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_is_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         wake_up_all(&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         /* The requests in committed_list always have smaller transnos than
 574          * the requests in replay_list */
 575         if (!cfs_list_empty(&imp->imp_committed_list)) {
 576                 tmp = imp->imp_committed_list.next;
 577                 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 578                 *transno = req->rq_transno;
 579                 if (req->rq_transno == 0) {
 580                         DEBUG_REQ(D_ERROR, req, "zero transno in committed_list");
 581                         LBUG();
 582                 }
 583                 return 1;
 584         }
 585         if (!cfs_list_empty(&imp->imp_replay_list)) {
 586                 tmp = imp->imp_replay_list.next;
 587                 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 588                 *transno = req->rq_transno;
 589                 if (req->rq_transno == 0) {
 590                         DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
 591                         LBUG();
 592                 }
 593                 return 1;
 594         }
 595         return 0;
 596 }
 597
 598 /**
 599  * Attempt to (re)connect import \a imp. This includes all preparations,
 600  * initializing CONNECT RPC request and passing it to ptlrpcd for
 601  * actual sending.
 602  * Returns 0 on success or error code.
 603  */
 604 int ptlrpc_connect_import(struct obd_import *imp)
 605 {
 606         struct obd_device *obd = imp->imp_obd;
 607         int initial_connect = 0;
 608         int set_transno = 0;
 609         __u64 committed_before_reconnect = 0;
 610         struct ptlrpc_request *request;
 611         char *bufs[] = { NULL,
 612                          obd2cli_tgt(imp->imp_obd),
 613                          obd->obd_uuid.uuid,
 614                          (char *)&imp->imp_dlm_handle,
 615                          (char *)&imp->imp_connect_data };
 616         struct ptlrpc_connect_async_args *aa;
 617         int rc;
 618         ENTRY;
 619
 620         spin_lock(&imp->imp_lock);
 621         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 622                 spin_unlock(&imp->imp_lock);
 623                 CERROR("can't connect to a closed import\n");
 624                 RETURN(-EINVAL);
 625         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 626                 spin_unlock(&imp->imp_lock);
 627                 CERROR("already connected\n");
 628                 RETURN(0);
 629         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 630                 spin_unlock(&imp->imp_lock);
 631                 CERROR("already connecting\n");
 632                 RETURN(-EALREADY);
 633         }
 634
 635         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 636
 637         imp->imp_conn_cnt++;
 638         imp->imp_resend_replay = 0;
 639
 640         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 641                 initial_connect = 1;
 642         else
 643                 committed_before_reconnect = imp->imp_peer_committed_transno;
 644
 645         set_transno = ptlrpc_first_transno(imp,
 646                                            &imp->imp_connect_data.ocd_transno);
 647         spin_unlock(&imp->imp_lock);
 648
 649         rc = import_select_connection(imp);
 650         if (rc)
 651                 GOTO(out, rc);
 652
 653         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 654         if (rc)
 655                 GOTO(out, rc);
 656
 657         /* Reset connect flags to the originally requested flags, in case
 658          * the server is updated on-the-fly we will get the new features. */
 659         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 660         /* Reset ocd_version each time so the server knows the exact versions */
 661         imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
 662         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 663         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 664
 665         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 666                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 667         if (rc)
 668                 GOTO(out, rc);
 669
 670         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 671         if (request == NULL)
 672                 GOTO(out, rc = -ENOMEM);
 673
 674         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 675                                       imp->imp_connect_op, bufs, NULL);
 676         if (rc) {
 677                 ptlrpc_request_free(request);
 678                 GOTO(out, rc);
 679         }
 680
 681         /* Report the rpc service time to the server so that it knows how long
 682          * to wait for clients to join recovery */
 683         lustre_msg_set_service_time(request->rq_reqmsg,
 684                                     at_timeout2est(request->rq_timeout));
 685
 686         /* The amount of time we give the server to process the connect req.
 687          * import_select_connection will increase the net latency on
 688          * repeated reconnect attempts to cover slow networks.
 689          * We override/ignore the server rpc completion estimate here,
 690          * which may be large if this is a reconnect attempt */
 691         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 692         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 693
 694 #ifndef __KERNEL__
 695         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 696 #endif
 697         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 698
 699         request->rq_no_resend = request->rq_no_delay = 1;
 700         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 701         /* Allow a slightly larger reply for future growth compatibility */
 702         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 703                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 704         ptlrpc_request_set_replen(request);
 705         request->rq_interpret_reply = ptlrpc_connect_interpret;
 706
 707         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 708         aa = ptlrpc_req_async_args(request);
 709         memset(aa, 0, sizeof *aa);
 710
 711         aa->pcaa_peer_committed = committed_before_reconnect;
 712         aa->pcaa_initial_connect = initial_connect;
 713
 714         if (aa->pcaa_initial_connect) {
 715                 spin_lock(&imp->imp_lock);
 716                 imp->imp_replayable = 1;
 717                 spin_unlock(&imp->imp_lock);
 718                 lustre_msg_add_op_flags(request->rq_reqmsg,
 719                                         MSG_CONNECT_INITIAL);
 720         }
 721
 722         if (set_transno)
 723                 lustre_msg_add_op_flags(request->rq_reqmsg,
 724                                         MSG_CONNECT_TRANSNO);
 725
 726         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
 727                   request->rq_timeout);
 728         ptlrpcd_add_req(request, PDL_POLICY_ROUND, -1);
 729         rc = 0;
 730 out:
 731         if (rc != 0) {
 732                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 733         }
 734
 735         RETURN(rc);
 736 }
 737 EXPORT_SYMBOL(ptlrpc_connect_import);
 738
 739 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 740 {
 741 #ifdef __KERNEL__
 742         int force_verify;
 743
 744         spin_lock(&imp->imp_lock);
 745         force_verify = imp->imp_force_verify != 0;
 746         spin_unlock(&imp->imp_lock);
 747
 748         if (force_verify)
 749                 ptlrpc_pinger_wake_up();
 750 #else
 751         /* liblustre has no pinger thread, so we wakeup pinger anyway */
 752         ptlrpc_pinger_wake_up();
 753 #endif
 754 }
 755
 756 static int ptlrpc_busy_reconnect(int rc)
 757 {
 758         return (rc == -EBUSY) || (rc == -EAGAIN);
 759 }
 760
 761 /**
 762  * interpret_reply callback for connect RPCs.
 763  * Looks into returned status of connect operation and decides
 764  * what to do with the import - i.e enter recovery, promote it to
 765  * full state for normal operations of disconnect it due to an error.
 766  */
 767 static int ptlrpc_connect_interpret(const struct lu_env *env,
 768                                     struct ptlrpc_request *request,
 769                                     void *data, int rc)
 770 {
 771         struct ptlrpc_connect_async_args *aa = data;
 772         struct obd_import *imp = request->rq_import;
 773         struct client_obd *cli = &imp->imp_obd->u.cli;
 774         struct lustre_handle old_hdl;
 775         __u64 old_connect_flags;
 776         int msg_flags;
 777         struct obd_connect_data *ocd;
 778         struct obd_export *exp;
 779         int ret;
 780         ENTRY;
 781
 782         spin_lock(&imp->imp_lock);
 783         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 784                 imp->imp_connect_tried = 1;
 785                 spin_unlock(&imp->imp_lock);
 786                 RETURN(0);
 787         }
 788
 789         if (rc) {
 790                 /* if this reconnect to busy export - not need select new target
 791                  * for connecting*/
 792                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 793                 spin_unlock(&imp->imp_lock);
 794                 ptlrpc_maybe_ping_import_soon(imp);
 795                 GOTO(out, rc);
 796         }
 797         spin_unlock(&imp->imp_lock);
 798
 799         LASSERT(imp->imp_conn_current);
 800
 801         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 802
 803         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 804                                    RCL_SERVER);
 805         /* server replied obd_connect_data is always bigger */
 806         ocd = req_capsule_server_sized_get(&request->rq_pill,
 807                                            &RMF_CONNECT_DATA, ret);
 808
 809         if (ocd == NULL) {
 810                 CERROR("%s: no connect data from server\n",
 811                        imp->imp_obd->obd_name);
 812                 rc = -EPROTO;
 813                 GOTO(out, rc);
 814         }
 815
 816         spin_lock(&imp->imp_lock);
 817
 818         /* All imports are pingable */
 819         imp->imp_pingable = 1;
 820         imp->imp_force_reconnect = 0;
 821         imp->imp_force_verify = 0;
 822
 823         imp->imp_connect_data = *ocd;
 824
 825         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
 826                imp->imp_obd->obd_name, ocd->ocd_instance);
 827         exp = class_conn2export(&imp->imp_dlm_handle);
 828
 829         spin_unlock(&imp->imp_lock);
 830
 831         /* check that server granted subset of flags we asked for. */
 832         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
 833             ocd->ocd_connect_flags) {
 834                 CERROR("%s: Server didn't granted asked subset of flags: "
 835                        "asked="LPX64" grranted="LPX64"\n",
 836                        imp->imp_obd->obd_name,imp->imp_connect_flags_orig,
 837                        ocd->ocd_connect_flags);
 838                 GOTO(out, rc = -EPROTO);
 839         }
 840
 841         if (!exp) {
 842                 /* This could happen if export is cleaned during the
 843                    connect attempt */
 844                 CERROR("%s: missing export after connect\n",
 845                        imp->imp_obd->obd_name);
 846                 GOTO(out, rc = -ENODEV);
 847         }
 848         old_connect_flags = exp_connect_flags(exp);
 849         exp->exp_connect_data = *ocd;
 850         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
 851         class_export_put(exp);
 852
 853         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 854
 855         if (aa->pcaa_initial_connect) {
 856                 spin_lock(&imp->imp_lock);
 857                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 858                         imp->imp_replayable = 1;
 859                         spin_unlock(&imp->imp_lock);
 860                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 861                                obd2cli_tgt(imp->imp_obd));
 862                 } else {
 863                         imp->imp_replayable = 0;
 864                         spin_unlock(&imp->imp_lock);
 865                 }
 866
 867                 /* if applies, adjust the imp->imp_msg_magic here
 868                  * according to reply flags */
 869
 870                 imp->imp_remote_handle =
 871                                 *lustre_msg_get_handle(request->rq_repmsg);
 872
 873                 /* Initial connects are allowed for clients with non-random
 874                  * uuids when servers are in recovery.  Simply signal the
 875                  * servers replay is complete and wait in REPLAY_WAIT. */
 876                 if (msg_flags & MSG_CONNECT_RECOVERING) {
 877                         CDEBUG(D_HA, "connect to %s during recovery\n",
 878                                obd2cli_tgt(imp->imp_obd));
 879                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 880                 } else {
 881                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 882                         ptlrpc_activate_import(imp);
 883                 }
 884
 885                 GOTO(finish, rc = 0);
 886         }
 887
 888         /* Determine what recovery state to move the import to. */
 889         if (MSG_CONNECT_RECONNECT & msg_flags) {
 890                 memset(&old_hdl, 0, sizeof(old_hdl));
 891                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 892                             sizeof (old_hdl))) {
 893                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
 894                                       "bad handle "LPX64"\n",
 895                                       obd2cli_tgt(imp->imp_obd),
 896                                       imp->imp_connection->c_remote_uuid.uuid,
 897                                       imp->imp_dlm_handle.cookie);
 898                         GOTO(out, rc = -ENOTCONN);
 899                 }
 900
 901                 if (memcmp(&imp->imp_remote_handle,
 902                            lustre_msg_get_handle(request->rq_repmsg),
 903                            sizeof(imp->imp_remote_handle))) {
 904                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
 905                                 D_HA : D_WARNING;
 906
 907                         /* Bug 16611/14775: if server handle have changed,
 908                          * that means some sort of disconnection happened.
 909                          * If the server is not in recovery, that also means it
 910                          * already erased all of our state because of previous
 911                          * eviction. If it is in recovery - we are safe to
 912                          * participate since we can reestablish all of our state
 913                          * with server again */
 914                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
 915                                 CDEBUG(level,"%s@%s changed server handle from "
 916                                        LPX64" to "LPX64
 917                                        " but is still in recovery\n",
 918                                        obd2cli_tgt(imp->imp_obd),
 919                                        imp->imp_connection->c_remote_uuid.uuid,
 920                                        imp->imp_remote_handle.cookie,
 921                                        lustre_msg_get_handle(
 922                                        request->rq_repmsg)->cookie);
 923                         } else {
 924                                 LCONSOLE_WARN("Evicted from %s (at %s) "
 925                                               "after server handle changed from "
 926                                               LPX64" to "LPX64"\n",
 927                                               obd2cli_tgt(imp->imp_obd),
 928                                               imp->imp_connection-> \
 929                                               c_remote_uuid.uuid,
 930                                               imp->imp_remote_handle.cookie,
 931                                               lustre_msg_get_handle(
 932                                               request->rq_repmsg)->cookie);
 933                         }
 934
 935
 936                         imp->imp_remote_handle =
 937                                      *lustre_msg_get_handle(request->rq_repmsg);
 938
 939                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
 940                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 941                                 GOTO(finish, rc = 0);
 942                         }
 943
 944                 } else {
 945                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 946                                obd2cli_tgt(imp->imp_obd),
 947                                imp->imp_connection->c_remote_uuid.uuid);
 948                 }
 949
 950                 if (imp->imp_invalid) {
 951                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
 952                                "marking evicted\n", imp->imp_obd->obd_name);
 953                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 954                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 955                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 956                                imp->imp_obd->obd_name,
 957                                obd2cli_tgt(imp->imp_obd));
 958
 959                         spin_lock(&imp->imp_lock);
 960                         imp->imp_resend_replay = 1;
 961                         spin_unlock(&imp->imp_lock);
 962
 963                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 964                 } else {
 965                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 966                 }
 967         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 968                 LASSERT(imp->imp_replayable);
 969                 imp->imp_remote_handle =
 970                                 *lustre_msg_get_handle(request->rq_repmsg);
 971                 imp->imp_last_replay_transno = 0;
 972                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 973         } else {
 974                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
 975                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
 976                 imp->imp_remote_handle =
 977                                 *lustre_msg_get_handle(request->rq_repmsg);
 978                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 979         }
 980
 981         /* Sanity checks for a reconnected import. */
 982         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 983                 CERROR("imp_replayable flag does not match server "
 984                        "after reconnect. We should LBUG right here.\n");
 985         }
 986
 987         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
 988             lustre_msg_get_last_committed(request->rq_repmsg) <
 989             aa->pcaa_peer_committed) {
 990                 CERROR("%s went back in time (transno "LPD64
 991                        " was previously committed, server now claims "LPD64
 992                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
 993                        "id=9646\n",
 994                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 995                        lustre_msg_get_last_committed(request->rq_repmsg));
 996         }
 997
 998 finish:
 999         rc = ptlrpc_import_recovery_state_machine(imp);
1000         if (rc != 0) {
1001                 if (rc == -ENOTCONN) {
1002                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1003                                "invalidating and reconnecting\n",
1004                                obd2cli_tgt(imp->imp_obd),
1005                                imp->imp_connection->c_remote_uuid.uuid);
1006                         ptlrpc_connect_import(imp);
1007                         imp->imp_connect_tried = 1;
1008                         RETURN(0);
1009                 }
1010         } else {
1011                 static bool warned;
1012
1013                 spin_lock(&imp->imp_lock);
1014                 cfs_list_del(&imp->imp_conn_current->oic_item);
1015                 cfs_list_add(&imp->imp_conn_current->oic_item,
1016                              &imp->imp_conn_list);
1017                 imp->imp_last_success_conn =
1018                         imp->imp_conn_current->oic_last_attempt;
1019
1020                 spin_unlock(&imp->imp_lock);
1021
1022                 if (!ocd->ocd_ibits_known &&
1023                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
1024                         CERROR("Inodebits aware server returned zero compatible"
1025                                " bits?\n");
1026
1027                 if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1028                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
1029                                         LUSTRE_VERSION_OFFSET_WARN ||
1030                      ocd->ocd_version < LUSTRE_VERSION_CODE -
1031                                         LUSTRE_VERSION_OFFSET_WARN)) {
1032                         /* Sigh, some compilers do not like #ifdef in the middle
1033                            of macro arguments */
1034                         const char *older = "older than client. "
1035                                             "Consider upgrading server";
1036 #ifdef __KERNEL__
1037                         const char *newer = "newer than client. "
1038                                             "Consider recompiling application";
1039 #else
1040                         const char *newer = "newer than client. "
1041                                             "Consider upgrading client";
1042 #endif
1043
1044                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
1045                                       "is much %s (%s)\n",
1046                                       obd2cli_tgt(imp->imp_obd),
1047                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1048                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1049                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1050                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1051                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
1052                                       newer : older, LUSTRE_VERSION_STRING);
1053                         warned = true;
1054                 }
1055
1056 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1057                 /* Check if server has LU-1252 fix applied to not always swab
1058                  * the IR MNE entries. Do this only once per connection.  This
1059                  * fixup is version-limited, because we don't want to carry the
1060                  * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
1061                  * need interop with unpatched 2.2 servers.  For newer servers,
1062                  * the client will do MNE swabbing only as needed.  LU-1644 */
1063                 if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1064                              !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
1065                              OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
1066                              OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
1067                              OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
1068                              strcmp(imp->imp_obd->obd_type->typ_name,
1069                                     LUSTRE_MGC_NAME) == 0))
1070                         imp->imp_need_mne_swab = 1;
1071                 else /* clear if server was upgraded since last connect */
1072                         imp->imp_need_mne_swab = 0;
1073 #else
1074 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1075 #endif
1076
1077                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
1078                         /* We sent to the server ocd_cksum_types with bits set
1079                          * for algorithms we understand. The server masked off
1080                          * the checksum types it doesn't support */
1081                         if ((ocd->ocd_cksum_types &
1082                              cksum_types_supported_client()) == 0) {
1083                                 LCONSOLE_WARN("The negotiation of the checksum "
1084                                               "alogrithm to use with server %s "
1085                                               "failed (%x/%x), disabling "
1086                                               "checksums\n",
1087                                               obd2cli_tgt(imp->imp_obd),
1088                                               ocd->ocd_cksum_types,
1089                                               cksum_types_supported_client());
1090                                 cli->cl_checksum = 0;
1091                                 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
1092                         } else {
1093                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1094                         }
1095                 } else {
1096                         /* The server does not support OBD_CONNECT_CKSUM.
1097                          * Enforce ADLER for backward compatibility*/
1098                         cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
1099                 }
1100                 cli->cl_cksum_type =cksum_type_select(cli->cl_supp_cksum_types);
1101
1102                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
1103                         cli->cl_max_pages_per_rpc =
1104                                 min(ocd->ocd_brw_size >> PAGE_CACHE_SHIFT,
1105                                     cli->cl_max_pages_per_rpc);
1106                 else if (imp->imp_connect_op == MDS_CONNECT ||
1107                          imp->imp_connect_op == MGS_CONNECT)
1108                         cli->cl_max_pages_per_rpc = 1;
1109
1110                 /* Reset ns_connect_flags only for initial connect. It might be
1111                  * changed in while using FS and if we reset it in reconnect
1112                  * this leads to losing user settings done before such as
1113                  * disable lru_resize, etc. */
1114                 if (old_connect_flags != exp_connect_flags(exp) ||
1115                     aa->pcaa_initial_connect) {
1116                         CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
1117                                "flags: "LPX64"\n", imp->imp_obd->obd_name,
1118                               ocd->ocd_connect_flags);
1119                         imp->imp_obd->obd_namespace->ns_connect_flags =
1120                                 ocd->ocd_connect_flags;
1121                         imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1122                                 ocd->ocd_connect_flags;
1123                 }
1124
1125                 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1126                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1127                         /* We need a per-message support flag, because
1128                            a. we don't know if the incoming connect reply
1129                               supports AT or not (in reply_in_callback)
1130                               until we unpack it.
1131                            b. failovered server means export and flags are gone
1132                               (in ptlrpc_send_reply).
1133                            Can only be set when we know AT is supported at
1134                            both ends */
1135                         imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1136                 else
1137                         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1138
1139                 if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1140                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1141                         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1142                 else
1143                         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1144
1145                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1146                         (cli->cl_max_pages_per_rpc > 0));
1147         }
1148
1149 out:
1150         imp->imp_connect_tried = 1;
1151
1152         if (rc != 0) {
1153                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1154                 if (rc == -EACCES) {
1155                         /*
1156                          * Give up trying to reconnect
1157                          * EACCES means client has no permission for connection
1158                          */
1159                         imp->imp_obd->obd_no_recov = 1;
1160                         ptlrpc_deactivate_import(imp);
1161                 }
1162
1163                 if (rc == -EPROTO) {
1164                         struct obd_connect_data *ocd;
1165
1166                         /* reply message might not be ready */
1167                         if (request->rq_repmsg == NULL)
1168                                 RETURN(-EPROTO);
1169
1170                         ocd = req_capsule_server_get(&request->rq_pill,
1171                                                      &RMF_CONNECT_DATA);
1172                         if (ocd &&
1173                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1174                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1175                            /* Actually servers are only supposed to refuse
1176                               connection from liblustre clients, so we should
1177                               never see this from VFS context */
1178                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1179                                         "(%d.%d.%d.%d)"
1180                                         " refused connection from this client "
1181                                         "with an incompatible version (%s).  "
1182                                         "Client must be recompiled\n",
1183                                         obd2cli_tgt(imp->imp_obd),
1184                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1185                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1186                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1187                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1188                                         LUSTRE_VERSION_STRING);
1189                                 ptlrpc_deactivate_import(imp);
1190                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1191                         }
1192                         RETURN(-EPROTO);
1193                 }
1194
1195                 ptlrpc_maybe_ping_import_soon(imp);
1196
1197                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1198                        obd2cli_tgt(imp->imp_obd),
1199                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1200         }
1201
1202         wake_up_all(&imp->imp_recovery_waitq);
1203         RETURN(rc);
1204 }
1205
1206 /**
1207  * interpret callback for "completed replay" RPCs.
1208  * \see signal_completed_replay
1209  */
1210 static int completed_replay_interpret(const struct lu_env *env,
1211                                       struct ptlrpc_request *req,
1212                                       void * data, int rc)
1213 {
1214         ENTRY;
1215         cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
1216         if (req->rq_status == 0 &&
1217             !req->rq_import->imp_vbr_failed) {
1218                 ptlrpc_import_recovery_state_machine(req->rq_import);
1219         } else {
1220                 if (req->rq_import->imp_vbr_failed) {
1221                         CDEBUG(D_WARNING,
1222                                "%s: version recovery fails, reconnecting\n",
1223                                req->rq_import->imp_obd->obd_name);
1224                 } else {
1225                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1226                                      "reconnecting\n",
1227                                req->rq_import->imp_obd->obd_name,
1228                                req->rq_status);
1229                 }
1230                 ptlrpc_connect_import(req->rq_import);
1231         }
1232
1233         RETURN(0);
1234 }
1235
1236 /**
1237  * Let server know that we have no requests to replay anymore.
1238  * Achieved by just sending a PING request
1239  */
1240 static int signal_completed_replay(struct obd_import *imp)
1241 {
1242         struct ptlrpc_request *req;
1243         ENTRY;
1244
1245         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1246                 RETURN(0);
1247
1248         LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
1249         cfs_atomic_inc(&imp->imp_replay_inflight);
1250
1251         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1252                                         OBD_PING);
1253         if (req == NULL) {
1254                 cfs_atomic_dec(&imp->imp_replay_inflight);
1255                 RETURN(-ENOMEM);
1256         }
1257
1258         ptlrpc_request_set_replen(req);
1259         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1260         lustre_msg_add_flags(req->rq_reqmsg,
1261                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1262         if (AT_OFF)
1263                 req->rq_timeout *= 3;
1264         req->rq_interpret_reply = completed_replay_interpret;
1265
1266         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
1267         RETURN(0);
1268 }
1269
1270 #ifdef __KERNEL__
1271 /**
1272  * In kernel code all import invalidation happens in its own
1273  * separate thread, so that whatever application happened to encounter
1274  * a problem could still be killed or otherwise continue
1275  */
1276 static int ptlrpc_invalidate_import_thread(void *data)
1277 {
1278         struct obd_import *imp = data;
1279
1280         ENTRY;
1281
1282         unshare_fs_struct();
1283
1284         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1285                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1286                imp->imp_connection->c_remote_uuid.uuid);
1287
1288         ptlrpc_invalidate_import(imp);
1289
1290         if (obd_dump_on_eviction) {
1291                 CERROR("dump the log upon eviction\n");
1292                 libcfs_debug_dumplog();
1293         }
1294
1295         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1296         ptlrpc_import_recovery_state_machine(imp);
1297
1298         class_import_put(imp);
1299         RETURN(0);
1300 }
1301 #endif
1302
1303 /**
1304  * This is the state machine for client-side recovery on import.
1305  *
1306  * Typicaly we have two possibly paths. If we came to server and it is not
1307  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1308  * state and reconnect from scratch.
1309  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1310  * We go through our list of requests to replay and send them to server one by
1311  * one.
1312  * After sending all request from the list we change import state to
1313  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1314  * and also all the locks we don't yet have and wait for server to grant us.
1315  * After that we send a special "replay completed" request and change import
1316  * state to IMP_REPLAY_WAIT.
1317  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1318  * state and resend all requests from sending list.
1319  * After that we promote import to FULL state and send all delayed requests
1320  * and import is fully operational after that.
1321  *
1322  */
1323 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1324 {
1325         int rc = 0;
1326         int inflight;
1327         char *target_start;
1328         int target_len;
1329
1330         ENTRY;
1331         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1332                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1333                           &target_start, &target_len);
1334                 /* Don't care about MGC eviction */
1335                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1336                            LUSTRE_MGC_NAME) != 0) {
1337                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1338                                            "by %.*s; in progress operations "
1339                                            "using this service will fail.\n",
1340                                            imp->imp_obd->obd_name, target_len,
1341                                            target_start);
1342                 }
1343                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1344                        obd2cli_tgt(imp->imp_obd),
1345                        imp->imp_connection->c_remote_uuid.uuid);
1346                 /* reset vbr_failed flag upon eviction */
1347                 spin_lock(&imp->imp_lock);
1348                 imp->imp_vbr_failed = 0;
1349                 spin_unlock(&imp->imp_lock);
1350
1351 #ifdef __KERNEL__
1352                 {
1353                 struct task_struct *task;
1354                 /* bug 17802:  XXX client_disconnect_export vs connect request
1355                  * race. if client will evicted at this time, we start
1356                  * invalidate thread without reference to import and import can
1357                  * be freed at same time. */
1358                 class_import_get(imp);
1359                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1360                                      "ll_imp_inval");
1361                 if (IS_ERR(task)) {
1362                         class_import_put(imp);
1363                         CERROR("error starting invalidate thread: %d\n", rc);
1364                         rc = PTR_ERR(task);
1365                 } else {
1366                         rc = 0;
1367                 }
1368                 RETURN(rc);
1369                 }
1370 #else
1371                 ptlrpc_invalidate_import(imp);
1372
1373                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1374 #endif
1375         }
1376
1377         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1378                 CDEBUG(D_HA, "replay requested by %s\n",
1379                        obd2cli_tgt(imp->imp_obd));
1380                 rc = ptlrpc_replay_next(imp, &inflight);
1381                 if (inflight == 0 &&
1382                     cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1383                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1384                         rc = ldlm_replay_locks(imp);
1385                         if (rc)
1386                                 GOTO(out, rc);
1387                 }
1388                 rc = 0;
1389         }
1390
1391         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1392                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1393                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1394                         rc = signal_completed_replay(imp);
1395                         if (rc)
1396                                 GOTO(out, rc);
1397                 }
1398
1399         }
1400
1401         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1402                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1403                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1404                 }
1405         }
1406
1407         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1408                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1409                        obd2cli_tgt(imp->imp_obd),
1410                        imp->imp_connection->c_remote_uuid.uuid);
1411
1412                 rc = ptlrpc_resend(imp);
1413                 if (rc)
1414                         GOTO(out, rc);
1415                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1416                 ptlrpc_activate_import(imp);
1417
1418                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1419                           &target_start, &target_len);
1420                 LCONSOLE_INFO("%s: Connection restored to %.*s (at %s)\n",
1421                               imp->imp_obd->obd_name,
1422                               target_len, target_start,
1423                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1424         }
1425
1426         if (imp->imp_state == LUSTRE_IMP_FULL) {
1427                 wake_up_all(&imp->imp_recovery_waitq);
1428                 ptlrpc_wake_delayed(imp);
1429         }
1430
1431 out:
1432         RETURN(rc);
1433 }
1434
1435 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1436 {
1437         struct ptlrpc_request *req;
1438         int rq_opc, rc = 0;
1439         int nowait = imp->imp_obd->obd_force;
1440         ENTRY;
1441
1442         if (nowait)
1443                 GOTO(set_state, rc);
1444
1445         switch (imp->imp_connect_op) {
1446         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1447         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1448         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1449         default:
1450                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1451                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1452                 RETURN(-EINVAL);
1453         }
1454
1455         if (ptlrpc_import_in_recovery(imp)) {
1456                 struct l_wait_info lwi;
1457                 cfs_duration_t timeout;
1458
1459
1460                 if (AT_OFF) {
1461                         if (imp->imp_server_timeout)
1462                                 timeout = cfs_time_seconds(obd_timeout / 2);
1463                         else
1464                                 timeout = cfs_time_seconds(obd_timeout);
1465                 } else {
1466                         int idx = import_at_get_index(imp,
1467                                 imp->imp_client->cli_request_portal);
1468                         timeout = cfs_time_seconds(
1469                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1470                 }
1471
1472                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1473                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1474                 rc = l_wait_event(imp->imp_recovery_waitq,
1475                                   !ptlrpc_import_in_recovery(imp), &lwi);
1476
1477         }
1478
1479         spin_lock(&imp->imp_lock);
1480         if (imp->imp_state != LUSTRE_IMP_FULL)
1481                 GOTO(out, rc);
1482
1483         spin_unlock(&imp->imp_lock);
1484
1485         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1486                                         LUSTRE_OBD_VERSION, rq_opc);
1487         if (req) {
1488                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1489                  * it fails.  We can get through the above with a down server
1490                  * if the client doesn't know the server is gone yet. */
1491                 req->rq_no_resend = 1;
1492
1493                 /* We want client umounts to happen quickly, no matter the
1494                    server state... */
1495                 req->rq_timeout = min_t(int, req->rq_timeout,
1496                                         INITIAL_CONNECT_TIMEOUT);
1497
1498                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1499                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1500                 ptlrpc_request_set_replen(req);
1501                 rc = ptlrpc_queue_wait(req);
1502                 ptlrpc_req_finished(req);
1503         }
1504
1505 set_state:
1506         spin_lock(&imp->imp_lock);
1507 out:
1508         if (noclose)
1509                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1510         else
1511                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1512         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1513         spin_unlock(&imp->imp_lock);
1514
1515         RETURN(rc);
1516 }
1517 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1518
1519 void ptlrpc_cleanup_imp(struct obd_import *imp)
1520 {
1521         ENTRY;
1522
1523         spin_lock(&imp->imp_lock);
1524         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1525         imp->imp_generation++;
1526         spin_unlock(&imp->imp_lock);
1527         ptlrpc_abort_inflight(imp);
1528
1529         EXIT;
1530 }
1531 EXPORT_SYMBOL(ptlrpc_cleanup_imp);
1532
1533 /* Adaptive Timeout utils */
1534 extern unsigned int at_min, at_max, at_history;
1535
1536 /* Bin into timeslices using AT_BINS bins.
1537    This gives us a max of the last binlimit*AT_BINS secs without the storage,
1538    but still smoothing out a return to normalcy from a slow response.
1539    (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1540 int at_measured(struct adaptive_timeout *at, unsigned int val)
1541 {
1542         unsigned int old = at->at_current;
1543         time_t now = cfs_time_current_sec();
1544         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1545
1546         LASSERT(at);
1547         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1548                val, at, now - at->at_binstart, at->at_current,
1549                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1550
1551         if (val == 0)
1552                 /* 0's don't count, because we never want our timeout to
1553                    drop to 0, and because 0 could mean an error */
1554                 return 0;
1555
1556         spin_lock(&at->at_lock);
1557
1558         if (unlikely(at->at_binstart == 0)) {
1559                 /* Special case to remove default from history */
1560                 at->at_current = val;
1561                 at->at_worst_ever = val;
1562                 at->at_worst_time = now;
1563                 at->at_hist[0] = val;
1564                 at->at_binstart = now;
1565         } else if (now - at->at_binstart < binlimit ) {
1566                 /* in bin 0 */
1567                 at->at_hist[0] = max(val, at->at_hist[0]);
1568                 at->at_current = max(val, at->at_current);
1569         } else {
1570                 int i, shift;
1571                 unsigned int maxv = val;
1572                 /* move bins over */
1573                 shift = (now - at->at_binstart) / binlimit;
1574                 LASSERT(shift > 0);
1575                 for(i = AT_BINS - 1; i >= 0; i--) {
1576                         if (i >= shift) {
1577                                 at->at_hist[i] = at->at_hist[i - shift];
1578                                 maxv = max(maxv, at->at_hist[i]);
1579                         } else {
1580                                 at->at_hist[i] = 0;
1581                         }
1582                 }
1583                 at->at_hist[0] = val;
1584                 at->at_current = maxv;
1585                 at->at_binstart += shift * binlimit;
1586         }
1587
1588         if (at->at_current > at->at_worst_ever) {
1589                 at->at_worst_ever = at->at_current;
1590                 at->at_worst_time = now;
1591         }
1592
1593         if (at->at_flags & AT_FLG_NOHIST)
1594                 /* Only keep last reported val; keeping the rest of the history
1595                    for proc only */
1596                 at->at_current = val;
1597
1598         if (at_max > 0)
1599                 at->at_current =  min(at->at_current, at_max);
1600         at->at_current =  max(at->at_current, at_min);
1601
1602         if (at->at_current != old)
1603                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1604                        "(val=%u) hist %u %u %u %u\n", at,
1605                        old, at->at_current, at->at_current - old, val,
1606                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1607                        at->at_hist[3]);
1608
1609         /* if we changed, report the old value */
1610         old = (at->at_current != old) ? old : 0;
1611
1612         spin_unlock(&at->at_lock);
1613         return old;
1614 }
1615
1616 /* Find the imp_at index for a given portal; assign if space available */
1617 int import_at_get_index(struct obd_import *imp, int portal)
1618 {
1619         struct imp_at *at = &imp->imp_at;
1620         int i;
1621
1622         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1623                 if (at->iat_portal[i] == portal)
1624                         return i;
1625                 if (at->iat_portal[i] == 0)
1626                         /* unused */
1627                         break;
1628         }
1629
1630         /* Not found in list, add it under a lock */
1631         spin_lock(&imp->imp_lock);
1632
1633         /* Check unused under lock */
1634         for (; i < IMP_AT_MAX_PORTALS; i++) {
1635                 if (at->iat_portal[i] == portal)
1636                         goto out;
1637                 if (at->iat_portal[i] == 0)
1638                         /* unused */
1639                         break;
1640         }
1641
1642         /* Not enough portals? */
1643         LASSERT(i < IMP_AT_MAX_PORTALS);
1644
1645         at->iat_portal[i] = portal;
1646 out:
1647         spin_unlock(&imp->imp_lock);
1648         return i;
1649 }