                         u32 req_portal;
1589                         int idx;
1590
1591                         req_portal = imp->imp_client->cli_request_portal;
1592                         idx = import_at_get_index(imp, req_portal);
1593                         timeout = at_get(&imp->imp_at.iat_service_estimate[idx]);
1594                 }
1595
1596                 timeout_jiffies = cfs_time_seconds(timeout);
1597                 lwi = LWI_TIMEOUT_INTR(max_t(long, timeout_jiffies, 1),
1598                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1599                 rc = l_wait_event(imp->imp_recovery_waitq,
1600                                   !ptlrpc_import_in_recovery(imp), &lwi);
1601
1602         }
1603
1604         spin_lock(&imp->imp_lock);
1605         if (imp->imp_state != LUSTRE_IMP_FULL)
1606                 GOTO(out, rc);
1607         spin_unlock(&imp->imp_lock);
1608
1609         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1610                                         LUSTRE_OBD_VERSION, rq_opc);
1611         if (req) {
1612                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1613                  * it fails.  We can get through the above with a down server
1614                  * if the client doesn't know the server is gone yet. */
1615                 req->rq_no_resend = 1;
1616
1617                 /* We want client umounts to happen quickly, no matter the
1618                    server state... */
1619                 req->rq_timeout = min_t(int, req->rq_timeout,
1620                                         INITIAL_CONNECT_TIMEOUT);
1621
1622                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1623                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1624                 ptlrpc_request_set_replen(req);
1625                 rc = ptlrpc_queue_wait(req);
1626                 ptlrpc_req_finished(req);
1627         }
1628
1629 set_state:
1630         spin_lock(&imp->imp_lock);
1631 out:
1632         if (noclose)
1633                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1634         else
1635                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1636         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1637         spin_unlock(&imp->imp_lock);
1638
1639         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1640                 rc = 0;
1641         RETURN(rc);
1642 }
1643 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1644
1645 void ptlrpc_cleanup_imp(struct obd_import *imp)
1646 {
1647         ENTRY;
1648
1649         spin_lock(&imp->imp_lock);
1650         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1651         imp->imp_generation++;
1652         spin_unlock(&imp->imp_lock);
1653         ptlrpc_abort_inflight(imp);
1654
1655         EXIT;
1656 }
1657
1658 /* Adaptive Timeout utils */
1659 extern unsigned int at_min, at_max, at_history;
1660
1661 /* Update at_current with the specified value (bounded by at_min and at_max),
1662  * as well as the AT history "bins".
1663  *  - Bin into timeslices using AT_BINS bins.
1664  *  - This gives us a max of the last at_history seconds without the storage,
1665  *    but still smoothing out a return to normalcy from a slow response.
1666  *  - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1667  */
1668 int at_measured(struct adaptive_timeout *at, unsigned int val)
1669 {
1670         unsigned int old = at->at_current;
1671         time64_t now = ktime_get_real_seconds();
1672         long binlimit = max_t(long, at_history / AT_BINS, 1);
1673
1674         LASSERT(at);
1675         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1676                val, at, (long)(now - at->at_binstart), at->at_current,
1677                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1678
1679         if (val == 0)
1680                 /* 0's don't count, because we never want our timeout to
1681                    drop to 0, and because 0 could mean an error */
1682                 return 0;
1683
1684         spin_lock(&at->at_lock);
1685
1686         if (unlikely(at->at_binstart == 0)) {
1687                 /* Special case to remove default from history */
1688                 at->at_current = val;
1689                 at->at_worst_ever = val;
1690                 at->at_worst_time = now;
1691                 at->at_hist[0] = val;
1692                 at->at_binstart = now;
1693         } else if (now - at->at_binstart < binlimit ) {
1694                 /* in bin 0 */
1695                 at->at_hist[0] = max(val, at->at_hist[0]);
1696                 at->at_current = max(val, at->at_current);
1697         } else {
1698                 int i, shift;
1699                 unsigned int maxv = val;
1700
1701                 /* move bins over */
1702                 shift = (u32)(now - at->at_binstart) / binlimit;
1703                 LASSERT(shift > 0);
1704                 for(i = AT_BINS - 1; i >= 0; i--) {
1705                         if (i >= shift) {
1706                                 at->at_hist[i] = at->at_hist[i - shift];
1707                                 maxv = max(maxv, at->at_hist[i]);
1708                         } else {
1709                                 at->at_hist[i] = 0;
1710                         }
1711                 }
1712                 at->at_hist[0] = val;
1713                 at->at_current = maxv;
1714                 at->at_binstart += shift * binlimit;
1715         }
1716
1717         if (at->at_current > at->at_worst_ever) {
1718                 at->at_worst_ever = at->at_current;
1719                 at->at_worst_time = now;
1720         }
1721
1722         if (at->at_flags & AT_FLG_NOHIST)
1723                 /* Only keep last reported val; keeping the rest of the history
1724                    for proc only */
1725                 at->at_current = val;
1726
1727         if (at_max > 0)
1728                 at->at_current =  min(at->at_current, at_max);
1729         at->at_current =  max(at->at_current, at_min);
1730
1731         if (at->at_current != old)
1732                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1733                        "(val=%u) hist %u %u %u %u\n", at,
1734                        old, at->at_current, at->at_current - old, val,
1735                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1736                        at->at_hist[3]);
1737
1738         /* if we changed, report the old value */
1739         old = (at->at_current != old) ? old : 0;
1740
1741         spin_unlock(&at->at_lock);
1742         return old;
1743 }
1744
1745 /* Find the imp_at index for a given portal; assign if space available */
1746 int import_at_get_index(struct obd_import *imp, int portal)
1747 {
1748         struct imp_at *at = &imp->imp_at;
1749         int i;
1750
1751         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1752                 if (at->iat_portal[i] == portal)
1753                         return i;
1754                 if (at->iat_portal[i] == 0)
1755                         /* unused */
1756                         break;
1757         }
1758
1759         /* Not found in list, add it under a lock */
1760         spin_lock(&imp->imp_lock);
1761
1762         /* Check unused under lock */
1763         for (; i < IMP_AT_MAX_PORTALS; i++) {
1764                 if (at->iat_portal[i] == portal)
1765                         goto out;
1766                 if (at->iat_portal[i] == 0)
1767                         /* unused */
1768                         break;
1769         }
1770
1771         /* Not enough portals? */
1772         LASSERT(i < IMP_AT_MAX_PORTALS);
1773
1774         at->iat_portal[i] = portal;
1775 out:
1776         spin_unlock(&imp->imp_lock);
1777         return i;
1778 }