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, 2013, 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 #ifndef __KERNEL__
  43 # include <liblustre.h>
  44 #endif
  45
  46 #include <obd_support.h>
  47 #include <lustre_ha.h>
  48 #include <lustre_net.h>
  49 #include <lustre_import.h>
  50 #include <lustre_export.h>
  51 #include <obd.h>
  52 #include <obd_cksum.h>
  53 #include <obd_class.h>
  54
  55 #include "ptlrpc_internal.h"
  56
  57 struct ptlrpc_connect_async_args {
  58          __u64 pcaa_peer_committed;
  59         int pcaa_initial_connect;
  60 };
  61
  62 /**
  63  * Updates import \a imp current state to provided \a state value
  64  * Helper function. Must be called under imp_lock.
  65  */
  66 static void __import_set_state(struct obd_import *imp,
  67                                enum lustre_imp_state state)
  68 {
  69         imp->imp_state = state;
  70         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
  71         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
  72                 cfs_time_current_sec();
  73         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
  74                 IMP_STATE_HIST_LEN;
  75 }
  76
  77 /* A CLOSED import should remain so. */
  78 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  79 do {                                                                           \
  80         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  81                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  82                       imp, obd2cli_tgt(imp->imp_obd),                          \
  83                       ptlrpc_import_state_name(imp->imp_state),                \
  84                       ptlrpc_import_state_name(state));                        \
  85                __import_set_state(imp, state);                                 \
  86         }                                                                      \
  87 } while(0)
  88
  89 #define IMPORT_SET_STATE(imp, state)                                    \
  90 do {                                                                    \
  91         spin_lock(&imp->imp_lock);                                      \
  92         IMPORT_SET_STATE_NOLOCK(imp, state);                            \
  93         spin_unlock(&imp->imp_lock);                                    \
  94 } while(0)
  95
  96
  97 static int ptlrpc_connect_interpret(const struct lu_env *env,
  98                                     struct ptlrpc_request *request,
  99                                     void * data, int rc);
 100 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
 101
 102 /* Only this function is allowed to change the import state when it is
 103  * CLOSED. I would rather refcount the import and free it after
 104  * disconnection like we do with exports. To do that, the client_obd
 105  * will need to save the peer info somewhere other than in the import,
 106  * though. */
 107 int ptlrpc_init_import(struct obd_import *imp)
 108 {
 109         spin_lock(&imp->imp_lock);
 110
 111         imp->imp_generation++;
 112         imp->imp_state =  LUSTRE_IMP_NEW;
 113
 114         spin_unlock(&imp->imp_lock);
 115
 116         return 0;
 117 }
 118 EXPORT_SYMBOL(ptlrpc_init_import);
 119
 120 #define UUID_STR "_UUID"
 121 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
 122 {
 123         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 124                 ? uuid : uuid + strlen(prefix);
 125
 126         *uuid_len = strlen(*uuid_start);
 127
 128         if (*uuid_len < strlen(UUID_STR))
 129                 return;
 130
 131         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 132                     UUID_STR, strlen(UUID_STR)))
 133                 *uuid_len -= strlen(UUID_STR);
 134 }
 135 EXPORT_SYMBOL(deuuidify);
 136
 137 /**
 138  * Returns true if import was FULL, false if import was already not
 139  * connected.
 140  * @imp - import to be disconnected
 141  * @conn_cnt - connection count (epoch) of the request that timed out
 142  *             and caused the disconnection.  In some cases, multiple
 143  *             inflight requests can fail to a single target (e.g. OST
 144  *             bulk requests) and if one has already caused a reconnection
 145  *             (increasing the import->conn_cnt) the older failure should
 146  *             not also cause a reconnection.  If zero it forces a reconnect.
 147  */
 148 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 149 {
 150         int rc = 0;
 151
 152         spin_lock(&imp->imp_lock);
 153
 154         if (imp->imp_state == LUSTRE_IMP_FULL &&
 155             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 156                 char *target_start;
 157                 int   target_len;
 158
 159                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 160                           &target_start, &target_len);
 161
 162                 if (imp->imp_replayable) {
 163                         LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
 164                                "lost; in progress operations using this "
 165                                "service will wait for recovery to complete\n",
 166                                imp->imp_obd->obd_name, target_len, target_start,
 167                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 168                 } else {
 169                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
 170                                "%.*s (at %s) was lost; in progress "
 171                                "operations using this service will fail\n",
 172                                imp->imp_obd->obd_name,
 173                                target_len, target_start,
 174                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 175                 }
 176                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 177                 spin_unlock(&imp->imp_lock);
 178
 179                 if (obd_dump_on_timeout)
 180                         libcfs_debug_dumplog();
 181
 182                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 183                 rc = 1;
 184         } else {
 185                 spin_unlock(&imp->imp_lock);
 186                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 187                        imp->imp_client->cli_name, imp,
 188                        (imp->imp_state == LUSTRE_IMP_FULL &&
 189                         imp->imp_conn_cnt > conn_cnt) ?
 190                        "reconnected" : "not connected", imp->imp_conn_cnt,
 191                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 192         }
 193
 194         return rc;
 195 }
 196
 197 /* Must be called with imp_lock held! */
 198 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 199 {
 200         ENTRY;
 201         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 202
 203         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 204         imp->imp_invalid = 1;
 205         imp->imp_generation++;
 206         spin_unlock(&imp->imp_lock);
 207
 208         ptlrpc_abort_inflight(imp);
 209         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 210
 211         EXIT;
 212 }
 213
 214 /*
 215  * This acts as a barrier; all existing requests are rejected, and
 216  * no new requests will be accepted until the import is valid again.
 217  */
 218 void ptlrpc_deactivate_import(struct obd_import *imp)
 219 {
 220         spin_lock(&imp->imp_lock);
 221         ptlrpc_deactivate_and_unlock_import(imp);
 222 }
 223 EXPORT_SYMBOL(ptlrpc_deactivate_import);
 224
 225 static unsigned int
 226 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
 227 {
 228         long dl;
 229
 230         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
 231               (req->rq_phase == RQ_PHASE_BULK) ||
 232               (req->rq_phase == RQ_PHASE_NEW)))
 233                 return 0;
 234
 235         if (req->rq_timedout)
 236                 return 0;
 237
 238         if (req->rq_phase == RQ_PHASE_NEW)
 239                 dl = req->rq_sent;
 240         else
 241                 dl = req->rq_deadline;
 242
 243         if (dl <= now)
 244                 return 0;
 245
 246         return dl - now;
 247 }
 248
 249 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
 250 {
 251         time_t now = cfs_time_current_sec();
 252         cfs_list_t *tmp, *n;
 253         struct ptlrpc_request *req;
 254         unsigned int timeout = 0;
 255
 256         spin_lock(&imp->imp_lock);
 257         cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 258                 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
 259                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 260         }
 261         spin_unlock(&imp->imp_lock);
 262         return timeout;
 263 }
 264
 265 /**
 266  * This function will invalidate the import, if necessary, then block
 267  * for all the RPC completions, and finally notify the obd to
 268  * invalidate its state (ie cancel locks, clear pending requests,
 269  * etc).
 270  */
 271 void ptlrpc_invalidate_import(struct obd_import *imp)
 272 {
 273         cfs_list_t *tmp, *n;
 274         struct ptlrpc_request *req;
 275         struct l_wait_info lwi;
 276         unsigned int timeout;
 277         int rc;
 278
 279         cfs_atomic_inc(&imp->imp_inval_count);
 280
 281         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 282                 ptlrpc_deactivate_import(imp);
 283
 284         LASSERT(imp->imp_invalid);
 285
 286         /* Wait forever until inflight == 0. We really can't do it another
 287          * way because in some cases we need to wait for very long reply
 288          * unlink. We can't do anything before that because there is really
 289          * no guarantee that some rdma transfer is not in progress right now. */
 290         do {
 291                 /* Calculate max timeout for waiting on rpcs to error
 292                  * out. Use obd_timeout if calculated value is smaller
 293                  * than it. */
 294                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 295                         timeout = ptlrpc_inflight_timeout(imp);
 296                         timeout += timeout / 3;
 297
 298                         if (timeout == 0)
 299                                 timeout = obd_timeout;
 300                 } else {
 301                         /* decrease the interval to increase race condition */
 302                         timeout = 1;
 303                 }
 304
 305                 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
 306                        timeout);
 307
 308                 /* Wait for all requests to error out and call completion
 309                  * callbacks. Cap it at obd_timeout -- these should all
 310                  * have been locally cancelled by ptlrpc_abort_inflight. */
 311                 lwi = LWI_TIMEOUT_INTERVAL(
 312                         cfs_timeout_cap(cfs_time_seconds(timeout)),
 313                         (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
 314                         NULL, NULL);
 315                 rc = l_wait_event(imp->imp_recovery_waitq,
 316                                   (cfs_atomic_read(&imp->imp_inflight) == 0),
 317                                   &lwi);
 318                 if (rc) {
 319                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 320
 321                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 322                                cli_tgt, rc,
 323                                cfs_atomic_read(&imp->imp_inflight));
 324
 325                         spin_lock(&imp->imp_lock);
 326                         if (cfs_atomic_read(&imp->imp_inflight) == 0) {
 327                                 int count = cfs_atomic_read(&imp->imp_unregistering);
 328
 329                                 /* We know that "unregistering" rpcs only can
 330                                  * survive in sending or delaying lists (they
 331                                  * maybe waiting for long reply unlink in
 332                                  * sluggish nets). Let's check this. If there
 333                                  * is no inflight and unregistering != 0, this
 334                                  * is bug. */
 335                                 LASSERTF(count == 0, "Some RPCs are still "
 336                                          "unregistering: %d\n", count);
 337
 338                                 /* Let's save one loop as soon as inflight have
 339                                  * dropped to zero. No new inflights possible at
 340                                  * this point. */
 341                                 rc = 0;
 342                         } else {
 343                                 cfs_list_for_each_safe(tmp, n,
 344                                                        &imp->imp_sending_list) {
 345                                         req = cfs_list_entry(tmp,
 346                                                              struct ptlrpc_request,
 347                                                              rq_list);
 348                                         DEBUG_REQ(D_ERROR, req,
 349                                                   "still on sending list");
 350                                 }
 351                                 cfs_list_for_each_safe(tmp, n,
 352                                                        &imp->imp_delayed_list) {
 353                                         req = cfs_list_entry(tmp,
 354                                                              struct ptlrpc_request,
 355                                                              rq_list);
 356                                         DEBUG_REQ(D_ERROR, req,
 357                                                   "still on delayed list");
 358                                 }
 359
 360                                 CERROR("%s: RPCs in \"%s\" phase found (%d). "
 361                                        "Network is sluggish? Waiting them "
 362                                        "to error out.\n", cli_tgt,
 363                                        ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
 364                                        cfs_atomic_read(&imp->
 365                                                        imp_unregistering));
 366                         }
 367                         spin_unlock(&imp->imp_lock);
 368                   }
 369         } while (rc != 0);
 370
 371         /*
 372          * Let's additionally check that no new rpcs added to import in
 373          * "invalidate" state.
 374          */
 375         LASSERT(cfs_atomic_read(&imp->imp_inflight) == 0);
 376         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 377         sptlrpc_import_flush_all_ctx(imp);
 378
 379         cfs_atomic_dec(&imp->imp_inval_count);
 380         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
 381 }
 382 EXPORT_SYMBOL(ptlrpc_invalidate_import);
 383
 384 /* unset imp_invalid */
 385 void ptlrpc_activate_import(struct obd_import *imp)
 386 {
 387         struct obd_device *obd = imp->imp_obd;
 388
 389         spin_lock(&imp->imp_lock);
 390         imp->imp_invalid = 0;
 391         spin_unlock(&imp->imp_lock);
 392         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 393 }
 394 EXPORT_SYMBOL(ptlrpc_activate_import);
 395
 396 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 397 {
 398         ENTRY;
 399
 400         LASSERT(!imp->imp_dlm_fake);
 401
 402         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 403                 if (!imp->imp_replayable) {
 404                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 405                                "auto-deactivating\n",
 406                                obd2cli_tgt(imp->imp_obd),
 407                                imp->imp_connection->c_remote_uuid.uuid,
 408                                imp->imp_obd->obd_name);
 409                         ptlrpc_deactivate_import(imp);
 410                 }
 411
 412                 CDEBUG(D_HA, "%s: waking up pinger\n",
 413                        obd2cli_tgt(imp->imp_obd));
 414
 415                 spin_lock(&imp->imp_lock);
 416                 imp->imp_force_verify = 1;
 417                 spin_unlock(&imp->imp_lock);
 418
 419                 ptlrpc_pinger_wake_up();
 420         }
 421         EXIT;
 422 }
 423 EXPORT_SYMBOL(ptlrpc_fail_import);
 424
 425 int ptlrpc_reconnect_import(struct obd_import *imp)
 426 {
 427         ptlrpc_set_import_discon(imp, 0);
 428         /* Force a new connect attempt */
 429         ptlrpc_invalidate_import(imp);
 430         /* Do a fresh connect next time by zeroing the handle */
 431         ptlrpc_disconnect_import(imp, 1);
 432         /* Wait for all invalidate calls to finish */
 433         if (cfs_atomic_read(&imp->imp_inval_count) > 0) {
 434                 int rc;
 435                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 436                 rc = l_wait_event(imp->imp_recovery_waitq,
 437                                   (cfs_atomic_read(&imp->imp_inval_count) == 0),
 438                                   &lwi);
 439                 if (rc)
 440                         CERROR("Interrupted, inval=%d\n",
 441                                cfs_atomic_read(&imp->imp_inval_count));
 442         }
 443
 444         /* Allow reconnect attempts */
 445         imp->imp_obd->obd_no_recov = 0;
 446         /* Remove 'invalid' flag */
 447         ptlrpc_activate_import(imp);
 448         /* Attempt a new connect */
 449         ptlrpc_recover_import(imp, NULL, 0);
 450         return 0;
 451 }
 452 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 453
 454 /**
 455  * Connection on import \a imp is changed to another one (if more than one is
 456  * present). We typically chose connection that we have not tried to connect to
 457  * the longest
 458  */
 459 static int import_select_connection(struct obd_import *imp)
 460 {
 461         struct obd_import_conn *imp_conn = NULL, *conn;
 462         struct obd_export *dlmexp;
 463         char *target_start;
 464         int target_len, tried_all = 1;
 465         ENTRY;
 466
 467         spin_lock(&imp->imp_lock);
 468
 469         if (cfs_list_empty(&imp->imp_conn_list)) {
 470                 CERROR("%s: no connections available\n",
 471                        imp->imp_obd->obd_name);
 472                 spin_unlock(&imp->imp_lock);
 473                 RETURN(-EINVAL);
 474         }
 475
 476         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 477                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 478                        imp->imp_obd->obd_name,
 479                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 480                        conn->oic_last_attempt);
 481
 482                 /* If we have not tried this connection since
 483                    the last successful attempt, go with this one */
 484                 if ((conn->oic_last_attempt == 0) ||
 485                     cfs_time_beforeq_64(conn->oic_last_attempt,
 486                                        imp->imp_last_success_conn)) {
 487                         imp_conn = conn;
 488                         tried_all = 0;
 489                         break;
 490                 }
 491
 492                 /* If all of the connections have already been tried
 493                    since the last successful connection; just choose the
 494                    least recently used */
 495                 if (!imp_conn)
 496                         imp_conn = conn;
 497                 else if (cfs_time_before_64(conn->oic_last_attempt,
 498                                             imp_conn->oic_last_attempt))
 499                         imp_conn = conn;
 500         }
 501
 502         /* if not found, simply choose the current one */
 503         if (!imp_conn || imp->imp_force_reconnect) {
 504                 LASSERT(imp->imp_conn_current);
 505                 imp_conn = imp->imp_conn_current;
 506                 tried_all = 0;
 507         }
 508         LASSERT(imp_conn->oic_conn);
 509
 510         /* If we've tried everything, and we're back to the beginning of the
 511            list, increase our timeout and try again. It will be reset when
 512            we do finally connect. (FIXME: really we should wait for all network
 513            state associated with the last connection attempt to drain before
 514            trying to reconnect on it.) */
 515         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 516                 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
 517                 if (at_get(at) < CONNECTION_SWITCH_MAX) {
 518                         at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
 519                         if (at_get(at) > CONNECTION_SWITCH_MAX)
 520                                 at_reset(at, CONNECTION_SWITCH_MAX);
 521                 }
 522                 LASSERT(imp_conn->oic_last_attempt);
 523                 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
 524                         "to %ds\n", imp->imp_obd->obd_name, at_get(at));
 525         }
 526
 527         imp_conn->oic_last_attempt = cfs_time_current_64();
 528
 529         /* switch connection, don't mind if it's same as the current one */
 530         if (imp->imp_connection)
 531                 ptlrpc_connection_put(imp->imp_connection);
 532         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 533
 534         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 535         LASSERT(dlmexp != NULL);
 536         if (dlmexp->exp_connection)
 537                 ptlrpc_connection_put(dlmexp->exp_connection);
 538         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 539         class_export_put(dlmexp);
 540
 541         if (imp->imp_conn_current != imp_conn) {
 542                 if (imp->imp_conn_current) {
 543                         deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 544                                   &target_start, &target_len);
 545
 546                         CDEBUG(D_HA, "%s: Connection changing to"
 547                                " %.*s (at %s)\n",
 548                                imp->imp_obd->obd_name,
 549                                target_len, target_start,
 550                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 551                 }
 552
 553                 imp->imp_conn_current = imp_conn;
 554         }
 555
 556         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 557                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 558                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 559
 560         spin_unlock(&imp->imp_lock);
 561
 562         RETURN(0);
 563 }
 564
 565 /*
 566  * must be called under imp_lock
 567  */
 568 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 569 {
 570         struct ptlrpc_request *req;
 571         cfs_list_t *tmp;
 572
 573         if (cfs_list_empty(&imp->imp_replay_list))
 574                 return 0;
 575         tmp = imp->imp_replay_list.next;
 576         req = cfs_list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 577         *transno = req->rq_transno;
 578         if (req->rq_transno == 0) {
 579                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 580                 LBUG();
 581         }
 582
 583         return 1;
 584 }
 585
 586 /**
 587  * Attempt to (re)connect import \a imp. This includes all preparations,
 588  * initializing CONNECT RPC request and passing it to ptlrpcd for
 589  * actual sending.
 590  * Returns 0 on success or error code.
 591  */
 592 int ptlrpc_connect_import(struct obd_import *imp)
 593 {
 594         struct obd_device *obd = imp->imp_obd;
 595         int initial_connect = 0;
 596         int set_transno = 0;
 597         __u64 committed_before_reconnect = 0;
 598         struct ptlrpc_request *request;
 599         char *bufs[] = { NULL,
 600                          obd2cli_tgt(imp->imp_obd),
 601                          obd->obd_uuid.uuid,
 602                          (char *)&imp->imp_dlm_handle,
 603                          (char *)&imp->imp_connect_data };
 604         struct ptlrpc_connect_async_args *aa;
 605         int rc;
 606         ENTRY;
 607
 608         spin_lock(&imp->imp_lock);
 609         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 610                 spin_unlock(&imp->imp_lock);
 611                 CERROR("can't connect to a closed import\n");
 612                 RETURN(-EINVAL);
 613         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 614                 spin_unlock(&imp->imp_lock);
 615                 CERROR("already connected\n");
 616                 RETURN(0);
 617         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 618                 spin_unlock(&imp->imp_lock);
 619                 CERROR("already connecting\n");
 620                 RETURN(-EALREADY);
 621         }
 622
 623         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 624
 625         imp->imp_conn_cnt++;
 626         imp->imp_resend_replay = 0;
 627
 628         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 629                 initial_connect = 1;
 630         else
 631                 committed_before_reconnect = imp->imp_peer_committed_transno;
 632
 633         set_transno = ptlrpc_first_transno(imp,
 634                                            &imp->imp_connect_data.ocd_transno);
 635         spin_unlock(&imp->imp_lock);
 636
 637         rc = import_select_connection(imp);
 638         if (rc)
 639                 GOTO(out, rc);
 640
 641         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 642         if (rc)
 643                 GOTO(out, rc);
 644
 645         /* Reset connect flags to the originally requested flags, in case
 646          * the server is updated on-the-fly we will get the new features. */
 647         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 648         /* Reset ocd_version each time so the server knows the exact versions */
 649         imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
 650         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 651         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 652
 653         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 654                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 655         if (rc)
 656                 GOTO(out, rc);
 657
 658         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 659         if (request == NULL)
 660                 GOTO(out, rc = -ENOMEM);
 661
 662         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 663                                       imp->imp_connect_op, bufs, NULL);
 664         if (rc) {
 665                 ptlrpc_request_free(request);
 666                 GOTO(out, rc);
 667         }
 668
 669         /* Report the rpc service time to the server so that it knows how long
 670          * to wait for clients to join recovery */
 671         lustre_msg_set_service_time(request->rq_reqmsg,
 672                                     at_timeout2est(request->rq_timeout));
 673
 674         /* The amount of time we give the server to process the connect req.
 675          * import_select_connection will increase the net latency on
 676          * repeated reconnect attempts to cover slow networks.
 677          * We override/ignore the server rpc completion estimate here,
 678          * which may be large if this is a reconnect attempt */
 679         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 680         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 681
 682 #ifndef __KERNEL__
 683         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 684 #endif
 685         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 686
 687         request->rq_no_resend = request->rq_no_delay = 1;
 688         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 689         /* Allow a slightly larger reply for future growth compatibility */
 690         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 691                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 692         ptlrpc_request_set_replen(request);
 693         request->rq_interpret_reply = ptlrpc_connect_interpret;
 694
 695         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 696         aa = ptlrpc_req_async_args(request);
 697         memset(aa, 0, sizeof *aa);
 698
 699         aa->pcaa_peer_committed = committed_before_reconnect;
 700         aa->pcaa_initial_connect = initial_connect;
 701
 702         if (aa->pcaa_initial_connect) {
 703                 spin_lock(&imp->imp_lock);
 704                 imp->imp_replayable = 1;
 705                 spin_unlock(&imp->imp_lock);
 706                 lustre_msg_add_op_flags(request->rq_reqmsg,
 707                                         MSG_CONNECT_INITIAL);
 708         }
 709
 710         if (set_transno)
 711                 lustre_msg_add_op_flags(request->rq_reqmsg,
 712                                         MSG_CONNECT_TRANSNO);
 713
 714         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
 715                   request->rq_timeout);
 716         ptlrpcd_add_req(request, PDL_POLICY_ROUND, -1);
 717         rc = 0;
 718 out:
 719         if (rc != 0) {
 720                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 721         }
 722
 723         RETURN(rc);
 724 }
 725 EXPORT_SYMBOL(ptlrpc_connect_import);
 726
 727 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 728 {
 729 #ifdef __KERNEL__
 730         int force_verify;
 731
 732         spin_lock(&imp->imp_lock);
 733         force_verify = imp->imp_force_verify != 0;
 734         spin_unlock(&imp->imp_lock);
 735
 736         if (force_verify)
 737                 ptlrpc_pinger_wake_up();
 738 #else
 739         /* liblustre has no pinger thread, so we wakeup pinger anyway */
 740         ptlrpc_pinger_wake_up();
 741 #endif
 742 }
 743
 744 static int ptlrpc_busy_reconnect(int rc)
 745 {
 746         return (rc == -EBUSY) || (rc == -EAGAIN);
 747 }
 748
 749 /**
 750  * interpret_reply callback for connect RPCs.
 751  * Looks into returned status of connect operation and decides
 752  * what to do with the import - i.e enter recovery, promote it to
 753  * full state for normal operations of disconnect it due to an error.
 754  */
 755 static int ptlrpc_connect_interpret(const struct lu_env *env,
 756                                     struct ptlrpc_request *request,
 757                                     void *data, int rc)
 758 {
 759         struct ptlrpc_connect_async_args *aa = data;
 760         struct obd_import *imp = request->rq_import;
 761         struct client_obd *cli = &imp->imp_obd->u.cli;
 762         struct lustre_handle old_hdl;
 763         __u64 old_connect_flags;
 764         int msg_flags;
 765         struct obd_connect_data *ocd;
 766         struct obd_export *exp;
 767         int ret;
 768         ENTRY;
 769
 770         spin_lock(&imp->imp_lock);
 771         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 772                 imp->imp_connect_tried = 1;
 773                 spin_unlock(&imp->imp_lock);
 774                 RETURN(0);
 775         }
 776
 777         if (rc) {
 778                 /* if this reconnect to busy export - not need select new target
 779                  * for connecting*/
 780                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 781                 spin_unlock(&imp->imp_lock);
 782                 ptlrpc_maybe_ping_import_soon(imp);
 783                 GOTO(out, rc);
 784         }
 785         spin_unlock(&imp->imp_lock);
 786
 787         LASSERT(imp->imp_conn_current);
 788
 789         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 790
 791         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 792                                    RCL_SERVER);
 793         /* server replied obd_connect_data is always bigger */
 794         ocd = req_capsule_server_sized_get(&request->rq_pill,
 795                                            &RMF_CONNECT_DATA, ret);
 796
 797         if (ocd == NULL) {
 798                 CERROR("%s: no connect data from server\n",
 799                        imp->imp_obd->obd_name);
 800                 rc = -EPROTO;
 801                 GOTO(out, rc);
 802         }
 803
 804         spin_lock(&imp->imp_lock);
 805
 806         /* All imports are pingable */
 807         imp->imp_pingable = 1;
 808         imp->imp_force_reconnect = 0;
 809         imp->imp_force_verify = 0;
 810
 811         imp->imp_connect_data = *ocd;
 812
 813         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
 814                imp->imp_obd->obd_name, ocd->ocd_instance);
 815         exp = class_conn2export(&imp->imp_dlm_handle);
 816
 817         spin_unlock(&imp->imp_lock);
 818
 819         /* check that server granted subset of flags we asked for. */
 820         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
 821             ocd->ocd_connect_flags) {
 822                 CERROR("%s: Server didn't granted asked subset of flags: "
 823                        "asked="LPX64" grranted="LPX64"\n",
 824                        imp->imp_obd->obd_name,imp->imp_connect_flags_orig,
 825                        ocd->ocd_connect_flags);
 826                 GOTO(out, rc = -EPROTO);
 827         }
 828
 829         if (!exp) {
 830                 /* This could happen if export is cleaned during the
 831                    connect attempt */
 832                 CERROR("%s: missing export after connect\n",
 833                        imp->imp_obd->obd_name);
 834                 GOTO(out, rc = -ENODEV);
 835         }
 836         old_connect_flags = exp_connect_flags(exp);
 837         exp->exp_connect_data = *ocd;
 838         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
 839         class_export_put(exp);
 840
 841         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 842
 843         if (aa->pcaa_initial_connect) {
 844                 spin_lock(&imp->imp_lock);
 845                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 846                         imp->imp_replayable = 1;
 847                         spin_unlock(&imp->imp_lock);
 848                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 849                                obd2cli_tgt(imp->imp_obd));
 850                 } else {
 851                         imp->imp_replayable = 0;
 852                         spin_unlock(&imp->imp_lock);
 853                 }
 854
 855                 /* if applies, adjust the imp->imp_msg_magic here
 856                  * according to reply flags */
 857
 858                 imp->imp_remote_handle =
 859                                 *lustre_msg_get_handle(request->rq_repmsg);
 860
 861                 /* Initial connects are allowed for clients with non-random
 862                  * uuids when servers are in recovery.  Simply signal the
 863                  * servers replay is complete and wait in REPLAY_WAIT. */
 864                 if (msg_flags & MSG_CONNECT_RECOVERING) {
 865                         CDEBUG(D_HA, "connect to %s during recovery\n",
 866                                obd2cli_tgt(imp->imp_obd));
 867                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 868                 } else {
 869                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 870                         ptlrpc_activate_import(imp);
 871                 }
 872
 873                 GOTO(finish, rc = 0);
 874         }
 875
 876         /* Determine what recovery state to move the import to. */
 877         if (MSG_CONNECT_RECONNECT & msg_flags) {
 878                 memset(&old_hdl, 0, sizeof(old_hdl));
 879                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 880                             sizeof (old_hdl))) {
 881                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
 882                                       "bad handle "LPX64"\n",
 883                                       obd2cli_tgt(imp->imp_obd),
 884                                       imp->imp_connection->c_remote_uuid.uuid,
 885                                       imp->imp_dlm_handle.cookie);
 886                         GOTO(out, rc = -ENOTCONN);
 887                 }
 888
 889                 if (memcmp(&imp->imp_remote_handle,
 890                            lustre_msg_get_handle(request->rq_repmsg),
 891                            sizeof(imp->imp_remote_handle))) {
 892                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
 893                                 D_HA : D_WARNING;
 894
 895                         /* Bug 16611/14775: if server handle have changed,
 896                          * that means some sort of disconnection happened.
 897                          * If the server is not in recovery, that also means it
 898                          * already erased all of our state because of previous
 899                          * eviction. If it is in recovery - we are safe to
 900                          * participate since we can reestablish all of our state
 901                          * with server again */
 902                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
 903                                 CDEBUG(level,"%s@%s changed server handle from "
 904                                        LPX64" to "LPX64
 905                                        " but is still in recovery\n",
 906                                        obd2cli_tgt(imp->imp_obd),
 907                                        imp->imp_connection->c_remote_uuid.uuid,
 908                                        imp->imp_remote_handle.cookie,
 909                                        lustre_msg_get_handle(
 910                                        request->rq_repmsg)->cookie);
 911                         } else {
 912                                 LCONSOLE_WARN("Evicted from %s (at %s) "
 913                                               "after server handle changed from "
 914                                               LPX64" to "LPX64"\n",
 915                                               obd2cli_tgt(imp->imp_obd),
 916                                               imp->imp_connection-> \
 917                                               c_remote_uuid.uuid,
 918                                               imp->imp_remote_handle.cookie,
 919                                               lustre_msg_get_handle(
 920                                               request->rq_repmsg)->cookie);
 921                         }
 922
 923
 924                         imp->imp_remote_handle =
 925                                      *lustre_msg_get_handle(request->rq_repmsg);
 926
 927                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
 928                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 929                                 GOTO(finish, rc = 0);
 930                         }
 931
 932                 } else {
 933                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 934                                obd2cli_tgt(imp->imp_obd),
 935                                imp->imp_connection->c_remote_uuid.uuid);
 936                 }
 937
 938                 if (imp->imp_invalid) {
 939                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
 940                                "marking evicted\n", imp->imp_obd->obd_name);
 941                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 942                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 943                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 944                                imp->imp_obd->obd_name,
 945                                obd2cli_tgt(imp->imp_obd));
 946
 947                         spin_lock(&imp->imp_lock);
 948                         imp->imp_resend_replay = 1;
 949                         spin_unlock(&imp->imp_lock);
 950
 951                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 952                 } else {
 953                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 954                 }
 955         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 956                 LASSERT(imp->imp_replayable);
 957                 imp->imp_remote_handle =
 958                                 *lustre_msg_get_handle(request->rq_repmsg);
 959                 imp->imp_last_replay_transno = 0;
 960                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 961         } else {
 962                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
 963                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
 964                 imp->imp_remote_handle =
 965                                 *lustre_msg_get_handle(request->rq_repmsg);
 966                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 967         }
 968
 969         /* Sanity checks for a reconnected import. */
 970         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 971                 CERROR("imp_replayable flag does not match server "
 972                        "after reconnect. We should LBUG right here.\n");
 973         }
 974
 975         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
 976             lustre_msg_get_last_committed(request->rq_repmsg) <
 977             aa->pcaa_peer_committed) {
 978                 CERROR("%s went back in time (transno "LPD64
 979                        " was previously committed, server now claims "LPD64
 980                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
 981                        "id=9646\n",
 982                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 983                        lustre_msg_get_last_committed(request->rq_repmsg));
 984         }
 985
 986 finish:
 987         rc = ptlrpc_import_recovery_state_machine(imp);
 988         if (rc != 0) {
 989                 if (rc == -ENOTCONN) {
 990                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 991                                "invalidating and reconnecting\n",
 992                                obd2cli_tgt(imp->imp_obd),
 993                                imp->imp_connection->c_remote_uuid.uuid);
 994                         ptlrpc_connect_import(imp);
 995                         imp->imp_connect_tried = 1;
 996                         RETURN(0);
 997                 }
 998         } else {
 999
1000                 spin_lock(&imp->imp_lock);
1001                 cfs_list_del(&imp->imp_conn_current->oic_item);
1002                 cfs_list_add(&imp->imp_conn_current->oic_item,
1003                              &imp->imp_conn_list);
1004                 imp->imp_last_success_conn =
1005                         imp->imp_conn_current->oic_last_attempt;
1006
1007                 spin_unlock(&imp->imp_lock);
1008
1009                 if (!ocd->ocd_ibits_known &&
1010                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
1011                         CERROR("Inodebits aware server returned zero compatible"
1012                                " bits?\n");
1013
1014                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1015                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
1016                                         LUSTRE_VERSION_OFFSET_WARN ||
1017                      ocd->ocd_version < LUSTRE_VERSION_CODE -
1018                                         LUSTRE_VERSION_OFFSET_WARN)) {
1019                         /* Sigh, some compilers do not like #ifdef in the middle
1020                            of macro arguments */
1021 #ifdef __KERNEL__
1022                         const char *older = "older. Consider upgrading server "
1023                                             "or downgrading client";
1024 #else
1025                         const char *older = "older. Consider recompiling this "
1026                                             "application";
1027 #endif
1028                         const char *newer = "newer than client version. "
1029                                             "Consider upgrading client";
1030
1031                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
1032                                       "is much %s (%s)\n",
1033                                       obd2cli_tgt(imp->imp_obd),
1034                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1035                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1036                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1037                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1038                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
1039                                       newer : older, LUSTRE_VERSION_STRING);
1040                 }
1041
1042 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1043                 /* Check if server has LU-1252 fix applied to not always swab
1044                  * the IR MNE entries. Do this only once per connection.  This
1045                  * fixup is version-limited, because we don't want to carry the
1046                  * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
1047                  * need interop with unpatched 2.2 servers.  For newer servers,
1048                  * the client will do MNE swabbing only as needed.  LU-1644 */
1049                 if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1050                              !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
1051                              OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
1052                              OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
1053                              OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
1054                              strcmp(imp->imp_obd->obd_type->typ_name,
1055                                     LUSTRE_MGC_NAME) == 0))
1056                         imp->imp_need_mne_swab = 1;
1057                 else /* clear if server was upgraded since last connect */
1058                         imp->imp_need_mne_swab = 0;
1059 #else
1060 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1061 #endif
1062
1063                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
1064                         /* We sent to the server ocd_cksum_types with bits set
1065                          * for algorithms we understand. The server masked off
1066                          * the checksum types it doesn't support */
1067                         if ((ocd->ocd_cksum_types &
1068                              cksum_types_supported_client()) == 0) {
1069                                 LCONSOLE_WARN("The negotiation of the checksum "
1070                                               "alogrithm to use with server %s "
1071                                               "failed (%x/%x), disabling "
1072                                               "checksums\n",
1073                                               obd2cli_tgt(imp->imp_obd),
1074                                               ocd->ocd_cksum_types,
1075                                               cksum_types_supported_client());
1076                                 cli->cl_checksum = 0;
1077                                 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
1078                         } else {
1079                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1080                         }
1081                 } else {
1082                         /* The server does not support OBD_CONNECT_CKSUM.
1083                          * Enforce ADLER for backward compatibility*/
1084                         cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
1085                 }
1086                 cli->cl_cksum_type =cksum_type_select(cli->cl_supp_cksum_types);
1087
1088                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
1089                         cli->cl_max_pages_per_rpc =
1090                                 min(ocd->ocd_brw_size >> PAGE_CACHE_SHIFT,
1091                                     cli->cl_max_pages_per_rpc);
1092                 else if (imp->imp_connect_op == MDS_CONNECT ||
1093                          imp->imp_connect_op == MGS_CONNECT)
1094                         cli->cl_max_pages_per_rpc = 1;
1095
1096                 /* Reset ns_connect_flags only for initial connect. It might be
1097                  * changed in while using FS and if we reset it in reconnect
1098                  * this leads to losing user settings done before such as
1099                  * disable lru_resize, etc. */
1100                 if (old_connect_flags != exp_connect_flags(exp) ||
1101                     aa->pcaa_initial_connect) {
1102                         CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
1103                                "flags: "LPX64"\n", imp->imp_obd->obd_name,
1104                               ocd->ocd_connect_flags);
1105                         imp->imp_obd->obd_namespace->ns_connect_flags =
1106                                 ocd->ocd_connect_flags;
1107                         imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1108                                 ocd->ocd_connect_flags;
1109                 }
1110
1111                 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1112                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1113                         /* We need a per-message support flag, because
1114                            a. we don't know if the incoming connect reply
1115                               supports AT or not (in reply_in_callback)
1116                               until we unpack it.
1117                            b. failovered server means export and flags are gone
1118                               (in ptlrpc_send_reply).
1119                            Can only be set when we know AT is supported at
1120                            both ends */
1121                         imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1122                 else
1123                         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1124
1125                 if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1126                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1127                         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1128                 else
1129                         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1130
1131                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1132                         (cli->cl_max_pages_per_rpc > 0));
1133         }
1134
1135 out:
1136         imp->imp_connect_tried = 1;
1137
1138         if (rc != 0) {
1139                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1140                 if (rc == -EACCES) {
1141                         /*
1142                          * Give up trying to reconnect
1143                          * EACCES means client has no permission for connection
1144                          */
1145                         imp->imp_obd->obd_no_recov = 1;
1146                         ptlrpc_deactivate_import(imp);
1147                 }
1148
1149                 if (rc == -EPROTO) {
1150                         struct obd_connect_data *ocd;
1151
1152                         /* reply message might not be ready */
1153                         if (request->rq_repmsg == NULL)
1154                                 RETURN(-EPROTO);
1155
1156                         ocd = req_capsule_server_get(&request->rq_pill,
1157                                                      &RMF_CONNECT_DATA);
1158                         if (ocd &&
1159                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1160                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1161                            /* Actually servers are only supposed to refuse
1162                               connection from liblustre clients, so we should
1163                               never see this from VFS context */
1164                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1165                                         "(%d.%d.%d.%d)"
1166                                         " refused connection from this client "
1167                                         "with an incompatible version (%s).  "
1168                                         "Client must be recompiled\n",
1169                                         obd2cli_tgt(imp->imp_obd),
1170                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1171                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1172                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1173                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1174                                         LUSTRE_VERSION_STRING);
1175                                 ptlrpc_deactivate_import(imp);
1176                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1177                         }
1178                         RETURN(-EPROTO);
1179                 }
1180
1181                 ptlrpc_maybe_ping_import_soon(imp);
1182
1183                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1184                        obd2cli_tgt(imp->imp_obd),
1185                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1186         }
1187
1188         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1189         RETURN(rc);
1190 }
1191
1192 /**
1193  * interpret callback for "completed replay" RPCs.
1194  * \see signal_completed_replay
1195  */
1196 static int completed_replay_interpret(const struct lu_env *env,
1197                                       struct ptlrpc_request *req,
1198                                       void * data, int rc)
1199 {
1200         ENTRY;
1201         cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
1202         if (req->rq_status == 0 &&
1203             !req->rq_import->imp_vbr_failed) {
1204                 ptlrpc_import_recovery_state_machine(req->rq_import);
1205         } else {
1206                 if (req->rq_import->imp_vbr_failed) {
1207                         CDEBUG(D_WARNING,
1208                                "%s: version recovery fails, reconnecting\n",
1209                                req->rq_import->imp_obd->obd_name);
1210                 } else {
1211                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1212                                      "reconnecting\n",
1213                                req->rq_import->imp_obd->obd_name,
1214                                req->rq_status);
1215                 }
1216                 ptlrpc_connect_import(req->rq_import);
1217         }
1218
1219         RETURN(0);
1220 }
1221
1222 /**
1223  * Let server know that we have no requests to replay anymore.
1224  * Achieved by just sending a PING request
1225  */
1226 static int signal_completed_replay(struct obd_import *imp)
1227 {
1228         struct ptlrpc_request *req;
1229         ENTRY;
1230
1231         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1232                 RETURN(0);
1233
1234         LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
1235         cfs_atomic_inc(&imp->imp_replay_inflight);
1236
1237         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1238                                         OBD_PING);
1239         if (req == NULL) {
1240                 cfs_atomic_dec(&imp->imp_replay_inflight);
1241                 RETURN(-ENOMEM);
1242         }
1243
1244         ptlrpc_request_set_replen(req);
1245         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1246         lustre_msg_add_flags(req->rq_reqmsg,
1247                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1248         if (AT_OFF)
1249                 req->rq_timeout *= 3;
1250         req->rq_interpret_reply = completed_replay_interpret;
1251
1252         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
1253         RETURN(0);
1254 }
1255
1256 #ifdef __KERNEL__
1257 /**
1258  * In kernel code all import invalidation happens in its own
1259  * separate thread, so that whatever application happened to encounter
1260  * a problem could still be killed or otherwise continue
1261  */
1262 static int ptlrpc_invalidate_import_thread(void *data)
1263 {
1264         struct obd_import *imp = data;
1265
1266         ENTRY;
1267
1268         unshare_fs_struct();
1269
1270         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1271                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1272                imp->imp_connection->c_remote_uuid.uuid);
1273
1274         ptlrpc_invalidate_import(imp);
1275
1276         if (obd_dump_on_eviction) {
1277                 CERROR("dump the log upon eviction\n");
1278                 libcfs_debug_dumplog();
1279         }
1280
1281         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1282         ptlrpc_import_recovery_state_machine(imp);
1283
1284         class_import_put(imp);
1285         RETURN(0);
1286 }
1287 #endif
1288
1289 /**
1290  * This is the state machine for client-side recovery on import.
1291  *
1292  * Typicaly we have two possibly paths. If we came to server and it is not
1293  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1294  * state and reconnect from scratch.
1295  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1296  * We go through our list of requests to replay and send them to server one by
1297  * one.
1298  * After sending all request from the list we change import state to
1299  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1300  * and also all the locks we don't yet have and wait for server to grant us.
1301  * After that we send a special "replay completed" request and change import
1302  * state to IMP_REPLAY_WAIT.
1303  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1304  * state and resend all requests from sending list.
1305  * After that we promote import to FULL state and send all delayed requests
1306  * and import is fully operational after that.
1307  *
1308  */
1309 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1310 {
1311         int rc = 0;
1312         int inflight;
1313         char *target_start;
1314         int target_len;
1315
1316         ENTRY;
1317         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1318                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1319                           &target_start, &target_len);
1320                 /* Don't care about MGC eviction */
1321                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1322                            LUSTRE_MGC_NAME) != 0) {
1323                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1324                                            "by %.*s; in progress operations "
1325                                            "using this service will fail.\n",
1326                                            imp->imp_obd->obd_name, target_len,
1327                                            target_start);
1328                 }
1329                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1330                        obd2cli_tgt(imp->imp_obd),
1331                        imp->imp_connection->c_remote_uuid.uuid);
1332                 /* reset vbr_failed flag upon eviction */
1333                 spin_lock(&imp->imp_lock);
1334                 imp->imp_vbr_failed = 0;
1335                 spin_unlock(&imp->imp_lock);
1336
1337 #ifdef __KERNEL__
1338                 {
1339                 cfs_task_t *task;
1340                 /* bug 17802:  XXX client_disconnect_export vs connect request
1341                  * race. if client will evicted at this time, we start
1342                  * invalidate thread without reference to import and import can
1343                  * be freed at same time. */
1344                 class_import_get(imp);
1345                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1346                                      "ll_imp_inval");
1347                 if (IS_ERR(task)) {
1348                         class_import_put(imp);
1349                         CERROR("error starting invalidate thread: %d\n", rc);
1350                         rc = PTR_ERR(task);
1351                 } else {
1352                         rc = 0;
1353                 }
1354                 RETURN(rc);
1355                 }
1356 #else
1357                 ptlrpc_invalidate_import(imp);
1358
1359                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1360 #endif
1361         }
1362
1363         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1364                 CDEBUG(D_HA, "replay requested by %s\n",
1365                        obd2cli_tgt(imp->imp_obd));
1366                 rc = ptlrpc_replay_next(imp, &inflight);
1367                 if (inflight == 0 &&
1368                     cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1369                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1370                         rc = ldlm_replay_locks(imp);
1371                         if (rc)
1372                                 GOTO(out, rc);
1373                 }
1374                 rc = 0;
1375         }
1376
1377         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1378                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1379                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1380                         rc = signal_completed_replay(imp);
1381                         if (rc)
1382                                 GOTO(out, rc);
1383                 }
1384
1385         }
1386
1387         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1388                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1389                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1390                 }
1391         }
1392
1393         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1394                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1395                        obd2cli_tgt(imp->imp_obd),
1396                        imp->imp_connection->c_remote_uuid.uuid);
1397
1398                 rc = ptlrpc_resend(imp);
1399                 if (rc)
1400                         GOTO(out, rc);
1401                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1402                 ptlrpc_activate_import(imp);
1403
1404                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1405                           &target_start, &target_len);
1406                 LCONSOLE_INFO("%s: Connection restored to %.*s (at %s)\n",
1407                               imp->imp_obd->obd_name,
1408                               target_len, target_start,
1409                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1410         }
1411
1412         if (imp->imp_state == LUSTRE_IMP_FULL) {
1413                 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1414                 ptlrpc_wake_delayed(imp);
1415         }
1416
1417 out:
1418         RETURN(rc);
1419 }
1420
1421 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1422 {
1423         struct ptlrpc_request *req;
1424         int rq_opc, rc = 0;
1425         int nowait = imp->imp_obd->obd_force;
1426         ENTRY;
1427
1428         if (nowait)
1429                 GOTO(set_state, rc);
1430
1431         switch (imp->imp_connect_op) {
1432         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1433         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1434         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1435         default:
1436                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1437                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1438                 RETURN(-EINVAL);
1439         }
1440
1441         if (ptlrpc_import_in_recovery(imp)) {
1442                 struct l_wait_info lwi;
1443                 cfs_duration_t timeout;
1444
1445
1446                 if (AT_OFF) {
1447                         if (imp->imp_server_timeout)
1448                                 timeout = cfs_time_seconds(obd_timeout / 2);
1449                         else
1450                                 timeout = cfs_time_seconds(obd_timeout);
1451                 } else {
1452                         int idx = import_at_get_index(imp,
1453                                 imp->imp_client->cli_request_portal);
1454                         timeout = cfs_time_seconds(
1455                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1456                 }
1457
1458                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1459                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1460                 rc = l_wait_event(imp->imp_recovery_waitq,
1461                                   !ptlrpc_import_in_recovery(imp), &lwi);
1462
1463         }
1464
1465         spin_lock(&imp->imp_lock);
1466         if (imp->imp_state != LUSTRE_IMP_FULL)
1467                 GOTO(out, 0);
1468
1469         spin_unlock(&imp->imp_lock);
1470
1471         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1472                                         LUSTRE_OBD_VERSION, rq_opc);
1473         if (req) {
1474                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1475                  * it fails.  We can get through the above with a down server
1476                  * if the client doesn't know the server is gone yet. */
1477                 req->rq_no_resend = 1;
1478
1479                 /* We want client umounts to happen quickly, no matter the
1480                    server state... */
1481                 req->rq_timeout = min_t(int, req->rq_timeout,
1482                                         INITIAL_CONNECT_TIMEOUT);
1483
1484                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1485                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1486                 ptlrpc_request_set_replen(req);
1487                 rc = ptlrpc_queue_wait(req);
1488                 ptlrpc_req_finished(req);
1489         }
1490
1491 set_state:
1492         spin_lock(&imp->imp_lock);
1493 out:
1494         if (noclose)
1495                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1496         else
1497                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1498         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1499         spin_unlock(&imp->imp_lock);
1500
1501         RETURN(rc);
1502 }
1503 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1504
1505 void ptlrpc_cleanup_imp(struct obd_import *imp)
1506 {
1507         ENTRY;
1508
1509         spin_lock(&imp->imp_lock);
1510         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1511         imp->imp_generation++;
1512         spin_unlock(&imp->imp_lock);
1513         ptlrpc_abort_inflight(imp);
1514
1515         EXIT;
1516 }
1517 EXPORT_SYMBOL(ptlrpc_cleanup_imp);
1518
1519 /* Adaptive Timeout utils */
1520 extern unsigned int at_min, at_max, at_history;
1521
1522 /* Bin into timeslices using AT_BINS bins.
1523    This gives us a max of the last binlimit*AT_BINS secs without the storage,
1524    but still smoothing out a return to normalcy from a slow response.
1525    (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1526 int at_measured(struct adaptive_timeout *at, unsigned int val)
1527 {
1528         unsigned int old = at->at_current;
1529         time_t now = cfs_time_current_sec();
1530         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1531
1532         LASSERT(at);
1533         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1534                val, at, now - at->at_binstart, at->at_current,
1535                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1536
1537         if (val == 0)
1538                 /* 0's don't count, because we never want our timeout to
1539                    drop to 0, and because 0 could mean an error */
1540                 return 0;
1541
1542         spin_lock(&at->at_lock);
1543
1544         if (unlikely(at->at_binstart == 0)) {
1545                 /* Special case to remove default from history */
1546                 at->at_current = val;
1547                 at->at_worst_ever = val;
1548                 at->at_worst_time = now;
1549                 at->at_hist[0] = val;
1550                 at->at_binstart = now;
1551         } else if (now - at->at_binstart < binlimit ) {
1552                 /* in bin 0 */
1553                 at->at_hist[0] = max(val, at->at_hist[0]);
1554                 at->at_current = max(val, at->at_current);
1555         } else {
1556                 int i, shift;
1557                 unsigned int maxv = val;
1558                 /* move bins over */
1559                 shift = (now - at->at_binstart) / binlimit;
1560                 LASSERT(shift > 0);
1561                 for(i = AT_BINS - 1; i >= 0; i--) {
1562                         if (i >= shift) {
1563                                 at->at_hist[i] = at->at_hist[i - shift];
1564                                 maxv = max(maxv, at->at_hist[i]);
1565                         } else {
1566                                 at->at_hist[i] = 0;
1567                         }
1568                 }
1569                 at->at_hist[0] = val;
1570                 at->at_current = maxv;
1571                 at->at_binstart += shift * binlimit;
1572         }
1573
1574         if (at->at_current > at->at_worst_ever) {
1575                 at->at_worst_ever = at->at_current;
1576                 at->at_worst_time = now;
1577         }
1578
1579         if (at->at_flags & AT_FLG_NOHIST)
1580                 /* Only keep last reported val; keeping the rest of the history
1581                    for proc only */
1582                 at->at_current = val;
1583
1584         if (at_max > 0)
1585                 at->at_current =  min(at->at_current, at_max);
1586         at->at_current =  max(at->at_current, at_min);
1587
1588         if (at->at_current != old)
1589                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1590                        "(val=%u) hist %u %u %u %u\n", at,
1591                        old, at->at_current, at->at_current - old, val,
1592                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1593                        at->at_hist[3]);
1594
1595         /* if we changed, report the old value */
1596         old = (at->at_current != old) ? old : 0;
1597
1598         spin_unlock(&at->at_lock);
1599         return old;
1600 }
1601
1602 /* Find the imp_at index for a given portal; assign if space available */
1603 int import_at_get_index(struct obd_import *imp, int portal)
1604 {
1605         struct imp_at *at = &imp->imp_at;
1606         int i;
1607
1608         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1609                 if (at->iat_portal[i] == portal)
1610                         return i;
1611                 if (at->iat_portal[i] == 0)
1612                         /* unused */
1613                         break;
1614         }
1615
1616         /* Not found in list, add it under a lock */
1617         spin_lock(&imp->imp_lock);
1618
1619         /* Check unused under lock */
1620         for (; i < IMP_AT_MAX_PORTALS; i++) {
1621                 if (at->iat_portal[i] == portal)
1622                         goto out;
1623                 if (at->iat_portal[i] == 0)
1624                         /* unused */
1625                         break;
1626         }
1627
1628         /* Not enough portals? */
1629         LASSERT(i < IMP_AT_MAX_PORTALS);
1630
1631         at->iat_portal[i] = portal;
1632 out:
1633         spin_unlock(&imp->imp_lock);
1634         return i;
1635 }