drivers/staging/lustre/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, 2012, 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         imp->imp_state = state;
  67         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
  68         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
  69                 cfs_time_current_sec();
  70         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
  71                 IMP_STATE_HIST_LEN;
  72 }
  73
  74 /* A CLOSED import should remain so. */
  75 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                 \
  76 do {                                                                       \
  77         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                           \
  78                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  79                       imp, obd2cli_tgt(imp->imp_obd),                     \
  80                       ptlrpc_import_state_name(imp->imp_state),         \
  81                       ptlrpc_import_state_name(state));                 \
  82                __import_set_state(imp, state);                           \
  83         }                                                                     \
  84 } while(0)
  85
  86 #define IMPORT_SET_STATE(imp, state)                                    \
  87 do {                                                                    \
  88         spin_lock(&imp->imp_lock);                                      \
  89         IMPORT_SET_STATE_NOLOCK(imp, state);                            \
  90         spin_unlock(&imp->imp_lock);                                    \
  91 } while(0)
  92
  93
  94 static int ptlrpc_connect_interpret(const struct lu_env *env,
  95                                     struct ptlrpc_request *request,
  96                                     void * data, int rc);
  97 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
  98
  99 /* Only this function is allowed to change the import state when it is
 100  * CLOSED. I would rather refcount the import and free it after
 101  * disconnection like we do with exports. To do that, the client_obd
 102  * will need to save the peer info somewhere other than in the import,
 103  * though. */
 104 int ptlrpc_init_import(struct obd_import *imp)
 105 {
 106         spin_lock(&imp->imp_lock);
 107
 108         imp->imp_generation++;
 109         imp->imp_state =  LUSTRE_IMP_NEW;
 110
 111         spin_unlock(&imp->imp_lock);
 112
 113         return 0;
 114 }
 115 EXPORT_SYMBOL(ptlrpc_init_import);
 116
 117 #define UUID_STR "_UUID"
 118 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
 119 {
 120         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 121                 ? uuid : uuid + strlen(prefix);
 122
 123         *uuid_len = strlen(*uuid_start);
 124
 125         if (*uuid_len < strlen(UUID_STR))
 126                 return;
 127
 128         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 129                     UUID_STR, strlen(UUID_STR)))
 130                 *uuid_len -= strlen(UUID_STR);
 131 }
 132 EXPORT_SYMBOL(deuuidify);
 133
 134 /**
 135  * Returns true if import was FULL, false if import was already not
 136  * connected.
 137  * @imp - import to be disconnected
 138  * @conn_cnt - connection count (epoch) of the request that timed out
 139  *           and caused the disconnection.  In some cases, multiple
 140  *           inflight requests can fail to a single target (e.g. OST
 141  *           bulk requests) and if one has already caused a reconnection
 142  *           (increasing the import->conn_cnt) the older failure should
 143  *           not also cause a reconnection.  If zero it forces a reconnect.
 144  */
 145 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 146 {
 147         int rc = 0;
 148
 149         spin_lock(&imp->imp_lock);
 150
 151         if (imp->imp_state == LUSTRE_IMP_FULL &&
 152             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 153                 char *target_start;
 154                 int   target_len;
 155
 156                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 157                           &target_start, &target_len);
 158
 159                 if (imp->imp_replayable) {
 160                         LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
 161                                "lost; in progress operations using this "
 162                                "service will wait for recovery to complete\n",
 163                                imp->imp_obd->obd_name, target_len, target_start,
 164                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 165                 } else {
 166                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
 167                                "%.*s (at %s) was lost; in progress "
 168                                "operations using this service will fail\n",
 169                                imp->imp_obd->obd_name,
 170                                target_len, target_start,
 171                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 172                 }
 173                 ptlrpc_deactivate_timeouts(imp);
 174                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 175                 spin_unlock(&imp->imp_lock);
 176
 177                 if (obd_dump_on_timeout)
 178                         libcfs_debug_dumplog();
 179
 180                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 181                 rc = 1;
 182         } else {
 183                 spin_unlock(&imp->imp_lock);
 184                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 185                        imp->imp_client->cli_name, imp,
 186                        (imp->imp_state == LUSTRE_IMP_FULL &&
 187                         imp->imp_conn_cnt > conn_cnt) ?
 188                        "reconnected" : "not connected", imp->imp_conn_cnt,
 189                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 190         }
 191
 192         return rc;
 193 }
 194
 195 /* Must be called with imp_lock held! */
 196 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 197 {
 198         LASSERT(spin_is_locked(&imp->imp_lock));
 199
 200         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 201         imp->imp_invalid = 1;
 202         imp->imp_generation++;
 203         spin_unlock(&imp->imp_lock);
 204
 205         ptlrpc_abort_inflight(imp);
 206         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 207 }
 208
 209 /*
 210  * This acts as a barrier; all existing requests are rejected, and
 211  * no new requests will be accepted until the import is valid again.
 212  */
 213 void ptlrpc_deactivate_import(struct obd_import *imp)
 214 {
 215         spin_lock(&imp->imp_lock);
 216         ptlrpc_deactivate_and_unlock_import(imp);
 217 }
 218 EXPORT_SYMBOL(ptlrpc_deactivate_import);
 219
 220 static unsigned int
 221 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
 222 {
 223         long dl;
 224
 225         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
 226               (req->rq_phase == RQ_PHASE_BULK) ||
 227               (req->rq_phase == RQ_PHASE_NEW)))
 228                 return 0;
 229
 230         if (req->rq_timedout)
 231                 return 0;
 232
 233         if (req->rq_phase == RQ_PHASE_NEW)
 234                 dl = req->rq_sent;
 235         else
 236                 dl = req->rq_deadline;
 237
 238         if (dl <= now)
 239                 return 0;
 240
 241         return dl - now;
 242 }
 243
 244 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
 245 {
 246         time_t now = cfs_time_current_sec();
 247         struct list_head *tmp, *n;
 248         struct ptlrpc_request *req;
 249         unsigned int timeout = 0;
 250
 251         spin_lock(&imp->imp_lock);
 252         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 253                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 254                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 255         }
 256         spin_unlock(&imp->imp_lock);
 257         return timeout;
 258 }
 259
 260 /**
 261  * This function will invalidate the import, if necessary, then block
 262  * for all the RPC completions, and finally notify the obd to
 263  * invalidate its state (ie cancel locks, clear pending requests,
 264  * etc).
 265  */
 266 void ptlrpc_invalidate_import(struct obd_import *imp)
 267 {
 268         struct list_head *tmp, *n;
 269         struct ptlrpc_request *req;
 270         struct l_wait_info lwi;
 271         unsigned int timeout;
 272         int rc;
 273
 274         atomic_inc(&imp->imp_inval_count);
 275
 276         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 277                 ptlrpc_deactivate_import(imp);
 278
 279         LASSERT(imp->imp_invalid);
 280
 281         /* Wait forever until inflight == 0. We really can't do it another
 282          * way because in some cases we need to wait for very long reply
 283          * unlink. We can't do anything before that because there is really
 284          * no guarantee that some rdma transfer is not in progress right now. */
 285         do {
 286                 /* Calculate max timeout for waiting on rpcs to error
 287                  * out. Use obd_timeout if calculated value is smaller
 288                  * than it. */
 289                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 290                         timeout = ptlrpc_inflight_timeout(imp);
 291                         timeout += timeout / 3;
 292
 293                         if (timeout == 0)
 294                                 timeout = obd_timeout;
 295                 } else {
 296                         /* decrease the interval to increase race condition */
 297                         timeout = 1;
 298                 }
 299
 300                 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
 301                        timeout);
 302
 303                 /* Wait for all requests to error out and call completion
 304                  * callbacks. Cap it at obd_timeout -- these should all
 305                  * have been locally cancelled by ptlrpc_abort_inflight. */
 306                 lwi = LWI_TIMEOUT_INTERVAL(
 307                         cfs_timeout_cap(cfs_time_seconds(timeout)),
 308                         (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
 309                         NULL, NULL);
 310                 rc = l_wait_event(imp->imp_recovery_waitq,
 311                                   (atomic_read(&imp->imp_inflight) == 0),
 312                                   &lwi);
 313                 if (rc) {
 314                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 315
 316                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 317                                cli_tgt, rc,
 318                                atomic_read(&imp->imp_inflight));
 319
 320                         spin_lock(&imp->imp_lock);
 321                         if (atomic_read(&imp->imp_inflight) == 0) {
 322                                 int count = atomic_read(&imp->imp_unregistering);
 323
 324                                 /* We know that "unregistering" rpcs only can
 325                                  * survive in sending or delaying lists (they
 326                                  * maybe waiting for long reply unlink in
 327                                  * sluggish nets). Let's check this. If there
 328                                  * is no inflight and unregistering != 0, this
 329                                  * is bug. */
 330                                 LASSERTF(count == 0, "Some RPCs are still "
 331                                          "unregistering: %d\n", count);
 332
 333                                 /* Let's save one loop as soon as inflight have
 334                                  * dropped to zero. No new inflights possible at
 335                                  * this point. */
 336                                 rc = 0;
 337                         } else {
 338                                 list_for_each_safe(tmp, n,
 339                                                        &imp->imp_sending_list) {
 340                                         req = list_entry(tmp,
 341                                                              struct ptlrpc_request,
 342                                                              rq_list);
 343                                         DEBUG_REQ(D_ERROR, req,
 344                                                   "still on sending list");
 345                                 }
 346                                 list_for_each_safe(tmp, n,
 347                                                        &imp->imp_delayed_list) {
 348                                         req = list_entry(tmp,
 349                                                              struct ptlrpc_request,
 350                                                              rq_list);
 351                                         DEBUG_REQ(D_ERROR, req,
 352                                                   "still on delayed list");
 353                                 }
 354
 355                                 CERROR("%s: RPCs in \"%s\" phase found (%d). "
 356                                        "Network is sluggish? Waiting them "
 357                                        "to error out.\n", cli_tgt,
 358                                        ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
 359                                        atomic_read(&imp->
 360                                                        imp_unregistering));
 361                         }
 362                         spin_unlock(&imp->imp_lock);
 363                   }
 364         } while (rc != 0);
 365
 366         /*
 367          * Let's additionally check that no new rpcs added to import in
 368          * "invalidate" state.
 369          */
 370         LASSERT(atomic_read(&imp->imp_inflight) == 0);
 371         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 372         sptlrpc_import_flush_all_ctx(imp);
 373
 374         atomic_dec(&imp->imp_inval_count);
 375         wake_up_all(&imp->imp_recovery_waitq);
 376 }
 377 EXPORT_SYMBOL(ptlrpc_invalidate_import);
 378
 379 /* unset imp_invalid */
 380 void ptlrpc_activate_import(struct obd_import *imp)
 381 {
 382         struct obd_device *obd = imp->imp_obd;
 383
 384         spin_lock(&imp->imp_lock);
 385         imp->imp_invalid = 0;
 386         ptlrpc_activate_timeouts(imp);
 387         spin_unlock(&imp->imp_lock);
 388         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 389 }
 390 EXPORT_SYMBOL(ptlrpc_activate_import);
 391
 392 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 393 {
 394         LASSERT(!imp->imp_dlm_fake);
 395
 396         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 397                 if (!imp->imp_replayable) {
 398                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 399                                "auto-deactivating\n",
 400                                obd2cli_tgt(imp->imp_obd),
 401                                imp->imp_connection->c_remote_uuid.uuid,
 402                                imp->imp_obd->obd_name);
 403                         ptlrpc_deactivate_import(imp);
 404                 }
 405
 406                 CDEBUG(D_HA, "%s: waking up pinger\n",
 407                        obd2cli_tgt(imp->imp_obd));
 408
 409                 spin_lock(&imp->imp_lock);
 410                 imp->imp_force_verify = 1;
 411                 spin_unlock(&imp->imp_lock);
 412
 413                 ptlrpc_pinger_wake_up();
 414         }
 415 }
 416 EXPORT_SYMBOL(ptlrpc_fail_import);
 417
 418 int ptlrpc_reconnect_import(struct obd_import *imp)
 419 {
 420         ptlrpc_set_import_discon(imp, 0);
 421         /* Force a new connect attempt */
 422         ptlrpc_invalidate_import(imp);
 423         /* Do a fresh connect next time by zeroing the handle */
 424         ptlrpc_disconnect_import(imp, 1);
 425         /* Wait for all invalidate calls to finish */
 426         if (atomic_read(&imp->imp_inval_count) > 0) {
 427                 int rc;
 428                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 429                 rc = l_wait_event(imp->imp_recovery_waitq,
 430                                   (atomic_read(&imp->imp_inval_count) == 0),
 431                                   &lwi);
 432                 if (rc)
 433                         CERROR("Interrupted, inval=%d\n",
 434                                atomic_read(&imp->imp_inval_count));
 435         }
 436
 437         /* Allow reconnect attempts */
 438         imp->imp_obd->obd_no_recov = 0;
 439         /* Remove 'invalid' flag */
 440         ptlrpc_activate_import(imp);
 441         /* Attempt a new connect */
 442         ptlrpc_recover_import(imp, NULL, 0);
 443         return 0;
 444 }
 445 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 446
 447 /**
 448  * Connection on import \a imp is changed to another one (if more than one is
 449  * present). We typically chose connection that we have not tried to connect to
 450  * the longest
 451  */
 452 static int import_select_connection(struct obd_import *imp)
 453 {
 454         struct obd_import_conn *imp_conn = NULL, *conn;
 455         struct obd_export *dlmexp;
 456         char *target_start;
 457         int target_len, tried_all = 1;
 458
 459         spin_lock(&imp->imp_lock);
 460
 461         if (list_empty(&imp->imp_conn_list)) {
 462                 CERROR("%s: no connections available\n",
 463                        imp->imp_obd->obd_name);
 464                 spin_unlock(&imp->imp_lock);
 465                 return -EINVAL;
 466         }
 467
 468         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 469                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 470                        imp->imp_obd->obd_name,
 471                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 472                        conn->oic_last_attempt);
 473
 474                 /* If we have not tried this connection since
 475                    the last successful attempt, go with this one */
 476                 if ((conn->oic_last_attempt == 0) ||
 477                     cfs_time_beforeq_64(conn->oic_last_attempt,
 478                                        imp->imp_last_success_conn)) {
 479                         imp_conn = conn;
 480                         tried_all = 0;
 481                         break;
 482                 }
 483
 484                 /* If all of the connections have already been tried
 485                    since the last successful connection; just choose the
 486                    least recently used */
 487                 if (!imp_conn)
 488                         imp_conn = conn;
 489                 else if (cfs_time_before_64(conn->oic_last_attempt,
 490                                             imp_conn->oic_last_attempt))
 491                         imp_conn = conn;
 492         }
 493
 494         /* if not found, simply choose the current one */
 495         if (!imp_conn || imp->imp_force_reconnect) {
 496                 LASSERT(imp->imp_conn_current);
 497                 imp_conn = imp->imp_conn_current;
 498                 tried_all = 0;
 499         }
 500         LASSERT(imp_conn->oic_conn);
 501
 502         /* If we've tried everything, and we're back to the beginning of the
 503            list, increase our timeout and try again. It will be reset when
 504            we do finally connect. (FIXME: really we should wait for all network
 505            state associated with the last connection attempt to drain before
 506            trying to reconnect on it.) */
 507         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 508                 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
 509                 if (at_get(at) < CONNECTION_SWITCH_MAX) {
 510                         at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
 511                         if (at_get(at) > CONNECTION_SWITCH_MAX)
 512                                 at_reset(at, CONNECTION_SWITCH_MAX);
 513                 }
 514                 LASSERT(imp_conn->oic_last_attempt);
 515                 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
 516                         "to %ds\n", imp->imp_obd->obd_name, at_get(at));
 517         }
 518
 519         imp_conn->oic_last_attempt = cfs_time_current_64();
 520
 521         /* switch connection, don't mind if it's same as the current one */
 522         if (imp->imp_connection)
 523                 ptlrpc_connection_put(imp->imp_connection);
 524         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 525
 526         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 527         LASSERT(dlmexp != NULL);
 528         if (dlmexp->exp_connection)
 529                 ptlrpc_connection_put(dlmexp->exp_connection);
 530         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 531         class_export_put(dlmexp);
 532
 533         if (imp->imp_conn_current != imp_conn) {
 534                 if (imp->imp_conn_current) {
 535                         deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 536                                   &target_start, &target_len);
 537
 538                         CDEBUG(D_HA, "%s: Connection changing to"
 539                                " %.*s (at %s)\n",
 540                                imp->imp_obd->obd_name,
 541                                target_len, target_start,
 542                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 543                 }
 544
 545                 imp->imp_conn_current = imp_conn;
 546         }
 547
 548         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 549                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 550                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 551
 552         spin_unlock(&imp->imp_lock);
 553
 554         return 0;
 555 }
 556
 557 /*
 558  * must be called under imp_lock
 559  */
 560 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 561 {
 562         struct ptlrpc_request *req;
 563         struct list_head *tmp;
 564
 565         if (list_empty(&imp->imp_replay_list))
 566                 return 0;
 567         tmp = imp->imp_replay_list.next;
 568         req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 569         *transno = req->rq_transno;
 570         if (req->rq_transno == 0) {
 571                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 572                 LBUG();
 573         }
 574
 575         return 1;
 576 }
 577
 578 /**
 579  * Attempt to (re)connect import \a imp. This includes all preparations,
 580  * initializing CONNECT RPC request and passing it to ptlrpcd for
 581  * actual sending.
 582  * Returns 0 on success or error code.
 583  */
 584 int ptlrpc_connect_import(struct obd_import *imp)
 585 {
 586         struct obd_device *obd = imp->imp_obd;
 587         int initial_connect = 0;
 588         int set_transno = 0;
 589         __u64 committed_before_reconnect = 0;
 590         struct ptlrpc_request *request;
 591         char *bufs[] = { NULL,
 592                          obd2cli_tgt(imp->imp_obd),
 593                          obd->obd_uuid.uuid,
 594                          (char *)&imp->imp_dlm_handle,
 595                          (char *)&imp->imp_connect_data };
 596         struct ptlrpc_connect_async_args *aa;
 597         int rc;
 598
 599         spin_lock(&imp->imp_lock);
 600         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 601                 spin_unlock(&imp->imp_lock);
 602                 CERROR("can't connect to a closed import\n");
 603                 return -EINVAL;
 604         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 605                 spin_unlock(&imp->imp_lock);
 606                 CERROR("already connected\n");
 607                 return 0;
 608         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 609                 spin_unlock(&imp->imp_lock);
 610                 CERROR("already connecting\n");
 611                 return -EALREADY;
 612         }
 613
 614         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 615
 616         imp->imp_conn_cnt++;
 617         imp->imp_resend_replay = 0;
 618
 619         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 620                 initial_connect = 1;
 621         else
 622                 committed_before_reconnect = imp->imp_peer_committed_transno;
 623
 624         set_transno = ptlrpc_first_transno(imp,
 625                                            &imp->imp_connect_data.ocd_transno);
 626         spin_unlock(&imp->imp_lock);
 627
 628         rc = import_select_connection(imp);
 629         if (rc)
 630                 GOTO(out, rc);
 631
 632         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 633         if (rc)
 634                 GOTO(out, rc);
 635
 636         /* Reset connect flags to the originally requested flags, in case
 637          * the server is updated on-the-fly we will get the new features. */
 638         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 639         /* Reset ocd_version each time so the server knows the exact versions */
 640         imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
 641         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 642         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 643
 644         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 645                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 646         if (rc)
 647                 GOTO(out, rc);
 648
 649         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 650         if (request == NULL)
 651                 GOTO(out, rc = -ENOMEM);
 652
 653         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 654                                       imp->imp_connect_op, bufs, NULL);
 655         if (rc) {
 656                 ptlrpc_request_free(request);
 657                 GOTO(out, rc);
 658         }
 659
 660         /* Report the rpc service time to the server so that it knows how long
 661          * to wait for clients to join recovery */
 662         lustre_msg_set_service_time(request->rq_reqmsg,
 663                                     at_timeout2est(request->rq_timeout));
 664
 665         /* The amount of time we give the server to process the connect req.
 666          * import_select_connection will increase the net latency on
 667          * repeated reconnect attempts to cover slow networks.
 668          * We override/ignore the server rpc completion estimate here,
 669          * which may be large if this is a reconnect attempt */
 670         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 671         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 672
 673         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 674
 675         request->rq_no_resend = request->rq_no_delay = 1;
 676         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 677         /* Allow a slightly larger reply for future growth compatibility */
 678         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 679                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 680         ptlrpc_request_set_replen(request);
 681         request->rq_interpret_reply = ptlrpc_connect_interpret;
 682
 683         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 684         aa = ptlrpc_req_async_args(request);
 685         memset(aa, 0, sizeof *aa);
 686
 687         aa->pcaa_peer_committed = committed_before_reconnect;
 688         aa->pcaa_initial_connect = initial_connect;
 689
 690         if (aa->pcaa_initial_connect) {
 691                 spin_lock(&imp->imp_lock);
 692                 imp->imp_replayable = 1;
 693                 spin_unlock(&imp->imp_lock);
 694                 lustre_msg_add_op_flags(request->rq_reqmsg,
 695                                         MSG_CONNECT_INITIAL);
 696         }
 697
 698         if (set_transno)
 699                 lustre_msg_add_op_flags(request->rq_reqmsg,
 700                                         MSG_CONNECT_TRANSNO);
 701
 702         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
 703                   request->rq_timeout);
 704         ptlrpcd_add_req(request, PDL_POLICY_ROUND, -1);
 705         rc = 0;
 706 out:
 707         if (rc != 0) {
 708                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 709         }
 710
 711         return rc;
 712 }
 713 EXPORT_SYMBOL(ptlrpc_connect_import);
 714
 715 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 716 {
 717         int force_verify;
 718
 719         spin_lock(&imp->imp_lock);
 720         force_verify = imp->imp_force_verify != 0;
 721         spin_unlock(&imp->imp_lock);
 722
 723         if (force_verify)
 724                 ptlrpc_pinger_wake_up();
 725 }
 726
 727 static int ptlrpc_busy_reconnect(int rc)
 728 {
 729         return (rc == -EBUSY) || (rc == -EAGAIN);
 730 }
 731
 732 /**
 733  * interpret_reply callback for connect RPCs.
 734  * Looks into returned status of connect operation and decides
 735  * what to do with the import - i.e enter recovery, promote it to
 736  * full state for normal operations of disconnect it due to an error.
 737  */
 738 static int ptlrpc_connect_interpret(const struct lu_env *env,
 739                                     struct ptlrpc_request *request,
 740                                     void *data, int rc)
 741 {
 742         struct ptlrpc_connect_async_args *aa = data;
 743         struct obd_import *imp = request->rq_import;
 744         struct client_obd *cli = &imp->imp_obd->u.cli;
 745         struct lustre_handle old_hdl;
 746         __u64 old_connect_flags;
 747         int msg_flags;
 748         struct obd_connect_data *ocd;
 749         struct obd_export *exp;
 750         int ret;
 751
 752         spin_lock(&imp->imp_lock);
 753         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 754                 imp->imp_connect_tried = 1;
 755                 spin_unlock(&imp->imp_lock);
 756                 return 0;
 757         }
 758
 759         if (rc) {
 760                 /* if this reconnect to busy export - not need select new target
 761                  * for connecting*/
 762                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 763                 spin_unlock(&imp->imp_lock);
 764                 ptlrpc_maybe_ping_import_soon(imp);
 765                 GOTO(out, rc);
 766         }
 767         spin_unlock(&imp->imp_lock);
 768
 769         LASSERT(imp->imp_conn_current);
 770
 771         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 772
 773         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 774                                    RCL_SERVER);
 775         /* server replied obd_connect_data is always bigger */
 776         ocd = req_capsule_server_sized_get(&request->rq_pill,
 777                                            &RMF_CONNECT_DATA, ret);
 778
 779         if (ocd == NULL) {
 780                 CERROR("%s: no connect data from server\n",
 781                        imp->imp_obd->obd_name);
 782                 rc = -EPROTO;
 783                 GOTO(out, rc);
 784         }
 785
 786         spin_lock(&imp->imp_lock);
 787
 788         /* All imports are pingable */
 789         imp->imp_pingable = 1;
 790         imp->imp_force_reconnect = 0;
 791         imp->imp_force_verify = 0;
 792
 793         imp->imp_connect_data = *ocd;
 794
 795         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
 796                imp->imp_obd->obd_name, ocd->ocd_instance);
 797         exp = class_conn2export(&imp->imp_dlm_handle);
 798
 799         spin_unlock(&imp->imp_lock);
 800
 801         /* check that server granted subset of flags we asked for. */
 802         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
 803             ocd->ocd_connect_flags) {
 804                 CERROR("%s: Server didn't granted asked subset of flags: "
 805                        "asked="LPX64" grranted="LPX64"\n",
 806                        imp->imp_obd->obd_name,imp->imp_connect_flags_orig,
 807                        ocd->ocd_connect_flags);
 808                 GOTO(out, rc = -EPROTO);
 809         }
 810
 811         if (!exp) {
 812                 /* This could happen if export is cleaned during the
 813                    connect attempt */
 814                 CERROR("%s: missing export after connect\n",
 815                        imp->imp_obd->obd_name);
 816                 GOTO(out, rc = -ENODEV);
 817         }
 818         old_connect_flags = exp_connect_flags(exp);
 819         exp->exp_connect_data = *ocd;
 820         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
 821         class_export_put(exp);
 822
 823         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 824
 825         if (aa->pcaa_initial_connect) {
 826                 spin_lock(&imp->imp_lock);
 827                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 828                         imp->imp_replayable = 1;
 829                         spin_unlock(&imp->imp_lock);
 830                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 831                                obd2cli_tgt(imp->imp_obd));
 832                 } else {
 833                         imp->imp_replayable = 0;
 834                         spin_unlock(&imp->imp_lock);
 835                 }
 836
 837                 /* if applies, adjust the imp->imp_msg_magic here
 838                  * according to reply flags */
 839
 840                 imp->imp_remote_handle =
 841                                 *lustre_msg_get_handle(request->rq_repmsg);
 842
 843                 /* Initial connects are allowed for clients with non-random
 844                  * uuids when servers are in recovery.  Simply signal the
 845                  * servers replay is complete and wait in REPLAY_WAIT. */
 846                 if (msg_flags & MSG_CONNECT_RECOVERING) {
 847                         CDEBUG(D_HA, "connect to %s during recovery\n",
 848                                obd2cli_tgt(imp->imp_obd));
 849                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 850                 } else {
 851                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 852                         ptlrpc_activate_import(imp);
 853                 }
 854
 855                 GOTO(finish, rc = 0);
 856         }
 857
 858         /* Determine what recovery state to move the import to. */
 859         if (MSG_CONNECT_RECONNECT & msg_flags) {
 860                 memset(&old_hdl, 0, sizeof(old_hdl));
 861                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 862                             sizeof (old_hdl))) {
 863                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
 864                                       "bad handle "LPX64"\n",
 865                                       obd2cli_tgt(imp->imp_obd),
 866                                       imp->imp_connection->c_remote_uuid.uuid,
 867                                       imp->imp_dlm_handle.cookie);
 868                         GOTO(out, rc = -ENOTCONN);
 869                 }
 870
 871                 if (memcmp(&imp->imp_remote_handle,
 872                            lustre_msg_get_handle(request->rq_repmsg),
 873                            sizeof(imp->imp_remote_handle))) {
 874                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
 875                                 D_HA : D_WARNING;
 876
 877                         /* Bug 16611/14775: if server handle have changed,
 878                          * that means some sort of disconnection happened.
 879                          * If the server is not in recovery, that also means it
 880                          * already erased all of our state because of previous
 881                          * eviction. If it is in recovery - we are safe to
 882                          * participate since we can reestablish all of our state
 883                          * with server again */
 884                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
 885                                 CDEBUG(level,"%s@%s changed server handle from "
 886                                        LPX64" to "LPX64
 887                                        " but is still in recovery\n",
 888                                        obd2cli_tgt(imp->imp_obd),
 889                                        imp->imp_connection->c_remote_uuid.uuid,
 890                                        imp->imp_remote_handle.cookie,
 891                                        lustre_msg_get_handle(
 892                                        request->rq_repmsg)->cookie);
 893                         } else {
 894                                 LCONSOLE_WARN("Evicted from %s (at %s) "
 895                                               "after server handle changed from "
 896                                               LPX64" to "LPX64"\n",
 897                                               obd2cli_tgt(imp->imp_obd),
 898                                               imp->imp_connection-> \
 899                                               c_remote_uuid.uuid,
 900                                               imp->imp_remote_handle.cookie,
 901                                               lustre_msg_get_handle(
 902                                               request->rq_repmsg)->cookie);
 903                         }
 904
 905
 906                         imp->imp_remote_handle =
 907                                      *lustre_msg_get_handle(request->rq_repmsg);
 908
 909                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
 910                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 911                                 GOTO(finish, rc = 0);
 912                         }
 913
 914                 } else {
 915                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 916                                obd2cli_tgt(imp->imp_obd),
 917                                imp->imp_connection->c_remote_uuid.uuid);
 918                 }
 919
 920                 if (imp->imp_invalid) {
 921                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
 922                                "marking evicted\n", imp->imp_obd->obd_name);
 923                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 924                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 925                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 926                                imp->imp_obd->obd_name,
 927                                obd2cli_tgt(imp->imp_obd));
 928
 929                         spin_lock(&imp->imp_lock);
 930                         imp->imp_resend_replay = 1;
 931                         spin_unlock(&imp->imp_lock);
 932
 933                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 934                 } else {
 935                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 936                 }
 937         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 938                 LASSERT(imp->imp_replayable);
 939                 imp->imp_remote_handle =
 940                                 *lustre_msg_get_handle(request->rq_repmsg);
 941                 imp->imp_last_replay_transno = 0;
 942                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 943         } else {
 944                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
 945                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
 946                 imp->imp_remote_handle =
 947                                 *lustre_msg_get_handle(request->rq_repmsg);
 948                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 949         }
 950
 951         /* Sanity checks for a reconnected import. */
 952         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 953                 CERROR("imp_replayable flag does not match server "
 954                        "after reconnect. We should LBUG right here.\n");
 955         }
 956
 957         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
 958             lustre_msg_get_last_committed(request->rq_repmsg) <
 959             aa->pcaa_peer_committed) {
 960                 CERROR("%s went back in time (transno "LPD64
 961                        " was previously committed, server now claims "LPD64
 962                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
 963                        "id=9646\n",
 964                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 965                        lustre_msg_get_last_committed(request->rq_repmsg));
 966         }
 967
 968 finish:
 969         rc = ptlrpc_import_recovery_state_machine(imp);
 970         if (rc != 0) {
 971                 if (rc == -ENOTCONN) {
 972                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 973                                "invalidating and reconnecting\n",
 974                                obd2cli_tgt(imp->imp_obd),
 975                                imp->imp_connection->c_remote_uuid.uuid);
 976                         ptlrpc_connect_import(imp);
 977                         imp->imp_connect_tried = 1;
 978                         return 0;
 979                 }
 980         } else {
 981
 982                 spin_lock(&imp->imp_lock);
 983                 list_del(&imp->imp_conn_current->oic_item);
 984                 list_add(&imp->imp_conn_current->oic_item,
 985                              &imp->imp_conn_list);
 986                 imp->imp_last_success_conn =
 987                         imp->imp_conn_current->oic_last_attempt;
 988
 989                 spin_unlock(&imp->imp_lock);
 990
 991                 if (!ocd->ocd_ibits_known &&
 992                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 993                         CERROR("Inodebits aware server returned zero compatible"
 994                                " bits?\n");
 995
 996                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 997                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
 998                                         LUSTRE_VERSION_OFFSET_WARN ||
 999                      ocd->ocd_version < LUSTRE_VERSION_CODE -
1000                                         LUSTRE_VERSION_OFFSET_WARN)) {
1001                         /* Sigh, some compilers do not like #ifdef in the middle
1002                            of macro arguments */
1003                         const char *older = "older. Consider upgrading server "
1004                                             "or downgrading client";
1005                         const char *newer = "newer than client version. "
1006                                             "Consider upgrading client";
1007
1008                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
1009                                       "is much %s (%s)\n",
1010                                       obd2cli_tgt(imp->imp_obd),
1011                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1012                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1013                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1014                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1015                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
1016                                       newer : older, LUSTRE_VERSION_STRING);
1017                 }
1018
1019 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1020                 /* Check if server has LU-1252 fix applied to not always swab
1021                  * the IR MNE entries. Do this only once per connection.  This
1022                  * fixup is version-limited, because we don't want to carry the
1023                  * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
1024                  * need interop with unpatched 2.2 servers.  For newer servers,
1025                  * the client will do MNE swabbing only as needed.  LU-1644 */
1026                 if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1027                              !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
1028                              OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
1029                              OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
1030                              OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
1031                              strcmp(imp->imp_obd->obd_type->typ_name,
1032                                     LUSTRE_MGC_NAME) == 0))
1033                         imp->imp_need_mne_swab = 1;
1034                 else /* clear if server was upgraded since last connect */
1035                         imp->imp_need_mne_swab = 0;
1036 #else
1037 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1038 #endif
1039
1040                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
1041                         /* We sent to the server ocd_cksum_types with bits set
1042                          * for algorithms we understand. The server masked off
1043                          * the checksum types it doesn't support */
1044                         if ((ocd->ocd_cksum_types &
1045                              cksum_types_supported_client()) == 0) {
1046                                 LCONSOLE_WARN("The negotiation of the checksum "
1047                                               "alogrithm to use with server %s "
1048                                               "failed (%x/%x), disabling "
1049                                               "checksums\n",
1050                                               obd2cli_tgt(imp->imp_obd),
1051                                               ocd->ocd_cksum_types,
1052                                               cksum_types_supported_client());
1053                                 cli->cl_checksum = 0;
1054                                 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
1055                         } else {
1056                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1057                         }
1058                 } else {
1059                         /* The server does not support OBD_CONNECT_CKSUM.
1060                          * Enforce ADLER for backward compatibility*/
1061                         cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
1062                 }
1063                 cli->cl_cksum_type =cksum_type_select(cli->cl_supp_cksum_types);
1064
1065                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
1066                         cli->cl_max_pages_per_rpc =
1067                                 min(ocd->ocd_brw_size >> PAGE_CACHE_SHIFT,
1068                                     cli->cl_max_pages_per_rpc);
1069                 else if (imp->imp_connect_op == MDS_CONNECT ||
1070                          imp->imp_connect_op == MGS_CONNECT)
1071                         cli->cl_max_pages_per_rpc = 1;
1072
1073                 /* Reset ns_connect_flags only for initial connect. It might be
1074                  * changed in while using FS and if we reset it in reconnect
1075                  * this leads to losing user settings done before such as
1076                  * disable lru_resize, etc. */
1077                 if (old_connect_flags != exp_connect_flags(exp) ||
1078                     aa->pcaa_initial_connect) {
1079                         CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
1080                                "flags: "LPX64"\n", imp->imp_obd->obd_name,
1081                               ocd->ocd_connect_flags);
1082                         imp->imp_obd->obd_namespace->ns_connect_flags =
1083                                 ocd->ocd_connect_flags;
1084                         imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1085                                 ocd->ocd_connect_flags;
1086                 }
1087
1088                 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1089                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1090                         /* We need a per-message support flag, because
1091                            a. we don't know if the incoming connect reply
1092                               supports AT or not (in reply_in_callback)
1093                               until we unpack it.
1094                            b. failovered server means export and flags are gone
1095                               (in ptlrpc_send_reply).
1096                            Can only be set when we know AT is supported at
1097                            both ends */
1098                         imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1099                 else
1100                         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1101
1102                 if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1103                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1104                         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1105                 else
1106                         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1107
1108                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1109                         (cli->cl_max_pages_per_rpc > 0));
1110         }
1111
1112 out:
1113         imp->imp_connect_tried = 1;
1114
1115         if (rc != 0) {
1116                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1117                 if (rc == -EACCES) {
1118                         /*
1119                          * Give up trying to reconnect
1120                          * EACCES means client has no permission for connection
1121                          */
1122                         imp->imp_obd->obd_no_recov = 1;
1123                         ptlrpc_deactivate_import(imp);
1124                 }
1125
1126                 if (rc == -EPROTO) {
1127                         struct obd_connect_data *ocd;
1128
1129                         /* reply message might not be ready */
1130                         if (request->rq_repmsg == NULL)
1131                                 return -EPROTO;
1132
1133                         ocd = req_capsule_server_get(&request->rq_pill,
1134                                                      &RMF_CONNECT_DATA);
1135                         if (ocd &&
1136                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1137                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1138                            /* Actually servers are only supposed to refuse
1139                               connection from liblustre clients, so we should
1140                               never see this from VFS context */
1141                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1142                                         "(%d.%d.%d.%d)"
1143                                         " refused connection from this client "
1144                                         "with an incompatible version (%s).  "
1145                                         "Client must be recompiled\n",
1146                                         obd2cli_tgt(imp->imp_obd),
1147                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1148                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1149                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1150                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1151                                         LUSTRE_VERSION_STRING);
1152                                 ptlrpc_deactivate_import(imp);
1153                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1154                         }
1155                         return -EPROTO;
1156                 }
1157
1158                 ptlrpc_maybe_ping_import_soon(imp);
1159
1160                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1161                        obd2cli_tgt(imp->imp_obd),
1162                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1163         }
1164
1165         wake_up_all(&imp->imp_recovery_waitq);
1166         return rc;
1167 }
1168
1169 /**
1170  * interpret callback for "completed replay" RPCs.
1171  * \see signal_completed_replay
1172  */
1173 static int completed_replay_interpret(const struct lu_env *env,
1174                                       struct ptlrpc_request *req,
1175                                       void * data, int rc)
1176 {
1177         atomic_dec(&req->rq_import->imp_replay_inflight);
1178         if (req->rq_status == 0 &&
1179             !req->rq_import->imp_vbr_failed) {
1180                 ptlrpc_import_recovery_state_machine(req->rq_import);
1181         } else {
1182                 if (req->rq_import->imp_vbr_failed) {
1183                         CDEBUG(D_WARNING,
1184                                "%s: version recovery fails, reconnecting\n",
1185                                req->rq_import->imp_obd->obd_name);
1186                 } else {
1187                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1188                                      "reconnecting\n",
1189                                req->rq_import->imp_obd->obd_name,
1190                                req->rq_status);
1191                 }
1192                 ptlrpc_connect_import(req->rq_import);
1193         }
1194
1195         return 0;
1196 }
1197
1198 /**
1199  * Let server know that we have no requests to replay anymore.
1200  * Achieved by just sending a PING request
1201  */
1202 static int signal_completed_replay(struct obd_import *imp)
1203 {
1204         struct ptlrpc_request *req;
1205
1206         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1207                 return 0;
1208
1209         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1210         atomic_inc(&imp->imp_replay_inflight);
1211
1212         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1213                                         OBD_PING);
1214         if (req == NULL) {
1215                 atomic_dec(&imp->imp_replay_inflight);
1216                 return -ENOMEM;
1217         }
1218
1219         ptlrpc_request_set_replen(req);
1220         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1221         lustre_msg_add_flags(req->rq_reqmsg,
1222                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1223         if (AT_OFF)
1224                 req->rq_timeout *= 3;
1225         req->rq_interpret_reply = completed_replay_interpret;
1226
1227         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
1228         return 0;
1229 }
1230
1231 /**
1232  * In kernel code all import invalidation happens in its own
1233  * separate thread, so that whatever application happened to encounter
1234  * a problem could still be killed or otherwise continue
1235  */
1236 static int ptlrpc_invalidate_import_thread(void *data)
1237 {
1238         struct obd_import *imp = data;
1239
1240         unshare_fs_struct();
1241
1242         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1243                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1244                imp->imp_connection->c_remote_uuid.uuid);
1245
1246         ptlrpc_invalidate_import(imp);
1247
1248         if (obd_dump_on_eviction) {
1249                 CERROR("dump the log upon eviction\n");
1250                 libcfs_debug_dumplog();
1251         }
1252
1253         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1254         ptlrpc_import_recovery_state_machine(imp);
1255
1256         class_import_put(imp);
1257         return 0;
1258 }
1259
1260 /**
1261  * This is the state machine for client-side recovery on import.
1262  *
1263  * Typicaly we have two possibly paths. If we came to server and it is not
1264  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1265  * state and reconnect from scratch.
1266  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1267  * We go through our list of requests to replay and send them to server one by
1268  * one.
1269  * After sending all request from the list we change import state to
1270  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1271  * and also all the locks we don't yet have and wait for server to grant us.
1272  * After that we send a special "replay completed" request and change import
1273  * state to IMP_REPLAY_WAIT.
1274  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1275  * state and resend all requests from sending list.
1276  * After that we promote import to FULL state and send all delayed requests
1277  * and import is fully operational after that.
1278  *
1279  */
1280 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1281 {
1282         int rc = 0;
1283         int inflight;
1284         char *target_start;
1285         int target_len;
1286
1287         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1288                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1289                           &target_start, &target_len);
1290                 /* Don't care about MGC eviction */
1291                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1292                            LUSTRE_MGC_NAME) != 0) {
1293                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1294                                            "by %.*s; in progress operations "
1295                                            "using this service will fail.\n",
1296                                            imp->imp_obd->obd_name, target_len,
1297                                            target_start);
1298                 }
1299                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1300                        obd2cli_tgt(imp->imp_obd),
1301                        imp->imp_connection->c_remote_uuid.uuid);
1302                 /* reset vbr_failed flag upon eviction */
1303                 spin_lock(&imp->imp_lock);
1304                 imp->imp_vbr_failed = 0;
1305                 spin_unlock(&imp->imp_lock);
1306
1307                 {
1308                 struct task_struct *task;
1309                 /* bug 17802:  XXX client_disconnect_export vs connect request
1310                  * race. if client will evicted at this time, we start
1311                  * invalidate thread without reference to import and import can
1312                  * be freed at same time. */
1313                 class_import_get(imp);
1314                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1315                                      "ll_imp_inval");
1316                 if (IS_ERR(task)) {
1317                         class_import_put(imp);
1318                         CERROR("error starting invalidate thread: %d\n", rc);
1319                         rc = PTR_ERR(task);
1320                 } else {
1321                         rc = 0;
1322                 }
1323                 return rc;
1324                 }
1325         }
1326
1327         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1328                 CDEBUG(D_HA, "replay requested by %s\n",
1329                        obd2cli_tgt(imp->imp_obd));
1330                 rc = ptlrpc_replay_next(imp, &inflight);
1331                 if (inflight == 0 &&
1332                     atomic_read(&imp->imp_replay_inflight) == 0) {
1333                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1334                         rc = ldlm_replay_locks(imp);
1335                         if (rc)
1336                                 GOTO(out, rc);
1337                 }
1338                 rc = 0;
1339         }
1340
1341         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1342                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1343                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1344                         rc = signal_completed_replay(imp);
1345                         if (rc)
1346                                 GOTO(out, rc);
1347                 }
1348
1349         }
1350
1351         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1352                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1353                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1354                 }
1355         }
1356
1357         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1358                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1359                        obd2cli_tgt(imp->imp_obd),
1360                        imp->imp_connection->c_remote_uuid.uuid);
1361
1362                 rc = ptlrpc_resend(imp);
1363                 if (rc)
1364                         GOTO(out, rc);
1365                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1366                 ptlrpc_activate_import(imp);
1367
1368                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1369                           &target_start, &target_len);
1370                 LCONSOLE_INFO("%s: Connection restored to %.*s (at %s)\n",
1371                               imp->imp_obd->obd_name,
1372                               target_len, target_start,
1373                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1374         }
1375
1376         if (imp->imp_state == LUSTRE_IMP_FULL) {
1377                 wake_up_all(&imp->imp_recovery_waitq);
1378                 ptlrpc_wake_delayed(imp);
1379         }
1380
1381 out:
1382         return rc;
1383 }
1384
1385 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1386 {
1387         struct ptlrpc_request *req;
1388         int rq_opc, rc = 0;
1389         int nowait = imp->imp_obd->obd_force;
1390
1391         if (nowait)
1392                 GOTO(set_state, rc);
1393
1394         switch (imp->imp_connect_op) {
1395         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1396         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1397         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1398         default:
1399                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1400                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1401                 return -EINVAL;
1402         }
1403
1404         if (ptlrpc_import_in_recovery(imp)) {
1405                 struct l_wait_info lwi;
1406                 cfs_duration_t timeout;
1407
1408
1409                 if (AT_OFF) {
1410                         if (imp->imp_server_timeout)
1411                                 timeout = cfs_time_seconds(obd_timeout / 2);
1412                         else
1413                                 timeout = cfs_time_seconds(obd_timeout);
1414                 } else {
1415                         int idx = import_at_get_index(imp,
1416                                 imp->imp_client->cli_request_portal);
1417                         timeout = cfs_time_seconds(
1418                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1419                 }
1420
1421                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1422                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1423                 rc = l_wait_event(imp->imp_recovery_waitq,
1424                                   !ptlrpc_import_in_recovery(imp), &lwi);
1425
1426         }
1427
1428         spin_lock(&imp->imp_lock);
1429         if (imp->imp_state != LUSTRE_IMP_FULL)
1430                 GOTO(out, 0);
1431
1432         spin_unlock(&imp->imp_lock);
1433
1434         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1435                                         LUSTRE_OBD_VERSION, rq_opc);
1436         if (req) {
1437                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1438                  * it fails.  We can get through the above with a down server
1439                  * if the client doesn't know the server is gone yet. */
1440                 req->rq_no_resend = 1;
1441
1442                 /* We want client umounts to happen quickly, no matter the
1443                    server state... */
1444                 req->rq_timeout = min_t(int, req->rq_timeout,
1445                                         INITIAL_CONNECT_TIMEOUT);
1446
1447                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1448                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1449                 ptlrpc_request_set_replen(req);
1450                 rc = ptlrpc_queue_wait(req);
1451                 ptlrpc_req_finished(req);
1452         }
1453
1454 set_state:
1455         spin_lock(&imp->imp_lock);
1456 out:
1457         if (noclose)
1458                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1459         else
1460                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1461         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1462         spin_unlock(&imp->imp_lock);
1463
1464         return rc;
1465 }
1466 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1467
1468 void ptlrpc_cleanup_imp(struct obd_import *imp)
1469 {
1470         spin_lock(&imp->imp_lock);
1471         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1472         imp->imp_generation++;
1473         spin_unlock(&imp->imp_lock);
1474         ptlrpc_abort_inflight(imp);
1475 }
1476 EXPORT_SYMBOL(ptlrpc_cleanup_imp);
1477
1478 /* Adaptive Timeout utils */
1479 extern unsigned int at_min, at_max, at_history;
1480
1481 /* Bin into timeslices using AT_BINS bins.
1482    This gives us a max of the last binlimit*AT_BINS secs without the storage,
1483    but still smoothing out a return to normalcy from a slow response.
1484    (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1485 int at_measured(struct adaptive_timeout *at, unsigned int val)
1486 {
1487         unsigned int old = at->at_current;
1488         time_t now = cfs_time_current_sec();
1489         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1490
1491         LASSERT(at);
1492         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1493                val, at, now - at->at_binstart, at->at_current,
1494                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1495
1496         if (val == 0)
1497                 /* 0's don't count, because we never want our timeout to
1498                    drop to 0, and because 0 could mean an error */
1499                 return 0;
1500
1501         spin_lock(&at->at_lock);
1502
1503         if (unlikely(at->at_binstart == 0)) {
1504                 /* Special case to remove default from history */
1505                 at->at_current = val;
1506                 at->at_worst_ever = val;
1507                 at->at_worst_time = now;
1508                 at->at_hist[0] = val;
1509                 at->at_binstart = now;
1510         } else if (now - at->at_binstart < binlimit ) {
1511                 /* in bin 0 */
1512                 at->at_hist[0] = max(val, at->at_hist[0]);
1513                 at->at_current = max(val, at->at_current);
1514         } else {
1515                 int i, shift;
1516                 unsigned int maxv = val;
1517                 /* move bins over */
1518                 shift = (now - at->at_binstart) / binlimit;
1519                 LASSERT(shift > 0);
1520                 for(i = AT_BINS - 1; i >= 0; i--) {
1521                         if (i >= shift) {
1522                                 at->at_hist[i] = at->at_hist[i - shift];
1523                                 maxv = max(maxv, at->at_hist[i]);
1524                         } else {
1525                                 at->at_hist[i] = 0;
1526                         }
1527                 }
1528                 at->at_hist[0] = val;
1529                 at->at_current = maxv;
1530                 at->at_binstart += shift * binlimit;
1531         }
1532
1533         if (at->at_current > at->at_worst_ever) {
1534                 at->at_worst_ever = at->at_current;
1535                 at->at_worst_time = now;
1536         }
1537
1538         if (at->at_flags & AT_FLG_NOHIST)
1539                 /* Only keep last reported val; keeping the rest of the history
1540                    for proc only */
1541                 at->at_current = val;
1542
1543         if (at_max > 0)
1544                 at->at_current =  min(at->at_current, at_max);
1545         at->at_current =  max(at->at_current, at_min);
1546
1547         if (at->at_current != old)
1548                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1549                        "(val=%u) hist %u %u %u %u\n", at,
1550                        old, at->at_current, at->at_current - old, val,
1551                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1552                        at->at_hist[3]);
1553
1554         /* if we changed, report the old value */
1555         old = (at->at_current != old) ? old : 0;
1556
1557         spin_unlock(&at->at_lock);
1558         return old;
1559 }
1560
1561 /* Find the imp_at index for a given portal; assign if space available */
1562 int import_at_get_index(struct obd_import *imp, int portal)
1563 {
1564         struct imp_at *at = &imp->imp_at;
1565         int i;
1566
1567         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1568                 if (at->iat_portal[i] == portal)
1569                         return i;
1570                 if (at->iat_portal[i] == 0)
1571                         /* unused */
1572                         break;
1573         }
1574
1575         /* Not found in list, add it under a lock */
1576         spin_lock(&imp->imp_lock);
1577
1578         /* Check unused under lock */
1579         for (; i < IMP_AT_MAX_PORTALS; i++) {
1580                 if (at->iat_portal[i] == portal)
1581                         goto out;
1582                 if (at->iat_portal[i] == 0)
1583                         /* unused */
1584                         break;
1585         }
1586
1587         /* Not enough portals? */
1588         LASSERT(i < IMP_AT_MAX_PORTALS);
1589
1590         at->iat_portal[i] = portal;
1591 out:
1592         spin_unlock(&imp->imp_lock);
1593         return i;
1594 }